Anomaly Detection in Rotating Machinery Vibration Signals Based on Domain Transformation and Transformer

Rotating components are essential in modern industrial production, and their safe and orderly operation is of great significance. Due to industrialization, the operating environment for rotating machinery is harsh, and vibration signals are easily disturbed by external factors. Additionally, vibration data is often unlabeled. This paper presents an anomaly detection method for vibration data of rotating machinery, using one-dimensional vibration data to create a dataset. Inspired by BERT, we randomly replace parts of the data before input, enabling self-supervised training on unlabeled data. Order analysis is used to transform non-stationary time-domain signals into stationary angle-domain signals, eliminating noise and unrelated periodic component interference. Subsequently, a transformer-based model is used for anomaly detection, thereby obtaining anomaly scores. Experimental results show that this method outperforms traditional transformer and LSTM algorithms in detecting anomalies in vibration data.