CN105675807A - Evaluation method of atrazine residue based on BP neural network - Google Patents

Abstract

The invention relates to an evaluation method of an atrazine residue based on a BP neural network. The method comprises the steps: with use of an atrazine residue based knowledge representation method, acquiring an original eigenvalue for predicting the atrazine residue; with use of a nerve cell quantity calculation method based on the BP neural network, discreting the original eigenvalue; and inputting the discreted eigenvalue into a BP neural network optimized by a generalization ability, modeling, acquiring the connection weight and threshold value of the BP neural network, and thus obtaining a more accurate atrazine residue evaluation model. With application of the evaluation model, a pesticide atrazine residue level can be more accurately predicted.

Description

A kind of atrazine based on BP neutral net remains evaluation methodology

Technical field

The present invention relates to BP neutral net, Detecting Pesticide field, particularly relate to a kind of Pesticide Atrazine Residue Grade assessment method based on BP neutral net.

Background technology

Crop can be produced poisoning by the pesticide residues in soil, particularly now widely used atrazine, and its pedo relict effect phase is long, and late stubble sensitive crop such as Semen Tritici aestivi, Semen sojae atricolor, Oryza sativa L. etc. are caused very big poisoning. Further, the pesticide residues in agricultural product exceed standard and also bring about great losses to agricultural products in China outlet. The pesticide residues evaluation methodology that current China adopts; it is usually and the pesticide to register is carried out environmental behaviour test; wasting substantial amounts of manpower and materials, the environmental behaviour experiment of pesticide simultaneously can cause environmental pollution, is unfavorable for environmental conservation and sustainable development.

Artificial neural network is a brain formula intelligent information handling system being intended to imitate human brain structure and function. By simulating thinking and the learning process of human brain, artificial neural network adjusts the connection weights between network neural unit and threshold value, completes reasoning and discrimination function. Artificial neural network has good self-organizing, self study and adaptive ability, it is adaptable to process challenge and open systems. In recent years, the research of artificial neural network and application start to penetrate into many fields, such as machine learning, specialist system, Based Intelligent Control, coding theory etc., present adaptability and higher accuracy widely.

Artificial neural network is being utilized to solve in the process of challenge, the relevant knowledge that key challenge is how problem of representation field faced. Relevant knowledge had both included the ABC of the principle function relevant with problem and configuration aspects, also included the experience etc. that expert accumulates in processing practical problem. In traditional specialist system, these knowledge are stored in knowledge base with the rule of production, and in neutral net, the knowledge got is stored among whole network with connection weights and the threshold value distribution of network.

Summary of the invention

It is an object of the invention to provide a kind of atrazine Residue Grade assessment method based on BP neutral net, utilize atrazine residual evaluation model that the Pesticide Atrazine Residue Grade in crops is estimated. For solving above-mentioned technical problem, the main technical content of the present invention is as follows:

A kind of knowledge representation method based on atrazine residual. The residual quantity of atrazine is relevant with several factors, such as original atrazine spray value, the target crop of spray medicine, field conditions factors etc., concrete field conditions factor includes annual precipitation, year-round average temperature, mean annual humidity, soil half-life, soil organic matter content, P in soil H-number, dust content (1.0-0.01mm), physical clay content (0.01-0.001mm) etc.The difference that selection and the weight of influence factor are chosen can directly influence the forecasting accuracy of atrazine Residue Grade. The knowledge representation method that the present invention remains based on a kind of atrazine, specifically comprising the following steps that of the method

(1) influence factor that atrazine residual quantity is relevant being carried out raw data acquisition, data are normalized expression by application data preprocess method.

(2) above-mentioned normalization result constitutes the eigenvalue for predicting, eigenvalue carries out the correlation analysis of atrazine residual quantity, deletes and the incoherent variable of atrazine residual quantity, and obtains the correlation coefficient of each eigenvalue.

(3) apply above-mentioned correlation coefficient and determine each eigenvalue weight, and eigenvalue is carried out sliding-model control.

(4) difference according to each eigenvalue discretization degree, application neuronal quantity determines formula, calculates the atrazine persistence characteristic value meeting neutral net pattern of the input, obtains the rule base of assessment atrazine residual.

Further, in above-mentioned steps 1, comparing for initial data is carried out dependency under unified dimension, data normalization adopts Z-score standardized method.

Further, in step 2 above, set atrazine residual quantity correlation coefficient threshold, when eigenvalue correlation coefficient is less than this threshold value, leave out this characteristic dimension; When eigenvalue correlation coefficient is be more than or equal to this threshold value, retains this characteristic dimension and this correlation coefficient is labeled as weight. Apply this eigenvalue weight in step 3 and carry out sliding-model control, when eigenvalue weight is higher, it was shown that this eigenvalue have higher can discrimination, the discretization degree of this eigenvalue is higher; When eigenvalue weight is relatively low, the discretization degree of this eigenvalue is relatively low.

In step 4, the discrete conditions number comprised in the neuron number thus eigenvalue that the neural tuple corresponding to certain eigenvalue comprises determines. Neuron number determines that formula is as follows:

N=[Log₂(M+1)+0.5]

In formula: M is the discrete conditions number that this eigenvalue comprises; N is the neuron number that this nerve tuple comprises. Coefficient 0.5 is for rounding the first half of formula. Utilize this formula can uniquely identify each eigenvalue, and make input neuron the least possible.

Initial data should be processed in aforementioned manners, obtain atrazine residual rule base. Being divided into two parts, a part to can be used for training neutral net the data in rule base, namely obtain the knowledge base of neutral net, another part is then used for the evaluation accuracy of system of verifying as test sample.

Utilize this knowledge base to obtain connection weights and the threshold value of BP neutral net, can accurately represent the relevant knowledge that atrazine remains.

A kind of Generalization Capability optimization method of BP neutral net. The generalization ability of neutral net depends on the structure of network and the characteristic of learning sample. In general hidden layer neuron quantity is more few, and the Generalization Capability of network is more good, but hidden layer neuron quantity can make network not restrain very little. This method is intended to choose between training precision and Generalization accuracy metastable equilibrium value, it is to avoid the situation that study is excessive occur.

This method, based on the Generalization Capability optimization method of a kind of BP neutral net, specifically comprises the following steps that

(1) according to the above-mentioned knowledge representation method based on atrazine residual, acquire the eigenvalue of training neutral net, and determine the initial neuron number of neutral net.

(2) by training sample input neural network, the computing network actual output from hidden layer to output layer.

(3) from output layer backwards calculation to input layer, application network performance function layer-by-layer correction each interneuronal connection weights and threshold value.

(4) multiple samples are carried out repeatedly repetition training, until the output of all samples all reaches training requirement.

This method is applicable to the neutral net of single hidden layer. Further, for determining the initial hidden neuron number of neutral net in step 1, following empirical equation is applied:

$h=\left\{\begin{array}{cc}n+0.618*(n-m)& n\&GreaterEqual;m\\ m-0.618*(m-n)& m<n\end{array}\right.$

In formula, n is input number of nodes, and m is output node number, and h is the number of hidden nodes. The number of hidden nodes obtained based on this formula is input layer nodes and the intermediate value of output layer neuron node number, and when the number of hidden nodes is near input number of nodes, faster, training result is better for the convergence rate of network.

Further, the possibility excessive for reducing study, improve the Generalization Capability of network, above-mentioned steps 3 adopts new network performance function. New network performance function is as follows:

$E_A=r\frac{1}{Q*N}\underset{q=1}{\overset{Q}{\&Sigma;}}\underset{n=1}{\overset{N}{\&Sigma;}}{\left(T_q\right(n)-O_q(n\left)\right)}^2+(1-r)\frac{1}{M}\underset{m=1}{\overset{M}{\&Sigma;}}{w}_m^2$

In formula, r is the proportionality coefficient between 0-1; Q is the sample number that training sample set comprises; N is output layer neuron node number; T_qN () is the target output value of the n-th output layer neuron node when sample q is trained; O_qN () is real output value; M is the number of the value that weight threshold matrix comprises; w_mFor m-th weight threshold.

When r is equal to 0, namely this performance function is equal to the performance function of standard neural network algorithm. When 0 < r, < when 1, the latter half of this performance function is improvement part. R value selects according to the training condition of network, now

$\frac{\&part;E_A}{\&part;w_{ij}}=r\frac{\&part;E}{\&part;w_{ij}}+(1-r)\frac{2}{M}{w}_{ij}$

After the BP neural network algorithm of standard is made above improvement, neutral net degree of convergence, pace of learning and Generalization Capability can have a degree of raising.

Advantages of the present invention:

1, this accuracy is good. Utilize the present invention that training sample is modeled, and sample will be tested as mode input, in the less situation of training sample, still reach higher accuracy.

2, this utilizes the generalization ability optimization method of BP neutral net of the present invention, and for the expert knowledge library of same atrazine residue detection, pace of learning has had and is comparatively obviously improved.

Accompanying drawing explanation

Fig. 1 is the present invention knowledge acquisition mechanism structure chart based on neutral net.

Fig. 2 is the discretization process schematic diagram of feature of present invention value.

Detailed description of the invention

Below in conjunction with accompanying drawing, the present invention is described further.

The data acquisition environment of the present embodiment is the milpa of Tong County, Beijing City and Guangxi province Nanning City, the residual quantity of atrazine is relevant with several factors, such as original atrazine spray value, the target crop of spray medicine, field conditions factors etc., concrete field conditions factor includes annual precipitation, year-round average temperature, mean annual humidity, soil half-life, soil organic matter content, P in soil H-number, dust content (1.0-0.01mm), physical clay content (0.01-0.001mm) etc. The difference that selection and the weight of influence factor are chosen can directly influence the forecasting accuracy of atrazine Residue Grade. Specifically comprising the following steps that of the present embodiment

(1) influence factor that atrazine residual quantity is relevant being carried out raw data acquisition, data are normalized expression by application data preprocess method.

(2) above-mentioned normalization result constitutes the eigenvalue for predicting, eigenvalue carries out the correlation analysis of atrazine residual quantity, deletes and the incoherent variable of atrazine residual quantity, and obtains the correlation coefficient of each eigenvalue.

(3) apply above-mentioned correlation coefficient and determine each eigenvalue weight, and eigenvalue is carried out sliding-model control.

(4) difference according to each eigenvalue discretization degree, application neuronal quantity determines formula, calculates the atrazine persistence characteristic value meeting neutral net pattern of the input, obtains the rule base of assessment atrazine residual.

Further, in above-mentioned steps 1, comparing for initial data is carried out dependency under unified dimension, data normalization adopts Z-score standardized method.

The discretization that Fig. 2 is feature of present invention value represents schematic diagram, it can be seen that primitive character value carries out sliding-model control according to the difference of correlation coefficient, and is expressed as the form of input layer according to discrete conditions number. Further, in step 2 above, set atrazine residual quantity correlation coefficient threshold, when eigenvalue correlation coefficient is less than this threshold value, leave out this characteristic dimension; When eigenvalue correlation coefficient is be more than or equal to this threshold value, retains this characteristic dimension and this correlation coefficient is labeled as weight. Sliding-model control in eigenvalue weights influence step 3. When eigenvalue weight is higher, it was shown that this eigenvalue have higher can discrimination, the discretization degree of this eigenvalue is higher; When eigenvalue weight is relatively low, the discretization degree of this eigenvalue is relatively low.

In step 4, the discrete conditions number comprised in the neuron number thus eigenvalue that the neural tuple corresponding to certain eigenvalue comprises determines. Neuron number determines that formula is as follows:

N=[Log₂(M+1)+0.5]

In formula: M is the discrete conditions number that this eigenvalue comprises; N is the neuron number that this nerve tuple comprises. Coefficient 0.5 is for rounding the first half of formula. Utilize this formula can uniquely identify each eigenvalue, and make input neuron the least possible.

Initial data should be processed in aforementioned manners, obtain atrazine residual rule base. Being divided into two parts, a part to can be used for training neutral net the data in rule base, namely obtain the knowledge base of neutral net, another part is then used for the evaluation accuracy of system of verifying as test sample.

Utilize this knowledge base to obtain connection weights and the threshold value of BP neutral net, can accurately represent the relevant knowledge that atrazine remains.

Primitive character value is inputted BP neutral net be trained, adopt the Generalization Capability optimization method of the BP neutral net of the present invention. The generalization ability of neutral net depends on the structure of network and the characteristic of learning sample. In general hidden layer neuron quantity is more few, and the Generalization Capability of network is more good, but hidden layer neuron quantity can make network not restrain very little. This method is intended to choose between training precision and Generalization accuracy metastable equilibrium value, it is to avoid the situation that study is excessive occur.

The present embodiment specifically comprises the following steps that

(1) according to the above-mentioned knowledge representation method based on atrazine residual, acquire the eigenvalue of training neutral net, and determine the initial neuron number of neutral net.

(2) by training sample input neural network, the computing network actual output from hidden layer to output layer.

(3) from output layer backwards calculation to input layer, application network performance function layer-by-layer correction each interneuronal connection weights and threshold value.

(4) multiple samples are carried out repeatedly repetition training, until the output of all samples all reaches training requirement.

This method is applicable to the neutral net of single hidden layer. Further, for determining the initial hidden neuron number of neutral net in step 1, following empirical equation is applied:

$h=\left\{\begin{array}{cc}n+0.618*(n-m)& n\&GreaterEqual;m\\ m-0.618*(m-n)& m<n\end{array}\right.$

In formula, n is input number of nodes, and m is output node number, and h is the number of hidden nodes.The number of hidden nodes obtained based on this formula is input layer nodes and the intermediate value of output layer neuron node number, and when the number of hidden nodes is near input number of nodes, faster, training result is better for the convergence rate of network.

Further, the possibility excessive for reducing study, improve the Generalization Capability of network, above-mentioned steps 3 adopts new network performance function. New network performance function is as follows:

$E_A=r\frac{1}{Q*N}\underset{q=1}{\overset{Q}{\&Sigma;}}\underset{n=1}{\overset{N}{\&Sigma;}}{\left(T_q\right(n)-O_q(n\left)\right)}^2+(1-r)\frac{1}{M}\underset{m=1}{\overset{M}{\&Sigma;}}{w}_m^2$

In formula, r is the proportionality coefficient between 0-1; Q is the sample number that training sample set comprises; N is output layer neuron node number; T_qN () is the target output value of the n-th output layer neuron node when sample q is trained; O_qN () is real output value; M is the number of the value that weight threshold matrix comprises; w_mFor m-th weight threshold.

When r is equal to 0, namely this performance function is equal to the performance function of standard neural network algorithm. When 0 < r, < when 1, the latter half of this performance function is improvement part. R value selects according to the training condition of network, now

$\frac{\&part;E_A}{\&part;w_{ij}}=r\frac{\&part;E}{\&part;w_{ij}}+(1-r)\frac{2}{M}{w}_{ij}$

After the BP neural network algorithm of standard is made above improvement, neutral net degree of convergence, pace of learning and Generalization Capability can have a degree of raising.

Choose Tong County, Beijing City data in low concentration spray medicine situation as test sample. Test result shows, participates in the significant difference between its ranking of sample of test, and maximum diagnosis probability has exceeded 99%, it was shown that application the method can effectively distinguish the atrazine residual of different stage.

Part that the present invention does not relate to is all same as the prior art maybe can adopt prior art to be realized.

Claims (8)

1. the knowledge representation method based on atrazine residual, it is characterized in that, primitive character for calculating atrazine residue includes: original atrazine spray value, the target crop of spray medicine, field conditions factors etc., concrete field conditions factor includes annual precipitation, year-round average temperature, mean annual humidity, soil half-life, soil organic matter content, P in soil H-number, dust content (1.0-0.01mm), physical clay content (0.01-0.001mm).

2. the knowledge representation method based on atrazine residual as claimed in claim 1, it is characterised in that comprise the steps of

(1) influence factor that atrazine residual quantity is relevant being carried out raw data acquisition, data are normalized expression by application data preprocess method.

(2) above-mentioned normalization result constitutes the eigenvalue for predicting, eigenvalue carries out the correlation analysis of atrazine residual quantity, deletes and the incoherent variable of atrazine residual quantity, and obtains the correlation coefficient of each eigenvalue.

(3) apply above-mentioned correlation coefficient and determine each eigenvalue weight, and eigenvalue is carried out sliding-model control.

(4) difference according to each eigenvalue discretization degree, application neuronal quantity determines formula, calculates the atrazine persistence characteristic value meeting neutral net pattern of the input, obtains the rule base of assessment atrazine residual.

3. the knowledge representation method based on atrazine residual as described in claim 1 and claim 2, it is characterised in that in above-mentioned steps 1, compares for initial data is carried out dependency under unified dimension, and data normalization adopts Z-score standardized method.

4. the knowledge representation method based on atrazine residual as described in claim 1 and claim 2, it is characterised in that in step 2 above, set atrazine residual quantity correlation coefficient threshold, when eigenvalue correlation coefficient is less than this threshold value, leave out this characteristic dimension; When eigenvalue correlation coefficient is be more than or equal to this threshold value, retains this characteristic dimension and this correlation coefficient is labeled as weight. In above-mentioned steps 3, applying this eigenvalue weight and carry out sliding-model control, when eigenvalue weight is higher, the discretization degree of this eigenvalue is higher;When eigenvalue weight is relatively low, the discretization degree of this eigenvalue is relatively low.

5. the knowledge representation method based on atrazine residual as described in claim 1 and claim 2, it is characterised in that in step 4, the discrete conditions number comprised in the neuron number thus eigenvalue that the neural tuple corresponding to certain eigenvalue comprises determines. Neuron number determines that formula is as follows:

N=[Log₂(M+1)+0.5]

In formula: M is the discrete conditions number that this eigenvalue comprises; N is the neuron number that this nerve tuple comprises. Coefficient 0.5 is for rounding the first half of formula. Utilize this formula can uniquely identify each eigenvalue, and make input neuron the least possible.

6. the generalization ability optimization method of a BP neutral net, it is characterised in that comprise the steps of

(1) according to the above-mentioned knowledge representation method based on atrazine residual, acquire the eigenvalue of training neutral net, and determine the initial neuron number of neutral net.

(2) by training sample input neural network, the computing network actual output from hidden layer to output layer.

(3) from output layer backwards calculation to input layer, application network performance function layer-by-layer correction each interneuronal connection weights and threshold value.

(4) multiple samples are carried out repeatedly repetition training, until the output of all samples all reaches training requirement.

7. the generalization ability optimization method of BP neutral net as claimed in claim 6, it is characterised in that for determining the initial hidden neuron number of neutral net in step 1, apply following empirical equation:

$h=\left\{\begin{array}{cc}n+0.618*(n-m)& n\&GreaterEqual;m\\ m-0.618*(m-n)& m<n\end{array}\right.$

In formula, n is input number of nodes, and m is output node number, and h is the number of hidden nodes. The number of hidden nodes obtained based on this formula is input layer nodes and the intermediate value of output layer neuron node number, has convergence rate faster.

8. the generalization ability optimization method of BP neutral net as claimed in claim 6, it is characterised in that the network performance function of optimizing application. New network performance function is as follows:

$E_A=r\frac{1}{Q*N}\underset{q=1}{\overset{Q}{\&Sigma;}}\underset{n=1}{\overset{N}{\&Sigma;}}{\left(T_q\right(n)-O_q(n\left)\right)}^2+(1-r)\frac{1}{M}\underset{m=1}{\overset{M}{\&Sigma;}}{w}_m^2$

In formula, r is the proportionality coefficient between 0-1; Q is the sample number that training sample set comprises; N is output layer neuron node number; T_qN () is the target output value of the n-th output layer neuron node when sample q is trained; O_qN () is real output value; M is the number of the value that weight threshold matrix comprises; w_mFor m-th weight threshold.