PROBABILISTIC EXTREME LEARNING COMMITTEE MACHINE FOR AUTOMOTIVE ENGINE SIMULTANEOUS FAULT DIAGNOSIS by Jia Siyu Thesis Supervisor: Prof. Wong Pak Kin Department of Electromechanical Engineering, Faculty of Science and Technology The automotive engine is prone to various faults because of its complex structure and wide range of running conditions. Development of a reliable fault diagnosis system with fast and accurate response for automotive engines is therefore of great significance. However, engine fault diagnosis has been a challenging problem due to the existence of simultaneous-faults (i.e. several single-faults appear concurrently) and the high cost in acquiring the exponentially increased simultaneous-fault signals. By studying the literature, it is learnt that signal-based fault diagnostic systems have the best potential to deal with the complicated engine fault diagnosis problem, but those available systems may not give reliable diagnostic results since they usually rely on single classifier and/or only one kind of engine signal. Aiming to enhance the reliability and the function of signal-based engine fault diagnosis systems, this thesis proposes a new diagnostic framework called probabilistic committee machine (PCM). The proposed PCM consists of three stages: 1) feature extraction for processing the engine signals, 2) fault classification using multiple classifiers, and 3) fault diagnosis by combining the classification results through a probabilistic ensemble method. In the first stage, empirical mode decomposition combined with sample entropy and domain knowledge (EMD+SampEn+DK) is used for extracting the useful information from the raw engine signals. In the second stage, multiple sparse Bayesian extreme learning machines (SBELM) networks are built, each of which is trained as an individual committee member for different kinds of engine signals. In the last stage, a new probabilistic ensemble method is proposed, in which each committee member is assigned with an optimal weight in accordance with their reliability and accuracy so that a reliable and widely-covered engine fault diagnostic results can be obtained from the weighted combination of the members. To verify the effectiveness of the proposed PCM, experiments are carried out to obtain representative sample data for classifier construction. The evaluation results show the proposed framework is superior to all the existing single probabilistic classifier. Moreover, while being trained by single-fault patterns only, the proposed system is shown to be capable of diagnosing both single- and simultaneous-faults for automotive engines effectively.