Application of Variational Auto-Encoder in Mechanical Fault Early Warning

Due to the non-stationary characteristic of the vibration signals of complex machinery, such as reciprocating compressors and aeroengines, the single feature threshold alarm method that is currently frequently utilized suffers from a shortage of inaccurate alarm. Therefore, a variational auto-encoder-based method for fault early warning is proposed according to the characteristics of the response signals of the mechanical response point. The method is based on high-dimensional feature parameters of mechanical vibration signals. It is used to compress the feature parameters and to self-learn the statistical distribution model from this compressed data via variational auto-encoder. The normal working condition model is used as benchmark model. The fault early warning is realized by comparing the difference between the real-time working condition model and the benchmark model with the adaptive early warning threshold. The advantages of the proposed method are verified via comparison between the single feature threshold alarm method and the state subspace-based early warning method. The results indicate that the method can improve the accuracy of mechanical fault early warning and can greatly advance the alarm time point of the fault. This method is highly timely and strongly adaptable.