The Research On Proactive Complex Event Processing For Cyber Physical System

Cyber Physical System(CPS)is a next-generation intelligent system that integrates computing,control and communication.CPS interacts with physical processes through a human-computer interface and uses cyberspace to manipulate a physical entity in a remote,real-time,reliable,secure,and collaborative way.CPS will have a great impact on the country's economic development and have increasing application fields in the future.The key issue of CPS that needs to be resolved is information processing.CPS systems may include a large variety of interconnected devices that continually generate massive amounts of raw events.Due to the existence of noise in the system,sensor error,network communication technology and other reasons,the uncertainty of event will occur.Meanwhile,CPS applications are mostly deployed in a distributed way.The events from different data sources are characterized by massiveness,heterogeneity and uncertainty.Current dataflow oriented processing platforms with functions such as filtering,transforming and counting have difficulty to satisfy the need of real-time data analysis and providing feedback effectively.As one of the core technologies of CPS,proactive complex event processing technology is capable of processing the raw data effectively and providing decision in real time.There are many problems in proactive complex event processing,including uncertainty event processing,context of event processing,predictive analysis and decision making.Taking the large-scale CPS intelligent transportation system as the background,this paper integrates classical algorithms such as Markov chain,fuzzy C-mean,Bayesian network,Markov decision process and Q-learning,and conducts studies of other parallel optimization techniques.The contributions of this paper mainly include the following aspects:(1)In the practical application of CPS,it is very common for data to be uncertain.Therefore,it is necessary to deal with uncertain events.This paper presents a Parallel Indeterminate Stream Complex Event Processing Method(PISCEP)for uncertain event flows.Based on the calculation of event probability,PISCEP processes a large number of CPS event flows concurrently by using matching tree and cached list,and solves the problem that the intermediate result sets of complex events can not be shared.The experiment results show that PISCEP has better performance.Compared with the classical RCEDA method,the average improvement rate of the complex events generated number by PISCEP is 25.65%,and the processing time of PISCEP is shortened by 17.66% compared with the SASE method.(2)Event context is general in CPS data stream.Therefore,contexts processing is also of great value in complex events processing for CPS.There are a great variety of contexts of events,which have significance value for separate processing after they are clustered.The traditional fuzzy C-means algorithm(FCM)requires that the cluster center to be determined first,and how to select the cluster center is the essential step of FCM.In order to solve this problem,this paper presents an Event Context Processing Method(ECPM).ECPM uses fuzzy ontology to model the context of events,and then cluster event context by improved FCM to provide the prediction models for learning and predicting.Compared with the classical FCM,K-means and SOM,the accuracy of ECPM is improved by 2.82%,6.34% and 9.86% respectively.(3)A fixed model cannot always have good predictive performance in different environments.In order to be more self-adapting,this paper presents a Context-aware Prediction Method for Complex Event(CPMCE).Based on event context clustering,CPMCE could learn corresponding Bayesian network models and provide appropriate Bayesian network model or combination of Bayesian network models for real-time prediction and analysis according to different context.In CPMCE,a hybrid enhanced learning method is used for Bayesian structure modeling,and the approximate inference is achieved through a Gaussian mixture model.The experiment results of real data show that the average accuracy of CPMCE is 6.06% and 4.08% higher than Bayesian network and deep belief network.(4)There is little literature on how to integrate MDP into proactive complex event processing and how to deal with the huge number of states that MDP generated in the complex event processing.Based on state partition of nodes,this pa per proposes Parallel MDP(PMDP)to support proactive complex event decisions.According to the results of predictive analysis and the node state partition,along with reward splitting method,PMDP achieves parallel optimization.A Parallel Q-Learning method(PQ)is proposed for improving performance in the situation with much larger data.Experiments in the intelligent traffic of CPS show that both PMDP and PQ methods can effectively reduce the congestion rate of nodes,and the performance of PQ-learning is better than PMDP when facing large-scale data.