Search

Intelligent fault diagnosis (IFD) has experienced tremendous progress owing to a great deal to deep learning (DL)-based methods over the decades. However, the "black box" nature of DL-based methods still seriously hinders wide applications in industry, especially in aero-engine IFD, and how to interpret the learned features is still a challenging problem. Furthermore, IFD based on vibration signals is often affected by the heavy noise, leading to a big drop in accuracy. To address these two problems, we develop a model-driven deep unrolling method to achieve ante-hoc interpretability, whose core is to unroll a corresponding optimization algorithm of a predefined model into a neural network, which is naturally interpretable and robust to noise attacks. Motivated by the recent multi-layer sparse coding (ML-SC) model, we herein propose to solve a general sparse coding (GSC) problem across different layers and deduce the corresponding layered GSC (LGSC) algorithm. Based on the ideology of deep unrolling, the proposed algorithm is unfolded into LGSC-Net, whose relationship with the convolutional neural network (CNN) is also discussed in depth. The effectiveness of the proposed model is verified by an aero-engine bevel gear fault experiment and a helical gear fault experiment with three kinds of adversarial noise attacks. The interpretability is also discussed from the perspective of the core of model-driven deep unrolling and its inductive reconstruction property., Competing Interests: Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper., (Copyright © 2022 ISA. Published by Elsevier Ltd. All rights reserved.)

In order to address the challenge of the small-scale, small-target, and complex scenes often encountered in offshore remote sensing image datasets, this paper employs an interpolation method to achieve super-resolution-assisted target detection. This approach aligns with the logic of popular GANs and generative diffusion networks in terms of super-resolution but is more lightweight. Additionally, the image count is expanded fivefold by supplementing the dataset with DOTA and data augmentation techniques. Framework-wise, based on the Faster R-CNN model, the combination of a residual backbone network and pyramid balancing structure enables our model to adapt to the characteristics of small-target scenarios. Moreover, the attention mechanism, random anchor re-selection strategy, and the strategy of replacing quantization operations with bilinear interpolation further enhance the model's detection capability at a low cost. Ablation experiments and comparative experiments show that, with a simple backbone, the algorithm in this paper achieves a mAP of 71.2% on the dataset, an improvement in accuracy of about 10% compared to the Faster R-CNN algorithm. [ABSTRACT FROM AUTHOR]

The rapid rise of the Internet of things (IoT) have brought the progress of electronic skin (e-skin). E-skin is used to imitate or even surpass the functions of human skin. Thermoregulating is one of the crucial functions of human skin, it is significant to develop a universal way to realize e-skin thermoregulating. Here, inspired by the sweat gland structure in human skin, we report a simple method for achieving dynamic thermoregulating, attributing to the temperature of microencapsulated paraffin remains unchanged when phase change occurs. Combining with the principle of triboelectric nanogenerator, a deep learning model is employed to recognize the output signals of handwriting different letters on ME-skin, and the recognition accuracy reaches 98.13%. Finally, real-time recognition and display of handwritings are successfully implemented by the ME-skin, which provides a general solution for thermoregulating e-skin and application direction for e-skin in the field of IoT. [ABSTRACT FROM AUTHOR]

Although traditional deep learning improves the accuracy of intelligent fault diagnosis, it suffers from a problem, which is that a change in working conditions may reduce the diagnostic accuracy. The reason for this phenomenon is that a change of working conditions influences the probability distributions. To solve this problem, domain adaptation is adopted to perform intelligent fault diagnosis. However, the design of regularization methods, such as maximum mean discrepancy (MMD), neglects the phenomenon of fault extension. Considering the property of fault extension, the paper sums up a concept named short-time consistency which means that ‘during stable operation, a failure does not expand over a short time period.’ Moreover, short-time consistent regularization is proposed to ensure that the output of the model meets the requirement for short-time consistency, and closed-set regularization is proposed to further solve the problem of ‘types of label drop’ when short-time consistent regularization is used. When the problem occurs, the number of predicted label types in the target domain is smaller than that in the source domain in the closed-set domain adaptation. Two types of regularization, namely entropy-based regularization and regularization based on the L 2 norm, are easily adopted in the final loss function. The proposed method is verified by experiments. [ABSTRACT FROM AUTHOR]

Post-hoc explaining approaches for deep learning (DL) models has attracted much attention in safety–critical applications such as rotating machinery intelligent fault diagnosis (IFD). However, with the help of the explanation techniques, the models are still fragile to domain shifts caused by varying speeds and loads without help for improving their cross-domain performance. Since humans in the decision-making loop are essential for a reliable diagnostic system to determine the reliability of diagnosis, this paper proposes a causal explaining guided domain generalization (CXDG) method to realize trust worthy IFD with human in the decision loop. Specifically, an explaining model is trained with the conditional mutual information, which is a causal strength metric, and utilized to tell the causal features in the input data as the attributions of the diagnostic model. A translation process of the attributions is proposed to make the explaining process understandable. Furthermore, the aim of this paper is not only explaining but also beyond that the diagnostic model is guided to focus on the causal features to improve its generalization ability in unseen domains. The effectiveness of the method is validated on two experiment datasets. The results show that the proposed method can both explain the attributions of the diagnosis model and be beneficial to the generalization ability. [ABSTRACT FROM AUTHOR]

Deep learning (DL) based methods have swept the field of mechanical fault diagnosis, because of the powerful ability of feature representation. However, many of existing DL methods fail in relationship mining between signals explicitly. Unlike those deep neural networks, graph convolutional networks (GCNs) taking graph data with topological structure as input is more efficient for data relationship mining, making GCN to be powerful for feature representation from graph data in non-Euclidean space. Nevertheless, existing GCNs have two limitations. First, most GCNs are constructed on unweighted graphs, considering importance of neighbors as the same, which is not in line with reality. Second, the receptive field of GCNs is fixed, which limits the effectiveness of GCNs for feature representation. To address these issues, a multireceptive field graph convolutional network (MRF-GCN) is proposed for effective intelligent fault diagnosis. In MRF-GCN, data samples are converted into weighted graphs to indicate differences in relationship of data samples. Moreover, MRF-GCN learns not only features from different receptive field, but also fuses learned features as an enhanced feature representation. To verify the efficacy of MRF-GCN for machine fault diagnosis, case studies are implemented, and the results show that MRF-GCN can achieve superior performance even under imbalanced dataset. [ABSTRACT FROM AUTHOR]

Recent progress on intelligent fault diagnosis (IFD) has greatly depended on deep representation learning and plenty of labeled data. However, machines often operate with various working conditions, or the target task has different distributions with the collected data used for training (the domain shift problem). Besides, the newly collected test data in the target domain are usually unlabeled, leading to unsupervised deep transfer learning (UDTL)-based IFD problem. Although it has achieved huge development, a standard and open source code framework and a comparative study for UDTL-based IFD are not yet established. In this article, we construct a new taxonomy and perform a comprehensive review of UDTL-based IFD according to different tasks. Comparative analysis of some typical methods and datasets reveals some open and essential issues in UDTL-based IFD, which is rarely studied, including transferability of features, the influence of backbones, negative transfer, physical priors, and so on. To emphasize the importance and reproducibility of UDTL-based IFD, the whole test framework will be released to the research community to facilitate future research. In summary, the released framework and comparative study can serve as an extended interface and basic results to carry out new studies on UDTL-based IFD. The code framework is available at https://github.com/ZhaoZhibin/UDTL. [ABSTRACT FROM AUTHOR]

Traditional deep learning assumes that the probability distributions of the test set and the training set are exactly the same. When the test samples and the training samples are from different working conditions, the distribution gap between different conditions might dramatically lower the accuracy. To address this problem, domain adaptation (DA) via joint using training and test sets is adopted to increase the final accuracy. The disadvantage of DA is that it requires test samples to participate in the training phase. However, sometimes it is impossible to collect the test samples beforehand. To solve this problem, this article proposes a conditional adversarial domain generalization with a single discriminator, which aims to extract domain-invariant features from different working conditions and generalize the features to unseen working conditions. To realize conditional adversarial training, a new conditional adversarial strategy that the feature extractor can allow the discriminator to distinguish the fault classes, but cannot distinguish the domains, is designed to better confuse the discriminator and generalize the features. Meanwhile, three loss functions, including mean square error (MSE), cross-entropy coupled with entropy, and cross-entropy combined with traditional adversarial methods, are proposed to achieve the new conditional adversarial strategy. The effectiveness of the conditional adversarial strategy is verified by training conditional generative adversarial networks (CGANs) on two image data sets, and the performance of the proposed domain generalization method for bearing fault diagnosis is also verified by bearing data sets. Compared with traditional conditional adversarial domain generalization, the proposed method can save computing resource consumption and improve accuracy that does not appear in the training phase. [ABSTRACT FROM AUTHOR]

Intelligent fault diagnosis is essential for downtime reduction in modern industries. However, domain (i.e., working condition) variants are unavoidable for fault diagnosis under changing environment. Though one can obtain a customized deep learning model for a given domain, it is not practical to design deep neural networks manually for multiple domains. Differentiable architecture search (DARTS), as an automatic machine learning technique, can automate the network design process for a specific domain efficiently by using hypernetwork and differentiable search strategy. Nevertheless, the representation of hypernetwork is restricted by several unfair factors, and the searched architecture of DARTS is overconfident. To address these issues, a one-shot neural architecture search approach, which involves two-stage learning, is proposed for efficient domain matching fault diagnosis. In the first training stage, the warmup and path-dropout strategies are taken to enhance the competitiveness of the parametric operators and alleviate the coadaptation problem to obtain an intuitively fair hypernetwork. In the second matching stage, variational inference is introduced into a differentiable search strategy to estimate the uncertainty of model matching, and a scale mixture prior is used to softly constrain the matching stage. Then, a candidate architectures set can be sampled and ordered from the posterior. Multidomain experiment is implemented by adding noise to the raw signal, and the proposed method outperforms four commonly used deep neural networks for aeroengine bevel gear fault diagnosis. [ABSTRACT FROM AUTHOR]

Rotating machinery intelligent diagnosis based on deep learning (DL) has gone through tremendous progress, which can help reduce costly breakdowns. However, different datasets and hyper-parameters are recommended to be used, and few open source codes are publicly available, resulting in unfair comparisons and ineffective improvement. To address these issues, we perform a comprehensive evaluation of four models, including multi-layer perception (MLP), auto-encoder (AE), convolutional neural network (CNN), and recurrent neural network (RNN), with seven datasets to provide a benchmark study. We first gather nine publicly available datasets and give a comprehensive benchmark study of DL-based models with two data split strategies, five input formats, three normalization methods, and four augmentation methods. Second, we integrate the whole evaluation codes into a code library and release it to the public for better comparisons. Third, we use specific-designed cases to point out the existing issues, including class imbalance, generalization ability, interpretability, few-shot learning, and model selection. Finally, we release a unified code framework for comparing and testing models fairly and quickly, emphasize the importance of open source codes, provide the baseline accuracy (a lower bound), and discuss existing issues in this field. The code library is available at: https://github.com/ZhaoZhibin/DL-based-Intelligent-Diagnosis-Benchmark. • Intelligent diagnosis occupies an important position in avoiding costly breakdowns. • The benchmark accuracy is provided based on seven datasets and nine models. • A code library is released to test the performance. • Existing issues are pointed out by specific-designed cases. [ABSTRACT FROM AUTHOR]

Deep learning with ability to feature learning and nonlinear function approximation has shown its effectiveness for machine fault prediction. While, how to transfer a deep network trained by historical failure data for prediction of a new object is rarely researched. In this paper, a deep transfer learning (DTL) network based on sparse autoencoder (SAE) is presented. In the DTL method, three transfer strategies, that is, weight transfer, transfer learning of hidden feature, and weight update, are used to transfer an SAE trained by historical failure data to a new object. By these strategies, prediction of the new object without supervised information for training is achieved. Moreover, the learned features by deep transfer network for the new object share joint and similar characteristic to that of historical failure data, which is beneficial to accurate prediction. Case study on remaining useful life (RUL) prediction of cutting tool is performed to validate effectiveness of the DTL method. An SAE network is first trained by run-to-failure data with RUL information of a cutting tool in an off-line process. The trained network is then transferred to a new tool under operation for on-line RUL prediction. The prediction result with high accuracy shows advantage of the DTL method for RUL prediction. [ABSTRACT FROM AUTHOR]

As the basic hydrological parameters, the concentration and type of suspended sediment particles greatly influence the aquatic ecological environment and the water conservancy infrastructure. However, because of the various composition and changing concentration, challenges still remain in low-cost and real-time sediment monitoring. Herein, we report a potential method to realize real-time monitoring of sediment particles parameters by introducing a particles-laden droplet-driven triboelectric nanogenerator (PLDD-TENG) combined with deep learning method. The mechanism of PLDD-TENG was proved to be induced by the liquid-PTFE electrification and particles-electrode electrostatic induction. The output signals of PLDD-TENG were measured under different particles types and mass fractions. The results indicated that the output signals were sensitive to the particles type and mass fraction. A convolutional neural network (CNN) deep learning model was adopted to identify the particles parameters based on the output signals of PLDD-TENG and high identifying accuracy was achieved. Meanwhile, this model can identify particles types in the mixed solution. An intelligent system was finally developed to realize visualization of real-time monitoring for sediment particles. These findings are crucial in both fundamental understanding and application prospect of triboelectric effect in two-phase flow. [Display omitted] This work presented an effective method for real-time monitoring of sediment particles parameters using a particles-laden droplet-driven triboelectric nanogenerator (PLDD-TENG) and deep learning. Compared with DI water, the output signals of PLDD-TENG show the reduction of major peaks and minor peaks generation, which is proved that the addition of particles will change the charge of droplet and rapid movement of charged particles will arise the electrostatic induction on copper electrode. Finite element simulations were conducted in order to better understand the mechanism we proposed. The output signals of PLDD-TENG are sensitive to particles type and concentration. Meanwhile, high accuracy for particles type and concentration identification are realized and the accuracies exceed 90% for all targets based on the CNN model. Besides, the proposed method can identify the particles types in the mixed solution. This work helps better understanding the mechanism of triboelectric effect in two-phase flow and provides a low-cost and simple method to monitor particles parameters in the rivers. • The mechanism of a particles-laden droplet-driven triboelectric nanogenerator (PLDD-TENG) is clearly demonstrated. • High accuracies of particle type and mass fraction recognition are realized using output of PLDD-TENG and deep learning. • An intelligent system is developed to visualize the process of sediment particles monitoring. [ABSTRACT FROM AUTHOR]

• Time-series augmentation and balanced loss to address minority overfitting. • Data-rebalanced sampler and boundary balancing to address majority domination. • Integrate minority overfitting-based and majority domination-based strategies. • A comprehensive model-agnostic framework for class-imbalanced diagnosis. Considering the difficulty of data acquisition in industry, especially for failure data of large-scale equipment, classification with these class-imbalanced datasets can lead to the problems of minority categories overfitting and majority categories domination. A model-agnostic framework towards class-imbalanced fault diagnosis requirement is proposed to systematically alleviate these problems. Four sub-modules, including Time-series Data Augmentation, Data-Rebalanced sampler, Balanced Margin Loss, and classifier with Dynamic Decision Boundary Balancing are performed to improve recognition accuracy of minority categories without performance degradation on majority categories. Meanwhile, the framework is compatible with general neural networks and provides flexible model candidates to meet the need of feature extraction for different data types. Three case studies on public datasets demonstrate that proposed framework outperformed various state-of-the-art methods. [ABSTRACT FROM AUTHOR]

• A novel multi-sensor features fusion approach for remaining useful life prediction is proposed;. • Multiple sensors are constructed to a sensor network to generate spatial-temporal graphs. • Hierarchical attention graph convolutional network is constructed to model the generated spatialtemporal graphs. • Regularized self-attention graph pooling is proposed to obtain more informative graph representations. Deep learning-based prognostic methods have achieved great success in remaining useful life (RUL) prediction, since degradation information of machine can be adequately mined by deep learning techniques. However, these methods suffer from following weaknesses, that is, 1) interactions among multiple sensors are not explicitly considered; 2) they are more inclined to model temporal dependencies while ignoring spatial dependencies of sensors. To address those weaknesses, the multiple sensors are constructed to a sensor network and hierarchical attention graph convolutional network (HAGCN) is proposed in this paper for modeling the sensor network. In HAGCN, the hierarchical graph representation layer is proposed for modeling spatial dependencies of sensors and bi-directional long short-term memory network is used for modeling temporal dependencies of sensor measurements. Moreover, a regularized self-attention graph pooling is designed in HAGCN to achieve effective information fusion of the sensors. To realize prognostics, the spatial-temporal graphs are firstly generated based on the sensor network. Then, HAGCN is applied to model the spatial and temporal dependencies of the graphs simultaneously. The experimental results of two case studies show the superiority of HAGCN over state-of-the-art methods for RUL prediction. [ABSTRACT FROM AUTHOR]

Powder bed quality is critical in Laser Powder Bed Fusion (LPBF) as defects in the powder bed will affect the quality of the forming parts. Recently, numerous powder bed defect detection methods based on semantic segmentation algorithms have been developed. However, due to the computational resource constraints and real-time requirements in industry scenarios, these heavy models face challenges in deploying into the LPBF process monitoring system. Meanwhile, existing lightweight models struggle to achieve sufficient performance on industrial data due to their limited feature extraction capabilities. To address the challenges in industrial deployment, this paper proposes a lightweight and efficient LPBF layer-wise defect detection paradigm to achieve a balance between performance and efficiency. Aiming at improving the performance of the lightweight model, this paper proposes a structured Knowledge Distillation framework, in which a pre-trained high-performance segmentation model (UNet++) is used to guide the training process of a lightweight model. To address the impact of model discrepancies on the effectiveness of Knowledge Distillation and further improve the inference speed, this paper constructs a specific lightweight model and applies Structural Re-parameterization to deeply compress the model. Moreover, we created a dataset comprising 406 images of powder-bed defects, with each image annotated at the pixel level. Massive experiments based on the dataset demonstrate the effectiveness of the proposed paradigm, with an excellent trade-off in performance and efficiency. The speed of the proposed model is approximately 5 times faster than that of high-performance models while maintaining comparable performance. • A lightweight and efficient paradigm for LPBF in-situ defect detection. • Introduced Structure Re-parameterization to enhance the Knowledge Distillation. • An excellent trade-off in performance and efficiency for model deployment. [ABSTRACT FROM AUTHOR]

• A domain generalization method is proposed. • A structural causal model is defined to represent fault diagnosis in varying speeds. • Independence and shift sparsity of features are viewed vital for generalization. • General correlation and sparsity measure are designed to realize the vital properties. • The superiority of the method is verified on two real datasets. Intelligent fault diagnosis (IFD) based on deep learning (DL) has demonstrated its powerful performance to promote the reliability and safe operation of rotating machinery. In industrial applications, working condition changes frequently, leading to models' performance degeneration due to out-of-distribution (OOD) samples, such as vibration signals with unseen rotating speeds, loads, etc. To address this issue, domain generalization fault diagnosis (DGFD) has attracted increasing attention. However, mainstream DGFD methods are implemented from statical dependence, which fail to excavate intrinsic causal mechanisms and have limited performance and reliability. From a causal view, the causal independence and sparse shift network (CIS2N) is proposed to diagnose via causal mechanisms. Specially, a structural causal model (SCM) representing causal relations of features in IFD is given. Based on the SCM, it is derived that features are jointly independent and feature deviation between intervened sample pairs is sparse. The dependence among features is measured via the covariance matrices constructed by high-level features mapped by random Fourier features (RFF). Meanwhile the L1-distance between pairs of features with the same causal features (intervened pair) is calculated and optimized to realize shift sparsity. The effectiveness of the proposed CIS2N on DGFD is demonstrated on two real helical gearbox datasets. [ABSTRACT FROM AUTHOR]

As a multi-resolution analysis method rooted rigorously in mathematics, wavelet transform (WT) has shown its great potential in rotary machine fault diagnosis, characterized by continued development and innovative new applications. In traditional fault diagnosis, WT has been widely used for fault feature extraction and extensively studied for performance improvement. With the emergence of data-driven intelligent fault diagnosis, especially deep learning techniques, WT has attracted renewed attention for its ability of adding interpretability into the intelligent diagnosis models. This paper aims to highlight the advancement of WT-based fault diagnosis research over the last decade. Toward this end, a comprehensive overview of WT method is given, followed by a summary of WT for fault diagnosis from two perspectives: traditional fault diagnosis and intelligent fault diagnosis. Finally, future research trends are discussed, including benchmarking, wavelet base design, integration with other methods, and enhancement through deep learning. [ABSTRACT FROM AUTHOR]

Cardiac abnormality detection from Electrocardiogram (ECG) signals is a common task for cardiologists. To facilitate efficient and objective detection, automated ECG classification by using deep learning based methods have been developed in recent years. Despite their impressive performance, these methods perform poorly when presented with cardiac abnormalities that are not well represented, or absent, in the training data. To this end, we propose a novel one-class classification based ECG anomaly detection generative adversarial network (GAN). Specifically, we embedded a Bi-directional Long-Short Term Memory (Bi-LSTM) layer into a GAN architecture and used a mini-batch discrimination training strategy in the discriminator to synthesis ECG signals. Our method generates samples to match the data distribution from normal signals of healthy group so that a generalised anomaly detector can be built reliably. The experimental results demonstrate our method outperforms several state-of-the-art semi-supervised learning based ECG anomaly detection algorithms and robustly detects the unknown anomaly class in the MIT-BIH arrhythmia database. Experiments show that our method achieves the accuracy of 95.5% and AUC of 95.9% which outperforms the most competitive baseline by 0.7% and 1.7% respectively. Our method may prove to be a helpful diagnostic method for helping cardiologists identify arrhythmias. • A novel ECG anomaly detection for the one-class classification analysis task. • We embed temporal information into GAN for better signal synthesis. • The model is more generalisable with the high-dimensional ECG data. [ABSTRACT FROM AUTHOR]

Three main challenges in industrial wear detection that limited data-availability and time-consuming annotations, small areas of initial wear, and sensitivity to variable light, have impeded the real-world applications of deep learning-based methods. To this end, we propose RobustFlow, an unsupervised method based on the normalizing flow and attention mechanism. In our work, only the wear-free images are required for training, and then the trained model can be employed to detect and segment wear. Extensive experiments have demonstrated that RobustFlow can achieve predominant robustness in real-world wear detection and segmentation, especially for wear with small regions and variable light. Overall, our work provides a promising paradigm for wear detection and segmentation in real-world industry. [ABSTRACT FROM AUTHOR]