Analysis of Applicability of Deep Learning Methods in Compressor Fault Diagnosis

The paper presents the results of work carried out on the applicability of deep learning techniques for the purpose of diagnostics of industrial rotary compressors. The paper focuses on the possibility of using the library TensorFlow by Google to build classifiers typical for this library, e.g. convolutional neural networks, as well as classical ones used in diagnostics, e.g. multilayer perceptron (MLP) or support vector machine (SVM). To provide a complete diagnostic tool was not the aim of the paper. Thus, only selected examplary faults were considered dips of the power supply voltage and surge. At the beginning, a description of test stand, from which the test data were collected, is given. The main part of the work contains a description of the implementation of classifiers, and the results of their tests conducted on the actual measurement data. The data, registered during the experiments on site, represented both, the fault free state, as well the state with selected faults. Finally, the concept of the software (functionality and structure) dedicated for using considered techniques for both off-line tasks of building classifiers, as well as on-line monitoring in the cloud is discussed.