Research And Implementation Of Health Monitoring Framework For Distributed Systems Supporting Model Continuous Optimization

In recent years,distributed systems have been widely used in various industries for their high reliability,high performance,and high scalability.Monitoring the health of distributed systems has become a key means to ensure their stable operation.Health monitoring of distributed systems needs to address challenges such as complex health definitions and noninstantaneous characteristics of abnormal states.To address the above problems,this thesis designs and implements a distributed system health monitoring framework supporting model continuous optimization.The framework can access operational data generated by distributed systems in multiple ways,optimize the health assessment model,continuously update the standard health assessment model to the production environment,and continuously improve the overall accuracy of the framework’s health assessment.Meanwhile,the framework supports model-based anomaly detection,continuously detects system anomalies,and issues alarms when anomalies are detected,notifying operation and maintenance personnel to take timely action.This thesis proposes an anomaly detection method based on Shapelet MultiHead Attention(SMHA-AD).The method first constructs a feature space based on Time2Vec for multi-dimensional time series data heterogeneity from different nodes,maps the multi-dimensional time series data into the feature space,and generates feature vectors of the same dimension as input to the encoder-decoder based on Shapelet Multi-Head Attention mechanism.The encoder-decoder based on Shapelet Multi-Head Attention mechanism alternately stacks Shapelet Multi-Head Attention layers,original Multi-Head Attention layers,and feedforward layers to effectively mine historical features and Shapelet features of time-series data through this stacking structure.To make the trained Shapelet interpretable in time and space,a time-position encoding is added to the input to ensure the consistency of the temporal information before and after the encoding process.Through comparative experiments,the SMHA-AD method outperforms baseline methods such as CTF-AD,Omni Anomaly,ROCKA in terms of model performance.Through ablation experiments,the Shapelet Multi-Head Attention mechanism,time-space encoding,and encoder-decoder structure in the SMHA-AD method are shown to improve the model’s performance.This thesis first introduces the background,research status,and research content of the distributed system health monitoring framework that supports continuous model optimization.It then introduces the relevant technologies of health assessment models and time series anomaly detection,conducts a survey of relevant products in the industry and performs requirement analysis,and then details the SMHA-AD method proposed in this thesis.This is followed by a detailed introduction of the design and implementation of this framework and finally testing to validate the effectiveness of the system.