CN102494979B - Soft measurement method for SVI (sludge volume index) - Google Patents

Abstract

A soft measurement method for an SVI (sludge volume index) belongs to the field of water treatment. Production conditions of sewage treatment process are severe, random disturbance of the sewage treatment process is serious, and the sewage treatment process is high in nonlinearity, large in time varying and serious in lagging. Almost all municipal sewage treatment plants and most industrial sewage treatment plants in China have sludge bulking to varying degrees every year, sludge bulking is mainly characterized by deterioration of sludge settleability, and the SVI is a parameter indicating the sludge settleability, but the key index is difficult to be measured on line. The soft measurement method based on a self-organizing RBF (radial basis function) neural network aims at solving the problem that the SVI cannot be measured on line in the sewage treatment process, the self-organizing RBF neural network is used for measuring the SVI on line in the sewage treatment process, normal operation of sewage treatment is ensured, cost is reduced, timely monitoring parameters are provided for realizing closed-loop control in the sewage treatment process, and the sewage treatment plants are promoted to run efficiently and stably.

Description

The flexible measurement method of a kind of sludge settling bulk index SVI

Technical field

The present invention utilizes self-organization RBF neural network to realize the soft measurement of sludge settling bulk index SVI in the sewage disposal process, and the concentration of SVI has directly determined the information of sludge settling in the sewage disposal process, and the normal operation of wastewater treatment is had material impact; Flexible measurement method is applied to sewage disposal system, both can have saved investment and operating cost, can in time monitor the wastewater treatment correlation parameter again, impel sewage treatment plant's efficient stable operation; The soft measurement of SVI is the important branch in advanced manufacturing technology field as the important step of wastewater treatment, has both belonged to water treatment field, belongs to the control field again.

Background technology

Water resources problems has become the subject under discussion of countries in the world government first concern, and the United Nations's " world's water resource comprehensive assessment report " points out: water problems will seriously restrict 21 century global economy and social development, and may cause conflicting between country; Therefore, set up sewage treatment plant, protect water environment to greatest extent, realize freshwater resources sustainable utilization and benign cycle, become the strategic measure of Chinese government's water resources comprehensive utilization.

Chinese Ministry of Environmental Protection pointed out that national wastewater emission amount in 2009 was 589.2 hundred million tons in 2010 in " China Environmental State Bulletin ", increase by 3.0% than the last year, and added up to dispose of sewage 343.33 billion cubic meters the whole year, and the water prevention and cure of pollution situation is still severe.Clearly propose in " Law of the People's Republic of China on the Prevention and Control of Water Pollution " that State Council implemented in 2008: country's encouragement, the scientific and technical research of held water prevention and cure of pollution and applying of advanced and applicable technology, the enterprise that requirement causes water to pollute undergoes technological transformation, take comprehensive preventive health measures, improve the recycling rate of waterused of water.Under this background, by the end of the year 2010, establish in the city, city in 654 in the whole nation, existing 607 cities have sewage treatment plant, account for 92.8% of city sum, the city, city is established in the whole nation, county's accumulative total is built up 2832 of urban wastewater treatment firms, and daily handling ability reaches 1.25 billion cubic meters.But the operation conditions of sewage treatment plant but allows of no optimist: because the facility operation rate of load condensate is low, Sewage Plant water inlet chemical oxygen demand (COD) concentration is low, reasons such as incomplete are supervised, detected to water quality, in sewage disposal process, be difficult to guarantee stability and the reliability of Sewage Plant operation.At present, nearly all municipal sewage plant and the most of industrial sewage all sludge bulking of various degrees in treatment plant's every year of China.Sludge bulking not only makes sludge loss, and effluent quality exceeds standard, even causes the collapse of entire sewage disposal system, endangers huge.Therefore, suppress sludge bulking and take place, guarantee quality of sewage disposal up to standard be current problem demanding prompt solution.

The principal character of sludge bulking is that sludge settling property worsens, and SVI is the parameter of expression sludge settling property, and sludge bulking takes place when SVI is higher than 150mL/g usually.This key index of SVI is difficult to on-line measurement, obtains by artificial chemical examination in the practical application, and its assay determination cycle generally needs a plurality of hours.The survey frequency of most of SVI of sewage treatment plant is 1-2 time weekly, is difficult to rely on the measured value of SVI in time to obtain sludge bulking information.Simultaneously, owing to the reason that causes sludge bulking is many-sided, and these factors influence each other, connect each other, mutual restriction, therefore, the modeling problem of sludge bulking is a global difficult problem, identify and predict that this is the relevant knowledge that an engineering problem has related to microorganism again for the sludge bulking that generally takes place, need a plurality of subject knowledges to intersect combination.Therefore, study the real-time measurement problem that new measuring method solves SVI, become the important topic of sewage control engineering area research, and had important practical significance.

The present invention proposes a kind of SVI flexible measurement method, by making up self-organization RBF neural network model, select one group both to maintain close ties with SVI, the auxiliary variable of measuring easily is as the input of self-organization RBF neural network again, realization is to the real-time measurement of SVI, thereby guarantee in time to find sludge bulking, reduce sludge bulking and take place, guarantee the normal operation of sewage treatment plant.

Summary of the invention

The present invention has obtained the flexible measurement method of sludge settling bulk index SVI in a kind of self-organization RBF Neural Network for Wastewater Treatment process; This method is by analyzing sewage disposal process, in numerous measurable variables, select one group the variable of close ties and measurement easily to be arranged as auxiliary variable with SVI, by structure self-organization RBF neural network, realize the on-line measurement of SVI, solved the problem that current SVI can't accurately obtain in real time;

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

The flexible measurement method of a kind of sludge settling bulk index SVI,

(1) sample data is proofreaied and correct;

Be provided with N primary data sample (x (1), x (2) ..., x (N)), average is χ, the deviation of each sample is D (j)=x (j)-χ, j=1,2 ..., N, calculate standard deviation:

$\mathrm{\σ}=\sqrt{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\frac{{(x\left(j\right)-\mathrm{\χ})}^2}{N}},---\left(1\right)$

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

|D(j)|≥3σ,j=1,2,...,N,(2)

Think that then sample x (j) is abnormal data, should give rejecting, the data after obtaining proofreading and correct, these data are as training sample and the test sample book of neural network;

(2) be designed for the self-organization RBF neural network initial topology structure of the soft measurement of SVI; Network is divided into three layers: input layer, hidden layer, output layer; Be input as auxiliary variable, be output as sludge settling bulk index SVI;

The initialization neural network: determine the connected mode of initial neural network, the neural network input layer is l, and the initial neuron of hidden layer is K, and the output layer neuron is one; Weights to neural network carry out random assignment; The input of neural network is expressed as x=(x ₁, x ₂..., x _l) ^T, (x ₁, x ₂..., x _l) ^TBe (x ₁, x ₂..., x _l) transposition, the desired output of neural network is expressed as y _dIf total M training sample, then t training sample is x (t)=(x ₁(t), x ₂(t) ..., x _l(t)) ^T, during with t training sample neural network training, the output of self-organization RBF neural network can be described as:

$y\left(t\right)=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{w}_k\left(t\right){\mathrm{\θ}}_k\left(x\left(t\right)\right),---\left(3\right)$

Wherein, K is the initial neuron number of hidden layer, x (t)=(x ₁(t), x ₂(t) ..., x _l(t)) ^TBe input vector, w _kBe the neuronic connection weights of k hidden layer neuron and output layer; θ _kBe the output of k hidden layer neuron, namely

${\mathrm{\θ}}_k\left(x\right)=e^{(-||x-{\mathrm{\μ}}_k||/{\mathrm{\σ}}_k^2)},---\left(4\right)$

Wherein, μ _kBe the central value of k hidden layer neuron, σ _kBe the variance of k hidden layer neuron,

k=1,2,…,K；

The definition error function is

$E\left(t\right)=\frac{1}{M}\underset{t=1}{\overset{M}{\mathrm{\Σ}}}{(y\left(t\right)-y_d\left(t\right))}^T(y\left(t\right)-y_d\left(t\right)),---\left(5\right)$

y _d(t) and y (t) be respectively system's actual measurement output and the neural network output of t training sample, the purpose of training self-organization RBF neural network is that the error function that makes formula (5) define reaches expected error value;

Its feature is further comprising the steps of:

(3) with the training sample neural network training after proofreading and correct, be specially:

1. a given RBF neural network, the initial neuron number of hidden layer is K, K is positive integer, initialization neural network weight w _k, its value is 0 to 1 random number; Initialization Center value μ _k, μ _kDimension be l, μ _kMiddle greatest member is less than 2, and least member is greater than-2; Expected error value is made as E _d, E _dBe generally less than 0.01; Variances sigma _k∈ [0.01,2];

2. adjust the weight w of neural network hidden layer neuron according to error function (5) _k, central value μ _kAnd variances sigma _kBy adjusting weight w _k, central value μ _kAnd variances sigma _kMake the value convergence E of E (t) _d, w _k, μ _kAnd σ _kTurn to step 3. after adjusting once;

3. calculate the strength of joint m between hidden layer neuron k and output layer neuron y,

Suppose that k and y are interconnective neurons, the intensity M (k of interactive information; Y) depend on average information between neuron k and y, the strength of joint between neuron k and y is:

M(k;y)=H(k)-H(k|y)=H(k)-H(y|k),(6)

Wherein, H (k) is the Shannon entropy of k, and H (y|k) is the Shannon entropy of y under the k condition; By formula (6) as can be known, when neuron k and y are separate, M (k; Y) value is 0; Otherwise, M (k; Y) be positive number; So, M (k; Y) 〉=0, and

M(k;y)≤min(H(k),H(y)).(7)

The intensity of regularization interactive information

$m(k;y)=\frac{M(k;y)}{\min (H\left(k\right),H\left(y\right))},---\left(8\right)$

0≤m (k wherein; Y)≤1, by calculating m (k; Y), can determine correlativity between neuron k and y, i.e. strength of joint; Set m ₀∈ [0,0.01] is in the RBF neural network, as m (k; Y) 〉=m ₀The time illustrate that then the information interaction between neuron k and y is stronger, think to have between k and y to be connected---Fig. 2 .1, the structure of neural network does not need to adjust; As m (k; Y)＜m ₀Shi Ze show between neuron k and y information interaction intensity a little less than, can ignore being connected between neuron k and y---Fig. 2 .2, be connected disconnection between neuron k and y, by structural adjustment, the redundant neuron of neural network hidden layer obtains pruning, suppose that original hidden layer neuron is K, the neuron that needs to prune is N _CutIndividual, then prune back neural network hidden layer neuron and become K-N _CutIndividual, thus the redundance of reduction neural network;

4. calculate the liveness NAi of neuron i,

${\mathrm{NA}}_i=m(i;y)\left|w_i\sqrt{{\mathrm{\θ}}_i\left(x\right)}\right|,---\left(9\right)$

Wherein, if the 3. in the step neural network structure prune, hidden layer neuron is K-N _CutIndividual, i=1 then, 2 ..., K-N _CutNA _iBe hidden layer i neuronic liveness, w _iBe the neuronic connection weights of i hidden layer neuron and output layer, θ _iBe hidden layer i neuronic output, m (i; Y) be hidden layer i neuronic interactive information intensity; If liveness NA _iGreater than liveness threshold value NA _o, NA _o∈ [0.01,0.2], division neuron i adjusts network structure, N _NewBe newly-increased neuron number, its value is 3; If there is 1 neuron to be split into N in the neural network _NewIndividual new neuron; Then the neural network hidden layer neuron becomes K-N after dividing _Cut+ N _New-1;

5. the value of E (t) in the error of calculation function (5) is if E (t) is less than or equal to anticipation error E _dIn time, stop to calculate; Otherwise turn to step 2. to continue training;

(4) test sample book is detected: with the input of test sample book data as the neural network that trains, the output of neural network is the soft measurement result of SVI;

Creativeness of the present invention is mainly reflected in:

(1) the present invention at first by calculating the interactive information relevance function, analyzes RBF neural network hidden layer neuron and the interneuronal strength of joint of output layer, thereby according to interactive information intensity network structure is pruned; Secondly, utilize neuronic liveness function to judge neuronic activity, the neuron stronger to liveness divides; This self-organization RBF neural network can satisfy the information requirement of handling object, improves computing velocity and the information processing capability of neural network;

(2) it is long measuring period to the present invention is directed in the current sewage disposal process sludge settling bulk index SVI, problem that can not online detection, can approach the characteristics of nonlinear function according to neural network, adopt self-organization RBF neural network that SVI is carried out online soft sensor, had characteristics such as real-time is good, good stability, precision height; Thereby saved the complex process of development sensor and reduced operating cost;

(3) the present invention realizes auxiliary variable discharge Q by having adopted self-organization RBF neural network _In, the mapping between dissolved oxygen concentration DO, acidity-basicity ph, biochemical oxygen demand BOD, chemical oxygen demand COD, total nitrogen TN and the SVI, obtained a kind of SVI and Q _In, DO, pH, BOD, COD, TN characteristic relation, be difficult to the problem described with mathematical model thereby solved SVI;

To note especially: the present invention just for convenience, employing be that SVI is carried out soft measurement, same this invention also is applicable to sludge density index (SDI)---SDI, needs only employing principle of the present invention and carries out soft measurement and all should belong to scope of the present invention.

Description of drawings

Fig. 1 is soft measurement neural network topology structure of the present invention;

Fig. 2 is connection layout between neuron k and y;

Fig. 3 is the soft measurement result figure of the present invention, and wherein solid line is the SVI measured value, and dotted line is the soft measured value of SVI;

Fig. 4 is the soft measuring result error figure of the present invention.

Embodiment

The present invention chooses the auxiliary variable Q that measures SVI _In, DO, pH, BOD, COD, TN, wherein Q _InBe discharge, DO is dissolved oxygen concentration in the aeration tank, and pH is the potential of hydrogen of water quality in the aeration tank, and BOD is the biochemical oxygen demand of effluent quality, and COD is the chemical oxygen demand (COD) of effluent quality, and TN is the total nitrogen concentration of effluent quality, Q _InUnit be m ³/ day, pH does not have unit, and other unit is mg/L;

The concrete steps of the soft measurement of SVI are as follows:

(1) sample data is proofreaied and correct; Experimental data is from certain sewage treatment plant water analysis daily sheet in 2008; Remaining 200 groups of data after the data based formula of experiment sample (1) and formula (2) pre-service, 200 groups of data samples are divided into two parts: wherein 100 groups of data are used as training sample, and all the other 100 groups of data are as test sample book, and test sample book is shown in table 1-8;

(2) be designed for the self-organization RBF neural network initial topology structure of the soft measurement of SVI; Fig. 1 is the SVI neural network soft sensor model, and its input is respectively Q _In, DO, pH, BOD, COD, TN, adopt the connected mode of 6-5-1, i.e. 6 input neurons, 5 hidden layer neuron, 1 output neuron is output as the value of SVI;

(3) with 100 groups of training sample data neural network trainings after proofreading and correct, in training process, self-organization RBF neural network algorithm concrete steps are as follows:

1. initialization neural network: determine the connected mode of neural network, the neural network input layer is 6, and the initial neuron of hidden layer is 5, and the output layer neuron is 1; Weight w to neural network _kCarry out random assignment, its value is 0 to 1 random number, w ₁=0.26, w ₂=0.41, w ₃=0.33, w ₄=0.31, w ₅=0.42; In order to save the training time, recommend to use central value μ ₁=[0.11 ,-0.23 ,-0.47 ,-0.61,0.12,0.26], μ ₂=[0.81,1.31,1.62 ,-1.43 ,-0.63,0.76], μ ₃=[1.46,0.78,0.13,1.55,0.65,0.23], μ ₄=[0.21,0.87,1.22 ,-0.21 ,-0.91,1.03], μ ₅=[1.26,0.59,1.15 ,-0.32 ,-0.82 ,-1.06]; In order to save the training time, recommend to use variance

σ ₁=σ ₂=σ ₃=σ ₄=σ ₅=0.25; Input is respectively Q _In, DO, pH, BOD, COD, TN value, be output as the value of SVI; Design expectation error E _d=0.005;

2. adjust the weight w of neural network hidden layer neuron according to error function (5) _k, central value μ _kAnd variances sigma _kTo adjust effect in order obtaining preferably, to adopt gradient descent algorithm here; w _k, μ _kAnd σ _kTurn to step 3. after adjusting once;

3. calculate the strength of joint m (k between hidden layer neuron k and output layer neuron y; Y), in order to obtain than compact structure, recommend to set m ₀=0.005, by calculating m (1; Y)=0.6781, m (2; Y)=0.0032, m (3; Y)=2.0117, m (4; Y)=2.5664, m (5; Y)=3.0124; Hidden layer neuron 2 and the interneuronal strength of joint m (2 of output layer; Y) less than m ₀, the interneuronal strength of joint of other hidden layer neuron and output layer is greater than m ₀, so neural network needs to prune, and is connected disconnection between hidden layer neuron 2 and output layer neuron y; Become 4 by pruning back hidden layer neuron number, the structure of neural network obtains adjusting; Fig. 2 .2 has provided structure pruning between hidden layer neuron and output layer neuron;

4. calculate the liveness NA of neuron i according to formula (9) _i, in order to obtain than compact structure, recommend to use liveness threshold value NA _o=0.15, by calculating, neuronic liveness is respectively NA ₁=0.1534, NA ₂=0.1166, NA ₃=0.1237, NA ₄=0.0342; The liveness of hidden layer neuron 1 is greater than liveness threshold value NA _oDivision hidden layer neuron 1, hidden layer neuron 1 splits into 3 hidden layer neuron, division back hidden layer neuron number becomes 6, and the new neuron of first after the division is filled up original neuronic position, and the new neuron of other divisions increases on the basis of original hidden layer neuron, therefore, first fills up the position of original neuron 13 neurons after the division, and other two is hidden layer neuron 5 and neuron 6, adjusts network structure;

5. the value of E (t) is E (t)=0.615 in the error of calculation function (5), E (t)〉0.005, do not reach anticipation error E _d, then turn to step 2. to continue training, stopped training up to E (t)≤0.005 o'clock;

(4) test data is detected: with the input of 100 groups of test sample book data as the neural network that trains, data are in table 1-6, and system's actual measurement output data are in table 7, and the output of neural network is the soft measurement result of SVI, and data are in table 8; Fig. 3 is soft measurement result, and Fig. 4 is soft measuring error;

Fig. 3 is the soft measurement situation map of SVI, X-axis: the input sample point, and Y-axis: sludge settling bulk index SVI, unit is mg/L, and solid line is the real system output valve, and dotted line is the neural network output valve; Fig. 4 is soft measuring error, and the result proves the validity of this method.

Table 1-8 is experimental data of the present invention, and table 1-6 is for detecting sample, and table 7 is the SVI measured value, and table 8 is soft measured value.

Table 1. auxiliary variable Q _InInput value (m ³/ day)

44101	39024	32229	35023	36924	38572	41115	36107	29156	39246
										42393	42857	42911	40376	40923	43830	39165	35791	37419	40983
42217	47665	44314	40841	41157	40078	44365	43080	29414	37312
										38568	38655	34193	36332	32484	37724	36446	35636	34746	34893
37102	41598	38058	40716	40868	36358	40879	44150	45779	41230
										37386	34535	32527	27760	36281	38055	34064	31447	32127	31059
36470	47449	43940	40347	40267	37976	47368	48086	47642	43174
										39891	32257	40498	40221	46669	34669	41824	51520	39421	36131
33251	35789	40106	45191	43308	37615	42596	41948	34647	36967
										34879	34365	34291	34886	38731	39308	44198	39003	34487	35198

The input value (mg/L) of table 2. auxiliary variable DO

1.5	3	2	2.5	1.5	3	3	2	2.5	2
										0.7	1.5	0.7	1.5	2.5	1.5	1.2	1.2	1.2	3

3.5	1.2	3	1	3	1.4	2.5	4.25	3	1
										0.7	1.5	2	3.5	0.9	1	1	1.2	1	1.2
2	1.2	1	3.5	1.5	2	1.2	1	3	0.35
										1.4	1	3	1.2	2	3.5	1	3.5	2	3.5
2.5	1.7	3.5	1.75	1.8	1	2	5	5	4.5
										2	3.5	1.5	2	1.75	1.2	1.2	2	1	1
1	1.5	0.6	2	1.4	1.2	3	1.5	1	1
										1	2	2	3	1.2	3	3	1.2	0.7	0.8

The input value of table 3. auxiliary variable pH

7.8	7.7	7.6	7.9	8	7.8	7.8	7.7	7.7	7.8
										7.9	7.7	7.6	8.1	7.6	7.8	7.4	7.8	7.6	7.6
7.5	7.7	7.8	7.6	8	7.9	7.9	7.8	7.6	8.1
										8.2	7.9	8	7.9	7.5	7.9	7.7	8	7.7	8
7.8	8.2	7.8	8.1	8.1	7.7	7.6	8.1	7.8	7.6
										7.9	7.8	7.8	7.6	7.8	7.8	8.1	7.9	7.7	7.8
7.8	7.8	7.8	7.7	7.9	7.7	7.9	8	7.9	7.7
										7.6	7.5	8.1	8.1	7.8	7.8	7.8	7.3	7.9	7.9
7.6	7.4	7.8	8	7.9	7.8	7.7	7.7	7.5	7.6
										7.5	7.6	7.9	7.7	7.5	7.8	7.7	7.8	7.9	7.7

The input value (mg/L) of table 4. auxiliary variable BOD

81	94.8	81	89.6	95.6	96.5	84.6	90.6	84.2	89
										71.2	51.9	84	78	78.2	90.7	92.1	89.4	76.1	88.3
92.9	88.5	92.5	88.1	89.1	90.4	90.6	82.4	83.3	80.7
										71.1	79.7	82.4	82.6	80.6	80.6	76	85.6	82.8	89
86.2	82.3	82.9	88.7	86.3	72.2	88.3	77.4	92.3	87.5
										89.4	85.9	83.3	86.6	85	79.3	85.1	88.5	87.7	90
80.7	82.9	79.6	79.5	86	87	82.9	82.5	91.2	89.2
										89.4	95	92.7	84.9	86.7	88	88.4	90.1	97.9	90.9
90.2	79.5	93.2	96.1	98	91.8	96.5	85.2	91.5	90.7

94.9

92.8

96

95.5

90.8

88.4

98.5

94.8

92.2

90.9

The input value (mg/L) of table 5. auxiliary variable COD

66.3	69.2	72.7	77.1	57.6	66.1	67.5	53.9	61.8	66.1
										62.5	31.3	64.7	74.4	63.6	71.9	57.4	65.3	70	58.1
67.4	71.7	49.4	72.1	61.7	64.7	65.2	74.5	69.9	72
										66	74.4	70.2	68.1	75.3	71.6	62.5	65.8	74.6	77.8
76.5	71.2	71.9	66.1	69.6	68.1	69.7	74.1	62.3	55.3
										70.8	71.3	66.1	81.8	65.7	77.8	69.7	71.1	62.8	78.3
75	69.8	66.7	68.8	75.9	73.5	67.1	65.9	52.8	68.8
										69.9	42.2	50.8	54.8	64.9	56.4	53.1	56.5	65.7	49.4
53.6	46.5	51.1	32.8	30	43.4	26.7	47.6	44.3	56.4
										50.5	50	50.5	57.5	63.3	65.4	36.4	55.6	55.6	64.6

The input value (mg/L) of table 6. auxiliary variable TN

4.5	6.5	4.5	4.2	4.5	4	6.5	7.5	4	6.5
										5	4.2	3	2.5	2	5.5	6.5	6	5	5.5
4	3.5	6.5	6	6	5.5	4.7	2	4.6	5.5
										5.5	4	5.5	5.5	4	6	6	4.5	7	8
3	4.5	4.5	4.5	6	4.5	7	5.5	7.5	3
										4.5	4	5.5	4	8	4.5	5.4	4.5	4	4.7
5.5	3.5	5	2.5	3.7	3.8	4.5	3.5	4.5	3.5
										3.5	7.5	3.5	4.6	4.5	3.5	2.5	4	3.3	7.5
6.5	6.5	7	3	2.5	3	7.5	5	6	7
										7.5	6	6.5	8.5	3	4	2.5	4.5	6	3

The actual measurement output valve (mg/L) of table 7.SVI

94.2	123.8	111.8	101.8	117.8	80	192	134.6	105	125.8
										123	125	129.6	108.6	126.4	134.8	111.4	120.8	133.4	122
121.6	125.8	128	117	119	123.4	134.6	138.2	135.4	100
										155.2	151.2	120.4	134	106.8	108.4	110.6	99.2	100	110.8

117	105.6	112.6	113	99.6	116.4	110.2	115	108	139
										125.2	128.6	118.8	132.6	160	136.2	163.2	111.8	95.4	119.2
133.4	159.4	113.6	80	91.6	101.4	117.6	107	82	106.4
										120	118.8	128.4	128.4	134.2	133.4	134.2	111.8	117.2	134.8
130.4	132.8	130.2	130.2	137.6	136	138	143.6	88.4	139
										76.4	114	128	134.6	119	151.2	142.2	123	138.2	141.2

The soft measured value (mg/L) of table 8.SVI

94.1	124.1	111.9	101.9	117.7	79.9	191.6	135.1	104.4	125.3
										122.9	124.9	129.5	108.6	125.9	134.6	111.4	120.8	133.2	121.8
121.6	124.3	128.4	116.9	119.2	123.5	134.5	138.2	135.4	100.1
										155.2	150.6	120.1	133.6	107.9	108.4	110.8	99.2	100.6	111.8
116.3	106.4	113.2	113.5	99.5	117.2	110.1	115.8	107.4	138.9
										126.2	128.6	119.3	132.1	159.9	135.2	162.5	112.2	95.5	119.4
132.2	159.4	113.4	80.1	91.7	101.5	117.4	106.7	82.2	106.4
										119.2	118.8	128.2	128.1	134.3	133.4	134.1	112.3	117.1	134.3
130.2	133.1	129.9	130.1	137.6	136.7	138.1	143.4	88.4	139.1
										76.4	113.9	127.8	134.7	119.1	151.1	142.2	123.3	137.6	141.2

Claims (1)

1. the flexible measurement method of a sludge settling bulk index SVI:

(1) sample data is proofreaied and correct;

Be provided with N primary data sample (x (1), x (2) ..., x (N)), average is χ, the deviation of each sample is D (j)=x (j)-χ, j=1,2 ..., N, calculate standard deviation:

$\mathrm{\σ}=\sqrt{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\frac{{(x\left(j\right)-\mathrm{\χ})}^2}{N}},---\left(1\right)$

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

|D(j)|≥3σ,j=1,2,...,N,(2)

Think that then sample x (j) is abnormal data, should give rejecting, the data after obtaining proofreading and correct, these data are as training sample and the test sample book of neural network;

(2) be designed for the self-organization RBF neural network initial topology structure of the soft measurement of SVI; Network is divided into three layers: input layer, hidden layer, output layer; Be input as auxiliary variable, be output as sludge settling bulk index SVI;

The initialization neural network: determine the connected mode of initial neural network, the neural network input layer is l, and the initial neuron of hidden layer is K, and the output layer neuron is one; Weights to neural network carry out random assignment; The input of neural network is expressed as x=(x ₁, x ₂..., x _l) ^T, (x ₁, x ₂..., x _l) ^TBe (x ₁, x ₂..., x _l) transposition, the desired output of neural network is expressed as y _dIf total M training sample, then t training sample is x (t)=(x ₁(t), x ₂(t) ..., x _l(t)) ^T, during with t training sample neural network training, the output of self-organization RBF neural network can be described as:

$y\left(t\right)=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{w}_k\left(t\right){\mathrm{\θ}}_k\left(x\left(t\right)\right),---\left(3\right)$

Wherein, K is the initial neuron number of hidden layer, x (t)=(x ₁(t), x ₂(t) ..., x _l(t)) ^TBe input vector, w _kBe the neuronic connection weights of k hidden layer neuron and output layer; θ _kBe the output of k hidden layer neuron, namely

${\mathrm{\θ}}_k\left(x\right)=e^{(-||x-{\mathrm{\μ}}_k||/{\mathrm{\σ}}_k^2)},---\left(4\right)$

Wherein, μ _kBe the central value of k hidden layer neuron, σ _kBe the variance of k hidden layer neuron, k=1,2 ..., K;

The definition error function is

$E\left(t\right)=\frac{1}{M}\underset{t=1}{\overset{M}{\mathrm{\Σ}}}{(y\left(t\right)-y_d\left(t\right))}^T(y\left(t\right)-y_d\left(t\right)),---\left(5\right)$

y _d(t) and y (t) be respectively system's actual measurement output and the neural network output of t training sample, the purpose of training self-organization RBF neural network is that the error function that makes formula (5) define reaches expected error value;

It is characterized in that further comprising the steps of:

(3) with the training sample neural network training after proofreading and correct, be specially:

1. a given RBF neural network, the initial neuron number of hidden layer is K, K is positive integer, initialization neural network weight w _k, its value is 0 to 1 random number; Initialization Center value μ _kμ _kDimension be l, μ _kGreatest member is less than 2 in the initialization, and least member is greater than-2; Expected error value is made as E _dE _dLess than 0.01; Variances sigma _k∈ [0.01,2];

2. adjust the weight w of neural network hidden layer neuron according to error function formula (5) _k, central value μ _kAnd variances sigma _kBy adjusting weight w _k, central value μ _kAnd variances sigma _kMake the value convergence E of E (t) _d, w _k, μ _kAnd σ _kTurn to step 3. after adjusting once;

3. calculate the strength of joint m between hidden layer neuron k and output layer neuron y,

Suppose that k and y are interconnective neurons, the intensity M (k of interactive information; Y) depend on average information between neuron k and y, the strength of joint between neuron k and y is:

M(k;y)=H(k)-H(k|y)=H(k)-H(y|k),(6)

Wherein, H (k) is the Shannon entropy of k, and H (y|k) is the Shannon entropy of y under the k condition; By formula (6) as can be known, when neuron k and y are separate, M (k; Y) value is 0; Otherwise, M (k; Y) be positive number; So, M (k; Y) 〉=0, and

M(k;y)≤min(H(k),H(y)).(7)

The intensity of regularization interactive information

$m(k;y)=\frac{M(k;y)}{\min (H\left(k\right),H\left(y\right))},---\left(8\right)$

0≤m (k wherein; Y)≤1, by calculating m (k; Y), can determine correlativity between neuron k and y, i.e. strength of joint; Set m ₀∈ [0,0.01] is in the RBF neural network, as m (k; Y) 〉=m ₀The time illustrate that then the information interaction between neuron k and y is stronger, think to have between k and y to be connected that the structure of neural network does not need to adjust; As m (k; Y)＜m ₀Shi Ze show between neuron k and y information interaction intensity a little less than, ignore being connected between neuron k and y, be connected disconnection between neuron k and y, by structural adjustment, the redundant neuron of neural network hidden layer obtains pruning, and supposes that original hidden layer neuron is K, and the neuron that needs to prune is N _CutIndividual, then prune back neural network hidden layer neuron and become K-N _CutIndividual, thus the redundance of reduction neural network;

4. calculate the liveness NA of neuron i _i,

${\mathrm{NA}}_i=m(i;y)\left|w_i\sqrt{{\mathrm{\θ}}_i\left(x\right)}\right|,---\left(9\right)$

Wherein, if the 3. in the step neural network structure do not adjust, i=1,2 ..., K; If the 3. in the step neural network structure take place to adjust and prune, hidden layer neuron is K-N _CutIndividual, then adjust, i=1,2 ..., K-N _CutNA _iBe hidden layer i neuronic liveness, w _iBe the neuronic connection weights of i hidden layer neuron and output layer, θ _iBe hidden layer i neuronic output, m (i; Y) be hidden layer i neuronic interactive information intensity; If liveness NA _iGreater than liveness threshold value NA _o, NA _o∈ [0.01,0.2], division neuron i adjusts network structure, N _NewBe newly-increased neuron number, its value is 3; If there is 1 neuron to be split into N in the neural network _NewIndividual new neuron; If the 3. in the step neural network structure do not adjust, then the neural network hidden layer neuron becomes K+N after by division _New-1; If the 3. in the step neural network structure adjust, then the neural network hidden layer neuron becomes K-N after dividing _Cut+ N _New-1;

5. the value of E (t) in the error of calculation function formula (5) is if E (t) is less than or equals to reach anticipation error E _dIn time, stop to calculate; Otherwise turn to step 2. to continue training;

(4) test sample book is detected: with the input of test sample book data as the neural network that trains, the output of neural network is the soft measurement result of SVI.

Images