Research On Deep Unsupervised Abnormal Workload Sequence Detection Algorithm In Cloud Edge Fusion Environment

The anomaly detection of the cloud platform workload sequence is a key technology to realize the intelligent operation and maintenance management of the cloud platform,and is of great significance to improving the efficiency of cloud resource operation and maintenance.Manual anomaly detection has problems such as low efficiency and heavy workload.The workload sequence anomaly detection algorithm can make abnormal decisions quickly and intelligently.However,with the rapid expansion of the scale of cloud platforms and the continuous increase of workload types,how to establish an accurate and efficient anomaly detection model for complex and diverse workload sequences is an important challenge for anomaly detection.With the widespread application of the Internet of Things,abnormal workload sequence detection is deployed at the edge of the data source to obtain faster response.In the case of limited edge computing and storage resources,how to design a lightweight workload sequence anomaly detection model to become the edge Computational intelligence services need to solve the key problems.In view of this,this paper aims at the problem of workload sequence anomaly detection in the cloud-edge fusion environment.Starting from the method of improving sequence feature extraction,the schemes of fusion of workload sequence spatiotemporal features,workload feature representation enhancement and inter-layer feature transfer enhancement are adopted respectively.Mining the hidden feature representation information of the workload sequence,and finally constructing three deep unsupervised abnormal load detection models.The main work include the following three aspects:(1)Due to the diversity of workload sequence change patterns in large-scale cloud platforms,the insufficient ability of sequence feature representation,and the difficulty of accurate anomaly detection,a deep unsupervised anomaly detection method with fusion of spatial and temporal features of workload sequence is proposed.In the network architecture described by deep support vector,this method introduces convolutional recurrent network to extract spatial and temporal features of the sequence.Firstly,the convolutional neural network module is used to extract the spatial features of workload sequence;Secondly,for the acquired spatial feature vectors,the bidirectional long short-term memory module extracts temporal features;Finally,the features vectors training support vector data description classifier fuses with the deep features of the workload sequence spatial and temporal features.The proposed method extracts the features of the workload sequences from two dimensions of spatial and temporal sequence in deep,realizes the comprehensive characterization of the inherent laws of the workload sequences,increases the distinction between normal sequences and abnormal sequences,and realizes the abnormal detection of the workload sequences more accurately.The simulation dataset and Google dataset were used for verification respectively.The results show that compared with the common unsupervised anomaly detection method,the proposed method can detect different abnormal sequences more accurately.(2)In view of the complexity of workload sequence change patterns,and the difficulty in capturing the sequence change characteristics comprehensively,a deep unsupervised anomaly sequence detection method based on the enhancement of workload sequence feature representation was proposed,to enhance the richness of sample information and improve the representation ability of sequence features.Firstly,the data enhancement technology is used to transform the workload sequence to mine the complex and hard to learn sequences,so that the network can capture more information and enhance the representation ability of data features.Then,the residual attention network was used to extract rich deep features.Finally,an anomaly detection model was constructed on this basis.Samples were weighted during the training of the model to reduce the influence of abnormal samples on the solution transformation center,and the product of the probabilities of transformed samples in their respective subspaces was calculated as the anomaly fraction.The simulation dataset and Google data set were used to verify the accuracy and universality of the proposed method.The results show that,compared with the common anomaly detection methods,the proposed method can deal with the workload sequence of different changing modes and has accurate anomaly detection effect.(3)In order to solve the problem that the number of feature representation parameters in deep neural networks is large,it is difficult to deploy at the edge,a lightweight method for deep unsupervised anomaly sequence detection with enhanced feature transfer between layers is proposed.This method satisfies the accuracy requirements of the edge,from the reduced model of the two aspects of quantity and improve the running speed,workload sequence feature extraction,referenced by the ideas of the network of the dense convolution structure design features between layers connected network structure,this structure makes the extracted features more fully,to strengthen the feature information.The deep separable convolution is applied to the network structure to reduce the number of parameters.Finally,the SVDD classifier is trained with the extracted deep features,and the anomaly detection model is constructed.The simulation dataset and Google dataset were used for verification respectively.The results show that,compared with the common unsupervised anomaly detection method,the proposed method can detect abnormal sequences with different changes more accurately and ensure the accuracy of the model.Meanwhile,compared with classical lightweight networks such as Inception and Res Net,the proposed model has a better lightweight effect in terms of number of parameters and speed.