Ridong Zhang

This article proposes a multiscale convolutional neural network (MSCNN)-based fault diagnosis method (AM-MAST) using a self-attention mechanism and soft thresholding. It aims to improve the accuracy of the industrial fault diagnosis model as well as the model’s immunity to interference. Unlike most current single-model fault diagnosis methods, this article employs an MSCNN and a multihead attention mechanism to extract features simultaneously. It also introduces an adaptive thresholding module at the end of the multiscale neural network to deal with interference noise. In addition, the features of the two networks are fused by a fusion module based on the attention mechanism; finally, the faults are accurately categorized by a classification module. This methodology effectively integrates both multiscale information and attention mechanisms, facilitating the proficient extraction of diverse features from the input dataset, while concurrently establishing interdependencies among distinct features. However, the incorporation of an adaptive soft threshold module efficiently mitigates noise interference. Experiments were conducted on the industrial coke furnace dataset, and it show that the method performs well and has a broad application prospect.

During industrial processes operation, fault diagnosis needs to rely on finite and imbalanced samples for learning because the equipment is usually in a normal operation state and the probability of faults is low. For this issue, we propose a new attention ensemble algorithm, DMSAM-OAdB. First, the multiscale sparse attention module (MSAM) extracts and fuses feature information from different scales to establish the correlation among features and adaptively adjusts the feature weights by soft thresholding operation to emphasize the critical features while filtering the redundant information. Subsequently, a deep network model (DMSAM) constructed by stacking multiple convolutional layers and MSAMs is iteratively trained as the base classifier for the optimized AdaBoost algorithm (OAdB). The OAdB dynamically increases the weights of misclassified samples while mitigating class imbalance by randomly zeroing out the weights of a certain number of correctly classified majority class samples. Finally, the improved focal loss (IFL) function enhances the model’s discrimination capability for minority class samples. Experiments on the industrial coking furnace validate the reliability and superiority of the proposed methodology.

Fault diagnosis is important for ensuring the safe and stable operation of industrial equipment. Due to the characteristics of high dimensionality, nonlinearity and timing in modern industrial processes, the problems brought about by traditional fault diagnosis methods are becoming increasingly evident. Deep-learning-based fault diagnosis demonstrates excellent performance. However, most fault diagnosis models are insufficient in feature extraction and cannot remove more redundant feature information. Therefore, a new model integrating dual-channel attention convolutional neural network (ACNN) and eXtreme gradient boosting (XGBoost) algorithm was developed. Among them, the ACNN can enhance features through channel attention and spatial attention, reduce redundant information, and improve diagnostic performance. To obtain the optimal model, the Adam optimizer with better reliability is used to optimize the model. Finally, the optimal feature information collected by the ACNN is passed to the XGBoost classifier for classification. Extensive experiments on two benchmark datasets, the tennessee eastman (TE) process and a real-world industrial coking furnace, demonstrate that the proposed method achieves superior diagnostic accuracy of 95.64% and 97.44%. The proposed framework shows great potential for deployment in practical industrial scenarios where fast and accurate fault detection is essential.

This paper proposes a unified open-set fault diagnosis (OSFD) strategy based on deep learning classification models and posterior detectors. We utilize a one-dimensional residual network (1D-ResNet) to build a classification model and introduce an out-of-distribution detection method that integrates the fusion of collective and extreme information (FCEI). Specifically, it includes calculating the distance matrix (FCDM) of the feature centers of each class of fault data, then calculating the Mahalanobis distance and cosine similarity between its features and the feature centers of each class for the input samples as the collective information, and finally, using the maximal softmax value of the model prediction class as the extreme information. By combining and analyzing cross-class and extreme information, the method can identify out-of-distribution samples more efficiently. Experiments on the Tennessee Eastman (TE) benchmark process show that the proposed method can not only correctly classify the known faults, but also accurately identify unknown faults.

In industrial fault diagnosis, traditional methods grapple with challenges, such as nonstationarity, nonlinearity, high dimensionality, and strong coupling. To address these issues, we propose an end-to-end fusion model based on multiscale residual convolutional channel attention and transformer model (MRCC-Transformer). This approach initially leverages a multiscale residual convolutional neural network (CNN) to extract data features across various scales, thereby preventing model degradation and autonomously learning and integrating abundant fault information from multiple monitoring variables. Subsequently, a channel attention mechanism (CAM) is introduced to prioritize focus on pertinent convolutional channels to enhance the network’s effectiveness and discriminative capacity. Furthermore, the Transformer is employed to establish dependencies among distinct features to enhance fault diagnosis accuracy. Lastly, the input data is classified for fault diagnosis. The efficacy of the proposed method was validated through simulation experiments on the Tennessee-Eastman (TE) process and an industrial coking furnace. Comparative results demonstrate that the proposed method significantly improves the accuracy of fault diagnosis.