Research And System Implementation Of Enhanced Adversarial Autoencoder Anomaly Detection Method

In the field of industrial operation and maintenance,time series anomaly detection has important research significance and application value.With the rapid growth of operation and maintenance data,it is not practical to label large-scale time series data with label information,so traditional supervised time series anomaly detection methods are no longer widely applicable to industrial scenarios.Therefore,this thesis focuses on the deep learning-based unsupervised time series anomaly detection model.The current mainstream deep anomaly detection models mainly reconstruct the normal sample data based on the autoencoder structure,and a new sample is judged to be abnormal if its reconstruction error is large.However,the above-mentioned models have the following problems:(1)If the abnormal sample data is close to the normal data,the reconstruction error of the abnormal sample will be insignificant,and abnormal misjudgment will easily occur,making the abnormal detection effect poor;(2)The complexity of the deep learning network The detection efficiency is low in the face of large-scale time series anomaly detection.Therefore,this thesis conducts the following research:Firstly,in order to solve the problem of poor detection effect caused by the insignificant reconstruction error of the autoencoder,this thesis proposes a temporal anomaly detection model(EAAE)based on the adversarial augmented autoencoder,which includes an adversarial generation module and a latent space adaptive learning module.part.By adding augmented adversarial training mechanisms for input samples and latent vectors for better sample reconstruction.The encoder in the adversarial generation module encodes the input sample data to generate a hidden vector.The hidden vector obtains the input vector of the decoder through the latent space adaptive learning module,and the hidden vector is decoded by the decoder to generate a reconstructed sample.In order to verify the effectiveness of this model,experiments are carried out on three public datasets of SMD,SMAP and MSL.The experimental results show that the EAAE model has good performance in time series anomaly detection.Secondly,in order to improve the efficiency of the EAAE model in large-scale time series anomaly detection,it is necessary to compress the EAAE model.In this thesis,knowledge distillation is used for model compression,and a lightweight network that is isomorphic to EAAE is designed,so that the lightweight network can learn the probability distribution of EAAE output,and realize knowledge transfer and compression.Experiments on the above three public datasets prove that the model after knowledge distillation not only has high classification performance,but also has a small amount of parameters,computation and time overhead,which promotes the application of deep networks in anomaly detection in the industry.deployment work.Finally,this thesis implements a large-scale anomaly detection system framework.The system framework includes three modules: data input simulation module,anomaly detection module,and storage output visualization.The framework has two advantages:on the one hand,the system implementation adopts distributed technology to build the framework,which has high system expansion performance;on the other hand,the system framework and algorithm use the same code system,which has high code implementation versatility.In the end,the system can detect the data volume of 80,000 time series per minute.The results prove the effectiveness of the above algorithm research,and also provide a reference idea for the practical application of industrial scenarios.