A hybrid fine-tuned VMD and CNN scheme for untrained compound fault diagnosis of rotating machinery with unequal-severity faults

In the case of a compound fault diagnosis of rotating machinery, when two failures with unequal severity occur in distinct parts of the system, the detection of a minor fault is a complicated and challenging task. In this case, the minor fault is overshadowed by the more severe one, and the characteristics of the compound fault are prone to the more severe one. Generally, the proposed methods in the literature consider compound failure as an individual fault type and unrelated to the corresponding single faults, either at the different locations of a sensitive component or in two separate parts, such as the bearing and gear, with approximately the same fault severity. Considering these issues, this study proposes a novel end-to-end fault diagnosis method based on fine-tuned VMD and convolutional neural network (CNN). The main idea is that CNN is trained only on a healthy and single fault dataset, without the use of compound fault data in training. In the test stage of the CNN model, the intelligent method alarms an untrained compound fault state if acquired probabilities of CNN output satisfy a set of probabilistic conditions. The performance of the fine-tuned VMD and the proposed hybrid method is evaluated by the decomposition of a simulated vibration signal and the analysis of a gearbox system with a compound fault scenario in such a way that one fault is minor and the other severe. The results obtained show the high accuracy of the proposed method in compound fault diagnosis and the feature extraction and classification of a minor fault in the presence of a more severe one.