An Unsupervised Deep Learning Framework for Anomaly Detection

In recent years, with the evolution of technology and hardware, people can per-form anomaly detection on machines by collecting immediate time series data, thereby realizing the vision of an unmanned chemical factory. However, the data is often collected from multiple sensors, and multivariate time series anomaly detection is a difficult and complex problem because of the different scales and the unclear interaction of each feature. In addition, there usually exist noises in the data, and those make it difficult to predict the trend of the data. Moreover, practically, it's hard to collect abnormal data, thus the imbalance is an important issue. Recently, with the rapid development of data science, unsupervised methods based on deep learning manner have gradually dominated the field of multivariate time series anomaly detection. In this paper, we propose a 3D-causal Temporal Convolutional Network based framework, namely TCN3DPredictor, to detect anomaly signals from sensors data. Our proposed TCN3DPredictor modifies multi-scale convolutional recurrent encoder-decoder by 3D-causal Temporal Convolutional Network which can learn the interaction and temporal correlation between features and even predict the next data. Based on the results of 3D-causal Temporal Convolutional Network, a new breed of statistical method is proposed in our proposed TCN3DPredictor to measure the anomaly score precisely. Through a series of experiments using dataset crawled from a computer numerical control (CNC) metal cutting machine tool in a precision machinery factory, we have validated the proposed TCN3DPredictor and shown that it has excellent effectiveness compared with state-of-the-art anomaly prediction methods under various conditions.