Anomaly detection in the time-series data of industrial plants using neural network architectures

Industrial process control systems collect a vast amount of production data, especially time-series data from various sensors in a plant. This data provides a plethora of potential use cases for AI and machine learning that can help plant operators better understand and predict complex plant situations and thereby run their plants more effectively and safely. A relevant use case is anomaly detection, due to the time and cost-intensive process of detecting and rectifying issues, e.g., plant equipment failures. In this paper, the anomaly detection problem for multivariate industrial time-series data is addressed with the help of neural network architectures. In particular, three architectures are evaluated and compared: Dense Autoencoders, LSTM Autoencoders, and LSTMs. The evaluation is performed to detect 20 simulated plant equipment failures for a real-world industrial process found in oil production fields, called the three-phase separator process. The evaluation is done by 1) measuring to what extent the anomaly detection models succeed in detecting the anomalies, and 2) how well the models are able to explain the anomaly root cause. The main contribution of this paper is to find an anomaly detection model that is best suited for the detection and explanation of various failure cases in such an industrial process setting. The evaluation results show the quantitative comparison of the three models metrics and their performances. Here, it is observed that the Dense Autoencoders performed best for the given 20 failure cases.