Optimization of Multi dimensional Time Series Data Anomaly Detection Model Based on Graph Deviation Network and Convolutional Neural Network

In China’s industrial structure, industry holds a pivotal role, yet machinery anomalies hinder factory efficiency and product quality. Traditional anomaly detection methods are inefficient, costly, and challenging to pinpoint issues. With technological advancements, AI has become crucial for anomaly detection. Factories now employ numerous sensors to gather time-series data, posing challenges due to complexity, temporal sequence, large volumes, high dimensionality, and noise. High dimensionality and data volume elevate algorithm time complexity, while noise impacts classifier performance, leading to false positives/negatives. Hence, designing real-time, noise-resistant anomaly detection models is vital. This article proposes a multidimensional temporal data-based anomaly detection algorithm, using public datasets and achieving real-time performance through dimension reduction and compression. A noise-resistant GDN model, integrating one-dimensional convolution modules, enhances data smoothness and model resistance. Experimental results demonstrate effective anomaly detection and localization. Additionally, we designed an equipment anomaly detection system with data acquisition, processing, and visualization functions, applied in papermaking equipment with 78% effective anomaly detection. However, parameter K selection in the GDN model and synchronous model calling need improvement, prompting future research into heuristic K-selection and asynchronous model calling to optimize detection performance and user experience.