Dimensional Attention Feature Representation Learning Method For Unsupervised Anomaly Detection

Anomaly detection methods based on deep learning have achieved excellent performance with powerful feature extraction capabilities and are now widely used in applications such as data mining,network intrusion detection and medical diagnosis.Unsupervised anomaly detection methods do not rely on data label training,but rather analyse data for anomalies by learning a low-dimensional potential representation of the original features of the data,which dominates the field of anomaly detection.However,existing unsupervised anomaly detection methods usually only perform coarse-grained analysis of the raw features of the data and neglect to analyse the contribution of fine-grained dimensional information within the latent representation from different perspectives to detect anomalies.To this end,this paper proposes a dimensional attention-based unsupervised deep anomaly detection model based on a dimensional attention feature representation learning approach.In order to obtain a potential representation of the original data containing finer-grained dimensional contribution information,this paper designs a Dimensional Attention Sparse Auto Encoder(DASAE)based on dimensional attention.Since the Sparse Auto Encoder(SAE)allows the network to be trained by only some of the neurons in the hidden layer by suppressing the output of some of the neurons,it can avoid the simple learning of the original data samples by traditional autoencoders.Therefore,this paper improves SAE based on the dimensional attention mechanism,which learns the importance weights of different dimensions of the potential representation to amplify the contribution information of important dimensions and reduce the contribution information of minor dimensions,in order to obtain a high-quality potential representation with fine-grained dimensional contribution information.In order to exploit the potential representation containing fine-grained dimensional contribution information and to analyse its role in anomaly detection more rationally from different perspectives,this paper jointly designs a dimensional attention-based sparse autoencoder DASAE and Affine Transformation(AT)for unsupervised anomaly detection model(Dimensional Attention Sparse Auto Encoder and Affine Transformation for Anomaly Detection(DASAE-AT).First,DASAE is used to obtain high-quality potential representations containing fine-grained dimensional contribution information;then,a transformation network is designed to affine transform the potential representations to analyse the dimensional contribution of potential representations in different transformation subspaces;finally,a feature extractor is designed to model these transformation subspaces into multiple hypersphere subspaces to combine the results of all subspaces to analyse anomalies from different perspectives in an integrated manner,so as to further optimise the anomaly detection performance of the model.The experimental results show that,compared with the latest anomaly detection methods,this paper significantly improves the anomaly detection performance of the model by analysing potential representations of the raw data containing finer-grained dimensional contribution information from different perspectives to detect anomalies.