Module for detection of electrical fault in the propulsion using mechanical and acoustical quantities

A software component for detection of a basic electric fault of the propulsion, implemented as a convolutional neural network, enabling the processing of heterogeneous information from the measurement of mechanical and acoustic quantities. Diagnostic data obtained by measuring the time series of vibrations and/or noise using accelerometers or microphones, or their combination, are normalized, divided into defined time periods and subsequently transformed into a 2D image in grayscale representing instantaneous values of acceleration or sound pressure. These prepared data are fed as input to a 2D convolutional neural network (2D-CNN), which represents a model of the diagnosed system. The neural network is trained on data from a training set created during measurements on a real electrical drive while emulating an electrical fault in the winding (inter-turn short circuit) and during different operating modes of the drive (torque, speed). Experimental data was obtained using several accelerometers located on the motor and one microphone located nearby. The created model therefore contains information from the acoustic area as well as from the area of mechanical vibrations. The CNN model was created in the TensorFlow/Keras environment in Python, and the training and validation of the model took place on a personal computer with a powerful nVidia GeForce RTX 2080T graphics card. The input data is defined as a byte type, and a number of trained network parameters is approximately 150 thousand. The model achieves fault detection accuracy of better than 99% when using input vibration data and over 90% when using data from acoustic measurements.