[論文レビュー] A SISA-based Machine Unlearning Framework for Power Transformer Inter-Turn Short-Circuit Fault Localization

N. Liu, Jingyi Yan+2 more|arXiv (Cornell University)|Mar 7, 2026

Power Transformer Diagnostics and Insulation被引用数 0

ひとこと要約

A SISA-based machine unlearning approach localizes ITSCF faults in power transformers and reduces retraining time by retraining only affected shards, while maintaining near-full accuracy.

ABSTRACT

In practical data-driven applications on electrical equipment fault diagnosis, training data can be poisoned by sensor failures, which can severely degrade the performance of machine learning (ML) models. However, once the ML model has been trained, removing the influence of such harmful data is challenging, as full retraining is both computationally intensive and time-consuming. To address this challenge, this paper proposes a SISA (Sharded, Isolated, Sliced, and Aggregated)-based machine unlearning (MU) framework for power transformer inter-turn short-circuit fault (ITSCF) localization. The SISA method partitions the training data into shards and slices, ensuring that the influence of each data point is isolated within specific constituent models through independent training. When poisoned data are detected, only the affected shards are retrained, avoiding retraining the entire model from scratch. Experiments on simulated ITSCF conditions demonstrate that the proposed framework achieves almost identical diagnostic accuracy to full retraining, while reducing retraining time significantly.

研究の動機と目的

• Motivate robust ITSCF localization when training data can be poisoned by sensor failures.
• Develop a SISA-based MU framework to isolate and remove poisoned data effects.
• Evaluate accuracy and computational efficiency against full retraining using simulated EMI-poisoned ITSCF data.

提案手法

• Adopt LSTM as the baseline model for ITSCF localization of six fault labels (HA, HB, HC, LA, LB, LC).
• Partition the training data into S shards and further into R slices to localize data influence per shard/slice.
• Train independent sub-models per shard and aggregate outputs with a softmax-probability averaging strategy.
• Use SISA-based unlearning to retrain only the poisoned shard(s) starting from the compromised slice, avoiding full retraining.
• Simulate EMI sensor failures to contaminate ITSCF data and assess recovery of accuracy and retraining time.
• Evaluate performance using accuracy and computational time on a PyTorch-based LSTM architecture.

実験結果

リサーチクエスチョン

• RQ1 Can SISA-based unlearning restore diagnostic accuracy after EMI-induced data poisoning in ITSCF localization?
• RQ2 How does the number of shards affect accuracy and retraining time in the SISA MU framework?
• RQ3 Is it possible to achieve substantial training-time speedups without sacrificing classification performance?

主な発見

• SISA unlearning restores accuracy close to full retraining after poisoned data removal.
• Increasing shards yields time reductions (e.g., 2x with two shards, up to ~4x with four shards) but can reduce accuracy when data diversity per shard becomes too limited.
• Non-SISA full retraining serves as a reference with higher training time, while SISA unlearning can achieve significant speedups.
• Confusion matrices show recovery of accuracy across single and multiple poisoned ITSCF conditions, with LV-side phases exhibiting more misclassifications when data is poisoned.
• Overall, SISA-based MU maintains satisfactory diagnostic accuracy while substantially reducing retraining time compared to full retraining.