Key points are not available for this paper at this time.
The complex noise environments found in industrial settings lead to a decline in the performance of deep learning-based anomaly detection models. While previous studies have primarily attempted to address this issue by increasing model complexity, this paper proposes an efficient approach that improves signal quality by incorporating the statistical characteristics of the data during the preprocessing stage. In this paper, we analyze the covariance matrix of the training dataset to establish a Global Whitening Baseline and propose a preprocessing pipeline utilizing Zero phase Component Analysis (ZCA) Whitening based on this baseline. This technique removes correlations between frequency bands in the Log-Mel Spectrogram and normalizes the variance by sphering, thereby suppressing the influence of dominant background noise and maximizing the model’s efficiency in extracting subtle anomaly patterns. Furthermore, by applying a multi-frame sliding window technique to the whitened features, we preserved temporal context and ensured robustness against local noise fluctuations. Experimental results show that the proposed preprocessing method improved anomaly detection accuracy Area Under the Curve (AUC) by more than 10% compared to existing Z -score and direct noise removal methods, without increasing the computational complexity of the model. In particular, by eliminating statistical redundancy in the data, it demonstrated excellent convergence speed and stability even in lightweight autoencoder models optimized for edge computing environments. This demonstrates its high potential for use as a highly efficient preprocessing module in real-time quality control and predictive maintenance systems.
Lee et al. (Fri,) studied this question.