Adaptive fault detection via machine unlearning

Data-driven models are widely relied upon for process fault detection. However, their performance is susceptible to the quality of training data. Anomalous data in training or online updates degrade detection models, increasing false alarms or missed detections, and retraining models on corrected datasets is impractical for real-time fault detection. To address this problem, we propose a machine unlearning based adaptive fault detection that updates the model parameters to selectively remove the influence of faulty data from trained models without retraining and compromising the accuracy on normal data. We implement blindspot unlearning on four deep learning-based fault detection models: Autoencoder (AE), Variational Autoencoder (VAE), LSTM-AE, and LSTM-VAE, and evaluate their performance on two benchmark datasets: the Tennessee Eastman Process (TEP) and the Wastewater Treatment Plant (WWTP). We evaluate the models’ fault detection performance before and after unlearning. Our findings demonstrate that unlearning improves fault detection performance while significantly reducing computational overhead. Compared to the original models, unlearned models showcased improved fault detection rate, achieving a 44% increase on the TEP dataset and reaching 90% on the WWTP dataset. Unlearned models achieved fault detection performance comparable to retrained models, reducing computational time by up to 46% on TEP and 33% on WWTP. This validates the effectiveness of machine unlearning for adaptive fault detection.