Research On Multimodal Anomaly Detection Method Based On Graph Attention Network And Temporal Convolutional Generative Adversarial Network

With the rapid development of our country’s industry,industrial production is gradually becoming more intelligent.In complex industrial scenarios,smart sensors are becoming more popular,and more and more reliable time-series data can be collected.These time-series data usually have multimodal characteristics,and each modality contains important equipment.information.Therefore,it is necessary to propose reasonable anomaly detection methods for this multimodal time series data.In factory scenarios,the complexity of multimodal time series data greatly increases the difficulty of anomaly detection.However,Existing deep learning methods face the following problems when dealing with multimodal industrial time-series data:(1)as the data dimensionality and modality increase,insufficient consideration is given to the relevant information between each modality,which is often key information required for multimodal anomaly detection;(2)existing deep learning methods rely more on labels in the dataset,while multimodal data in industrial scenarios usually do not have labels;(3)the sampling frequency of industrial big data is very high,and existing deep learning methods cannot effectively retain longer-term time-series information.This dissertation proposes a multimodal anomaly detection model called GAT-MAD,which addresses the issue of insufficient consideration of inter-modality correlations in existing deep anomaly detection methods.The GAT-MAD model utilizes graph attention networks to capture spatial dependency relationships between different modalities.Specifically,the GATMAD model includes three multi-head attention modules,with the first module capturing the correlations between all sensors,and the latter two modules capturing the correlations between different time series within the same modality and between time series of all different modalities,respectively.Additionally,the dissertation introduces time embedding vectors to better determine the inter-modality correlations for each sensor.Finally,comparative experiments are conducted on four multimodal datasets,and the results demonstrate that the GAT-MAD model outperforms other baseline models.Furthermore,ablation experiments verify the importance of each module of the GAT-MAD model.This dissertation proposes a multimodal anomaly detection model,GATCN-MAD,based on time convolutional generative adversarial network,to address the challenge of highfrequency and unlabeled multimodal data.The proposed model retains the adversarial training structure of GAN and replaces the generator and discriminator with time convolutional neural networks,which can capture longer time-series information while maintaining unsupervised training.Moreover,in the anomaly inference stage,the study introduces a joint generatordiscriminator anomaly score that detects anomalies through discrimination and reconstruction.Finally,comparative experiments are conducted on two datasets,SWa T and SMAP,demonstrating the superiority of GATCN-MAD in capturing temporal dependencies of time series.