MFPPG: Multi-feature fusion and patch-wise perturbation generation for anomalous sound detection

Anomalous sound detection (ASD) plays a crucial role in industrial systems by enabling early fault diagnosis and ensuring operational reliability. However, conventional ASD methods often rely solely on log-Mel spectrograms, which lack sensitivity to transient patterns and fine-grained structural variations. Additionally, shallow data augmentation is unable to model semantic perturbations in the high-level feature space, thereby limiting generalization. To address these challenges, we propose Multi-feature Fusion and Patch-wise Perturbation Generation (MFPPG), a lightweight method for unsupervised ASD. First, MFPPG introduces a multi-feature attention fusion method that integrates four complementary time–frequency representations—log-Mel, Teager-Kaiser Energy Operator (TKEO), Temporal Spectral Residual (TSR), and Frequency Spectral Residual (FSR)—to capture diverse acoustic cues. To enable semantic interaction across modalities, we design an attention-based fusion module that dynamically enhances complementarity. Then, a Patch-wise Perturbation Generation (PPG) module injects structured, localized noise into semantic feature maps, enhancing contrastive learning by promoting intra-class variability while preserving spatial and semantic consistency. Moreover, a dual-objective Angular–Contrastive Loss (ACL) supports local compactness and global separation across the normalized hypersphere. Extensive experiments on the DCASE 2022, 2024, and 2025 datasets demonstrate robust performance, outperforming baseline by 0.098, 0.075, and 0.103 AUC, respectively, while maintaining lightweight deployment. MFPPG delivers a + 0.071 AUC improvement over the log-Mel baseline with angular loss. Ablation studies confirm the effectiveness of multi-feature fusion, PPG, and ACL. Additional evaluations show strong generalization across backbones and robustness to domain shifts without fine-tuning. MFPPG offers a compact, effective, and scalable solution for real-world ASD with fewer parameters.

Recommended citation: Lu J, Guan W, Zhang M, et al. MFPPG: Multi-feature fusion and patch-wise perturbation generation for anomalous sound detection[J]. Measurement, 2025: 118915.
Download Paper