Anomaly Detection of Sensor Measurements During a Turbo-Machine Prototype Testing - An Integrated ML Ops, Continual Learning Architecture

Building a reliable Machine Learning infrastructure in the Energy domain is a complex task; indeed, it requires a full data integration, continual learning, continual prediction, and the integration of human feedback. Continual learning is a challenging task due to the risk of machine learning to forget former data while learning from new ones (catastrophic forgetting). We present a Machine Learning Operations (MLOPS) architecture able to perform continual learning every day on more than 500 models, to perform inference on new data using such models and to take human feedback and data shift into account. More in detail, we continuously train a recurrent Deep Neural Network to build a virtual sensor from other signals and we compare the prediction versus the real signal to raise (in case) an anomaly. Furthermore, Kullback-Leibler (KL) divergence is used to estimate the overlap between the input distributions available at training time and the distributions seen at test time to estimate the confidence level of the prediction. Finally, we integrate human feedback to tune model retraining. The tool has been applied on a set of about 500 sensors on a three-months long test campaign. Results report 100% recall and 94% accuracy. Moreover, using a recurrent neural network, the system is self-explainable (XAI) by design: indeed, the user can compare the predicted vs real signal to understand the performance of the model. We propose a data intensive MLOPS architecture integrating continual learning, anomaly detection and user feedback. The architecture is based on a standard cloud, it is event-driven and integrates retraining based on GPU capabilities.