An Uncertainty-Aware Framework for Fault Diagnosis Using Low and High Sampling Rate Signals

Abstract

With the development of the 4th industrial revolution, industrial systems are growing in size and complexity, and automation systems are increasingly being introduced to manage and control them. However, unexpected fault in these systems can have significant social, economic, and human consequences. To prevent and diagnose such failures, researchers are focusing on failure diagnosis techniques for various components of the system. These techniques aim to analyze potential failures in the system, quantify them using health indicators, and manage the health status of the system. In industry, a signal system is being developed to control and manage the industrial systems using various signals such as temperature, pressure, operation, vibration, and acoustic emission signals. Among these, vibration and acoustic emission signals are considered highly sensitive in evaluating the health of the system. These signals are typically acquired using data acquisition sensors and systems with a high sampling frequency of 20 kHz or more. When such high-frequency signals are measured, the main method used for evaluating the health of the system is to perform spectrum analysis and compare the results with those expected in a normal state. The high-sampling signal-based spectrum analysis method is particularly effective for diagnosing early-stage gradual failure or small energy changes such as fine cracks. This approach has been validated in numerous studies and has even been successfully applied in industrial settings. However, for large-scale systems such as modern industrial systems, utilizing high-sampling signals for diagnosis is challenging. The data acquisition systems used for each sensor typically have limited computational capabilities, only able to perform simple calculations such as pre-amplification and linear frequency filtering. Analyzing high-sampling signals through time frequency analysis and other similar techniques requires additional computational facilities that are not readily available in these systems. To overcome the computational burden of analyzing high-sampling signals, low-sampling signals such as root mean square and band pass energy are commonly used in industrial systems for fault diagnosis. However, the applicability of such low-sampling signal-based methods is limited to detecting only radical or large-scale faults that increase energy in all frequency bands, and cannot detect specific frequency reactions. Additionally, a large number of sensors may be reacted to the fault because applicable range for fault diagnosis is restricted on the severe fault, causing inefficiency in maintenance after fault detection. Lastly, the signals measured in industrial sites are often affected by noise or external signals, which results in relatively greater uncertainty compared to signals acquired in laboratories. As a result, it is important to consider uncertainty in the methodology applied to industrial systems in order to estimate or prevent the degree of error that may occur during actual application. Given the current state of industrial systems, it is possible to conduct a fault diagnosis study that efficiently utilizes both low-sampling and high-sampling signals. However, conventional fault diagnosis techniques for industrial systems face three main problems that need to be addressed. Firstly, in low-sampling signal based fault diagnosis, a large number of sensors could respond, making it necessary to estimate the location of the fault for efficient maintenance. Secondly, high-sampling signal-based diagnosis requires a large amount of computation, so a technique capable of robust fault diagnosis is needed even when using a limited amount of data. Lastly, both low-sampling and high-sampling signal techniques need to consider the uncertainty of signals measured at industrial sites. Based on these considerations, this dissertation propose a framework for uncertainty-based fault diagnosis in industrial systems. The first study proposes a methodology for estimating the location of a fault using low-sampling signals. The proposed method presents an energy probability model of the signal measured by a sensor when a fault signal is converted into a low-sampling signal, taking into account the energy difference between the normal state and the fault state signals. When a fault signal occurs at a specific location, the energy measured by numerous sensors can be probabilistically quantified. Then, the fault location can then be estimated probabilistically by deriving a probability value at various specific locations based on the energy ratio between the measured sensors, using the Bayesian inversion. In the second study, a methodology for robust fault diagnosis using high sampling rate signals is proposed. The first step is to evaluate signal similarity using Kullback-Leibler divergence and group similar signals reflecting the operating condition. Then, a probabilistic model of the time-frequency expression of signals is developed to handle variations in operating conditions. Even for a newly measured signal with an unknown operating condition, it can be compared with a group of similar operating conditions using this model. Next, a new feature is proposed to discriminate the fault state by comparing the newly measured signal with the probability models estimated from the signals from the normal state. An adaptive threshold is also suggested, which reflects the range of time-frequency result and corresponding proposed features that vary for the state of signal, to perform robust fault diagnosis. Finally, the data sampling technique is applied based on the result values obtained through the suggested low-sampling signal based approach. This enables the use of the suggested high-sampling rate signal based technique, which helps to reduce the computational time required for analysis while ensuring the stochastic robustness of the final results. By utilizing this combined approach, the diagnostic process becomes more efficient and reliable, leading to improved fault detection and characterization in industrial systems.