Search

Neural network-based nonlinear system identification is crucial for various multi-step ahead prediction tasks, including model predictive control and digital twins. These applications demand models that are not only accurate but also efficient in training and deployment. While current state-of-the-art neural network-based methods can identify accurate models, they often become prohibitively slow when scaled to achieve high accuracy, limiting their use in resource-constrained or time-critical applications. We propose FranSys, a Fast recurrent neural network-based method for multi-step ahead prediction in non-autoregressive System Identification. FranSys comprises three key innovations: 1) the first non-autoregressive RNN model structure for multi-step ahead prediction that enables much faster training and inference compared to autoregressive RNNs by separating state estimation and prediction into two specialized sub-models, 2) a state distribution alignment training technique that enhances generalizability and 3) a prediction horizon scheduling method that accelerates training by progressively increasing the prediction horizon. We evaluate FranSys on three publicly available benchmark datasets representing diverse systems, comparing its speed and accuracy against state-of-the-art RNN-based multi-step ahead prediction methods. The evaluation includes various prediction horizons, model sizes, and hyperparameter optimization settings, using both our own implementations and those from related work. Results demonstrate that FranSys is 10 to 100 times faster in training and inference with the same and often higher accuracy on test data than state-of-the-art RNN-based multi-step ahead prediction methods, particularly with long prediction horizons. This substantial speed improvement enables the application of larger neural network-based models with longer prediction horizons on resource-constrained systems in time-critical tasks, such as model predictive control and online learning of digital twins. The code of FranSys is publicly available.

Recent advancements in robotics have pushed the development of active exoskeletons and orthoses for assistive, augmentative, and rehabilitative purposes. Deep Learning approaches, particularly in motion analysis and prediction, hold promise for enhanced control of lower limb exoskeletons, offering the potential for improved outcomes in assistive and rehabilitative interventions. This review paper explores recent advancements in deep learning approaches for controlling lower limb exoskeletons. The study encompasses papers published from 2018 to the present, focusing on various aspects of deep learning models, including recognition, prediction, and other related tasks. The paper includes 103 papers covering various tasks and parameters like the gait phase, kinematics, and kinetics of the lower limb exoskeleton. Each aspect is thoroughly examined, highlighting the parameters utilized in the respective models. Moreover, the results obtained from these approaches are evaluated and compared against classical control strategies, shedding light on their effectiveness and potential benefits. The review also addresses the limitations of current deep learning techniques in lower limb exoskeleton control and outlines potential avenues for future research and improvement. By consolidating the latest findings and advancements in this field, this review paper provides valuable insights into the application of deep learning in the control of lower limb exoskeletons, paving the way for enhanced rehabilitation and assistive technologies in the future.

In the era of large-scale data, the role of image compression in computer vision(CV) and computer graphics(CG) tasks is increasingly critical. Traditional methods of image compression have reached their potential limits, leading to increased interest in deep learning-based techniques. However, these modern methods often compromise image quality and require extensive decoding times. This paper introduces the EICNet, which features the innovative Quick Depth-Residual Attention Module (Q-DRAM), an optimized post-processing module, and a checkerboard context model. This design aims to overcome typical shortcomings of deep learning-based compression, enhancing both training and compression efficiency as well as the quality of images at equivalent bit rates. The findings suggest that EICNet improves both the quality and efficiency of image compression. This approach marks a significant advancement in image compression technology, potentially benefiting future applications in the field. The code for this research can be accessed at: https://github.com/ziyicheng427/EICNet.

Global Illumination (GI) is a technique that is employed in computer graphics to enhance realism. Various methods have been used to achieve this using computer-generated imagery. The most precise method involves conventional ray tracing, which yields highly realistic results but is computationally intensive and unsuitable for real-time applications. Alternatively, faster algorithms utilize post-processing on rasterization, making them more suitable for real-time scenarios. However, these algorithms are also resource-intensive and may produce inaccurate lighting due to limited information on screen-space features. our proposal involves utilizing a Generative Adversarial Network (GAN) approach to achieve real-time GI effects, following the methodology of conventional screen-space GI techniques. We take surrounding graphical information into account by going beyond screen-space and producing consistent GI effects that are comparatively closer to their physically correct ray-tracing counterpart. Moreover, our model provides a better quality of generated output than the other recent model which utilized a similar approach by scoring 0.90811 in SSIM, 0.00093 in MSE, and 30.30576 dB in PSNR on our developed dataset.

Areas of computational mechanics such as uncertainty quantification and optimization usually involve repeated evaluation of numerical models that represent the behavior of engineering systems. In the case of complex non-linear systems however, these models tend to be expensive to evaluate, making surrogate models quite valuable. Artificial neural networks approximate systems very well by taking advantage of the inherent information of its given training data. In this context, this paper investigates the improvement of the training process by including sensitivity information, which are partial derivatives w.r.t. inputs, as outlined by Sobolev training. In computational mechanics, sensitivities can be applied to neural networks by expanding the training loss function with additional loss terms, thereby improving training convergence resulting in lower generalisation error. This improvement is shown in two examples of linear and non-linear material behavior. More specifically, the Sobolev designed loss function is expanded with residual weights adjusting the effect of each loss on the training step. Residual weighting is the given scaling to the different training data, which in this case are response and sensitivities. These residual weights are optimized by an adaptive scheme, whereby varying objective functions are explored, with some showing improvements in accuracy and precision of the general training convergence.

In a recent study, we have uncovered the potential of utilizing impact sounds generated when objects fall freely and strike a plane for acoustic-based object recognition (AOR). Building upon this discovery, we address the practical scenario where multiple objects fall simultaneously, leading to overlapped impact sounds. This investigation marks the inception of our work on recognizing multiple objects using overlapped impact sounds, termed acoustic-based multiple object recognition (AMOR). To tackle the challenges posed by overlapped sounds in AOR, we propose a novel two-branch network named OANet, designed with overlap awareness. This network integrates transformer encoding blocks, with the first branch pre-trained using the supervised contrastive learning approach to predict the number of objects. The outputs of the first and second branches are combined and fed into a classifier for the AMOR task. We conducted experiments and evaluations on a self-collected dataset of 10 distinct LEGO objects, including instances where a maximum of three objects fell and hit the designated plane almost concurrently. This dataset contains both non-overlapped and overlapped samples across 175 classes or object combinations (i.e., 10 individual objects, 45 combinations of two objects, and 120 combinations of three objects). Experimental results underscore the viability and promising outcomes of AMOR, showcasing the potential for accurate recognition in complex scenarios. Notably, OANet exhibits superior performance and robustness compared to baseline networks without overlap awareness. Furthermore, upon investigating AMOR in more complex scenarios with mixed overlapped datasets containing samples of a maximum of 5, 7, and 10 objects, we observed the remarkable efficiency of the proposed OANet. It consistently achieved accuracies surpassing 91%, which were 14-24% higher than the results obtained by the baseline models. Additionally, other important findings include the effectiveness of feature combination, the supervised contrastive learning approach for feature learning in the overlap detection network, and the use of mixed data during the training phase.

As a consequence of increased complexity, technical requirements of services and expectations of end-users, telecommunication networks and systems must still increase their efficiency. Research currently focuses on introducing automation and intelligence into network control and management. Neural Networks form a class of Machine Learning techniques utilizing the supervised learning paradigm. The class is mature and comprises a significant number of algorithms. As a result, Neural Networks-based solutions are widely applicable, also in the context of communications and networking. The aim of this tutorial is twofold. Firstly, to provide fundamentals regarding the Neural Network (NN) method. Secondly, to comprehensively study the examples of applying NN-based solutions to solve problems in different aspects of communications and networking. Studies are supplemented with additional explanations, figures and critical considerations, including pros and cons of using selected methods for particular purposes. This part uniquely complements the tutorial one and facilitates in-depth understanding of NN. Based on the conducted studies, we draw a comparative analysis, summaries and expected future research topics and challenges of using NN in communications and networking.

The advent of artificial intelligence (AI) has catalyzed transformative changes across diverse domains, establishing itself as a disruptive technology. Traditional methods often fall short in solving intricate information engineering problems, whereas AI methods can offer superior efficiency, speed, and accuracy. This paper provides a comprehensive overview of the research and practical applications of artificial intelligence in addressing information engineering challenges over the past two decades. It focuses on four critical information engineering fields: intelligent control, pattern recognition and intelligent systems, intelligent signal processing and communication, and optimization algorithms. This paper covers the backgrounds, driving forces, technologies, and applications of these fields, as well as the applications of AI in these areas. We believe that this study will be a valuable resource for both novice and seasoned researchers interested in leveraging AI to tackle information engineering challenges, offering a comprehensive understanding of AI integration and its potential in these areas.

Dyslexia is a condition that reduces the cognitive abilities of the individuals. Dyslexic individuals (DI) face challenges in recognizing letters and words. Healthcare centers employ multiple modalities, including functional magnetic resonance imaging (MRI), Electroencephalogram (EEG), eye movements, and handwritten images for dyslexia detection (DD). A limited number of research studies employed multi-modalities for DD. The existing DD models demand substantial computational resources to identify dyslexia patterns. Ensemble learning (EL) approaches integrate the efficiency of multiple models to produce an optimal outcome. In this study, the authors developed an EL-based DD model for detecting dyslexia using MRI and EEG. An improved SWIN transformer was used to extract features from MRI images. A closed form continuous time neural networks (CFC) model was fine-tuned to detect crucial patterns from EEG data. The authors employed CatBoost and Dartbooster XGBoost (DXB) as base models for dyslexia prediction. An extremely randomized tree was used as meta-model to make a decision using base models outcomes. The model was developed based on two MRI datasets with an EEG dataset. The authors compared the proposed DD model’s performance against single and EL-based DD models. The findings indicated a significant improvement in DD using the model. The proposed model delivered an exceptional performance by achieving 98.5% and 98.7% on MRI and EEG datasets. Educational institutions and healthcare centers can benefit from the proposed DD model by rendering better services to DI. Generalizing the proposed model in diverse datasets can improve its performance. The proposed DD model’s performance can be enhanced using genome data.

Classification of volatile organic compounds (VOCs) is of interest in many fields. Examples include but are not limited to medicine, detection of explosives, and food quality control. Measurements collected with so-called electronic noses can be used for classification and analysis of VOCs. One type of electronic noses that has seen considerable development in recent years is Differential Mobility Spectrometry (DMS). DMS yields measurements that are visualized as dispersion plots that contain traces, also known as alpha curves. Current methods used for analyzing DMS dispersion plots do not usually utilize the information stored in the continuity of these traces, which suggests that alternative approaches should be investigated. In this work, for the first time, dispersion plots were interpreted as a series of measurements evolving sequentially. Thus, it was hypothesized that time-series classification algorithms can be effective for classification and analysis of dispersion plots. An extensive dataset of 900 dispersion plots for five chemicals measured at five flow rates and two concentrations was collected. The data was used to analyze the classification performance of six algorithms. The highest classification accuracy of 88% was achieved by a Long-Short Term Memory neural network, which supports the hypothesis that interpreting DMS measurements as sequential data is beneficial and outperformed classification algorithms traditionally used for DMS-based VOC identification.

Embedded with many advantages, the Wireless Access in Vehicular Environment (WAVE) was created to facilitate communication between vehicles and infrastructure (V2I) and among vehicles (V2V). Its ability to preserve equipment and lives was one of its most important and anticipated benefits. However, the complete integration of this life-saving technology into contemporary automobiles still faces technological obstacles after more than 10 years of intensive work by scientists and academia. The success of the WAVE completely relies on the regular exchange of Cooperative Awareness Messages (CAMs) and Decentralized Environmental Notification Messages (DENMs). Unfortunately, since CAMs and DENMs are broadcast by each mobile device in the network frequently, with the limited channel band of 10 MHz, communication resources quickly get exhausted due to broadcast storm issues, rendering further communication impossible. In any given scenario, if every mobile device involved could predict the Packet Error Rate (PER) based on factors such as mobility, Signal-to-Noise Ratio (SNR), and available bandwidth, then controlling the rate at which messages are broadcast might be possible, thus preventing or reducing broadcast storms. Consequently, the current work focuses on designing, developing, and modeling the PER in the VANET environment using Neural Network (NN). Based on the type and complexity of the challenge, the Levenberg-Marquardt Algorithm (LMA) was selected and used to develop the proposed system model. The response of the created and verified model demonstrated some efficacy and efficiency in predicting the PER based on the SNR and available bandwidth with up to 97% confidence.

This paper investigates the exponential synchronization problem for the proposed state-dependent spiking switched neural network. Under certain conditions, it is proven that state-dependent spiking switched systems can be transformed into fixed-time spiking switched systems, and exponential synchronization of homologous comparison systems implies the exact synchronization of the considered systems. Then, an exponential synchronization criterion is obtained for the proposed systems. Finally, three numerical examples are given to illustrate the validity of our results.

Cardiac motion tracking from echocardiography can be used to estimate and quantify myocardial motion within a cardiac cycle. It is a cost-efficient and effective approach for assessing myocardial function. However, ultrasound imaging has the inherent characteristics of spatially low resolution and temporally random noise, which leads to difficulties in obtaining reliable annotation. Thus it is difficult to perform supervised learning for motion tracking. In addition, there is no end-to-end unsupervised method currently in the literature. This paper presents a motion tracking method where unsupervised optical flow networks are designed with spatial reconstruction loss and temporal-consistency loss. Our proposed loss functions make use of the pair-wise and temporal correlation to estimate cardiac motion from noisy background. Experiments using a synthetic 4D echocardiography dataset has shown the effectiveness of our approach, and its superiority over existing methods on both accuracy and running speed.

One of the most common causes of maternal deaths during pregnancy is preeclampsia. As a result, it is critical to consider the resources required to provide timely care and avoid health complications. One of those resources is hospital beds. Thus, to explore in-depth the inpatient beds occupation to treat preeclampsia at the IESS Los Ceibos Hospital in Guayaquil-Ecuador, Descriptive, Diagnostic, Predictive, and Prescriptive (DDPP) analytics will be applied together to the medical records collected retrospectively. In the descriptive analysis, the effects of the COVID-19 pandemic can be seen during the period March-September 2020. The diagnostic analysis helps us to show a consistent upward trend in the time series. In the predictive analysis, techniques such as Dynamic Bayesian Networks (DBN), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM) are considered to forecast the future inpatient bed occupation. The model trained with MLP achieved the lowest MAPE, which was 11.58%. It predicts that in September and November of 2024, there will be peaks of 274 and 258 inpatient beds, respectively. Finally, in our prescriptive analysis, we show that closing the hospitalization unit would produce, on average, 130.33 monthly referrals. To our knowledge, we do not find any work that applies all the DDPP analytics together in preeclampsia. Moreover, to our knowledge, this is the first study that makes a joint application of DDPP analytics in univariate time series. Therefore, this work may serve as a basis for future studies in the joint application of all DDPP analytics to time series and to health.

Breast cancer is a prominent reason of death amidst females and is considered one of the most widespread types of cancer. Early detection is crucial for effective treatment. Research has demonstrated that computerized approaches can aid in the detection procedure and offer benefits beyond accustomed evaluation schemes. The study presents a contemporary approach for classifying breast cancer images by virtue of histopathology. This novel approach combines the fusion of ESRGAN and NSCT algorithms with a multi-attention-based CNN that utilizes wavelet transform. The original size of an image from the Patch Camelyon (PCam) dataset is $96\times 96$ , which is too small to extract features for the Convolutional Neural Network (CNN) effectively. An enhanced-super-resolution-generative-adversarial network, also known as the ESRGAN model, and the adaptive nonsubsampled contourlet transform were employed to rectify image quality. The proposed model was designed using Symlet, Coiflet, and a combination of both wavelets to identify which preserves better classification features and yields a promising result. The model proposed in this study utilizes the multi-head attention mechanism combined with CNN layers to operate in parallel multiple times. The model’s training, validation, and testing were conducted on the PCam dataset, which includes 327,680 histopathological images. The trained model was assessed using test data and achieved an accuracy of 93.67% and AUC of 94%, surpassing the performance of several other existing models.

Recent years witnessed increasing average annual precipitation in northern Xinjiang, amplifying the region’s vulnerability to natural disasters due to its fragile ecology and limited ground infiltration capacity. Therefore, the study of precipitation forecast in northern Xinjiang area is of vital significance to prevent these disasters. This study aims to capture the internal feature dependencies of complex multivariate meteorological data, build a PFformer model based on deep learning by reducing the algorithm time complexity of the short-term precipitation forecast model, use piecewise attention to reduce the algorithm complexity, and enhance the point-by-point dot product attention mechanism in the self-attention model. Data from five meteorological stations in Urumqi, Karamay, Tacheng, Altay, and Hami were used for the experiment. Comparing PFformer with mainstream time series prediction models, performance evaluation using Mean Squared Error (MSE) and Mean Absolute Error (MAE) indicates superior accuracy. Compared with the benchmark model, PFformer’s MSE improved by 7.85% on average and MAE improved by 9.74% on average. It shows potential for improving the accuracy of short-term precipitation forecasting, offering valuable assistance in disaster mitigation and management.

This review explores the integration of human posture recognition and classification technologies in performing arts education, focusing on the advancements in deep learning, neural networks, and computer vision. It traces the evolution of these technologies and highlights their significance in developing personalized teaching methods for performing arts. Following this, we examine various posture recognition technologies, emphasizing data acquisition and processing methods while revealing their strengths and limitations, particularly in their adaptability to performing arts. Moreover, the review discusses posture classification methods, addressing the processing and interpretation of data, and the challenges in achieving classification accuracy and efficiency. Illustrative case studies demonstrate the application of these technologies in enhancing teaching and creativity in performing arts, significantly contributing to the skill development and expressiveness of artists and students. Concluding with a critical assessment of the current feasibility of these technologies in performing arts education, the review suggests potential improvements and future directions. It emphasizes the indispensable role of human posture recognition in advancing performing arts education and underscores the need for continued interdisciplinary research in this evolving field.

The main drive system of the rolling mill often experiences torsional vibrations, which severely affect product quality, precision, and the service life of the transmission equipment. This paper investigates the torsional vibration suppression problem in the main drive system of the rolling mill, considering actuator faults, nonlinear friction, nonlinear damping, and model uncertainties. Based on the high-order fully actuated (HOFA) system approach, the main drive system of the rolling mill is transformed into a rolling mill main drive fully actuated system (RMMDFAS). Adaptive neural networks are introduced to address unknown uncertainties, and a continuous differentiable Gaussian error function is used to handle actuator faults. An adaptive neural network fault-tolerant control law for motor torque is proposed. The stability of the designed main drive torsional vibration system is rigorously proven, while maintaining the performance of the transformed states. Finally, the effectiveness and superiority of the proposed algorithm are verified through simulations.

In the past few decades, there have been multiple algorithms proposed for the purpose of solving optimization problems including Machine Learning (ML) applications. Among these algorithms, metaheuristics are an appropriate tool to solve these real problems. Also, ML is one of the advanced tools in Artificial Intelligence (AI) including different learning strategies to teach new tasks according to data. Therefore, proposing an efficient meta-heuristic to improve the inputs of the trainer in ML would be significant. In this study, a new idea centered on seed growth, Seed Growth Algorithm (SGA), as a conditional convergent evolutionary algorithm is proposed for optimizing several discrete and continuous optimization problems. SGA is used in the process of solving optimization test problems by neural networks. The problems are solved by the same neural network with and without SGA, computational results prove the efficiency of SGA in neural networks. Finally, SGA is proposed to solve very extensive test problems including IoT optimization problems. Comparative results of applying the SGA on all of these problems with different sizes are included, and the proposed algorithm suggests effective solutions within a reasonable timeframe.

In modern semiconductor manufacturing, optimizing plasma etching processes is key for advancing technology and achieving profitable production. In this study, we propose an efficient process flow to calibrate the etching mechanisms and optimize the etch profile by combining three-dimensional plasma simulations with machine learning (ML). Our proposed workflow speeds up the model calibration process and effectively addresses data scarcity issues by integrating extensive simulation data with TEM and other data sources. We have developed a three-dimensional Cl2/HBr/O2 plasma etching model for shallow trench isolation (STI), which is based on physical and chemical reaction mechanisms and external parameters consistent with actual production requirements, such as gas ratio, time duration, and chamber conditions. Machine learning is employed to tackle both forward and inverse problems. In the forward problem, the model predicts etching profiles under varying process conditions, with accuracy and similarity assessed using multiple criteria, including the structural similarity index (SSIM). Our approach achieves a notable reduction in prediction time, completing forecasts in 0.5 seconds compared to the 6500 CPU seconds required by TCAD. In the inverse problem, the model adjusts physical parameters to match observed results, improving its predictive accuracy. After model calibration, predictions are validated with wafer test outcomes, showing relative errors below 7%, thereby confirming the accuracy and robustness of the model parameters and the effectiveness of the entire process.

Predicting the closing price of the stock market with accuracy is highly uncertain and volatile. Deep learning (DL) can analyze vast amounts of historical stock data to identify patterns and correlations, aiding in predictive modeling. By learning from past market behavior, deep learning algorithms can potentially forecast future price movements with some degree of accuracy. This research presents a hybrid Long Short-Term Memory (LSTM) and Deep Neural Network (DNN) model designed to tackle the complexities of stock market prediction, including market volatility and intricate patterns. Initially the proposed model is being trained on Bajaj’s stock dataset. Getting noteworthy performance in single dataset does not prove robustness of the model. The research validates the model’s robustness and scalability through rigorous comparative analysis on 26 company’s stock datasets, achieving an average R-squared (R2) score of 0.98606, a Mean Absolute Error (MAE) of 0.0210, and a Mean Squared Error (MSE) of 0.00111. To assess the contribution of our proposed model, we retrained previously used deep learning models alongside our new approach, utilizing a shared dataset for validation. Additionally, we provide ablation study of LSTM-DNN model which provides insights into the individual contributions of each component towards detecting closing price of stocks, offering valuable information for optimizing future stock market prediction models. The model’s exceptional performance sets a new standard in stock market prediction, offering promising implications for investors, traders, and financial analysts. Making our work available open-source on https://github.com/codewithkhurshed/SMP-IUB can enhance its accessibility and promote future research opportunities in stock market price prediction.

Hydraulic jump is a common physical phenomenon in the field of hydraulic engineering. The essence of hydraulic jump is the conversion and dissipation of a large amount of energy due to the interaction between vortex structures, mainly released in the form of turbulence and water waves. This process significantly reduces the kinetic energy of water flow, thereby mitigating downstream erosion and protecting hydraulic structures, which in turn extends their service life. As a crucial factor in the energy dissipation design of discharge structures, the length of the hydraulic jump is influenced by various factors, including flow velocity, upstream and downstream water depths, riverbed roughness height, and Froude number. In this study, we applied dimensional analysis to identify the key parameters influencing hydraulic jumps on the dataset provided by literature. We utilized a multi-task learning strategy, incorporating a shared feature extraction layer for characteristic modeling of hydraulic jumps within Physics-Informed Neural Networks (PINNs). Furthermore, we compared the performance of PINNs with other data-driven models such as Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), and Transformers. The results demonstrated that these models are effective in estimating the length of hydraulic transitions and distinguishing between steady and unsteady hydraulic jump processes. Notably, the PINNs model exhibited better performance than other models, achieving an $\mathbf {R^{2}}$ score of $\mathbf {0.8818}$ , RMSE of 4.4627(cm), MAE of 3.3784(cm), precision of $\mathbf {0.9677}$ and recall of 0.9677 on the test set. These findings are significant for elucidating the characteristics and effects of hydraulic jumps in hydraulic structures, providing a scientific basis for the safe operation and design of practical hydraulic engineering projects.

Reinforcement Learning (RL) has been responsible for some of the most impressive advances in the field of Artificial Intelligence (AI). Research in competitive RL has shown that multiple agents competing in an adversarial environment can learn simultaneously in order to discover their optimal decision-making policies. Competitive RL algorithms have been used to train performant AI for a variety of games and optimization problems. Cybersecurity is a domain where the emerging research in competitive RL is being considered for its real-world application. In order to develop Automated Cyber Operations (ACO) tools using RL, various open-source environments are available to simulate network security incidents. However, the existing research in these environments is typically one-sided: a Red or Blue agent is trained to optimize their decision-making against a static opponent. Competitive RL has not been attempted in these emerging environments. In this work, we trained agents using competitive RL to approximate their game theory optimal policies in a simulated ACO environment. We showed that near-optimal behavior was reached gradually through fictitious play demonstrating that these strategies can be used to approximate the optimal policies for agents involved in sophisticated sequential decision-making during a cyber attack.

Plants play a crucial role in human history by providing essential resources such as food. However, they are susceptible to various diseases like leaf blight, grey spots, and rust, posing significant challenges for farmers and ranchers due to the potential damage and financial losses incurred. The current method of diagnosing plant diseases relies heavily on manual inspection by trained professionals, which becomes increasingly complex and resource-intensive, particularly on large farms. To address this issue, a novel approach is proposed, integrating clustering methods and neural networks to enhance the efficiency of identifying and studying plant disease samples. This method aims to expedite the diagnostic process and accurately quantify disease-induced damage. It introduces a formula for assessing damage, calculating the diseased leaf area as a fraction of the total leaf area. With an impressive accuracy rate of 96–97%, the system successfully identifies diseases such as Alternaria, Anthracnose, Bacterial Blight, and Cercospora Leaf Spot. This advancement signifies a notable improvement over previous methods, highlighting the evolution of the system. By harnessing the capabilities of clustering techniques and neural networks, this innovative approach seeks to revolutionize the detection and monitoring of plant diseases in agriculture, offering a more accessible and precise solution.

Missing data is a critical challenge in industrial data analysis, particularly during anomaly incidents caused by system equipment malfunctions or, more critically, by cyberattacks in industrial systems. It impedes effective imputation and compromises data integrity. Existing statistical and machine learning techniques struggle with heavily missing data, often failing to restore original data characteristics. To address this, we propose Anomaly Signal Imputation Using Latent Coordination Relations, a framework employing a variational autoencoder (VAE) to learn from complete data and establish a robust imputation model based on latent space coordination points. Experimental results from a water treatment testbed show significant improvements in output signal fidelity despite substantial data loss, outperforming conventional techniques.

Agriculture has been completely transformed by Deep Learning (DL) techniques, which allow for quick object localization and detection. However, because weeds and crops are similar in color, form, and texture, weed detection and categorization can be difficult. Advantages in object detection, recognition, and image classification can be obtained with deep learning (DL), a vital aspect of machine learning (ML). Because crops and weeds are similar, ML techniques have difficulty extracting and choosing distinguishing traits. This literature review demonstrates the potential of various DL methods for crop weed identification, localization, and classification. This research work investigates the present status of Deep Learning based weed identification and categorization systems. The majority of research employs supervised learning strategies, polishing pre-trained models on sizable, labeled datasets to achieve high accuracy. Innovations are driven by the need for sustainable weed management methods, and deep learning is demonstrating encouraging outcomes in image-based weed detection systems. To solve issues like resource scarcity, population increase, and climate change, precision agriculture holds great promise for the integration of AI with IoT-enabled farm equipment.

Recent strides in integrating artificial intelligence (AI) with production systems align with the trend towards highly automated manufacturing, demanding smarter machinery. This dovetails with the overarching vision of Industry 4.0, moving beyond conventional models towards employing AI for real-time modeling of production processes, enabling adaptable and learning-enabled models. This study focuses on leveraging cutting-edge deep learning techniques to monitor and classify tool wear using authentic image data from machining processes. Various deep learning algorithms, including CNN, AlexNet, EfficientNetB0, MobileNetV2, CoAtNet-0, and ResNet18, are explored for monitoring and measuring wear through images of machining chips. The collected images of machining chips are categorized as ‘Accepted’, ‘Unaccepted’, and ‘Optimal’. Due to imbalanced datasets, the study investigates two distinct strategies: upsampling and downsampling. The study also aimes to enhance sensitivity for a specific minority class to meet industrial requirements. The study showed that upsampling enhanced accuracy and almost fulfilled the stated requirements, whereas downsampling did not achieve the desired outcomes. The study evaluates and compares the effectiveness of recently introduced deep learning algorithms with other CNN-based architectures in classifying tool wear states in real-world scenarios. It sheds light on the challenges faced by the machining industry, particularly the prevalent issue of class imbalance in real-world machining data. The observed results indicate that ResNet18 and AlexNet outperform other algorithms, achieving a weighted average accuracy of 96% for both multiclass and binary classification problems when considering upsampled datasets. Consequently, the study concludes that both ResNet18 and AlexNet demonstrate adaptability to class imbalances, generalization to real-world machining scenarios, and competitive accuracy.

In non-cooperative communication systems such as radio spectrum resource regulation and modern electronic warfare, automatic modulation recognition is a key technology. Traditional modulation recognition methods mainly rely on manual feature extraction, decision theory, and recognition selection. The Deep Leaning (DL) algorithm automatically obtains signal features directly from massive data, and realizes feature extraction and recognition at the same time. However, most of the research on DL-AMR methods focuses on single input single output (SISO) systems, while there are few studies on DL-AMR methods in multiple-input, multiple-output (MIMO) systems, so the integration of deep learning models into modulation recognition of MIMO systems has attracted the attention of many researchers. The purpose of this paper is to provide a comprehensive review of modulation recognition methods for MIMOsystems based on DL. Firstly, the basic theory of MIMO and its derivative systems and modulation recognition is introduced in detail, then the traditional modulation recognition algorithms and deep learning-based modulation recognition algorithms of MIMO systems are introduced, and finally, on the basis of discussion and summary, the problems to be solved, the challenges and potential research directions are proposed.

Accurate tumor segmentation in PET/CT imaging is essential for the diagnosis and treatment of cancer, impacting therapeutic outcomes and patient management. Our study introduces a new approach integrating a Weighted Fusion Transformer Network to enhance the segmentation of tumor volumes. This method synergizes PET and CT modalities through a Fusion FormerU-Net architecture that employs convolutional neural networks alongside transformer blocks, aiming to leverage the unique advantages of each imaging modality. We evaluated the proposed approach using a multi-institutional dataset, applying key performance metrics such as Dice Similarity Coefficient aggregate, Jaccard Index, Volume Correlation, and Average Surface Distance to assess segmentation precision. The results indicate that the CT/PET/Fusion strategy significantly improves tumor delineation, outperforming traditional segmentation methods. The main findings suggest that this integrative approach could potentially redefine the standard for tumor segmentation in clinical practice. Lastly, the proposed approach offers a promising direction for enhancing the accuracy of oncological imaging, with implications for the improvement of patient-specific treatment strategies.

Gears are widely recognized as crucial components for power connections and transmission owing to their substantial loading capacity, transmission efficiency, and consistent power output. Given their complexity, they are prone to faults and require continuous condition monitoring (CM). The study proposes an innovative approach that combines signal analysis with data-driven techniques to monitor gear conditions accurately. By employing a shallow neural network (NN) model, which incorporates statistical time features and principal component analysis (PCA), this strategy achieves precise detection of seven distinct levels of gear wear. The feature generation technique utilizes a sliding and overlapping window to extract 15 statistical time features from the tri-axial vibration signals of the gearbox. The PCA has been applied to the feature set (FS) to retain relevant features, avoid overfitting during model training, and visualize the underlying structure of the training data. A comparative analysis with other leading classification methods, applied to both the original and PCA-reduced feature sets reveals that the shallow NN model, particularly when using the PCA-reduced FS, stands out for its exceptional performance. It achieves a remarkable test accuracy of 99.1% on experimental data having a linear time complexity, highlighting the model’s practicality for real-world applications.

Earthquakes pose a significant threat to urban areas, necessitating accurate forecasting models to mitigate their impact. This study focuses on earthquake forecasting in Los Angeles, a region with high seismic activity and limited research. We established a feature matrix for forecasting earthquakes within a 30-day period by analyzing the most predictive patterns from recent studies. Our model developed a subset of features capable of forecasting the highest magnitude of an earthquake. Using advanced machine learning algorithms and neural networks, our model achieved an accuracy of 69.14% in forecasting the maximum magnitude earthquake as one of the 6 categories. We aim to provide a useful guideline for future researchers.

Predictive analytics is concerned with making predictions of future outcomes based on past data using data statistics, machine learning, dynamic models and filtering algorithms. This paper provides a review of the theory of predictive analytics from an optimization perspective, including some new developments in convex optimization and quadratic neural networks. Several examples from many different areas show how the theory can be applied to analyze data and provide useful information to decision makers.

A model predictive control-based heuristic dynamic programming (HDP-MPC) online strategy is proposed in this paper, to further improve the fuel economy for the series hybrid loaders. The back propagation neural network (BPNN) is utilized to establish the dynamics model for series hybrid loader, to effectively characterize its real working process. Next, a HDP-MPC approximate optimal algorithm is proposed. And its convergence and the controlled system stability are separately proved. Besides, an online energy management strategy based on HDP-MPC is designed to implement the energy management controller. Finally, the BPNN dynamics model and online HDP-MPC strategy are simulated and validated. Through the experimental results analysis, contrasted with the analytic model, the established dynamics model can more effectively reflect the actual motion process of the series hybrid loader with higher accuracy. Moreover, compared with the fuzzy rule-based and adaptive equivalent consumption minimization strategy strategies, the designed HDP-MPC strategy can respectively achieve fuel-savings of 28.65% and 10.38%, supplying a new idea for online energy management of the series hybrid loaders.

Retinal optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) have emerged as promising, non-invasive, and cost-effective modalities for the early diagnosis of Alzheimer’s disease (AD). However, a comprehensive review of automated deep learning techniques for diagnosing AD or mild cognitive impairment (MCI) using OCT/OCTA data is lacking. We addressed this gap by conducting a systematic review using the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) guidelines. We systematically searched databases, including Scopus, PubMed, and Web of Science, and identified 16 important studies from an initial set of 4006 references. We then analyzed these studies through a structured framework, focusing on the key aspects of deep learning workflows for AD/MCI diagnosis using OCT-OCTA. This included dataset curation, model training, and validation methodologies. Our findings indicate a shift towards employing end-to-end deep learning models to directly analyze OCT/OCTA images in diagnosing AD/MCI, moving away from traditional machine learning approaches. However, we identified inconsistencies in the data collection methods across studies, leading to varied outcomes. We emphasize the need for longitudinal studies on early AD and MCI diagnosis, along with further research on interpretability tools to enhance model accuracy and reliability for clinical translation.

It has become a fundamental component of the electrical networking system, both in residential and industrial settings, to adopt advanced power meter architecture. With traditional smart meters, a bi-channel communication network is established between homes and utility companies, providing consumers with information regarding their daily power consumption. Manual meter reading, however, may result in inaccurate meter logging and incorrect billing criteria, which will lead to an increase in overhead costs associated with deploying meter readers and billing power consumption for each site within metropolitan and large urban areas. Moreover, the current smart metering system does not enable consumers to predict their future energy consumption, only providing insights into their current power consumption and accumulative costs. In order to address these issues, we propose a novel intelligent Software-Defined Control (SDC) super cluster with a comprehensive architecture based on SDN routing capabilities, which differs from conventional commercial smart meters. The developed micro cluster is enhanced to run full availability and high performance compared to traditional metering system. Moreover it deploys intelligent capabilities to predict the consumption of power per home, we implemented a polynomial model experimentally. Furthermore, we propose an intelligent Software-Defined Controller Gateway (SDN-GW) to serve as a traffic predictor between distributed metering nodes and the cloud data warehouse, eliminating congestion caused by the large volumes of traffic data generated periodically by the metering nodes. Based on the experimental results, the software-defined control system was estimated to have 97.75% percent accuracy in power prediction, and the traffic flow predictor demonstrated 98.79% percent accuracy in network traffic prediction. Furthermore, the proposed SDN-GW achieved 29.37% power consumption rate compared to standard routing engine.

This study presents the dynamic analysis for the fractional order economic and financial supply chain dynamical system using the supervised neural network performances aided with scale conjugate gradient. The investigations based on the fractional derivatives have been implemented to achieve the realistic and accurate performances of the fractional order economic and financial supply chain dynamical system. The mathematical form of the fractional order economic and financial supply chain dynamical system is categorized into three dynamics: rate of interest, investment cost, and price index. Three different variations based on the fractional order form of the economic and financial supply chain dynamical system have been numerically presented using the supervised neural network performances based on the scale conjugate gradient scheme. The selection of the data for solving the economic and financial supply chain dynamical system is taken as 80% for training, and 12% for testing, and 8% for endorsement. The accuracy of the proposed stochastic scheme is presented using the obtained and referenced Adam results. Rationality, capability and perfection are performed through the supervised neural network with scale conjugate gradient scheme performances-based together with the performances of correlation/regression, mean square error, state transition measures, and error histograms. Finally, a comparison of numerical results is examined and observed that the range of absolute error is between $10^{-05}$ to $10^{-07}$ which indicates that the proposed stochastic computing model can effectively analyze the economic and financial supply chain dynamical system.

In recent years, Explainable Artificial Intelligence (XAI) has attracted considerable attention from the research community, primarily focusing on elucidating the opaque decision-making processes inherent in complex black-box machine learning systems such as deep neural networks. This spike in interest originates from the widespread adoption of black-box models, particularly in critical domains like healthcare and fraud detection, highlighting the pressing need to understand and validate their decision-making mechanisms rigorously. In addition, prominent XAI techniques, including LIME (Local Interpretable Model-Agnostic Explanations) and SHAP (Shapley Additive exPlanations), rely on heuristics and cannot guarantee the correctness of the explanations provided. This article systematically addresses this critical issue associated with machine learning and deep learning models, underscoring XAI’s pivotal role in promoting model transparency to enhance decision-making quality. Furthermore, this study advocates integrating Formal Methods to provide correctness guarantees for black-box internal decision-making. The proposed methodology unfolds in three pivotal stages: firstly, training black-box models using neural networks to generate synthetic datasets; secondly, employing LIME and SHAP techniques to interpret the models and visualize their internal decision-making processes; and finally, training decision trees on the synthetic datasets to implement Formal Methods for ensuring the correctness of the black-box model’s decision-making. To validate this proposed approach, experimentation was conducted on four widely recognized medical datasets, including the Wisconsin Breast Cancer and Thyroid Cancer (TC) datasets, which are available in the UCI Machine Learning Repository. Specifically, this research represents a significant contribution by pioneering a novel approach that seamlessly integrates XAI and Formal Methods, thereby furnishing correctness guarantees for internal decision-making processes within the healthcare domain.

This study explores the contemporary landscape of integrating numerical algorithms, artificial intelligence (AI), and expert methodologies within the domain of nephrology. Focusing on automation and decision support, we scrutinize the impact of numerical algorithms on the precise evaluation of kidney function. Furthermore, we delve into the transformative potential of artificial intelligence, particularly in the realms of machine learning and deep learning applications, elucidating its role in early disease detection and the formulation of personalized treatment strategies. The synergy between computer-based tools and expert-driven approaches is examined, underscoring their collaborative role in enhancing the accuracy and dependability of diagnoses. Additionally, we address the ethical considerations and challenges associated with the incorporation of automation in nephrology. The paper provides a concise overview of in-depth analysis while illustrating the promising prospects of these innovative methodologies for reshaping nephrology research and patient care.

Fence intrusion detection system (FIDS) must ideally detect all malicious intrusions without producing false alarms arising from environmental sources. Classification of intrusions can provide security personnel with additional information regarding the level of threat. Modern FIDS are equipped with several sensing channels capable of recording fence vibrations either optically or mechanically. Introducing autonomy to each channel can improve the robustness of the FIDS as well as introduce resilience to changing conditions such as visibility/weather conditions. In this work, an autonomous accelerometer-based FIDS edge device capable of detecting and classifying intrusions is presented. The FIDS consists of three stages. First, threshold detection is used to flag potential intrusions allowing the microcontroller (MCU) to save power during idle state of fence. In the second stage, the threshold exceedence probability is evaluated to discriminate between background noise and human intrusions. An oscillator model was fitted to derive the parameters of the first two stages based on physical properties of the fence. Third, a convolutional neural network (CNN) was trained to classify the detected disturbances into two types namely rattling and climbing. The intrusion detection stages generated only a single false alarm from 17 hours of storm data while the classification stage produced a 5-fold cross validation accuracy of $\approx 90.5$ %. The intrusion detection and classification was implemented by rounding weights and using a custom CNN inference engine on an 8-bit MCU. The implementation showed no degradation in classification accuracy and no drift in sampling frequency during real-time operation.

In deep learning, various optimizers use a starting value for the learning rate. In general, this value is the one defined by default in the optimizer or chosen by the user after searching for a value giving the best performances of the studied neural network. Finding this value in the case of complex models with large-scale data, either by fine-tuning it manually or by applying a grid search algorithm requires several training sessions, which is penalizing in terms of learning time. To solve this problem, we propose an efficient optimizer using an automatic setting of the learning rate value. Learning then requires a single training session. The proposed optimizer is based on the inexact linear search of Armijo’s rule that we adapt for neural networks. We have experimented with this optimizer for several types of loss functions such as mean squared error, mean absolute error, and cross-entropy. Different types of models such as regression and classification are studied. Also, the robustness of this optimizer against the noisy labels is verified. Empirical results demonstrate that the proposed optimizer works well in practice, it outperforms other state-of-the-art optimizers that use default learning rate in terms of accuracy in addition to the time saving when searching for the optimal value of the learning rate.

This paper proposes a fault location and protection method for power distribution system with Distributed generation (DG) resources, considering fault impedance. A multi-layer perceptron (MLP) Artificial Neural Network (ANN)-based approach using the Levenberg-Marquardt algorithm to train the neural network for fault location and identification within distribution networks with DGs is developed. The proposed method emphasizes the incorporation of fault impedance in the analytic process. It also leverages the robust computational capabilities and inherent simplicity of ANNs to accurately detect the fault type and subsequently identify the fault location. Furthermore, it implements a zoning strategy, which partitions the distribution system into multiple independent sections to facilitate a more organized and effective fault management process. Simulation results and conducted analyses confirmed the effectiveness of the proposed approach in localizing faults within the network thereby minimizing downtime during faulty conditions and improving power system’s capabilities and reliability. Among the key advantages of the proposed fault detection approach are its ability to adeptly handle the complexities introduced by the wide integration of DGs in distribution networks.

The current study presents a novel, non-invasive method for estimating both systolic and diastolic blood pressure by combining photoplethysmogram (PPG) signals with physiological data, such as sex, age, weight, height, heart rate, and BMI, using two Gated Recurrent Units (GRUs) models. The first model processes dynamic patterns in PPG signals, while the second model incorporates physiological parameters. Both models are connected through a series of dense layers. To prepare the datasets for the GRU framework, rigorous preprocessing was conducted. This resulted in a robust architecture capable of accurately predicting systolic and diastolic blood pressure. The proposed method achieved a Mean Absolute Error (MAE) of 1.458 for systolic and 1.164 for diastolic blood pressure. These findings demonstrate the potential of this approach for continual and non-intrusive blood pressure monitoring in wearable health technology. The study’s results also make a significant contribution to the field of medical monitoring technology. The proposed solution addresses a major limitation in traditional blood pressure measurement practices and paves the way for advancements in personalized health monitoring, particularly for managing hypertension and cardiovascular conditions.

Accurate estimation of the state of charge (SoC) is crucial for optimizing battery performance, battery health estimation, and ensuring reliable operation. In recent years, deep learning techniques have shown promising results in capturing complex non-linear relationships between input features and estimated the SoC as they do not rely on battery models. However, a detailed comparison of these techniques is not available. Therefore, this paper presents a comprehensive comparison of SoC estimation using ANN (artificial neural network), LSTM (long short-term memory), BiLSTM (bidirectional LSTM), and GRU (gated recurrent unit) models applied to lithium-ion batteries. These models are trained using five distinct drive cycles and tested against four additional drive cycles from a publicly available dataset. The effectiveness is evaluated with RMSE (root mean squared error) and R2 score (coefficient of determination). The results provide insights into the applicability of ANN, LSTM, BiLSTM, and GRU for SoC estimation in lithium-ion battery systems, highlighting the advantages and limitations of each architecture. Additionally, the neuron effect is considered to examine the impact of the number of neurons in the hidden layer on the prediction of SoC. Comparing all models, it is observed that the BiLSTM network shows promising results, showing RMSE of 0.0097 and R2 value of 0.9987 when evaluated on one of the testing datasets, surpassing the performance of the other models. Further, the addition of temperature as a feature and the effect of external noise on the SoC estimation using the BiLSTM network are analyzed, highlighting the significance of accurate SoC estimation.

This paper presents a comprehensive study on fault identification in Hall sensors within Brushless Direct Current (BLDC) motor drives using neural networks. Detecting these faults is critical for optimizing motor performance, enhancing energy efficiency, and ensuring overall reliability. Our objective is to propose an accurate, compact, and efficient fault detection neural network model for real-time monitoring and swift responses to Hall sensor faults. Conventional methods are often computationally demanding and fail to detect subtle faults, leading to significant performance decline and potential catastrophic failures. To address this, we leverage MobileNet-based compact models and compare them with state-of-the-art neural network models to identify the most accurate, lightweight, and fast inference option. Our findings demonstrate that MobileNet models excel in detecting Hall sensor faults, achieving over 90% accuracy with significantly fewer parameters (less than five million) and an impressive inference time of under 25 milliseconds. This highlights MobileNet as a robust and efficient choice for Hall sensor fault detection in BLDC motor drives.

Speech remains a prevalent mode of communication powering various intelligent functions in human-computer interaction applications, notably in Speaker/Person Identification (PID) systems. However, there is a considerable population of Non-speaking and Minimal-speaking (NMS) individuals, who heavily rely on nonverbal vocalizations for communication, and the existing speech-based PID systems may not be suitable for users from this community. This study delves into the use of nonverbal vocalizations to identify NMS subjects, termed as NMS-PID, and explores the feasibility of developing an identification system, namely S-NMS-PID, that accommodates both speaking users (with speech input) and NMS users (with nonverbal-vocalization input). Leveraging the recently published ReCANVo dataset of NMS nonverbal vocalizations and our speech dataset, our experiments with multiple networks and acoustic features demonstrate promising results for NMS-PID and S-NMS-PID, evident in average accuracies ranging from 70% to 92%. The proposed convolutional recurrent neural network-based model, despite its smaller size, achieves results nearly on par with much deeper models such as VGG16 and ResNet50. Our findings also highlight the efficacy of Mel-frequency cepstral coefficients features compared to the spectrogram features. Furthermore, a two-step training strategy involving supervised contrastive learning for representation learning followed by fine-tuning with cross-entropy loss significantly enhances robustness and accuracy, particularly in classifying data from minority classes, enhancing overall performance. This study’s outcomes hold potential for tailoring human-computer interaction applications specifically for NMS users. Implementing NMS-PID and S-NMS-PID in security and authentication processes ensures secure and reliable user identification across diverse platforms, transcending sole reliance on speech-based methods.

Convolutional Neural Networks are being increasingly applied to the detection of anomalies in Computed Tomographies (CTs). The goal of this paper is to implement an automated Computer-Aided Detection (CADe) system for spinal lesions using CTs and Convolutional Neural Networks pre-trained on commercial datasets. The proposed pipeline works as follows. The CADe takes in input CT scans and is equipped with an intuitive GUI to allow physicians to use it as a support tool for diagnoses. From the CT scans, the CADe selects volumes containing the vertebrae and extracts 2D slices of these volumes. These slices are pre-processed and then analyzed using a VGG19 Convolutional Neural Network and a tailored, binary classifier. The neural network identifies healthy vertebrae and vertebrae containing lesions (e.g. metastases, primary tumors, lytic and sclerotic lesions). For training and testing purposes, we generated a dataset from CTs of vertebrae from patients treated at IRCCS Rizzoli Orthopaedics Institute of Bologna, Italy, between 2009 and 2019. Both healthy and lesioned vertebrae retrieved with different tomography machines and setups are considered. The dataset has been enlarged using data augmentation techniques and subsequently used to train a wide range of deep learning models. We perform an in-depth benchmark study to assess the performance of the considered classifier against different Convolutional Neural Networks pre-trained on the ImageNet dataset and exploiting Transfer Learning techniques. Leveraging Transfer Learning techniques, we reached 93.43% accuracy and a recall of 92.99%.

Controlling closed architecture industrial robot manipulators poses significant challenges due to limited access to inner controller configurations and specific control gain structures. The absence of open torque or voltage interfaces further compounds these difficulties. Consequently, traditional methods such as the computed-torque approach often prove inadequate when applied to closed architecture robots, widening the gap between advanced control algorithms and practical industrial requirements. In response, this paper introduces a unified framework that utilizes adaptive neural networks to tackle these challenges in controlling closed architecture industrial manipulators. Our approach operates independently of the robot’s dynamics, inner controller configuration, and control gain structure. We provide thorough evidence showcasing the boundedness of all control variables. Moreover, the proposed approach is versatile, allowing for the use of a single joint velocity controller across robotic manipulators employing closed control architecture, even under varying conditions. Our strategy streamlines implementation without requiring complex calculations for updating control variables. Experimental results and comparative studies are provided to illustrate the applicability and effectiveness of our proposed control strategy compared to existing approaches.

Deep Neural Networks have been shown to perform poorly or even fail altogether when deployed in real-world settings, despite exhibiting excellent performance on initial benchmarks. This typically occurs due to relative changes in the nature of the production data, often referred to as distributional shifts. In an attempt to increase the transparency, trustworthiness, and overall utility of deep learning systems, a growing body of work has been dedicated to developing distributional shift detectors. As part of our work, we investigate distributional shift detectors that utilize statistical tests of neural network-based representations of data. We show that these methods are prone to fail under sample-bias, which we argue is unavoidable in most practical machine learning systems. To mitigate this, we implement a novel distributional shift detection framework which explicitly accounts for sample-bias via a simple sample-selection procedure. In particular, we show that the effect of sample-bias can be significantly reduced by performing statistical tests against the most similar data in the training set, as opposed to the training set as a whole. We find that this improves the stability and accuracy of a variety of distributional shift detection methods on both covariate- and semantic-shifts, with improvements to balanced accuracy typically ranging between 0.1 and 0.2, and false-positive-rates often being eliminated altogether under bias.

The recommendation accuracy of traditional product recommendation systems is insufficient. Therefore, an improved deep factor decomposition machine algorithm combining adaptive regularization and attention mechanisms is proposed, and big data components are integrated to enable the algorithm to support more data input types. The publicly available datasets Criteo and Avazu from the Kaggle competition were selected for testing experiments in the study. The experimental results are as follows. During the training phase, after the convergence of each model, the loss function value of the model designed in this study is 1.26, which is the lowest among all comparison models. Moreover, when the number of recommended products is 7, the overall recommendation effect of each model is the best. The area under the curve of the subject operation characteristic curve of the model designed in this study on the Criteo dataset is 0.809, which is significantly higher than all comparison models. It is believed that this model has higher recommendation accuracy and can be used in application scenarios that require higher recommendation quality.

Network Slicing is an enabler for new use cases and an improved network performance, especially for 5G private networks, which opens new business opportunities for vendors and applications for customers. On the other hand, the slicing mechanism adds another level of complexity to network management that significantly increases total cost of ownership. Full automation is a must, which is also evident in the standardization work on autonomous and zero-touch mobile networks under the umbrella of 3GPP and ITU organizations. Moreover, there is a clear methodological gap in research related to mobile network slicing, i.e. capacity dimensioning and planning for such infrastructure. The concept of the network modeling tool has been updated with an emphasis on adding functionality of mobile network capacity dimensioning and planning, which is presented in this article. Data-driven framework with thoroughly verified methods is outlined (e.g., Prophet, Neural Networks, VARMAX and its univariate equivalent - ARMA). Special attention is paid to traffic forecasting as the basis for the dimensioning and planning process. We evaluate how to use the framework as a scenario simulator to estimate the impact of traffic changes in any slice on quality of service (namely throughput and delay) of all. Finally, we explain how this solution realizes the concept of Digital Twin-based network simulator.