Intelligent fault diagnosis is vital in smart manufacturing. And deep learning-based fault diagnosis has become a hot topic due to its strong feature extraction ability. However, traditional deep learning-based methods exist two limitations. One is that a large number of labeled samples are required to construct effective diagnosis models. Another is that these methods lack flexibility, especially for homologous multi-task problems.In this paper, a novel domain knowledge-based deep-broad learning framework (DK-DBLF) is proposed to overcome above limitations. DK-DBLF consists of two parts: task-specific feature extractor and flexible fault recognizer. The first part is constructed by several convolutional neural networks to obtain abstract features automatically, and the second part employs broad learning system to improve flexibility of DK-DBLF. To combine two parts more effectively, bridge label-based strategy is designed, which is a key connection that can integrate domain knowledge into learning process. The performance of DK-DBLF is tested on motor bearing and pipeline defect datasets, which are health condition classification and homologous multi-task estimation problems respectively. The results have demonstrated that our framework can significantly reduce the usage of labeled samples in learning process, and architecture adjustment can be easily performed when compared with traditional deep methods.

Intelligent fault diagnosis has attracted much attention in industrial processes. The difficulty of collecting fault samples and high price of labeling data, has led to a relative scarcity of labeled data for deep learning tasks in the field. To address this gap, we propose a domain knowledge-guided contrastive learning framework based on complementary data views for fault diagnosis with limited data. Seven data views of either time- or frequency-domains are introduced and designed first. Then, the framework extracts task-specific features by 1) considering complementary information provided by multiple data views to each other, and 2) embedding a domain knowledge-involved space as the guide for the learning process. The results on two bearing datasets show the proposed framework can produce diagnosis accuracies of 96.60% and 94.24% when just 5% of samples have labels. This study determines two pairs of complementary data views that can boost the performance of the proposed framework.

Given that the total labeled fault signals of rolling bearings are difficult to be collected, it is urgent to address the challenge of limited labeled vibration signals for intelligent fault diagnosis. In this case, the combination of artificial features and machine learning (ML) techniques is a potential way to handle this problem, because ML techniques have simple structures so that the large amounts of labeled signals are not demanded in the training process when compared to deep learning (DL)-based approaches. However, the artificial features usually have bias since they are based on experts’ subjective experience, which may leads to performance degradation of the ML-based approaches. To address this gap, a self-supervised framework of artificial feature bias rectification (SFAFBR) is proposed. Specifically, the considered artificial features of vibration signals are from wavelet packet (WP), which are combined with deep networks by serving as the guides to coax the deep networks learning. The learned features are closely relevant to the original artificial features but the bias is rectified due to the sparsely expression of deep model. Then, on the top of that, a ML-based classifier is constructed to perform the final diagnosis. The proposed framework achieved the state-of-the-art diagnosis accuracy of 97.79% and 94.84% on two bearing datasets with just 5 labeled samples. The result shows that using deep networks to assist in rectifying artificial feature bias is a promising way to address fault diagnosis problem of rolling bearings under limited labeled vibration signals.