Research on fault diagnosis and early warning of reciprocating compressor based on stacked convolutional autoencoder optimized by gradient differential evolution

Abstract

The reciprocating compressor is one of the key equipment in the process industrial field. Due to its complex structure and motion state, the bearing bush of the connecting rod is prone to wear failure. In the early stage of wear failure, the monitoring signal signs are very weak. As a result, it has produced bad results that identify the fault signs by using traditional data processing and spectrums analytical methods. Aiming at the early fault identification of the bearing bush, unsupervised feature mining based on auto-encoder principle and super-parameter optimization based on Gradient-Differential-Evolution are utilized, and an early-warning-model based on Gradient-Differential-Evolution and Stacked-Convolutional-Autoencoder is proposed. In order to study the sensitivity of the vibration signal and piston rod settlement signal to the early stage of wear failure, the two signals are input into the early warning model for comparison. In addition, they are fused to verify the improvement ability of multi-source signal on early warning. Moreover, to verify the early fault recognition performance of the proposed methods, the proposed method is compared with the other two early-warning-models based on Stacked-Autoencoder and Convolutional-Neural-Networks. The actual fault case analysis results show that based on the Gradient-Differential-Evolution optimization model, the difficulty of parameter setting can be effectively reduced and the proposed method has significant advantages to detect the early warning timely and effectively.