In practical scenarios, the working conditions of bearings change with the variation of work tasks, making it extremely challenging to collect a large number of fault samples for each working condition. This difficulty hinders the existing fault diagnosis methods from achieving satisfactory diagnostic results. To address this issue, this article proposes a novel long short-term memory (LSTM)-based node-gated graph neural network (LNGGN) for cross-condition few-shot bearing fault diagnosis. LNGGN is primarily composed of four modules: the graph initialization module, the LSTM-based node-gated update module (LNGUM), the mean-based node attention module (MNAM), and the edge update module (EUM). The graph initialization module is responsible for converting raw data into graph data. LNGUM is proposed to implement gated updates for node features, and it effectively retains the essential information of nodes, alleviating the problem that node features become indistinguishable with the increasing number of updates. Moreover, LNGUM facilitates adaptive filtering of aggregated information, thereby reducing the impact of potential data noise on nodes. MNAM is designed to emphasize the more expressive feature dimensions within nodes for improved similarity computation. Within the EUM, two metric networks are employed to calculate the similarity and dissimilarity between nodes, which are then used to update the edge features. Experimental evaluations are conducted on two bearing datasets to validate the effectiveness of LNGGN. The results demonstrate that LNGGN outperforms baseline methods significantly in cross-condition few-shot bearing fault diagnosis tasks.

To address the problem of traditional diagnosis methods being ineffective due to the feature distribution shift of rolling bearings under cross-conditions, a new method called the dual domain adversarial network (DDAN) has been proposed in this study. The DDAN is integrated with a multi-channel parallel feature extractor, which can mine domain-invariant features, extracting as many useful features as possible from both the frequency domain and the time-frequency domain perspectives. The L _1,2 -norm based Wasserstein Discrepancy (L _1,2 WD) is then introduced as the domain difference value to improve the stability and computation speed of the diagnosis model. After that, a dual domain adversarial paradigm is constructed to jump out of local superiority and improve the generalization of the model by correcting the overconfidence of the model and expanding the confidence interval, respectively. Finally, the verification is performed on two bearing datasets in comparison with several unsupervised domain adaptation (UDA) methods, and the outcomes demonstrate the excellence of DDAN in resolving cross-conditions rolling bearing fault diagnosis issues.

To address the poor performance of traditional rolling bearing fault diagnosis models in cross-condition tasks due to significant feature differences, a Transitional Domain Adversarial Network (TDAN) proposed in this paper. This model initially builds a multi-channel, multi-feature extractor to obtain the frequency domain phase spectrum of vibration signals. It then integrates this data with spectral and time-domain features to extract deep, domain-invariant characteristics from various perspectives. Transition units are also designed to derive both domain and class transitional zones. The domain transitional zone aims to mitigate the loss of certain features caused by forced alignment between source and target domains. Meanwhile, the class transitional zone enhances feature granularity from the perspective of inter-class variation, thereby improving class-specific representation, smoothing the adversarial process, and boosting model generalization. Additionally, to address the target-oriented adversarial loss function, a re-adversarial module is introduced. This process equips the model with the capability to escape local optima and optimize parameters adaptively during training, resulting in stronger robustness and adaptability. Comparative experiments with other unsupervised domain adaptation (UDA) methods on two bearing datasets demonstrate TDAN’s effectiveness and superiority in rolling bearing cross-condition fault diagnosis. It also demonstrates the model’s potential for application in real industrial scenarios where varying operating conditions lead to differences in vibration signals.

Recently, bearing fault diagnosis methods based on deep learning have achieved significant success. However, in practical engineering applications, the limited labeled data and various working conditions severely constrain the widespread application of most deep-learning-based fault diagnosis methods. Additionally, many methods focus solely on the amplitude information of samples, neglecting the rich relational information between samples. To address these issues, this paper proposes a novel cross-condition few-shot fault diagnosis method based on an adaptive dynamic threshold graph neural network (ADTGNN). The aim of the proposed method is to rapidly identify fault types after they occur only a few times or even once. The adaptive threshold computation module (ATCM) in ADTGNN dynamically assigns thresholds to each edge based on edge confidence, optimizing the graph structure and effectively alleviating the over-smoothing issue. Furthermore, a dynamic threshold adjustment strategy (DTAS) is introduced to gradually increase the threshold with the training iterations, preventing the model from prematurely discarding crucial edges due to insufficient performance. The proposed model’s effectiveness is demonstrated using three bearing datasets. The experimental results indicate that the proposed approach significantly outperforms other comparison methods in cross-condition bearing fault diagnosis.

Rolling bearings are critical in maintaining smooth operation of rotating machinery and considerably influence its reliability. The signals collected from rolling bearings in field conditions are often subjected to noise, creating a challenge to extract weaker fault features. This paper proposes a rolling bearing fault diagnosis method that addresses the above-mentioned problem through the moth-flame optimization algorithm optimized variational mode decomposition (MFO-optimized VMD) and an ensemble differential evolution online sequential extreme learning machine (DE-OSELM). By using the dynamic adaptive weight factor and genetic algorithm cross operator, the optimization accuracy and global optimization ability of the moth-flame optimization (MFO) are improved, and the two basic parameters of VMD decomposition level and quadratic penalty factor are adaptive selected. Since the vibration characteristics of the signal cannot be fully interpreted by a single index, The effective weighted correlation sparsity index (EWCS) is utilized to extract the relevant intrinsic mode functions (IMF) of VMD decomposition and extract their energies as features. In order to improve the classification accuracy, The energy feature set is subsequently inputted into DE-OSELM for training and classification purposes, and the proposed method is assessed via a sample set with four different health states of actual rolling bearings. Our proposed method results are compared with other diagnosis methods, proving its feasibility to diagnose rolling bearing faults with higher classification accuracy.

In the Industrial Internet of Things (IIoT), detecting and identifying faults in machinery is crucial as it helps to eliminate potential failures and reduce losses caused by these faults. Deep learning models employing domain adaptation methods have proven highly effective for diagnosing faults in bearings. These models, however, require the incorporation of examples from the target domain during training, resulting in an excessive dependence on information from the target domain. In real-world applications, target domain data sources are often discrete, diverse, and inaccessible, requiring the model to be re-trained for each diagnosis and thus hindering real-time diagnostics. To tackle this problem, this research introduces adversarial generalized models aimed at real-time failure detection of rolling elements without requiring samples from the target domain. First, this model enhances domain-invariant feature extraction through multi-scale joint weighting augmentation. Additionally, it utilizes random norm mutual coupling to determine the Wasserstein distance for assessing domain variation, thereby improving convergence and operational efficiency. To further enhance domain confusion ability and robustness, the Label Smoothing Regularization approach is introduced for calculating model class classification loss. Finally, tests were performed using three publicly available datasets and one private dataset, achieving optimal accuracies of 98.08%, 85.40%, and 91.90% for the public datasets, along with a 97.69% accuracy for the private dataset. These results confirm the robustness and effectiveness of the method introduced in this study.