| Today’s society highly relies on information technology(IT)infrastructure.In order to meet the technical needs of various enterprises and organizations for conducting business,IT systems have become larger and more complex.This trend makes traditional manual maintenance and management methods no longer able to meet the needs of enterprises for efficient,intelligent,secure,and reliable IT services.More and more enterprises are exploring how to use automation and intelligent technology to improve maintenance and management efficiency.In this context,Artificial Intelligence for IT Operations(AIOps)emerged as an emerging technology for operations and management.AIOps aims to study the application of artificial intelligence(AI)in IT service automation management,helping operation and maintenance teams improve the quality and reliability of IT services by using intelligent algorithms and monitoring the large amount of data provided by infrastructure.AIOps relies on data-driven technologies such as machine learning,big data,data mining,data analysis,visualization,and human-machine interaction to observe the operational status of infrastructure,minimize the impact of daily failures,and proactively manage the allocation of computer resources.This thesis delves into two key scenario technologies in AIOps,namely fault correlation and prediction,and root cause analysis of anomalies:(1)For hardware alarms and hardware indicator data in IT operation and maintenance,this thesis delves into the correlation analysis method between alarm logs and hardware indicator data,and designs a fault prediction model based on correlation analysis to help operation and maintenance personnel deeply understand the correlation between alarms and time series indicators,and be able to grasp fault information in advance.IT operation and maintenance systems typically collect a large amount of time series indicator data and alarm log files to monitor and manage the operational status of the system.However,on the one hand,due to the large amount of data,the distribution types of temporal data are diverse,and it is difficult to obtain anomaly labels related to temporal data;On the other hand,due to different collection methods,the correlation between timing indicator data and alarm logs is generally unknown,which will bring great trouble to system operation and maintenance.Most traditional fault prediction models have the attribute of "black box",and their interpretability is poor.There is also no in-depth analysis of the relationship between alarms and hardware equipment indicators,which makes the predicted results lack credibility.In response to the above issues,this thesis designs a fault prediction model based on correlation analysis.This model first uses an automatic anomaly detection algorithm selection framework to detect anomalies in multi distribution type temporal indicator data.Then,based on the correlation analysis method in correlation analysis,the correlation between alarm logs and hardware indicators is established.The XGBoost model is used to classify hardware indicator anomalies and calculate the indicator threshold for alarm occurrence.Finally,time series prediction methods such as ARIMA are used,Cooperate with the indicator threshold to achieve the effect of fault prediction.(2)Furthermore,based on the KPI data in IT operations and maintenance,this thesis delves into the root cause analysis algorithm for anomalies after faults or anomalies occur,helping operations and maintenance personnel grasp the causes of anomalies at a finer granularity.With the popularization of the internet,the number of users of internet products is increasing rapidly,and the number of KPI data is exploding.The attribute dimensions of the data are also increasing,making it difficult to attribute changes to KPIs solely based on traditional decision trees constructed by operation and maintenance personnel.Therefore,researchers have turned to studying automatic root cause analysis algorithms,and many root cause analysis algorithms have been proposed one after another.However,most of these root cause analysis algorithms have strong limitations due to their strong assumptions about root causes,and their effectiveness is often poor when the data does not meet their root cause assumptions.This thesis proposes a root cause analysis algorithm based on co directional proportion,which can capture the consistency of fine-grained data and its root cause combination change trend,thus obtaining effective root cause results.In order to verify the effectiveness of the fault prediction model and root cause analysis algorithm in this thesis,this thesis conducted model testing on the fault prediction model based on real datasets,and compared the root cause analysis algorithm with typical root cause analysis algorithms on two real datasets.The test and experimental results show that the fault prediction model proposed in this thesis has good model interpretability and good correlation analysis and prediction performance for hardware alarms;Compared with typical root cause analysis algorithms,this root cause analysis algorithm has stronger robustness to different data due to its weaker data assumptions,and can more accurately discover root causes. |
