Answers are grounded in this paper and cite their sources.
Reconstruction-based autoencoder models have been proven effective and feasible for unsupervised industrial anomaly detection. However, they still face challenges such as the lack of supervision information, impure training data, uncontrolled generalization capabilities, and incomplete anomaly evaluation criteria. In this paper, we propose a knowledge-informed density-aware reconstruction framework as a potential solution. We utilize industrial domain knowledge to obtain a pure training dataset and a self-annotated validation dataset that serves as a supervisor. An effective long short-term memory autoencoder and a dual local outlier factor module are designed for reconstruction of temporally correlated data and density-aware information fusion. Our framework considers both the individual shape patterns and distribution characteristics of industrial data, offering high anomaly detection precision and adaptability to enhance the performance of various autoencoder models. Comparative analysis results on a real-world lithium-ion power battery sorting application demonstrate the validity and advantages of our approach, which serves as a useful solution for identifying unqualified battery cells as well as improving the consistency of battery packs.
No takes yet. Share an insight, caveat, or question.
Bai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: