BEARING FAULT DETECTION BY USING AUTOENCODER CONVOLUTIONAL NEURAL NETWORK

Abstract

Context. Bearings are an important part for the functioning of various means of transportation. They have the property of wear and failure, which requires high-quality and timely detection of faults. Failures are not always easy to detect, so the use of traditional detection methods may not be effective enough. The use of machine learning methods well-suited to the task can effectively solve the problem of detecting bearing faults. The object of study is the process of non-destructive diagnosis of bearings. The subject of study is methods of selecting hyperparameters and other optimization for building a diagnostic model based on a neural network according to observations.
Objective. The goal of the work is to create a model based on a neural network for detecting bearing faults based on the ZSL.
Method. A proposed filter smooths the data, preserving key characteristics such as peaks and slopes, and eliminates noise without significantly distorting the signal. A normalization method vibration data is proposed, which consists of centering the data and distributing the amplitude within optimal limits, contributing to the correct processing of this data by the model architecture. A model based on a neural network is proposed to detect bearing faults by data processing and subsequent binary classification of their vibrations. The proposed model works by compressing the vibration data into a latent representation and its subsequent recovery, calculating the error between the recovered and original data, and determining the difference between the errors of healthy and faulty bearing vibration data. The Zero-Shot Learning machine learning method involves training, validating the model only on healthy vibration data, and testing the model only on faulty vibration data. Due to the proposed machine learning method, the model based on a neural network is able to detect faulty bearings present in the investigated fault class and theoretically new fault classes, that is, the model can detect different classes of data that it did not see during training. The architecture of the model is built on the convolutional and max-pooling layers of the encoder, and the reverse convolutional layers for the decoder. The best hyperparameters of the model are selected using a special method.
Results. Using the Pytorch library, a model capable of binary classification of healthy and faulty bearings was obtained through training, validation, and testing in the Kaggle software environment.
Conclusions. Testing of the constructed model architecture confirmed the model's ability to classify healthy and fault bearings binaryly, allowing it to be recommended for use in practice to detect bearing faults. Prospects for further research may include testing the model through integration into predictive maintenance systems for timely fault detection