Compound faults are common in modern intelligent manufacturing systems and pose severe risks to critical machinery. Accurate diagnosis is therefore essential but remains hindered by two challenges: the limited interpretability of deep learning models and the neglect of intrinsic dependencies among co-occurring faults in existing multi-label frameworks. To tackle these issues, we propose an interpretable time-frequency network that integrates physics-guided signal processing with deep learning. Specifically, a multi complex wavelet kernel convolution layer generates explainable time-frequency representations using five distinct wavelets, while a kernel weight module dynamically integrates them through adaptive weighting. Subsequently, a hybrid multi-scale convolution further enhances cross scale feature interactions, while a dynamic time-frequency fusion module adaptively balances complementary information. Furthermore, an asymmetric joint loss is introduced to explicitly establish label correlations and capture the joint probability of compound faults. The experimental results of two case studies highlight that the proposed method achieves superior diagnostic performance and significantly outperforms other advanced diagnostic methods.

Cardiovascular diseases (CVDs) are the leading cause of mortality worldwide. While deep learning (DL) has shown potential in automated CVDs diagnosis, domain shifts due to variations in acquisition devices and environments hinder generalization and reliability. This paper proposes an interpretable cross-domain diagnostic framework, named CFTResNet, to mitigate domain shifts and enhance diagnostic interpretability. In contrast to traditional transfer learning methods that typically fine-tune fully connected layers (FC), the proposed CFTResNet uses a strategy called Module Robustness Criticality (MRC) to evaluate which parts of the pre-trained model are weak in robustness and then fine-tunes only those specific weak modules instead of adjusting the entire model, thus enhancing adaptability and interpretability. Additionally, to enhance feature representation, we integrate a Temporal-Channel Fusion Module (TCFM) with the ResNet architecture, which effectively captures characteristic information of different channels from heart sound (HS) signals, enhancing the model's capability to discern subtle pathological patterns in cardiac auscultation. Experiments on two public HS datasets demonstrate that CFTResNet outperforms conventional methods in diagnostic accuracy, interpretability, and cross-domain generalization. Highlighting its potential as a reliable AI-assisted (artificial intelligence assisted) tool for clinical CVDs diagnosis.

Multi-adversarial domain adaptation (MADA) techniques enable the effective transfer of knowledge across multiple domain classifiers and have shown significant potential in transfer learning-based fault diagnosis of rotating machinery. Nevertheless, their performance is highly dependent on the availability of abundant labelled source domain data, and their feature extraction capability degrades severely when there exists a substantial domain discrepancy between source and target distributions. This challenge is particularly acute in the fault diagnosis of closed-loop hydraulic pumps with incipient or hidden faults, where the risk of misclassification is considerably elevated. To overcome these limitations, this paper presents a novel transfer fault diagnosis framework that integrates a dynamic simulation model with an enhanced multi-adversarial domain adaptation network, termed DS-IMADA (Dynamic Simulation-Improved Multi-Adversarial Domain Adaptation). Specifically, a dynamic pressure simulation model is established to simulate representative fault scenarios of hydraulic pumps, providing synthetic labelled data in the source domain under various fault conditions. Subsequently, a deep convolutional cross-branch parallel architecture with embedded self-attention mechanisms is employed to strengthen domain-invariant feature extraction. Additionally, an improved semi-supervised multi-adversarial pre-adaptation strategy is proposed to mitigate the negative transfer effect and enhance domain alignment. Finally, comprehensive transfer diagnosis experiments using simulated signals as source domain data and real measured signals as target domain data are conducted, and the results validate the effectiveness and robustness of the proposed method.

In recent years, there has been a surge focusing on advanced heart sound (HS) signal analysis and automated diagnosis based on deep learning (DL) for cardiovascular diseases (CVDs). However, the untrust of users in decision-making caused by the complex nonlinear transformation within the model and unclear feature extraction mechanism remains a huge challenge. For the diagnosis issue of CVDs involving human life and health, if the reasons why the model obtains the final conclusion cannot be known in advance, taking actions rashly will conceal significant risks. In this paper, an interpretable wavelet convolution transformer, named WCFormer, is proposed for HS signal analysis and automated diagnosis of CVDs. This method aims to enhance the interpretability of the traditional transformer and realize the high-accuracy diagnosis of CVDs by embedding wavelet knowledge information and improving its structure. Specifically, a wavelet convolution kernel is first designed to capture disease-related information with a clear physical meaning. Then, a global–local feature extractor is designed by removing the position encoding of the transformer and combining it with the convolution module. Two case studies involving HS signals are implemented to validate the efficacy of the proposed WCFormer and the results are compared with several widely used approaches, revealing that the WCFormer can achieve more excellent performance than other comparison methods.

Playing a vital role in keeping gearbox working reliably and safely, smart fault diagnosis (FD) technology has attracted much attention in recent years. However, in practical industrial applications, owing to the imbalance of healthy state data and fault state data and various unpredictable compound fault modes, it is still extremely challenging to fulfill the high-accuracy and effective FD of gearbox based on existing intelligent diagnostic models. In this paper, a new method is proposed to tackle these problems by integrating an improved deep residual network (IDRN) and wavelet transform (WT). The proposed method mainly contains two parts: In the first part, the vibration signals are transformed into images by WT. In the second part, the IDRN is applied to realize an accurate FD of gearbox. Compared with the original deep residual network, the main differences of IDRN are as follows: first of all, a new loss function is used to replace the commonly used logistic loss function by adding a class-balanced re-weighting term and combining with multi-label classification. Then, attention modules focusing operations on specific regions and enhancing the features of some regions are combined with residual blocks. Finally, a spatial transformer module is inserted into the original deep residual network to scale images and clip the region of interest. Two trials are conducted to validate the effectiveness of WT-IDRN method. The experimental consequences demonstrate that the WT-IDRN method has more excellent performance than the existing intelligent FD method in accuracy and generalization ability.

Transfer learning of multi-source domain adaptation seems a promising way for fault diagnosis of roller element bearings under variable working conditions. Data imbalance affects the performance of multi-source domain adaptation greatly and is expected to be solved by GAN. However, GAN-based transfer learning diagnosis models suffer pattern collapse and training instability, leading to unsatisfying diagnosis results in practical engineering. This paper proposes a denoising diffusion multi-source domain adaptation model (DDMDA). The proposed model uses diffusion denoising, which has better performance and is simpler to train than GAN, to generate shifted source domains for solving the data imbalance problem. A new noise prediction structure in diffusion denoising named Utrans-net, is constructed to restore the data distribution in the shifted source domain. Also, a multiple-domain discriminator structure is designed to extract features from multiple source domains to solve the issue of variable working conditions. Advanced models are used in this paper to compare with the proposed model for validation. Experimental demonstrations show that the proposed model is superior to the comparison models with satisfying performance.