CN101334395B - COD soft sensing process - Google Patents

Abstract

The invention discloses a COD soft measurement method which belongs to the sewage treatment field. Water quality indicators are difficult to carry out the on-line measurement due to the non-linearity, the time-varying property and the complexity of the sewage biological treatment process, but the water quality indicators are often important for the sewage treatment. The COD soft measurement method is designed against the problem of difficult COD on-line measurement and provides the method which applies the rapid EFAST method for pruning redundant neurons, simplifying the neural network structure and carrying out the soft measurement of COD according to the characteristic of the non-linearity of the sewage treatment process. A great deal of materials matching for processing the biochemical reactions of an aeration tank are timely adjusted according to the soft measurement result, thereby facilitating the better removal of the COD and avoiding complicated projects for research and development of sensors. The method can also be expanded to carry out the research of other water quality indicators for guiding the actual production and operation.

Description

The method of the soft measurement of COD

Technical field

The present invention relates to the flexible measurement method of effluent quality index in the sewage disposal process, especially utilize quick EFAST method to simplify neural network structure carries out soft measurement to COD method.

Background technology

Wastewater treatment is exactly in fact to adopt necessary processing method and treatment scheme, pollutants in sewage is separated or is translated into harmless material, thereby sewage is purified.Modern sewage water treatment method can be divided into physical treatment process, method of chemical treatment, physicochemical treatment method and biochemical treatment method four big classes by its mechanism of action.Wherein biochemical treatment method is the most frequently used sewage water treatment method of a class, and its major function is to utilize the metabolism of microorganism, makes the organic contaminant that is dissolving and colloidal state in the sewage be converted into stable innoxious substance.Its advantage is the organic removal rate height, and operating cost is low, does not still have the method that can compare favourably with it in municipal effluent and biodegradable Industrial Wastewater Treatment.Activated sludge process wherein is to cause in recent years both at home and abroad extensively to pay attention to, one of aerobic biochemical treatment process that research and application increase day by day.But on the one hand, because the working condition of biological wastewater treatment process is abominable, random disturbance is serious, has many inputs, many output, uncertainty, strong nonlinearity, characteristics such as change greatly the time, makes extremely complexity of this process, is difficult to describe with mathematical model; On the other hand, the good course of reaction that will help to describe activated Sludge System at the intelligent model of nonlinear system, help to simulate the dynamic change of activated Sludge System and to the influence of every water-quality guideline to instruct actual production run, fundamentally improve the efficient stable and the economic rationality of wastewater treatment.Therefore, set up the activated sludge sewage processing system model of more reliable perfect practicality, become the important topic of sewage control engineering area research, and had important practical significance.

In sewage disposal system, there are some water-quality guideline to be difficult to on-line measurement, this is because non-linear, the time variation of biological wastewater treatment process and complicacy cause.And these water-quality guideline are often very important to the qualified discharge or the on-line monitoring in the operation of sewage disposal system of wastewater treatment, as BOD5, COD, SVI.

Develop the Process meter of novel example, in hardware, though can directly solve the detection problem of various sewage disposal process variablees and water quality parameter, but because organism is very complicated in the sewage, research and develop these sensors and will be one expensive big, last long engineering.

Summary of the invention

The objective of the invention is problem,, propose the quick EFAST method of a kind of utilization and prune redundant neuron, simplify neural network structure according to the sewage disposal process nonlinear feature at COD on-line measurement difficulty, and the method for COD being carried out soft measurement.According to soft measurement result, in time adjust the material collocation of countless processing aeration tank biochemical reaction, so that COD is better removed, avoided the complex engineering of research and development sensors.Also this method can be expanded, other water-quality guideline is studied, to instruct actual production run.

The present invention has adopted following technical scheme and performing step:

1. the method for the soft measurement of COD is characterized in that, may further comprise the steps:

(1). set up three layers of feedforward neural network forecast model of the soft measurement of COD; Be input as sewage regulating reservoir influent quality index, be output as chemical oxygen demand COD;

Initialization neural network: determine the connected mode of l-p-1, the weights of neural network are carried out random assignment;

Promptly an input layer has l neuron, and hidden layer has three layers of feedforward neural network of p neuronic single output, x ₁, x ₂..., x _lThe input of expression neural network, y _dThe desired output of expression neural network; Total k training sample, (it is that field of neural networks is known that the size of the number k of training sample is set) establishes t training sample is x ₁(t), x ₂(t) ..., x _l(t), y _d(t), when then using t training sample neural network training, hidden layer j neuronic output is expressed as:

$Z_j\left(t\right)=\mathrm{\ψ}\left(\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{w}_{i,j}^I{x}_i\left(t\right)\right)---\left(1\right)$

Wherein, x _iBe the input of neural network, w _{I, j} ^IBe input layer weights, Z _jBe hidden layer j neuronic output, ψ is the sigmoid function, and its form is:

$\mathrm{\ψ}\left(n\right)=\frac{1}{1+e^{-n}}---\left(2\right)$

The pass of hidden layer neuron output and neural network output is:

$y\left(t\right)=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_j^O{Z}_j\left(t\right)---\left(3\right)$

Wherein, w _j ^OBe the output layer weights, y is the actual output of neural network;

The definition error function is

$E=\frac{1}{2}\underset{t=1}{\overset{K}{\mathrm{\Σ}}}{(y\left(t\right)-y_d\left(t\right))}^T(y\left(t\right)-y_d\left(t\right))---\left(4\right)$

The purpose of neural network training is to make the error function of formula (4) definition reach minimum;

(2). sample data is proofreaied and correct;

If k data sample x (1), x (2) ..., x (k), average is x, the deviation of each sample is v (t)=x (t)-x, calculates standard deviation according to the Bessel formula:

$\mathrm{\σ}=\sqrt{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}\frac{{(x\left(t\right)-\stackrel{\‾}{x})}^2}{k-1}}---\left(5\right)$

If the deviation of some sample x (t) satisfies:

|v(t)|≥3σt＝1，2，...，k (6)

Think that then sample x (t) is an abnormal data, should give rejecting, the data after obtaining proofreading and correct;

(3). with the data neural network training after proofreading and correct, and in training process, utilize quick EFAST that the redundant hidden neuron of neural network is pruned,, increase generalization ability and predetermined speed of neural network to simplify neural network structure;

Be specially: with the output Z of p hidden neuron ₁, Z ₂..., Z _pAs p input parameter of quick EFAST algorithm, with the output of neural network as the output of EFAST algorithm fast, by calculating input parameter Z ₁, Z ₂..., Z _pFor the sensitivity of output Y, prune the little neuron of sensitivity;

Write down the output valve of each hidden layer neuron in the training process, find out maximal value and minimum value;

J (j=1,2 .., p) maximal value of individual hidden neuron output in the training m step is b _j, minimum value is a _j, then suppose Z _jWith assigned frequency w _jAt [a _j, b _j] interior vibration, that is:

$Z_j\left(s\right)=\frac{b_j+a_j}{2}+\frac{b_j-a_j}{\mathrm{\π}}\arcsin \left(\sin \left(w_{j}s\right)\right)---\left(7\right)$

The output y of convolution (3) neural network is expressed as:

$y\left(s\right)=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_j^O{Z}_j\left(s\right)---\left(8\right)$

(5). calculate each neuronic sensitivity:

If calculate the sensitivity of h hidden neuron, then the fourier coefficient of this neuron correspondence is:

$A_h=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}y\left(s\right)\cos \left(w_{h}s\right)\mathrm{ds}---\left(9\right)$

$B_h=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}y\left(s\right)\cos \left(w_{h}s\right)\mathrm{ds}---\left(9\right)$

Wherein, w _h=8p;

All the other neuron fourier coefficient A except that h neuron _jAnd B _jBe expressed as:

$A_j=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}y\left(s\right)\cos \left(w_{j}s\right)\mathrm{ds}---\left(11\right)$

$B_j=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}y\left(s\right)\cos \left(w_{j}s\right)\mathrm{ds}---\left(12\right)$

Wherein, j=1,2 ..., h-1, h+1 ..., p, and w _j=j;

Output Z by each hidden layer neuron of feedforward neural network ₁, Z ₂..., Z _pBetween do not interact, the Fourier amplitudes value mainly concentrates on the fundamental frequency, adopts following formula to calculate the total sensitivity of h hidden neuron;

${\mathrm{ST}}_h=\frac{A_h^2+B_h^2}{A_h^2+B_h^2+\underset{j=1}{\overset{p}{\mathrm{\Σ}}}(A_j^2+B_j^2)}---\left(13\right)$

Wherein, denominator is to comprise A _h ²+ B _h ²In all interior hidden neuron fourier coefficient sums; ST _hComprise Z _hBe independent of effect and the Z of other input parameter to output _hSynergy with other input parameter is called total sensitivity;

Then the total sensitivity sum of p hidden neuron is:

$\mathrm{Sum}\_S=\underset{h=1}{\overset{p}{\mathrm{\Σ}}}S{T}_h---\left(14\right)$

The sensitivity of h input parameter is:

S _h＝ST _h/Sum_S (15)

Neural network is pruned, promptly delete S _hHidden layer neuron less than 5%;

(6). continue neural network training, then every certain step number repeating step (2)-(5), the sensitivity of all hidden layer neuron all stops to prune greater than 5% in new neural network;

(7). neural network training up to error E less than specification error E _dThis error is generally less than 0.01;

(8). forecast sample is predicted: with the input of forecast sample data as the neural network that trains, the output of neural network is predicting the outcome of water outlet COD.

Creativeness of the present invention is mainly reflected in:

(1) the present invention is directed to the problem that chemical oxygen demand COD is difficult to on-line measurement,, adopted neural network that COD is carried out soft measurement, saved the complex process of development sensor, thereby had more convenience according to the characteristics that neural network can be approached nonlinear function.

(2) the present invention adopts quick EFAST method that the structure of neural network is pruned, and has solved the problem that the neural network initial configuration is difficult to determine, has avoided neural network to cause generalization ability poor owing to too complicated, and fitting precision is low.

To note especially: the present invention just for convenience for the purpose of, so what adopt is that COD is carried out soft measurement, equally also, should belong to scope of the present invention as long as adopted principle of the present invention to measure all applicable to BOD5, SVI etc.

Description of drawings

The soft measurement neural net model establishing of Fig. 1 .COD.

Fig. 2. the quick EFAST algorithm of neural network is pruned back prediction COD value and actual value matched curve figure.

Embodiment

Experimental data is from certain sewage treatment plant's water analysis daily sheet.Fig. 1 has provided the COD neural network prediction model, and its input is respectively mixed liquor suspended solid, MLSS (MLSS), oil, pH, ammonia nitrogen, is output as the COD of water outlet in the sewage disposal process.Wherein MLSS is meant the weight of the contained dewatered sludge of unit volume biochemistry pool mixed liquor; The soda acid degree of pH reflection influent quality; Oil is the content of the oil pollutant of water into; Ammonia nitrogen content in the ammonia nitrogen representative water inlet.COD is meant the oxidized material needed oxygen amount by the chemical oxidizing agent oxidation time of energy in the waste water.Except that pH, above unit is mg/litre.Adopt the connected mode of 4-32-1, i.e. 4 input neurons, 32 hidden neurons, 1 output neuron is seen Fig. 1.Each 100 groups of training sample and forecast samples are pruned neural network with quick EFAST method.

Neural network structure pruning algorithm concrete steps are as follows:

(1). the initialization neural network: determine the connected mode of 4-32-1, the weights of neural network are carried out random assignment, the weights initial value of this experiment is 0 to 1 random number.

(2). sample data is proofreaied and correct.

(3). go on foot to c with the training sample data neural network training after proofreading and correct.If the c choosing is too small, then the quantity of information of Cai Jiing is sufficient inadequately; If the c choosing is excessive, can increase the training time.It is 600 that c is chosen in the experiment of this paper.

(4). write down the output valve of each hidden neuron in the training process, find out maximal value and minimum value.

(5). calculate each neuronic sensitivity: be the output appointment fundamental frequency separately of each hidden neuron.This tests neuronic number is 32, if calculate the sensitivity of the 10th hidden neuron, so just the frequency of each hidden neuron is appointed as [1,2,3 ..., 9,256,11 ..., 32], for each hidden neuron,, calculate this neuronic sensitivity S T according to formula (13) again with maximal value, minimum value and the assigned frequency substitution thereof of its output _h

(6). repeating step (5), up to the total sensitivity of having calculated 32 hidden neuron.

(7). calculate each neuronic sensitivity S according to formula (15) _h(h=1,2 ..., p), deletion S _hHidden layer neuron less than 5%.

(8). continue neural network training, then every certain step number t repeating step 2-5, the sensitivity of all hidden layer neuron all stops to prune greater than 5% in new neural network.It is excessive that t should not select, otherwise neural network might finish with regard to training in the step at t; If the t choosing is too small, then the training burden of Cai Jiing is abundant inadequately.It is 100 that t is chosen in this experiment.

(9). neural network training up to deviation less than specification error E _dE is chosen in this experiment _d=0.01.。

(10). forecast sample is predicted: with the input of forecast sample data as the neural network that trains, the output of neural network is predicting the outcome of water outlet COD.

Fig. 2 is that the quick EFAST method of neural network is pruned back prediction COD value and actual value matched curve figure.Hence one can see that, can predict water outlet COD effectively based on the structural self-organizing neural network of quick EFAST method.

Claims (1)

1. the method for the soft measurement of COD is characterized in that, may further comprise the steps:

(1). set up three layers of feedforward neural network forecast model of the soft measurement of COD; Be input as sewage regulating reservoir influent quality index, be output as chemical oxygen demand COD;

Initialization neural network: determine the connected mode of l-p-1, the weights of neural network are carried out random assignment;

Promptly an input layer has l neuron, and hidden layer has three layers of feedforward neural network of p neuronic single output, x ₁, x ₂..., x _lThe input of expression neural network, y _dThe desired output of expression neural network; Total k training sample, establishing t training sample is x ₁(t), x ₂(t) ..., x _l(t), y _d(t), when then using t training sample neural network training, hidden layer j neuronic output is expressed as:

$Z_j\left(t\right)=\mathrm{\ψ}\left(\underset{i=1}{\overset{l}{\text{\Σ}}}{w}_{i,j}^I{x}_i\left(t\right)\right)---\left(1\right)$

Wherein, x _iBe the input of neural network,

Be input layer weights, Z _jBe hidden layer j neuronic output, ψ is the sigmoid function, and its form is:

$\mathrm{\ψ}\left(n\right)=\frac{1}{1+e^{-n}}---\left(2\right)$

The pass of hidden layer neuron output and neural network output is:

$y\left(t\right)=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_j^O{Z}_j\left(t\right)---\left(3\right)$

Wherein,

Be the output layer weights, y is the actual output of neural network;

The definition error function is

$E=\frac{1}{2}\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{(y\left(t\right)-y_d\left(t\right))}^T(y\left(t\right)-y_d\left(t\right))---\left(4\right)$

The purpose of neural network training is to make the error function of formula (4) definition reach minimum;

(2). sample data is proofreaied and correct;

If k data sample x (1), x (2) ..., x (k), average is

The deviation of each sample is Calculate standard deviation according to the Bessel formula:

$\mathrm{\σ}=\sqrt{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}\frac{{(x\left(t\right)-\stackrel{\‾}{x})}^2}{k-1}}---\left(5\right)$

If the deviation of some sample x (t) satisfies:

|v(t)|≥3σt＝1，2，...，k (6)

Think that then sample x (t) is an abnormal data, should give rejecting, the data after obtaining proofreading and correct;

(3). with the data neural network training after proofreading and correct, and in training process, utilize quick EFAST that the redundant hidden neuron of neural network is pruned,, increase generalization ability and predetermined speed of neural network to simplify neural network structure;

Be specially: with the output Z of p hidden neuron ₁, Z ₂..., Z _pAs p input parameter of quick EFAST algorithm, with the output of neural network as the output of EFAST algorithm fast, by calculating input parameter Z ₁, Z ₂..., Z _pFor the sensitivity of output Y, prune the little neuron of sensitivity;

Write down the output valve of each hidden layer neuron in the training process, find out maximal value and minimum value;

The maximal value of j hidden neuron output in the training m step is b _j, minimum value is a _j, j=1 wherein, 2 ..., p then supposes Z _jWith assigned frequency w _jAt [a _j, b _j] interior vibration, that is:

$Z_j\left(s\right)=\frac{b_j+a_j}{2}+\frac{b_j-a_j}{\mathrm{\π}}\arcsin \left(\sin \left(w_{j}s\right)\right)---\left(7\right)$

The output y of convolution (3) neural network is expressed as:

$y\left(s\right)=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_j^O{Z}_j\left(s\right)---\left(8\right)$

(4). calculate each neuronic sensitivity:

If calculate the sensitivity of h hidden neuron, then the fourier coefficient of this neuron correspondence is:

$A_h=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}y\left(s\right)\cos \left(w_{h}s\right)\mathrm{ds}---\left(9\right)$

$B_h=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}y\left(s\right)\sin \left(w_{h}s\right)\mathrm{ds}---\left(10\right)$

Wherein, w _h=8p;

All the other neuron fourier coefficient A except that h neuron _jAnd B _jBe expressed as:

$A_j=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}y\left(s\right)\cos \left(w_{j}s\right)\mathrm{ds}---\left(11\right)$

$B_j=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}y\left(s\right)\sin \left(w_{j}s\right)\mathrm{ds}---\left(12\right)$

Wherein, j=1,2 ..., h-1, h+1 ..., p, and w _j=j;

Output Z by each hidden layer neuron of feedforward neural network ₁, Z ₂..., Z _pBetween do not interact, the Fourier amplitudes value mainly concentrates on the fundamental frequency, adopts following formula to calculate the total sensitivity of h hidden neuron;

${\mathrm{ST}}_h=\frac{A_h^2+B_h^2}{A_h^2+B_h^2+\underset{j=1}{\overset{P}{\mathrm{\Σ}}}(A_j^2+B_j^2)}---\left(13\right)$

Wherein, denominator is to comprise

In all interior hidden neuron fourier coefficient sums; ST _hComprise Z _hBe independent of effect and the Z of other input parameter to output _hSynergy with other input parameter is called total sensitivity;

Then the total sensitivity sum of p hidden neuron is:

$\mathrm{Sum}\_S=\underset{h=1}{\overset{p}{\mathrm{\Σ}}}{\mathrm{ST}}_h---\left(14\right)$

The sensitivity of h input parameter is:

S _h＝ST _h/Sum_S (15)

Neural network is pruned, promptly delete S _hHidden layer neuron less than 5%;

(5). continue neural network training, then every certain step number repeating step (2)-(4), the sensitivity of all hidden layer neuron all stops to prune greater than 5% in new neural network;

(6). neural network training up to error E less than specification error E _d

(7). forecast sample is predicted: with the input of forecast sample data as the neural network that trains, the output of neural network is predicting the outcome of water outlet COD.

Images