TW201437945A - Distant data classification system and method thereof - Google Patents

Abstract

A distant data classification system and method thereof are provided. The system and method apply to collect the parameter-related data relating to an environment within a remote region. It builds several monitoring results by these parameter-related data via an artificial neural network and a classification model, and provides alarm message corresponding the monitoring results. The system and method enable users to manage the conditions of the distant environment conveniently and efficiently without being present on the spot or performing measurement manually on the spot, and consider all factors at the same time to generate the objective result.

Description

Remote environmental data classification system and method thereof

The invention relates to a data classification system and a method thereof, in particular to a remote environment data classification system and method thereof using an artificial intelligence neural network and a classification algorithm for classification.

There are more than 4,000 species of melon flies in the world, and there are 148 species in Taiwan. These melon flies are mostly pests, and most of them spawn for specific parasitic plants. The larvae of the egg can cause damage to the feeding of the parasitic plants, especially when the larva eats the fruit part of the parasitic plant. Therefore, when a large number of pests occur, it means that the harvest of local farmers will be destroyed. In addition to seriously affecting the yield and quality of fruits and vegetables, it will become an obstacle to the export of fruits and vegetables in China due to quarantine problems. Therefore, the prevention and control of pests is often one of the important topics in the management and research of crop cultivation. However, only by recording the number of pests can not effectively interpret the degree of damage and distribution, and must be combined with temperature, relative humidity, solar radiation, wind speed, wind direction, rainfall and other arable land environmental parameters for further integration analysis.

Collect environmental factors that affect crop growth, such as the number of pests per unit area, total number of pests, temperature, relative humidity, amount of sunshine, wind speed, wind direction, rainfall, global geographic location, etc. The monitoring of the surrounding environment and the management of crop cultivation is an extremely important task. Management researchers can use this data to perform the necessary controls. Traditionally, the collection of data for these environmental factors is usually done by human means. For example, the calculation of the number of pests is manually captured by the insect net, and the number of pests per unit area or the total number of pests is counted by visual counting. Others such as temperature, relative humidity, amount of sunshine, The collection of data such as air volume, wind speed, rainfall, global geographical location, etc. is to install sensors on site to collect relevant data, and then management researchers can make statistical reports based on these data, so as to treat crops according to the contents of this report. Cultivation and management work.

However, the above-mentioned management and control methods must rely on a large amount of manpower, which is quite time-consuming, laborious and inefficient. If such data collection is not sorted out, it is difficult to perform the interpretation, and the information presented by the data can only be reviewed afterwards, so that the surrounding environment of the cultivated crops cannot be immediately detected, such as sensor failure and the number of pest populations. Surge, pests are happening, etc., the alarm message will be sent out immediately, so that the management researchers can immediately respond to the corresponding treatment, or let the farmers as the basis for spraying pesticides. In addition, in terms of data collection, it is also impossible to obtain data on environmental factors in large quantities at the same time. In addition, the management researchers interpret it by manual means, and the factors of consideration are limited, and it is not possible to consider the complicated factors at the same time. Manual interpretation is easy to subject to the subjective judgment of management researchers, and there are unobjective results, resulting in the inability to make appropriate pest control and management. The high level of professionalism in the interpretation of data makes it difficult for average farmers to use.

At this point, it is necessary to provide an innovative and negatively progressive remote environmental data. The classification system and its methods enable management researchers to visit the site without having to manually collect data. The system and method can calculate the current situation on the spot according to the collected data, and consider various factors to make an immediate reward or warning action, and have a self-learning mechanism to improve the accuracy of the monitoring alarm and data collection. The accuracy is to solve the aforementioned problems.

The present invention provides a remote environment data classification system, comprising: a sensing network, comprising a plurality of sensing node devices, respectively disposed at a plurality of monitoring points in the remote area to sense the surrounding environment of the monitoring points a situation in which a plurality of environmental monitoring parameter data are generated correspondingly; the front-end gateway device is coupled to the sensing network for receiving the environmental monitoring parameter data to transmit the environmental monitoring parameter data through the wireless communication system; And a back-end main control servo device, comprising: a database module, wherein the environmental monitoring parameter data is received and stored through the wireless communication system; and the neural network training module is connected to the database module to read The environmental monitoring parameter data, the environmental monitoring parameter data is input into the neural network for training, and the plurality of monitoring events are outputted by the neural network; and the classification model training module performs the monitoring events. Classification training to obtain multiple monitoring results.

The invention provides a remote environment data classification method, the steps comprising: (a) forming a plurality of environmental monitoring parameter data into a complex array multidimensional coordinate vector data, and then inputting the multidimensional coordinate vector data into a neural network for training. , after the training is completed, the plurality of monitoring is outputted by the neural network (b) classifying the monitored events according to the classification model to generate a plurality of monitoring results; (c) judging the untrained environmental monitoring parameter data with the monitoring results, and generating a judgment result; and (d) A warning message corresponding to the result of the determination is generated to transmit the warning message to the wireless communication device via the wireless communication system.

According to the system and method provided by the present invention, the collected environmental monitoring parameter data can be trained by using a neural network, and the results of the training through the neural network can be trained by a classification to obtain a plurality of Monitoring results. The monitoring results can be used to determine the untrained environmental monitoring parameters, and corresponding warning messages are generated and transmitted to a wireless communication device via a wireless communication module connected to the wireless communication system. The system and method provided by the invention can not only perform accurate analysis according to the collected data, but also have a self-learning mechanism, and can simultaneously consider various parameters in the data to calculate The situation on the spot makes an immediate warning action. Not only does the researcher not need to visit the site, the data collection does not need to be carried out manually, and there is no unobjective situation in the interpretation of the data. The operation of simple automation is also convenient for ordinary farmers.

10‧‧‧Remote environmental data classification system

100‧‧‧Sensing network

110‧‧‧Sensor node device

120‧‧‧Automatic statistical device for the number of animals

20‧‧‧Wireless communication system

200‧‧‧ front-end gateway device

30‧‧‧Wireless communication device

300‧‧‧Backend master control servo

301‧‧‧Database Module

302‧‧‧ class neural network training module

303‧‧‧Classification Model Training Module

304‧‧‧ Warning Module

305‧‧‧Parameter data normalization module

306‧‧‧Wireless communication module

307‧‧‧Parameter data inspection module

308‧‧‧Diagnostic Module

40‧‧‧Network Topology

401‧‧‧Superplane

S01~S06‧‧‧Steps

1 is a schematic structural diagram of a remote environment data classification system according to the present invention; FIG. 2A is a schematic structural diagram of a rear-end master control servo device; FIG. 2B is another embodiment of a rear-end master control servo device of the present invention; Schematic diagram of the structure; 3 is a flow chart of a method for classifying remote environment data according to the present invention; FIG. 4A is a schematic diagram of a network topology for generating a plurality of monitoring events according to the present invention; and FIGS. 4B and 4C are diagrams for classifying network topologies into plural numbers according to the present invention; A schematic diagram of the monitoring results.

Hereinafter, embodiments of the remote environmental data classification system and method thereof according to the present invention will be described in detail in conjunction with the drawings.

Please refer to FIG. 1 , which is a schematic diagram of the architecture of the remote environment data classification system of the present invention. The remote environmental data classification system 10 of the present invention is used for the monitoring of crop pests in a practical example, in particular, the crops of the orchard and the oriental fruit fly, but the invention is not limited thereto. The remote environment data classification system 10 is configured to allow a user at a local end location to perform environmental monitoring on an orchard at a remote location, and the remote environmental data classification system 10 can automatically diagnose the orchard at the remote location. The status of the site, instant return to users located at the local end. Therefore, the remote environment data classification system 10 includes a wireless communication device 30, a sensing network 100, a front-end gateway device 200, and a back-end master servo device 300, and is integrated into a wireless communication system 20.

The wireless communication system 20 can be a standardized GSM (Global System for Mobile Communications) compatible wireless communication system, or CDMA2000, GPRS, LTE, WiMAX, WCDMA, and the like. Since the remote environmental data classification system 10 of the present invention is constructed based on a decentralized system architecture, including a local end location rear end architecture and a remote end location The front-end architecture, therefore, uses the wireless communication system 20 of the various mobile communication standards described above as a bridge for data exchange between the local end location rear end architecture and the remote end location front end architecture.

The sensing network 100 can be wireless or wired, but a wireless sensor network (WSN) is the preferred embodiment of the present invention (hereinafter, the WSN wireless type will be taken as an example). The wireless sensing network 100 is composed of a plurality of sensing node devices 110, and each sensing node device 110 can be a wireless sensing node device controlled by a TinyOS operating system, and wirelessly connected to each other. The connection is based on a standardized wireless network protocol such as ZigBee or Bluetooth.

The sensing node devices 110 are respectively disposed at monitoring points to be monitored in the distal region, such as the growth sites of the plants in the farmland. In practical applications, each sensing node device 110 is configured to sense the condition of the surrounding environment of the monitoring point of the device, and correspondingly generate a plurality of environmental monitoring parameter data, such as temperature, relative humidity, sunshine amount, wind speed, and rainfall. Monitoring parameters such as volume, or global location. The sensing node devices 110 can be separately connected to a cluster animal automatic counting device 120 for counting the number of cluster pests of the monitoring points where the sensing node devices 110 are located, for example, the number of oriental fruit flies.

The cluster animal quantity automatic counting device 120 can induce the oriental fruit fly to a trap space to trap and prevent it from escaping, and count the number of oriental fruit flies entering the trap space one by one. The detailed structure of the cluster animal quantity automatic counting device 120 has been disclosed in the present applicant. The previously proposed Republic of China patent application "multi-level card sensing type of animal automatic counting device", so this detailed structure and function will not be further explained.

In an embodiment, the WSN wireless sensing network 100 adopts multi-hop linking and routing method to transmit the environmental monitoring parameter data sensed by each sensing node device 110 to the front gate. Track device 200. The multi-layer hopping and routing method connects all the sensing node devices 110 in the WSN wireless sensing network 100 into an integrated network architecture, such as a ring, tree, star, and other network architecture. The invention is not limited to this.

The front-end gateway device 200 is used to provide a gateway function for data exchange between the WSN wireless sensing network 100 and the back-end master servo device 300. The WSN wireless sensing network 100 can also be collected. The environmental monitoring parameter data sensed by each sensing node device 110 is transmitted to the backend master control device 300 through the wireless communication system 20 in a wireless manner. The backend master control device 300 can also transmit control commands (including user-specified control commands and automatically generated control commands) to the WSN wireless sensing network 100 through the wireless communication system 20. Node device 110.

In an embodiment, the front-end gateway device 200 can have a GPS (Global Positioning System) geolocation function, which can automatically detect the global geographical position (ie, longitude and latitude) where the front-end gateway device 200 is located, and Correspondingly, a set of electronic type geographic location parameter data is generated. The front-end gateway device 200 can also optionally include the built-in sensing of the machine. Function, such as temperature, relative humidity, amount of sunshine, wind speed, rainfall, and other environmental condition data. The user can selectively set the collection of the environmental monitoring parameter data by the front-end gateway device 200 or the sensing node devices 110 in the WSN wireless sensing network 100 according to the requirements. In the transmission of the data, the front-end gateway device 200 integrates and formats the environmental monitoring parameter data into a predetermined wireless communication data transmission format packet, such as a standardized SMS (Short Message Service) message format or GPRS. The (General Packet Radio Service) wireless packet format is used to transmit the packet to the backend master control device 300 via the wireless communication system 20.

In an embodiment, the front-end gateway device 200 first executes a wireless sensing network deployment program to connect all the sensing node devices 110 in the WSN wireless sensing network 100 to each other. (preferably a tree network architecture), the other sensing node devices 110 can sense the sensing node devices 110 through a multi-layer linking and routing method through the network architecture. The measured environmental monitoring parameter data is transmitted to the front-end gateway device 200, and the connection and data communication between the front-end gateway device 200 and all the sensing node devices 110 in the WSN wireless sensing network 100 adopts S-MAC. (Sensor Media Access Control) sensor media access control specification.

In addition, in another embodiment, the front-end gateway device 200 can perform a time synchronization process on all the sensing node devices 110 in the WSN wireless sensing network 100 before the actual operation, so that all the sensing node devices are used. 110 built-in timing functions are set to the same time to provide the same time base The environmental monitoring parameter data generated by each sensing node device 110 can be located on the same time reference. The foregoing time synchronization program is a time synchronization method such as RBS (Reference Broadcast Synchronization) or TPSN (Timing-sync Protocol for Sensor Networks).

Moreover, the front-end gateway device 200 further includes an automatic routing function, which can cause a fault in any of the sensing node devices in the WSN wireless sensing network 100, thereby causing other devices connected to the faulty sensing node device. When the sensing node device cannot normally return the sensing data, the situation can be detected by the front-end gateway device 200, and the automatic routing function is executed to construct the WSN wireless sensing network 100. The routing method performs a re-routing procedure, and the routing path of the other sensing node devices connected to the faulty sensing node device is changed to be connected to other good sensing node devices, so that the sensing node devices can The sensor data is again normally returned to the front-end gateway device 200.

In an embodiment, the front-end gateway device 200 can be implemented by using a personal computer platform or a programmable embedded microprocessor; for example, the implementation of the personal computer platform has the advantages of better scalability, but disadvantages. For more power-consuming implementations, such as the implementation of a programmable embedded microprocessor, the advantage is that it saves power, but the scalability is worse than that of a personal computer platform.

Please refer to FIG. 2A at the same time, which is a schematic diagram of the architecture of the main control servo device at the rear end of the present invention. The backend master control device 300 is a personal computer platform with a microprocessor, and includes a database module 301, a kind of god The network training module 302, a classification model training module 303 and a parameter data normalization module 305 are provided. The database module 301 receives and stores the environmental monitoring parameter data through the wireless communication system 20. In an embodiment, the database module 301 can be a MySQL database.

The neural network training module 302 is coupled to the database module 301 to read the environmental monitoring parameter data, and input the environmental monitoring parameter data into a neural network for training to be output by the neural network. Multiple monitoring events. The neural network training module 302 adopts a self-organizing map (SOM) network training method of an unsupervised learning network. The so-called unsupervised learning network is to obtain learning examples from external materials, to find intrinsic rules from the learning paradigm, and then to use the internal rules to find new rules. The SOM network is a kind of neural network based on the Competitive Learning rule. The competitive learning method is a self-organizing learning method. It searches for a group of unlabeled samples. Features are integrated into the same class. In a general neural network competition, only one neuron will be activated to an active state, while other neurons will be suppressed to a resting state, and when the competition is over, only the winning neurons can learn. After the SOM network competes for learning, because SOM incorporates the concept of a neighborhood, more than one winning neuron can learn, but neurons in the vicinity of the winning neuron are eligible to learn. Therefore, in SOM, the output layers are usually arranged in a matrix in a two-dimensional space (also applicable to high dimensions), competing with each other according to the input vector, in order to obtain the right to be activated, and finally based on the input vector. The feature is topological in the output space. The biggest difference between the SOM network and the general neural network is that the SOM network has only the input layer and the output layer, and there is no hidden layer. The input layer is used to input the data for training. In the present invention, the complex array multidimensional coordinate vector data composed of the environmental monitoring parameter data is input, and after performing the dimensional reduction operation, the output layer outputs one having at least two dimensions. In the above network topology, the network topology is a coordinate system composed of output layer-like neurons, which represents the clustering of the input vectors, that is, the relative positions of the output layer-like neurons are meaningful, which makes the SOM network and other The neural network has the biggest difference. Therefore, in the present invention, the network topology output by the output layer includes a plurality of monitoring events clustered according to the environmental monitoring parameter data, such as normal operation, sensor failure, or pest occurrence.

Since the neurons output by the SOM network are presented in a meaningful topology in the output space, the neurons compete with each other to obtain the activated right. The SOM network calculation step is as follows, firstly, the input data is defined to exist in the space R ⁿ , the output layer topological coordinates are set to two dimensions (may also be multi-dimensional), the number of neurons is n , and the initial link weight is randomly given. The value m _i =[ μ _{i 1} , μ _{i 2} ,..., μ _in ] ^T R ⁿ , also extract some data from the sample data as training data x =[ λ ₁ , λ ₂ ,..., λ _n ] ^T R ⁿ . In an embodiment of the present invention, the input vector may be a temperature having a time parameter, a relative humidity, an illumination, and a pest number, and is defined as x ( i , t )= [ T ( i , t ), H ( i , t ), I ( i , t ) P ( i , t )], and the input temperature range can be from 5 to 50 degrees C, and the relative humidity range can be from 35~ 95%, and the amount of sunshine can be from 0 to 80000 lumens (Lux). However, since the value of the sunshine quantity parameter is larger than other values, before inputting the neural network training module 302, the parameter data normalization module 305 must be used to logarithmically calculate the numerical value of the sunshine quantity; According to the number of pests to analyze the pests are occurring, the number of pests is one of the most important parameters in the analysis process. In order to make the network topology results output by the final output layer easy to operate normally or the pests are happening, the characteristics are more To be obvious, the value of the number of pests must first be squared by the parameter data normalization module 305 to the number of pests, and then input to the neural network training module 302. Therefore, the vector input in one embodiment of the present invention is x ( i , t )=[ T ( i , t ), H ( i , t ), log I ( i , t ), p ² ( i , t )], in this way, it can highlight the network topology after the neural network operation, which represents the clustering of the input vector. Only four parameters are listed as input vectors in this embodiment, but the invention is not limited thereto.

After the parameter vector adjusted by the parameter data normalization module 305 is input into the neural network training module 302, each neuron in the network is compared with each other to obtain the right to be activated, and the reference system is compared. The degree of similarity to the input data shall prevail. And calculating the distance between the input training data and each initial link weight value in the network topology, the mathematical formula is Dis ( x , j )=∥ x ( i , k )- m _j ∥, Dis ( x , j ) This is a commonly used measure of similarity for the Euclidean Distance. And the mathematical formula for determining the winning neurons is . Where c is the winner (Winner), k represents the number of trainings, indicating the neuron most similar to the input data x , and the highest similarity indicates that the input vector has the shortest Euclidean distance from the corresponding neuron. After finding the winner, the connection weighting value will be updated, that is, in addition to the winner update, the weighted values of all the neuron in the vicinity of the winner will be updated, and the mathematical expression is m _j ( k +1 _{) = m j (k) +} α (k) h cj (k) [x (k) - m j (k)], which is adjacent to the function h _cj, α (k) is the learning rate. For the purpose of convergence, the proximity function decreases with the number of trainings. The present invention uses the Gaussian function as the proximity function, which is defined as follows: Where r _j and r _c refer to the other neurons in the adjacent region and the winning neurons, respectively, ∥ r _j - r _c ∥ refers to the distance between the two neurons, and σ is the range of the adjacent region (region) . Therefore, the results of the training will gradually reduce the size of the adjacent area as time goes by. Please refer to FIG. 4A. FIG. 4A is a schematic diagram of a network topology for generating a plurality of monitoring events. The network topology 40 includes areas of several sensing area failures (i.e., SF), areas where several pests are occurring (i.e., PO), and areas of normal operation (NS). Because of the SOM network calculus, it can be seen that the same monitoring events are clustered together, making the network topology 40 contain meaningful structures.

After the neural network training module 302 generates the network topology of the plurality of monitoring events that can be distinguished, in addition to the artificial method, the automatic classification method can be adopted, and the classification model training module 303 is reused. The classification of these monitoring events is conducted to obtain a plurality of monitoring results. The classification model training module 303 adopts a calculation method of a support vector machine (SVM), which performs classification training according to a network topology with a plurality of monitoring events, and generates at least one or more nonlinear boundary. The Hyper Plane is used to segment the plurality of monitoring events within the network topology to obtain the monitoring results.

The calculation method of the so-called support vector machine is to calculate the "+1" or "-1" respectively for the converted input vector and its associated input vector [ r _n , y _n ], [11] Is Where j=1, 2, ... , N , N are the number of hyperplanes, and the hyperplane can divide the network topology generated by the neural network training module 302 into several parts, and w _j and b _j are weight vectors (vector of weight) and the jth offset. By listing the data as a positive category (+1) or a negative category (-1), the data can be split into two categories by hyperplane. Another SVM can propose a calculation method with a maximum margin (maxinizing margin): . r _n 〉+ b _j ±1 for y _n =±1 n . In this embodiment, the monitoring events of three normal operations, sensor failures, or pests are taken as an example, and the monitoring vector machine generates at least one hyperplane having a nonlinear boundary, which can be monitored. The event is divided. For example, a hyperplane of two different boundaries can be generated, as shown in Figures 4B and 4C, respectively. Referring to FIG. 4B, the network topology 40 divides the area of the sensor fault (ie, SF) with a hyperplane 401, and another area separated by the hyperplane 401 contains the pest (ie, PO). And the area of normal operation (ie, NS), and the hyperplane 401 is a hyperplane of a nonlinear boundary; see FIG. 4C, the network topology 40 is an area that will operate normally (ie, NS) with a hyperplane 401 The other region segmented by the hyperplane 401 contains an area where pests are occurring (i.e., PO) and sensor failure (i.e., SF), and the hyperplane 401 is also a hyperplane of a nonlinear boundary. Of course, the present invention is not limited to the above-mentioned monitoring result, and the monitoring result of the sensor fault can be further divided into the sensor fault of the sensing node device where the device is located, or the segmentation is different according to the season. Monitoring results. Since the parameter data before the input neural network training module 302 is adjusted by the parameter data normalization module 305, the network topology trained by the neural network training module 302 can make the classification model. The support vector machine adopted by the training module 303 can more accurately generate at least one hyperplane with a nonlinear boundary, so that the error value of the segmented monitoring result can be minimized. By the super-plane generated by the classification model training module 303, the network topology segmentation of the monitoring result can be transformed into a classification model, which can be used as a basis for subsequent diagnostic environment monitoring parameter data. If you input other environmental monitoring parameter data that has not been trained by the system, you can diagnose which monitoring result is located in the classification model. The classification model can also produce different classification models for supply according to the season.

The training method of the self-organizing map (SOM) network of the unsupervised learning network and the calculation method of the support vector machine (SVM) are based on the LabVIEW software. Written. The present invention is accomplished with LabVIEW in conjunction with a MySQL database and a GSM network. The MySQL database also provides a web interface that allows users to directly browse or download the monitored data over the network.

In another embodiment of the present invention, referring to FIG. 2B , the back-end main control servo device 300 further includes a diagnostic module 308 , an alert module 304 , and a wireless communication module 306 . After the classification model training module 303 completes the segmentation of the plurality of monitoring results, the diagnostic module 308 of the classification model training module 303 is connected, and the untrained environment can be obtained according to the monitoring results. The monitoring parameter data is judged to obtain the judgment result corresponding to the untrained environmental monitoring parameter data in the monitoring results, and the warning module 304 can generate a corresponding warning message according to the determination result. For example, the foregoing segmentation results of a monitoring result of the occurrence and normal operation of the pest, and another classification model with the monitoring result of the sensor failure, when the untrained environmental monitoring parameter data is fed into the classification model, if The trained environmental monitoring parameter data is classified into the categories of "insect pests are occurring" and "sensor faults", which can generate warning messages such as "pest pests are occurring" and "sensor faults"; for example, segmentation has one Monitoring results of pests or sensor failures and monitoring results of normal operation. If the untrained environmental monitoring parameters are classified as "system anomalies or pests are occurring" and "normal operation" In the case of categories, warning messages of "system anomalies or pests are occurring" and "normal operation" may be generated separately. The present invention is not limited thereto, and it is to be seen which category the untrained environmental monitoring parameter data is classified into. , that is, send a warning message to be conveyed in this category. The warning message format can be a standard SMS (Short Message Service) message format or a GPRS (General Packet Radio Service) wireless packet format, so the alert module 304 can pass through a wireless communication module 306 that is connected to the wireless communication system 20. The warning message is transmitted to a wireless communication device 30, thereby achieving the effects of remote monitoring, diagnosis, and warning. The wireless communication device 30 can be a mobile phone.

In an embodiment, please refer to FIG. 2B again. The backend master control device 300 further includes a parameter data checking module 307. The parameter data checking module 307 checks the environmental monitoring parameters read from the database module 301. Whether the data has been trained by the neural network training module 302 or whether new data is stored in the database module 301. If the parameter data is untrained, the parameter data is input to the neural network training module 302 for training; if the parameter data is trained, the warning module 304 can send a warning message to remind The manager checks; if no new data is stored in the database module 301, there may be problems such as network disconnection, sensor failure, etc., and the database module 301 has no new data. The warning message is sent through the warning module 304; if there is new data, the diagnosis module 308 can be used for classification. Such a mechanism can maintain the neural network training module 302 to perform training according to the new parameter data, and can also be used as a method for checking whether the data is abnormal. For example, the original system sets new data into the database every 30 minutes. If the parameter data checking module 307 has not read the new data from the database module 301 after a certain period of time, the system may also have problems, such as the wireless communication system 20 disconnection, the front-end gateway device 200 failure, and the like. At this time, the parameter data checking module 307 can send an alert message through the alert module 304 to remind the administrator to perform the check.

Please refer to FIG. 3, which is a flow chart of a method for classifying remote environmental data according to the present invention, which is applied to the monitoring of crop pests. Step S02 is to form a plurality of environmental monitoring parameter data into a multi-array multi-dimensional coordinate vector data, and the environmental monitoring parameter data is read from a database, and the environmental monitoring parameter data includes pest quantity, temperature, relative humidity, and sunshine. Volume, wind speed, direction, rainfall, or global location. In an embodiment of the present invention, the plurality of environmental monitoring parameter data is composed of multiple complex arrays Before dimension vector data, the number of pests in the environmental monitoring parameter data must be squared and the value of the sunshine quantity should be logarithmically calculated, and then the multi-dimensional coordinate vector data of the complex array is formed, as shown in step S01. First, the number of pests and the amount of sunshine are processed first, in order to highlight the clustering of the input vectors in the network topology after the neural network operation.

After the complex array multidimensional coordinate vector data is formed, the training is input into a type of neural network, and after the training is completed, the plurality of monitoring events are outputted by the neural network, as shown in step S03. In this type of neural network, a training method of a self-organizing map network of an unsupervised learning network is adopted, and dimension reduction operations are performed on the multidimensional coordinate vector data to reduce the vector dimension of the multidimensional coordinate vector data. Generating a network topology having at least two dimensions, the network topology including the monitoring events. The calculation formula of this training method is as described above, and will not be described again here.

After generating a network topology that includes a plurality of monitoring events, in addition to the artificial way of distinguishing, an automated classification can be adopted. In step S04, the monitoring events are classified according to a classification model to generate a plurality of monitoring results. The classification model is based on the calculation method of the support vector machine and is trained based on the monitoring events. The classification model may generate at least one hyperplane having a nonlinear boundary according to the network topology to segment a plurality of monitoring events in the network topology to obtain the monitoring results.

In this embodiment, the monitoring events of three normal operations, sensor failures, or pests are taken as an example, and the monitoring vector machine generates at least one hyperplane having a nonlinear boundary, and the monitoring events are performed. Divide Cutting, produces a classification model that can be used for diagnosis. As shown in FIG. 4C, for example, a hyperplane 401 is generated, which divides the network topology 40 into a monitoring result of pest occurrence (ie, PO) and sensor failure (ie, SF), and another has normal operation. (i.e., NS) monitoring results; or as shown in Figure 4B, another hyperplane 401 is generated that separates the network topology 40 from a monitoring result of pest occurrence (i.e., PO) and normal operation (i.e., NS). And another monitoring result with a sensor fault (ie SF).

After the monitoring result is obtained, the monitoring result can be judged, and the untrained environmental monitoring parameter data is judged, and the judgment result is generated (step S05). A warning message is generated corresponding to the result of the judgment, so that the warning message is transmitted to the wireless communication device via the wireless communication system, as shown in step S06. For example, based on the classification model of the monitoring results of "pest disease is occurring", "sensor failure" or "normal operation", the untrained environmental monitoring parameter data is judged to generate the untrained environmental monitoring parameter data system. What kind of monitoring results are used? If the untrained environmental monitoring parameter data is a monitoring result of "pest is happening", a warning message "Pests are occurring" will be generated. The warning message format can be a standardized SMS (Short Message Service) message format or a GPRS (General Packet Radio Service) wireless packet format, so the warning message can be transmitted to a wireless communication device such as a mobile phone via a wireless communication system, thereby reaching a far distance. The effect of monitoring, diagnosis, and warning.

In another embodiment, the remote environment data classification method of the present invention can also achieve the effects of automatic diagnosis and learning. When reading environmental monitoring parameter data from the database, it can be judged whether it is the latest resource for the data. Material, or check for information. If it is not the latest information or no data, you can immediately send a warning message to the wireless communication device to remind the administrator to pay attention. If it is the latest data, the training of the neural network and the classification model of the above steps S01 to S04 is started to obtain a plurality of monitoring results. The monitoring results include "Pests are happening", "Sensor failure" or "Normal operation", etc., then you can proceed to steps S05~S06 to determine whether the system is operating normally. If not, send warning messages; or judge pests. Whether it occurs or not, if it is happening, send a warning message.

According to the system and method of the present invention, the self-organizing map network learning of the unsupervised learning network and the classification of the support vector machine can be performed according to the collected data, and several monitoring results are generated, and according to the These monitoring results judge the untrained environmental monitoring parameter data to generate the judgment result and generate a warning message correspondingly. Such a system and method is more suitable for automatically detecting the degree of damage to the oriental fruit fly in the orchard, to issue an alarm message, and to provide farmers with more precise management information, which can achieve the effect of agricultural automation. The warning message "Pests are happening" is also used as a basis for farmers to spray pesticides to prevent the oriental fruit flies from destroying the farmers' crops and indirectly improving and improving the yield and quality of fruits.

The above is only the preferred embodiment of the present invention, and is not intended to limit the scope of the technical content of the present invention. The technical contents of the present invention are broadly defined in the following claims. Any technical entity or method performed by any other person that is identical to, or equivalent to, the ones defined in the scope of the claims below will be considered to be included in the scope of the invention.

300‧‧‧Backend master control servo

301‧‧‧Database Module

302‧‧‧ class neural network training module

303‧‧‧Classification Model Training Module

304‧‧‧ Warning Module

305‧‧‧Parameter data normalization module

306‧‧‧Wireless communication module

307‧‧‧Parameter data inspection module

308‧‧‧Diagnostic Module

Claims (10)

A remote environment data classification system, comprising: a sensing network, comprising a plurality of sensing node devices, respectively disposed at a plurality of monitoring points in the remote area to sense the condition of the surrounding environment of the monitoring points, Correspondingly generating a plurality of environmental monitoring parameter data; the front-end gateway device is connected to the sensing network for receiving the environmental monitoring parameter data to transmit the environmental monitoring parameter data through the wireless communication system; and the back-end main The control servo device comprises: a database module, which receives and stores the environmental monitoring parameter data through the wireless communication system; the neural network training module connects the database module to read the environmental monitoring Parameter data, the environmental monitoring parameter data is input into the neural network for training, and the plurality of monitoring events are outputted by the neural network; and the classification model training module is used to classify the monitoring events to obtain Multiple monitoring results. The remote environmental data classification system according to claim 1, wherein the environmental monitoring parameter data includes the number of pests, temperature, relative humidity, amount of sunshine, wind speed, wind direction, rainfall, or global geographic location. For example, the remote environmental data classification system described in claim 2, wherein the back-end main control servo device comprises a formal parameter data The parameterization module of the parameter data firstly squares the value of the number of pests and calculates the value of the amount of sunshine, and then inputs the neural network training module. For example, the remote environmental data classification system described in claim 1 wherein the neural network training module is a self-organizing map network training mode of an unsupervised learning network, the unsupervised learning The self-organizing map network of the network comprises an input layer and an output layer, wherein the input layer is used to input a multi-array coordinate vector data composed of the environmental monitoring parameter data, and the output layer output is performed after the dimension reduction operation is performed. A network topology having at least two dimensions, the network topology including the monitoring events. The remote environment data classification system according to claim 4, wherein the classification model training module adopts a calculation method of a support vector machine, and the support vector machine generates at least one or more according to the network topology. The hyperplane of the nonlinear boundary is used to segment the monitoring events within the network topology to obtain the monitoring results. The remote environmental data classification system of claim 1, wherein the back-end main control servo device comprises: a diagnostic module, wherein the untrained environmental monitoring parameter data is determined according to the monitoring results. And obtaining a warning message corresponding to the untrained environmental monitoring parameter data in the monitoring results; and the warning module generates a corresponding warning message according to the determination result, and transmits the warning message to the wireless communication system Transfer to wireless On the communication device. A remote environment data classification method comprises: (a) composing a plurality of environmental monitoring parameter data into a complex array multidimensional coordinate vector data, and then inputting the multidimensional coordinate vector data into a neural network for training, after the training is completed, The neural network outputs a plurality of monitoring events; (b) classifying the monitoring events according to the classification model to generate a plurality of monitoring results; (c) judging the untrained environmental monitoring parameter data by using the monitoring results And generating a determination message; and (d) generating a warning message corresponding to the result of the determination to transmit the warning message to the wireless communication device via the wireless communication system. For example, the remote environment data classification method described in claim 7 is characterized in that the neural network in step (a) is a training method of the self-organizing map network of the unsupervised learning network, The multidimensional coordinate vector data performs a dimensional reduction operation and generates a network topology having at least two dimensions, the network topology includes the monitoring events, and the classification model in the step (b) adopts a calculation of the support vector machine And obtaining, according to the network topology, at least one hyperplane having a nonlinear boundary to segment the monitoring events in the network topology to obtain the monitoring results. The method for classifying remote environmental data as described in claim 7 wherein the environmental monitoring parameter data includes the number of pests, temperature, relative humidity, amount of sunshine, wind speed, wind direction, rainfall, or Global location. The remote environmental data classification method according to claim 9 , wherein before the step (a), the method further comprises: square the value of the pest and calculating the numerical value of the sunshine amount, and then calculating The multidimensional coordinate vector data is composed to input the steps of training in the neural network of the type.