Three-layer Approach to Detect Anomalies in Industrial Environments based on Machine Learning
This paper introduces a general approach to design a tailored solution to detect rare events in different industrial applications based on Internet of Things (IoT) networks and machine learning algorithms. We propose a general framework based on three layers (physical, data and decision) that defines the possible designing options so that the rare events/anomalies can be detected ultra-reliably. This general framework is then applied in a well-known benchmark scenario, namely Tennessee Eastman Process. We then analyze this benchmark under three threads related to data processes: acquisition, fusion and analytics. Our numerical results indicate that: (i) event-driven data acquisition can significantly decrease the number of samples while filtering measurement noise, (ii) mutual information data fusion method can significantly decrease the variable spaces and (iii) quantitative association rule mining method for data analytics is effective for the rare event detection, identification and diagnosis. These results indicates the benefits of an integrated solution that jointly considers the different levels of data processing following the proposed general three layer framework, including details of the communication network and computing platform to be employed.