DEEP LEARNING BASED FAULT DIAGNOSIS AND PROGNOSIS TECHNIQUES FOR INDUSTRIAL EQUIPMENT USING THE ACOUSTIC EMISSION SIGNAL

Abstract

An unappealing property of all practical industrial systems is the fact that they are vulnerable to malfunction since they are often operated in extreme working conditions. More generally, the unexpected modes of behaviour can cause unexpected failures leading to unanticipated interruptions in industrial production. Specifically, the bearing in rotary machine and boiler tubes in a power plant are the critical components in industrial systems, which usually have faults leading to catastrophic loss in the economy. The diagnosis and prognosis to detect the faults and predict the lifetime of these components are needed for scheduling maintenance. The failure in these machines usually goes with the symptoms such as the generating abnormally vibration or generating of the elastic wave in form of acoustic emission (AE) which can be measured by vibration acceleration sensors or AE sensors. The AE-based signal processing techniques are more efficient for capturing low-energy fault signatures that make them more attractive and appealing to many researchers. Hence, advanced data acquisition and signal processing techniques have been developed for effective and early fault detection and estimate future health in critical systems. Moreover, with the ever-increasing amounts of data, the need for automated methods for data analysis continues to grow. Machine learning and deep learning are thus developed to automatically detect a pattern in data and use the uncovered pattern to predict the future or other outcomes of interest. This dissertation considers these issues related to the fault diagnosis and health prognosis of bearings and boiler tubes using the deep learning methodology. Firstly, a new algorithm of non-mutual exclusive classifier deep neural network (NMEC-DNN) is proposed for the field of detecting the simultaneous occurrence of various types of defects in bearings. The proposed methodology uses a deep neural network architecture based on stacked de-noising auto-encoder (SDAE) and a non-mutually exclusive classifier (NMEC). The NMEC-DNN is trained only using data for single faults but it affords to classifies both single faults and multiple combined faults. The proposed method is implemented and evaluated using experimental bearing data, achieves good classification performance with a maximum accuracy of 95%; and yields better diagnostic performance in comparison to the multi-class classifier based on support vector machines. Secondly, to handle a specific issue in a boiler tube that is the effect of random and high-amplitude impulses from the interaction between the coal fuel stream and the boiler tube membrane, and to propose an impulse detection method, the deep learning flexible boundary regression (DLFBR) architecture is effectively exploited. Thereby, the shape extraction (SE) pre-processing technique is implemented to yield the shape signal, which contains intrinsic information about the impulse from the raw AE signal. Then, DLFBR extracts and generates both the feature map and the confidence mask from the shape signal to regress a boundary box, which specifies the position of the impulse. The proposed algorithm is applied to an experimental leakage detection dataset recorded from a sub-critical boiler unit with a tube membrane. Experimental results show that the proposed method is effective for detecting impulses in a boiler tube test-bed signal and improve the accuracy for leakage detection. Thirdly, a reliable health indicator (HI) methodology for bearing fault prognosis is proposed. The HI model is constructed using historical data, with considering its nonstationary nature, that is decomposed to the different levels of resolution as sub-band; smoothed via a locally weighted regression; evaluated using a gradient-based method. His showing the best trends among all the sub-bands are iteratively accumulated to construct an HI with the best trend over the entire life of the bearing. The efficiency of the proposed approach is validated using several experiments from the run-to-failure test data of the center for Intelligent Maintenance Systems (IMS) and PRONOSTIA dataset, which yields accurate estimates of the RUL of a bearing at different points during its life. Finally, in order to estimate the remaining useful life (RUL), an efficient evaluation method using a recursive extreme learning machine (RELM) integrate the sequential Monte-Carlo framework is proposed for the correct prediction of the lifetime of bearings. The parameters of models are recursively updated to capture the evolving trend in the run-to-failure data by the training process with a particle filter, an instance of the sequential Monte-Carlo framework. Then, the RELM can automatically projection the health state values and estimate the RUL. The effectiveness of the proposed degradation prediction approach is evaluated through the PRONOSTIA dataset of bearing faults and its prognostic performance is compared with existing techniques.