Beyond reconstruction: A deep convolutional critic with spectral normalization for superior bearing fault anomaly detection

Unsupervised anomaly detection is paramount for the predictive maintenance of rotating machinery. However, prevalent reconstruction-based models like Autoencoders (AEs) often exhibit architectural complexity and can yield suboptimal discriminative performance. To address these limitations, this study proposes a novel, non-reconstructive framework centered on a Deep Convolutional Critic (DCC)—a deep 1D Convolutional Neural Network (CNN) designed to directly learn a “normality score” from raw time-series data. Critically, we apply Spectral Normalization (SN) to all convolutional and linear layers, a key technique that regularizes the model to ensure a smooth, stable, and highly discriminative scoring function. The efficacy of the proposed method was rigorously evaluated through extensive experiments on three diverse public bearing fault datasets (CWRU, SEU, and MFPT), benchmarking against prominent methods including Autoencoders and Deep SVDD. The DCC model demonstrated exceptional and generalizable performance, achieving a perfect Area Under the ROC Curve (AUC) of 1.0000 across all three datasets, significantly outperforming all baselines. Ablation studies confirmed that spectral normalization is essential for preventing score explosion during training and achieving this state-of-the-art performance. Furthermore, visual analysis revealed that the model learns a perfectly separated feature space, mapping normal and anomalous data into distinct, non-overlapping clusters. This study demonstrates that a direct, well-regularized scoring approach is a more potent, efficient, and robust strategy for unsupervised time-series anomaly detection than indirect, reconstruction-based modeling. The proposed DCC framework offers a powerful, high-performance, and practical solution for deployment in industrial predictive maintenance systems. Problem–Solution–Contribution Pathway for Bearing Fault Detection Using the Lightweight DCC Model • Beyond reconstruction: A deep critic for superior fault detection. • Spectral normalization creates a stable, robust scoring function. • Perfect anomaly detection (1.0000 AUC) across diverse bearing datasets. • Direct scoring learns a perfectly separated feature space. • Demonstrates exceptional robustness to signal noise in fault diagnosis.