Wavelet Knowledge-Driven Transformer for Intelligent Machinery Fault Detection With Zero-Fault Samples

Extensive investigations have been conducted to address the challenge of machinery fault detection (FD) under zero-fault samples, with a particular focus on exploring transfer learning and data generation methods. However, the fault samples generated from a laboratory cannot well represent the real-world conditions of machinery in real scenarios, leading to significant challenges in the online application of an offline-trained FD model. To overcome the limitations of insufficient actual fault samples, a novel FD approach that does not rely on any fault samples is proposed. First, a continuous wavelet transform knowledgebase is established using different mother wavelet functions to process the monitoring signals from multiple perspectives. Then, a novel contrastive loss function is formulated to optimize the feature encoder by measuring the similarity between distinct time-frequency features. Finally, an FD algorithm is developed based on the extracted representative features of the monitoring signal to detect anomaly conditions in machinery. The effectiveness of this approach is evaluated using two real-world fault datasets obtained from the industrial robot and turbine test rig. The results demonstrate that the proposed method achieves a high-performance FD under zero-fault samples.