Fault Detection of Reciprocating Air Compressors Based on Electrical Signature Analysis Using Long Short-Term Memory Networks

The widespread application of reciprocating air compressors endows the condition monitoring and fault diagnosis of these systems with pronounced value for several industries to reduce costs related to maintenance and repair. In this paper, a fault detection scheme for electrically-driven reciprocating air compressors based on Electrical Signature Analysis (ESA) is developed in two different styles. Firstly, a brief description of the electromechanical coupling between the electrical and mechanical sub-systems of the electrically-driven compressor is provided that justifies the idea of using electrical signals as an indicator of mechanical faults. Next, two fault detection schemes using Long Short-Term Memory (LSTM) deep neural networks are proposed. In the first method, temporal sequences of features are used as the network inputs, but the second method merely utilizes raw signals. Experimental results show the ability of these approaches in detecting mechanical compressor faults with high accuracy.