A Service-based Approach to Situational Correlation and Analyses of Stream Sensor Data

Transcription

1 A Service-based Approach to Situational Correlation and Analyses of Stream Sensor Data Zhongmei Zhang 1,2,3, Xiaohong Li 1, Chen Liu 2,3, Shen Su 2,3, Yanbo Han 1,2,3 1 School of Compute Science and Technology of Tianjin University, Tianjin, China 2 Beijing Key Laboratory on Integration and Analysis of Large-scale Stream Data, North China University of Technology, Beijing China 3 Cloud Computing Research Center, North China University of Technology, Beijing, China gloria_z@126.com, xiaohongli@tju.edu.cn, liuchen@ncut.edu.cn, sushen@ncut.edu.cn, hanyanbo@ncut.edu.cn Abstract IoT service and service composition provide an effective means to develop IoT applications based on correlating multiple sensor data. The change of specific sensor data can cause others changes under uncertain situations. It makes difficult for defining service composition plan in advance to build IoT application. This paper proposes a datadriven service composition method based on our previous proactive data service model. We regard service events frequently happen together with given service event as its situation, and the service events happen next as reacted actions under the situation. We analyze two kinds of correlation among service events via an improved FP-tree algorithm, and realize the service composition at runtime based on the realtime service events. Based on the real sensor data set in a coalfired power plant, a series of experiments demonstrate that our method can effectively detect new service events based on current service events. Keywords- IoT service; service composition; sensor data; data-driven I. INTRODUCTION Today, a growing number of sensors are deployed in physical world with the development of IoT (Internet of Things) applications. IoT service plays a more and more important role for sensor data sharing and analyzing for IoT applications in field of intelligent transportation [1, 2], health-care [3], smart home [4] and equipment monitoring [5]. Following the thought of software-defined sensors, we proposed a novel type of service abstraction, called as proactive data service in our previous works [1, 6]. We can map one or several sensors into proactive data services in the software-level. With proactive data service, the sensor events generated directly by sensor can be transformed into multiple service events which are more meaningful information. The sensors deployed in physical world are not independent from each other. For obtaining more valuable information, many IoT applications demand to correlate multiple sensor data. Service composition based on proactive data service provides an effective means to build valueadded applications via the evolution of service events. When a service event is generated, we can use service composition to generate new service events triggered by it, and eventually generate the required events step by step. The generation of new service events is situational, since the correlations among sensors are quite complicated. Recently, many studies proposed their situation-aware service composition methods to help construct the service composition at runtime. Models for context-awareness service composition [7, 8] or situationawareness service composition [9] have been developed to support the modeling and specification of context/situation requirements. Some works support declarative specification for situation constrains and the requirements for composition [10, 11]. However, the situation of service cannot be modeled statically and specific in advance. As the mapping of sensors in software-level, proactive data services also have complicated relations. Firstly, a situation of one service event can be formed by different service event generated by other services, and keeps changing with the changing of sensor data. For example, an import air pressure under 65 kpa alone may cause a bearing vibration over 0.118mm/s, while an import air pressure under 65 kpa with a valve degree under 30% will cause an electricity under 80A. Secondly, the service events triggered by current service event are also changeable, which cannot be predefined statically. For example, an import air pressure under 65 kpa can cause a bearing vibration over 0.118mm/s, and can also cause a primary air pressure under 62 kpa. In this case, one service should compose with different services in different situations for obtaining new service events. Therefore, we focus on analyzing the correlations of sensor data as a key perspective to compose proactive data services under uncertain situation. In this paper, we aim to provide a data-driven service composition method based on the analyzing of two kinds of correlations of service events. On one hand, we regard service events which frequently happen together with a given service event as its situations, and propose to analyze service events frequently cooccurrence to obtain the situations for given service event. On the other hand, we regard service events happen after specific situation as its reacted actions, and propose to analyze successive service events to obtain potential service events triggered by specific situation. The service composition method can keep composing services and generating new service events proactively based on current service events and its real-time situations.

2 In summary, the contributions of this paper are as follows: (1) We propose a data-driven method to compose situational services oriented to IoT applications. We obtain real-time situations of given service event through the analyses of co-occurrence service events, advise service composition based on the correlations of successive service events. (2) We realize the data-driven service composition method with two phases. We firstly utilize a modified FP-tree algorithm to mine frequently cooccurrence service event sets with their subsequent events. Then we realize the service composition at runtime based on service events generated in realtime. (3) We verify our method based on a real sensor data set from a coal-fired power plant. Through a series of experiments, we verify the effectiveness of our method. The remainder of the paper is organized as follows. In section II, we describe a motivation scenario and define the problem. In section III, we introduce our service composition method based on the analyses of sensor data. In section IV, we present our evaluations. In section V, we introduce the related work. Finally, in section VI, we draw conclusions and discuss future work. II. MOTIVATION SCENARIO The motivation example of this paper is a real-time anomaly detection application for coal-fired power plant. There are over ten thousands of sensors deployed on hundreds of equipment in a coal-fired power plant. These sensors can reflect running indicators for equipment and affect the changes of each other, such as the increase of temperature will cause an increase of pressure. The affection can be further transmitted from sensors to sensors and eventually expose specific anomaly. The anomaly detection application expects to alert relevant staff in real-time once the sensors expose any anomaly. Figure 1 shows some potential anomalies in specific equipment, i.e. the Primary Air Fan (PAF) and the Coal Mill (CM). In practice, anomaly detection in coal-fired power plant generally relies on specifying threshold for each sensor or conditions for multiple sensors. For example, specify the maximum of valve degree as 80%, or specify sensors air pressure, bearing vibration and motor electricity in PAF with conditions when the air pressure is under 88kPa, and bearing vibration is over 0.125mm/s, and the motor electricity is less than 65A, trigger a fan stall anomaly. However, as showed in the upper part of Figure 1, the correlations among sensors are much more complicated in practice. For one sensor, one change of its sensor data can cause the change of different sensor data under different situations, and eventually expose different anomalies. Take air pressure under 88 kpa as an example, it can cause a motor electricity under 70A under the situation of valve degree over 60%, and eventually expose a fan stall anomaly. And it can also cause a primary volume under 1.5t in different equipment under the situation of valve degree over 60% and import air volume under 1.8t, and eventually expose a coal block anomaly in CM. The situation of certain sensor is quiet complex, because it can be formed by different sensor data generated by different sensors. What s more, the situation of certain sensor and the sensors affected by it are both changeable since the sensor data is keep changing. pressure under 88 kpa valve degree import air pressure motor temperature bearing vibration motor electricity export air pressure fan stall fan stall fan surge import air volume primary air volume_cm import coal volume_cm export pressure_cm coal block 08:01:25 30% 08:02:45 30% valve degree control valve air inlet 08:04:03 60% 08:02: kpa air outlet 08:04: kpa air pressure 08:05: kpa primary air inlet coal inlet pulverized coal outlet fore bearing rear bearing primary air fan electric motor 08:02:25 88 A coal mill 08:04:45 08:06:03 86 A 80 A electricity_c Figure 1. Partial Anomaly Detection in Partial Equipment of Coal-fired Power Plant.

3 Service composition is an effective means to expose the affection of sensors and detect anomalies. Based on our proactive data service model, we can create one proactive data service for each indicator. For example, an air pressure service can be created with sensor data generate by air pressure sensor. To obtain new service events triggered by current service event, the service composition method needs to detect the situation of current service event and compose relative services to generate new service events at runtime. Based on the above analyses, the situation of certain service events and its reacted service events keep changing dynamically, it is practically impossible to predefine the situations and the actions caused by them. Hence, the analysis of sensor data is required to advise the detection of real-time situations and the composition next at runtime. III. DATA-DRIVEN METHOD FOR SERVICE COMPOSITION A. Overview of Service Composition Method We propose a data-driven service composition method based on the analyses of sensor data. We regard service events which frequently happen together with specific event as its situations and service events happen next as the reacted actions. The real-time situation and the correlations between situations and services are the major bases for our method. Figure 2 shows the process of our service composition method. Each proactive service is mapped by certain set of sensor though receiving sensor event streams as inputs. There are two kinds of correlations of service events that can advise the service composition when any new service event is generated by one service. For any current service event, the correlations among co-occurrence service events can advise the finding of situations, and then if any situation is satisfied, the correlations among successive service events can indicate the services to compose and the new service events to generate. The sensor data-driven service composition method includes two major steps: 1) generate the two kinds of correlations among service events to obtain the situations for given service event and reacted service events for satisfied situations, and 2) composite services based on the real-time service event and its situations. In this paper, we obtain correlations among service events based on the periodical analysis and mining of history sensor data. We discretize each sensor data in to different classes, and the values in the same class are regard as belonging to the same service event. We adopt FP-tree algorithm [12], a common algorithm for frequent items mining, to generate the frequent cooccurrence service event sets with their followed service events. Different from traditional FP-tree algorithm, in this paper the transactions contain two parts, i.e. the situation and reacted service event. We modify the FP-tree algorithm and realize it in parallel. B. Core Definitions In our previous works [6], we introduced event mechanism in service model and gave some basic definitions. We classified event in our model as sensor event and service event. Service event generated by service is transformation of sensor event. In this paper, we use the correlation of service events to advise the service composition at runtime. Follows give the definition of two kind correlations of service events and introduce their utility in our method. Definition 1. (Co-occurrence Service Event Correlation): co-occurrence service event correlation is the correlation among service events E si happen in the same moment. We use frequency of co-occurrence to measure the correlation of service events. The co-occurrence frequency of a set of service event can be calculated as follow: in which, m is the co-occurrence times of the service events, and n is the total moments. For a given service event, we regard the set of service events as its situation if, in which is their cooccurrence frequency and is a given threshold. Figure 3 shows an example. Suppose there are four services {S 1, S 2, S 3, S 4 },, and the total number of moments as 10. For service event E 11 generated by service S 1, its situation can be formed by {E 22, E 42 }. time service t 1 t 2 t 3 t i t i+1 t k t k+1 t m t m+1 S1 E11 E11 E11 E11 E11 E13 E12 E11 E12 S2 E21 E24 S3 E31 E31 E33 E32 E33 E33 E32 E34 E33 S4 E43 Figure 3. Example for Service Event Correlations. Definition 2. (Successive Service Event Correlation): successive service event correlation is the correlation between service event set E si(t) and E si(t+1) happen in two successive moments, in which E si(t) and E si(t+1) satisfy ( ) ( ). Figure 2. Overview of Data-driven Service Composition Method.

4 We use the confidence ratio to measure the correlation of successive service events. The confidence ratio of service events happen next E si(t+1) can be calculated as follow: In which, k is the happen times of E si(t+1) together with E si(t), and m is the total happen times of E si(t). For a specific situation E si(t) of given service event E s, we regard E si(t) is the reacted actions if, in which is their confidence ratio and is a given threshold. Based on the mapping of sensors and proactive data services, E si(t) can indicate the services to compose under real-time situation of given service event. In the example shown in Figure 3, suppose that. For E 11, service events {E 33 } are the reacted action under situation {E 22, E 42 }. C. Generation of Correlations among Service Events In this paper, we utilize an FP-tree algorithm to generate the two kind correlations among service events according to the above definitions. There are mainly three procedures, i.e. the transformation of transactions, the pre-processing of transactions, and the generation of correlations between frequent service event sets and their subsequent service events. 1) Transformation of transactions We firstly transform the history sensor data into transactions by regarding values of history sensor data in two successive moments as one transaction. Table 1 shows some transaction instances. The former part of the transaction is used to generate the co-occurrence service event correlation, and the latter part is used to generate the successive service event correlation. tid TABLE I. EXAMPLES OF TRANSACTIONS transaction {valve degree = 10, air pressure = 88.3, vibration = 0.125, time = 08:00} {valve degree = 10, air pressure = 87.5, vibration = 0.124, time = 08:03} {valve degree = 10, air pressure = 87.5, vibration = 0.125, time = 08:03} {valve degree = 10, air pressure = 86.5, vibration = 0.123, time = 08:06} {valve degree = 10, air pressure = 86.8, vibration = 0.123, time = 08:06} {valve degree = 10, air pressure = 84.5, vibration = 0.121, time = 08:09} {valve degree = 10, air pressure = 84.5, vibration = 0.121, time = 08:09} {valve degree = 10, air pressure = 82.5, vibration = 0.112, time = 08:12} It is observed that the value of sensor data is very discrete. It is difficult to found the laws between sensor data and mining corresponding correlations. Hence, a pre-process is needed to classify discrete value into meaningful data range. We adopt a K-means algorithm to decentralize sensor data into 4-8 data ranges referred to experts advices. Table II shows some examples of sensor data division. The data ranges can be represented by sensor events, and values belong to the same data range can be represented by the same service events. Table III shows the transactions examples after the discretization. TABLE II. EXAMPLES OF SENSOR DATA DIVISION sensor cid data range unit d1 (0 30] valve d2 (30 60] degree d3 (60 100) % p1 (0 69] p2 (69 85] air p3 ( ] pressure p4 ( ] kpa p5 (107 + ) v1 ( ] vibration v2 ( ] v3 ( ] mm/s v4 ( ) TABLE III. tid EXAMPLES OF TRANSACTIONS AFTER DISCRETIZATION transaction {valve degree d1, air pressure p3, vibration v3, time = 08:00} {valve degree d1, air pressure p3, vibration v3, time = 08:03} {valve degree d1, air pressure p3, vibration v3, time = 08:03} {valve degree d1, air pressure p3, vibration v3, time = 08:06} {valve degree d1, air pressure p3, vibration v3, time = 08:06} {valve degree d1, air pressure p2, vibration v3, time = 08:09} {valve degree d1, air pressure p2, vibration v3, time = 08:09} {valve degree d1, air pressure p2, vibration v2, time = 08:12} As showed in table III, a transaction can be transformed as: ( ) ( ) In which, ( ) is the service events generated in moment t, and ( )is the service events generated in the next moment (t+1). 2) Pre-processing of transactions After the obtaining of the transactions, we can adopt the FP-tree algorithm to generate the frequently co-occurrence service event sets and their subsequent service events. However, there are two problems when we directly utilize the FP-tree algorithm. Firstly, as showed in table III, ( )and ( ) in one transaction may be the same or contain the same service events. With this kind of transaction, meaningless successive service event correlations will be generated, because the reacted service events and the frequent service event set will contain the same events. And there is no need to trigger corresponding services to generate these events because they are already being generated. Secondly, if both parts of transaction contain service events transformed by all the sensor data, the frequent service event set and its reacted service events may contain events generated by uncorrelated sensors. It will eventually affect the effectiveness of the correlations among service events. What s more, if the event s number in each transaction is double of the number of sensor data, which will increase the computational complexity with plentiful sensors. Hence, we need to preprocess the transactions before the mining of correlations. We firstly delete the redundant transactions by removing the transactions which have the same service events, and delete the repetitive events from the

5 latter events. Then we divide the each transaction into different partitions based on the relativity of sensors, and generate corresponding association rules for each partition. In this paper, we also utilize a K-means to divide the sensors into different groups, and the measurement to clustering is refer to Pearson Correlation Coefficient (PCC) [12], which is common method to measure the relativity of two variables. With the pre-processing of the transactions, we will obtain more effective correlations. 3) Generation of Correlations Then we adopt FP-tree algorithm to mine the frequent subsets of service event ( ) and the correlations between service event ( )and service event ( ). According to the Definition 1 and Definition 2, we set the minimal frequency degree for the former events in transactions for generating co-occurrence service event correlations, and set the minimal confidence ratio and for the latter part of transactions for generating successive service event correlations. We firstly compute the frequency degree for each event in the former part of all transactions and delete events whose frequency are smaller than the minimal frequency degree. Then we scan all the transactions to construct a FP-tree, in which each transaction can add to the FP-tree by add a new path or add count for existing node. And lastly, for each path in the FP-tree, we calculate the confidence ratio of the service events in the latter part and delete the service events whose confidence ratios are smaller than the minimal confidence ratio. Each path in the FP-tree can represent the successive service event correlations, and the former part of the path can represent the co-occurrence service event correlations. Algorithm 1 shows the preudocode of our algorithm. Algorithm 1: FPTreeBuilding Input: Output: T={ t 1, t 2,, t n },the transaction set;, the minimal frequency degree;, the minimal confidence ratio FP-tree, the FP-tree contains all the correlations; 1. FES= Ø; 2. root=new Node(); 3. FP-tree.root=root; 4. for each e ij in each t i in T 5. if FES.contains(e ij) 6. e ij.count++; 7. else if e ij belong to t i.former 8. FES.former.add(e ij.); 9. e ij.count=1; 10. else if e ij belong to t i.latter 11. FES.latter.add(e j.); 12. e ij.count=1; 13. end if 14. end for 15. for each e i in FES.former 16. fre= e i.count/t.n; 17. if e i. fre < 18. FES.delete(e i); 19. end if 20. end for 21. sort FES.former & FES.latter in desc order by e.count; 22. for each transaction t in T 23. for each e i in t.former 24. if!fes.former.contains(e i) 25. delete e i; 26. end if 27. end for 28. sort t.former& t.latter in desc order by e.count 29. insertfptree(t, root); 30. end for 31. for each path p in FP-tree 32. m is the count of the last service event in the former part of p; 33. for each service event e i in the latter part of p 34. if e i.count/m< 35. delete e i; 36. end if 37. end for 38. end for 39. return FP-tree; In Algorithm 1, we firstly generate the frequent event set (line 4-20), which contains two parts as the same as the transactions. We calculate the count for each events (line 4-14), and calculate the support degree of events in the former events and delete ones whose support degree is lower than given minimal support degree (line 15-20). We sort the two parts of frequent event set in order of the count of events respectively (line 21). Then we delete the infrequent events in the former events for each transaction (23-27) and sort both events in each transaction in order of the count of events, too (line 28). Lastly, we insert each processed transaction into the FP-tree via algorithm insertfptree (line 29). For each event in transaction, we either add the count of existing node or insert a new node into the FP-tree. At last, we prune the FP-tree through calculating the confidence ratio of latter service events (line 31-38). D. Realization of Service Composition After generating the FP-Tree based on the analysis of sensor data, for a given service event, its situations and reacted service events can be obtained from the path of FPtree. Based on the real-time situation and its reacted service events, we can realize our service composition method in real time. Each proactive service will be set to generate specific service events when it is created. When a service event is generated, our method searches the path of FP-tree whose former parts contain the service event, and obtains the situations of the service. If any situation is satisfied, the latter part of the path can indicate the services to compose next and new service events to generate. The preudo-code of our algorithm is showed as followed:

6 Algorithm 2: servicecomposition Input: Output: S, the proactive data service set; FP-tree, the FP-tree contains all the association rules; E, the current service even; null; 1. PATH = FP-tree.getPathContains(E); 2. if PATH!=null 3. for each path i in PATH 4. C i = path i.former- E; 5. if C i is stratified 6. E i = path.latter; 7. for each E ij in E i 8. find S for E i from S; 9. send E to S 10. S.triggerOperations(E i); 11. end for 12. end if 13. end for 14. end if In Algorithm 2, for the current service event, we firstly find all the paths from the FP-tree whose former parts contain it (line 1), then we obtain the situation for the service event from each path through taking the service event away from the former part of the path (line 4). If the situation is satisfied, we obtain the service events contained in the latter part of the path as service events happen next (line 6). For each service event, we find service to generate it based on the mapping of sensors and services and trigger corresponding operations to generate the service events (line 9, 10). IV. EVALUATION In this section, we will verify the effectiveness our service composition method through a set of experiments. We compare our modified FP-tree algorithm (with the preprocessing of transactions) with traditional FP-tree algorithm to verify that our method can compose service and generate new service events with higher precision and recall. A. Experiment Setup 1) Data set The datasets used in our experiment is the real sensor data from a coal-fired power plant. We select sensor data from 5 systems of the plant. There are 1751 sensors and each sensor generated one value per 3 minutes from :00:00 to :59:59. 2) Environment We implemented our method in a cluster consisting of 5 nodes, and the nodes are running in virtual machines with CentOS release 6.4, four Intel Core i CPUs 3.10G Hz and 4.00 GB RAM. All the algorithms are implemented in Java with JDK ) Experiment method We random select 10 different days data and simulate stream data for each sensor strictly according to the real sensor data s timestamp. And we use the data before as training data to generate the correlations among service events. To verify the effectiveness of our method we design following criterion: (1) Precision: the precision of our method can be calculated as follow: (2) Recall: the recall of our method can be calculated as follow: In which L={l 1, l 2,, l m } be the list of service events which are expected to generate by service composition, C = {c 1, c 2,, c n } be the service events happened in practice. B. Effectiveness Evaluation Figure 4 (a) shows the numbers of co-occurrence service event correlations generated with different minimal frequency degree, and Figure 4 (b) shows the numbers of successive service event correlations generated with different (a) the numbers of co-corrence service event correlations (b) the numbers of successive service event correlations (c) the precision and recall with different parameters Figure 4. The Result of Experiment with Different Parameters.

7 minimal confidence ratio. The higher threshold is set, the less number of correlations will be generated. This is because more service events will be filtered out as the infrequent items, so there will be fewer the frequent service event sets. Figure 4(c) shows the precision and recall of new service events generated by service composition with different minimal confidence ratio. We randomly select 100 service events from 10 days data as current service events, and calculate the average precision and recall. The precision and recall are both decreased with the increasing of minimal confidence ratio. Although the smaller minimal frequency degree and minimal confidence ratio will obtain higher precision and recall, the number of correlations will be increased. And it will bring burden for the updating and searching of correlations. For balancing the number of correlations and the precision and recall of generated service events, we set the minimal frequency degree as 0.05 and the minimal confidence ratio as 0.4. Then we set the training data size as 1 week, 2 weeks, 3 weeks, 4 weeks and 5 weeks, and calculated the precision and recall for our improved FP-tree algorithm with or without the pre-processing of transactions. Figure 5 shows the comparison of the precision and recall of different FP-tree algorithm with different training data size. As showed in Figure 5, the highest of precision of our method is 81% and the highest recall is 67%. And the precision and recall for our method are higher than FP-tree algorithm without the pre-processing of transactions with different training data size. That is because through the preprocessing, we removed the redundant events from the transactions and divided them into different partitions, in (a) the precision with different data size (b) the recall with different data size Figure 5. The Result of Experiment with Different Data Size. which service events in the same partition are relative and service events in different partitions are irrelative. Therefore the generated correlations are much more effective and can obtain higher precision and recall than that without the preprocessing. V. RELATED WORK Service composition is one of the hottest research problems in service-oriented computing [14]. Several approaches have been proposed in recent years regarding situational service composition. Most of these works can be classified into two kinds, data-driven and situation modeling. In our previous works [15, 16], we proposed an end-user supported exploratory service composition method oriented to application in research collaboration, telemedicine and emergency areas. This method can recommend to users available services at runtime based two manners. One is to do semantic label for the input and output of service, the other is to utilize the existing composition plans. Liu et al [17] presents a systematic data-driven approach to assisting situational application development through extracting useful information from multiple sources to abstract service capabilities with a set of tag. Besides the above work, many researchers proposed situation-aware approaches based on their situation/context model. Bai et al. [7] considered the context of user when mining the history composition plan, which can support the individual service composition. Doulkeridis et al. [18, 19] firstly classified the services, and use the same dimensionality to describe the context of services in the same class. It aimed to help user to select one service from a class based on current context. Driver et al. [20] presented an application framework for mobile, context-aware trails-based application development. The framework supports trail generation through context-based activity set reduction and trail reconfiguration point identification through identification of significant context events. However, it unpractical to predefine unified and centralized situational model, we need to define different model rely on the application characteristic from different domain. Most of current works focus on static data source, while sensor data in IoT applications are typical stream data and the situation keeps changing with the change of sensor data. Because of the dynamical of IoT applications, history composition plans are not always effective for new service composition. Azadeh et al. [21, 22] propose a new framework for composing sensor-cloud services based on dynamic features in spatio-temporal aspects. Different from the above works, this paper aims to propose a sensor datadriven service composition method facing the dynamic and uncertain situations. VI. CONCLUSION AND FUTURE WORK In this paper, we proposed a data-driven service composition method based on our previous proactive data service model to handle the uncertain and dynamic situations in IoT environment. We regard service events that frequently happen together with specific service event as its situations, and service events happen next as reacted actions. We utilize

8 a modified FP-tree algorithm to generate two kind service event correlations, i.e. co-occurrence service event correlation and successive service event correlation. Then we utilize these correlations to realize our service composition method, which can detect the real-time situation for current service event and automatically compose service and generate new service events next. The experiments demonstrate that the effectiveness of our method. In the future, we plan to further refine our method by considering more complex situations, such as situations that contain service events with temporal relations. ACKNOWLEDGMENT This work was supported in part by a grant from the National Natural Science Foundation of China (Grant No ), Beijing Natural Science Foundation (Grant No ), The Program for Youth Backbone Individual, supported by Beijing Municipal Party Committee Organization Department, Research of Instant Fusion of Multi-Source and Large-scale Sensor Data. REFERENCES [1] Han Y, Wang G, Yu J, et al. A Service-Based Approach to Traffic Sensor Data Integration and Analysis to Support Community-Wide Green Commute in China[J]. IEEE Transactions on Intelligent Transportation Systems, 2016:1-10. [2] Chu V W, Wong R K, Liu W, et al. Traffic Analysis as a Service via a Unified Model[C]// IEEE International Conference on Services Computing. IEEE, 2014: [3] Zhang J, Radia N, Li Z, et al. An Infrastructure Supporting Considerate Sensor Service Provisioning[C]// The 6th IEEE International Conference on Service Oriented Computing and Applications (SOCA). 2013: [4] Guilly T L, Olsen P, Ravn A P, et al. HomePort: Middleware for heterogeneous home automation networks[c]// IEEE International Conference on Pervasive Computing and Communications Workshops. 2013: [5] Pinto J, Martins R, Sousa J B. Towards a REST-style architecture for networked vehicles and sensors[c]// IEEE International Conference on Pervasive Computing and Communications Workshops. IEEE, 2010: [6] Han YB, Liu C, Su S, et al. A Decentralized and Service-based Apporach to Proactively Correlating Stream Data[C]//.International Conference on Internet of Things, [7] Bai L, Wei J, Huang X, Ye D, Huang T. An exploratory service composition approach for mobile application[j]. Ruan Jian Xue Bao/Journal of Software, 2015,26(9): [8] Sheng QZ, Benatallah B. ContextUML: a UML-based modeling language for model-driven development of context-aware Web services development [C]//. Proc. of Int'l Conf. on Mobile Business, [9] Yau SS, Liu J. Incorporating situation awareness in service specifications [C]//. Proc. of 9th IEEE Int'l Symp. on Object and Component-Oriented Real-Time Distributed Computing [10] Yau SS, Yao Y, Yan M. Development and runtime support for situation-aware security in autonomic computing [C]//. Proc. of 3rd Int'l Conf. on Autonomic and Trusted Computing [11] Yau SS, Davulcu H, Mukhopadhyay S, Gong H, Huang D, Singh P, Gelgi F. Automated situation-aware service composition in serviceoriented computing [J]. Int'l Journal on Web Services Research, 2007, 4(4): [12] Han J, Pei J, Yin Y. Mining frequent patterns without candidate generation[c]// ACM SIGMOD International Conference on Management of Data. ACM, 2000:1-12. [13] Nahler G. Pearson correlation coefficient[j]. Springer Topics in Signal Processing, 2009, 2:1-4. [14] Medjahed B, Bouguettaya A, Elmagarmid A K. Composing Web services on the Semantic Web[J]. The VLDB Journal, 2003, 12(4): [15] Han YB, Wang GL, Ji G, et al. Situational data integration with data services and nested table[j]. Service Oriented Computing and Applications, 2013, 7(2): [16] Han YB, Wang HC, Wang JW, Yan SY, Zhang C. An end-useroriented approach to exploratory service composition [J]. Journal of Computer Research and Development, 2006,43(11): [17] Liu X, Ma Y, Huang G, et al. Data-Driven Composition for Service- Oriented Situational Web Applications[J]. Services Computing IEEE Transactions on, 2015, 8(1):2-16. [18] Doulkeridis C, Vazirgiannis M. CASD: Management of a contextaware service directory[j]. Pervasive Mobile Computing, 2008,4(5): [19] Doulkeridis C, Vazirgiannis M. Querying and updating a contextaware service directory in mobile environments[c]// 2004 IEEE/WIC/ACM Int l Conf. on Web Intelligence (WI 2004). Washington: IEEE Computer Society, [20] Driver C, Clarke S. An application framework for mobile, contextaware trails[j]. Pervasive Mobile Computing, 2008,4(5): [21] Neiat A G, Bouguettaya A, Sellis T. Spatio-Temporal Composition of Crowdsourced Services[M]// Service-Oriented Computing. Springer Berlin Heidelberg, [22] Neiat A G, Bouguettaya A, Sellis T, et al. Spatio-temporal Composition of Sensor Cloud Services[C]// IEEE International Conference on Web Services. 2014: