| Fault diagnosis is an essential area in prognostics and health management, to identify and isolate the developing faults in a machine. It provides diagnostic information that allows spare parts to be ordered and replaced before failure, and can also help identify the root cause of the fault to prevent the problem from reoccurring. Many classification algorithms have been applied for machine fault diagnosis, but few studies benchmark two or more methods at a time, and most fail to describe the strengths of each method. For that reason, it is difficult to know which fault diagnosis methods to apply for different applications, since there is no single algorithm that will perform the best for all scenarios. Choosing the appropriate method is critical for maximizing the diagnosis accuracy. Therefore, this research focuses on reviewing and evaluating a number of selected classification algorithms for machinery fault diagnosis modelling, in order to establish a knowledge base which can be used to develop more accurate PHM systems in a shorter time period.;The models evaluated in this study include k-nearest neighbors, naive Bayes classifier, Bayesian belief network, self-organizing map, support vector machine, and random forests. In order to understand the capability of each model for different scenarios, this study benchmarks with two datasets from two different applications. The first dataset contains simulated healthy and faulty data from a large stationary natural gas engine used for power generation applications. The simulated faults include a seized wastegate valve, clogged injector, poor fuel quality, and all possible combinations of the 3 failure modes. The challenge of the case study is that only a few features are available and the system is complex with many interacting components. Compared to each model, the Bayesian belief network performs the best, because it can incorporate domain knowledge of the complex relationships amongst the features and failure modes. It provides the highest diagnosis accuracy with an average accuracy rate of 92.6%, and also gives the most consistent results.;The second dataset is from a ball screw driven linear motion system testbed in a sensor-less environment, where only controller data is available (motor speed and motor torque). The objective is to diagnose the preload loss of two different components, ball screw and guideway. Three levels are defined for the preload condition, new (green), worn (yellow), and faulty (red). Thus, 3 ball screws and 3 guideways with 3 different preload levels are tested to simulate 9 possible combinations. Since motor signals are often noisy and unstable, the case study aims to evaluate each models robustness to noisy and irrelevant features. It also compares how many training samples are required for each model to perform well, considering that faulty samples are difficult to obtain, since machines do not fail often and induced failure tests are expensive. The results show the random forest model is the most robust model with the highest diagnosis accuracy, achieving an average accuracy of 99.6%, and requires very few training samples. |
