Bearing fault diagnosis using weakly supervised long short-term memory

Anomaly detection in vibration signals is an important technique for fault diagnosis, monitoring, and maintenance in nuclear power plants. Therefore, various signal-analysis methods that apply statistical, machine-learning, and deep-learning techniques have been proposed. In particular, deep neural networks (DNNs) have excellent recognition accuracy and do not require the designing of a feature extractor. However, to apply a DNN model for the analysis of time-series data, its parameters must be optimized. This requires not only signal data acquired from real systems, but also data labels that explain any abnormality in the signals. This requires data preparation, and it is time consuming and difficult for humans to annotate manually, especially when the data includes complicated features. Therefore, to extract abnormal features latent in time-series data automatically, we devised a DNN-model training method. To train the DNN model, we propose a novel weakly supervised training method by devising a loss function. We confirmed through experiments that the proposed approach can be used to detect, identify, and localize anomalies in vibration signal data. Furthermore, by applying this method to a fault-classification problem, we confirmed that it can be used to extract features that represent each type of the failures of rotating machinery.