Intelligent vibration-monitored data-driven anomaly detection for rolling bearings

In the actual industrial environment, the equipment is usually in normal working condition, resulting in the lack of failure modes and data, so it is difficult to establish a mechanical equipment classification and recognition model based on labeled samples. Considering the significant advantages of unsupervised learning in handling unlabeled data, this paper compares four typical detection methods. The article utilizes Fourier Transform to extract 12 frequency domain features from the original vibration signals, representing fault information to assist anomaly detection models. Subsequently, healthy data is used to train the models to obtain anomaly detection thresholds; finally, the detection results are obtained by calculating the scores of the test samples. In the experimental study, the XJTU-SY bearing dataset is selected to compare the detection performance of four typical anomaly detection methods One-Class Support Vector Machine (OCSVM), K-Nearest Neighbors (KNN), Principal Component Analysis (PCA) and Isolation Forest (iForest). In addition, the early fault location, false alarm number and missing alarm number are used as the index evaluation model. Among them, the iForest anomaly detection method shows excellent anomaly detection performance.