CSA-LSTF: An Improved Method for Time Series Data Anomaly Detection

With the development of industrial Internet of Things (IoT), a large number of sensors have been deployed in practical environments to collect data. These data possess characteristics such as large volume, high dimensionality, strong time series nature, and high real-time requirements, making anomaly detection difficult. Traditional anomaly detection mainly focuses on single data streams, which leads to high model maintenance costs and an inability to determine dependencies be-tween multidimensional time series data, thus failing to meet the practical requirements of current industrial IoT scenarios. This paper proposes an improved algorithm CSA-LSTF(Convolutional Sparse Self Attention-Long Sequence Time-Series Forecasting) based on LSTF, which utilizes convolutional sparse self-attention mechanism to enhance LSTF and applies this algorithm to the framework of Generative Adversarial Network (GAN) for training. Anomalies are detected through discrimination results and reconstruction errors. Experimental results demonstrate that the CSA-LSTF method can timely detect anomalies and take preventive measures, enhancing the risk resistance capability of industrial IoT systems. Compared with existing anomaly detection models, it enhances model stability, improves anomaly detection accuracy, and surpasses existing multidimensional un-supervised time series data anomaly detection models.