Research And Application Of Deep Learning Based Method For Time Series Anomaly Detection

With the rapid development of information technology,the scale of various types of data continues to expand,and time series data as an important data type has increasingly received extensive attention and research.When mining these time series data,there will be some data that is inconsistent with the data model or general laws.These abnormal data mostly contain important information and knowledge in different scenarios.Ignoring these abnormal data will often cause irreparable economics loss,the study of these small amounts of abnormal data is becoming more and more important,which makes time series anomaly detection an important research content in the field of data mining.In recent years,deep learning technology has also been widely studied and applied in various industries and has made good progress.Some scholars have tried to apply deep learning technology to anomaly detection of time series data,but the current research in this area is not deep enough.Based on previous research,this article researches and innovates the convolutional neural network with strong feature expression ability in deep learning,the long short-term memory network with strong memory ability,and apply the improved algorithm to time series for anomaly detection tasks,it also combines other intelligent algorithms to improve the calculation efficiency and detection accuracy,and has achieved good results.First,this paper proposes a supervised time series anomaly detection algorithm HPCSLSFCN based on a fully convolutional neural network.For labeled time series data,we built a binary classification model CSLSFCN that is suitable for it.In order to improve the feature learning ability of the algorithm,improve the relatively important features and suppress the features that are not useful for the current task,improved SENet is embedded to obtain the importance of each feature channel.At the same time,in view of the extremely imbalanced positive and negative samples in the time series anomaly detection problem,we propose a cost-sensitive loss function to improve the classification accuracy of imbalanced data sets.Finally,from the perspective of algorithm calculation efficiency,using the great parallel advantages of membrane system,the combination of CSLSFCN and membrane system is proposed based on the hybrid chain P system,which improves the efficiency of the algorithm.Secondly,this paper constructs an unsupervised time series anomaly detection algorithm IPSO-LSTM-AE based on long short-term memory network.This method combines the complementary advantages of long short-term memory networks and autoencoders,and proposes a two-stage time series anomaly detection model LSTM-AE.At the same time,in view of the problem that the parameter selection in the proposed network relies mostly on manual experience,through the combination with the particle swarm algorithm with adaptive inertia factor,the key hyperparameters in the network are automatically optimized,and the detection accuracy of the algorithm is improved and the consumption of human resources is reduced.Finally,artificial intelligence for IT operation(AIOps)is a common application scenario of time series anomaly detection algorithms.We try to apply the two types of algorithms HPCSLSFCN and IPSO-LSTM-AE proposed in this article to the key performance indicators(KPI)abnormalities detection problem.This article introduces in detail the background and significance of AIOps and KPI anomaly detection problems,and applies the proposed deep learning-based time series anomaly detection algorithm to the detection of KPI anomalies.