A Self-Commissioning Edge Computing Method for Data-Driven Anomaly Detection in Power Electronic Systems

TL;DR

This work tackles the challenge of deploying data-driven anomaly detection for power electronic converters in field settings by addressing lab-to-field transfer and edge-device memory constraints. It proposes a self-commissioning method that online-trains a neural network on streaming data using a fixed-size buffer with a prioritized, high-error sample replacement strategy, aiming to mitigate catastrophic forgetting. Compared with online baselines and memory-replay variants, the proposed buffer-selection method achieves higher AUC and faster training, approaching offline performance, and it runs efficiently on edge hardware. The approach enables reliable, field-adaptive, self-commissioning condition monitoring for power converters, reducing reliance on cloud resources and enabling broader, low-cost deployment in industrial environments.

Abstract

Ensuring the reliability of power electronic converters is a matter of great importance, and data-driven condition monitoring techniques are cementing themselves as an important tool for this purpose. However, translating methods that work well in controlled lab environments to field applications presents significant challenges, notably because of the limited diversity and accuracy of the lab training data. By enabling the use of field data, online machine learning can be a powerful tool to overcome this problem, but it introduces additional challenges in ensuring the stability and predictability of the training processes. This work presents an edge computing method that mitigates these shortcomings with minimal additional memory usage, by employing an autonomous algorithm that prioritizes the storage of training samples with larger prediction errors. The method is demonstrated on the use case of a self-commissioning condition monitoring system, in the form of a thermal anomaly detection scheme for a variable frequency motor drive, where the algorithm self-learned to distinguish normal and anomalous operation with minimal prior knowledge. The obtained results, based on experimental data, show a significant improvement in prediction accuracy and training speed, when compared to equivalent models trained online without the proposed data selection process.

Figures (8)

• Figure 1: Simplified circuit diagram of the monitored frequency converter in the experimental setup, as part of a Danfoss FC 302 variable-frequency motor drive.
• Figure 2: Picture of the motor test bench used to collect experimental data.
• Figure 3: Block diagram depicting the experimental test bench. Solid lines represent electrical connections, dotted lines represent drive-internal measurements and communication, and dashed lines represent external communication. Motor 1 (M1) is speed-controlled, while motor 2 (M2) is torque-controlled. The two inverters are connected to a common dc-link.
• Figure 4: A portion of the collected data set, for normal operation (top) and anomalous operation (bottom). The output RMS average phase current $i_{out}$ (blue line, solid) is plotted on the left axes, while the heat sink temperature $T_{hs}$ (orange line, dashed) is plotted on the right axes.
• Figure 5: Experimental results in a test run, using a model trained with online learning with the proposed sample selection method. In this example, the false positive rate is 3.7%, and the true positive rate is 85.7%.