Noise-resilient process fault detection via multi-scale pls and distribution monitoring metrics

Rapid and accurate fault detection is crucial for maintaining safe and reliable operations in modern process industries. However, significant measurement noise often obscures subtle fault signatures, degrading the performance of conventional fault detection (FD) methods. This study proposes an efficient fault detection strategy that integrates wavelet-based multiscale filtering with Partial Least Squares (PLS) modeling and employs the Hellinger Distance (HD) as a distribution-based fault indicator. The multiscale PLS (MSPLS) approach effectively suppresses measurement noise while preserving essential process features, and the HD metric enhances sensitivity to small deviations often missed by traditional monitoring indices. A kernel density estimation (KDE) technique is used to derive robust statistical thresholds for fault decision-making. The proposed MSPLS–HD technique is evaluated using a benchmark Continuous Stirred-Tank Reactor (CSTR) process, under both low (Signal-to-Noise Ratio (SNR) = 15) and high noise (SNR = 5) conditions, and tested across three categories of faults. In addition, we performed a comprehensive performance analysis with different wavelet types and assessed the MSPLS–HD technique under varying decomposition depths. Comparisons with machine learning methods, including Gradient Boosting Regressor (GBR), Random Forest Regressor (RF), and Multi-Layer Perceptron (MLP) Regressor combined with Exponentially Weighted Moving Average (EWMA) chart monitoring, were also conducted to further validate the effectiveness of the proposed approach. The method consistently outperforms conventional PLS-based monitoring schemes, achieving an average F1-score exceeding 98% in high-noise scenarios. These results demonstrate the effectiveness and robustness of the proposed approach for fault detection in complex industrial environments.