A variable selection‐driven global–local probabilistic SFA with Jensen–Shannon divergence for incipient fault detection

Incipient faults present a significant challenge to conventional data‐driven approaches due to their subtle variations, low amplitudes, and susceptibility to noise interference. These faults tend to evolve gradually into severe process anomalies, posing potential risks to system operation. To address this issue, this paper proposes a novel variable selection‐based global–local slow feature analysis method (VS‐GLSFA) for incipient fault detection in multivariate complex processes. First, process variables are partitioned into two subspaces according to their fault‐related significance using a probability‐driven variable selection strategy. Then, the Jensen–Shannon divergence (JSD) is incorporated into the slow feature analysis (SFA) framework to extract discriminative fault features from a probabilistic perspective across multiple scales. Finally, a Bayesian inference mechanism is employed to integrate the detection results from different sub‐blocks, yielding a more comprehensive and reliable decision. The effectiveness and superiority of the proposed method are validated using the Tennessee Eastman process (TEP) benchmark dataset. Experimental results demonstrate that the proposed VS‐GLSFA method achieves a superior average fault detection rate (FDR) of 99.66% while maintaining a low average false alarm rate (FAR) of 5.13%, significantly outperforming conventional principal component analysis and SFA approaches.