CN101685506B - Expert diagnosis decision method of inorganic waste water processing scheme - Google Patents

Abstract

The invention relates to an expert diagnosis decision method of an inorganic waste water processing scheme, which can process an inorganic waste water quality index and a reuse water quality index by a neural network structure so as to output a waste water processing scheme. The method comprises: preprocessing a first vector containing the inorganic waste water quality index and the reuse water quality index in sections to obtain a second vector; according to a trained first weighted value matrix, calculating the weighed value sum of each element of the second vector; using a non-linear excitation function to convert to obtain a third vector; in an output layer, calculating the weighed value sum of each element of the third vector according to the trained second weighed value matrix; and using a threshold value type excitation function to convert to output a fourth vector, wherein the fourth vector comprises a selected waste water processing scheme.

Description

The expert diagnosis decision-making technique of inorganic wastewater processing scheme

Technical field

The present invention relates to a kind of decision-making technique of inorganic wastewater processing scheme, relate in particular to the expert diagnosis decision-making technique of automatic generation inorganic wastewater processing scheme.

Background technology

Industry such as metallurgy, iron and steel, automobile, electronics, plating can produce a large amount of inorganic wastewaters.Industrial waste water is converted into the universal demand that can reusable new resources have become enterprise; Yet inorganic industrial waste water is of a great variety; Often contain materials such as toxic heavy metal, grease, acid, alkali in the waste water; The treatment process process is complicated, and enterprise has nothing in common with each other to the requirement of resource product, is difficult to solve with single scheme.Therefore need to confirm corresponding wastewater treatment and response scheme to different enterprise wastewater sources.

For wastewater treatment and reuse scheme determination, mostly rely on general knowledge and experience to come initial setting at present at present, verify through artificial lab scale or pilot scale again, go into operation at last and use.Therefore the deficiency of experience and the restriction of experiment condition often cause the experimental result deviation, are difficult to guarantee the stable treated effect, finally possibly draw the wrong scheme that causes failure in investment, are necessary to take more scientific methods to carry out the preliminary of scheme and confirm.

Summary of the invention

Technical matters to be solved by this invention provides a kind of expert diagnosis decision-making technique of inorganic wastewater processing scheme; Utilize the outstanding self study adaptive ability of artificial neural network technology; Network distributes and stores knowledge; Concurrent operation characteristic and superior non-linear mapping capability realize the automatic decision of wastewater treatment scheme.

The present invention solves the problems of the technologies described above the expert diagnosis decision-making technique that the technical scheme that adopts is a kind of inorganic wastewater processing scheme of proposition; Via a neural network structure inorganic wastewater water-quality guideline and quality of reused water index are handled; With output wastewater treatment scheme; This neural network structure comprises input layer, pretreatment layer, hidden layer and output layer, and this method may further comprise the steps:

Primary vector to the said pretreatment layer that comprises said inorganic wastewater water-quality guideline and said quality of reused water index through said input layer input;

In said pretreatment layer, each water-quality guideline in the said primary vector is carried out the branch section and handle, and the output secondary vector;

In said hidden layer,, and calculate a plurality of said and functions under a non-linear excitation respectively according to each element sum of the said secondary vector of the trained first weighted value matrix computations, export comprise a plurality of said and the 3rd vector of function; And

Each the element sum that in said output layer, accordings to said the 3rd vector of the trained second weighted value matrix computations; And calculate a plurality of said and functions under threshold-type excitation respectively; Output comprise a plurality of said and the four-way amount of function, said four-way amount comprises the wastewater treatment scheme of selection.

In above-mentioned method, also be included in the said pretreatment layer each water-quality guideline after minute section processing is carried out the normalization processing.

In above-mentioned method, said non-linear excitation is Sigmoid excitation function or RBF.

In above-mentioned method; Train the method for said first weighted value matrix and the said second weighted value matrix to comprise: to import the training battle array according to expert's scheme rule list structure of prevision and train battle array to train, utilize the δ learning algorithm of error back propagation to adjust said first weighted value matrix and the said second weighted value matrix with exporting.

The present invention proposes a kind of expert diagnosis decision-making technique of inorganic wastewater processing scheme in addition; Via a neural network structure inorganic wastewater water-quality guideline and quality of reused water index are handled; With output wastewater treatment scheme; This neural network structure comprises total input layer, a plurality of sub-network and total output layer, and wherein each sub-network comprises input layer, pretreatment layer, hidden layer and output layer respectively, and the method may further comprise the steps:

The primary vector that comprises said inorganic wastewater water-quality guideline and said quality of reused water index through said total input layer input;

In said each sub-network, carry out following steps:

Through said input layer input inorganic wastewater water-quality guideline and the quality of reused water index relevant with said sub-network;

In said pretreatment layer, said each relevant water-quality guideline is carried out the branch section and handle, and the output secondary vector;

In said hidden layer,, and calculate a plurality of said and functions under a non-linear excitation respectively according to each element sum of the said secondary vector of the trained first weighted value matrix computations, export comprise a plurality of said and the 3rd vector of function; And

Each the element sum that in said output layer, accordings to said the 3rd vector of the trained second weighted value matrix computations; And calculate a plurality of said and functions under threshold-type excitation respectively; Export a plurality of said and function to said total output layer, said four-way amount comprises the wastewater treatment scheme of selection; And

Via the wastewater treatment scheme of said total output layer output by each sub-network selection.

In above-mentioned method, said neural network structure comprises a plurality of sub-networks, in said pretreatment layer, also comprises, each water-quality guideline after minute section is handled is carried out normalization and handled.

In above-mentioned method, said non-linear excitation is Sigmoid excitation function or RBF.

In above-mentioned method; Train the method for said first weighted value matrix and the said second weighted value matrix to comprise: to import the training battle array according to expert's scheme rule list structure of prevision and train battle array to train, utilize the δ learning algorithm of error back propagation to adjust said first weighted value and the said second weighted value matrix with exporting.

The present invention makes it compared with prior art owing to adopt above technical scheme, has following remarkable advantage:

1, diagnosis decision method can generate tentative programme according to the raw water quality that detects and the reuse water-quality guideline of input automatically with reference to expert's scheme of maturation, in order to verification experimental verification, has promoted the efficient of definite scheme greatly.

2, owing to adopt neural network structure; The present invention has overcome to a certain extent that knowledge acquisition problem, the learning ability of normal expert system is relatively poor, " narrow step effect " knowledge is the bottleneck problems such as contradiction of poor fault tolerance and knowledge storage capacity and travelling speed; Thereby further improve expert system learning ability and the ability of handling large complicated problem; Make knowledge base have good expandability, the operation of system has higher reliability.

Description of drawings

For let above-mentioned purpose of the present invention, feature and advantage can be more obviously understandable, elaborate below in conjunction with the accompanying drawing specific embodiments of the invention, wherein:

Fig. 1 is an expert diagnosis decision system structured flowchart according to an embodiment of the invention.

Fig. 2 is the neural network structure figure of expert diagnosis decision-making module according to an embodiment of the invention.

Fig. 3 is the neural network structure figure that has the expert diagnosis decision-making module of a plurality of sub-networks according to an embodiment of the invention.

Fig. 4 is an expert diagnosis decision-making technique process flow diagram according to an embodiment of the invention.

Fig. 5 is an expert diagnosis decision-making technique process flow diagram according to another embodiment of the present invention.

Embodiment

Among the embodiment below, utilize the outstanding self study adaptive ability of artificial neural network technology, network distributes and stores knowledge, and concurrent operation characteristic and superior non-linear mapping capability realize the automatic decision of wastewater treatment scheme.

Fig. 1 is an expert diagnosis decision system structured flowchart according to an embodiment of the invention.This expert diagnosis decision system 100 comprises expert diagnosis decision-making module 110 and expert knowledge library 120.In one embodiment, expert diagnosis decision system 100 also can comprise data analysis module 130.

To the character of the inorganic wastewater of different industry, different process and the water quality standard of different disposal water,, can formulate the wastewater treatment scheme of maturation in advance according to existing wastewater recycling process.These ripe schemes and expert's scheme rule is stored in the expert knowledge library 120, supplies expert diagnosis decision-making module 110 to call.Expert diagnosis decision-making module 110 can pass through these expertises of training study.Use the knowledge of having learnt, expert diagnosis decision-making module 110 can be confirmed preliminary inorganic wastewater processing scheme according to the inorganic wastewater water-quality guideline and the desired quality of reused water index of input from existing wastewater treatment scheme.

This preliminary inorganic wastewater processing scheme can be passed through testing equipment and made an experiment, and obtains the actual water-quality guideline of recycle-water.The water-quality guideline that requires of the actual water-quality guideline of recycle-water and recycle-water is understood input data analysis module 130; Carry out the analysis of actual processing effect and the assessment of result; And can make amendment to expert diagnosis decision-making module 110 determined wastewater treatment schemes according to analysis and assessment result; Produce the wastewater treatment scheme scheme of optimizing, and additional expert knowledge library.

The structure of at first describing the expert decision-making module of one embodiment of the invention below realizes.

Expert diagnosis decision-making module 110 is made up of neural network structure as shown in Figure 1, and the neural network of this expert diagnosis decision-making module 110 adopts feedforward network, can be divided into four layers, is respectively input layer 111, pretreatment layer 112, hidden layer 113 and output layer 114.These layers can be realized by computer executable program.

In one embodiment, each layer neuron number confirmed as follows: the water-quality guideline that inorganic wastewater is set has the R item, and then the neuron number of input layer each item water-quality guideline of equaling Inlet and outlet water (waste water and recycle-water) is counted sum, is 2R; The neuron of pretreatment layer is to divide and decide (following detailed description) at the different condition section in when decision-making by the Inlet and outlet water index; And two times of the desirable pretreatment layer node of the neuron number of hidden layer number; It is S that the neuron number of output layer equals expert's scheme number.

Because expert's scheme of difference in functionality is only relevant with different piece Inlet and outlet water index; In one embodiment; For the pace of learning and simplification network structure that improves neural network; Can a macroreticular be divided into a plurality of sub-networks according to the Inlet and outlet water index classification of association, thus the expert diagnosis decision-making module that to generate a scale be the plurality of subnets network.

For instance; When 13 schemes are only relevant with 2 Inlet and outlet water indexs; Shown in table one, an input layer then can setting up is 2 neurons (conductivity of promptly intaking with go out water conductivity), and output layer is four layers of forward direction sub-network of 13 neurons (i.e. 13 schemes).Because the different range of the corresponding different Inlet and outlet water indexs of different processes scheme; Water inlet conductivity from 0 to 2999 μ m/cm is divided into 0-199,200-399,400-999,1000-2999 totally 4 sections in will showing, and goes out water conductivity from 0.2 to 19 and is divided into 0.2-0.9,1-9,10-19 totally 3 sections.When if sub-network is only got into the water conductivity and gone out water conductivity and be 2 neurons of input layer 111, then to get the section sum be 7 to pretreatment layer 112 neuron numbers, and the neuron number of getting hidden layer 113 is 14.Other sub-network design by that analogy.

Table one expert scheme rule list

Accept above-mentionedly, when there was a plurality of sub-network in module, mode that can similar parallel connection connected.Fig. 2 illustrates the neural network structure figure of the expert diagnosis decision-making module with a plurality of sub-networks.The input quantity of total input layer 211 of this expert diagnosis decision-making module 210 is whole Inlet and outlet water index parameter, and total input layer 211 each neuronic input connect each Inlet and outlet water index parameter respectively.The input of each child network 2121-212n connects again and comprises the neuron output of the input layer 211 of relevant Inlet and outlet water index parameter with it.Each child network can have similar structure shown in Figure 2, comprises input layer, pretreatment layer, hidden layer and output layer.The output quantity of the total output layer 213 of module is whole schemes number, and total output layer 213 each neuronic output correspond to scenarios number.And the output of each child network connects the scheme relevant with these sub-networks number neuronic input of pairing those total output layers.Thereby realize the expert diagnosis decision-making module that a scale is the plurality of subnets network.

Be example with single network shown in Figure 1 below, describe the inference mechanism of expert diagnosis decision-making module 110.

Expert diagnosis decision-making module 110 adopts the forward reasoning methods, by the input as module of each item water-quality guideline of Inlet and outlet water (waste water and recycle-water), carries out reasoning through the neural network propagated forward, and the network output of calculating is decision scheme number.Concrete forward reasoning method is following:

Input layer 111 is as module interface, and each neuron of input layer adopts the linear incentive function, and input and output directly are the primary vector X={x that comprises each item water-quality guideline of Inlet and outlet water (waste water and recycle-water) ₁..., x _i..., x _2R, i is a natural number, 2R is input water-quality guideline number.

Pretreatment layer 112 is the output x with input layer _iCarrying out the branch section on request handles.In order to overcome the saturated phenomenon of the s type function in the study, the Inlet and outlet water parameter value of different range is carried out normalization handle, adopt 0-1-0 square type excitation function to be:

A, b are threshold value in the formula, are decided by concrete scheme rule.

So each water-quality guideline x _iAll can be divided into one or more normalizing value y _jThereby, obtain secondary vector Y={y ₁..., y _j..., y _T, j=1,2 ..., T.T is the neuronal quantity of pretreatment layer 112, and it depends on the processing of branch section.

The neuron of hidden layer 113 is input as the weighting sum of all pretreatment layer outputs:

$z_k=\underset{j}{\mathrm{\Σ}}{w}_{\mathrm{jk}}{y}_j---\left(2\right)$

W wherein _JkBe first weighted value, all w _JkForm the first weighted value matrix W ₁w _JkInitial value can be set arbitrarily.After the study of the neural network that warp is stated later, w _JkTo level off to accurately.k＝1，2，...，2T。2T is the neuronal quantity of hidden layer.The neuronal quantity that it is pointed out that hidden layer 113 is not defined as 2 times of pretreatment layer.

Nonlinear function is adopted in the neuron output of hidden layer 113, Sigmoid excitation function (also claiming the S type function) for example commonly used:

$z_k^{\′}=f\left(z_k\right)=\frac{1}{1+e^{-z_k}}---\left(3\right)$

The 3rd vectorial Z is formed in all outputs of hidden layer 113

.

In another example, the output of the neuron of hidden layer 113 also can be adopted RBF.

The neuronic of output layer 114 is output as:

$a_l=\underset{k}{\mathrm{\Σ}}{w}_{\mathrm{kl}}{z}_k^{\′}---\left(4\right)$

W wherein _KlBe second weighted value, all w _KlForm the second weighted value matrix W ₂w _KlInitial value can be set arbitrarily.After the study of the neural network that warp is stated later, w _KlTo level off to accurately.

But passing threshold type excitation function further makes the neuronic of output layer 114 be output as:

C is a threshold value in the formula, between the desirable 0.8-0.9.Four-way amount A is formed in these outputs

; 1=1 wherein; 2 ..., S; S is the neuronal quantity of output layer, also is expert's scheme quantity of module.For each neuron output

, 1=1,2 ..., S, if $a_l^1=1$ , corresponding expert's scheme number has been selected in expression. approaches 1 more, can think that corresponding expert's scheme is number reliable more.

Be understood that easily for the module with a plurality of sub-networks as shown in Figure 3, the inference mechanism of its sub-network is similar.Difference only is that just from total input layer of network, get relevant with it part water-quality guideline carries out above-mentioned reasoning to each sub-network, exports simultaneously on the neuron relevant in corresponding expert's reuse scheme to total output layer of network.

To describe the foundation of the expert knowledge library of neural network expert diagnosis decision system below, the foundation of expert knowledge library comprises knowledge acquisition and two processes of knowledge store.

1) knowledge obtains

Obtaining of knowledge shows as obtaining and selecting of training sample.According to existing expert decision-making result, list the into corresponding relation (being the Expert Rules scheme) of water index and effluent index and water technology scheme, can generate the input training battle array P and an output training battle array T of neural network.

If Expert Rules scheme table shown in aforementioned table one, has two condition projects in the table: water inlet conductivity and water outlet conductivity indices, be divided into 4 sections and 3 sections respectively, the water technology scheme has 13.If in the Expert Rules scheme table 13 rules are arranged, wherein rule 1 is: water inlet conductivity indices condition 400-2999 drops on the 3rd subregion, and water outlet conductivity indices condition 0.2-9 strides the 1st, 2 two subregion, and corresponding scheme as a result is No. 1.The raw column data of then corresponding neural metwork training battle array P and T battle array is shown in following table two.

Table two

The P battle array	The T battle array
			Water inlet conductivity subregion	Go out the water conductivity subregion	Scheme number as a result
1?2?3?4	1 2	1?2?3?4?5?6?7?8?9?10 11?12?13
			0?0?1?0	1 0	1?0?0?0?0?0?0?0?0?0 ?0 0 0
0?0?1?0	0 1	1?0?0?0?0?0?0?0?0?0?0?0 0 0
			......	......	......

Other all possible situation is all carried out similar processing, note guaranteeing the completeness and the expandability of Expert Rules, thereby generate the P battle array and the T battle array of training usefulness.

2) storage of knowledge

The expert decision-making knowledge store of neural network expert diagnosis decision system is implicitly to disperse to be stored in each neuron of neural network to connect in weights and the threshold value.The storing process of knowledge is exactly the learning process of neural network.

According to the neural network structure that design generates, the parameter that the neural network of expert decision-making can be learnt to adjust is the weight w of hidden layer _JkWeight w with output layer _KlAdopt the δ learning algorithm of error back propagation, adjust the weights of each interlayer, the learning algorithm that can derive neural network is following:

If the error of neural network individual output and corresponding desired output

is:

$e_l=a_l^0-a_l^1---\left(6\right)$

The error performance target function of p sample is:

$E_p=\frac{1}{2}\underset{l=1}{\overset{N}{\mathrm{\Σ}}}{e_l}^2---\left(7\right)$

Wherein N is the neuron number of network output layer.

According to the gradient descent method, the learning algorithm of weights is following:

The connection weights learning algorithm of output layer and hidden layer is:

${\mathrm{\Δw}}_{\mathrm{kl}}=-\mathrm{\η}\frac{\∂E}{{\∂w}_{\mathrm{kl}}}=-\mathrm{\η}\·e_l\·\frac{{\∂a}_l}{{\∂w}_{\mathrm{kl}}}=-\mathrm{\η}\·e_l\·z_k^{\′}---\left(8\right)$

The t+1 weights of network constantly is:

w _kl(t+1)＝w _kl(t)+Δw _kl(t+1) (9)

Hidden layer and pretreatment layer connect the weights learning algorithm:

${\mathrm{\Δw}}_{\mathrm{jk}}=-\mathrm{\η}\frac{\∂E}{{\∂w}_{\mathrm{jk}}}=-\mathrm{\η}\·e_l\·\frac{{\mathrm{\ηa}}_l}{{\∂w}_{\mathrm{jk}}}---\left(10\right)$

Wherein

$\frac{{\∂a}_l}{{\∂w}_{\mathrm{jk}}}=\frac{{\∂a}_l}{{\∂z}_k^{\′}}\·\frac{{\∂z}_k^{\′}}{{\∂z}_k}\·\frac{{\∂z}_k}{{\∂w}_{\mathrm{jk}}}=w_{\mathrm{kl}}\·z_k^{\′}(1-z_k^{\′})\·y_j---\left(11\right)$

The k+1 weights of network constantly is:

w _jk(k+1)＝w _jk(k)+Δw _jk(t+1) (12)

If consider the influence that last time, weights changed these weights, add factor of momentum, the weights of this moment are:

w _kl(k+1)＝w _kl(k)+Δw _kl(t+1)+α(w _kl(k)-w _kl(k-1)) (13)

w _jk(t+1)＝w _jk(t)+Δw _jk(t+1)+α(w _jk(t)-w _jk(t-1)) (14)

Wherein η is a learning rate, and α is a factor of momentum.Get α ∈ [0,1] η ∈ [0,1].

By the training battle array P and the T battle array that generate according to the Expert Rules table, neural network is through 5000 training studies, the error criterion function can reach E 0.02, accomplish the foundation of neural network expert decision-making knowledge base at this moment.Via the reasoning decision-making of training back neural network, the requirement of its input/output relation and Expert Rules scheme table reaches in full accord.To train the back neural network to embed in the expert decision system, operation result can reach consistent with expectation.

Therefore, above-mentioned neural network expert decision system can be carried out computing according to input water-quality guideline (being the waste water quality index) and output water-quality guideline (being the quality of reused water index), and output waste water recycling processing scheme is in order to verification experimental verification.

According to above-mentioned expert diagnosis decision system, can summarize the expert diagnosis decision-making technique of a kind of inorganic wastewater processing scheme of the present invention, please combine Fig. 2 and shown in Figure 4, this method comprises the steps:

Step 310: the primary vector X={x that comprises inorganic wastewater water-quality guideline and quality of reused water index through input layer 111 inputs ₁..., x _i..., x _2RTo pretreatment layer 112;

Step 320: in pretreatment layer 112, each water-quality guideline among the primary vector X is carried out the branch section and handle, and output secondary vector Y={y ₁..., y _j..., y _T, 0-1-0 square type excitation function wherein capable of using carries out normalization with the Inlet and outlet water index of different range to be handled.

Step 330: in hidden layer 113, according to the trained first weighted value matrix W ₁Calculate each element weighting sum z of secondary vector _k:

$z_k=\underset{j}{\mathrm{\Σ}}{w}_{\mathrm{jk}}{y}_j$ (2)

Use the Sigmoid excitation function $z_k^{\′}=f\left(z_k\right)=\frac{1}{1+e^{-z_k}}$ Is z with these with nonlinear transformation _k', export the 3rd vectorial Z={z ₁' ..., z _k' ..., z _2T'.

Step 340: in output layer 114, according to the trained second weighted value matrix W ₂Calculate each element weighting sum a of the 3rd vector _l:

$a_l=\underset{k}{\mathrm{\Σ}}{w}_{\mathrm{kl}}{z}_k^{\′}---\left(4\right)$

And calculate a respectively _lThrough the value a after the conversion of threshold-type excitation function _l ¹, output comprises the four-way amount A={a of these functional values ₁ ¹..., a ₁ ¹..., a _s ¹, its intermediate value is the wastewater treatment scheme that 1 element representation is selected.

According to above-mentioned expert diagnosis decision system, can summarize the expert diagnosis decision-making technique of another kind of inorganic wastewater processing scheme of the present invention, please combine Fig. 3 and shown in Figure 5, this method comprises the steps:

Step 410: the primary vector X={x that comprises inorganic wastewater water-quality guideline and said quality of reused water index through total input layer 211 inputs ₁..., x _i..., x _2R;

Step 420: carry out forward inference at each sub-network 2121-212n, to confirm relevant wastewater treatment scheme, it further may further comprise the steps:

Step 421: through input layer input inorganic wastewater water-quality guideline and the quality of reused water index relevant with sub-network, these relevant water-quality guideline are that selectivity is imported from primary vector X;

Step 422: in pretreatment layer, each relevant water-quality guideline is carried out the branch section and handle, and export the secondary vector Y ' that comprises the index after the branch section is handled, its processing procedure is similar to aforesaid step 320;

Step 423: in hidden layer, according to the trained first weighted value matrix W ₁Calculate each element weighting sum of secondary vector Y ', and with these with through after the non-linear excitation functional transformation, export the 3rd vectorial Z ', its processing procedure is similar to aforesaid step 330;

Step 424: in output layer, according to the trained second weighted value matrix W ₂Calculate each element weighting sum of the 3rd vector, and through after the conversion of threshold-type excitation function, the value behind the output transform is to total output layer 213, and its intermediate value is the wastewater treatment scheme that 1 function representation is selected, and its processing procedure is similar to aforesaid step 340.

At last, in step 430, via the wastewater treatment scheme of total output layer 213 outputs by each sub-network 2121-212n selection.

Thus, can from existing wastewater treatment scheme, confirm one or more preliminary wastewater treatment schemes, supply verification experimental verification.

In sum, the present invention utilizes the outstanding self study adaptive ability of artificial neural network technology, and network distributes and stores knowledge, and concurrent operation characteristic and superior non-linear mapping capability realize the automatic decision of wastewater treatment scheme.Knowledge acquisition problem, the learning ability of normal expert system are relatively poor because neural network expert system has overcome to a certain extent, " narrow step effect " knowledge is the bottleneck problems such as contradiction of poor fault tolerance and knowledge storage capacity and travelling speed; Thereby further improve expert system learning ability and the ability of handling large complicated problem; Make knowledge base have good expandability, the operation of system has higher reliability.

Though the present invention discloses as above with preferred embodiment; Right its is not that any those skilled in the art are not breaking away from the spirit and scope of the present invention in order to qualification the present invention; When can doing a little modification and perfect, so protection scope of the present invention is when being as the criterion with what claims defined.

Claims (8)

1. the expert diagnosis decision-making technique of an inorganic wastewater processing scheme; Via a neural network structure inorganic wastewater water-quality guideline and quality of reused water index are handled; With output wastewater treatment scheme; Said neural network structure comprises input layer, pretreatment layer, hidden layer and output layer, said method comprising the steps of:

Primary vector to the said pretreatment layer that comprises said inorganic wastewater water-quality guideline and said quality of reused water index through said input layer input;

In said pretreatment layer, each water-quality guideline in the said primary vector is carried out the branch section and handle, and the output secondary vector;

In said hidden layer,, and calculate a plurality of said and functions under a non-linear excitation respectively according to each element sum of the said secondary vector of the trained first weighted value matrix computations, export comprise a plurality of said and the 3rd vector of function; And

Each the element sum that in said output layer, accordings to said the 3rd vector of the trained second weighted value matrix computations; And calculate a plurality of said and functions under threshold-type excitation respectively; Output comprise a plurality of said and the four-way amount of function, said four-way amount comprises the wastewater treatment scheme of selection.

2. the method for claim 1 is characterized in that, also is included in the said pretreatment layer each water-quality guideline after minute section processing is carried out the normalization processing.

3. the method for claim 1 is characterized in that, said non-linear excitation is Sigmoid excitation function or RBF.

4. the method for claim 1 is characterized in that, trains the method for said first weighted value matrix and the said second weighted value matrix to comprise:

Import the training battle array according to expert's scheme rule list structure of prevision and train battle array to train, utilize the δ learning algorithm of error back propagation to adjust said first weighted value matrix and the said second weighted value matrix with exporting.

5. the expert diagnosis decision-making technique of an inorganic wastewater processing scheme; Via a neural network structure inorganic wastewater water-quality guideline and quality of reused water index are handled; With output wastewater treatment scheme; Said neural network structure comprises total input layer, a plurality of sub-network and total output layer, and wherein each sub-network comprises input layer, pretreatment layer, hidden layer and output layer respectively, said method comprising the steps of:

The primary vector that comprises said inorganic wastewater water-quality guideline and said quality of reused water index through said total input layer input;

In said each sub-network:

Through said input layer input inorganic wastewater water-quality guideline and the quality of reused water index relevant with said sub-network;

In said pretreatment layer, said each relevant water-quality guideline is carried out the branch section and handle, and the output secondary vector;

In said hidden layer,, and calculate a plurality of said and functions under a non-linear excitation respectively according to each element sum of the said secondary vector of the trained first weighted value matrix computations, export comprise a plurality of said and the 3rd vector of function;

Each the element sum that in said output layer, accordings to said the 3rd vector of the trained second weighted value matrix computations; And calculate a plurality of said and functions under threshold-type excitation respectively; Export a plurality of said and function to said total output layer, said four-way amount comprises the wastewater treatment scheme of selection; And

Via the wastewater treatment scheme of said total output layer output by each sub-network selection.

6. method as claimed in claim 5 is characterized in that, also is included in the said pretreatment layer each water-quality guideline after minute section processing is carried out the normalization processing.

7. method as claimed in claim 5 is characterized in that, said non-linear excitation is Sigmoid excitation function or RBF.

8. method as claimed in claim 5 is characterized in that, trains the method for the said first weighted value matrix and said many group second weighted value matrixes to comprise:

Import the training battle array according to expert's scheme rule list structure of prevision and train battle array to train, utilize the δ learning algorithm of error back propagation to adjust said first weighted value matrix and the said second weighted value matrix with exporting.

Images