Improved Mel Frequency Cepstral Coefficients for Compressors and Pumps Fault Diagnosis with Deep Learning Models

Abstract

Compressors and pumps are machines frequently used in petroleum and chemical industries for fluid transportation through flow systems to keep industrial processes running permanently. As their failure can produce costly disruption, developing fault detection and diagnosis tools is essential for accurately detecting and diagnosing faults. This research proposes a bi-dimensional representation of the vibration signal corresponding to the Mel Frequency Cepstral Coefficients (MFCC) and their first two derivatives as features. The pseudo-periodic nature of the fault signature in rotating machines is exploited to put forward an efficient and accurate patch-wise fault classification method. This approach enables the classification of 13 combined types of faults in a multi-stage centrifugal pump and 17 faults in a reciprocating compressor. Classification is performed using the Long Short-Term Memory (LSTM) network, the bidirectional Long Short-Term Memory (BiLSTM) neural network, and the Convolutional Neural Network (CNN). Accurate classification over 99% is attained, showing that the proposed feature extraction procedure correctly classifies a large set of faults simultaneously appearing in such rotating machines.