Cognitive Machine Learning Systems For Real-Time Anomaly Detection In Industrial Iot Monitoring Applications

Abstract

Anomaly detection in IIoT monitoring systems has to be able to identify anomalies in multivariate sensor streams in real time; however, the current anomaly detectors used in practice have been trained once using historic data and with fixed parameters thereafter, making them susceptible to sensor drift, fault signatures, and changing operating conditions common in real-world industrial settings. In this paper, we propose the Cognitive Machine Learning System (CMLS), which combines the reconstruction error-based anomaly detector using an LSTM autoencoder with the Isolation Forest outlier detector using a model-based approach through an ensemble consensus layer, improves the anomaly score using a transformer-based temporal attention, and implements a cognitive self-retraining process that enables model updates automatically using a dual data-driven and model-based metric as soon as sensor drift or degraded reconstruction is detected. The framework is validated using statistically modelled multivariate data from sensors on the SWaT industrial water treatment testbed environment under four different states of operation: normal state, slow sensor drift, a fault event, and a new operating state. When comparing the proposed CMLS with a static Isolation Forest benchmark and a static LSTM-autoencoder benchmark, CMLS has an F1 score of 85.9% under a new operating state, as opposed to 55.3% and 63.8% obtained by the two benchmarks respectively, while also keeping the false alarms below 3.3%. These results suggest that the fusion of ensemble-based anomaly scoring with an autonomous retraining mechanism is very effective in improving robust anomaly detection in IoT environments.