CN102968664A - Mobile robot variable pesticide applying intelligent decision-making method and system - Google Patents

Abstract

The invention relates to a mobile robot variable pesticide applying intelligent decision-making method and a system. The mobile robot variable pesticide applying intelligent decision-making system comprises a camera, an ultrasonic sensor, a flow sensor and a computer, the camera is used for sending an acquired greenhouse crop image to the computer through a universal serial bus (USB) interface, the ultrasonic sensor is used for measuring the distance of a sprayed target and sending the information of the distance to the computer, the flow sensor is used for measuring actual sprayed pesticide flow, the computer is connected with the camera, the ultrasonic sensor and the flow sensor, a fuzzy neural network decision-making device is configured in the computer, the fuzzy neural network decision-making device is used for obtaining target area, spraying distance and plant diseases and insect pests grade information, preset pesticide applying amount is calculated according to the obtained target area, spraying distance and plant diseases and insect pests grade information, a control amount is calculated and output according to the preset pesticide applying amount, the actual pesticide applying amount is obtained, the actual pesticide applying amount is compared with the preset pesticide applying amount, and the control amount is adjusted.

Description

A kind of mobile robot's variable farm chemical applying Using Intelligent Decision-making Method and system

Technical field

The present invention relates to the application of fuzzy neural network, relate in particular to mobile robot's variable farm chemical applying Using Intelligent Decision-making Method and system for the greenhouse band crop.

Background technology

Extensive mode is mainly taked in current agricultural chemicals dispenser, and the dispenser scheme is mostly formulated according to experience by the operating personnel, and workload is large, and is subjected to subjective factor affecting.For a long time, the technique in using agriculture chemical of China still rests on large capacity, the large droplet spray technique level, and the agricultural chemicals utilization factor is low, easily causes the dispenser personnel to poison, and has seriously restricted China's sustainable development of agriculture.Therefore, active demand variable farm chemical applying technology in agricultural production process is to improve the agricultural chemicals utilization factor and to reduce agricultural chemicals to operating personnel's injury.The greenhouse is as the important component part of modern installations agricultural, and its environment is airtight, temperature is high, high humidity, easily cause that the dispenser personnel poison in the canopy.For greenhouse band crop growth pattern and plantation characteristics carry out the research of variable farm chemical applying intelligent decision to effective minimizing agricultural chemicals waste, improve product yield and quality, the assurance operating personnel is healthy has great importance.

Document " based on the variable farm chemical applying control system of fuzzy control " (Shao Lushou etc., agricultural mechanical journal. the 36th volume o. 11th, 110-112 page or leaf) a kind of fuzzy Decision Making Method proposed.But fuzzy system lacks self study and adaptive ability, is not suitable for the changeable greenhouse dispenser of environmental factor; Document " design of variable spray medicine adaptive neural-networks fuzzy controller and emulation " (old rear people etc., mechanical facilities of irrigation and drainage engineering journal .2011 03 phase of the 29th volume, the 272-276 page or leaf) designed a kind of dual input based on the adaptive neural network fuzzy reasoning, single o controller. take weeds area and the speed of a motor vehicle as input quantity, spray value is output quantity, select gaussmf type subordinate function, utilize measured data that the initial fuzzy inference system that generates is trained, but its essence or a fuzzy control, just adopt neural network theory to extrapolate 5 fuzzy rules, and the variation of its travel speed easily causes spraying the medicine system pressure and changes, and causes the dispenser performance to be difficult to control.

Application number be CN201010239778.X, the name be called " greenhouse is to target dispenser robot system " Patent Application Publication a kind of for the greenhouse to target dispenser robot system.To utilize machine vision to obtain disease information and be passed to industrial computer to target dispenser robot, industrial computer sends instruction as host computer to PLC, when PLC receives the disease positional information that host computer sends, start moving chassis truck and move to the disease ridge, mechanical arm drops to the predefine position with spray boom, the camera that simultaneously PLC triggering is installed on the line slideway carries out information acquisition to current disease ridge, host computer transmits PLC with result, PLC is in spray boom motion process from left to right, in certain nozzle enters the disease occurrence scope, control this nozzle and open, otherwise close; Nozzle needs the single or multiple unit of target are carried out operation in single pass, realizes the single or multiple open and close controlling, realizes target spraying.This invention focuses on product structure design, does not relate to for the decision making algorithm.In addition, consult domestic Patent aspect variable farm chemical applying as seen, the variable farm chemical applying patent is the new structure design of various variable farm chemical applying system, so far, has no pertinent literature variable farm chemical applying is proposed more satisfactory Decision Control method.

Can find out that from above-mentioned correlation technique existing research has obtained some achievements, but have certain limitation.Fuzzy Decision Making Method lacks self-learning capability, in the variable farm chemical applying decision system, can not adjust online decision rule, and select with certain subjectivity at the subordinate function of decision-making device input quantity output quantity, adaptive ability is limited, is difficult to obtain desirable variable farm chemical applying effect; Though the Decision of Neural Network method has powerful self-learning capability, it can not utilize existing experimental knowledge well in variable spray medicine decision process, and the mode of knowledge representation and processing is difficult for understanding and use for operating personnel.In addition, above-mentioned correlation technique is failed effective integration crop area to be painted, disease and pest and range information etc. and is conducive to the decision information of variable farm chemical applying, have stronger probabilistic autonomous mobile robot for working in the greenhouse, be difficult to obtain pesticide supplying effect accurate, efficient, environmental protection.

Summary of the invention

For overcoming above-mentioned the deficiencies in the prior art, the present invention is directed to Greenhouse Robot and designed a kind of intelligent decision system based on fuzzy neural network, according to the target characteristic processing information of obtaining, calculate corresponding flow by the fuzzy neural network decision-making device, thereby realize the amount of spraying with the variable farm chemical applying of target characteristic information real-time change, be used for realizing that the decision-making of greenhouse variable farm chemical applying can obtain better pesticide supplying effect.The present invention is applied to the fuzzy neural network decision system of fuzzy control and neural network combined together formation in the greenhouse variable farm chemical applying decision-making first, realize that by neural network fuzzy logic has improved the precision of variable farm chemical applying decision system, utilize simultaneously the self-learning capability of neural network, adjust subordinate function so that the decision making system has stronger adaptive ability.

According to a kind of mobile robot's variable farm chemical applying of the object of the invention Using Intelligent Decision-making Method, comprising:

Obtain the target area, spray distance and be and the disease and pest class information; According to the target area that obtains, spray distance for and the disease and pest class information, calculate default formulation rate; Calculate and the output controlled quentity controlled variable according to default formulation rate; Obtain actual formulation rate, actual formulation rate and default formulation rate are compared, adjust controlled quentity controlled variable.

In the technique scheme, the default formulation rate step of described calculating comprises:

Set up the fuzzy neural network step, comprising:

Structure take the target area, spray distance for and the fuzzy neural network of disease and pest grade as inputting;

Learning procedure comprises:

The fuzzy neural network that makes up is carried out off-line training, obtain every weights and the threshold value of fuzzy neural network;

Fuzzy neural network after the training is carried out decision accuracy to be detected;

The fuzzy neural network that obtains training;

Determining step comprises:

The fuzzy neural network that the programming realization trains in mobile robot's computer system;

The fuzzy neural network that trains of operation, according to the target area that obtains, spray distance for and the disease and pest class information,

Judge default formulation rate.

As a further improvement on the present invention, describedly judge default formulation rate step and specifically comprise:

Take the target area, spray distance and 3 input variables of the sick worm extent of injury as the input neuron of network;

The subordinate function of analog input variable calculates degree of membership corresponding to each input variable;

The coupling fuzzy rule, the former piece that carries out every fuzzy rule calculates;

Calculate the relevance grade of every fuzzy rule;

Calculate the amount of spraying.

According to a kind of mobile robot's variable farm chemical applying intelligent decision system of the present invention, comprising:

Camera is used for by USB interface the chamber crop image that collects being delivered to computing machine;

Ultrasonic sensor is used for recording the distance that sprays target and being sent to computing machine;

Flow sensor is used for measuring actual spray medicine flow;

Computing machine, it is connected with camera, ultrasonic sensor and flow sensor, wherein disposes the fuzzy neural network decision-making device, is used for:

Obtain the target area, spray distance and be and the disease and pest class information;

According to the target area that obtains, spray distance for and the disease and pest class information, calculate default formulation rate;

Calculate and the output controlled quentity controlled variable according to default formulation rate;

Obtain actual formulation rate, actual formulation rate and default formulation rate are compared, adjust controlled quentity controlled variable.

As a further improvement on the present invention, described fuzzy neural network decision-making device adopts five layers of structure of fuzzy neural network, and ground floor is input layer, take the target area, spray distance, 3 input variables of the sick worm extent of injury as the input neuron of network; The second layer is the obfuscation input layer, is used for calculating degree of membership corresponding to each input variable; The 3rd layer is the fuzzy rule layer, and the former piece that is used for every fuzzy rule calculates; The 4th layer is the obfuscation output layer, is used for calculating the relevance grade of every rule; Layer 5 is output layer, is used for calculating the amount of spraying.

The present invention is applied to fuzzy-neural network method greenhouse spray medicine mobile robot's dispenser decision making first, and its characteristics and beneficial effect are:

1, for greenhouse embark on journey planting patterns, growth characteristic and the disease and pest grade situation of plant, institute's invention intelligent decision system not only merges target target area information, range information but also merges the disease and pest class information.

The variable farm chemical applying decision system based on fuzzy neural network of 2, inventing, adopt neural network that the fuzzy decision rule is carried out rapid extraction, solved the difficult problem that fuzzy rule is difficult to obtain in the variable farm chemical applying decision system, utilize simultaneously the self-learning capability of neural network, constantly improve decision rule, improve constantly the decision accuracy of system, adopted the fuzzy neural network decision-making device can effectively control the precision of formulation rate, satisfied the needs of greenhouse variable farm chemical applying.

3, institute's invention fuzzy neural network decision-making technique owing to adopt the mode of off-line training, therefore can realize the real-time decision-making of dispenser system; Because designed variable farm chemical applying intelligent decision system has stronger adaptive faculty and generalization ability, therefore can adapt to better and have stronger probabilistic greenhouse working environment.

4, designed dispenser decision making system can reduce the use amount of agricultural chemicals, improves the service efficiency of agricultural chemicals, environmental contamination reduction.

Description of drawings

Fig. 1 is greenhouse band crop variable farm chemical applying system construction drawing.

Fig. 2 is variable farm chemical applying decision system block diagram.

Fig. 3 is structure of fuzzy neural network figure.

Fig. 4 is the intelligent decision process flow diagram.

Fig. 5 is the network training result.

Fig. 6 is the actual amount of spraying and the expectation amount of spraying error.

Fig. 7 is the part sample data.

Embodiment

Such as Fig. 1, at first make up and merge the off-line training fuzzy neural network decision-making device with self study and adaptive ability that sprays target range, area and disease and pest class information; Secondly, collect the on-the-spot true-time operation data of greenhouse dispenser composing training sample pair, choose therefrom that 2/3 data are carried out off-line training as training sample to the fuzznet that makes up and with 1/3 the data sample of reserving as detecting sample to training rear fuzzy neural network to carry out the decision accuracy detection; Then in variable farm chemical applying mobile robot's computer system, the fuzzy neural network that the programming realization trains detects target area to be sprayed, range information, and the disease and pest grade n that manually inputs hothouse plants by computer user interface; On this basis, decision-making device goes out accordingly according to the target information calculations obtained that the amount of spraying q(unit is Ls ^-1), record actual flow by controller and flow sensor and carry out computing and comparison, and regulate in real time the PWM dutycycle according to comparative result, drive solenoid valve, realize that variable sprays.Concrete grammar is as follows:

1, makes up the off-line training fuzzy neural network with self study and adaptive ability that fusion sprays target range, area and disease and pest class information

As shown in Figure 2, the designed fuzzy neural network of the present invention comprises: neural network module, for obfuscation, fuzzy reasoning, fuzzy decision rule and the ambiguity solution of realizing input quantity; The off-line learning module is used for carrying out the study of described neural network module, learns the weights of each network node.

The present invention adopts five layers of structure of fuzzy neural network, represents such as Fig. 3.Ground floor is input layer: target area s(unit is m ²), to spray apart from d(unit be m), sick worm extent of injury n, 3 input variables are the input neuron of network; The second layer is the obfuscation input layer: calculate degree of membership corresponding to each input variable; The 3rd layer is the fuzzy rule layer: the former piece of realizing every rule calculates; The 4th layer is the obfuscation output layer: the relevance grade of calculating every rule; Layer 5 is output layer (de-fuzzy layer): calculate the amount of spraying.

The ground floor input layer.3 nodes are arranged, and corresponding target area s(unit is m respectively ²), to spray apart from d(unit be m), sick worm extent of injury n.Each neuron represents an input signal in this layer, and neuron number equals the variable number that occurs in the fuzzy rule prerequisite, and it is input vector X=(x ₁, x ₂, x ₃... x _n) ^TBe directly passed to lower one deck.I the unit x of i neuron and input variable X _iLink to each other;

Second layer obfuscation input layer.8 nodes are arranged, represent 8 membership functions, finish asking for of membership function value.Each node represents a linguistic variable: area, the m of unit ², give 3 linguistic variable values, i.e. " greatly ", " in ", " little "; Distance, the m of unit gives 3 linguistic variable values, i.e. " closely ", " in ", " far "; The disease and pest extent of injury is given 2 linguistic variable values, i.e. " seriously ", " not serious ".Each neuron in this one deck is used for a subordinate function of analog input variable, and its effect is to calculate respectively to input the membership function that component belongs to each linguistic variable value fuzzy set.Adopt Gaussian function as subordinate function, that is:

$y=e^{-\frac{{(x-c)}^2}{{\mathrm{\σ}}^2}---\left(1\right)}$

$I_{\mathrm{isi}}^{\left(2\right)}=-\frac{{(x_i^1-c_{\mathrm{isi}})}^2}{{\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="2"\; label="isi"\; filenames="121010164251.png,121010164256.png"\; state="\{\{state\}\}">isi</figure-callout>}}^2},$ $O_{\mathrm{isi}}^{\left(2\right)}={\mathrm{\&mu;}}_i=e^{-\frac{{(x_1^{\left(1\right)}-c_{\mathrm{isi}})}^2}{{\mathrm{\&sigma;}}_{\mathrm{isi}}^2}}---\left(2\right)$

Wherein, c _Isi, o _IsiThe center and the width that represent respectively membership function;

The 3rd layer of fuzzy rule layer.18 nodes are arranged, represent 18 fuzzy rules, the former piece of finishing fuzzy rule calculates.Its each node represents a fuzzy rule, and its effect is the anterior layer that mates fuzzy rule, calculates the relevance grade of every rule.That is:

$\mathrm{aj}={\mathrm{\&mu;}}_1^{s_{1j}}{\mathrm{\&mu;}}_2^{{\mathrm{\&mu;}}_{2j}}{s}_{1j}\&Element;\{\mathrm{1,2},...,m_1\},$ s _2j∈{1,2,…,m ₂} (3) ${\mathrm{<figure-callout\; id="3"\; label="I\; k"\; filenames="121010164256.png"\; state="\{\{state\}\}">I</figure-callout>}}_k^{}={\mathrm{\&mu;}}_1^3{\mathrm{\&mu;}}_2^3=a_k,$ ${\mathrm{<figure-callout\; id="3"\; label="O\; k"\; filenames="121010164256.png"\; state="\{\{state\}\}">O</figure-callout>}}_k^{}=I_k^3---\left(4\right)$

The 4th layer of obfuscation output layer.18 nodes are arranged, represent the relevance grade of 18 degrees of membership.Each node carries out normalized, also can be used to regard as the weights of rear layer network output layer.

$I_j^{\left(4\right)}=\stackrel{\&OverBar;}{a_j}=\frac{a_j}{\underset{i=1}{\overset{N_A}{\&Sum;}}{a}_i}{,O}_j^{\left(4\right)}=I_j^{\left(4\right)}=a_j---\left(5\right)$

The layer 5 output layer is 1 node, the representative amount of the spraying result of decision.

2, collect the on-the-spot true-time operation data of greenhouse dispenser composing training sample pair, therefrom choose 2/3 data and as training sample the fuzzy neural network that makes up is carried out off-line training, determine weights and the threshold value of fuzzy neural network.

Because the fuzzy partition number of input component determines that the parameter that need to learn only has the connection weight ω of last one deck _Ij(i=1,2 ... r; J=1,2 ..., m), and the central value c of the subordinate function of the second layer _IjAnd width cs _Ij(i=1,2 ..., r; J=1,2 ..., m).Getting the error cost function is:

$E=\frac{1}{2}\underset{i=1}{\overset{r}{\mathrm{\&Sigma;}}}{(y_{\mathrm{di}}-y_i)}^2---\left(6\right)$

Wherein, y _DiAnd y _iThe actual output and the desired output that represent respectively neural network.Utilize error back propagation algorithm to calculate

With

Then utilize the gradient search algorithm to regulate ω _Ijc _IjAnd σ _IjThe learning algorithm of the parameter adjustment that provides at last is:

${\mathrm{\&omega;}}_{\mathrm{ij}}(k+1)={\mathrm{\&omega;}}_{\mathrm{ij}}\left(k\right)-\mathrm{\&beta;}\frac{\mathrm{dE}}{{\mathrm{d\&omega;}}_{\mathrm{ij}}},$ j＝1,2,…,m _i (7)

$c_{\mathrm{ij}}(k+1)=c_{\mathrm{ij}}\left(k\right)-\mathrm{\&beta;}\frac{\mathrm{dE}}{{\mathrm{dc}}_{\mathrm{ij}}},$ j＝1,2,…,m _i (8)

${\mathrm{\&sigma;}}_{\mathrm{ij}}(k+1)={\mathrm{\&sigma;}}_{\mathrm{ij}}\left(k\right)-\mathrm{\&beta;}\frac{\mathrm{dE}}{{\mathrm{d\&sigma;}}_{\mathrm{ij}}},$ j＝1,2,…,m _i (9)

Wherein, i=1,2 ..., n; β〉0 be learning rate.

The learning training process of neural network is by the weight coefficient between continuous each layer of adjustment, makes the output of neural network minimum with the error of final expectation value, until satisfy application request.In database, preserve the fuzzy neural network data of learning success.After network training was complete, every weights and the threshold value of downloading fuzzy neural network arrived decision-making device.

3, with 1/3 the data sample of reserving as detecting sample to training rear fuzzy neural network to carry out 1/3 test sample book input that decision accuracy detects to reserve as the input of the fuzzy neural network that trains, whether the error of observing its output and sample expectation value meets the demands, to finish test process.

4, in mobile robot's computer system, the fuzzy neural network that the programming realization trains

By training, test, obtained every weights and the threshold value of fuzzy neural network.In mobile spraying machine device people computer system, carry out software programming, realize the input/output relation of fuzzy neural network institute match.

5, camera is delivered to computing machine by USB interface with the image information that collects, and utilizes image processing techniques that crop row is extracted from background, and calculates crop row information, sprays area in order to differentiation

The crop map that collects is looked like to deliver to computing machine, process by image, crop row is extracted from background, the image area coverage on crop blade face can with the method for the zoning area of simplifying, be calculated the pixel that belongs to the crop zone exactly in every width of cloth image.

Image is traveled through, utilize following formula to the zone of crop wherein (green) statistics as the rope number:

$\mathrm{Totalg}=\underset{(x,y)\&Element;R}{\&Sum;}1---\left(10\right)$

Wherein: R is all pixels of crop row image; (x, y) is 255 pixels for all R components of image.Obtain after the crop actual pixels number, utilize following formula can obtain the actual area coverage of crop:

$\mathrm{Area}=\frac{{\mathrm{Tota}1}_g}{{\mathrm{Total}}_p}*P---\left(11\right)$

In the formula: Area is the actual area coverage of crop;

Totalg is the actual pixel count that should obtain of crop;

Totalp is the image total pixel number;

P is the corresponding sampling area of piece image.

6, adopt ultrasonic sensor to record the range information that sprays target

According to formula (12) can calculate to be painted apply thing apart from d.

d＝(c*t)/2 (12)

Wherein,

Be the aerial velocity of propagation of ultrasound wave, t is the time of timer record.

7, by the disease and pest grade of computer user interface by artificial input hothouse plants

If the disease and pest grade is obtained by video camera, image processing, sensor transmissions time and computer console deal with data time will cause system responses to lag behind, and affect the efficiency of decision-making.In order to improve spray drug effect rate and real-time, and consider the characteristics of general greenhouse band crop generation disease and pest, by plant protection personnel setting disease and pest plague grade, manually input the disease and pest grade n of hothouse plants according to pest species information by computer user interface.

8, utilize to have made up and merge the fuzzy neural network that sprays target range, area and the disease and pest extent of injury information amount of spraying is made a strategic decision, calculate the accordingly amount of spraying, thereby realize the variable farm chemical applying decision-making.

As the input of fuzzy neural network, calculate accordingly spray amount by the fuzzy neural network decision-making device with actual target distance, area and the disease and pest level data of recording.

The inventive method is mainly put forth effort on a kind of fuzzy neural network decision-making technique and is solved the quick, intelligent decision problem of greenhouse band crop mobile robot variable farm chemical applying.The implementing procedure of movement-based robot as shown in Figure 4.The specific implementation method of the method is as follows:

1, makes up, trains and the test fuzzy neural network;

For the greenhouse band crop, choose area, distance and the disease and pest class information of crop target as the distinguishing rule of formulation rate, so fuzzy neural network is three inputs, five layer network structures.Wherein, area, unit are m ², give 3 linguistic variable values, i.e. " greatly ", " in ", " little "; Distance, unit is m, gives 3 linguistic variable values, i.e. " closely ", " in ", " far "; The disease and pest grade is given 2 linguistic variable values, i.e. " seriously ", " not serious ".

Collect the data that dispenser scene, greenhouse records, analyze on-the-spot real time data, choose 180 pairs of data and carry out normalized as sample data, the part sample data sees Table 1, wherein 120 pairs of sample datas are as training sample, training result as shown in Figure 5, hands-on result and deviation from the desired value are very little, show that this fuzzy neural network has reflected decision laws preferably, training result is more satisfactory.60 pairs of samples reserving are used for the test fuzzy neural network.After satisfying accuracy requirement after tested, fuzzy neural network namely becomes the neural network that trains.

After network training was complete, every weights and the threshold value of downloading fuzzy neural network arrived the computer decision-making device.

2, the fuzzy neural network that the programming realization trains in mobile robot's computer system

The fuzzy neural network that has trained is made the variable farm chemical applying decision-making with higher precision to the greenhouse band crop.In mobile robot's computer system, the making software program realizes the fuzzy neural network that has trained.

3, utilize the fuzzy neural network Real-time Decision greenhouse band crop variable amount of spraying that trains

Camera is delivered to computing machine by USB interface with the chamber crop image that collects, and processes by carrying out image, and crop row is extracted from background, and (unit is m to calculate the image area coverage s on crop blade face in every width of cloth image ²), the image area coverage on crop blade face can with the method for the zoning area of simplifying, be calculated the pixel that belongs to the crop zone exactly in every width of cloth image.Image is traveled through, utilize formula (10) that the zone of crop wherein (green) statistics is obtained picture rope number, (unit is m to utilize formula (11) can calculate the actual area coverage s of crop ²); Adopt the appended ultrasonic sensor of mobile robot to record the range information that sprays target, can calculate according to formula (12) that to be painted what apply thing is m apart from d(unit); Manually input the disease and pest extent of injury n of chamber crop by computer user interface.More than the data transmission that collects to the fuzzy neural network decision-making device, the dispenser system is 0.00656m by the best amount of spraying of can making a strategic decision out in line computation such as the target area ², distance is 0.2394m, and the disease and pest grade is 0, and the best amount of spraying is 0.02125Ls ^-1

Record actual flow by controller and flow sensor and carry out computing and comparison, and regulate in real time the PWM dutycycle according to comparative result, drive solenoid valve, realize that variable sprays.

4, repeat 3 steps until dispenser is finished.

Fig. 6 utilizes fuzzy neural network the decision-making device amount of spraying of making a strategic decision out and the Error Graph of expecting the amount of spraying, analytical error as can be known, fuzzy neural network decision-making technique provided by the present invention can realize fast and effectively formulation rate decision-making, can satisfy well greenhouse spray medicine mobile robot's variable precision dispenser requirement.

In sum, fuzzy neural network decision-making technique provided by the present invention, with target target area information, range information and disease and pest extent of injury information are decision-making foundation, it is more reasonable accurately that decision-making sprays the result, can simulate people's actual thinking and operating process, make the more realistic greenhouse of the adjustment process situation of system, owing to adopt the fuzzy neural network of off-line training, unnecessary process of carrying out sample data collection and analysis before variable is implemented, thereby can improve the rapidity of intelligent decision and the real-time of robot decision system, by can obtain best output in line computation, show very flexible.Fuzzy neural network intelligent variable dispenser decision-making technic proposed by the invention, easier requirement of real time can effectively reduce pesticide dosage, has significant practical significance to promoting food and job safety, reduction environmental pollution.

Claims (9)

1. mobile robot's variable farm chemical applying Using Intelligent Decision-making Method comprises:

Obtain the target area, spray distance and be and the disease and pest class information;

According to the target area that obtains, spray distance for and the disease and pest class information, calculate default formulation rate;

Calculate and the output controlled quentity controlled variable according to default formulation rate;

Obtain actual formulation rate, actual formulation rate and default formulation rate are compared, adjust controlled quentity controlled variable.

2. the method for claim 1 is characterized in that, the described step of obtaining the target area comprises:

Obtain crop map picture to be painted;

Carry out image and process, crop row is extracted from background, by the pixel that belongs to the crop zone being calculated the image area coverage on crop blade face in every width of cloth image.

3. the method for claim 1 is characterized in that, the default formulation rate step of described calculating comprises:

Set up the fuzzy neural network step, comprising:

Structure take the target area, spray distance for and the fuzzy neural network of disease and pest grade as inputting;

Learning procedure comprises:

The fuzzy neural network that makes up is carried out off-line training, obtain every weights and the threshold value of fuzzy neural network;

Fuzzy neural network after the training is carried out decision accuracy to be detected;

The fuzzy neural network that obtains training;

Determining step comprises:

The fuzzy neural network that the programming realization trains in mobile robot's computer system;

The fuzzy neural network that trains of operation, according to the target area that obtains, spray distance for and the disease and pest class information, judge default formulation rate.

4. method as claimed in claim 3 is characterized in that, described learning procedure comprises:

Getting the error cost function is:

$E=\frac{1}{2}\underset{i=1}{\overset{r}{\&Sum;}}{(y_{\mathrm{di}}-y_i)}^2---\left(1\right)$

Wherein, y _DiAnd y _iThe actual output and the desired output that represent respectively neural network; Utilize error back propagation algorithm to calculate

With

Then utilize the gradient search algorithm to regulate ω _Ijc _IjAnd σ _Ij, the learning algorithm of the parameter adjustment that provides at last is:

${\mathrm{\&omega;}}_{\mathrm{ij}}(k+1)={\mathrm{\&omega;}}_{\mathrm{ij}}\left(k\right)-\mathrm{\&beta;}\frac{\mathrm{dE}}{{\mathrm{d\&omega;}}_{\mathrm{ij}}},$ j＝1,2,…,m _i (2)

$c_{\mathrm{ij}}(k+1)=c_{\mathrm{ij}}\left(k\right)-\mathrm{\&beta;}\frac{\mathrm{dE}}{{\mathrm{dc}}_{\mathrm{ij}}},$ j＝1,2,…,m _i (3)

${\mathrm{\&sigma;}}_{\mathrm{ij}}(k+1)={\mathrm{\&sigma;}}_{\mathrm{ij}}\left(k\right)-\mathrm{\&beta;}\frac{\mathrm{dE}}{{\mathrm{d\&sigma;}}_{\mathrm{ij}}},$ j＝1,2,…,m _i (4)

Wherein, i=1,2 ..., n; β〉0 be learning rate.

5. method as claimed in claim 3 is characterized in that, the fuzzy neural network that described operation trains, according to the target area that obtains, spray distance for and the disease and pest class information, judge default formulation rate step and specifically comprise:

Take the target area, spray distance and 3 input variables of the sick worm extent of injury as the input neuron of network;

The subordinate function of analog input variable calculates degree of membership corresponding to each input variable;

The coupling fuzzy rule, the former piece that carries out every fuzzy rule calculates;

Calculate the relevance grade of every fuzzy rule;

Calculate the amount of spraying.

6. method as claimed in claim 4 is characterized in that,

The subordinate function of described analog input variable calculates in degree of membership step corresponding to each input variable, adopts Gaussian function as subordinate function, that is:

$y=e^{-\frac{{(x-c)}^2}{{\mathrm{\&sigma;}}^2}---\left(5\right)}$

$I_{\mathrm{isi}}^{\left(2\right)}=-\frac{{(x_i^1-c_{\mathrm{isi}})}^2}{{\mathrm{\&sigma;}}_{\mathrm{isi}}^2},$ $O_{\mathrm{isi}}^{\left(2\right)}={\mathrm{\&mu;}}_i=e^{-\frac{{(x_1^{\left(1\right)}-c_{\mathrm{isi}})}^2}{{\mathrm{\&sigma;}}_{\mathrm{isi}}^2}---\left(6\right)}$

Wherein, formula (5) is the Gaussian function expression;

In the formula (6), c _Isi, σ _IsiRepresent respectively i input variable x _iCentral value and the mean square deviation of Gaussian subordinate function of j input; o _IsiIt is the second layer output valve of fuzzy neural network;

Described coupling fuzzy rule carries out in the former piece calculation procedure of every fuzzy rule, adopts following formula to calculate the relevance grade of every rule:

$a_j={\mathrm{\&mu;}}_1^{s_{1j}}{\mathrm{\&mu;}}_2^{s_{2j}}---\left(7\right)$

Wherein, in the formula (7), a _jJ neuronic input; The linguistic variable value of the former piece part of j rule; s _2j∈ 1,2 ..., m ₁, s _2j∈ 1,2 ..., m ₂.

In the relevance grade step of every fuzzy rule of described calculating, adopt formula as follows:

$I_j^{\left(4\right)}=\stackrel{\&OverBar;}{a_j}=\frac{a_j}{\underset{i=1}{\overset{N_A}{\&Sum;}}{a}_i}---\left(8\right)$

Wherein, in the formula (8), a _jJ neuronic input;

Neuron a _iAt interval [1, N _A] directly and.

7. method as claimed in claim 2, it is characterized in that, the described image that carries out is processed, crop row is extracted from background, in the image area coverage step that the pixel that belongs to the crop zone is calculated crop blade face in every width of cloth image, comprise: image travels through, utilize following formula to the zone of crop wherein (green) statistics as the rope number:

$\mathrm{Totalg}=\underset{(x,y)\&Element;R}{\&Sum;}1---\left(9\right)$

Wherein: R is all pixels of crop row image, and (x, y) is 255 pixels for all R components in the image;

Obtain after the crop actual pixels number, utilize following formula can obtain the actual area coverage of crop:

$\mathrm{Area}=\frac{{\mathrm{Tota}1}_g}{{\mathrm{Total}}_p}*P---\left(10\right)$

In the formula: Area is the actual area coverage of crop;

Totalg is the actual pixel count that should obtain of crop;

Totalp is the image total pixel number;

P is the corresponding sampling area of piece image.

8. mobile robot's variable farm chemical applying intelligent decision system comprises:

Camera is used for by USB interface the chamber crop image that collects being delivered to computing machine;

Ultrasonic sensor is used for recording the distance that sprays target and being sent to computing machine;

Flow sensor is used for measuring actual spray medicine flow;

Computing machine, it is connected with camera, ultrasonic sensor and flow sensor, wherein disposes the fuzzy neural network decision-making device, is used for:

Obtain the target area, spray distance and be and the disease and pest class information;

According to the target area that obtains, spray distance for and the disease and pest class information, calculate default formulation rate;

Calculate and the output controlled quentity controlled variable according to default formulation rate;

Obtain actual formulation rate, actual formulation rate and default formulation rate are compared, adjust controlled quentity controlled variable.

9. system as claimed in claim 8, it is characterized in that, described fuzzy neural network decision-making device adopts five layers of structure of fuzzy neural network, and ground floor is input layer, take the target area, spray distance, 3 input variables of the sick worm extent of injury as the input neuron of network; The second layer is the obfuscation input layer, is used for calculating degree of membership corresponding to each input variable; The 3rd layer is the fuzzy rule layer, and the former piece that is used for every fuzzy rule calculates; The 4th layer is the obfuscation output layer, is used for calculating the relevance grade of every rule; Layer 5 is output layer, is used for calculating the amount of spraying.

Images