The prediction of the remaining useful life (RUL) of bearings is of great significance for reducing cost and increasing efficiency of mechanical equipment and ensuring healthy operation. Most of the traditional RUL prediction methods based on deep learning only analyze single sensor data, which makes the prediction accuracy fluctuate greatly and the reliability is low, and the contributions of data collected by different sensors to RUL prediction are inconsistent, which makes the low utilization of data lead to poor prediction results. To solve the above problems, a new RUL prediction framework with convolutional attention mechanism and temporal convolutional network (CAMTCN) is proposed. Firstly, the vibration signals collected by multi-sensor are normalized and processed directly as the input of the network. Secondly, an Efficient Adaptive Shrinkage (EAS) model is designed to adaptively eliminate noise interference and improve the accuracy of RUL prediction. Finally, we use the attention mechanism to design the convolutional attention subnet and combine it with the TCN to obtain the contribution degree of bearing degradation features collected by different sensors. Adaptive weight allocation captures more important degradation features to achieve end-to-end RUL prediction. Experiments were conducted on IEEE PHM 2012 dataset and XJTU-SY dataset respectively, and compared with some advanced in-depth learning methods. The experimental results show that this method can predict RUL more accurately on different bearing data sets, and has good generalization.

Many current fault diagnosis methods tend to ignore the temporal correlation in signals, leading to a loss of critical fault information. Additionally, traditional diagnostic models often face challenges in terms of noise immunity, generalization, and handling non-Euclidean structured data. To address these issues, we propose a novel fault diagnosis approach that combines Graph Neural Networks (GNN) with the Markov Transform Field (MTF). We first use the Markov Transform Field to convert vibration signals into two-dimensional images, preserving temporal correlation and preventing the loss of crucial fault information. Next, we use a Graph Convolutional Neural Network (GCN) to process graph-structured data, capturing global structural information. Finally, we introduce the Graph Attention Network (GAT) to dynamically adjust node weights based on their relative importance, enhancing the overall model performance. In this paper, we introduce a new fault diagnosis model, GCN-GAT, and evaluate it using the CWRU bearing dataset and a custom-built planetary gearbox dataset. The results show that our model maintains high fault detection accuracy even in the presence of significant noise and variable load conditions. This indicates that our approach demonstrates strong robustness and generalization, providing an effective solution for complex fault diagnosis tasks.

In practical engineering scenarios, machines are seldom in a faulty operating state, so it is difficult to get enough available sample data to train the fault diagnosis model, leading to the problem of the small and unbalanced number of rotating machinery fault samples and low fault diagnosis accuracy. To solve this problem, this paper introduces a novel approach to machinery fault diagnosis. This approach involves the integration of a Convolutional Attention Residual Network (CBAM-ResNet) with a Graph Convolutional Neural Network (GCN). Firstly, to comprehensively exploit time-domain information from one-dimensional vibration signals, this study utilize Gram Angular Field (GAF) coding to transform traits of vibration signals into two-dimensional image characteristics. The resultant two-dimensional image is then expanded by applying the Wasserstein Distance Gradient Penalty Generation Adversarial Network (WGAN-GP) to produce a representative sample image. Secondly, the image is input to CBAM-ResNet to perform focused feature extraction and construct the feature matrix. Lastly, the adjacency matrix is derived through Graph Generation Layer (GGL); subsequently, the feature matrix and adjacency matrix are utilized as inputs for the GCN. After deep feature extraction, fault feature classification is executed via Softmax. Performance tests were conducted using the Case Western Reserve University bearing dataset and the planetary gearbox dataset. The method demonstrated remarkable results, achieving an accuracy of over 99% on the unbalanced dataset and surpassing 98% in 0dB noise compared to various other models. This illustrates the effectiveness and feasibility of the proposed method.

In recent years, the Unsupervised Domain Adaptation (UDA) technique has achieved remarkable success in cross-domain fault diagnosis of rotating machinery. In UDA, three pivotal information—namely, class labels, domain labels, and data structures, play a critical role in establishing a connection between labeled samples of the source domain and unlabeled samples of the target domain. Most research methods use only one or two of these types of information, ignoring the importance of data structure. In addition, global domain adaptive techniques are typically used, ignoring the relationships between subdomains. The conventional convolutional neural network exhibits limited capability in extracting essential fault-related information, thereby significantly affecting the accuracy of fault identification. To address this problem, we propose the Split-Attention Mechanism and Graph Convolutional Adversarial Network (SPGCAN) as a novel approach for the intelligent diagnosis of faults in rotating machinery. A classifier and a domain discriminator are used to extract the first two types of information. Using residual networks with multi-channel split attention mechanism, graph convolutional neural networks for the modelling of data structures. We use a combination of Local Maximum Mean Difference (LMMD) and Adversarial Domain Adaptation methods are used to align the subdomain distributions and reduce the distributional differences between the relevant subdomains and the global. Using CWRU bearing dataset and Planetary Gearbox dataset for cross-domain fault diagnosis and comparing with current mainstream UDA methods. Ultimately, SPGCAN demonstrates better fault identification accuracy across 24 cross-domain fault diagnosis tasks on both datasets, thus substantiating the method’s effectiveness and superiority.

The remaining useful life (RUL) prediction is important for improving the safety, supportability, maintainability, and reliability of modern industrial equipment. The traditional data-driven rolling bearing RUL prediction methods require a substantial amount of prior knowledge to extract degraded features. A large number of recurrent neural networks (RNNs) have been applied to RUL, but their shortcomings of long-term dependence and inability to remember long-term historical information can result in low RUL prediction accuracy. To address this limitation, this paper proposes an RUL prediction method based on adaptive shrinkage processing and a temporal convolutional network (TCN). In the proposed method, instead of performing the feature extraction to preprocess the original data, the multi-channel data are directly used as an input of a prediction network. In addition, an adaptive shrinkage processing sub-network is designed to allocate the parameters of the soft-thresholding function adaptively to reduce noise-related information amount while retaining useful features. Therefore, compared with the existing RUL prediction methods, the proposed method can more accurately describe RUL based on the original historical data. Through experiments on a PHM2012 rolling bearing data set, a XJTU-SY data set and comparison with different methods, the predicted mean absolute error (MAE) is reduced by 52% at most, and the root mean square error (RMSE) is reduced by 64% at most. The experimental results show that the proposed adaptive shrinkage processing method, combined with the TCN model, can predict the RUL accurately and has a high application value.