Deep SVDD-based anomaly detection of machine sound signals using active signal refinement

Detecting abnormal industrial audio is essential but often hindered by irregular background noise and data imbalance. This study proposes a Deep SVDD-based framework integrated with novel preprocessing to enhance detection robustness. The methodology consists of two core modules. First, Shifted Sequential Segmentation applies overlapping segmentation and time-shifting to Log-Mel Spectrograms. This augments training data and ensures temporal invariance by capturing structural features independent of frame boundaries. Second, Active Signal Refinement utilizes the Sobel operator to distinguish signal regions based on gradient magnitude, enabling effective noise subtraction to isolate valid dynamic patterns. Experimental results on the MIMII dataset demonstrate superior performance, achieving an AUC of 0.83 compared to 0.70 for existing Deep SVDD and 0.61 for autoencoders.