Fault Mechanism Knowledge-Driven Fault Detection of Rolling Bearings with Zero Samples

Deep neural networks have significant advantages in mass data learning, feature mining, accurate pattern recognition, etc., which have received extensive attention and in-depth research, and achieved prominent application performance. However, the network training depth depends on the labeled training set, and in practice, rotating mechanical equipment fault samples are scarce, or even zero samples, there is an imbalance in the training samples, and network training faces difficulties. To complete above challenges, this paper proposes a zero-sample intelligent fault detection method for rolling bearing driven by the knowledge of vibration fault spectrum. Firstly, an improved Hilbert envelope analysis is used to analyze monitored state signal of rolling bearing to obtain spectral features. Considering the fault frequencies of the inner and outer rings of rolling bearings are highly correlated with their structural dimensions, Ix, 2x and 3x frequency data under inner ring and outer ring faults are intercepted respectively as the training samples. Then, a new contrastive loss function is designed to optimize the network-based encoder by introducing the assumption that two samples acquired under continuous time series have high similarity. Furthermore, a fault detection algorithm is established to realize the intelligent detection of rolling bearing fault states. Finally, two standard rolling bearing datasets are used to analyze the effectiveness and reliability of proposed method. The results show that the proposed method can realize accurate detection for inner and outer ring faults with zero fault sample. This provides a new paradigm to carry out the zero-sample fault diagnosis for machinery.