Using Deep-Learning Approach to Detect Anomalous Vibrations of Press Working Machine

In recent years, there has been a demand for advanced maintenance in factories. Data collection from factory equipment is being carried out, and the collected sensor data is widely used for statistical analysis in quality control and failure prediction by machine learning. For example, if it is possible to detect an abnormality using vibration data obtained from an equipment, increase in the operation rate of the plant can be expected. In this research, we aim at early detection of equipment failure by finding signs of abnormality from vibration data, using a deep-learning technique, particularly an autoencoder. In this paper, the following two methods were tested. The first scheme is based on the reconstruction error in an autoencoder. An autoencoder is trained using normal data only. Looking at the difference between input data and reconstructed data, we can regard the data having higher difference as abnormal. In the second approach, given the input data, values of the middle layer of the autoencoder are extracted, and we calculate the degree of abnormality using a Gaussian Mixture Model (GMM), representing a data set by superposition of a mixture of Gaussian distributions. In this framework, regarding an autoencoder structure, we tested both full-connection networks and convolutional networks. In this work, we chose a press machine. Frequency characteristics were acquired from the data in production mode of a press machine. Then using each method, we evaluated whether abnormality could be found by calculating the degree of abnormality. We employed two-day data without failure as training data, and another data set was prepared as forecast data obtained on the following days; on one of the days the machine stopped due to a sudden abnormality. Similar to time-series signal processing, we applied framing processing so that we can analyze data even in the case we can only get a small amount of data. As a result, our method succeeded in finding the day when the abnormality occurred and the machine stopped. In addition, the degree of abnormality became higher before the abnormality occurs, indicating we can detect signs of abnormality. In conclusion, the degree of abnormality could be calculated using the reconstruction error using an autoencoder from the vibration data during production, and the method using GMM from the middle layer of autoencoder. We consequently conclude it is possible to detect a sudden abnormality in which the device stopped, from actual vibration data. These results provide new solutions for equipment failure estimation.