Ensemble support vector machines and dempster-shafer evidence theory for machinery multi fault diagnosis

Abstract

Machinery fault diagnosis is essential for ensuring the integrity of machinery. To this end, vibration analysis has been proven to be the most effective method. However, its effectiveness is highly dependent on the experience and knowledge of the machine operator due to abundance of various machine parameters and the complexity of machinery. Thus, artificial intelligence (AI) or the machine learning approach provides a more consistent diagnostic result based on a trained machine learning model and hence leads to a more automated fault diagnosis system that minimizes human intervention. Support vector machines (SVM) are frequently used in automated machinery fault diagnosis to classify multiple machinery faults by handling a high number of input features with small data sets. However, SVM is well known for binary fault classifications only (i.e., healthy vs. faulty). When SVM is used for multi-fault diagnostics and classification, it results in decreased classification accuracy; this is due to the adaptation of SVM for multi-fault classification which requires the reduction of multiple classification problems into multiple subsets of binary classification problems, producing many contradictory results from each individual SVM model. Thus, this research aims to improve the multi-fault classification accuracy of SVM by the adaptation of DempsterShafer (DS) evidence theory which is referred as Ensemble SVM-DS. Besides, a novel feature selection tree (FST) is proposed to improve the computation time of a wrapper-based feature selection algorithm such as a genetic algorithm (GA) as part of the improvement for the proposed model. In order to fulfil the objectives of this study, the scope of the work is divided into two parts: the algorithm development and the experimental study. The initial model of feature selection and fault diagnosis algorithm is developed by using a bearing dataset downloaded from the Case Western Reserve University Bearing Data Center website specifically to represent healthy and faulty ball bearing conditions. Then, the proposed algorithms are validated with two sets of vibration signals which are recorded in the laboratory at a measured velocity with a sampling frequency of 2.56 kHz from the belt-driven machinery and SpectraQuest rotating machinery, respectively. The analysis showed that the FST is 13 times faster than the GA at selecting an optimal feature subset. The novel Ensemble SVM-DS model is developed to resolve conflicting results generated from each SVM model and thus increase the multi-fault classification accuracy. The analysis showed that the proposed Ensemble SVM-DS model improved the fault diagnostic accuracy of bearings (from 76% to 94%), belt-driven machinery (from 52% to 82%), and SpectraQuest rotating machinery (from 48% to 72%), as the Ensemble SVM-DS continuously refined and eliminated all conflicting results from traditional SVM models. The proposed Ensemble SVM-DS model was found to be more accurate and effective at handling multi-fault diagnostic and classification problems commonly faced by industry, and was found to be capable of general-purpose machinery fault diagnosis.