Feature Learning Method Research Articles

Acceleration-Guided Acoustic Signal Denoising Framework Based on Learnable Wavelet Transform Applied to Slab Track Condition Monitoring

Acoustic monitoring has recently shown great potential in the diagnosis of infrastructure condition. However, due to the severe noise interference in acoustic signals, meaningful features tend to be difficult to infer. It creates a considerable obstacle for an extensive application of acoustic monitoring. To tackle this problem, we propose an acceleration-guided acoustic signal denoising framework (AG-ASDF) based on learnable wavelet transform to automatically denoise the acoustic signal and extract the relevant features based on the acceleration signal. This denoising framework requires the acceleration signal only for the training stage. Therefore, only acoustic sensors (non-intrusive) need to be installed during the application phase, which is convenient and crucial for the condition monitoring of safety-critical infrastructure. A comparative study is conducted among the proposed AG-ASDF and other feature learning / extraction methods, by using a multi-class support vector machine to evaluate the detection effectiveness of slab track condition based on acoustic signals. Different healthy and unhealthy states of slab tracks are imitated with three types of slab track supporting conditions in a railway test line. The classification based on the proposed AG-ASDF features outperforms other feature extraction and learning methods with a significant accuracy improvement.

Center-based Transfer Feature Learning With Classifier Adaptation for surface defect recognition

Surface defect recognition using Deep Learning based computer vision techniques is an important task in industrial manufacturing. However, surface images have different distributions due to different environments in industrial manufacturing, where the distribution difference between images will degrade the accuracy of Deep Learning based computer vision techniques. In addition, existing Transfer Feature Learning (TFL) methods reduced the distribution difference only by the location parameters of distributions, which ignored the effect of scale parameters in representing the distribution. To overcome these problems, we propose Center-based Transfer Feature Learning with Classifier Adaptation (CTFLCA) for surface defect recognition. First, to eliminate the distribution difference at the feature layer from the location parameters and scale parameters of distributions, we utilize centers as bases to propose the Center-based Transfer Feature Learning method (CTFL) by minimizing Center-based Distribution Difference of Location Parameters (CDDLP) and Center-based Distribution Difference of Scale Parameters (CDDSP). Second, to reduce the distribution difference at the classifier layer from the location parameters and scale parameters, we establish the Center-based Classifier Adaptation method (CCA) using a similar idea of CTFL, where the optimization objective of CCA is formulated by minimizing the classification errors, CDDLP in classification results, and CDDSP in classification results. Next, under the guidance of the class-wise sample selection, we establish CTFLCA for integrating CTFL with CCA. Finally, sufficient results on four datasets (NEU-C, PR-C, Office-Caltech, and ImageCLEF-DA) illustrate the effectiveness of CTFLCA, where the average classification accuracies of CTFLCA are 99.6%, 98.0%, 95.1%, and 91.1%.

Lead Separation and Combination: A Novel Unsupervised 12-Lead ECG Feature Learning Framework for Internet of Medical Things

The development of healthcare industry, especially Internet of Medical Things (IoMT), has generated considerable unlabeled electrocardiogram (ECG) signals. This article proposes a new unsupervised feature learning method for these unlabeled 12-lead ECGs, a type of 12-channel 1-D time series. Based on contrastive predictive coding (CPC), it considers the characteristics of 12-lead ECGs and develops novel lead-separation CPC (LSCPC) and lead-combination CPC (LCCPC). Specifically, LSCPC captures intralead features for each lead, while LCCPC combines all the leads and explores interlead relationships. Furthermore, a fusion model of LSCPC and LCCPC generates final representations. The Physikalisch-Technische Bundesanstalt (PTB)-XL database that contains 21837 12-lead records is used for unsupervised feature learning. Using learned features, linear classifiers are trained to accomplish the downstream tasks. 448 ECG records from 148 myocardial infarction (MI) and 52 healthy control subjects of the PTB database are used for MI detection. 6877 records from the CPSC-2018 database are used for atrial fibrillation (AF) detection, including 918 normal records, 1098 AF records, and 4861 other records. Using fivefold cross-validation, our model achieves 90.38% and 73.27% accuracy in MI and AF detection, respectively. Compared with existing models, it improves the performances by at least 2.32% for MI detection and 3.99% for AF detection. The model has been deployed on a lightweight embedded system (800-MHz ARM processor, 1-GB RAM). The maximum latency is only 465.34 ms, which can satisfy the real-time constraints. Overall, all the results have demonstrated the potential of our method for real-world healthcare, and lightweight IoMT applications.

Weakly-Supervised Part-Attention and Mentored Networks for Vehicle Re-Identification

Vehicle re-identification (Re-ID) aims to retrieve images with the same vehicle ID across different cameras. Current part-level feature learning methods typically detect vehicle parts via uniform division, outside tools, or attention modeling. However, such part features often require expensive additional annotations and cause sub-optimal performance in case of unreliable part mask predictions. In this paper, we propose a weakly-supervised Part-Attention Network (PANet) and Part-Mentored Network (PMNet) for Vehicle Re-ID. Firstly, PANet localizes vehicle parts via part-relevant channel recalibration and cluster-based mask generation without vehicle part supervisory information. Secondly, PMNet leverages teacher-student guided learning to distill vehicle part-specific features from PANet and performs multi-scale global-part feature extraction. During inference, PMNet can adaptively extract discriminative part features without part localization by PANet, preventing unstable part mask predictions. We address this Re-ID issue as a multi-task problem and adopt Homoscedastic Uncertainty to learn the optimal weighing of ID losses. Experiments are conducted on two public benchmarks, showing that our approach outperforms recent methods, which require no extra annotations by an average increase of 3.0% in CMC@5 on VehicleID and over 1.4% in mAP on VeRi776. Moreover, our method can extend to the occluded vehicle Re-ID task and exhibits good generalization ability.

A neural network approach for short-term water demand forecasting based on a sparse autoencoder

Abstract This paper presents a backpropagation neural network (BPNN) approach based on the sparse autoencoder (SAE) for short-term water demand forecasting. In this method, the SAE is used as a feature learning method to extract useful information from hourly water demand data in an unsupervised manner. After that, the extracted information is employed to optimize the initial weights and thresholds of the BPNN. In addition, to enhance the effectiveness of the proposed method, data reconstruction is implemented to create suitable samples for the BPNN, and the early stopping method is employed to overcome the BPNN overfitting problem. Data collected from a real-world water distribution system are used to verify the effectiveness of the proposed method, and a comparison with the BPNN and other BPNN-based methods which integrate the BPNN with particle swarm optimization (PSO) and the mind evolutionary algorithm (MEA), respectively, is conducted. The results show that the proposed method can achieve fairly accurate and stable forecasts with a 2.31% mean absolute percentage error (MAPE) and 320 m3/h root mean squared error (RMSE). Compared with the BPNN, PSO–BPNN and MEA–BPNN models, the proposed method gains MAPE improvements of 5.80, 3.33 and 3.89%, respectively. In terms of the RMSE, promising improvements (i.e., 5.27, 2.73 and 3.33%, respectively) can be obtained.

Researches Advanced in the Development and Application of Transformers

The basic task of feature learning is to use algorithms to allow machines to automatically learn useful data and its features during the model building process. The quality of the learned features will greatly affect the results of downstream tasks. Early feature learning methods relied on handcrafted features. Thanks to the development of deep learning, feature learning methods based on convolutional neural networks have greatly improved the quality of features. However, with the increasing scale of training data and the increasing complexity of modeling tasks, deep neural network Transformer based on self-attention mechanism and parallel data processing has gradually become a new research hotspot. Transformer can adaptively and selectively select contextual information and key semantic information in a scene by covering attention networks and fully connected layers and has become an important research area for computer vision and natural language processing tasks. This paper reviews the basic principles and development of Transformer, focuses on its application in CV and NLP, and analyzes effective Transformer-based models. Finally, the challenges faced, and future trends of transformer models are summarized.

A novel information fusion approach using weighted neural networks for intelligent multi-class diagnostics of rotating machinery with unseen working conditions

Analytical diagnostics tools have been proved effective. However, these techniques are unable to be adaptively implemented for diverse machines with various defect modes. Big data-based diagnostics tools have been demonstrated efficient, but they exploit massive data with a high computational cost and time that is impractical for online condition monitoring. Furthermore, the health signals may be acquired under unseen working conditions in real machines. These are the major challenges for previous diagnostics methods and make them infeasible in real-world applications. So, a novel technique is proposed in this research to solve mentioned problems. The main novelty of this method is to fuse information of various processing functions for multi-sensor signals and to apply an efficient feature bank for performing an efficacious feature learning method. The other novelty of this research is to select transferable features for diagnostics of rotating machinery under unseen working conditions. Mean squared error function for the improvement of diagnostics accuracy and transferable feature space with semidefinite dimension for the improvement of diagnostics speed and better visualization for the source and target machines with different working conditions are minimized jointly via weighted neural networks. This approach is verified by two case studies of machinery vibration datasets for the robustness analysis of bearing multi-fault diagnosis problems under various operational conditions for the sensitivity analysis. Results indicate that the proposed algorithm achieves good performance in real rotating machinery diagnostics with unseen working conditions.

Few-Shot Class-Incremental Learning via Compact and Separable Features for Fine-Grained Vehicle Recognition

Most of the existing deep learning-based fine-grained vehicle recognition methods collect a large-scale training set in advance and train a model based on the closed-world assumption. However, in the real world, new classes of vehicles are released over time but it is difficult to collect sufficient labeled data for new classes, which results in a typical few-shot class-incremental learning problem (FSCIL). To solve this problem, this work proposes a compact and separable feature learning method (CSFL) which exploits a decoupled learning scheme to prevent the feature extractor from updating during class-incremental learning. CSFL trains an initial model to learn discriminative features of fine-grained vehicles using deep metric learning. Then an incremental linear discriminant analysis algorithm is applied to the learned features to further discriminate potential confused classes. Specifically, the decoupled components share the same objective of enhancing intra-class compactness and inter-class separability, which is beneficial for classification. Extensive experiments on three fine-grained vehicle datasets demonstrate that the proposed CSFL achieves better results than state-of-the-art incremental learning methods, validating the importance of compact and separable features in the problem of FSCIL.

Binary feature learning with local spectral context-aware attention for classification of hyperspectral images

The classification of hyperspectral images (HSIs) has achieved success in applications. For many approaches, features are directly extracted from whole spectral pixels, which can not well describe local characteristics. These methods are also susceptible to noise since each feature code is learned individually. Accordingly, a binary feature learning method with local spectral context-aware attention (BFLSC) is proposed for the classification. Specifically, for training samples, we first build the local spectrum models (LSMs) to describe local spectral properties, where each training sample is segmented into some parts and the difference between the central value and its neighborhoods is calculated in each part. Then, we construct the BFLSC model to learn a projection and binary features of training samples. In such model, the spectral context-awareness attention is established to collaboratively learn binary feature codes by enforcing one shift between 0/1 of each LSM, which enhances the robustness and stability of binary leaning. We also introduce the loss constraint, even distribution constraint, and variance constraint to reduce information loss and improve the quality of learned feature distribution. Additionally, an optimization scheme is designed to obtain the solution of the BFLSC model. Further, the learned binary features are added to train the support vector machine (SVM). For each testing sample, the LSMs are first extracted, and then mapped into binary features by the learned projection. The trained SVM is finally used for the mapped binary features to predict the label of the testing sample. Experimental results validate that our BFLSC realizes the better performance compared with some advanced approaches.

A Novel Spatiotemporal Process Feature Learning Method Based On the Pseudo-Siamese Network for Complex Chemical Process Concurrent Condition Monitoring.

The deep learning-based process monitoring method has attracted great attention due to its ability to deal with nonlinear correlation. However, the further modeling of learned deep features from process data to better depict typical process features to obtain more precise monitoring results remains a challenge. In this paper, a novel nonlinear spatiotemporal process feature learning method is proposed to extract high-value slow-varying spatiotemporal process features, with an explicit temporal relationship model for the concurrent monitoring of the static deviation and the dynamic anomaly of complex chemical processes. Different from directly mixed spatiotemporal information methods, the pseudo-Siamese autoencoder network is designed with two different subencoders to separately describe the nonlinear spatial and temporal relationships of the nonlinear dynamic input data. Correspondingly, a cost function including three losses and one orthogonal constraint is proposed to make sure that the extracted spatiotemporal process features change as slowly as possible and contain the most nonlinear dynamic information on the input data. With the explicit spatial and temporal relationship submodel, predictions are utilized to shrink the variability of the nonlinear temporal correlated data and focus on the unpredictable variabilities to improve process monitoring performance. Meanwhile, the linear dynamic information is further extracted in the reconstructed residual space by the general slow feature analysis (SFA) method to provide a more detailed analysis of the process characteristics and improve the monitoring results. The case study monitoring results demonstrate the effectiveness and superiority of the proposed method over other compared methods for concurrent process monitoring.

Three-Stream and Double Attention-Based DenseNet-BiLSTM for Fine Land Cover Classification of Complex Mining Landscapes

The fine classification of land cover around complex mining areas is important for environmental protection and sustainable development. Although some advances have been made in the utilization of high-resolution remote sensing imagery and classification algorithms, the following issues still remain: (1) how the multimodal spectral–spatial and topographic features can be learned for complex mining areas; (2) how the key features can be extracted; and (3) how the contextual information can be captured among different features. In this study, we proposed a novel model comprising the following three main strategies: (1) design comprising a three-stream multimodal feature learning and post-fusion method; (2) integration of deep separable asymmetric convolution blocks and parallel channel and spatial attention mechanisms into the DenseNet architecture; and (3) use of a bidirectional long short-term memory (BiLSTM) network to further learn cross-channel context features. The experiments were carried out in Wuhan City, China using ZiYuan-3 imagery. The proposed model was found to exhibit a better performance than other models, with an overall accuracy of 98.65% ± 0.05% and an improvement of 4.03% over the basic model. In addition, the proposed model yielded an obviously better visual prediction map for the entire study area. Overall, the proposed model is beneficial for multimodal feature learning and complex landscape applications.

Skeleton-based deep pose feature learning for action quality assessment on figure skating videos

Most of the existing Action Quality Assessment (AQA) methods for scoring sports videos have deeply researched how to evaluate the single action or several sequential-defined actions that performed in short-term sport videos, such as diving, vault, etc. They attempted to extract features directly from RGB videos through 3D ConvNets, which makes the features mixed with ambiguous scene information. To investigate the effectiveness of deep pose feature learning on automatically evaluating the complicated activities in long-duration sports videos, such as figure skating and artistic gymnastic, we propose a skeleton-based deep pose feature learning method to address this problem. For pose feature extraction, a spatial–temporal pose extraction module (STPE) is built to capture the subtle changes of human body movements and obtain the detail representations for skeletal data in space and time dimensions. For temporal information representation, an inter-action temporal relation extraction module (ATRE) is implemented by recurrent neural network to model the dynamic temporal structure of skeletal subsequences. We evaluate the proposed method on figure skating activity of MIT-skate and FIS-V datasets. The experimental results show that the proposed method is more effective than RGB video-based deep feature learning methods, including SENet and C3D. Significant performance progress has been achieved for the Spearman Rank Correlation (SRC) on MIT-Skate dataset. On FIS-V dataset, for the Total Element Score (TES) and the Program Component Score (PCS), better SRC and MSE have been achieved between the predicted scores against the judge’s ones when compared with SENet and C3D feature methods.

SDNMF: Semisupervised discriminative nonnegative matrix factorization for feature learning

As one of the most effective feature learning methods, Nonnegative Matrix Factorization (NMF) has been widely used in many scientific fields, such as computer vision, data mining, and bioinformatics. However, NMF is an unsupervised method that cannot fully utilize the label information of data. Thus, its performance is limited in some recognition and classification problems. To remedy this shortcoming, this paper proposes a Semisupervised Discriminative NMF (SDNMF) method. First, we design a Soft-Labeled NMF (SLNMF) model by introducing a soft-label matrix-based regression term into the original NMF, so that the relationship between the soft-label matrix and low-dimensional features can be constructed to improve the discriminative ability of low-dimensional features. Second, to effectively estimate the soft-label matrix, a Label Propagation (LP) model is adopted to fully explore the spatial distribution relationship between the labeled and unlabeled samples. Third, an Adaptive Graph Learning (AGL) model is proposed to exploit the geometric relationship of samples well, which could enhance the performance of LP. Finally, the above three models (i.e., SLNMF, LP, and AGL) are integrated into a unified framework for effective feature learning, which can not only effectively explore the structural relationship matrix between data, but also predict the labels for unknown samples. Moreover, an iterative optimization algorithm is presented to solve our objective function. The convergence and computational complexity analysis of the proposed SDNMF method are also provided. Extensive experiments are conducted on several standard data sets. Compared with related methods, the experimental results verify that the proposed SDNMF method achieves better performance.

Genetic Programming for High-Level Feature Learning in Crop Classification

Information on crop spatial distribution is essential for agricultural monitoring and food security. Classification with remote-sensing time series images is an effective way to obtain crop distribution maps across time and space. Optimal features are the precondition for crop classification and are critical to the accuracy of crop maps. Although several approaches are available for extracting spectral, temporal, and phenological features for crop identification, these methods depend heavily on domain knowledge and human experiences, adding uncertainty to the final crop classification. This study proposed a novel Genetic Programming (GP) approach to learning high-level features from time series images for crop classification to address this issue. We developed a new representation of GP to extend the GP tree’s width and depth to dynamically generate either fixed or flexible informative features without requiring domain knowledge. This new GP approach was wrapped with four classifiers, i.e., K-Nearest Neighbor (KNN), Decision Tree (DT), Naive Bayes (NB), and Support Vector Machine (SVM), and was then used for crop classification based on MODIS time series data in Heilongjiang Province, China. The performance of the GP features was compared with the traditional features of vegetation indices (VIs) and the advanced feature learning method Multilayer Perceptron (MLP) to show GP effectiveness. The experiments indicated that high-level features learned by GP improved the classification accuracies, and the accuracies were higher than those using VIs and MLP. GP was more robust and stable for diverse classifiers, different feature numbers, and various training sample sets compared with classification using VI features and the classifier MLP. The proposed GP approach automatically selects valuable features from the original data and uses them to construct high-level features simultaneously. The learned features are explainable, unlike those of a black-box deep learning model. This study demonstrated the outstanding performance of GP for feature learning in crop classification. GP has the potential of becoming a mainstream method to solve complex remote sensing tasks, such as feature transfer learning, image classification, and change detection.

COVID-19 Infection Segmentation and Severity Assessment Using a Self-Supervised Learning Approach.

Background: Automated segmentation of COVID-19 infection lesions and the assessment of the severity of the infections are critical in COVID-19 diagnosis and treatment. Based on a large amount of annotated data, deep learning approaches have been widely used in COVID-19 medical image analysis. However, the number of medical image samples is generally huge, and it is challenging to obtain enough annotated medical images for training a deep CNN model. Methods: To address these challenges, we propose a novel self-supervised deep learning method for automated segmentation of COVID-19 infection lesions and assessing the severity of infection, which can reduce the dependence on the annotation of the training samples. In the proposed method, first, many unlabeled data are used to pre-train an encoder-decoder model to learn rotation-dependent and rotation-invariant features. Then, a small amount of labeled data is used to fine-tune the pre-trained encoder-decoder for COVID-19 severity classification and lesion segmentation. Results: The proposed methods were tested on two public COVID-19 CT datasets and one self-built dataset. Accuracy, precision, recall, and F1-score were used to measure classification performance and Dice coefficient was used to measure segmentation performance. For COVID-19 severity classification, the proposed method outperformed other unsupervised feature learning methods by about 7.16% in accuracy. For segmentation, when the amount of labeled data was 100%, the Dice value of the proposed method was 5.58% higher than that of U-Net.; in 70% of the cases, our method was 8.02% higher than U-Net; in 30% of the cases, our method was 11.88% higher than U-Net; and in 10% of the cases, our method was 16.88% higher than U-Net. Conclusions: The proposed method provides better classification and segmentation performance under limited labeled data than other methods.

Discriminative feature learning using a multiscale convolutional capsule network from attitude data for fault diagnosis of industrial robots

Effective fault diagnosis is important to ensure the reliability, safety, and efficiency of industrial robots. This article proposes a simple yet effective data acquisition strategy based on transmission mechanism analysis, using only one attitude sensor mounted on an end effector or an output component to monitor the attitude of all transmission components. Unlike widely used vibration-monitoring signals, attitude signals can provide fault features reflecting spatial relationships. Using one attitude sensor facilitates the data collection, but weakens fault features and introduces strong background noise in attitude signals. To learn discriminative features from the attitude data collected by the attitude sensor, a multiscale convolutional capsule network (MCCN) is proposed. In MCCN, integrating low-level and high-level features in a convolutional neural network (CNN) as multiscale features is conductive to noise reduction and robust feature extraction, and a capsule network (CapsNet) is used to recognize the spatial relationships in attitude data. The extracted multiscale features in CNN and the spatial-relational features in CapsNet are fused for effective fault diagnosis of industrial robots. The performance of MCCN is evaluated by attaching a softmax-based classifier and integrating it into different transfer learning frameworks to diagnose faults in industrial robots under single and variable working conditions, respectively. Fault diagnosis experiments were conducted on a 6-axis series industrial robot and a parallel robot-driven 3D printer. The superiority of the proposed MCCN was demonstrated by comparing its performance with the other feature learning methods.

Cross-Attribute adaptation networks: Distilling transferable features from multiple sampling-frequency source domains for fault diagnosis of wind turbine gearboxes

Vibration signals of wind turbine gearboxes are often collected under various sampling frequencies. However, most traditional domain adaptation methods, which are applied to improve fault diagnosing accuracy with limited or unlabeled datasets, only consider a single source domain with the same sampling frequency. In this paper, a novel domain-invariant feature learning method, i.e., cross-attribute adaptation networks (CAAN), is developed, in which each source domain with a particular sampling frequency has individual feature learning and adaptation modules. Specifically, a multi-branch framework with the attention mechanism is developed to learn and weigh the characteristics of multiple sampling-frequency data separately. Then, the discrepancies between multiple classifiers and the domains are minimized to build a precise decision boundary. Extensive experimental analysis on real wind turbine gearboxes datasets is performed to demonstrate the effectiveness and advantage of the proposed CAAN.

A feature learning method for rotating machinery fault diagnosis via mixed pooling deep belief network and wavelet transform

Deep learning has extensive application in fault diagnosis regarding the health monitoring of machinery core components. Although deep networks have better nonlinear representation ability, they inevitably introduce a large number of parameters, resulting in slow model training, as well as poor generalization. Additionally, the compound fault tends to be neglected in conventional fault diagnosis. For this purpose, a bearing fault diagnosis method based on mixed pooling deep belief network (MP-DBN) is suggested. Firstly, the Morlet wavelet is adopted for obtaining the corresponding time–frequency representation to enhance analysis efficiency. Then, a new MP-DBN model with a mixed pooling layer and restricted Boltzmann machines (RBM) is constructed to reduce the effects of overfitting and the parameters scale. Finally, the MP-DBN-based method is employed for the bearings in the laboratory and actual working circumstances to verify its performance. The results show that MP-DBN is a powerful compound fault diagnosis technique for rolling bearings with superior feature extraction ability and diagnosis efficiency.

XGBoost-Based Feature Learning Method for Mining COVID-19 Novel Diagnostic Markers.

In December 2019, an outbreak of novel coronavirus pneumonia spread over Wuhan, Hubei Province, China, which then developed into a significant global health public event, giving rise to substantial economic losses. We downloaded throat swab expression profiling data of COVID-19 positive and negative patients from the Gene Expression Omnibus (GEO) database to mine novel diagnostic biomarkers. XGBoost was used to construct the model and select feature genes. Subsequently, we constructed COVID-19 classifiers such as MARS, KNN, SVM, MIL, and RF using machine learning methods. We selected the KNN classifier with the optimal MCC value from these classifiers using the IFS method to identify 24 feature genes. Finally, we used principal component analysis to classify the samples and found that the 24 feature genes could effectively be used to classify COVID-19-positive and negative patients. Additionally, we analyzed the possible biological functions and signaling pathways in which the 24 feature genes were involved by GO and KEGG enrichment analyses. The results demonstrated that these feature genes were primarily enriched in biological functions such as viral transcription and viral gene expression and pathways such as Coronavirus disease-COVID-19. In summary, the 24 feature genes we identified were highly effective in classifying COVID-19 positive and negative patients, which could serve as novel markers for COVID-19.

Mechanical compound fault diagnosis via suppressing intra-class dispersions: A deep progressive shrinkage perspective

Compound faults and their involved single faults often have severe overlap in traditional feature spaces, and the strong background noise unavoidably exacerbates the degree of overlap. Aiming at the problem, this article constructs a multi-level discriminative feature learning method, namely deep progressive shrinkage learning, to progressively suppress intra-class dispersions using a few feature-level shrinkage modules and a decision-level shrinkage module for separating compound faults from single faults. First, soft thresholding is embedded as a key part of feature-level shrinkage modules to gradually eliminate noise-related information in the multi-layer feature learning process, in which thresholds are adaptively set using attention mechanism. Second, in the decision-level shrinkage module, high penalties are imposed on the samples that are far from their class centers. Finally, the efficacy of the method in compound fault diagnosis along with single faults has been verified through a variety of experiments.