Collaboration Better Than Integration: A Novel Time-Frequency-Assisted Deep Feature Enhancement Mechanism for Few-Shot Transfer Learning in Anomaly Detection

Abstract

Deep transfer learning has achieved significant success in anomaly detection over the past decade, but data acquisition challenges in practical engineering hinder high-quality feature representation for few-shot learning tasks. To address this issue, a novel time-frequency-assisted deep feature enhancement (TFE) mechanism is proposed. Unlike traditional methods that integrate time-frequency analysis with deep neural networks, TFE employs a wavelet scattering transform to establish a parallel time-frequency feature space, where a dual interaction strategy facilitates collaboration between deep feature and time-frequency spaces through two operations: 1) Enhancement, where a frequency-importance-driven contrastive learning (FICL) network transfers physically-aware information from wavelet scattering features to deep features, and 2) Feedback, which uses a detection rule adaptation module to minimize bias in wavelet scattering features based on deep feature performance. TFE is applied to a domain-adversarial anomaly detection framework and, through alternating training, significantly enhances both deep feature discriminative power and few-shot anomaly detection. Theoretical analysis confirms that the proposed dual interaction strategy reduces the upper bound of classification error. Experiments on benchmark datasets and a real-world industrial dataset from a large steel factory demonstrate TFE's superior performance and highlight the importance of frequency saliency in transfer learning. Thus, collaboration is shown to outperform integration for few-shot transfer learning in anomaly detection.