A Semi-supervised Fault Diagnosis Method Based on Deep Adaptation Autoencoder and Manifold Learning for Rolling Bearings

Due to the safety and stability requirements of industrial operations, bearing as a vital part of rotating machinery, its faults diagnosis has received increasing attention. The fault diagnosis can be carried out effectively with the aid of artificial intelligence. However, traditional machine learning methods need abundance of labeled data, which is inconsistent with the facts. Lately, transfer learning, a significant branch of machine learning, has been introduced to overcome this barrier. Transfer learning methods could efficiently utilize datasets with no labels or a small number of labels to help train the classification model. In this paper, a semi-supervised model combining deep adaptive autoencoder and manifold learning is proposed to solve bearing faults diagnosis in a small amount of labeled data. The proposed approach combines the deep adaptation autoencoder (DAA) and uniform manifold approximation training process. In addition, a learn-forget (LF) mechanism is added to the model, which selects and removes unlabeled data based on the confidence coefficient generated by the k Nearest Neighbor (KNN) algorithm to expand labeled data rapidly. The effectiveness of the proposed method is verified by the experiments on bearing vibration signals datasets from Case Western Reserve University (CWRU) and Xi'an Jiao-tong University (XJ).