Search

Traffic flow prediction remains an ongoing hot topic in the field of Intelligent Transportation System. The state-of-the-art traffic flow prediction models can effectively extract both spatial and temporal features of traffic flow data, but ignore the correlation and external interference between traffic nodes. To this end, this paper proposes a novel method based on Spatial–Temporal Complex Graph Convolution Network (ST-CGCN) for traffic flow prediction. Specifically, we first constructs the distance matrix, the data correlation matrix, and the comfort measurement matrix according to the geographical locations, the historical data record, and the external interference between traffic nodes. Then, these three matrices are fused into a complex correlation matrix by introducing self-learning dynamic weights to improve the joint modeling ability of spatial–temporal features and external factors. Next, a spatial feature extraction module and a temporal feature extraction module are designed to characterize dynamic spatial–temporal features. The spatial feature extraction module consists of a graph convolution operator with a proposed complex correlation matrix and a residual unit. The temporal feature extraction module consists of a 3D convolution operator and a Long Short-Term Memory (LSTM). Experiments constructed on five real-world datasets demonstrate that the new proposed ST-CGCN is more effective than several existing deep learning based traffic flow prediction models. The key source code and data are available at https://github.com/Bounger2/ST-CGCN. [ABSTRACT FROM AUTHOR]

The development of an automated rail line defect classification system is of great benefit, as railway tracks must be periodically monitored and inspected to guarantee the safety of rail transportation. In this paper, an effective multi-scale residual convolutional network (MSRConvNet) model is proposed to classify the different types of railway track defects. The skip connections with residual learning blocks are used to increase the effectiveness of the network. The multi-scale convolutions are connected with parallel and two skip connections in the structure to distribute detailed feature maps with each other. Therefore, different scale feature maps can be extracted. The data augmentation method is performed to ensure a balanced class distribution and to eliminate the negative effect of the imbalanced dataset. The proposed model is compared with both benchmark deep learning models and the different variations of the designed network. The results verify that the proposed model can reach superior classification fulfillment, and the MSRConvNet provides an overall accuracy of 99.83%, precision of 99.83%, sensitivity of 99.83%, specificity of 99.94%, F1-score of 99.83%, and Matthew's correlation coefficient of 99.78% for four defect classes. • Multi-scale residual network is designed for railway track defect classification. • Hierarchical features are extracted with different level kernels. • Data augmentation is used to increase network performance. • Obtained results are compared with variations of multi-scale structures. [ABSTRACT FROM AUTHOR]

Electric vehicles (EVs) are considered emerging popularity due to their cost-saving and eco-friendliness nature. The charging of EVs with elevated efficiency is a major issue for users as the hardware facilities, like charging stations are limited. Thus, the EV charging scheduling model is required to discover the charging schedule for each EV. An optimization-driven framework for EV charge scheduling in a Vehicular ad hoc network (VANET) topology is presented in this paper. First, the network simulation with the charge station (CS) and the EV locations are performed. The Deep Maxout network (DMN) is employed for predicting the power. To charge at the charging station, the vehicles are scheduled based on various parametric factors. The proposed Fractional Feedback Tree Algorithm (FFTA) is used to schedule charges in the EV network. The Feedback Artificial Tree (FAT) algorithm and fractional calculus (FC) are combined to create the proposed FFTA. Additionally, a new model of the fitness function is created taking into account variables like the average waiting time, the distance, the predicted power, and the quantity of EVs that are requested to be charged. With the shortest distance of 11.00 km, the highest power of 8.17 J, the shortest average waiting time of 0.47 s, and less number of EVs charged at 4.4, the proposed FFTA offered improved performance. [Display omitted] [ABSTRACT FROM AUTHOR]

Accurate multi-node load forecasting is the key to the safe, reliable, and economical operation of the power system. However, the dynamic nature of load and the coupling nature of networks are difficult to extract, making consistent and accurate forecasting of node load rather difficult. In this regard, this paper proposes a soft sharing multi-task deep learning method for multi-node load forecasting in the power system. It has the following aspects: (1) Considering the coupling characteristics of the node network, a multi-modal feature module, based on the inception strategy and gated temporal convolutional network (GTCN), is firstly designed to explore the coupling features implied in the node load data. (2) A novel multi-objective neural network model is proposed to achieve simultaneous prediction of multi-node load by integrating the multi-modal feature module and gated recurrent unit (GRU). For sharing the learning information of sub-networks, this paper uses the soft sharing mechanism to capture load features, which can better optimize the prediction task for each node load simultaneously. Load data from the New Zealand distribution network and AEMO are used to compare the proposed model's performance in various scenarios using regression metrics such as mean absolute percentage error (MAPE), Weighted Mean Accuracy (WMA), root mean squared logarithmic error (RMSLE), and Diebold–Mariano (DM). The simulation results show that the proposed method can explore the spatial–temporal coupling characteristics in multi-node load data. Compared with existing state-of-the-art multi-node load prediction methods, our proposed method's MAPE decrease 17.04% and 3.92% in Non-aggregation and Aggregation situations. • A novel deep multi-task learning method is proposed for the multi-node forecasting problem. • An IGTCN module is designed to extract the coupling features from different node loads. • The multi-task deep neural network exhibits the accuracy of multi-node load forecasting. [ABSTRACT FROM AUTHOR]

To estimate the parameters of an induction motor in a data-based manner, this paper proposes a new offline method to estimate rotor resistance and excitation inductance based on the deep-Q-learning approach. In this method, parameter estimation can be facilitated without disturbing the model error or operating state. To achieve this goal, three important elements, namely observation, action, and reward, are appropriately designed. To improve the robustness and accelerate the convergence, a new concept, denoted as Q-sensitivity, is proposed and investigated in detail. The experimental results show that a high-Q-sensitivity design can allow the proposed method to obtain a fast and torque-maximized estimation. Results from the comparative studies confirm the accuracy and robustness of the proposed method. [ABSTRACT FROM AUTHOR]

After 2007–2008 crisis, it is clear that corporate credit scoring is becoming a key role in credit risk management. In this paper, we investigate the performances of credit scoring models applied to CDS data sets. The classification performance of deep learning algorithm such as deep belief networks with Restricted Boltzmann Machines are evaluated and compared with some popular credit scoring models such as logistic regression, multi-layer perceptron and support vector machine. The performance is assessed using the classification accuracy and the area under the receiver operating characteristic curve. It is found that DBN yields the best performance. [ABSTRACT FROM AUTHOR]

Since the generative adversarial network (GAN) was proposed in 2014, it has rapidly become a hot topic in the field of deep learning. In recent years, there are many optimizations for GAN, which are divided into three kinds, including the optimization of loss functions, external structure of network and internal structure of network. Few people optimizes GAN from internal structure of network, and because of the process of game, slow training speed is one of the biggest problems of GAN. This paper introduces the ideology of broad learning algorithm (Chen and Liu, 2017) to put forward the multi-judge generative adversarial network method based on different random features to enhance the quality of generative result and increase the training speed of GAN (Goodfellow et al., 2014). This model provides a method which reduce the input information of each discriminator to optimize the training process of GAN. The experiments on cifar10 and anime face dataset find that our model obtains a better performance than the GMAN model and Base model. This paper finds a group of hyper-parameters to enhance the BroadGAN. [ABSTRACT FROM AUTHOR]

The paper proposes a surrogate model for liquid moulding of structural composites. A methodology is presented to simulate the dual-phase Darcy's flow in a heterogeneous porous medium. The approach is an encoder–decoder that receives as input a matrix of permeabilities and produces two scalar fields that represent the pressure and front flow. This model is trained with synthetic data generated with a computer fluid dynamics simulator. In this context, the lack of robustness of models trained with the popular L 2 and L 1 losses is highlighted and several enhancements to these baseline approaches are introduced. First, the study provides a piece-wise power-logarithmic loss that improves training in the presence of the bimodal distribution of error residuals produced by the dual-phase flow predictions. A non-uniform sampling strategy for the selection of time training snapshots is also included, which contributes to improve the prediction accuracy. The estimation of the front flow field is further refined with a multi-task training strategy. The introduction of these improvements in the baseline models reduce the relative error of the pressure and front flow fields by more than 50%, performing these simulations in a record time of 50 ms. The surrogate model is further evaluated as a digital twin to predict – in a real experiment – the location and spatial extent of race-tracking channels and regions with dissimilar degrees of permeability. [ABSTRACT FROM AUTHOR]

Visual tracking methods based on correlation filter (CF) are to estimate the location and the scale of tracking object within video sequences. However, there are still several inadequate hypotheses within the CF framework. These inadequate hypotheses include the constant Gaussian label map, the predetermined object location and the insufficient utilization for features. To further address the problems caused by these hypotheses, this paper proposes a novel separable-programming based probabilistic-iteration and restriction-resolving correlation filter (PRCF). The main innovation points of this work are: 1) the separable-programming based tracking framework is adopted to achieve adaptive regulation incorporation and to simultaneously address location prediction and filter training; 2) the probabilistic-iteration scheme is suggested to update the Gaussian label for each frame to solve the problem of ad-hoc label map; 3) the adaptive feature fusion measure is introduced to abate the effects of insufficient utilization for numerous features; 4) the convergence behavior of PRCF is proved and discussed through theoretical analysis and worst-case convergence rate calculation. Experiments have also been conducted to test the efficiency of suggested PRCF on 7 benchmarks: OTB100, TC128, VOT2016, VOT2019, UAV123, NFS and LaSOT. The results have indicated that: 1) the PRCF obtains favorable performances compared with other 15 state-of-the-art (SOTA) CF-based trackers; 2) the PRCF reports comparable results against 4 deep learning based trackers; 3) the PRCF achieves a real-time speed of 52 frames-pre-second (FPS) on 7 benchmarks averagely. Thus, the PRCF is qualified for practical target tracking scenarios such as video surveillance and unmanned aerial vehicles (UAVs). [ABSTRACT FROM AUTHOR]

Internet of Vehicles (IoV) is a paradigm of ITS that incorporates automobiles, transportation, information sharing, and traffic infrastructure management with the aim to improve safety on roads. Federated learning enables the Internet of Vehicles (IoV) networks with minimal data communication overhead and privacy infringement. The vehicles participate in model training for improving the model accuracy at the global level and enhancing road safety to a great extent. In this paper, we apply federated learning to driver behavior analysis in which the vehicles in the network collaborate to train CNN-LSTM and CNN-Bi-LSTM deep learning models for driver behavior classification (safe, unsafe, or fatigue) without sharing raw data. The CNN-LSTM and CNN-Bi-LSTM models have been implemented and analyzed using real-time driver's behavior data, collected using inbuilt smartphone sensors or vehicle onboard devices. The proposed model has been evaluated using the performance parameters viz., accuracy, Area Under Curve (AUC), loss, precision, and recall. Both IID and non-IID types of datasets have been considered for analyzing the diversity of the proposed model. It is evident that implementation of federated learning architecture provided similar parametric outcomes on IID and non-IID datasets in comparison to contemporary deep learning implementations on a non-distributed dataset. The proposed work aims towards designing a reliable intelligent transportation system using federated learning. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel EEG classification approach based on the Spatio-spectral feature, aiming to design a motor movement/imagery classification model that extracts multi-domain features with promising performance. Firstly, the difference between motor imagery (MI) tasks and real execution tasks was analyzed by calculating complexity measures of different frequency bands. Then, different connection patterns between MI tasks and real execution tasks were investigated by constructing the PLV and PLI networks. The results of the PLV and PLI brain networks showed that real execution tasks' network connections intensity was stronger than MI tasks, which meant these two networks could be used to distinguish the EEG signal features of different tasks. Afterward, to fully explain the multi-domain features of EEG signals, we fused the Phase Locking Value (PLV) and Phase-Lag Index (PLI) matrices (spatial-domain) under the subsets of frequency bands (frequency-domain) into a 3-D feature, namely the Spatio-spectral feature. Finally, a 3-D convolutional neural network combined with a long short-term memory (3DCNN-LSTM) was utilized to decode the feature. The results showed that the average accuracy of 10 subjects, 20 subjects, 50 subjects, 80 subjects, and 103 subjects was 85.88%, 83.09%, 76.30%, 75.02%, and 74.54%. Taken together, the proposed method provided promising classification accuracies, superior multi-domain features extraction ability, simpler structure, and robustness to classify the different motor movement/imagery tasks. The results contribute to our understanding of applying the deep learning method to decode EEG multi-domain features in the brain-computer interface (BCI) systems (e.g., MI, emotion recognition, and epileptic seizure classification). • Analyzed the difference between MI tasks and corresponding real execution tasks by calculating complexity measures of different frequency bands. The results showed that the differences between different tasks were more significant in the alpha (α) band in the PLV network and the gamma (γ) band in the PLI network. • Investigated different connection patterns between MI tasks and real execution tasks by constructing the PLV and PLI networks after thresholding. The results showed that real execution tasks' network connections intensity was stronger than MI tasks, which meant these two networks could be used to distinguish EEG signal features of different tasks. • Fused the PLV and PLI matrices (spatial-domain) under the subsets of frequency bands (frequency-domain) into a 3-dimensional feature, namely the Spatio-spectral feature. • Utilized a 3-dimensional convolutional neural network combined with a long short-term memory (3DCNN-LSTM) to decode and classify the Spatio-spectral feature. The average classification accuracy of 10 subjects, 20 subjects, 50 subjects, 80 subjects, and 103 subjects was 85.88%, 83.09%, 76.30%, 75.02%, and 74.54%, respectively. [ABSTRACT FROM AUTHOR]

Timely and effective flight regime recognition is one of the key tasks for structural usage monitoring, it can provide early warning to the dangerous regime. However, existing methods for flight regime recognition either rely on expert knowledge or ignore the continuity and decision boundary of the regimes, which limits their performance for complex regimes and hinders their deployment in practice. In this paper, we provided a brand-new solution for this problem and designed a deep temporal segmentation neural network to realize intelligent regime segmentation. Meanwhile, we revealed the long-tailed distribution of flight regimes and proposed class-wise dynamic group rebalance loss to keep inter-class accuracy balanced. To evaluate the effectiveness of the proposed model, we collected and elaborately annotated plentiful actual flight sorties data, including 11 flight regimes. Test results demonstrated that the model can automatically separate different regimes in a continuous flight sortie without any pre-processing and post-processing while extracting the accurate regime boundary and achieving 95.98% recognition accuracy. • Flight regime recognition from the perspective of temporal segmentation. • DTS-Net realizes continuous regimes recognition and boundaries positioning. • Class-wise dynamic group rebalance loss deals with long-tailed distributed regimes. • Without additional prior knowledge and post-processing. [ABSTRACT FROM AUTHOR]

Currently, most of the existing point cloud registration methods use stacked edge convolution as feature extractor, ignoring the importance of deep semantic information, meanwhile, the use of attention mechanism tends to increase the computational cost of the model and limits the matching effect. This paper proposes a dynamic graph aggregation framework for point cloud registration by building a dynamic deep network, which can capture richer semantic information and shape properties of point clouds. First, a hybrid feature extractor, which fully fuses local graph information and global graph information, is designed to obtain more discriminative feature descriptors. Second, a local graph neighborhood scoring module to update local graph features and achieve feature enhancement is built by learning the difference of similar point neighborhood features. Finally, a dual-constrained matching module is designed, which is used to measure the similarity of point pairs from the two aspects of Euclidean distance and affinity between features. The proposed framework achieves excellent registration performance, as demonstrated by experimental results on challenging 3D point cloud benchmarks. Especially in partial point cloud registration, MAE(R) achieved excellent results of 0.2614 and 0.2619 under unseen shapes and unseen categories. [ABSTRACT FROM AUTHOR]

Speaker counting is an important research area in sound forensics. There are limited speaker counting papers in the literature, as it is challenging to collect datasets. This work aims to collect a new overlapping speech signal dataset for speaker counting and propose a novel feature engineering model. In this work, textural feature extraction is based on the iconic James Webb space telescope; hence, this pattern is named James Webb Pattern (JWPat). A new speaker counting speech dataset comprising 3,121 speeches divided into 32 classes (the class number corresponded to the number of speakers) was collected. A new framework that mimics the deep learning model has been proposed to classify the collected speech classes. The proposed feature engineering model is self-organized and uses various mother wavelet functions to generate features at both low and high levels. We have obtained the best classification accuracy of 86.74% using the symlet4 mother wavelet function. Using our proposed framework, eight classification results have been calculated with accuracy ranging from 75.94% to 86.74%. This range is over 10% accuracy, and it demonstrates the effect of the mother wavelet function on the classification performance. Moreover, the feature extraction capability of the mirror of the James Webb telescope has been demonstrated. Our proposed method yielded 86.74% accuracy on a large dataset and indicated the success of our proposed model. [ABSTRACT FROM AUTHOR]

Human body pose estimation represented by joint rotations is essential for driving the virtual characters. The present paper developed a novel end-to-end point-to-pose mesh fitting network (P2P-MeshNet) to directly estimate the body joint rotations. P2P-MeshNet provided a strong collaboration between the deep learning network, an inverse kinematics network for body pose estimation (IKNet-body), and the self-correcting network, an iterative error feedback network (IEF). The introduced P2P-MeshNet was then applied to the free mocap (FreeMocap) dataset covering OpenPose 3D joint locations reconstructed from multi-view OpenPose 2D joint locations. The generated joint rotations were tested using the mean per joint position error (MPJPE), as well as the percentage of correct keypoints (PCK) along with the area under the PCK curve (AUC) with a threshold range of 0–60 mm after Procrustes aligned. Based on the compared metrics, P2P-MeshNet with 11.31 mm and 99.7% in estimate error and success rate as well as an AUC of 80.9 demonstrated a more consistent tool for future human body pose estimation. The runtime performance of 100 frames per second implied its potential application prospects. • A novel end-to-end point-to-pose mesh fitting network (P2P-MeshNet) was proposed. • P2P-MeshNet was applied to estimate the body joint rotations directly. • The challenge of real-time estimation and driving the virtual characters was solved. • P2P-MeshNet provided a significant qualitative and quantitative improvement over the existing methods. [ABSTRACT FROM AUTHOR]

TV program recommendation is important for users in the face of a huge amount of information data. The existing TV program recommendation mainly relies on a collaborative filtering method to recommend through interactive data between users and programs. Although some methods utilize auxiliary information to enrich semantic features, most of them only use a single data type, which cannot capture a more diverse feature representation of the user and program. In this paper, we propose a neural TV program recommendation model with multi-source heterogeneous data, which makes full use of the multi-source heterogeneous auxiliary information. Specifically, we combine heterogeneous features derived from auxiliary information to learn a deep program representation in the program encoder module. To more accurately capture user preferences, we further utilize the personalized attention mechanism to determine the importance of different programs to the user representation based on the interaction between users and programs in the user encoder module. Extensive experiments on a real dataset of the Chinese capital show that our model can effectively improve the performance of TV program recommendations compared to the existing models. [ABSTRACT FROM AUTHOR]

Building occupancy information could aid energy preservation while simultaneously maintaining the end-user comfort level. Energy conservation becomes essential since energy resources are scarce and human dependency on appliances is only exponentially increasing. While intrusive sensors (i.e., cameras and microphones) can raise privacy concerns, this paper presents an innovative non-intrusive occupancy detection approach using environmental sensor data (e.g., temperature, humidity, carbon dioxide (CO 2), and light sensors). The proposed scheme transforms multivariate time-series data into images for better encoding and extracting relevant features. The utilized image transformation method is based on data normalization and matrix conversion. Precisely, by representing time-series in 2D space, an encoding kernel can move in two directions while it can move only in one direction when applied to a 1D signal. Moreover, machine learning (ML) and deep learning (DL) techniques are utilized to classify occupancy patterns. Several simulations are used to evaluate the approach; mainly, we investigated pre-trained and custom convolutional neural network (CNN) models. The latter attained an accuracy of 99.00%. Additionally, pixel data are extracted from the generated images and subjected to traditional ML methods. Throughout the numerous comparison settings, it was observed that the latter strategy provided the optimal balance of 99.42% accuracy performance and minimal training time across the occupancy datasets. • Exhaustive literature review on occupancy detection schemes based on sensor data. • Image transformation technique based on data normalization and matrix conversion. • Assessing a variety of pre-trained and custom CNN models to predict occupancy status. • Pixel extraction to gather values and integrate them with conventional ML algorithms. [ABSTRACT FROM AUTHOR]

Deep anomaly detection (DAD) is essential in optimizing building energy management. Nonetheless, most existing works concerning this field consider unsupervised learning and involve the analysis of sensor readings through a one-dimensional (1D) energy time series, which limits the options for detecting anomalies within the building's energy consumption. To the best of the authors' knowledge, this paper presents the first study that explores using two-dimensional (2D) image representations as features of a supervised deep transfer learning (DTL) approach. Specifically, using 2D image representations allows taking advantage of any underlying data within the feature set, providing more possibilities to encode data and detect pertinent features which may not be considered in standard 1D time-series. Furthermore, the effects of using CNN activations as machine learning (ML) model features are also investigated to combine the advantages of both techniques. Additionally, the concept of layer and hyperparameter variation for the CNN model is also studied, with the objective of reducing the overall time computation and resource requirements of the proposed system. Hence, this makes our approach a candidate for edge-based applications. As per the conducted experiments, the top methodology rests at the F1-scores of 93.63% and 99.89% under simulated and real-world energy datasets, respectively. This involves using grayscale 2D image representations that combine CNN activations extracted from AlexNet and GoogleNet pre-trained models as features to a linear support vector machine (SVM) classifier. Finally, the comparison analysis with the state-of-the-art has shown the superiority of the proposed method in various assessment criteria. • Deep anomaly detection of energy consumption using energy time-series images. • Converting 1D energy time series into 2D representations. • 2D energy representations as features of a neural network-based transfer learning. • Using neural network activations as machine learning model features. • Combining neural network activations extracted from AlexNet and GoogleNet. [ABSTRACT FROM AUTHOR]

Stroke, categorized under cardiovascular and circulatory diseases, is considered the second foremost cause of death worldwide, causing approximately 11% of deaths annually. Stroke diagnosis using a Computed Tomography (CT) scan is considered ideal for identifying whether the stroke is hemorrhagic or ischemic. However, most methods for stroke classification are based on a single slice-level prediction mechanism, meaning that the most imperative CT slice has to be manually selected by the radiologist from the original CT volume. This paper proposes an integration of Convolutional Neural Network (CNN), Vision Transformers (ViT), and AutoML to obtain slice-level predictions as well as patient-wise prediction results. While the CNN with inductive bias captures local features, the transformer captures long-range dependencies between sequences. This collaborative local-global feature extractor improves upon the slice-wise predictions of the CT volume. We propose stroke-specific feature extraction from each slice-wise prediction to obtain the patient-wise prediction using AutoML. While the slice-wise predictions helps the radiologist to verify close and corner cases, the patient-wise predictions makes the outcome clinically relevant and closer to real-world scenario. The proposed architecture has achieved an accuracy of 87% for single slice-level prediction and an accuracy of 92% for patient-wise prediction. For comparative analysis of slice-level predictions, standalone architectures of VGG-16, VGG-19, ResNet50, and ViT were considered. The proposed architecture has outperformed the standalone architectures by 9% in terms of accuracy. For patient-wise predictions, AutoML considers 13 different ML algorithms, of which 3 achieve an accuracy of more than 90%. The proposed architecture helps in reducing the manual effort by the radiologist to manually select the most imperative CT from the original CT volume and shows improvement over other standalone architectures for classification tasks. The proposed architecture can be generalized for volumetric scans aiding in the patient diagnosis of head and neck, lungs, diseases of hepatobiliary tract, genitourinary diseases, women's imaging including breast cancer and various musculoskeletal diseases. Code for proposed stroke-specific feature extraction with the pre-trained weights of the trained model is available at: https://github.com/rishiraj-cs/StrokeViT_With_AutoML. [Display omitted] • Most methods for stroke classification are based on slice-level prediction mechanism. • An integration of CNN and ViT is proposed for slice-level prediction. • For patient-wise predictions, stroke-specific feature extraction is proposed. • AutoML is trained on the proposed domain-specific features extracted. [ABSTRACT FROM AUTHOR]

Multi-object tracking (MOT) is a highly valued and challenging research topic in computer vision. To achieve more robust tracking performance, recently published MOT methods tend to use anchor-free object detectors, which have the advantage of dealing with the identity ambiguity problem encountered by anchor-based methods in learning appearance features. However, in practical applications, it is found that the detection accuracy of the anchor-free object detector based on classical convolutional neural networks in crowded scenes will be significantly reduced. In order to have better detection and tracking performance in crowded scenes, this paper proposes an anchor-free joint detection and embedding (JDE) MOT method based on Transformer architecture, called Swin-JDE. The proposed method includes a novel Patch-Expanding module, which can improve the spatial information of feature maps by up-sampling processing through neural network learning and Einops Notation-based rearrangement to enhance the detection and tracking performance of the MOT model. In terms of training method, we propose a two-step training method that trains the detection branch separately from the appearance branch to enhance the detection robustness of anchor-free predictors. Furthermore, during the training process, we also propose an examination method to remove occluded targets from the training dataset to improve the accuracy of the appearance embedding layer. In terms of data association, we propose a new post-processing method, which simultaneously considers the three factors of detection confidence, appearance embedding distance, intersection-over-union (IoU) distance to match each tracklet and the detection information to improve the tracking robustness of the MOT model. Experimental results show that the proposed method achieves 70.38% multiple object tracking accuracy (MOTA) and 69.53% identification F1-score (IDF1) results in the MOT20 benchmark dataset, and the identification switch (ID Switch) is reduced to 2026. Compared with FairMOT, the proposed method improves MOTA and IDF1 by 8.58% and 2.23%, respectively. [ABSTRACT FROM AUTHOR]

Identification of exfoliated graphene flakes and classification of the thickness are important in the nanomanufacturing of advanced materials and devices. This paper presents a deep learning method to automatically identify and classify exfoliated graphene flakes on Si/SiO 2 substrates from optical microscope images. The presented framework uses a hierarchical deep convolutional neural network that is capable of learning new images while preserving the knowledge from previous images. The deep learning model was trained and used to classify exfoliated graphene flakes into monolayer, bi-layer, tri-layer, four-to-six-layer, seven-to-ten-layer, and bulk categories. Compared with existing machine learning methods, the presented method showed high accuracy and efficiency as well as robustness to the background and resolution of images. The results indicated that the pixel-wise accuracy of the trained deep learning model was 99% in identifying and classifying exfoliated graphene flakes. This research will facilitate scaled-up manufacturing and characterization of graphene for advanced materials and devices. • A machine intelligence approach is developed to classify exfoliated graphene flakes. • A hierarchical deep convolutional neural network is applied for the classification task. • The neural network can learn from new images while preserving the learned knowledge. • The approach is robust to the background and resolution of microscopy images. • Advanced vision techniques are incorporated to handle highly imbalanced datasets. [ABSTRACT FROM AUTHOR]

Remaining useful life (RUL) estimation, a key component in predictive maintenance (PdM), aims to reduce maintenance cycles in the prognostic health of mechanical equipment(s). Research directions using deep-learning-based RUL estimation often suffer from limited availability of degradation signals resulting inaccurate predictions. Until now, state-of-the-art models trained on large sets of natural images to classify objects have not been re-used to improve regression-based RUL estimation accuracies of mechanical equipment. Actually, this is a rarely researched topic in PdM. Inspired by transfer learning, we showcase that the knowledge learned by popular pre-trained models can be transferred to improve industrial machinery-based-maintenance decision-making. Accordingly, this paper proposes a novel multi-model data-fusion-based deep transfer learning (MMF-DTL) framework for improved RUL estimation of rolling bearings through degradation images (DI) and pre-trained deep convolutional neural networks (CNNs). After procuring the degradation signals, we obtain DIs incorporating sufficient deterioration information using Markov Transition Fields. Next, these DIs are input into a DTL network comprising three pre-trained CNNs, i.e., DenseNet201, VGG16, and ResNet50, designed in a parallel fashion, where each constituent fine-tunes a different count of layers. Subsequently, features extracted from each component model are fused in a weighted manner and passed onto several fully connected layers. The effectiveness of the proposed framework is validated using the PHM Challenge 2012 bearing degradation dataset. Compared to several state-of-the-art works, our approach improves ∼ 12. 57 % on error rate and ∼ 26. 04 % on MAE, suggesting it is practically feasible to grasp transferable attributes from a general-purpose related dataset to another with minimal dataset size. [ABSTRACT FROM AUTHOR]

Recently, developing automated video surveillance systems (VSSs) has become crucial to ensure the security and safety of the population, especially during events involving large crowds, such as sporting events. While artificial intelligence (AI) smooths the path of computers to think like humans, machine learning (ML) and deep learning (DL) pave the way more, even by adding training and learning components. DL algorithms require data labeling and high-performance computers to effectively analyze and understand surveillance data recorded from fixed or mobile cameras installed in indoor or outdoor environments. However, they might not perform as expected, take much time in training, or not have enough input data to generalize well. To that end, deep transfer learning (DTL) and deep domain adaptation (DDA) have recently been proposed as promising solutions to alleviate these issues. Typically, they can (i) ease the training process, (ii) improve the generalizability of ML and DL models, and (iii) overcome data scarcity problems by transferring knowledge from one domain to another or from one task to another. Although the increasing number of articles proposed to develop DTL- and DDA-based VSSs, a thorough review that summarizes and criticizes the state-of-the-art is still missing. To that end, this paper introduces, to the best of the authors' knowledge, the first overview of existing DTL- and DDA-based video surveillance to (i) shed light on their benefits, (ii) discuss their challenges, and (iii) highlight their future perspectives. [ABSTRACT FROM AUTHOR]

Human activity recognition has become a pillar of today intelligent Human–Computer Interfaces as it typically provides more comfortable and ubiquitous interaction. This paper proposes a novel fuzzy-based deep learning-based algorithm to predict future sequences of activities from a given sequence of daily living activities of a subject wearing a lower limb exoskeleton. The engineering application concerns the challenging task of recognizing locomotion activities of the wearer in real-time, which is needed to ensure appropriate control of the robot during daily living activities. Indeed, real-time locomotion activity recognition is very challenging for controlling lower-limb exoskeletons. The model proposes a new adaptive kernel, based on the data features derived from the fuzzy rules on the input sequences to enrich the features of the activity sequences. Then, a CNN is applied to extract local subsequences from the whole sequences to identify local patterns in the convolution window. Finally, an attention-based GRU is incorporated into the model to extract meaningful parts of the time-series sequences. The results show high accuracy in the estimation of the transition between gait modes which is critical to ensure smooth control of the exoskeleton. The performance of the model is evaluated using the dynamic activity data gathered from different subjects. The proposed model outperforms the traditional models used in the literature. [ABSTRACT FROM AUTHOR]

Assembly is an important process of aerospace industry which involves significant number of fastening elements such as bolts, washers and nuts, etc. Presently, the identification of these elements is done manually by humans. Since human operators are prone to errors, any fault in the assembly line can have a major impact on the overall efficiency and safety of an aerospace vehicle. In this paper, we present a deep learning and image processing-based approach for identification of mechanical fixation/fastening elements used in the aerospace assembly line. We propose YOLO-v5 algorithm to classify the components based on their head and lateral shape. We also propose an image processing method to estimate the spatial dimensions of the assembly line components including thread pitch. Moreover, this study also features the development of an image acquisition platform with two cameras mounted and proper lighting mechanism to capture quality images. Despite the challenges associated with such systems, the proposed deep learning (YOLO-v5) algorithm has shown promising results with a mAP@0.5 of 0.996 for component classification. Whereas, the proposed image processing mechanism achieved an accuracy of 100% for standard size assignments with a maximum error of 0.05 mm for pitch calculation. [ABSTRACT FROM AUTHOR]

As a significant extension of rough set theory, three-way decision (3WD) theory plays a crucial role in the data mining of uncertain information and decision-making analysis. Transfer learning (TL) is also a powerful knowledge discovery and deep learning strategy that has attracted the attention of many scholars. At present, many achievements have been made in related studies based on the concepts of sample transfer, feature transfer, and parameter transfer. However, there are few studies on multi-granularity fusion and TL for multisource data from the perspective of the stochastic dominance (SD) relation. In this paper, an intuitionistic fuzzy three-way transfer learning (IF3WTL) model based on rough almost stochastic dominance (RASD) is proposed. In this scenario, we first introduce the concept of a rough marginal information measure and the corresponding calculation method. Then, an RASD method is proposed to generate the multi-granularity distribution of same-category information in the source domain. In addition, 3WD theory is introduced to classify the target domain objects into positive, negative, and boundary regions based on the relation between marginal minimum risk information and the corresponding multi-granularity distribution. Furthermore, a secondary decision strategy with an SVM algorithm is implemented to iteratively process boundary region objects. The proposed method can reduce the differences in the data distribution among domains through RASD, improve noise tolerance, and obtain the degree of roughness of common information at different scales. Moreover, the proposed method adopts an iterative learning strategy, which can reduce the decision-making cost in cases with insufficient information and improve the accuracy of classification for objects in the boundary region. The rationality and effectiveness of the proposed model are verified through experiments with the ABIDE dataset and comparative analyses with the existing state-of-the-art methods. [ABSTRACT FROM AUTHOR]

While constructing predictive model for industrial process quality prediction, the selection of an appropriate input variable set is crucial to the online prediction performance. In this paper, the issue is solved through a deep causality analysis method, where the deep learning feature extracting unit, i.e., the Gated Recurrent Unit (GRU), is combined with the Granger Causality (GC) to extract the causal relationships between process variables. In this way, the nonlinear and dynamic features contained in the process data can be effectively learned in the GC analyzing framework, resulting in a more reliable causal relationship structure. Furthermore, through applying the attention mechanism to the input layer of the GRU cells, the causal relationship between process variables and the quality variable can be extracted, even the long time-lags exist. Based on the direct and indirect causal structure, the quality-related causal input variables are selected, and the predictive model is established through a deep dynamic regression network with GRU cells. A numerical example and a real chemical process case are provided to verify the effectiveness of the proposed causal variable selection method, where the prediction performance is significantly improved comparing with the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Community detection is a field of research with increasing influence due to its importance in dimensionality reduction and revealing the organizational patterns in networks. A community or cluster in a network reflects the interactions among the objects that are strongly connected and distinguish them from the ones in other clusters. Over the past decades, various algorithms have been developed to detect the community structure in networks. However, the performance of these algorithm is affected by the existence of noise or nonlinear information in the network. Consequently, deep learning algorithms arise to tackle this issue. In this paper, we propose a novel deep robust auto-encoder nonnegative matrix factorization (DRANMF) approach to detect the community structure in networks. In particular, DRANMF consists of a deep structured decoder and encoder components to transform the high-order proximity matrix of the network to and from the cluster membership space, respectively. Moreover, we enhance the robustness of the proposed approach against noise and outliers by adopting the l 2 , 1 -norm to formulate the objective function. The experiments on multiple synthetic and real-world networks demonstrate that the proposed DRANMF has a better performance and robustness than other existing methods. [ABSTRACT FROM AUTHOR]

Deep Reinforcement Learning systems are now a hot topic in Machine Learning for their effectiveness in many complex tasks, but their application in safety-critical domains (e.g., robot control or self-autonomous driving) remains dangerous without mechanism to detect and prevent risk situations. In Deep RL, such risk is mostly in the form of adversarial attacks, which introduce small perturbations to sensor inputs with the aim of changing the network-based decisions and thus cause catastrophic situations. In the light of these dangers, a promising line of research is that of providing these Deep RL algorithms with suitable defenses, especially when deploying in real environments. This paper suggests that this line of research could be greatly improved by the concepts from the existing research field of Safe Reinforcement Learning, which has been postulated as a family of RL algorithms capable of providing defenses against many forms of risks. However, the connections between Safe RL and the design of defenses against adversarial attacks in Deep RL remain largely unexplored. This paper seeks to explore precisely some of these connections. In particular, this paper proposes to reuse some of the concepts from existing Safe RL algorithms to create a novel and effective instance-based defense for the deployment stage of Deep RL policies. The proposed algorithm uses a risk function based on how far a state is from the state space known by the agent, that allows identifying and preventing adversarial situations. The success of the proposed defense has been evaluated in 4 Atari games. [ABSTRACT FROM AUTHOR]

This paper describes a novel DNN-based system, named PD3net, that detects multiple people from a single depth image, in real time. The proposed neural network processes a depth image and outputs a likelihood map in image coordinates, where each detection corresponds to a Gaussian-shaped local distribution, centered at each person's head. This likelihood map encodes both the number of detected people as well as their position in the image, from which the 3D position can be computed. The proposed DNN includes spatially separated convolutions to increase performance, and runs in real-time with low budget GPUs. We use synthetic data for initially training the network, followed by fine tuning with a small amount of real data. This allows adapting the network to different scenarios without needing large and manually labeled image datasets. Due to that, the people detection system presented in this paper has numerous potential applications in different fields, such as capacity control, automatic video-surveillance, people or groups behavior analysis, healthcare or monitoring and assistance of elderly people in ambient assisted living environments. In addition, the use of depth information does not allow recognizing the identity of people in the scene, thus enabling their detection while preserving their privacy. The proposed DNN has been experimentally evaluated and compared with other state-of-the-art approaches, including both classical and DNN-based solutions, under a wide range of experimental conditions. The achieved results allows concluding that the proposed architecture and the training strategy are effective, and the network generalize to work with scenes different from those used during training. We also demonstrate that our proposal outperforms existing methods and can accurately detect people in scenes with significant occlusions. • Robust system to detect people only using depth information from a depth camera. • System outperforms state-of-the-art methods in different datasets without fine-tuning. • Proposal runs in real time using conventional GPUs. • Computational demands are independent of the number of people in the scene. • Generated database is available to the research community. [ABSTRACT FROM AUTHOR]

Recently Deep Denoising Autoencoders (DDAE) have shown state-of-the-art performance on various machine learning tasks. In this paper, the authors extended this approach to whispered speech recognition which is one of the most challenging problems in Automatic Speech Recognition (ASR). Namely, due to the profound differences between acoustic characteristics of neutral and whispered speech, the performance of traditional ASR systems trained on neutral speech degrades significantly when whisper is applied. This mismatch between training and testing is successfully alleviated with the new proposed system based on deep learning, where DDAE is applied for generating whisper-robust cepstral features. This system was tested and compared in terms of word recognition accuracy with conventional Hidden Markov Model (HMM) speech recognizer in an isolated word recognition task with a real database of whispered speech (WhiSpe). Three types of cepstral coefficients were used in the experiments: MFCC (Mel-Frequency Cepstral Coefficients), TECC (Teager-Energy Cepstral Coefficients) and TEMFCC (Teager-based Mel-Frequency Cepstral Coefficients). The experimental results showed that the proposed system significantly improves whisper recognition accuracy and outperforms traditional HMM-MFCC baseline, resulting in an absolute 31% improvement of whisper recognition accuracy. The highest word recognition rate of 92.81% in whispered speech was achieved with TECC feature. [ABSTRACT FROM AUTHOR]

In the pathological diagnosis of early, late and non-membranous nephropathy, direct immunofluorescence is highly likely to present potentially specific lesions, while it is often overlooked due to the difficulty of screening with naked eyes. With the advanced progress of deep learning, they have shown powerful abilities in detecting potential lesions. In this paper, we propose an adaptive weighted confidence calibration fusion framework (Ada-CCFNet) consisting of a preprocessing module, an adaptive weighted confidence calibration fusion (Ada-CCF) module and a classification module for diagnosis of membranous nephropathy by classifying the multimodal direct immunofluorescence images. In the preprocessing module, we use the well-known U-Net to segment individual glomeruli and standardize their luminance appearance by the average luminance difference method, allowing the subsequent modules to focus more on the diseased glomerular region. Subsequently, in the Ada-CCF module, six confidence calibration methods are utilized for two main direct immunofluorescence images, IgG and C3, and the comprehensive calibration scores are obtained based on the adaptive weighted fusion of six confidence calibration methods to obtain more reliable confidence level, in which the adaptive weights are related with expected calibration error reductions. For the classification module, the weighted probability scores of IgG and C3 are jointly fed into the module to achieve the classification by random forest. Experimental results showed that Ada-CCFNet achieves the classification accuracy of 73.52%, surpassing the methods of using single IgG or C3 images and positive grade indicator with 8.24%, 8.94% and 22.76%, and outperforming the compared methods in the classification of membranous nephropathy. • An adaptive weighted confidence calibrated fusion network (Ada-CCFNet) is designed to classify early, late and non-membranous nephropathy. • Classify membranous nephropathy stages by multi-modal direct immunofluorescence images. • A preprocessing method to segment glomerular region effectively and standardize luminance. • An adaptive weighted method by Excepted Calibration Error reductions is proposed for fusion. • Our method provides higher classification accuracy than the compared methods. [ABSTRACT FROM AUTHOR]

Breast ultrasound images segmentation is one of the key steps in clinical auxiliary diagnosis of breast cancer, which seriously threatens women's health. Currently, deep learning methods have been successfully applied to breast tumors segmentation. However, blurred boundaries, heterostructure and other factors can cause serious missed detections and false detections in the segmentation results. In this paper, we developed a novel refinement residual convolutional network to segment breast tumors accurately from ultrasound images, which mainly composed of SegNet with deep supervision module, missed detection residual network and false detection residual network. In SegNet, we add six side-out deep supervision modules to guide the network to learn to predict precise segmentation masks scale-by-scale. In missed detection residual network, the receptive field provided by different dilation rates can provide more global information, which is easily lost in deep convolutional layer. The introduction of false detection and missed detection residual network can promotes the network to make more efforts on those hardly-predicted pixels to help us obtain more accurate segmentation results of the breast tumor. To evaluate the segmentation performance of the network, we compared with several state-of-the-art segmentation approaches using five quantitative metrics on two public breast datasets. Experimental results demonstrate that our method achieves the best segmentation results, which indicates that our method has better adaptability on breast tumors segmentation. [ABSTRACT FROM AUTHOR]

Deep learning models have gained a reputation of high accuracy in many domains. Convolutional Neural Networks (CNN) are specialized towards image recognition and have high accuracy in classifying objects within images. However, CNNs are an example of a black box model, meaning that experts are unsure how they work internally to reach a classification decision. Without knowing the reasoning behind a decision, there is low confidence that CNNs will continue to make accurate decisions, so it is unsafe to use them in high-risk or safety–critical​ fields without first developing methods to explain their decisions. This paper is a survey and analysis of the available explainability methods for showing the reasoning behind CNN decisions. [ABSTRACT FROM AUTHOR]

The high-speed railway is a deeply coupled system with strong and weak electrical equipment, while complex electromagnetic interference (EMI) consequently brings potential risks and hazards to signaling safety. Since the incident reports on signaling failure intrinsically reflect the generation and evolution mechanism of equipment failures, relying on text mining technology, this paper tries to extract failure-related entities and constructs a knowledge graph to clarify the negative impact of the on-site electromagnetic environment. Firstly, based on convolutional neural networks (CNN), a supervised deep learning model for Chinese text classification is established to generate a corpus containing only railway failures caused by EMI. Then, the bidirectional long short-term memory (BiLSTM) and bidirectional encoder representations from transformers (BERT) algorithms are adopted to build the named entity recognition (NER) model. A NER algorithm more suitable for Chinese text features is proposed through ensemble modeling, training verification, and comparative evaluation. Finally, the knowledge storage and visualization of relational graph construction based on the Neo4j database are realized according to the obtained failure-related entities. This knowledge topology network effectively explores the inherent relationship between EMI factors and railway safety, as well as provides support for improving the safety assessment and enhancing the anti-interference performance of the equipment. • Electromagnetic interference as a hidden risk may lead to railway safety incidents. • Text mining-based approach for understanding railway incident reports is proposed. • BERT is applied to improve the performance of NER for Chinese text. • A knowledge graph of fault-related entities is implemented based on Neo4j. [ABSTRACT FROM AUTHOR]

Modern power systems are experiencing the challenge of high uncertainty with the increasing penetration of renewable energy resources and the electrification of heating systems. In this paradigm shift, understanding electricity users' demand is of utmost value to the retailers, aggregators, and policymakers. However, behind-the-meter (BTM) equipment and appliances at the household level are unknown to the other stakeholders mainly due to privacy concerns and tight regulations. In this paper, we seek to identify residential consumers based on their BTM equipment, mainly rooftop photovoltaic (PV) systems and electric heating, using imported/purchased energy data from utility meters. To solve this problem with an interpretable, fast, secure, and maintainable solution, we propose an integrated method called Interpretable Refined Motifs And binary Classification (IRMAC). The proposed method comprises a novel shape-based pattern extraction technique, called Refined Motif (RM) discovery, and a single-neuron classifier. The first part extracts a sub-pattern from the long time series considering the frequency of occurrences, average dissimilarity, and time dynamics while emphasising specific times with annotated distances. The second part identifies users' types with linear complexity while preserving the transparency of the algorithms. With the real data from Australia and Denmark, the proposed method is tested and verified in identifying PV owners and electrical heating system users. The performance of the IRMAC is studied and compared with various state-of-the-art methods. The proposed method reached an accuracy of 96% in identifying rooftop PV users and 94.4% in identifying electric heating users, which is comparable to the best solution based on deep learning, while the speed of the inference model is a thousand times faster. Last but not least, the proposed method is a transparent algorithm, which can tackle the concerns regarding the agnostic decision-making process when policies prohibit some machine learning methods. [ABSTRACT FROM AUTHOR]

The progress of machine learning (ML) in the past years has opened up new opportunities to the field of gas turbine (GT) modelling. However, successful implementation of ML algorithms remains challenging, particularly for complex problems such as multi-mode faults. An important tool for enabling applications are the feature selection and feature learning (FSFL) techniques. In particular, FL techniques have recently facilitated and improved the applicability of ML to GT modelling. This review paper conducts a review on 46 studies that utilised FSFL for GT modelling with ML. The purpose of this review is to investigate how FSFL techniques can help address GT modelling challenges and when researchers should deploy them. Therefore, the theories behind the techniques are illustrated in depth along with practical application examples from the analysed literature. The advantages and limitations of FSFL are discussed, the computational costs of different techniques are compared, and trends in the field are highlighted. Consequently, a novel categorisation framework for FSFL techniques and recommendations regarding when and how to implement them are provided. A new knowledge accumulation, extraction, and transfer concept is proposed to address GT modelling challenges. [ABSTRACT FROM AUTHOR]

In case the spatial shape of the feature maps of the teacher in feature-based knowledge distillation (KD) is significantly greater than the student model, first, they cannot be compared directly. Second, the knowledge of these complex feature maps cannot be quite apprehensible for the student. This paper proposed a new KD, in which Tucker decomposition was used to decompose the large-dimension feature maps of a teacher to obtain core tensors from the feature maps of the teacher. The knowledge of these tensors can be easily understood by students due to their low complexity. Furthermore, in the proposed KD, an adaptor function is suggested, which balances the spatial shape of the core tensors of the teacher and student and helps compare them using a convolution regressor. Finally, a hybrid loss based on adaptor function is suggested to distill the knowledge of the core tensors of the teacher to the student. Both teacher and student models were implemented on smartphones used as edge devices, and the experiments were evaluated in terms of recognition rate and complexity. According to the results, the student model designed by ResNet-18 architecture has ∼ 65.44 million fewer parameters, ∼ 6.45 GFLOPs less computational complexity, ∼ 1.12 G less GPU memory use, and ∼ 265.67 times greater compression rate than its teacher model designed by ResNet-50 architecture. While the recognition rate of the student model merely dropped down to 1.5% in the benchmark dataset. • Introducing a teacher–student learning with decomposed knowledge distillation. • Distillation is based on the decomposition of feature maps within the middle layers. • Implementing and evaluating the student model on a smartphone as an edge. • The results of the student model showed competitive performance on benchmark datasets. [ABSTRACT FROM AUTHOR]

In modern industrial processes, the data-driven soft sensor technology has been widely used for the prediction of key quality variables. Due to the important of dynamics and nonlinearity in industrial process data, deep learning models like long short-term memory (LSTM) network are well suited for temporal sequence dynamic modeling due to their excellent long-term memory function and feature extraction capability. Furthermore, industrial processes generate a large amount of process data with irregular sampling frequencies. However, traditional LSTM cannot fully utilize the process data with irregular sampling frequency and the guidance value of historical data samples for feature learning. To address these issues, a novel semi-supervised LSTM with history feature fusion attention (HFFA-SSLSTM) model is proposed in this paper. First, the semi-supervised learning strategy is implemented in LSTM to fully utilize the unlabeled data and mine the temporal sequence features of labeled samples and unlabeled samples with irregular sampling frequencies. Then, a novel historical feature fusion attention (HFFA) mechanism is developed, which utilizes historical hidden features to learn attention scores for obtaining weighted historical information-related features. Finally, the extracted features are combined to form the soft sensor model to perform time series prediction tasks for key quality variables in industrial processes. The experimental results on the actual industrial hydrocracking data set demonstrate the effectiveness of the proposed HFFA-SSLSTM model and its possibility of applicating in real industrial processes. [ABSTRACT FROM AUTHOR]

The chiller fault diagnosis is of great significance to maintain the normal operation of the HVAC system and indoor comfort. Due to the difficulty in collecting the chiller's fault data, we usually face the data imbalance problem of less fault data and more normal data. To tackle this problem and further improve the fault diagnosis accuracy, this paper proposes a novel augmented data driven self-attention based deep learning model for the chiller fault diagnosis. Firstly, to solve the data imbalance problem, a stable synthetic minority oversampling technique (SSMOTE) is presented to generate the artificial fault data, which are then mixed with the raw fault data to achieve data augmentation. Further, to enhance the diagnosis accuracy, the self-attention mechanism based temporal convolutional network (STCN) is developed to classify normal and fault datasets. The developed STCN is realized by stacking the modified blocks which can dynamically pay attention to different data information through combining the self-attention mechanism and the traditional residual block together. What is more, in the STCN, the skip connection structure is added to the self-attention mechanism to solve the gradient disappearance problem. Finally, detailed experiments and comparisons are performed. Experimental results show that, compared with other data augmentation methods, the SSMOTE can complete a large scale expansion of the minority fault datasets, and its augmented data have better effect on handling the data imbalance problem. Moreover, the skip self-attention mechanism based deep learning model can achieve better diagnosis accuracy compared with some popular deep and shallow models, such as the LSTM, ELM and SVM, etc. [Display omitted] • The data expanded by the SSMOTE method is more conducive for model training. • The STCN with the SSA mechanism achieves efficient data analysis and improves diagnostic accuracy. • The improved SSA mechanism via jump connection can facilitate solving the gradient disappearance. [ABSTRACT FROM AUTHOR]

This paper proposes two novel machine learning algorithms, namely Random Vector Functional Link Forests and Extreme Learning Forests, to develop an improved unmanned aerial vehicles automatic target recognition system. Such models take advantage of the stochastic procedure followed by Random Forests, where random subsets of instances and features are selected to build diverse Decision Trees. However, different from the usual uni-variate split criterion from Decision Tree algorithms, we propose and employ the novel Random Vector Functional Link Tree or Extreme Learning Tree, where each decision split is performed using the fast non-linear mapping of multiple features provided by either Random Vector Functional Link or Extreme Learning Machines. To prove the efficacy of the novel algorithm, experiments are performed using 90 binary classification problems to compare the performance of the proposed algorithm against other state-of-the-art ensemble learning techniques. Statistical analysis indicates the success of the proposed algorithms in terms of both predictive performance and computational complexity. While the model with deeper trees outperforms classical ensembles in terms of predictive performance (1.41% error reduction) and has similar results to state-of-the-art ensemble models, the model with shallow trees outperforms all ensembles in terms of computational burden (at least 36% faster). Finally, the novel methods are applied to develop an automatic target recognition system for unmanned aerial vehicles, achieving a valuable trade-off in terms of area under receiver operating characteristic curve (0.9309), F 1 score (0.8190), accuracy (0.8646), and computational time (4.14 s). [ABSTRACT FROM AUTHOR]

The application of Artificial Intelligence (AI) is a popular trend to make damage inspection and analysis in structural health monitoring more intelligent and automatic. However, the existing AI-based approaches, especially vision-based methods, mainly focus on damage identification and quantification from images without further analysis to obtain structural load-carrying performance. This paper proposes a Damage-T Generative Adversarial Network (Damage-T GAN) to achieve fast translation from real-world crack images to numerical damage contours. To verify its applicability and effectiveness, two datasets from different reinforced concrete beams were built, and the performance of the trained GAN model was evaluated against the metrics IS, FID, etc. After obtaining the damage contour, a purely visual approach was applied to quantify the damage. As a result, the proposed framework greatly helps field engineers to quickly judge the damage stage of beams in site scenes by simultaneous acquisition of real-world crack images and the generated damage contours of numerical model. [ABSTRACT FROM AUTHOR]

The reliable control of wave energy devices highly relies on the forecasts of wave heights. However, the dynamic characteristics and significant fluctuation of waves' historical data pose challenges to precise predictions. Neural networks offer a promising solution to forecast the wave heights by extracting meaningful features from historical observations. This paper proposes a novel hybrid random vector functional link network with the ensemble and deep learning benefits. Hierarchical stacks of hidden layers are constructed to enforce the deep representations of the time series. Individual output layers follow all enhancement layers to adopt ensemble learning. A neuron pruning strategy is proposed to remove the noisy information from the random features and boost the network's performance. Besides, the proposed network is further utilized to forecast the additive and multiplicative residuals from the ARIMA method. Finally, the ensemble of additive-ARIMA-edRVFL, multiplicative-ARIMA-edRVFL, and edRVFL achieves the best average rankings around two for three forecasting horizons. The proposed ensemble achieves an average ranking of 1.33 on four-hours ahead of forecasting in terms of root mean square error and mean absolute scaled error. Extensive experiments are conducted on twelve time series of the significant wave height. The comparative results demonstrate the superiority of the proposed model over other state-of-the-art methods. The source codes are available on https://github.com/P-N-Suganthan/CODES. [ABSTRACT FROM AUTHOR]

Accurate indoor location information in multi-building/floor environments is essential for establishing many indoor location-based services (LBS). Wi-Fi fingerprinting with received signal strength indicators (RSSI) has become one of the most practical techniques to localize users indoors. However, the fluctuation of wireless signals caused by fading, the multipath effect, and device heterogeneity leads to considerable variations in RSSIs, which poses a challenge to accurate localization. This paper proposes an indoor positioning method, which is constructed based on a convolutional neural network (CNN) framework. Specifically, a novel model by combining an extreme learning machine autoencoder (ELM-AE) with a two-dimensional CNN is proposed. The ELM-AE extracts critical features by reducing input dimensions, while the CNN is trained to effectively achieve significant performance in the positioning phase. To increase positioning accuracy and deal with data shortages, a data augmentation strategy by frequently adding noisy data to the original fingerprint map is devised. The statistical properties namely, the RSS values of each MAC address, are utilized to adjust input noise. The performance of the proposed system is evaluated on two Tampere and UJIIndoorLoc datasets. The experimental results show that EA-CNN achieves better performance than CNN by decreasing average positioning error up to 40.95% and 43.74% in Tampere and UJIIndoorLoc datasets, respectively. Compared to state-of-the-art deep learning-based methods, the positioning performance improves up to 68.36% in Tampere and 67.56% in the UJIIndoorLoc dataset by exploiting only 25% of the training samples. Compared to several state-of-the-art deep learning models, EA-CNN achieves higher accuracy in both positioning and floor estimation. • EA-CNN outperform the state-of-the-art CNN-based procedures. • ELM-AE decrease input dimension, and increase positioning performance of CNN. • Augmentation strategy dealing with data shortages, improve the training efficiency. • Large scale environments and crowdsourced data are utilized for evaluation. [ABSTRACT FROM AUTHOR]

Tumor has become a hot topic in the field of image processing and pattern recognition. Gene expression data is an important way to study tumor. Since gene expression data is characterized by high-dimensional small samples, it is one of the key steps to extract discriminative gene features to distinguish different tumor types. Nonnegative matrix factorization (NMF) is an unsupervised feature representation method that does not depend on the label information of data. NMF can achieve nonlinear dimension reduction and is widely used in tumor recognition. Considering that some data may carry label information, models' feature representation capability will be improved if the label information can be used effectively. Therefore, this paper proposes a semisupervised NMF model based on label consistency (LC-NMF), which uses both labeled and unlabeled data to obtain feature representations. Furthermore, to alleviate the sensitivity of the NMF model to initial values and excavate the deep features of data, a label consistency-based deep semisupervised NMF model (LC-DNMF) is constructed, which combines the LC-NMF model with the layer-by-layer pretraining and multilayer representation strategy in deep learning. The performance of the proposed models (i.e., LC-NMF and LC-DNMF) is verified by applying them to the tumor recognition tasks. The experimental results on seven datasets show that the two models achieve good results and can obtain competitive recognition accuracies compared with the state-of-the-art methods. Furthermore, the performance of the LC-DNMF model outperforms that of the LC-NMF model, which verifies the effectiveness of introducing the layer-by-layer pretraining and multilayer representation strategy. [ABSTRACT FROM AUTHOR]

The application of test cases for detecting the faults within the software is called software testing. Manual testing is laborious and time-consuming hence automation tools to test software were introduced. Despite the use of automation tools at the User Interface (UI) level of the test pyramid, the limitations of current automation tools like automated test case generation and automated repairing of fragile tests still force us to carry out a large amount of manual testing. In this paper, we propose a novel method using AI to address the given challenges. With our proposed method test cases are automatically generated from the structure of the UI using a pipelined architecture of object detection, text detection and NLP models. We show that the test cases generated by the proposed framework can be translated into executable test scripts using a simple parser. The proposed method generates an average of 98.8% correct executable test cases for the applications under study. We also show the capability of the proposed method in generating new tests automatically when the application is modified. The proposed method generates an average of 98.605% correct executable test cases when the UI is modified for the applications under study. We also empirically prove that a GPU implementation of the proposed framework results in just an additional average runtime of 0.92 seconds per test case which is significantly low given the benefits of automated generation of test scripts and automated repairing of fragile tests. [ABSTRACT FROM AUTHOR]

With the rapid development of high-speed railways (HSRs) throughout the world, the fault diagnosis systems of vehicle on-board equipments (VOBEs) for high speed trains have received increasing attention. Since the faults of VOBEs in HSRs are usually uncertain and complex, the current fault diagnosis methods are mainly based on manual judgement in real-world operations, which is generally inefficient and insecurity with the big rail traffic data. In this paper, we propose an automated diagnosis network of VOBE for high-speed train via a deep learning approach. First, we propose a mathematical model to formulate the fault diagnosis problem in HSRs, involving the definition of fault evidence vectors and reason vectors by analyzing the real-world fault data that are collected in Wuhan-Guangzhou high speed railway. Then, a deep belief network (DBN) and its training procedures are developed on the basis of Restricted Boltzmann Machine (RBM). Finally, the proposed diagnosis network is trained and validated with real-world data. Furthermore, we compare the DBN-based fault diagnosis network with k-nearest neighbor (KNN) and ANN-BP (artificial neural network with back propagations). The results indicate that, the developed DBN outperforms both KNN and ANN-BP, and improves the accuracy of fault diagnosis for VOBEs to 90–95% in HSRs. [ABSTRACT FROM AUTHOR]

Deep Neural Network (DNN) models have been extensively developed for intelligent bearing fault diagnosis. The superior performance of DNN-based fault diagnosis methods is attributed to the deepening network structure, which increases the complexity of the deep models and causes difficulties to deploy them in industrial environments with constrained resources. To address the problems, this paper proposes a multi-hierarchical compression to compact the deep neural network for intelligent fault diagnosis. It combines network pruning, parameter quantization, and matrix compression to process the deep model from different perspectives, which achieves a considerable reduction of parameter volume and acceleration of training and response. Firstly, structured pruning is employed to prune the inconsequential filters in the convolutional layer, and unstructured pruning is utilized to eliminate the trivial connections in the fully connected layer. Then, parameter quantization is applied to minimize the number of bits that are required for parameter representation. Finally, the matrix compression storage is taken advantage of to reduce the requirements of storage capacity and further alleviate the excessive demands for the monitoring system. Experimental results and comparisons on two bearing datasets validate the effectiveness of the proposed method. The results show that for two CNN networks with different depths, the integrated network compression method can achieve considerable reductions in the number of parameters and the floating-point operations while almost no decrease in recognition accuracy. The proposed method makes an inspiring exploration for the application of intelligent fault diagnosis in the practical industry. • It proposes a multi-hierarchical network compression method and achieves a considerable reduction of parameters and FLOPs. • A compound pruning process is introduced to reduce the parameters and accelerate the training and response. • Combining parameter quantization and matrix compression decreases the requirement for deployment in the monitoring device. • It improves performance and inspires the application of intelligent fault diagnosis in practical industrial environments. [ABSTRACT FROM AUTHOR]

Detection transformer (DETR) combines convolutional neural network (CNN) with transformer, providing a more advanced idea. In this paper, an object detection model based on DETR is proposed for fire and smoke detection. Compared with other methods that based on deep learning, the proposed one simplifies the pipeline of detection and builds an end-to-end detector. At the same time, the original DETR model usually requires long training time and large amount of computation, resulting in relatively poor performance in detection speed and accuracy, and it shows to be not friendly for small or early fire detection. Therefore, when designing the proposed network model, a normalization-based attention module is added in the feature extraction stage to highlight the effective features, being beneficial to the process of encoding. A multiscale deformable attention is also used in the encoder–decoder structure, accelerating therefore, the process of convergence during the training of the model, which includes the enhancement of the detection of small objects. Also, considering the cost of computation, the number of layers in the encoder–decoder structure is redefined to reduce the complexity of the model, and that also reduces the requirements of the application equipment. Detailed experiments are conducted on three self-built datasets and two public video sets. The results show that, the proposed method has an excellent performance on all of the datasets considered here. [ABSTRACT FROM AUTHOR]

Many current intrinsically interpretable machine learning models can only handle the data that are linear, low-dimensional, and relatively independent attributes and often with discrete attribute values, while the models that are capable of handling high-dimensional nonlinear data, like deep learning, have very poor interpretability. Based on the geometric characteristics, a new idea of accurately wrapping the data region with minimum-volume geometry is proposed for pattern classification. The Variance Hyper Rectangle (VHR) model presented in this paper is a realization of the idea. The VHR model uses the minimum-volume hyper rectangles, obtained through projection variance calculation, to wrap the regions occupied by a category of data, hence it has strong and clear geometric interpretability. In addition, the VHR model is well suited for large data volume, as it approaches the linear complexity in both time and space. Extensive qualitative and quantitative experiments are performed on seven real-world data sets, demonstrating that VHR outperforms the state-of-the-art interpretable methods while running quickly. [Display omitted] • VHR has strong geometric interpretability, and is much reliable and trustworthy. • VHR can provide a clear range of values in each direction for a category of data. • VHR naturally supports incremental learning without any extra processing. • VHR has great performance and stability, and is able to handle big data. [ABSTRACT FROM AUTHOR]