چكيده به لاتين
Due to the progress made in the industry and the development of industrial automation, there has always been an effort to do things automatically. In this thesis, the troubleshooting of one of the most used components of the industry, i.e. rotating machine, is discussed. Rotating machines such as induction motors are widely used in industries due to easy installation, maintenance process and low cost. Therefore, induction motors in critical areas or production lines need timely fault diagnosis to avoid low productivity and critical breakdowns. Short circuit fault is one of the most frequent faults in induction motors and accounts for more than 37% of failures. Therefore, short-circuit fault detection in the early stages has become critical to prevent catastrophic failures and production losses. Intelligent methods of fault detection have been the use of traditional machine learning algorithms, but with the development of data science in all fields and the progress of signal processing methods, the amount of data has increased, so today deep learning has found a more widely used place. In this thesis, the combined structure of convolutional neural network - long short-term memory as a deep network and multilayer perceptron structure as a shallow neural network to detect the primary short circuit fault and also to separate the fault from ambiguous conditions such as load changes and the frequency changes have been investigated and the performance of these two networks has been compared. In this way, two feature models have been extracted from the three-phase current data of the stator, and with both deep and shallow networks, fault detection once in the 7-class mode and once in the two-class mode for both facial feature extraction models. has taken. The results show that the deep network with the first feature extraction model in which the data has been transferred to the time-frequency domain using only discrete wavelet transformation shows better results (in terms of accuracy, sensitivity, and specificity). to give and the shallow network with the second feature extraction model where a number of features are extracted from the data as more preferable features and the dimensions of the data are reduced to 37, better results (in terms of accuracy, sensitivity, and feature) shows. And in general, regarding this data set and in the presence of these disturbances, the shallow network has performed better than the deep network (in terms of accuracy, sensitivity, specificity, and test time). But the advantage of deep network approach is that it does not need manual feature extraction and feature extraction is done unsupervised.
