| Summary: | 碩士 === 國立交通大學 === 網路工程研究所 === 106 === With the rapid advances of machine learning algorithms and sensing technologies, machine prognostics and health management (PHM) via data-driven approaches have become a trend in sophisticated machine tool industry. Recurrent neural network (RNN) is an important technique to handle the sequential sensing data. The popular PHM approaches for machine tool are two types: 1) remaining useful life (RUL) prediction, and 2) anomaly detection (AD). These two PHM approaches have different requirements. Run-to-failure data are necessary to train RUL-based models while AD does not need. Before run-to-failure data are collected, a machine may face abnormal situations, such as impulse signals. Therefore, AD-based approaches are more suitable for machine tool industry compared to RUL approaches. However, due to vanishing gradient problem in RNN, it is difficult to train an RNN model. To this end, we propose a raw data to decision end-to-end framework, called QUAntized Recurrent neural network autoencoder for Time-series anomaly detection (QUART), to accelerate the anomaly detection RNN model training for the machine tools with time-series data. The main idea of QUART is to reduce data size by mapping raw data into finite classes according to probability mass in the raw data. To learn the anomaly states without considering the sequence order of sensing values in a short interval, we adopt a data shuffling method to increase the variance of the training data. Our experiment results show that the proposed framework can accelerate almost 120 times faster than the existing RNN methods.
|
|---|
