计算机科学
异常检测
模糊逻辑
算法
人工智能
模式识别(心理学)
模糊控制系统
数据挖掘
信号处理
模糊集
噪音(视频)
异常(物理)
模糊规则
算法设计
作者
Lei Chen,Bowen Hu,Ying Zou,Weihua Tan,Jiaao Peng,Ming Lu
标识
DOI:10.1109/tbdata.2026.3735650
摘要
Identifying anomalies from multi-sensor signals on resource-limited edge devices has become a new research focus for industrial production safety. However, most existing detection models always ignore timeliness and resource constraints, which results in inevitable trade-offs among speed, accuracy, and resource consumption. To address this issue, a fuzzy KAN-based unsupervised bidirectional fast time-series anomaly detection model, namely FuKAN, is proposed for multi-sensor signals.Firstly, inspired by fuzzy inference and the Kolmogorov-Arnold Network (KAN), a interpretable two-layer Fuzzy-KAN with few parameters is constructed.Secondly, the lightweight FuKAN model is built with only four parallelizable Fuzzy-KAN networks to ensure high speed and low resource consumption.Thirdly, a bidirectional contrastive learning mechanism is developed to capture forward and backward contextual changes of each timestamp in the fuzzy domain instead of the temporal domain, thereby enhancing detection accuracy.Finally, a customized anomaly scoring strategy is designed based on bidirectional contextual changes to generate the final score for each timestamp. To the best of our knowledge, this is the first study to explore combining fuzzy neural networks with KAN networks for time-series anomaly detection. Extensive experiments on eight datasets and three industrial devices demonstrate that FuKAN outperforms multiple state-of-the-art(SOTA) deep models in detection speed, accuracy, and resource efficiency.
科研通智能强力驱动
Strongly Powered by AbleSci AI