An Evaluating Study of Using Thermal Imaging and Convolutional Neural Network for Fault Diagnosis of Reciprocating Compressors

As an essential mechanical device in many industrial applications, reciprocating compressors may be subject to thermal performance failures, mechanical function failures and motor faults resulting in extremely severe catastrophic collapses. Generally, the presence of such faults affects the temperature field distribution of the device. Infrared thermography technology can detect the thermal radiation signal of an object and converts it into images, which is sensitive and reliable to monitor the condition of reciprocating compressor systems. In this paper, three kinds of faults are simulated in an uncontrolled temperature environment. The temperature distribution signal of a reciprocating compressor is captured by a remote infrared camera in the form of a heat map during the experimental process. A slight shaking window is employed to crop the photographed range of experimental equipment, and 30% of each type of images are flipped to prevent the image position information from affecting the classification results. A convolutional neural networks (CNN) is involved for evaluating the monitoring by classifying three common faulty operations. The results demonstrate that thermal images contains the full information and can be a promising technique to diagnose the faults of reciprocating compressors under various operating conditions with a classification accuracy of more than 98.59%.