JP5026324B2 - Event recognition system and event recognition method - Google Patents

Description

  The present invention relates to an event recognition system including a plurality of sensor nodes having a sensing function and a communication function, and a server that collects data from these sensor nodes and recognizes events.

  In recent years, there are those that can collect information in real time using sensors connected to a network in fields such as environmental measurement, disaster prevention, crime prevention, and manufacturing. Among these, in addition to simply transmitting the response value of the sensor to a remote location via a communication line, there are those that select a communication line according to the characteristics of the detected data, for example, the degree of urgency. In addition, in order to simplify information collection, there is one using a sensor network including sensor nodes having a battery-driven and wireless data communication function.

  In such a battery-driven wireless sensor network, a technology has been developed that achieves both a sophisticated sensing function and a long drive time. For example, the average of the sum of the power consumed by the hardware module having the sensing function of the sensor node and the power consumed by the hardware module having the communication function is controlled to be constant, Sensing function for sensor nodes that sense highly urgent information or important information by controlling the nodes to be equal or controlling the sum of power to be proportional to the remaining battery capacity There has been proposed a sensor network that increases the power consumed by the hardware module of No. 1 and increases the power consumed by the hardware module of other communication functions (see Patent Document 1).

JP 2005-208719 A

  However, the sensor network proposed in Patent Document 1 has the following problems. That is, in this sensor network, the sensor node detects a single event such as the presence of a person, the concentration of particles, or the presence of strain. However, this sensor network identifies and evaluates sensor nodes that detect important information when multiple sensor nodes detect changing events such as human behavior patterns that are spatial movements. It was not possible to cope with the temporal change of the sensing event.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide an event recognition system and an event recognition method capable of recognizing an event by selecting a sensor node important for event recognition. .

The present invention relates to an event recognition system comprising a plurality of sensor nodes having a sensing function and a communication function for detecting an event, and a server for collecting data from these sensor nodes. A decision tree generating means for generating a feature vector from time-series sensor data of the detected sensor nodes, and generating a decision tree for event recognition from the feature vector and a desired recognition result of the sample event; and the decision tree during learning Sensor node function selecting means for selecting the function of each sensor node according to the above, sensor node function notifying means for notifying each sensor node of the function selected by the sensor node function selecting means at the time of learning, and an event in the operation after learning When sensor data is obtained from the detected sensor node, the decision tree is used. It is characterized in further comprising a event recognition means for recognizing an elephant.
Also, in one configuration example of the event recognition system of the present invention, the sensor node function selecting means divides each sensor node into a relay node that performs data relay and a sensing node that performs both event detection and data relay. It is characterized by.
In one configuration example of the event recognition system of the present invention, the sensor node includes a sensor for detecting an event, a cost determining unit for determining a communication cost and a sensing cost of the own node, and a communication determined by the cost determining unit. Cost notification means for notifying other sensor nodes of cost and sensing cost, and receiving communication cost and sensing cost data from other sensor nodes within the communication range when a direct communication path to the server cannot be secured Then, among the sensor nodes within the communication range of the own node, when the sensor detects an event, the relay node selection means for selecting the sensor node having the smallest sum of the communication cost and the sensing cost as the relay node, Sensor data transmission means for transmitting sensor data indicating that an event has been detected, and other sensor nodes When the sensor data is received, if a direct communication path to the server can be secured, the received sensor data is transferred to the server, and if the direct communication path to the server cannot be secured, It comprises data relay means for transferring the received sensor data to the relay node selected by the relay node selection means.

In the event recognition method of the present invention, the server generates a feature vector from the time-series sensor data of the sensor node that has detected the sample event for learning, and from this feature vector and a desired recognition result of the sample event, A decision tree generation step for generating a decision tree for event recognition, a sensor node function selection step for the server to select a function of each sensor node according to the decision tree during learning, and a server for the sensor node function selection step for learning. A sensor node function notifying step for notifying each sensor node of the selected function, and when the server obtains sensor data from the sensor node that detected the event in the operation after learning, the event is recognized using the decision tree. And an event recognition step.
Also, in one configuration example of the event recognition method of the present invention, the sensor node function selection step divides each sensor node into a relay node that performs data relay and a sensing node that performs both event detection and data relay. It is characterized by.
In addition, according to one configuration example of the event recognition method of the present invention, a sensor node determines a communication cost and a sensing cost of its own node, and a communication cost and a sensing cost determined by the sensor node in the cost determination step. A cost notification step for notifying other sensor nodes, and if the sensor node cannot secure a direct communication path to the server, it receives communication cost and sensing cost data from other sensor nodes within the communication range. When the sensor node detects an event, the relay node selection step of selecting the sensor node having the smallest sum of the communication cost and the sensing cost as the relay node among the sensor nodes within the communication range of the own node, A sensor data transmission step for transmitting sensor data indicating that an event has been detected; When a sensor node has received sensor data from another sensor node and has secured a direct communication path to the server, it transfers the received sensor data to the server and A data relay step of transferring the received sensor data to the relay node selected in the relay node selection step when the communication path cannot be secured is provided.

  According to the present invention, a feature vector is generated from time-series sensor data of a sensor node that has detected a sample event for learning, and a decision tree is generated from the feature vector and a desired recognition result of the sample event. By selecting the function of each sensor node and notifying each sensor node of the selected function, it is possible to identify the minimum necessary sensing node for event recognition and the relay node that does not perform sensing It becomes. Therefore, in the present invention, if an example of an event to be recognized is given as a sample event for learning in advance, and the event recognition system is trained and a decision tree is generated, the event recognition system is at least important for event recognition. Since the necessary sensing nodes are automatically selected, it is possible to deal with changing events such as human behavior patterns. In the present invention, it is possible to configure a sensor network with a minimum number of sensor nodes without sacrificing the sensing capability.

  Further, in the present invention, when a sensor node determines its own communication cost and sensing cost and notifies other sensor nodes, and a direct communication path to the server cannot be secured, other sensor nodes within the communication range Since the data of the communication cost and the sensing cost is received from the sensor node, the sensor node having the smallest sum of the communication cost and the sensing cost is selected as the relay node among the sensor nodes within the communication range of the own node. The sensor network can be reconfigured, and both the sensing function and the communication function can be achieved without sacrificing the sensing function. Further, if it is performed in the operation after learning, the sensor network can be reconfigured according to the latest communication status and the latest resource of the sensor node.

  Hereinafter, an event recognition system and an event recognition method according to an embodiment of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a sensor node according to the present embodiment. The sensor node 1 transmits a state measurement sensor 2 such as a temperature sensor, a gas sensor, a human sensor or a switch, a detection result of the sensor node 1 or a unique ID to the sensor node 1, and receives a signal from a server which will be described later. A wireless communication function module 3, a memory 4 that stores sensor data and a program, and a CPU (Central Processing Unit) 5 that controls the state measurement sensor 2 and the wireless communication function module 3 in accordance with the program stored in the memory 4 are provided. ing.

  As the state measuring sensor 2, a human sensor or a switch is adopted when used for sensing indoor human behavior, and a gas sensor or a temperature sensor is adopted when used for environmental measurement. Examples of the human sensor include those using infrared rays, ultrasonic waves, visible light, and the like. As described above, the state measuring sensor 2 is appropriately selected depending on the sensing purpose.

  Since the sensor node 1 is small and can be driven by an internal battery (not shown), the user can easily install the sensor node 1. Although a commercial power source can be used as the power source of the sensor node 1, it can be installed anywhere by battery drive. As wireless communication between the wireless communication function module 3 and a server to be described later, a wireless local area network (LAN), a specific low power wireless, a weak wireless, ZigBee (registered trademark), Bluetooth (registered trademark), or the like can be used. When these wireless communications are used, the wireless communication distance is several meters to several tens of meters. However, since the sensor node 1 has a relay function, long-distance communication can be realized by relaying communication between other sensor nodes 1 and the server.

  FIG. 2 is a block diagram showing the configuration of the server according to the present embodiment. The server 6 can be configured as a personal computer. The server 6 includes a memory 7 that stores data and programs, a CPU 8 that controls the entire server in accordance with a program stored in the memory 7, a communication interface 10, and the like. Connected to the communication interface 10 is a sensor node transceiver 9 for performing wireless communication with the sensor node 1.

  As the communication interface 10, for example, a serial port, a parallel port, a LAN port, a USB (Universal Serial Bus) port, a CF (CompactFlash) card slot, or the like can be used. In either case, a similar interface is required on the sensor node transceiver 9 side.

  FIG. 3 is a diagram showing the configuration of the event recognition system according to the present embodiment. The plurality of sensor nodes 1-1 to 1-3 installed around the user (not shown) have detected the presence of the user by detecting the presence of the user with the state measuring sensor 2 such as a human sensor. Is transmitted to the server 6 or the ID of the sensor node is transmitted to the server 6 instead of the detected data. Hereinafter, the detection data and ID obtained from the sensor node 1 are collectively referred to as sensor data.

  The server 6 receives sensor data from the sensor node 1 via the sensor node transceiver 9. And the server 6 recognizes the event (a user's action and the state of an environment) detected by the sensor node 1 by performing pattern recognition based on the received sensor data. Furthermore, the server 6 divides each sensor node 1 into a sensing node and a relay node, and notifies each sensor node 1 of the function.

  FIG. 4 is a block diagram showing functional means realized by the CPU 8 of the server 6. Functional means realized by the CPU 8 include a sensor node control unit 12, a database 13, and an event recognition unit 14. The sensor node control unit 12 includes a decision tree generation unit 15, a sensor node function selection unit 16, and a sensor node function notification unit 17.

  The decision tree generation unit 15 stores the sensor data obtained from the sensor node 1 at the time of learning and the desired recognition result input by the user in the memory 7 and constructs the database 13 in the memory 7. A decision tree for recognizing an event detected by the sensor node 1 is constructed in the database 13 from the recognition result.

The sensor node function selection unit 16 selects the function of each sensor node 1 by selecting the minimum necessary sensor node 1 important for event recognition based on the decision tree constructed in the database 13.
The sensor node function notifying unit 17 notifies each sensor node 1 of the function selected by the sensor node function selecting unit 16 through the communication interface 10.

The event recognition unit 14 refers to the database 13 based on the sensor data obtained from the sensor node 1 in the recognition operation after learning, and recognizes the event by performing pattern recognition using a decision tree.
The CPU 8 implements the sensor node control unit 12, the database 13, and the event recognition unit 14 by executing a program stored in the memory 7.

  Note that the sensor node control unit 12, the database 13, and the event recognition unit 14 do not necessarily need to be built in the server 6, and the sensor node control unit 12 is provided outside the system in order to use higher performance computing resources. The database 13 and the event recognition unit 14 may be arranged to communicate with the server 6 via a communication network. Further, the database 13 can be prepared for each user.

FIG. 5 is a plan view illustrating an arrangement example of the sensor nodes 1-1 to 1-10 in a certain environment, for example, a living room. The sensor nodes 1-1 to 1-10 are arranged along the passage 51 from the entrance 50 of the living room. As the sensor 2 for measuring the state of each of the sensor nodes 1-1 to 1-10, a contact sensor for detecting contact by a user, a human sensor for detecting approach of the user, or the like is used.
The server 6 recognizes the user's behavior, controls the environment such as lighting and air conditioning of the living room, and monitors the user.

  Hereinafter, the operation at the time of learning in which learning data (data from sensor node 1 and desirable recognition result of sample event) is acquired, decision tree is generated, function selection of sensor node 1 and function notification to sensor node 1 will be described. To do. FIG. 6 is a flowchart showing the operation of the server 6 during learning. FIG. 7 shows a detection example of the sensor nodes 1-1 to 1-10 for a user's action (sample event). When the sensor nodes 1-1 to 1-10 detect the presence of the user, the sensor nodes 1-1 to 1-10 transmit sensor data indicating that the presence of the user is detected to the server 6. In the example of FIG. 7, the sensor data transmitted from the sensor nodes 1-1 to 1-10 is represented by a pulse 60. As described above, the sensor nodes 1-1 to 1-10 may transmit a unique ID as sensor data indicating that the presence of the user is detected.

  The user entered the room and moved the passage 51 toward the desk 52. As a result, as shown in FIG. 7, the sensor nodes 1-1, 1-2, 1-3, 1-4, 1-5, 1-6, 1-7, 1-8, 1-9, 1- 10 in turn detected the presence of the user. And since the user worked at the desk 52 for a while, the sensor node 1-10 continuously detected the presence of the user during this work. Thereafter, the user leaves the desk 52, moves through the passage 51, and leaves the room. Sensor nodes 1-9, 1-8, 1-7, 1-6, 1-5, 1-4, 1-3, 1-2, 1-1 detect the presence of the user in sequence by leaving and leaving the room. did.

  Focus on the four actions of entering, sitting, leaving, and leaving the series of actions. These four actions are highly likely to be related to environmental control by the user, such as turning on the lights when entering the room, turning on the individual lights on the desk when sitting, turning off the individual lights when leaving, and turning off the lights when leaving the room. Therefore, consider recognizing these four actions.

  The sensor node 1-1 detects the user's entry, the sensor node 1-10 detects the seating, the sensor node 1-9 detects the leaving, and the sensor node 1-1 detects the leaving. It can be seen that the sensors 1-1, 1-9, and 1-10 are closely related to the four actions. However, these sensors alone cannot recognize four actions. That is, it is not known whether the detection of the sensor node 1-1 indicates the user's entry or the exit, and the work at the desk indicates whether the detection of the sensor node 1-10 indicates the user's seating. It is not known whether it is a thing, and it is not known whether the detection of the sensor node 1-9 indicates a detection before sitting or a leaving.

  Here, the sensor data of the ten sensor nodes 1-1 to 1-10 are S1 to S10, the sensor detection vectors are (S1, S2, S3, S4, S5, S6, S7, S8, S9, S10), The state that the presence of the user is currently detected is r, the state that the presence of the user is detected x time before the present is px, and the state that the presence of the user is not detected from the time t to the present is indicated. Let n. t and x are positive numbers (t ≧ x), and t is called an order. When t = 3, the sensor detection vector can be expressed as shown in FIG. 8 from the time-series detection results of the sensor nodes 1-1 to 1-10. However, in the example of FIG. 8, it is assumed that the user takes one hour to move between the nodes.

When the decision tree generation unit 15 of the server 6 receives sensor data from the sensor nodes 1-1 to 1-10 during learning, the decision tree generation unit 15 stores this data in the memory 7 (step S11 in FIG. 6).
Further, the user inputs, as a desired recognition result of the sample event, the user's attention behavior (entrance, seating, leaving, leaving) to the server 6 in association with the sensor data (step S12). That is, the action “entrance” is associated with the sensor detection vector when the user first passes the sensor node 1-1, and the action “sitting” is associated with the sensor detection vector when the user passes the sensor node 1-10. Are associated with the sensor detection vector when the user passes through the sensor node 1-9, and the action “leave” is associated with the sensor detection vector when the user last passes the sensor node 1-1. Corresponding actions. In this way, the database 13 in which the data of the sensor nodes 1-1 to 1-10 and the desired recognition result of the sample event are associated is constructed.

  In the above example, the user's four actions of entering, leaving, sitting, and leaving can be distinguished by the sensor detection vector generated from the time-series detection results of the sensors 1-1 to 1-10. Further, as an example, it can be understood that the user's behavior can be recognized by using the information of the sensor nodes 1-1, 1-2, 1-9, and 1-10 intuitively. Humans can make such a selection relatively easily, but it takes too much time to think about them in various cases. For example, how an individual wants to adjust and control the environment, such as lighting, will vary depending on the individual, and will vary depending on the environment. It takes a lot of labor to install and operate the sensor node while the user thinks about it. Therefore, it is necessary to automatically realize such sensor selection without human intervention.

  Here, a feature quantity in pattern recognition is selected based on the information gain. Specifically, the decision tree method of J.R.Quinlan is used (see the document “J.R.Quinlan,“ Induction of Decision Trees ”, Machine Learning 1: 81-106, 1986)). The information gain represents the importance of an attribute (here, a sensor node) that becomes a branch node of the decision tree, and the feature amount can be considered as a detection result of each sensor node 1.

  The decision tree generation unit 15 constructs a decision tree in the database 13 using a data mining tool WEKA including decision tree software developed using the Quinlan ID3 algorithm (step S13 in FIG. 6). The data mining tool WEKA is disclosed in, for example, the document “http://www.cs.waikato.ac.nz/ml/weka/”. In this decision tree construction process, the decision tree generator 15 generates a sensor detection vector, which is a feature vector, from the time-series sensor data of the sensor node 1 under a certain order, and calculates the information gain of each sensor node 1. The sensor node 1 having the maximum information gain is determined to be a branch node of the decision tree.

  FIG. 9 shows an example of a decision tree generated by the decision tree generation unit 15. According to the decision tree of FIG. 9, it can be seen that if S1 = r and S2 = n, it can be recognized that the user has entered the room. Similarly, if S1 = r, S2 = p1, it can be recognized that the user has left the room, and if S1 = n, S10 = r, S9 = p1, it can be recognized that the user is seated, and S1 = n, S10 = p1. Can be recognized as leaving the user. In addition, according to the decision tree of FIG. 9, the sensor data of the six sensor nodes 1-3 to 1-8 other than the sensor nodes 1-1, 1-2, 1-9, and 1-10 are used to recognize the user's behavior. It can be seen that the sensor node that is not used and therefore the minimum necessary for action recognition can be selected.

As described above, it is understood that both the generation of the event recognition rule and the determination of the sensor node 1 that is the minimum necessary for the event recognition can be realized by the process of step S13.
Subsequently, the sensor node function selection unit 16 selects the sensor node 1 that is a branch node in the decision tree constructed in the database 13 as the minimum necessary sensor node 1 for event recognition. The function of each sensor node 1 is selected (step S14 in FIG. 6). In the example of FIG. 9, sensor nodes 1-1, 1-2, 1-9, and 1-10 are selected as the sensor nodes 1 that are the minimum necessary for event recognition.

  In a sensor network composed of a plurality of sensor nodes 1, it is important to achieve both a sensing function and a communication function, and a sensor network cannot be configured even if one of them is missing. When a direct wireless communication path from one sensor node 1 to the server 6 cannot be secured, a wireless communication path can be secured by using another sensor node 1 as a relay node. The problem is how to determine such a relay node. Since the sensor node 1 serving as the relay node relays communication of other sensor nodes 1 other than itself, the power consumption increases. If both sensing and relaying are performed, a large amount of power is consumed, and the function of the sensor node 1 may be stopped. Therefore, it is important to use the selection function of the sensor node 1 and classify the sensor node 1 into a sensing node and a relay node so that the function of the entire sensor network is not impaired.

  That is, the sensor node function selection unit 16 detects an event of the sensor nodes 1-1, 1-2, 1-9, and 1-10 that are necessary for event recognition, and also relays data depending on the case. The sensing nodes are selected, and the other sensor nodes 1-3 to 1-8 are selected as relay nodes (step S14). Thus, the sensor node 1 is divided into the sensing node and the relay node, and the function of each sensor node 1 can be selected.

  The sensor node function notifying unit 17 notifies each sensor node 1 of the function selected by the sensor node function selecting unit 16 through the communication interface 10 (step S15 in FIG. 6). According to the example of FIG. 9, the sensor node function notification unit 17 notifies the sensor nodes 1-1, 1-2, 1-9, and 1-10 that the function to be realized is a sensing function. Further, the sensor nodes 1-3 to 1-8 are notified that the function to be realized is the relay function.

Finally, each sensor node 1 calculates the sensing cost by the function notified from the sensor node function notifying unit 17 of the server 6, selects the relay node having the smallest sum of the communication cost and the sensing cost, and sets the sensor network. Reconfiguration is performed (step S16 in FIG. 6).
FIG. 10 is a block diagram showing functional means realized by the CPU 5 of the sensor node 1. As functional means realized by the CPU 5, there are a sensor data transmission unit 18, a function setting unit 19, a cost determination unit 20, a cost notification unit 21, a relay node selection unit 22, and a data relay unit 23. The CPU 5 realizes these functional means by executing a program stored in the memory 4.

  When each sensor node 1 cannot secure a direct wireless communication path to the server 6, each sensor node 1 has several sensor nodes 1 that are within the wireless communication range of its own node and have a wireless communication path to the server 6. Select as a relay node. As a selection criterion for this relay node, first, communication cost, that is, frequency of communication packet loss due to noise or the like when communicating to the server 6 can be mentioned. In the sensor network, in order to transmit data to the server 6 with the highest possible accuracy, it is important to select a wireless communication path with less packet loss.

  Further, another selection criterion for the relay node is sensing cost, that is, resources necessary for sensing important information. Here, the resource indicates not only the capacity of the battery but also the usage frequency of the CPU 5 of the sensor node 1 and the usage frequency of the memory 4. In a sensor network, securing a sensing function is fundamental, and the node 1 sensing important information consumes resources accordingly. Since resource consumption due to communication may affect the sensing function, it is important to concentrate on the sensing function as much as possible. If possible, the sensor node 1 with a low sensing cost may be selected as a relay node. is important. Therefore, the sensor node 1 having the smallest sum of the communication cost and the sensing cost is selected as the relay node.

  The function setting unit 19 of each sensor node 1 sets the function notified from the server 6 as a function to be realized by the own node. Thereby, each sensor node 1 operates as a sensing node or a relay node. The sensing node performs sensing as a main operation, but may also operate as a relay node in some cases.

  The cost determination unit 20 of each sensor node 1 determines the sensing cost of the own node based on the function notified from the server 6 and the resource of the own node. When the function notified from the server 6 is a relay function, the sensing cost of the sensor node 1 is zero. Further, when the function notified from the server 6 is a sensing function, the cost determination unit 20 determines the sensing cost of the own node based on the current resource of the own node. Furthermore, the cost determination unit 20 determines the communication cost of the own node based on, for example, the loss frequency of transmission / reception data.

  The cost notification unit 21 of each sensor node 1 wirelessly transmits the communication cost and sensing cost of the own node determined by the cost determination unit 20 to another sensor node 1 through the wireless communication function module 3 at regular intervals, for example.

The relay node selection unit 22 of each sensor node 1 receives communication cost and sensing cost data from other sensor nodes 1 within the wireless communication range when a direct wireless communication path to the server 6 cannot be secured. Among the sensor nodes 1 within the wireless communication range of the node, the sensor node 1 that minimizes the sum of the communication cost and the sensing cost is selected as a relay node (step S16 in FIG. 6). The relay node selection unit 22 of each sensor node 1 wirelessly transmits a signal to the server 6 and determines that a direct wireless communication path to the server 6 cannot be secured when there is no response from the server 6. To do.
As described above, the sensor network can be reconfigured, and the learning operation is completed.

  FIG. 11 shows an example of the operation for reconfiguring the sensor network in the event recognition system. FIG. 11 illustrates a case where the sensor node 1 -A that cannot directly transmit sensor data to the server 6 transmits data to the server 6. The sensor node 1-A selects any one of the sensor nodes 1-B, 1-C, and 1-D within the wireless communication range as a relay node. At this time, the sensor node 1-B is notified from the server 6 that it should operate as a sensing node. The sensor nodes 1-C and 1-D are notified from the server 6 that they should operate as relay nodes.

  The sensing cost of the sensing node is 10, and the sensing cost of the relay node that does not perform sensing is 0. Further, the communication cost of the sensor node 1-B to the server 6 is 3, the communication cost of the sensor node 1-C to the server 6 is 5, and the communication cost of the sensor node 1-D to the server 6 is 4. In this case, the sensor node 1-A selects the sensor node 1-D having the minimum communication cost and sensing cost as a relay node. The sensor node 1-C does not contribute to sensing or communication, and transitions to a sleep state as a relay node instead of the sensor node 1-D, or is excluded from the sensor network as an unused sensor node.

  FIG. 12 shows another example of the operation for reconfiguring the sensor network. Here, in the same configuration as the example of FIG. 11, the communication costs of the sensor nodes 1-B, 1-C, and 1-D are set to 3, 15, and 20, respectively. Since the sensor nodes 1-C and 1-D are included in the noise region 120 where there is much noise turbidity, the communication quality is poor and the communication cost is high. Therefore, even if the sensor nodes 1-C and 1-D are selected as relay nodes, there is a high possibility that the sensor data does not reach the server 6.

  The sensor node 1-A selects the sensor node 1-B having the smallest communication cost and sensing cost as a relay node. In this case, resources are consumed, but the sensor node 1-B operates as a relay node at the same time as the sensing node. The sensor nodes 1-C and 1-D are installed in other places or excluded from the sensor network.

  In the initial sensor network before the function is notified to each sensor node 1 in step S15, each sensor node 1 operates as a sensing node and also as a relay node. In the initial sensor network, when the relay node selection unit 22 of each sensor node 1 cannot secure a direct wireless communication path to the server 6, only the communication cost is considered and the sensor node 1 within the wireless communication range of the own node is considered. Among them, the sensor node 1 with the lowest communication cost is selected as the relay node.

  As for the normal recognition operation at the time of learning and after learning, when the sensor data transmission unit 18 of each sensor node 1 operates as a sensing node, the state measuring sensor 2 of its own node detects the presence of the user as described above. Sometimes, sensor data is wirelessly transmitted through the wireless communication function module 3. At this time, when the sensor data transmission unit 18 can secure a direct wireless communication path to the server 6, the sensor data transmission unit 18 transmits sensor data to the server 6 and cannot secure a direct wireless communication path to the server 6. Transmits the sensor data to the relay node selected by the relay node selection unit 22 of the own node.

  For normal recognition operation during learning and after learning, when the data relay unit 23 of each sensor node 1 operates as a sensing node or a relay node, the sensor data received from other sensor nodes 1 is transmitted through the wireless communication function module 3. Forward. At this time, when the data relay unit 23 can secure a direct wireless communication path to the server 6, the data relay unit 23 transfers the sensor data to the server 6 and cannot secure a direct wireless communication path to the server 6. Then, the sensor data is transferred to the relay node selected by the relay node selection unit 22 of the own node.

  As described above, in the present embodiment, a feature vector is generated from time-series sensor data of the sensor node 1 that has detected a sample event for learning, and a decision tree is generated from the feature vector and a desired recognition result of the sample event. Since the function of each sensor node is generated and selected according to the decision tree and the selected function is notified to each sensor node 1, the minimum necessary sensing node for event recognition and the relay node that does not perform sensing Can be specified. Therefore, in this embodiment, if an example of an event to be recognized is given as a sample event for learning in advance, the event recognition system is learned and a decision tree is generated, the event recognition system is important for event recognition. Since the minimum necessary sensing node is automatically selected, it is possible to cope with a changing event such as a human behavior pattern. In the present embodiment, it is possible to configure a sensor network with a minimum number of sensor nodes without sacrificing the sensing capability.

Note that the reconfiguration of the sensor network may be executed not only during learning but also in the operation after learning.
For example, the cost determination unit 20 of each sensor node 1 determines the latest communication cost and the latest sensing cost of its own node at regular intervals, and the cost notification unit 21 wirelessly transmits the communication cost and the sensing cost. When the relay node selection unit 22 of each sensor node 1 cannot secure a direct wireless communication path to the server 6, the latest communication cost and the latest sensing cost received from other sensor nodes 1 within the wireless communication range are set. Based on this, the relay node is selected again. Thereby, the sensor network can be reconfigured according to the latest communication status and the latest resource of the sensor node 1.

  In the present embodiment, the user's behavior is exemplified as an event recognized by the event recognition system. However, the present invention is not limited to this, and may be an event that recognizes an environmental state such as an indoor state.

  The present invention can be applied to a technique for recognizing an event using a plurality of sensor nodes.

It is a block diagram which shows the structure of the sensor node which concerns on embodiment of this invention. It is a block diagram which shows the structure of the server which concerns on embodiment of this invention. It is a figure which shows the structure of the event recognition system which concerns on embodiment of this invention. It is a block diagram which shows the function means implement | achieved by CPU of a server in embodiment of this invention. It is a top view showing the example of arrangement | positioning of the sensor node in embodiment of this invention. It is a flowchart which shows the operation | movement at the time of the learning of the server which concerns on embodiment of this invention. It is a figure which shows the example of detection of the sensor node in embodiment of this invention. It is a figure which shows the example of the sensor detection vector in embodiment of this invention. It is a figure which shows the example of the decision tree in embodiment of this invention. It is a block diagram which shows the function means implement | achieved by CPU of a sensor node in embodiment of this invention. It is a figure which shows one example of the operation | movement which reconfigure | reconfigures a sensor network in embodiment of this invention. It is a figure which shows another example of the operation | movement which reconfigure | reconfigures a sensor network in embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Sensor node, 2 ... State measurement sensor, 3 ... Wireless communication functional module, 4 ... Memory, 5 ... CPU, 6 ... Server, 7 ... Memory, 8 ... CPU, 9 ... Transmitter / receiver for sensor node, 10 ... Communication Interface: 12 ... Sensor node control unit, 13 ... Database, 14 ... Event recognition unit, 15 ... Decision tree generation unit, 16 ... Sensor node function selection unit, 17 ... Sensor node function notification unit, 18 ... Sensor data transmission unit, 19 A function setting unit, 20 a cost determination unit, 21 a cost notification unit, 22 a relay node selection unit, and 23 a data relay unit.

Claims (6)

In an event recognition system comprising a plurality of sensor nodes having a sensing function and a communication function for detecting an event, and a server for collecting data from these sensor nodes,
The server
Decision tree generating means for generating a feature vector from time-series sensor data of a sensor node that has detected a sample event for learning, and generating an event recognition decision tree from the feature vector and a desired recognition result of the sample event; ,
Sensor node function selection means for selecting the function of each sensor node according to the decision tree during learning;
Sensor node function notifying means for notifying each sensor node of the function selected by the sensor node function selecting means during learning;
An event recognition system comprising: event recognition means for recognizing an event using the decision tree when sensor data is obtained from a sensor node that has detected the event in an operation after learning. The event recognition system according to claim 1,
The sensor node function selecting means classifies each sensor node into a relay node that performs data relay and a sensing node that performs both event detection and data relay. The event recognition system according to claim 1,
The sensor node is
A sensor for detecting the event;
A cost determining means for determining the communication cost and sensing cost of the own node;
Cost notification means for notifying other sensor nodes of the communication cost and sensing cost determined by this cost determination means,
When a direct communication path to the server cannot be secured, communication cost and sensing cost data is received from other sensor nodes within the communication range, and the communication cost among the sensor nodes within the communication range of the own node And a relay node selection means for selecting a sensor node having the smallest sum of sensing costs as a relay node,
Sensor data transmitting means for transmitting sensor data indicating that an event has been detected when the sensor detects an event;
When a direct communication path to the server can be secured when sensor data is received from another sensor node, the received sensor data is transferred to the server, and the direct communication path to the server is set. An event recognition system comprising: data relay means for transferring received sensor data to the relay node selected by the relay node selection means when it cannot be secured. In an event recognition method for recognizing an event using a plurality of sensor nodes having a sensing function and a communication function for detecting an event and a server for collecting data from these sensor nodes,
A decision tree in which a server generates a feature vector from time-series sensor data of a sensor node that has detected a sample event for learning, and generates an event recognition decision tree from the feature vector and a desired recognition result of the sample event Generation step;
A sensor node function selection step in which the server selects a function of each sensor node according to the decision tree during learning;
A sensor node function notification step in which the server notifies each sensor node of the function selected in the sensor node function selection step during learning;
An event recognition method comprising: an event recognition step of recognizing an event using the decision tree when sensor data is obtained from a sensor node that has detected the event in an operation after learning. The event recognition method according to claim 4,
The sensor node function selecting step divides each sensor node into a relay node that performs data relay and a sensing node that performs both event detection and data relay. The event recognition method according to claim 4,
A cost determination step in which the sensor node determines the communication cost and sensing cost of the node;
A cost notification step in which the sensor node notifies the other sensor nodes of the communication cost and the sensing cost determined in the cost determination step;
When the sensor node cannot secure a direct communication path to the server, it receives communication cost and sensing cost data from other sensor nodes in the communication range, and the sensor node in the communication range of its own node Among them, a relay node selection step of selecting a sensor node having a minimum sum of communication cost and sensing cost as a relay node;
A sensor data transmission step of transmitting sensor data indicating that an event has been detected when the sensor node detects the event;
When a sensor node has received sensor data from another sensor node and has secured a direct communication path to the server, the received sensor data is forwarded to the server, and the direct communication to the server is performed. And a data relay step of transferring the received sensor data to the relay node selected in the relay node selection step when the communication path cannot be secured.