음향 전력스펙트럼밀도를 이용한 컨볼루션 신경망 기반 기계 결함 진단 시스템

Abstract

Machine fault diagnosis is a process that automatically detects the states of machine. In fact, the efficiency of machining and the quality of products in the manufacturing process are affected by the tool state, and wear and damaged tools can cause more serious problems in process performance and quality degradation of products, from the viewpoint of automation, it is necessary to develop a system for the progress of tool wear and the prevention of breakage during the process so that the tool can be replaced at an appropriate time. In order to automatically detect the operational status of the entire machine system, including idle drills as well as abnormal drills for diagnosing machine defects, this thesis proposes a system that diagnoses the fault states of tool using the acoustic signal generated by tool that cuts the aluminium work piece on milling machine. In this thesis, CNN (Convolutional Neural Network) technology is used to diagnose machine faults. The sound generated when machine cuts a workpiece is one-dimensional signal. Therefore, it converts one-dimensional signal into two-dimensional signal and uses it as an input from CNN. There are two ways to represent one-dimensional data as an image of two-dimensional data. Both time-series data and frequency-based PSD(Power Spectral Density) are used, both of which apply histogram equalization to increase the homogeneity of data in all data sets. The generated image is used as CNN input. CNN is a good way to handle the complexity of image classification by effectively recognizing and emphasizing the characteristics of adjacent images while preserving the spatial information of images. It is an objective feature extraction and diagnosis method that excludes uncertainties and biases by automatically learning features with a high level of learning method. This thesis provides two methods to diagnose faults using CNN. first is to classify the five states simultaneously using one CNN to diagnose the five states of the collected tools. Next, the five states of the collected tools are divided into three categories, and each state is diagnosed using CNN. This is a hierarchical structure with three CNN classifiers. Both methods showed similar diagnostic results. As a result of comparing CNN model learning with time-series Sound-Image and frequency-based PSD-Image, the accuracy of CNN fault detection using the proposed PSD-Image was excellent. In addition, a comparative study of traditional feature extraction methods and deep learning methods based on the latest technology was conducted, with CNN using the proposed PSD-Image showed high accuracy compared to other feature extraction methods, and showing the results of a robust fault diagnosis system even when noise was injected. Therefore, the proposed system will enable real-time maintenance of industrial field by reducing economic losses by preventing machine failures and improving the quality of machining accuracy.