The operational precision, stability, and lifespan of servo motors are directly influenced by the healthy operation of rolling bearings. Early diagnosis of bearing faults holds significant importance in enhancing the reliability of servo motors and preventing substantial economic losses and personal injuries. However, strong noise interference in industrial environments poses challenges to bearing fault diagnosis. Additionally, the difficulty in detecting single faults in bearings often leads to the occurrence of compound faults. To achieve compound fault diagnosis under intense noise interference in industrial environments, this paper proposes the nonlinear convolutional sparse filtering (NCSF). First, generalized sigmoid and Z-score normalization are utilized to amplify fault features and reduce the influence of irrelevant interference. Subsequently, multidimensional filters are learned based on a nonlinear objective function to achieve the separation of compound faults. Then, the Gaussian function is utilized to fit the trained filters, reducing the impact of low-frequency noise. Finally, the filtered signals are envelope demodulated to analyze fault types. The capability of NCSF to distinguish compound faults amid noise interference is confirmed by both experimental and simulation results. The proposed NCSF is an effective method to realize compound fault diagnosis under strong noise interference.

Abstract Early fault signatures of rolling bearings are typically very weak and are often contaminated by strong background noise and random impulsive interference. Traditional kurtosis-based feature extraction methods are highly sensitive to non-fault-related large random impacts, which may result in incorrect frequency-band selection. To address this issue, an improved sparse filtering (SF) method is proposed in this paper. First, to overcome the unstable convergence caused by random initialization in conventional SF, a linear uniform initialization strategy is introduced, in which a uniformly distributed band-pass filter bank is constructed as the initial weight matrix to ensure full-band feature extraction. Second, the envelope spectrum harmonic-to-noise ratio (ESHNR) is employed as an evaluation criterion to adaptively select the optimal filter containing periodic fault characteristics from multiple filters generated during iteration, thereby effectively suppressing random interference. The key novelty of this method is that the fault-characteristic frequency and its harmonic structure are incorporated into the sparse-filtering-based component selection process, enabling the selected component to reflect fault-related periodicity rather than merely high impulsiveness. The effectiveness of the proposed method is verified using simulation signals and experimental inner- and outer-race bearing fault signals under strong random impulsive interference.

Fault diagnosis in mechanical systems is limited to single-modality measurements, which may lead to incomplete feature representation. Vibration and acoustic signals contain substantial redundant and complementary information, yet the effective extraction and utilization of these multi-dimensional features remain challenging. To address these limitations and obtain a more complete and representative feature set, a Multichannel Adaption guided Vision Transformer (MAViT) with acoustic-vibration signal feature fusion is proposed, which employs acoustic-vibration signal feature fusion technology to achieve more comprehensive and robust feature extraction. Meanwhile, the MAViT incorporates the FReLU activation function, which considers the contextual information of activation features, thereby enhancing the description of vibro-acoustic signals. Additionally, the fusion mechanism integrates multi-channel features, enabling the model to capture multi-dimensional attributes without losing important information, thus improving the accuracy and stability of diagnosis. A series of 12 different cross-condition transfer diagnosis tasks were designed to validate the effectiveness of the MAViT. Experimental results demonstrate that the MAViT can not only effectively identify various types of fault samples but also exhibit the highest stability in multiple tests compared with other methods. The diagnostic accuracy of this method can reach over 98 %.

In recent years, unsupervised domain adaptation has attracted increasing attention for learning transferable representations from related but distribution-shifted datasets. In bearing fault diagnosis, conventional deep models often exhibit degraded performance under varying operating conditions due to insufficient cross-domain feature alignment. To address this issue, we propose a multiwavelet kernel network with distance fusion domain adaptation (MWKN-DFDA). The model employs a multiwavelet kernel convolution strategy to extract rich temporal and time-frequency-sensitive features from vibration signals, followed by an SE (Squeeze-and-Excitation)-style channel attention module that adaptively recalibrates multiwavelet feature channels and emphasizes fault-relevant responses. On this basis, we design a hierarchical distribution alignment strategy with a three-stage collaborative optimization scheme, integrating sliced Wasserstein distance (SWD) for marginal alignment, conditional SWD for class-conditional alignment with confidence-filtered pseudo-labels, and random Fourier feature-maximum mean discrepancy for nonlinear global alignment in an approximated kernel space. Only source-domain labels are used for training, while target-domain labels are used solely for evaluation. The learned representations are then fed into a classifier and a domain discriminator for joint optimization. Experiments on bearing datasets with varying rotational speeds and noise conditions demonstrate that MWKN-DFDA consistently outperforms competing methods, achieving up to 99.01% accuracy and improved robustness across multiple transfer scenarios.

Despite remarkable advancements in few-shot transferable fault diagnosis, most studies remain restricted to homogeneous signals; meanwhile, fault diagnosis methods have grown increasingly complex to ensure robust transfer performance, imposing higher computational demands. To address these issues, this paper proposes the light cross domain model-agnostic meta-learning for bearing few-shot transferable fault diagnosis. The method constructs a hierarchical interactive feature encoder based on cross-layer channel attention, which breaks single-layer perspective limitations, extracts complementary channel features, and enhances generalization-meeting heterogeneous signal diagnosis needs in few-shot transfer scenarios. Additionally, replacing fully connected layers with GAP modules reduces model size and improves computational efficiency. Validation using bearing vibration and acoustic signals across two datasets confirms the method's effectiveness.

Object detection in unmanned aerial vehicle (UAV) imagery has significant application value in various fields, including traffic monitoring, disaster rescue, agricultural surveys, and military reconnaissance. However, the variable flight altitudes and perspectives of UAV platforms often results in targets occupying an extremely small pixel proportion, exhibiting limited appearance characteristics, and being susceptible to complex background interference. These factors pose considerable challenges for small object detection. To address these challenges, this paper proposes a novel small object detection framework called CGD-YOLO, based on YOLOv11n. The framework introduces a C3k2-ConvFormer module into the backbone network, which uses depthwise SepConv for efficient spatial feature interaction and a dual amplitude modulation mechanism to alleviate gradient vanishing. Furthermore, a global edge information enhancement module is designed by embedding an edge extractor in shallow layers and enabling cross-layer propagation of edge features, improving the capture of fine contours of small objects. In addition, a multi-dimensional attention mechanism integrating scale, spatial, and task awareness is incorporated into the detection head, substantially enhancing its adaptability and discriminative performance for objects of various sizes and spatial distributions. Moreover, an improved Inner-PIoUv2 loss function is proposed, which integrates a target-size penalty term with a scale factor adjustment strategy to guide anchor box regression more accurately and optimize detection performance across different object scales. Extensive experiments on the VisDrone2019 dataset demonstrate that the proposed method outperforms existing state-of-the-art approaches. It achieves a mean average precision of 31.5%, which is 4.2% higher than the baseline, with notable gains in small object detection, while maintaining real-time inference speed. This study not only provides an efficient and robust solution for small object detection in UAV imagery but also offers valuable insights for enhancing tiny object detection in other visual domains through its global edge feature enhancement strategy.

Due to its broad receptive field and ability to eliminate future information leakage, the temporal convolutional network (TCN) has been widely used in the remaining useful life (RUL) estimation of bearings. However, the predictive capability of TCN is constrained by the distortion during degraded feature acquisition and the loss during feature extraction. To address these issues, a hybrid TCN with feature enhancement and adaptive threshold (HTCN-FA) is proposed. Firstly, a plug-and-play resilient nonlinear blind deconvolution (RNBD)-net subnetwork without reliance on prior knowledge is designed to enhance the representation of degraded features. Secondly, a theoretically rigorous and interpretable metric, harmonic distribution clarity, is introduced to monitor the operational condition of bearings. Thirdly, TCN blocks with adaptive thresholds and hybrid branches are employed to avoid the loss of degraded characteristics. Finally, datasets from twelve bearings under three working conditions are employed to validate the prediction and generalization performance of the proposed HTCN-FA. The results demonstrate that the proposed HTCN-FA outperforms existing methods in fault monitoring and RUL prediction while exhibiting superior noise adaptability.

The Cognitive Industrial Twin (CIT) has emerged as an advanced evolution of the Digital Twin (DT) paradigm for Industry 4.0 environments. By embedding learning, reasoning, and decision-making capabilities, CITs extend conventional DTs from passive monitoring toward adaptive and intelligent representations of industrial systems. This survey systematically examines the current state of CIT research, clarifies its definition, and distinguishes it from traditional DTs and other intelligent twin frameworks. A generic five-layer reference architecture is presented, encompassing data sensing, information fusion, knowledge cognition, autonomous decision-making, and feedback optimization. The paper further reviews key enabling technologies—including multi-modal data fusion, reinforcement learning, knowledge graphs, causal inference, and edge intelligence—and discusses their roles in supporting cognitive and autonomous twin functionalities. In addition, industrial applications across smart manufacturing, prescriptive maintenance, autonomous logistics, and sustainable production are analyzed, revealing a paradigm shift from open-loop monitoring to closed-loop, cognition-driven autonomy. Finally, emerging research directions—such as brain-inspired computing, large language model (LLM) integration, and hybrid physical–cognitive modeling—are outlined, along with the key challenges that must be addressed to enable the broader industrial deployment of CITs.

Fault diagnosis of rotating machinery often suffers from severe class imbalance because fault samples are scarce under practical operating conditions. Existing data augmentation methods still face challenges in simultaneously preserving sample quality, diversity, and structural consistency. This study aims to develop an effective diffusion-based generation framework for imbalanced fault diagnosis. A local–global residual enhanced diffusion model (LG-ResDiff) is proposed. The denoising network is improved by introducing a Local Information Perception Feed-Forward Network to strengthen local feature extraction and global dependency modeling. A residual-enhanced bottleneck structure is further designed to stabilize feature propagation during reverse diffusion. Conditional time–label embeddings are incorporated to guide class-aware generation of minority-class samples. The generated samples are then combined with original training data to construct balanced datasets for fault classification. Experiments conducted on bearing and gear datasets demonstrate that the proposed method can generate high-quality time-frequency representations with strong structural fidelity and distribution consistency. Compared to various generative methods, LG-ResDiff achieves superior generation quality and consistently improves fault diagnosis accuracy under various imbalance ratios. Under fully balanced conditions, the proposed method achieves a diagnostic accuracy of 99.5

Abstract Digital twin (DT) technology, increasingly applied in bearing fault diagnosis, effectively overcomes the key limitations of traditional bearing fault diagnosis methods, such as over-reliance on samples, poor adaptability to complex conditions, and inadequate lifecycle management, thereby enabling high-precision and intelligent fault diagnosis for bearings. This paper provides a systematic review of the latest developments in DT-assisted bearing fault diagnosis. It begins by reviewing the key technologies in the application of DT for bearing fault diagnosis, specifically examining the methods for constructing bearing DT (BDT) models and the post-processing of these models. Regarding the modeling methods of BDT, this paper offers a comprehensive review and comparative analysis, classifying them into physics-driven models, phenomenological models, and data-driven models. Besides, the post-processing of BDT models is summarized in two key aspects: strategies for model correction and updates, and model verification. Subsequently, this paper discusses the core applications of DT technology in bearing fault diagnosis, categorizing them into four key scenarios: small-sample or imbalanced-sample diagnosis, cross-domain or variable-operating-condition diagnosis, fault evolution and remaining useful life prediction, and real-time condition monitoring. Furthermore, the key challenges of DT technology in bearing fault diagnosis are explored, including model accuracy, real-time data integration, and system complexity. In response to these challenges, the paper offers targeted recommendations for the future development of DT applications. Overall, this study clarifies the technological framework and development path, aiming to provide comprehensive theoretical references and practical guidance for future academic research and engineering applications of DT technology in the field of bearing fault diagnosis.

Multi-target domain adaptive (MTDA) methods, which take into account the variability of working condition and identify faults in different target domains more accurately, are becoming a growing focus of fault diagnosis. However, the neglect of structural information within the key data, the robustness impact of the disturbing information strongly correlated with operating conditions, and the inconsistencies in matching the source and multi-target domains (SMTD) are the limitations of MTDA. Therefore, a transfer graph feature alignment multi-target domains adaptive network (TGAMDN) is proposed to reduce the disturbing information and extract structural information. Firstly, the data structure between the SMTD is modeled through the construction of a new transfer graph sample generation module (GSG), with the data structure information being learned by a shared graph neural network. Secondly, a novel weighting mechanism and the corresponding training framework are constructed to effectively reduce the impact of interfering information and the boundary difference between different data classes is enhanced using the fitted circle method. Finally, the weighted hybrid alignment strategies are used to minimize differences between domains and resolve inconsistent matches. The performance of the TGAMDN is validated using a rotating machinery dataset across various transfer tasks under different rotational speed and load conditions.

The fault diagnosis of rotating components such as bearings and gearboxes is crucial for ensuring the safe operation of machinery. However, traditional deep learning diagnostic models often exhibit poor performance when facing changes in operating conditions. To address this issue, this paper proposes an adaptive multi-scale attention adversarial network (AMSAAN). First, a combination of one-dimensional (1-D) wide convolution and two-dimensional (2-D) multi-scale convolution is employed to extract long-term temporal features and short-term multi-scale features from the samples. Second, to reduce the impact of noise, a multi-layer attention feature refinement mechanism is utilized to progressively refine the extracted features layer by layer. Cross-domain adaptation is carried out in two parts: domain shift reduction and feature alignment. The first part uses a dual-domain distance approach combining correlation alignment (CORAL) distance and Wasserstein distance to reduce domain distribution differences. The second part integrates contrastive loss and domain adversarial mechanisms to achieve domain feature alignment. This method comprehensively considers both supervised domain adaptation (SDA) and unsupervised domain adaptation (UDA). The proposed AMSAAN effectively extracts domain-invariant features, reduces noise impact, and achieves domain adaptation. The effectiveness of AMSAAN is demonstrated through experimental results based on two datasets of bearings and planetary gearboxes. AMSAAN achieves a diagnostic accuracy of over 97 % across multiple transfer tasks under varying operating conditions, and the noise resistance of the model has been demonstrated to be superior to that of other methods.

Existing intelligent fault diagnosis methods commonly rely on extensive and balanced dataset. However, the collection of fault samples significantly lags behind that of normal samples in practical industrial scenarios, resulting in data imbalance problem that severely impacts diagnostic accuracy. To address this challenge, this paper presents a novel approach, termed the Hybrid Distance Generative Adversarial Network with gradient penalty (HDGAN-GP). Initially, a stacked autoencoders (SAE) is incorporated into the original generator to form an auxiliary generator, facilitating the production of high-quality samples. Subsequently, a loss function is devised based on a hybrid distance metric comprising cosine similarity and maximum mean difference, supplemented by a gradient penalty term to ensure stable model training. Finally, experimental validation is conducted using gear dataset. Comparative analysis with existing generative adversarial network models demonstrates that the proposed method generates superior quality fault samples, effectively addressing the challenge posed by data imbalance in fault diagnosis.

During industrial processes, strong noise often hinders the reliable extraction of features from mechanical equipment, which is crucial for effective fault detection. convolutional neural networks (CNNs) are widely employed in mechanical fault diagnosis due to their powerful capability for autonomous feature learning. However, CNNs suffer from limitations in interpretability and robustness to noise. To address these issues, this paper proposes a wavelet attention and time attention-guided stochastic resonance network (WATA-SRN), which integrates traditional signal processing techniques with CNNs to incorporate theoretical foundations and physical interpretability. By combining attention mechanisms in both the wavelet and time domains, the proposed network fully exploits time-frequency information, thereby significantly enhancing its ability to recognize complex signal patterns. Furthermore, the incorporation of the classical bistable stochastic resonance mechanism strengthens the model's feature extraction capability and improves its resilience to noise, ultimately boosting diagnostic accuracy and generalization performance. The integration of discrete wavelet transform (DWT) and inverse DWT into the CNN architecture enables multi-scale feature extraction and enhances model interpretability. In addition, adaptive noise injection and a frequency-domain data augmentation strategy based on the wavelet domain further improve the model's robustness and generalization. Experimental results on bearing and gear fault datasets demonstrate that WATA-SRN outperforms traditional CNNs in terms of noise robustness and feature extraction capability, especially under high-noise conditions. This advancement enhances the reliability of fault detection in noisy industrial environments, contributing to improved maintenance efficiency and operational safety.

Many of the current fault diagnosis methods rely on time-domain signals. While the richest information are contained in these signals, their complexity poses challenges to network learning and limits the ability to fully characterize them. To address these issues, a novel multi-channel fused vision transformer network (MFVTN) is proposed in this paper. Firstly, the overlapping patch embedding module is introduced to overlap the time-domain map with edge information, preserving the global continuous features of the time-domain map and adding positional encoding for sorting. This integration helps the vision transformer merge detailed features and construct the global mapping. Secondly, multiple dimensional time domain signal features are extracted and fused in parallel, enabling multi-domain fault diagnosis of bearings. In order to enhance the network ability to extract domain-invariant features, an adversarial training strategy combined with Wasserstein distance is utilized. The results demonstrate that the diagnostic accuracy of the proposed MFVTN can reach 98.2%.

In the aerospace and high-speed rail industries, carbon fiber reinforced polymer (CFRP) has seen widespread application. CFRP plates and connectors in operation are often subjected to impacts that can cause damage. The unpredictable nature of the impacts introduces uncertainties in both the location and extent of the damage, posing significant challenges to traditional supervised learning models, which often struggle with missed detections or misclassifications when identifying unknown damages. To address the issue, a deep learning model based on temporal convolutional network-gated recurrent unit (TCN-GRU) is proposed. TCN extracts features from the raw time domain signals, and GRU selectively retains the significant features and completes sequence modeling. A center loss function is incorporated into the fully connected layer to improve the effects of intra-class aggregation and inter-class separation. An unknown detection module is introduced to realize the identification and classification of unknown damages based on a predefined threshold. The experimental results indicate that the proposed method can achieve effective unknown damage diagnosis in the open set case. This study provides a feasible solution for open set unknown damage diagnosis in CFRP plates and connectors.

The fast kurtogram is a very useful tool in the field of fault diagnosis, but it also contains two drawbacks. On the one hand, its kurtosis indicator is susceptible to random impulse interference, leading to frequency band mis-selection. On the other hand, the frequency band division method determined by the tree filter bank may cause under or over decomposition problems. Therefore, in this paper, a novel fault diagnosis theory is constructed combining the adaptive recombination empirical wavelet transform (AREWT) with the envelope spectral energy ratio (ESER). Adaptive decomposition of frequency bands is first performed using AREWT. Afterwards, the ESER is proposed as a statistical indicator to choose the optimal demodulation frequency band which overcomes the effects of non-Gaussian noise interference and improves diagnostic accuracy. To further highlight the fault elements of the selected components, an adaptive sparse coding shrinkage algorithm is introduced for sparsely denoising sensitive components. Correspondingly, clear fault feature frequency components can be extracted from the envelope spectrum. Finally, the practicability and superiority of the proposed AREWT-ESER approach are fully validated through numerical simulation signals and case studies.

In the practical application of rotating machinery, the change of working conditions can meet different manufacturing requirements. When fault diagnosis is performed on monitoring data with different working conditions, the change of data distribution will bring interference information which is highly related to working conditions and inconsistent matching problems in the process of multi-target domain transfer. In order to solve these problems, a working condition decoupling adversarial network (WCDAN) is proposed for multi-target domain fault diagnosis. Specifically, the prototype discrepancy alignment module is constructed following a weight-shared wavelet convolution feature extractor to ensure a clear prototype representation boundary. Then, the adaptive domain discriminator weight, along with the acquired multi-domain discrepancy, are utilized to decouple the working conditions. This process filters out interference information that highly associated with the source domain working conditions while preserving the inherent fault characteristics. Furthermore, the strategy of multi-domain hybrid alignment aims to minimize the disparity between different domains and solve the inconsistent matching issue. Based on two gearbox fault datasets under stable and unstable conditions, the comparative experimental results show that the WCDAN can be generalized from a single source domain to multiple target domains at the same time and achieve excellent fault diagnosis performance.

Abstract This paper presents a comprehensive review of recent advancements in bearing health monitoring and remaining useful life (RUL) prediction. It highlights key innovations in anomaly detection, health indicator construction, degradation modeling, and RUL estimation, examining developments across statistical, machine learning, and deep learning approaches while analyzing their strengths, limitations, and application contexts. Special emphasis is placed on the role of deep learning in capturing complex degradation patterns from multi-dimensional time series data and improving predictive accuracy in dynamic industrial settings. Additionally, this review explores multi-source data fusion techniques, which enhance anomaly detection robustness by integrating information from diverse sensor modalities. By identifying critical challenges and suggesting future research directions, this study aims to advance the development of robust and adaptive prediction models for intelligent maintenance in industrial applications.

Rolling bearings are prone to failure due to harsh working environments, which can affect production efficiency. Given the background environmental noise interference, it is difficult to extract the fault characteristics of rolling bearings in the early stages of weak fault signals. Based on adaptive VMD combined with adaptive IMCKD (AVMD-AIMCKD), a rolling bearing fault diagnosis method is proposed in this paper. The vibration signal from the rolling element bearing is processed by adaptive variational modal decomposition (AVMD) to extract and select the time-frequency domain signal characteristics. The adaptive improved maximum correlated kurtosis deconvolution (AIMCKD) algorithm is used for noise reduction optimization to obtain the best extraction results. Because the traditional method requires input parameter values for variable mode decomposition (VMD) and improved maximum correlated kurtosis deconvolution (IMCKD), it is not adaptive. The particle swarm optimization algorithm is utilized in this study to optimize two variables: the penalty factor and decomposition mode number of VMD. The filter length and shift order of the IMCKD are optimized to make it adaptive. The viability of the method is substantiated through simulations, and the efficacy and precision are validated by comparative studies. The experimental results show that the AVMD-AIMCKD method has high accuracy and robustness in rolling bearing fault diagnosis and provides an accurate reference for rolling bearing health monitoring in engineering practice. The AVMD-AIMCKD method effectively extracts the early fault features through adaptive optimization, overcomes the restrictions of traditional methods, and provides a more accurate and reliable tool for bearing fault signal diagnosis.