Anomaly Detection Algorithm Optimization And Circuit Design Based On Deep Neural Network Feature Extraction

The basic goal of anomaly detection is to detect scarce data that deviates from the overall data characteristics.In recent years,with the rapid development of Internet technology and the advent of the era of big data,anomaly detection problems have become more and more attention.Anomaly detection technology is increasingly used in the fields of network intrusion,credit fraud,medical diagnosis,etc.Unsupervised anomaly detection technology has wider application scenarios and more technical difficulties,so it is of great significance to study and optimize it.Related technologies and existing anomaly detection models in the field of anomaly detection are summarized by this thesis.There are two difficult problems in the field of anomaly detection: the problem of reduced detection efficiency when processing high-dimensional data,and the difficulty of ensuring the robustness of the detection model when the proportion of anomaly data changes.This thesis improves the structure of the deep auto-encoder,and retains its advantages of anomaly detection technology based on reconstruction errors,and uses the high reliability of the Gaussian process regression algorithm to make up for the deep auto-encoder when the proportion of anomaly data is high.A new type of unsupervised anomaly detection algorithm for high-dimensional data and high anomaly rate is proposed,thus providing a solution that can effectively alleviate the above two difficult problems at the same time.This thesis further optimizes the training methods of deep auto-encoder and Gaussian process regression,and gives a self-supervised training algorithm.The algorithm iteratively trains with the prediction results as labels,and effectively converges after 20 iterations.The simulation experiment code was written and simulated in Python language.Compared with traditional anomaly detection algorithms,the new hybrid anomaly detection algorithm has a significant improvement in detection performance,and the average AUC of the anomaly detection performance indicators on the MNIST,CIFAR-10,Fashion MNIST,and STL-10 benchmark datasets reached0.961,0.669,0.865,0.684,respectively.Especially in the case of high input data dimensions and large changes in the proportion of anomaly data,the new hybrid anomaly detection algorithm has higher detection efficiency,less impact on detection performance,and can remain relatively stable.This thesis also completed the logical design of the new anomaly detection algorithm,including the logical design of the convolution calculation module,fully connected calculation module,activation function module and exponent calculation module.Completed the simulation and synthesis of digital circuits by writing Verilog HDL code,Vivado simulation results show that the function of each module is correct.It has been verified on the PYNQ-Z2 development platform.Although the accuracy has a certain loss in the calculation,the average value of the AUC for anomaly detection on the MNIST data set still reaches 0.873,which is superior to that of the oneclass support vector machine and the deep self-encoder algorithm.