Algorithm Research And FPGA Verification Of Anomaly Detection Based On Deep Generative Models

Anomaly detection technology is playing an increasingly important role in social security and stability as well as people’s healthy life.Due to the higher difficulty to obtain abnormal data compared to normal data in real life and the great quantities of unknown anomalies as well as the high dimension of the data,the traditional anomaly detection method based on discriminative models often has a poor performance and it is necessary to study more effective algorithm.To solve the above problems,deep generative models are studied in this thesis,which could model on normal data to study their pattern distribution and detect anomalies according to the test sample’s deviation from normal pattern.Based on the analysis of the advantages and disadvantages of autoencoders and generative adversarial networks,this thesis improves the structure of the autoencoder by adding a discriminator to enhance the autoencoder’s ability to model the normal data.By making the latent vectors fit the standard normal distribution,the decoder could utilize the vectors sampled from it to generate new data and improve its reconstruction ability of normal data.This thesis also improves the anomaly score by using the reconstruction error of features instead the pixels,so as to alleviate the problem that the latter one tends to miss inspection of anomalies with simple patterns.The improved deep generative model reaches an average AUC value of 0.871 on the benchmark dataset,which is better than other anomaly detection methods.The implementations of the nonlinear activation functions and the convolutional layers are analyzed in detail and a simulation experiment is performed on Xilinx’s Vivado HLS to verify their functions.The verification result on the PYNQ-Z2 board shows that it still obtains an average AUC value of 0.851,which is still better than other methods.