以機器學習演算法預測銀行虛擬資源使用模式

Abstract

隨著數位科技的進步與發展，產業開始應用大型且複雜系統，監控這些大型資源和突發事件會產生大量數據，再加上大數據分析技術普及，智能維運的概念隨之興起，效率管理、異常偵測與預測以及異常自動化處理為其的主要功能，能夠幫助企業降低維運成本、提升客戶體驗，因此導入智能維運技術為系統維運部門非常關切的議題，故本研究蒐集銀行虛擬資源監控資料，聚焦於資源使用率和異常預測。 過去的研究中，針對虛擬資源使用率和異常預測，主要透過蒐集監控資料，以統計模型和機器學習模型進行預測，而本研究則實驗適合處理時間序列資料的深度學習模型：卷積神經網路、長短期記憶模型以及將兩者結合的演算法，同時實驗不同特徵之下模型的預測效果，並與過去的研究方法進行比較，改善未來針對資源使用率以及異常預測模型。根據實驗結果，本研究採用長短期記憶模型結合卷積神經網路的演算法無論在資源使用率或異常預測上，相較於過去研究使用的機器學習演算法，都有更突出的表現。 透過本研究成果，希望能幫助銀行更有效率的管理內部資源，更能提前預測系統異常，降低服務中斷的機率，減少內部系統維運的人工成本、同時提升管理效率以及系統穩定度。

Owing to new technological advances, most of the industries have implemented some large-scale and complicated systems as their infrastructure. Because of the improvement of Big Data Analysis and the numerous data produced by monitoring these systems, the concept of Artificial Intelligence for IT Operations (AIOps) was born. The main purposes of AIOps are to efficiently manage resources, and predict abnormal of the system and automatically deal with emergencies. As a result, introducing AIOps into IT infrastructure is an urgent demand for many companies. In our work, we collected monitoring data of bank’s IaaS platform and focused on the methods to implement AIOps in the bank. In the past, many research implemented statistical models and machine learning models to predict virtual resource utilization and abnormal situation. In our work, we applied deep learning models, Convolutional Neural Networks(CNN), Long Short-Term Memory(LSTM) and the combination of these two models, and experiment different length of features to improve the performance of the prediction. According to the result of the experiments, the combination of LSTM and CNN is the most effective model to predict utilization and irregular situation among the algorithm used in the previous research. Our research has favorable results which are able to predict resource usage and unusual patterns. By achieving these goals, our work could help the bank to manage the resource more efficiently, reduce the possibility of the interruption of the services as well as the cost of the maintenance, and also ensure the stability of the systems.