Research And Application Of KPI Anomaly Detection Based On Improved MALSTM-FCNs Model

Anomaly detection of KPI is an important part of Artificial Intelligence for IT Operations,and the improvement of its prediction accuracy is conducive to the timely detection of abnormalities in KPI,so as to ensure the stability and efficiency of application services.KPI data collected during IT operations are typical multivariate time series data.In the past,researches on KPI anomaly detection are mostly based on empirical rules and indicators dismantling to bulid models.Based on the real data generated by the server in the process of operation and maintenance,this paper studies the problem of KPI anomaly detection based on the improved time series neural network model.1.The background,research status of KPI anomaly detection and the research methods of multivariate time series anomaly detection are summarized,and data preprocessing methods and related deep learning theories are introduced.2.Based on the real data of IT operations monitoring platform,LSTM-FCNs model is built by Python,Squeeze-and-Excited block and Attention mechanisms are introduced to enhance the local feature of data to detect the anomalies of the multivariate time series KPI.The experimental results show that F1-score of LSTM-FCNs model with two mechanisms is improved from 0.9713 to 0.9742 on the train dataset,and from 0.9636 to 0.9681 on the test dataset.3.Aiming at the problem of "dead neuron" of ReLU activation function,the activation function is improved,and the performance of improved activation function is verified by using the common dataset.Compared with the exciting activation function,the improved activation function has higher accuracy and faster calculation speed,and the improved MALSTM-FCNs model is applied to the real data of IT operations monitoring platform.The experimental results show that compared with ReLU activation function,Fl-score of the improved activation function is improved from 0.9742 to 0.9780 on the train dataset,and from 0.9681 to 0.9687 on the test dataset.At the same time,two mechanisms are introduced into LSTM-FCNs model and the activation function is improved,which can improve model’s F1-score from 0.9713 to 0.9780 on the train dataset and from 0.9636 to 0.9687 on the test dataset,and reduce the calculation time of the model.