Anomaly Detection of Industrial Data Based on Multivariate Multi Scale Analysis

Abstract

Anomaly detection stands as a crucial facet within the domain of data quality assurance. Notably, significant strides have been made within the realm of existing anomaly detection algorithms, encompassing notable techniques such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and anomaly detection models founded upon Generative Adversarial Networks (GANs). However, a notable gap lies in the inadequate consideration of interdependencies and correlations inherent in multidimensional time-series data. This becomes particularly pronounced within the context of industrial evolution, where industrial data burgeons in complexity. To address this lacuna, a novel hybrid model has been introduced, synergizing the capabilities of GRU with structural learning methodologies and graph neural networks. The model capitalizes on graph structural learning to unearth dependencies linking data points across distinct spatial dimensions. Concurrently, GRU extracts temporal correlations embedded within data along a single dimension. Through the incorporation of graph attention networks, the model employs a dual-faceted correlation perspective for data prediction. Discrepancies between predicted values and ground truth are utilized to gauge errors. The amalgamation of predictive and scoring mechanisms enhances the model’s versatility. Empirical validation on two authentic sensor datasets unequivocally demonstrates the superior efficacy of this approach in anomaly detection compared to alternative methodologies. A notable augmentation is observed particularly in the recall rate, underscoring the method’s potency in identifying anomalies.