변화점 시공간 확률과정 모형을 활용한 이미지 기반 이상 진단 및 분류

Abstract

Along with the advancement of sensing and data collection technology, data-based diagnosis and fault detection based on sensor data have been developed in manufacturing industry. From de- vices or equipment, various format of condition data can be obtained, and anomaly detection and classification for repair or replacement is conducted through the analysis. Representatively, signal and image data can be used for examining the current state of facilities. To deal with these types of condition data, signal processing (SP) methods are applied throughout the entire maintenance procedure, from data preprocessing to failure diagnostics. There are two types of representative SP methods; Fourier transform (FT) and wavelet trans- form (WT). FT converts the signals at time domain into frequency components having periodic function. Using the basis function, the forward and inverse transform are conducted for the transfor- mation. All Fourier basis functions have global supports, but do not preserve local or time-varying representation. As an alternative to consider the local regularity, WT is developed. WT decomposes the signal and image data using wavelets, which are a collection of basis functions with respect to time and frequency domain, and obtains the coefficients for each domain. WT is effective for the data containing irregular patterns since this method extracts both of time and frequency features from the observations. Using two types of key functions, wavelet function and scaling function, signal decomposition in frequency and time is conducted, respectively. To conduct the abnormal detection and classification based on WT and FT, various methods for noise removal, feature-based fault diagnosis, and degradation modeling were proposed, improv- ing the overall maintenance performance. Data reduction and de-noising should be conducted to refine the original measurements for better detection or classification results. After converting the observations into Fourier or wavelet coefficients, insignificant coefficients can be considered as noise and removed, which are less than the specific threshold value. In addition, based on the statistical characteristics of converted coefficients, the primary information can be extracted to distinguish the abnormal observations from normal data. Based on the self-similar property from transformed coef- ficients, energy-based feature can be extracted as health indicator, which represents the irregularity or variation of given observations. Aside from these, SP can be integrated into stochastic process for degradation model, providing improved representation and computation. Based on the decom- position of large-scale observations into the summation of basis function, the degrading pattern and its change-point to anomalies can be effectively estimated. In this thesis, we aim to propose SP-based abnormal detection and fault classification approaches for prognostics and health management (PHM) of equipment and machinery. Analyzing the signal or image sensor data obtained from facilities, we propose diagnostic methods for noise removal, fea- ture extraction, and degradation modeling. The proposed methodologies provide efficient condition diagnosis methods based on the key information from complex equipment, and effective detection results identifying the abnormal observations. The contribution of this thesis are summarized as follows: Firstly, we suggest a robust noise reduction method for signal data to deal with irregular and biased measurements in sensor data. To provide a stable data reduction result, the concept of scale- invariant order statistics is introduced to the thresholding rule. In addition, the false discovery rate arising from type I errors is considered by providing a flexible and customized attributes to control the degree of data reduction. Secondly, we provide an energy-based feature extraction and fault classification for image data. Based on the self-similar characteristics of hierarchically decomposed coefficients, spectral analysis is conducted to derive the energy based features. The energy spectrum can capture the variations at both of high- and low-frequency domain. The extracted feature can be utilized for fault diagnosis and classification to evaluate the status of facilities. Finally, we suggest a spatial and temporal degradation modeling to comprehend the deterio- rating patterns in equipment and detect abnormal status based on time-series image data. To con- sider diverse degrading aspects in normal and abnormal states, we establish a change-point-based spatio-temporal process (CP-STP). We also developed the sequential procedures for determining the optimal number of change points (CPs) to define abnormal state.