New AI Generates Bearing Fault Signals for Improved Maintenance Decisions.

This research introduces two novel approaches for synthesizing bearing vibration data, specifically focusing on generating signals that correspond to intermediate fault probabilities. These "gray-zone" samples are typically rare in real-world datasets but are critical for understanding decision boundaries in predictive maintenance and for situations requiring careful inspection. The first method, Probability-Regularized Generative Adversarial Network (PR-GAN), modifies real signals using a residual generator to push classifier outputs towards a target probability. The second, a training-free counterfactual (CF) procedure, directly optimizes input signals to achieve the desired probability while staying close to the original. The study evaluated both methods on standard bearing datasets, measuring their accuracy in hitting target probabilities and the magnitude of changes required. The counterfactual method consistently outperformed PR-GAN, achieving significantly lower mean absolute probability errors and higher success rates in steering signals to the target. While PR-GAN had a lower runtime in some cases, the CF method demonstrated superior reliability and required smaller modifications to the original signals, making it more effective for generating precise, borderline fault data.