CN105760934A - Bridge abnormity monitoring restoration method based on wavelet and BP neural network - Google Patents

Abstract

The invention discloses a bridge abnormity monitoring restoration method based on wavelet and a BP neural network, and the method comprises the following steps: 1), carrying out the extraction of monitoring signal change trend information according to the monitoring data condition of one certain piece of monitoring data on a large-span bridge; 2), carrying out measurement point correlation analysis based on a wavelet low-frequency general picture coefficient; 3), carrying out the training of the BP neural network; 4), enabling the monitoring data of a measurement point B which is the highest in relevancy with an abnormal measurement point A and has no abnormality to serve as the input of the BP neural network according to the correlation between the measurement points, and obtaining the monitoring data of the abnormal measurement point after restoration. According to the invention, the method carries out the restoration of the abnormal data through the correlation between the monitoring data of the bridge and the BP neural network.

Description

A kind of bridge thundering observed data restorative procedure based on small echo Yu BP neutral net

Technical field

The present invention relates to Loads of Long-span Bridges Analysis on monitoring data and process field, design a kind of method repaired based on the bridge thundering observed data of small echo with BP neutral net.

Background technology

China's " highway grow up bridge tunnel operation safety management way (exposure draft) " proposes national highway, the operation safety management of provincial highway grand bridge should implement the working policy of " safety first, put prevention first ", suggestion pipe is supported unit and is adopted modern information technologies, progressively set up bridge tunnel safety monitoring system of growing up, grasp overall technology state and the operation condition of bridge tunnel of growing up in time, provide foundation for the research of bridge tunnel operation management of growing up, maintenance, reliability assessment and related science.In structure monitoring process, owing to bridge health monitoring system measurement point is more, moment all needs to automatically pick up substantial amounts of measurement data, and bridge health monitoring systems designer is often concerned with how utilizing these huge measurement data to analyze the health condition judging bridge.However, when these data occur abnormal, just influence whether the judged result that bridge health monitoring is final, cause false alarm.Bridge structure monitor signal abnormal conditions can be largely classified into following four aspect: Outlier Data point；Monitoring missing data；Anomaly trend data；Noisy excessive data.Different data exception Producing reason is also different.Wherein, with anomaly trend data, follow-up Data Analysis Services is had the greatest impact.Long-span bridges health monitoring systems occurs the monitoring point of anomaly trend data is often less simultaneously, these measuring points are all insecure during data exception to data recovers normally, if carried out follow-up data analysis and trend prediction by this section of abnormal data, structure monitoring system overall evaluation to bridge operation situation within this time period will certainly be had a strong impact on.Therefore, we must flow through certain abnormal data restorative procedure anomaly trend data to being likely to occur in structure monitoring process and are repaired, thus obtaining Monitoring Data more accurately.The strain of reflection bridge entirety stress change, deflection monitoring data not only by the effect of self or load of bridge, are additionally subjected to the impact of the environmental factors such as temperature, humidity.So this type of Monitoring Data often shows as non-linear in time domain, the change of timing curve tends not to the function representation determined.In existing research, it is described by neutral net for the expression that this nonlinear data are best.

Summary of the invention

In order to overcome the situation being likely to occur anomaly trend data in structure monitoring process, the present invention provides a kind of bridge thundering observed data restorative procedure based on small echo Yu correlation analysis, and abnormal data is also repaired by the method by dependency and neutral net between bridge monitoring data.

The technical solution adopted for the present invention to solve the technical problems is:

A kind of bridge thundering observed data restorative procedure based on small echo Yu BP neutral net, comprises the steps:

Step 1) according to a certain specific Monitoring Data Monitoring Data situation on Loads of Long-span Bridges, it is monitored the extraction of signal intensity tendency information；

Signal one layer decomposition is as follows:

X (t): bridge structure monitor signal；

cA₁(k): the 1st layer of wavelet low frequency general picture decomposition coefficient；

cD₁(k): the 1st layer of small echo high frequency detail decomposition coefficient；

1st layer of wavelet basis；

Wherein, cA₁(k)、cD₁K () can by below equation gained:

${\mathrm{cA}}_1\left(k\right)=\underset{k}{\Σ}H(2t-k){\mathrm{cA}}_{j-1}\left(k\right)---\left(2\right)$

${\mathrm{cD}}_1\left(k\right)=\underset{k}{\Σ}G(2t-k){\mathrm{cA}}_{j-1}\left(k\right)---\left(3\right)$

In formula (2), (3), t is discrete time series t=1,2 ..., N, k is discrete wavelet offset parameter coefficient, and j is the number of plies of wavelet decomposition, j=1,2 ..., N, H and G is wavelet decomposition wave filter；

By Multiscale Wavelet Decomposition, obtain the low frequency profile information of monitor signal, by the low frequency general picture coefficient input sample as follow-up correlation analysis Yu neural metwork training；

Step 2) based on the measuring point correlation analysis of wavelet low frequency general picture coefficient；

Provide the definition of the related degree model of a kind of different sensors node monitors data:

$G(X_i,Y_i)=\frac{\underset{k=1}{\overset{M}{\Σ}}\left(x_i\right(k)-E(X_i\left)\right)\left(y_i\right(k)-E(Y_i\left)\right)}{{\left(\underset{k=1}{\overset{M}{\Σ}}{\left(x_i\left(k\right)-E\left(X_i\right)\right)}^2\underset{k=1}{\overset{M}{\Σ}}{\left(y_i\left(k\right)-E\left(Y_i\right)\right)}^2\right)}^{\frac{1}{2}}}---\left(4\right)$

Wherein, X_iAnd Y_iRepresent the time series of two different sensors Monitoring Data respectively, amount to M sampled data, x_i(k) and y_iK () represents the Monitoring Data in k moment, E (X respectively_i) and E (Y_i) represent two sensors sampling average respectively, here, degree of association coefficient 0 < G (X_i,Y_i) < 1, its value shows that more greatly the two is more similar in time T；When the two sensors measuring point that script correlation coefficient is significantly high occurs that suddenly correlation coefficient reduces, then judge that this group measuring point Monitoring Data occurs abnormal, it is necessary to detect；

Step 3) training to BP neutral net

Adopting measuring point A Monitoring Data to be analyzed and the wavelet low frequency coefficient the highest with the A degree of association to be trained as inputoutput pair BP neutral net, training step is as follows:

3.1) according to the feature of Monitoring Data and finally require that performance indications carry out netinit, including input layer i, hidden layer node j, output layer node k with connect weight w_ij、w_jk；

3.2) hidden layer output calculates:

$H_j=f(\underset{i=1}{\overset{n}{\&Sigma;}}{w}_{ij}{x}_i-a_j),j=1,2...,l---\left(5\right)$

Wherein, j is node in hidden layer, a_jFor hidden layer threshold value, f is hidden layer excitation function:

$f\left(x\right)=\frac{1}{1+e^{-x}}---\left(6\right)$

3.3) carry out output layer output to calculate, export H according to hidden layer_j, connect weight w_ijWith threshold value b_k, computing network prediction output O_k:

$O_k=\underset{j=1}{\overset{l}{\&Sigma;}}{H}_j{w}_{jk}-b_k,k=1,2,...m---\left(7\right)$

3.4) carry out the calculating of error, export O and desired output Y, computing network forecast error e according to neural network forecast_k:

e_k=Y_k-O_k, k=1,2..., m (8)

3.5) carry out the renewal of weights, update network according to neural network forecast error e and connect weight w_ijAnd w_jk:

w_jk=w_jk+ηH_je_k(9)

$w_{ij}=w_{ij}+{\mathrm{\&eta;H}}_j(1-H_j)x\left(i\right)\underset{k=1}{\overset{m}{\&Sigma;}}{w}_{jk}{e}_k---\left(10\right)$

Wherein: i=1,2..., n；J=1,2..., l；K=1,2 ..., the nodes of m respectively input layer, hidden layer and output layer；

3.6) renewal of threshold value is carried out, according to neural network forecast error e_kUpdate Node B threshold a_j、b_k:

b_k=b_k+e_k(11)

$a_j=a_j+{\mathrm{\&eta;H}}_j(1-H_j)\underset{k=1}{\overset{m}{\&Sigma;}}{w}_{jk}{e}_k---\left(12\right)$

Step 4) according to the dependency between measuring point, using with abnormal measuring point A degree of association is the highest the input as neutral net of the abnormal measuring point B Monitoring Data does not occur, obtain the Monitoring Data after abnormal measuring point is repaired.

The technology of the present invention is contemplated that: wavelet analysis overcomes short time discrete Fourier transform defect in single resolution and has the feature of multiresolution analysis, has the ability characterizing signal local message in time domain and frequency domain.Time window and frequency domain window dynamically can adjust according to the concrete form of signal, and this is wavelet transformation sharpest edges to conventional Fourier Transform just also.Therefore wavelet transformation is highly suitable for extracting low frequency profile information mostly concerned in deflection data.

Beneficial effects of the present invention is mainly manifested in: 1, bridge monitoring data carry out wavelet analysis, extracts the trend profile information paid close attention to the most in Monitoring Data.By extracting the low frequency wavelet general picture coefficient of Monitoring Data, under the premise reducing Monitoring Data surrounding enviroment noise jamming to greatest extent, significantly reduce the time that subsequent samples is analyzed；2, on the basis based on bridge deflection measurement dot relation analysis, the Monitoring Data using normal operation measurement point recovers the abnormal data of other measurement points being associated, set up the neural network model of the single output of single input (can also multi input), and utilize neutral net powerful data approximation characteristic to recover in time domain the data such as non-linear amount of deflection, strain.So the introducing of wavelet low frequency profile information can not only well improve the performance of denoising and substantially increase the efficiency of process in the present invention.It is simultaneous for the feature of bridge monitoring data, the method of this combination can not only well suppress the impact that abnormal data is repaired by Monitoring Data surrounding enviroment noise, the tendency information paid close attention to the most can also be retained simultaneously, can better abnormality sensor node trend data be repaired.

Accompanying drawing explanation

Fig. 1 is the temperature after wavelet decomposition and strain signal low frequency general picture coefficient time-histories figure, and wherein, (a) is temperature monitoring data low frequency general picture sequence；B () is strain monitoring data low frequency general picture sequence.

Fig. 2 is BP neutral net topological diagram.

Fig. 3 is that anomaly trend data repair flow chart.

Fig. 4 is the base plate strain of middle section, 5# pier crossbeam downstream, temperature time-history curves schematic diagram, and wherein, (a) is structure temperature monitoring data sequent；B () is structural strain monitoring data sequent.

Fig. 5 is that strain monitoring abnormal data repairs Comparative result figure.

Fig. 6 is that strain monitoring abnormal data repairs error comparison diagram.

Detailed description of the invention

Below in conjunction with accompanying drawing, the invention will be further described.

With reference to Fig. 1～Fig. 6, a kind of bridge thundering observed data restorative procedure based on small echo Yu BP neutral net, comprise the steps:

Step 1) according to a certain specific Monitoring Data Monitoring Data situation on Loads of Long-span Bridges, it is monitored the extraction of signal intensity tendency information.

Owing to small echo has the feature of multiresolution, by the wavelet decomposition to signal, the localization property of signal can be obtained at high band, and higher frequency resolution and relatively low temporal resolution is obtained in low-frequency range so that the overall trend information of signal is substantially retained in low frequency decomposition coefficient.In bridge monitoring process, the low frequency profile information of signal more can embody the relatedness between each Sensor.The training process of follow-up correlation analysis and neutral net then can be produced impact by low frequency detail section to a great extent.Therefore, it can be obtained the tendency information of each measuring point by wavelet decomposition, thus laying the first stone for subsequent data analysis and process.Signal one layer decomposition is as follows:

X (t): bridge structure monitor signal；

cA₁(k): the 1st layer of wavelet low frequency general picture decomposition coefficient；

cD₁(k): the 1st layer of small echo high frequency detail decomposition coefficient；

1st layer of wavelet basis；

Wherein, cA₁(k)、cD₁K () can by below equation gained:

${\mathrm{cA}}_1\left(k\right)=\underset{k}{\&Sigma;}H(2t-k){\mathrm{cA}}_{j-1}\left(k\right)---\left(2\right)$

${\mathrm{cD}}_1\left(k\right)=\underset{k}{\&Sigma;}G(2t-k){\mathrm{cA}}_{j-1}\left(k\right)---\left(3\right)$

In formula (2), (3), t is discrete time series t=1,2 ..., N, k is discrete wavelet offset parameter coefficient, and j is the number of plies of wavelet decomposition, j=1,2 ..., N, H and G is wavelet decomposition wave filter.

By Multiscale Wavelet Decomposition, we may be monitored the low frequency profile information of signal, by the low frequency general picture coefficient input sample as follow-up correlation analysis Yu neural metwork training.Can farthest reduce, for training sample, the noise jamming that external environment condition difference causes with low frequency general picture coefficient, reduce the model training time simultaneously.

Fig. 1 is the low frequency general picture coefficient time-histories figure after the backplane fiber-optic grating sensor wavelet transformation of middle section, certain bridge 5# pier crossbeam downstream.Respectively strain, deflection signals are carried out wavelet decomposition, and extract the low frequency general picture coefficient training dataset as neutral net, farthest remain and can reflect the bridge corresponding profile information of actual mechanical structure and eliminate the impact on signal of the factors such as external environmental noise.Training sample set is reduced to 350 from 700, and by filter high frequency noise, the degree of association of temperature data Yu deflection data is increased to 0.7719 from 0.6041.

Step 2) based on the measuring point correlation analysis of wavelet low frequency general picture coefficient

Loads of Long-span Bridges is the mechanical system of an infinite degrees of freedom, and is subject to the interaction of various external loads.Therefore, all monitoring responses thereon are all mutually related.Compare to having High relevancy measuring point, it is possible to the effective existence identifying measuring point data unusual condition.According to sensor node physical location, conventional study general determines that it is associated, this method exists very big subjectivity and uncertainty, and the degree for association lacks quantization, also lacks the theory analysis of necessity.Provide the definition of the related degree model of a kind of different sensors node monitors data:

$G(X_i,Y_i)=\frac{\underset{k=1}{\overset{M}{\&Sigma;}}\left(x_i\right(k)-E(X_i\left)\right)\left(y_i\right(k)-E(Y_i\left)\right)}{{\left(\underset{k=1}{\overset{M}{\&Sigma;}}{\left(x_i\left(k\right)-E\left(X_i\right)\right)}^2\underset{k=1}{\overset{M}{\&Sigma;}}{\left(y_i\left(k\right)-E\left(Y_i\right)\right)}^2\right)}^{\frac{1}{2}}}---\left(4\right)$

Wherein, X_iAnd Y_iRepresent the time series of two different sensors Monitoring Data respectively, amount to M sampled data, x_i(k) and y_iK () represents the Monitoring Data in k moment, E (X respectively_i) and E (Y_i) represent two sensors sampling average respectively.Here, degree of association coefficient 0 < G (X_i,Y_i) < 1, its value shows that more greatly the two is more similar in time T.So, degree of association coefficient can reflect each Sensor and measure the degree that is associated of point with other: numerical value is more big and for just, illustrating that two are measured some positive correlations, and namely variation tendency is got over and reached unanimity；Numerical value is more little and be negative, illustrates that two are measured some negative correlation, and namely variation tendency more tends to contrary.Owing to occurring that suddenly correlation coefficient reduces when the significantly high two sensors measuring point of correlation coefficient originally, then we may determine that this group measuring point Monitoring Data occurs extremely, it is necessary to detects.

Step 3) training to BP neutral net

In conjunction with the degree of association of the actual measuring point of bridge, adopting measuring point A Monitoring Data to be analyzed and the wavelet low frequency coefficient the highest with the A degree of association to be trained as inputoutput pair BP neutral net, training step is as follows:

3.1) according to the feature of Monitoring Data and finally require that performance indications carry out netinit, including input layer i, hidden layer node j, output layer node k, weight w is connected_ij、w_jkDeng.

3.2) hidden layer output calculates:

$H_j=f(\underset{i=1}{\overset{n}{\&Sigma;}}{w}_{ij}{x}_i-a_j),j=1,2...,l---\left(5\right)$

Wherein, j is node in hidden layer, a_jFor hidden layer threshold value, f is hidden layer excitation function:

$f\left(x\right)=\frac{1}{1+e^{-x}}---\left(6\right)$

3.3) carry out output layer output to calculate, export H according to hidden layer_j, connect weight w_ijWith threshold value b_k, computing network prediction output O_k:

$O_k=\underset{j=1}{\overset{l}{\&Sigma;}}{H}_j{w}_{jk}-b_k,k=1,2,...m---\left(7\right)$

3.4) carry out the calculating of error, export O and desired output Y, computing network forecast error e according to neural network forecast_k:

e_k=Y_k-O_k, k=1,2..., m (8)

3.5) carry out the renewal of weights, update network according to neural network forecast error e and connect weight w_ijAnd w_jk:

w_jk=w_jk+ηH_je_k(9)

$w_{ij}=w_{ij}+{\mathrm{\&eta;H}}_j(1-H_j)x\left(i\right)\underset{k=1}{\overset{m}{\&Sigma;}}{w}_{jk}{e}_k---\left(10\right)$

Wherein: i=1,2..., n；J=1,2..., l；K=1,2 ..., the nodes of m respectively input layer, hidden layer and output layer.

3.6) renewal of threshold value is carried out, according to neural network forecast error e_kUpdate Node B threshold a_j、b_k:

b_k=b_k+e_k(11)

$a_j=a_j+{\mathrm{\&eta;H}}_j(1-H_j)\underset{k=1}{\overset{m}{\&Sigma;}}{w}_{jk}{e}_k---\left(12\right)$

Trained BP neutral net topological diagram such as Fig. 2.

Step 4) according to the dependency between measuring point, using with abnormal measuring point A degree of association is the highest the input as neutral net of the abnormal measuring point B Monitoring Data does not occur, obtain the Monitoring Data after abnormal measuring point is repaired.

If obtaining the monitor signal after abnormal measuring point is repaired, only the prediction that the measuring point data the highest with abnormal measuring point degree of association in the data exception period carries out neutral net as input sample need to be repaired.The overall flow figure of this method is as shown in Figure 4.Also demonstrate the present invention by experiment and really can meet the reparation requirement for bridge monitoring anomaly trend data.

Experimental verification: in bridge monitoring, main deck strain, temperature data are the leading indicators of reflection bridge force-bearing deformation.Being limited mainly by dead load, Wind effects, the monitoring sampling period is short, and trend is relatively stable.Middle section, 5# pier crossbeam downstream base plate is strained with certain long-span bridges health monitoring fiber-optic grating sensor for object of study, temperature data is studied.Data time series is on March 1st, 2015, amounts to 720 sampled datas.Light grating sensor sample frequency is that 1 time/s. chooses front 700 Monitoring Data here as training sample set, and 20 Monitoring Data of 701-720 are test sample set.Temperature and strain time history curve are as shown in Figure 3.

Here respectively strain, deflection signals are carried out wavelet decomposition, and extract the low frequency general picture coefficient training dataset as neutral net, farthest remain and can reflect the bridge corresponding profile information of actual mechanical structure and eliminate the impact on signal of the factors such as external environmental noise.Wavelet decomposition low frequency general picture coefficient is as shown in Figure 1.

It can be seen that extract after low frequency general picture coefficient through wavelet decomposition, tendency information retains complete, and sample size is reduced to 350 from 700.Can be calculated before and after extraction Monitoring Data general picture coefficient through relevance model, the dependency of temperature and amount of deflection measuring point respectively 0.6041 and 0.7719.This also reflects that extraction low frequency general picture coefficient is effectively reduced, as the method for sample, the impact that different measuring points produces because external environment condition is different from the side.

Here with 20 Monitoring Data of measuring point 701-720 for test sample set, the prediction output after improving with classical BP neutral net, small echo respectively is predicted export structure such as Fig. 5.

From fig. 5, it can be seen that the effect of the abnormal restorative procedure after small echo extraction general picture coefficient is greatly improved compared to classical BP neural network forecast method.Prediction output is fine with desired output fitting effect, and within absolute error substantially remains in 0.3 microstrain, and classic BP neural network forecast output maximum absolute error is close to 1 microstrain.Two kinds of method error comparison diagram such as Fig. 6.

By comparing it appeared that the present invention is in being applied to the reparation of bridge monitoring anomaly trend data, go abnormal data to repair precision to be significantly improved, and obtained the present invention by emulation experiment and can greatly improve the efficiency of operation, therefore can be applied greatly in the middle of Real-time System.Method is run before solving the deficiency of overlong time.

Claims (1)

1. the bridge thundering observed data restorative procedure based on small echo Yu BP neutral net, it is characterised in that: comprise the steps:

Step 1) according to a certain specific Monitoring Data Monitoring Data situation on Loads of Long-span Bridges, it is monitored the extraction of signal intensity tendency information；

Signal one layer decomposition is as follows:

X (t): bridge structure monitor signal；

cA₁(k): the 1st layer of wavelet low frequency general picture decomposition coefficient；

cD₁(k): the 1st layer of small echo high frequency detail decomposition coefficient；

: the 1st layer of wavelet basis；

Wherein, cA₁(k)、cD₁K () is by below equation gained:

${\mathrm{cA}}_1\left(k\right)=\underset{k}{\&Sigma;}H(2t-k){\mathrm{cA}}_{j-1}\left(k\right)---\left(2\right)$

${\mathrm{cD}}_1\left(k\right)=\underset{k}{\&Sigma;}G(2t-k){\mathrm{cA}}_{j-1}\left(k\right)---\left(3\right)$

In formula (2), (3), t is discrete time series t=1,2 ..., N, k is discrete wavelet offset parameter coefficient, and j is the number of plies of wavelet decomposition, j=1,2 ..., N, H and G is wavelet decomposition wave filter；

By Multiscale Wavelet Decomposition, obtain the low frequency profile information of monitor signal, by the low frequency general picture coefficient input sample as follow-up correlation analysis Yu neural metwork training；

Step 2) based on the measuring point correlation analysis of wavelet low frequency general picture coefficient；

Provide the definition of the related degree model of a kind of different sensors node monitors data:

$G(X_i,Y_i)=\frac{\underset{k=1}{\overset{M}{\&Sigma;}}\left(x_i\right(k)-E\left(X_i\right))\left(y_i\right(k)-E\left(Y_i\right))}{{\left(\underset{k=1}{\overset{M}{\&Sigma;}}{\left(x_i\right(k)-E(X_i\left)\right)}^2\underset{k=1}{\overset{M}{\&Sigma;}}{\left(y_i\right(k)-E(Y_i\left)\right)}^2\right)}^{\frac{1}{2}}}---\left(4\right)$

Wherein, X_iAnd Y_iRepresent the time series of two different sensors Monitoring Data respectively, amount to M sampled data, x_i(k) and y_iK () represents the Monitoring Data in k moment, E (X respectively_i) and E (Y_i) represent two sensors sampling average respectively, here, degree of association coefficient 0 < G (X_i,Y_i) < 1, its value shows that more greatly the two is more similar in time T；When the two sensors measuring point that script correlation coefficient is significantly high occurs that suddenly correlation coefficient reduces, then judge that this group measuring point Monitoring Data occurs abnormal, it is necessary to detect；

Step 3) training to BP neutral net

Adopting measuring point A Monitoring Data to be analyzed and the wavelet low frequency coefficient the highest with the A degree of association to be trained as inputoutput pair BP neutral net, training step is as follows:

3.1) according to the feature of Monitoring Data and finally require that performance indications carry out netinit, including input layer i, hidden layer node j, output layer node k with connect weight w_ij、w_jk；

3.2) hidden layer output calculates:

$H_j=f(\underset{i=1}{\overset{n}{\&Sigma;}}{w}_{ij}{x}_i-a_j),j=1,2...,l---\left(5\right)$

Wherein, j is node in hidden layer, a_jFor hidden layer threshold value, f is hidden layer excitation function:

$f\left(x\right)=\frac{1}{1+e^{-x}}---\left(6\right)$

3.3) carry out output layer output to calculate, export H according to hidden layer_j, connect weight w_ijWith threshold value b_k, computing network prediction output O_k:

$O_k=\underset{j=1}{\overset{l}{\&Sigma;}}{H}_j{w}_{jk}-b_k,k=1,2,...m---\left(7\right)$

3.4) carry out the calculating of error, export O and desired output Y, computing network forecast error e according to neural network forecast_k:

e_k=Y_k-O_k, k=1,2..., m (8)

3.5) carry out the renewal of weights, update network according to neural network forecast error e and connect weight w_ijAnd w_jk:

w_jk=w_jk+ηH_je_k(9)

$w_{ij}=w_{ij}+{\mathrm{\&eta;H}}_j(1-H_j)x\left(i\right)\underset{k=1}{\overset{m}{\&Sigma;}}{w}_{jk}{e}_k---\left(10\right)$

Wherein: i=1,2..., n；J=1,2..., l；K=1,2 ..., the nodes of m respectively input layer, hidden layer and output layer；

3.6) renewal of threshold value is carried out, according to neural network forecast error e_kUpdate Node B threshold a_j、b_k:

b_k=b_k+e_k(11)

$a_j=a_j+{\mathrm{\&eta;H}}_j(1-H_j)\underset{k=1}{\overset{m}{\&Sigma;}}{w}_{jk}{e}_k---\left(12\right)$

Step 4) according to the dependency between measuring point, using with abnormal measuring point A degree of association is the highest the input as neutral net of the abnormal measuring point B Monitoring Data does not occur, obtain the Monitoring Data after abnormal measuring point is repaired.