Research On Anomaly Detection Algorithm Based On Boosting Adversarial Learning And Variational Autoencoder

With the advancement of modernization,the construction of smart factories is receiving increasing attention.In order to improve the intelligence of factory software,the industrial system must have the ability of automatic data analysis.One of the most important data analysis tasks is anomaly detection.Deep learning has made a breakthrough in the field of anomaly detection,but there are still several problems in the study of multivariate anomaly detection in factory environment.On the one hand,small and medium-sized factories deploy lightweight models due to computational problems,while the existing models cannot take into account the accuracy and speed.Secondly,the location of small and medium-sized factories is flexible,and the pipeline often changes.However,the existing models lack adaptability.On the other hand,complex neural networks deployed in large factories cannot capture the temporal features and the interdimensional relationship features well at the same time,so the effect of anomaly detection needs to be improved.Secondly,the model lacks interpretability,which cannot help engineers quickly locate the specific device causing the exception.To address the issue that the existing models cannot take into account both detection accuracy and speed,this paper proposes a lightweight anomaly detection algorithm DBNBAAE based on enhanced adversarial autoencoder.In this paper,an improved deep belief network is proposed for pre-training to improve the stability of autoencoder training.Secondly,ensemble learning was proposed to improve the representation ability of the encoder.Then,the decoder reconstruction error was amplified based on adversarial learning.Finally,a dynamic threshold was proposed to make the model adaptive to alleviate the computational overhead of retraining.Experiments show that the F1 score of the model reaches 0.82,and the training speed is 2.2 times faster than the comparison object with the best performance.To address the issue that the the existing models cannot capture two types of temporal features well at the same time and cannot perform anomaly localization,this paper proposes an interpretable anomaly detection algorithm MCA-VAE based on variational autoencoder.In this paper,multi-scale weight shared convolutional neural network is proposed to extract time sequence dependent features.Secondly,the multi-head attention mechanism was proposed to extract the correlation features between dimensions.Then,an improved variational autoencoder was proposed to calculate the data reconstruction error.Finally,anomaly interpretation based on reconstruction probability was proposed.Experiments show that the F1 score of the model is 4% higher than that of the best performing comparison object,and it has good interpretability.Finally,combining the characteristics of the above two algorithms,an industrial control anomaly detection system for multi-scene is designed.The system provides efficient anomaly detection based on deep learning algorithm,and provides anomaly visualization,real-time data display,model management and other functions.The test proves that the system has better performance and detection effect.