Fault Diagnosis and Fault Severity Prediction Based on Computational Intelligence Techniques for Industrial Processes

Abstract

As industrial processes become more integrated and complex, the impact of faults rises, and process monitoring, and maintenance becomes more challenging. Though fault detection is crucial for mitigating risk, a comprehensive understanding of fault diagnosis and prognosis is essential for identifying the root causes and ensuring systemic safety. Previously, fault diagnosis based on reconstruction methods using principal component analysis (PCA) has been oriented towards sensor faults outside of the control loop, neglecting the intricacies of process fault that impact multiple variables due to intricate correlations between these variables. This thesis bridges this gap by offering enhanced statistical and machine learning methodologies for diagnosing and prognosing process faults. Typically, a process fault will cause process measurements to move in a specific direction within the measurement space. The first principal component (loading vector) of a PCA model, when applied to known faulty data, can effectively capture the primary characteristics of the fault, thereby delineating the direction of deviation. This study enhances fault reconstruction methodologies by applying PCA to historical fault data, thereby establishing a fault direction matrix that encapsulates the multidimensional nature of process faults. It is imperative to note that this approach presupposes the availability of historical fault data for model training and is primarily effective for previously encountered faults. For emergent faults, without historical data, alternative methods or the incorporation of new data types would be necessary for effective diagnosis. In this study, fault prognosis focuses on the long-range prediction of the reconstructed fault magnitudes. In this thesis, time series prediction models based on machine learning are developed, utilizing reconstructed fault magnitude data from historical process data and are used to predict future fault magnitudes. The thesis presents using extreme learning machine (ELM) models and long short-term memory (LSTM) networks to build fault magnitude prediction models, as well as autoregressive (AR) models as a baseline for comparison. ELM models show improved prediction results compared to AR models. Nevertheless, during the training phase, the adjustment of the ELM network's parameters is confined to the output weights. The hidden layer weights and biases, initially set at random, remain unchanged. This design choice, made by the developer, aims to streamline the training process, but it may lead to less consistent predictive performance across different datasets. In contrast, LSTM networks, with their memory units, are better suited for modelling dynamic relationships and handling long-term dependencies. To enhance the reliability and accuracy of long-range prediction, this study employs an LSTM network model for multi-step prediction of fault magnitude. The applicability of this method is illustrated through a simulated continuous stirred tank reactor (CSTR) process, suggesting that such model could be beneficial in industries such as chemical production, pharmaceutical manufacturing, and energy generation. This model could be integrated into monitoring systems to optimize predictive maintenance, enabling early fault detection, and preserving the integrity of production processes. While this thesis primarily focuses on the individual contributions and applications of PCA-based methods, ELM, and LSTM models in process fault prediction, it acknowledges the potential of their integrated use. The combination of these methods, although not explored in depth within this work, poses a promising avenue for future research. Such a hybrid approach could synergistically utilize the early detection strengths of PCA-based methods, the rapid prediction capabilities of ELM, and the long-term dependency learning of LSTM models. This integrated model could potentially lead to more advanced, robust, and comprehensive fault prediction systems, forming an exciting and valuable direction for future research endeavours.