Applied Anomaly Detection: a Bayesian Approach to Improve Robustness

Abstract

The main drivers of the development and production of new energy devices are energy device efficiency and machinery maintenance strategies. The former to minimize pollutant emissions with a view to future carbon neutrality. Condition-Based Maintenance (CBM), on the other hand, can help improve machinery reliability and reduce downtime by monitoring equipment conditions and addressing potential problems before they become serious. It can also save companies money by reducing the number of unnecessary repairs, minimizing the need for spare parts, and optimizing maintenance schedules. In this paper, the authors propose a deep learning methodology to automatically detect anomalies on a real Combined Heat and Power (CHP) unit supplying a school in Germany. The core of the work is a convolutional autoencoder trained on the normal behavior of the energy generator. The autoencoder is enhanced with a Bayesian technique, the Monte Carlo dropout, used to add a stochastic component to the model to quantify the uncertainty degree of the detection. This information is crucial to determine if or when action is actually needed, optimizing the service and maintenance strategy. The proposed approach was applied to a real case study and was found to be effective, heat exchanger fouling was detected 5 weeks before the standard detection system. The algorithm returns high confidence in system anomalies and low detection confidence for minor alterations in behavior, less risky for the machine.