AI-Driven Anomaly Detection for Industrial Data

Abstract

This paper proposes a structure-aware AI-driven anomaly detection method to address challenges such as anomaly scarcity, complex structural features, and blurred decision boundaries in industrial data. The method builds a multi-layer residual feature enhancement module to extract multi-scale temporal features and incorporates a structural attention mechanism to dynamically model the importance of different channels and time positions. This improves the model's ability to perceive potential abnormal regions. A dual-branch architecture is designed to capture temporal consistency and reconstruction error from the input sequence, forming a comprehensive anomaly score signal that enables unsupervised detection without explicit labels. The model adopts an end-to-end training framework and applies a sliding window mechanism to construct decision trajectories over continuous time segments, enhancing the detection of both sudden and evolving anomalies in industrial systems. A sensitivity evaluation scheme is developed across multiple dimensions, including anomaly ratio, noise perturbation, and input sequence length, to verify the proposed method's discriminative stability and structural robustness under varying operational conditions. Results show that the method performs well across multiple evaluation metrics, demonstrating strong adaptability and practical engineering value.