Research On Anomaly Detection Methods Of One-Class Classification Model

Anomaly detection,as the main link of data analysis and data processing,is an important research topic in machine learning.According to the characteristics of anomaly detection tasks,that normal data is easy to obtain,but abnormal data is difficult.One class classification algorithm is one of the main methods to deal with this kind of problem.With the development of information technology and the diversification of social relations,the data form is becoming more and more complex.And the traditional one class classification algorithm has been difficult to apply in high-dimensional scenarios due to weak computing scalability and the existence of problems such as the disaster of dimensionality in the anomaly detection task.In order to deal with the problem of high-dimensional disasters,it is often necessary to use feature engineering technology to reduce the dimensionality of the original data,and then perform anomaly analysis on the data feature vectors after dimensionality reduction.This operation belongs to two-stage anomaly detection,namely dimensionality reduction and data analysis.It is completed in two steps independently.There may be a lot of invalid information in the data after dimension reduction,and the data after dimension reduction is not the optimal feature vector that is conducive to later analysis.In recent years,many end-to-end anomaly detection methods designed by joint deep learning have been proposed.Feature learning and feature analysis are carried out simultaneously.The loss of feature analysis calculation will affect the update of network model parameters,such as deep support vector data description,deep Gaussian mixture model,etc.method.However,in terms of implementation,these methods still have insufficient data feature learning,slow model training,and low anomaly detection accuracy.Therefore,this paper conducts in-depth research around the above issues,and the main research work is as follows:(1)A new method named One-class anomaly detection with redundancy reduction and momentum mechanism is proposed.For the traditional one class classification algorithm,the support vector data description is mainly based on the assumption that the data distribution is spherical,and the learning of the data feature vector is insufficient,which usually contains a lot of redundant information.In order to deal with those problems,a redundancy reduction mechanism is proposed.This mechanism calculates the similarity matrix between each dimension of the feature vector in the pre-training stage,and optimizes the linearity and independence of different dimensions through constraint optimization.So that the limited feature vector can store more original data information,and limit each dimension to be independent in the feature space.At the same time,in view of the fact that the traditional support vector data description algorithm only calculates the corresponding centroid hyperparameter C in the early stage of the model,a momentum mechanism is proposed to jointly consider the feature center vector of every cycle.So the finally obtained centroid parameter C has a characteristic globality.(2)Another method named one class classification based on Wide Res Net and selfsupervised training is proposed.In view of the fact that many existing methods use deep networks to extract data features and obtain high accuracy in multi-classification tasks.But lack suitable objective functions for one class classification model training,a selfsupervised training strategy is added to train feature extraction networks.This method mainly processes the original target data by designing different data conversion operations,constructs the corresponding pseudo-category labels,and then trains the corresponding network to identify which conversion have been used for the input data.At the same time,this model calculates the anomaly score of each sample by calculating the similarity between each transformed sample and the center vector of the corresponding transformed subspace.The higher the similarity,the more normal the sample.Which follow the main assumption that the converted data of normal samples will be mapped to the subspace of the corresponding shape,while the abnormal samples have a large difference from the center of the corresponding subspace because the target contour is different from the normal samples.In order to verify the accuracy of the proposed method in identifying anomalies,relevant comparative experiments and strategy analysis experiments were carried out on the open image data sets.The final experimental results show that the proposed methods have improved performance.