PE-TSFM: Self-Supervised Time-Series Learning for Generalizable Power Converter Health Monitoring under Unseen Conditions

TL;DR

This work introduces PE-TSFM, a domain-specific time-series foundation model designed for noninvasive health monitoring of power converters under unseen conditions. By pre-training on a large corpus of unlabeled converter data using a patch-based Transformer with a dual-attention mechanism that captures both temporal and inter-sensor correlations, the model learns robust PE-specific representations and is fine-tuned on scarce labeled degradation data. Empirical results show PE-TSFM achieving up to 92% accuracy in OOD scenarios, substantially outperforming traditional methods and generic TSFMs, with ablations confirming the critical role of channel attention and pre-training. The approach enables scalable, interpretable health monitoring and has practical potential for edge deployment in industrial PE systems.

Abstract

Data-driven health monitoring of power converters remains limited by poor generalization to unseen operating conditions. This work addresses this out-of-distribution (OOD) challenge by building a domain-specific time-series foundation model (PE-TSFM) that learns representations directly from large-scale unlabeled converter data. Unlike generic TSFMs trained on broad time-series datasets, the proposed PE-TSFM is pre-trained entirely on domain data, enabling it to learn the physical relationships unique to power electronics. To further tailor the model to this domain, we introduce a dual-attention mechanism that captures both temporal patterns and inter-channel dependencies. While generic TSFMs primarily model temporal dependencies, the added channel attention captures inter-sensor physical relationships essential for converter degradation analysis. A dataset containing 141 million unlabeled timestamps from an operating power converter is used for pre-training. Experiments show that PE-TSFM achieves 92% accuracy under unseen operating conditions. In contrast, generic TSFMs achieve around 60% and conventional time-series models achieve around 40% accuracy. This result confirms the strong OOD generalization of the proposed PE-TSFM. Ablation studies further verify that the introduced channel attention mechanism significantly improves model performance. In addition, we conduct detailed studies on model scalability, hyperparameter sensitivity, and interpretability to provide a comprehensive understanding of the proposed approach.

Figures (8)

• Figure 1: Experimental setup of a 5-kW traction converter used for health monitoring, focusing on the degradation of the power module. All signals (H1–H6) are captured noninvasively through the existing system infrastructure. The photos show the physical testbench and the same power module at four different aging levels (Level 0 – Level 3). Level 0 denotes a new device, while Level 3 approaches the end-of-life condition. The identical power module was stressed by the standard power cycling testing zhang2025power, and each aging state was subsequently retested in the converter platform to collect data.
• Figure 2: Euclidean norm of three-phase current $\|I\|_2$ under four standard driving cycles: NYCC, LA92, UDDS, and HWEFT. The heat map on the lower right shows the similarity matrix of different mission profiles, where smaller matrix values indicate higher similarity. The HWFET profile, which exhibits the largest deviation from the others, is reserved for OOD validation and excluded from model training. The similarity matrix is quantified by the dynamic time warping (DTW) algorithm salvador2007toward.
• Figure 3: PE-TSFM architecture with the self-supervised pre-training and supervised fine-tuning pipelines for power module degradation monitoring.
• Figure 4: Classification performance, which was evaluated under three different random seeds, of full PE-TSFM and comparison models without channel attention and without pre-training.
• Figure 5: The relationship between GPU memory usage and model performance, where the bigger bubble denotes the model has a larger variance under different random seeds. Note that the GPU memory usage is measured during inference on 1 sample with FP32.