Industrial large model-driven online early fault warning: A novel incremental anomaly detection framework with wavelet scattering contrastive transformer

Abstract

Despite substantial potential of industrial large model (ILM), there are still some challenges for deploying ILM in specific industry scenarios: (1) Current ILMs generally adopt learning mode with large-scale parameters, which tends to raise over-fitting on limited data in real-world engineering and fails to generalize to various working conditions; (2) Current ILMs usually employ static deployment based on fine-tuning technique, which is hard to establish a dynamic downstream task with sequentially-collected monitoring data. To address these challenges, this paper selects the early fault detection (EFD) of rotating machinery as an example, and designs an ILM-driven incremental anomaly detection framework that includes a pre-training model and a downstream task. In the pre-training stage, a new wavelet scattering contrastive Transformer (WaveCT) is proposed by introducing wavelet scattering transform to expand physically-meaningful time-frequency feature set from limited data. Through exploring the contextual dependency among normal state data, WaveCT effectively extracts the representation of common one-class detection rules under complex working conditions such as being interference by irregular noise. In the downstream task, an online EFD network is designed based on fine-tuning the sequential data, while an incremental fault alarm indicator is built based on support vector data description (SVDD) generalization error upper bound. Then the statistical distribution shift caused by weak state changes is perceived to achieve early fault warning. Experiments on the vibration signals from our bearing cage testbed and the actual finishing mill bearing of steel enterprise from China validate the effectiveness of the proposed framework.