CN103557884A - Multi-sensor data fusion early warning method for monitoring electric transmission line tower - Google Patents

Abstract

A multi-sensor data fusion early warning method for monitoring an electric transmission line tower comprises the following steps that 1) current electric transmission line parameters of detection sensors are obtained in real time; 2) first-class BP neural networks are fused; 3) second-class D-S evidence theories are fused, output of the BP neural networks is evaluated between 0 and 1 and processed to be elementary probabilities to serve as evidences of the D-S evidence theories, and danger classes are obtained. According to the second step, 2.1) fault monitoring is carried out on a single detection sensor: a BP network is adopted to construct a detection sensor output sequence predication model, and assuming n time output samples of the observed sensors as x(1), x(2), x(3),..., x (n), (n+1) time sensor output values are predicted; 2.2) the BP neural network of each detection sensor carries out time data fusion; 2.3) input signals are normalized; 2.4) fusion identification is carried out on characteristic layers. The multi-sensor data fusion early warning method is good in stability, high in reliability and good in real-time performance.

Description

A kind of Fusion method for early warning of electric power line pole tower monitoring

Technical field

The present invention relates to a kind of transmission line of electricity security monitoring field, especially a kind of method for early warning of electric power line pole tower monitoring.

Background technology

In recent years, along with China's economy develops rapidly, rural area is urbanization day by day, urbanize day by day in city, the annual civilian power consumption of China is explosive increase, the rapid expansion length of high-low pressure transmission line of electricity is brought paradox problem, on the one hand, the extravagance of transmission line of electricity allows cities and towns citizen live away from urban district commercial center, reduce urban population concentrations, on the other hand, electric pole distributes and stands in great numbers, particularly traffic route both sides electric wire arrangement bar is as easy as rolling off a log is subject to traffic circulation vehicle and pedestrian because Traffic Accidents damages, thereby cause interrupting along road transmission line of electricity, near residential electricity consumption safety impact, particularly, transmission line of electricity is often subject to disaster, as: typhoon, strong rain, the attack of heavy snow, therefore the transmission line of electricity on-line monitoring early warning mechanism that builds safety and stability is imperative.

Summary of the invention

In order to overcome, less stable, the reliability of the method for early warning of existing electric power line pole tower monitoring is strong, the poor deficiency of real-time, the invention provides a kind ofly have good stability, the Fusion method for early warning of reliability is higher, real-time is good electric power line pole tower monitoring.

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

A Fusion method for early warning for monitoring, described method for early warning comprises the following steps:

1) by installation and measuring sensor on described electric power line pole tower, in order to detect n transmission line parameter, the current transmission line parameter of detecting sensor described in Real-time Obtaining;

2) one-level BP neural network fusion, specific as follows:

2.1) single detecting sensor malfunction monitoring: adopt BP network to set up detecting sensor output sequence forecast model, suppose that sensor t the moment output sample having observed is x (1), x (2), x (3), x (t), prediction t+1 is sensor output value constantly, and detailed process is as follows:

2.1.1) training of BP neural network: sensor collection moment data are as the input of BP neural network, under the effect of initial weight, the predicted value of a monitoring variable of output, predicted value and the actual monitoring variable recording compare, if their error is greater than the global error mean value E of setting, continue input training data, until the value that the error between measured value and network output valve is less than setting with regard to deconditioning or when frequency of training is greater than setting value also deconditioning, so far network training completes, and each weights of BP neural network are determined;

2.1.2) off-line is set up after Network Prediction Model, with the front m step sample x (k-m+1) of the actual output of sensor, x (k-m+2) ..., x (n) prediction sensor n+1 step

, sensor n+1 step actual output x (n+1) and prediction output

compare, this value and measured value are carried out to trend analysis, if difference judges that over threshold values this detecting sensor breaks down between measured value and predicted value, these detecting sensor data are invalid data, otherwise described detecting sensor data are valid data;

2.2) the BP neural network of single detecting sensor merges time data: establish detecting sensor buffer memory n wheel data, n value regulates according to demand, definition of T is the matching cycle, it represents the sensor node collection n wheel needed time of data, each node monitoring of environmental periodically in network, and Monitoring Data is stored in nodal cache; After buffer memory is full of by data, node utilizes the Monitoring Data collection (x in buffer memory _j, y _j), j=1,2 ..., a BP neural network of n structure training, wherein, time x _jas input parameter, the transmission line parameter y relative with this time _jas output parameter, due to the parameter of BP neural network, the transport function between layer, before node is sent out, arrange, detecting sensor only need send to aggregation node by the weights and bias of the BP neural network training, then, detecting sensor empties buffer memory, for next round Data Collection is prepared;

2.3) input signal is normalized, its average of input signal that makes all samples is close to 0 or compare very little with its mean square deviation variance; Adopting normalized method is linear transformation

illustrate:

the value that is respectively conversion front and back, max, min are respectively maximal value and the minimum value of sample;

2.4) characteristic layer fusion recognition: BP network classifier is first identified the sample in training set, is written into network topology and parameter setting, and initialization weights, are written into characteristic to be identified, and network carries out forward calculation, provides recognition result;

3) secondary D-S evidence theory fusion, the output of described BP neural network value between 0-1, conduct is that elementary probability is as the evidence of D-S evidence theory after treatment; Specific as follows:

3.1) establish neural network and be output as y ₁, y ₂..., y _n,, get

i=1,2 ..., n.By y' _iafter the output of BP network being normalized as elementary probability assignment, can be used as framework of identification;

3.2) adopt the decision-making technique based on elementary probability assignment, as follows:

$m\left(A_i\right)=\frac{y\left(A_i\right)}{s}---\left(8\right)$

$s=\underset{i=1}{\overset{5}{\mathrm{\Σ}}}y\left(A_i\right)+E_n---\left(9\right)$

M(A _i) be the elementary probability value of i fault mode of sample, the error amount of BP neural network, as the uncertainty m (θ) of D-S evidence theory, has been realized the abridged edition probability assignments of D-S evidence theory, wherein, A _ifor danger classes, i=1,2, ..., n, y (A _i) be the diagnosis output of BP neural network; E _nnetwork sample error.

Further, in described step 1), described transmission of electricity parameter comprises temperature, humidity and wind speed, wind direction, icing, conductor temperature, windage yaw, sag, waves, around condition of construction and shaft tower angle of inclination.Also can adopt other parameters.

Further again, described step 2.2), single detecting sensor is carried out data-gathering process:

2.2.1) numerical value and threshold values initialization, give each connection weight { w _i, { θ _iand threshold values { θ _i, { q} gives the random value between [0,1], and neural network propagated forward signal is calculated in given input and output;

2.2.2) input---the output signal of-hidden layer is

$c_i=f[{\mathrm{\Σ}}_{i=1}^m{w}_{i}x-{\mathrm{\θ}}_i]---\left(1\right)$

Wherein neural transferring function is

The output signal of 2.2.3) hidden layer---output layer and input layer is:

$b=\underset{i=1}{\overset{r}{\mathrm{\Σ}}}{v}_i{c}_i-q---\left(2\right)$

2.2.4) revise weights: from output layer, by error signal along connecting path backpropagation, in order to revise weights and bias,

v _i(N+1)＝v _i(N)+adc _i

w _i(N+1)＝w _i(N)+βe _ix (3)

q(N+1)＝q(N)-ad

p _i(N+1)＝p _i(N)-βe _i

Wherein, 0 < α, β < 1 is learning coefficient, general error d=(y-b) b (1-b) of each unit of output layer, the general error e of each unit of hidden layer _i=dv _ic _i(1-c _i).

2.2.5) network carries out learning training until reach error precision requirement,

$E\left(n\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{E}_j=\frac{1}{n}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{(y_j-b_j)}^2---\left(4\right)$

E (n) < ε wherein, 0≤ε≤1, is error precision requirement, comes as required given.

Technical conceive of the present invention is: adopt based on BP neural network+D-S evidence theory electric power line pole tower multi-sensor data-fusion system, can be effectively early warning electric power line pole tower abnormal conditions in time, thereby for guaranteeing that whole transmission line of electricity safe and highly efficient operation contributes.

BP neural network is counterpropagation network, it is a multilayer feedforward neural networks, it utilizes error backpropagation algorithm to train network, in monitoring analysis work, BP neural network only need to be considered the selection of factor of influence, and not needing to know mathematical expression relation definite between factor of influence and the output factor, between pairing effect amount and the amount of impact, relation is indefinite or may exist in the work transmission line application of complex nonlinear relation as suitable.Fig. 1 is BP neural network structure figure, three-layer network, consists of: input layer, hidden layer and output layer.

The learning process of BP neural network is comprised of two parts: forward-propagating and backpropagation.In the time of forward-propagating, input message is through input layer, and the neuron of hidden layer is successively processed, and propagates into forward output layer, provides result; If can not get desired output at output layer, proceed to reverse communication process, error is returned along original neuron connecting path.In return course, by each layer of interneuronal connection weights, error signal is reduced, and then proceed to forward-propagating process, iterate, until error is less than specified value.BP neural network model, in the indefinite situation of the mathematical relation between input layer and output layer, can be by the learning simulation of Monitoring Data is gone out to its related law, as long as find out the situation of its measured data and influence factor, can be to data analysis and prediction according to this.

D-S evidence theory is the work of doing based on A.P.Dempster the earliest, attempts to go simulation uncertain by a probable range rather than single probable value.In evidence theory, introduced belief function and measured uncertainty, and quoted likelihood function and process the uncertainty causing due to " not knowing ", and needn't provide in advance the prior probability of knowledge.If there are two evidence body A ₁,

Ω is identification framework, and each element objectionable intermingling in Ω meets

$\begin{array}{c}m\left(A_1\right)=\max \{m\left(A_i\right),A_i\&Subset;\mathrm{\&Omega;}\}\\ m\left(A_2\right)=\max \{m\left(A_i\right),A_i\&NotEqual;A_1\}\end{array}---\left(5\right)$

If have $\left\{\begin{array}{c}m\left(A_1\right)-m\left(A_2\right)>{\mathrm{\&epsiv;}}_1\\ m\left(\mathrm{\&Omega;}\right)<{\mathrm{\&epsiv;}}_2\\ m\left(A_1\right)>m\left(\mathrm{\&Omega;}\right)\end{array}\right.---\left(6\right)$

A ₁for court verdict, i.e. A ₁break down, wherein, ε ₁, ε ₂for predefined thresholding.Due to the basic reliability distribution of evidence theory be expert on the basis of obtained evidence, the digitized representations according to individual experience to the degree of support of the different propositions of identification framework, subjectivity is very strong.Therefore the output of BP neural network solves combination effectively as the source of evidence of D-S evidence theory.

Beneficial effect of the present invention is mainly manifested in: have good stability, reliability is higher, real-time is good.

Accompanying drawing explanation

Fig. 1 is the schematic diagram of three layers of BP neural network.

Fig. 2 is the schematic diagram of multi-sensor data-fusion system.

Fig. 3 is the schematic diagram of secondary fusion structure.

Fig. 4 is the process flow diagram of the fault monitoring method based on the single detecting sensor of BP neural network.

Fig. 5 is the training schematic diagram of BP neural network, wherein, (a) is BP training process convergence curve, is (b) that actual value and predicted value error are in (0.05-0.05) scope.

Fig. 6 is the schematic diagram of BP neural network in the regional electric power line pole tower data fusion of monitoring.

Fig. 7 utilizes the BP neural network training to carry out the schematic diagram of data fusion.

Embodiment

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

With reference to Fig. 1～Fig. 7, a kind of Fusion method for early warning of electric power line pole tower monitoring, described method for early warning comprises the following steps:

1) by installation and measuring sensor on described electric power line pole tower, in order to detect n transmission line parameter, n is more than or equal to 5 natural number; The current transmission line parameter of detecting sensor described in Real-time Obtaining;

2) one-level BP neural network fusion, specific as follows:

2.1) single detecting sensor malfunction monitoring: adopt BP network to set up detecting sensor output sequence forecast model, suppose that sensor t the moment output sample having observed is x (1), x (2), x (3),, x (t), prediction t+1 is sensor output value constantly

detailed process is as follows:

2.1.1) training of BP neural network: sensor collection moment data are as the input of BP neural network, under the effect of initial weight, the predicted value of a monitoring variable of output, predicted value and the actual monitoring variable recording compare, if their error is greater than the global error mean value E of setting, continue input training data, until the value that the error between measured value and network output valve is less than setting with regard to deconditioning or when frequency of training is greater than setting value also deconditioning, so far network training completes, and each weights of BP neural network are determined;

2.1.2) off-line is set up after Network Prediction Model, with the front m step sample x (k-m+1) of the actual output of sensor, x (k-m+2) ..., x (n) prediction sensor n+1 step sensor n+1 step actual output x (n+1) and prediction output

output compares, this value and measured value are carried out to trend analysis, if difference judges that over threshold values this detecting sensor breaks down between measured value and predicted value, these detecting sensor data are invalid data, otherwise described detecting sensor data are valid data;

2.2) the BP neural network of single detecting sensor merges time data: establish detecting sensor buffer memory n wheel data, n value regulates according to demand, definition of T is the matching cycle, it represents the sensor node collection n wheel needed time of data, each node monitoring of environmental periodically in network, and Monitoring Data is stored in nodal cache; After buffer memory is full of by data, node utilizes the Monitoring Data collection (x in buffer memory _j, y _j), j=1,2 ..., n, a BP neural network of structure training, wherein, time x _jas input parameter, the transmission line parameter y relative with this time _jas output parameter, due to the parameter of BP neural network, the transport function between layer, before node is sent out, arrange, detecting sensor only need send to aggregation node by the weights and bias of the BP neural network training, then, detecting sensor empties buffer memory, for next round Data Collection is prepared;

2.3) input signal is normalized, its average of input signal that makes all samples is close to 0 or compare very little with its mean square deviation variance; Adopting normalized method is linear transformation

illustrate:

the value that is respectively conversion front and back, max, min are respectively maximal value and the minimum value of sample;

2.4) characteristic layer fusion recognition: BP network classifier is first identified the sample in training set, is written into network topology and parameter setting, and initialization weights, are written into characteristic to be identified, and network carries out forward calculation, provides recognition result;

3) secondary D-S evidence theory fusion, the output of described BP neural network value between 0～1, conduct is that elementary probability is as the evidence of D-S evidence theory after treatment; Specific as follows:

3.1) establish neural network and be output as y ₁, y ₂..., y _n, get by y' _iafter the output of BP network being normalized as elementary probability assignment, can be used as framework of identification;

3.2) adopt the decision-making technique based on elementary probability assignment, as follows:

$m=\left(A_i\right)=\frac{y\left(A_i\right)}{s}---\left(8\right)$

$s=\underset{i=1}{\overset{5}{\mathrm{\&Sigma;}}}y\left(A_i\right)+E_n---\left(9\right)$

M(A _i) be the elementary probability value of i fault mode of sample, the error amount of BP neural network, as the uncertainty m (θ) of D-S evidence theory, has been realized the abridged edition probability assignments of D-S evidence theory, wherein, A _ifor danger classes, i=1,2, ..., n, y (A _i) be the diagnosis output of BP neural network; E _nnetwork sample error.

Further, in described step 1), described transmission of electricity parameter comprises temperature, humidity and wind speed, wind direction, icing, conductor temperature, windage yaw, sag, waves, around condition of construction and shaft tower angle of inclination.Also can adopt other parameters.

In the present embodiment, transmission line online monitoring system is mainly by wireless (GSM/GPRS/CDMA) transmission mode, temperature to transmission line of electricity environment, humidity and wind speed, wind direction, icing, conductor temperature, windage yaw, sag, wave, condition of construction around, the parameters such as shaft tower inclination are carried out Real-Time Monitoring, the early warning of the perimeter circuit unusual condition centered by shaft tower is provided, can improve the management level to transmission line of electricity safety and economic operation, and provide necessary reference for the repair based on condition of component work of transmission line of electricity, the required sensor of shaft tower transmission line of electricity has shaft tower obliquity sensor, wire tension sensor, temperature sensor, humidity sensor etc., it is various in style, the sensing data amount gathering is huge.Especially, some shaft towers are under extreme climate environment, as goaf with easily wash away location, for preventing because these shaft towers are toppled over the generation that causes the accident of falling rod disconnection, just need to grasp in time shaft tower inclination development, to take appropriate measures in time.Generally, shaft tower inclination sensor sends to monitoring center by data such as the shaft tower transverse pitch collecting, fore-and-aft tilt, composite inclined by 3G/GPRS/EDGE/CDMA1X, monitoring center carries out data storage, demonstration, statistical report form and analyzes in conjunction with shaft tower self design parameter state parameters such as transverse pitch, fore-and-aft tilts, completes the multiparameter warning function that shaft tower tilts.Can judge in time the development trend that shaft tower tilts, when reaching alarm condition, process in time, be a kind of effective means that on-line monitoring is carried out in pit mining and rainwater Piao brush more regions.For the monitoring of inclination of transmission line tower or the situation correct information that the collapses multi-Sensor Information Fusion Approach of need to being correlated with, carry out early warning differentiation.Fig. 2 is multi-sensor data-fusion system generally.Fig. 2 adopts secondary emerging system structural drawing for Multiple Source Sensor.

In Fig. 3, BP neural network information fusion refers to that the information exchange of multisensor crosses Processing with Neural Network, draw preliminary fusion results, it is that process evidence theory comprehensively produces decision information using the output of BP neural network is not as evidence in the same time that evidence theory information merges.What sensor one-level BP neural network fusion completed is the fusion on feature level, and secondary evidence information fusion is the fusion in decision level.This one deck is to adopt D-S evidence theory inference method, by each evidence on same identification framework, carries out fusion reasoning, finally forms the result of decision.Concrete secondary blending algorithm process is as follows:

1) pick-up transducers data, and carry out pre-service.

2) create BP neural network, from experience database, extract characteristic and carry out neural metwork training.

3) utilize the neural network train to data analysis, during different sensors different, segment data obtains by neural network

To Different Results.Select evidence theory rule merging a plurality of sensing datas effectively.A plurality of sensor collections are from the monitoring parameter signal of complicated transmission line of electricity, for example: conductor temperature, wind speed, shaft tower inclination pulling force, wire tension etc., first these signals pass through pre-service (outlier identification, denoising etc.) and data fusion, form data fusion level.Secondly, the information that data grades of fusion was processed is carried out feature extraction fusion, forms Fusion Features level.Finally on the basis of feature extraction, carry out situation estimation and formulate corresponding dangerous Precaution Tactics according to assessment result.Wherein, in data fusion level security monitoring preprocessing process, except carrying out denoising multi-sensor data, there are class data to should be noted that especially, that is: exceptional value or be called singular point.These type of data are often submerged in the data noise of sensor and are not easy to be found, and the reason that they produce is likely that sensor breaks down.

Based on BP neural network single-sensor node fault monitoring method: adopt BP network to set up sensor output sequence forecast model, suppose that sensor t the moment output sample having observed is x (1), x (2), x (3), x (t), prediction t+1 is sensor output value constantly

, detailed process is as follows:

A) training of BP neural network: sensor collection moment data are as the input of BP neural network, under the effect of initial weight, the predicted value of a monitoring variable of output, predicted value and the actual monitoring variable recording compare, if their error is greater than the global error mean value E of setting, continue input training data, until the value that the error between measured value and network output valve is less than setting with regard to deconditioning or when frequency of training is greater than setting value also deconditioning, so far network training completes, and each weights of BP neural network are determined.Algorithm flow as shown in Figure 4.

B) off-line is set up after Network Prediction Model, during central online application, and with the front m step sample x (k-m+1) of the actual output of sensor, x (k-m+2) ..., x (n) prediction sensor n+1 step , sensor n+1 step actual output x (n+1) and prediction output

compare, this value and measured value are carried out to trend analysis, if difference surpasses threshold values and determines sensor and break down between measured value and predicted value.

The BP neural network of single-sensor node is to time data blending algorithm: the sensor node of take monitoring transmission line of electricity wind speed is explained as example.If node can be taken turns data by buffer memory n, n value can regulate according to demand, and definition of T is the matching cycle, and its represents sensor node collection n wheel needed time of data.Each node monitoring of environmental periodically in network, and Monitoring Data is stored in nodal cache.After buffer memory is full of by data, node utilizes the Monitoring Data collection (x in buffer memory _j, y _j), j=1,2 ..., n, according to a BP neural network of method construct training cited below, wherein, time x _jas input parameter, with right wind speed y of this time _jas output parameter, parameter due to BP neural network, transport function between layer, before sending out, node arranges, like this, node only need by the weights and bias of the BP neural network training and correlation parameter as the initial time of node identification, monitoring periods, this fitting data and closing time certain routing algorithm send to aggregation node.Then, node empties buffer memory, for next round Data Collection is prepared.That individual node carries out data collection algorithm process below:

1. numerical value and threshold values initialization, gives each connection weight { w _i, { v _iand threshold values { θ _i, { q} gives the random value between [0,1], and neural network propagated forward signal is calculated in given input and output.

2. input---the output signal of-hidden layer is

$c_i=f[{\mathrm{\&Sigma;}}_{i=1}^m{w}_{i}x-{\mathrm{\&theta;}}_i]---\left(1\right)$

Wherein $\left(x\right)=\frac{2}{1e^{-2x}}-1$

3. hidden layer---the output signal of output layer and input layer is:

$b=\underset{i=1}{\overset{r}{\mathrm{\&Sigma;}}}{v}_i{c}_i-q---\left(2\right)$

4. revise weights: from output layer, by error signal along connecting path backpropagation, in order to revise weights and bias,

v _i(N+1)＝v _i(N)+adc _i

w _i(N+1)＝w _i(N)+βe _ix

q(N+1)＝q(N)-ad (3)

p _i(N+1)＝p _i(N)-βe _i

5. wherein, 0 < α, β < 1 is learning coefficient, general error d=(y-b) b (1-b) of each unit of output layer, the general error e of each unit of hidden layer _i=dv _ic _i(1-c _i).

6. network carries out learning training until reach error precision requirement,

$E\left(n\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{E}_j=\frac{1}{n}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{(y_j-b_j)}^2---\left(4\right)$

E (n) < ε wherein, 0≤ε≤1, is error precision requirement, comes as required given.

It is after network training process finishes that said process finishes, and network has obtained one group of best weight value and threshold values.This group best weight value is the fusion numerical value to Monitoring Data, and the matching numerical value that can obtain the temperature data of corresponding time after this group numerical value by propagated forward is received by Surveillance center.In order to improve the speed of convergence of neural network, need to be normalized the data that gather.Normalized concrete effect is the statistical distribution of concluding unified samples, first basic measuring unit will be unified, neural network be with sample the statistics in event respectively probability train and predict, normalization is same statistical probability distribution between 0-1: when all sample input signals be all on the occasion of time, the weights that are connected with the first hidden layer neuron can only increase simultaneously or reduce, thereby cause pace of learning very slow.For fear of there is this situation, accelerate e-learning speed, can be normalized input signal, its average of input signal that makes all samples is close to 0 or compare very little with its mean square deviation variance.Adopting normalized method is linear transformation,

illustrate:

the value that is respectively conversion front and back, max, min are respectively maximal value and the minimum value of sample.Utilize the premnmax function in matlab to be normalized, the data after processing are for network training.

Table 1330kv extra high voltage network test sample book (some moment measure)

Numbering	1	2	3	4	5	6	7	8	9	10
											Wind speed (m/s)	0.2	0.2	0.2	0.2	3.0	1.5	1.5	1.3	0.1	3.8
Numbering	11	12	13	14	15	16	17	18	19	20
											Wind speed (m/s)	1.3	1.9	2.3	0.6	1.8	1.0	0.8	2.2	3.2	2.0

Table 2330kv extra high voltage network training sample (normalization)

The Fusion of electric power line pole tower Collapse prediction characteristic layer model: merge in target identification at characteristic layer, each sensor extracts target signature according to the raw measurement data obtaining separately, the target signature vector that fusion center provides each sensor merges, and based on fusion feature vector, target is classified.Fig. 6 is that a kind of electric power line pole tower is at the BP neural network model of key monitoring Regional And Multi-source data fusion.

In this model, using the knowledge data relevant to guarded region, radio sensing network Monitoring Data and weather data as data source.Weather data comprises: gather wind speed, temperature, humidity etc.1) wait for that the area monitoring situation merging is not only relevant to radio sensing network Monitoring Data, but also relevant with geographic position and the monitoring meteorological condition at that time of monitoring; 2) adopt neural net method, because neural network is a kind of effective non-linear approach method, and this just with between radio sensing network Monitoring Data to be merged and remote sensing weather data and key monitoring region relevant knowledge data has nonlinear relation and is consistent.For transmission line of electricity monitoring information, gather territory and process, introduce multisource information fusion technology, can realize between multisensor message complementary sense in the time and space, the information of low-density, low resolution monitoring net of solving is not enough and damaged problem.

In the realization of characteristic layer fusion recognition, BP algorithm partly has two control flows: training flow process and identification process.BP network classifier is first identified the sample in training set.Identification process is namely trained the forward calculation of back-propagation algorithm in flow process: 1) be written into network topology and parameter setting, initialization weights; 2) be written into characteristic to be identified; 3) network carries out forward calculation, provides recognition result.Being calculated as follows of error in BP algorithm: have M output neuron, for P input sample, the total error E of network output.Wherein, y _pjoutput layer j the neuronic network output that p input sample provides, d _pjrepresent p output layer j the neuronic desired output that input sample is corresponding.

$E=\frac{1}{2}\underset{p=0}{\overset{p=P}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{j=M}{\mathrm{\&Sigma;}}}{(y_{\mathrm{pj}}-d_{\mathrm{pj}})}^2---\left(5\right)$

W _ij(t) one deck j neuronic connection weights before representing in the t time training, are used LMS algorithm to carry out backward error propagation, have following BP weights to adjust scheme formula, and wherein t is iterations, and n is Learning Step.

$w_{\mathrm{ij}}(t+1)=w_{\mathrm{ij}}\left(t\right)-\mathrm{\&eta;}\frac{\&PartialD;E}{w_{\mathrm{ij}}}\left(t\right)---\left(6\right)$

By formula, calculated after E, the evaluation of network output error precision adopts and calculates relative error by following formula:

$\mathrm{relativeErr}=\frac{2E}{\underset{p=0}{\overset{p=P}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{j=M}{\mathrm{\&Sigma;}}}{\left(d_{\mathrm{pj}}\right)}^2}---\left(7\right)$

What this patent BP neural network Multiple Source Sensor data fusion adopted is three-layer network model.An i.e. input layer, a hidden layer and an output layer.Input parameter has 5, the sensors such as wind speed, environment temperature, humidity, wire tension, shaft tower inclination.And output dangerous situation has 5, be respectively low, slightly low, normal, middle height, high 5 grades, separately with 1,2,3,4,5 numerals, design 5 input ends and with the BP neural network of 5 output terminals.

Fig. 7 is for utilizing the BP neural network training to carry out data fusion.In some areas, select 13 to there is characteristic ultra-high-tension power transmission line:

Table 3330KV extra high voltage network training sample (normalization)

Numbering	Temperature	Humidity	Wind speed	Wire tension	Tower inclination angle	State status
							1	0.8068	0.0526	0.0271	0.8756	0.0487	1
2	0.9659	0	0.0271	0.9656	0.0957	1
							3	0.4204	0.6316	0.0271	0.9114	0.6380	3
4	0	0	0.0271	1.000	0.1785	1
							5	0.5113	0.5263	0.7838	0.1592	1.0000	3
6	0.5568	0.8070	0.3784	0.6156	0.4186	4
							7	0.5625	0.8070	0.3784	0.5065	0.1436	4
8	0.4034	0.9123	0.3243	0.5592	0.2311	3
							9	0.1932	1.0	0	0.9253	0.7132	2
10	1.0	0.5263	1.00	0	0.4918	4
							11	0.4034	0.9123	0.3243	0.5487	0.3592	3
12	0.3352	0.6491	0.4865	0.4128	0.6756	3
							13	0.8522	0.5614	0.5946	0.3504	0.3479	3
14	0.2386	0.8947	0.1351	0.7464	0	2
							15	0.3466	0.6140	0.4595	0.4750	0.6274	3
16	0.1364	0.4035	0.2433	0.7519	0.1969	2

Electric power line pole tower collapses and monitors the realization of Fusion decision-making level: integration technology is divided from time-space domain and can be divided into spatial information fusion and temporal information fusion, spatial information fusion belongs to a kind of horizontal integration technology, it carries out overall treatment by a plurality of sensor multi-source informations, in conjunction with the characteristic information of each sensor, thereby draw more accurate, reliable conclusion.Owing to existing certain deviation between neural network Output rusults and actual value, easily affect final differentiation result, thereby we consider the method that adopts two-stage to merge, output value between 0-1 due to neural network, can be used as be after treatment elementary probability as the evidence of evidence theory, concrete grammar is as follows:

If neural network is output as y ₁, y ₂..., y _n, get

so just can be by y' _ias elementary probability assignment, after being normalized, the output of BP network can be used as framework of identification, by BP neural network output numerical value, have:

Wind speed:

y _i＝3.1778 3.1778 3.1778 3.1778 4.0948 4.1533 4.1449 3.1202 4.9984 4.14494.0601 3.9650 3.8354 4.0902 4.0602

Humidity:

y _i＝3.000 3.1302 2.0066 3.1302 0.9976 0.9871 0.9871 1.2465 0.9984 0.99761.2465 1.9690 1.5918 1.2135 1.9965 0.3585

Temperature: y _i=4.4172 5.0225 4.0868 2.8866 4.0394 3.996 3.9959 4.0785 3.14354.9957 4.0785 3.9349 4.7197 3.4333 3.9757 3.0424

Wire tension: y _i=1.093 1.093 1.093 1.093 0.0122 2.0024 2.0024 1.9887 1.0004.0519 1.9887 1.9994 1.9755 0.8024 1.9992 1.6358

The defeated transmission tower of high pressure inclination angle: y _i=1.0489 0.9519 2.5524 1.2958 3.9985 1.9873 0.8822 1.04283.0452 2.0112 2.0433 2.8815 1.9878 0.9811 2.5176 1.8073

Utilize evidence theory that aforementioned two spaces are merged, can reduce the impact of other uncertain factor, adopt afterwards the decision-making technique based on elementary probability assignment.The danger classes of this patent is divided into 5 grades, and the identification framework of corresponding evidence theory just comprises 5 kinds of states.Meanwhile, each output of BP neural network, as a corroboration of evidence theory, through conversion, is called the belief assignment of various danger classess under this evidence by the output valve of BP neural network.Output rusults based on BP network is normalized.

$\begin{array}{c}m\left(A_i\right)=\frac{y\left(A_i\right)}{s}---\left(8\right)\\ s=\underset{i=1}{\overset{5}{\mathrm{\&Sigma;}}}y\left(A_i\right)+E_n---\left(9\right)\end{array}$

Be m (A _i) be the elementary probability value of i fault mode of sample, the error amount of BP neural network, as the uncertainty m (θ) of D-S evidence theory, has so just been realized the abridged edition probability assignments of D-S evidence theory.Wherein, A _ifor danger classes (i=1,2,3,4,5); y(A _i) be the diagnosis output of BP neural network; E _nnetwork sample error.In view of reliability and the rationality of fusion results, consider, under sometime, electric power line pole tower is monitored, sensor sample parameter is respectively wind speed, environment temperature, humidity, wire tension, shaft tower inclination angle.And will record parameter and deliver to local fusion center and process.According to DS evidence theory, above-mentioned Output rusults is carried out to probability assignments assignment, basic probability assignment value function is respectively m (A ₁), m (A ₂), m (A ₃), m (A ₄), m (A ₅), as shown in the table:

Table 4

Transmission line of electricity	m(A 1)	m(A 2)	m(A 3)	m(A 4)	m(A 5)
						1	0.0307	0.9530	0.7166	0.1571	0.0535
2	0.0307	1.00	1.000	0.1571	0.0224
						3	0.0307	0.5946	0.5619	0.1571	0.5360
4	0.0307	1.00	0.0	0.1571	0.1327
						5	0.5189	0.2306	0.5397	0.6781	1.0000
6	0.5500	0.2268	0.5211	0.3414	0.3546
						7	0.5456	0.2268	0.5194	0.3414	0
8	0	0.3204	0.5580	0.3485	0.0515
						9	1.00	0.2309	0.1203	0.0235	0.6941
10	0.5456	0.2306	0.9875	1.0000	0.3623
						11	0.5004	0.3204	0.5580	0.3485	0.3726
12	0.4498	0.5811	0.4908	0.3439	0.6416
						13	0.3808	0.4450	0.8582	0.3409	0.3548
14	0.5165	0.3085	0.2138	0.1091	0.0317
						15	0.5005	0.5910	0.5099	0.3438	0.5248
16	0.4616	0	0.0729	0	0.2949

Adopt decision-making level's data fusion evidence theory rule: to an evidence collection S, this evidence is concentrated and contained n bar evidence.Order $q=\frac{1}{n}\underset{1\&le;i\&le;n}{\mathrm{\&Sigma;}}{m}_i\left(A\right),k=\mathrm{\&Sigma;}{A}_i\&cap;B_j\&cap;C_k.....,$ There is composite formula:

m(φ)＝0 (10)

$m\left(A\right)=\underset{A_i\&cap;B_j\&cap;C_k...}{\mathrm{\&Sigma;}}{m}_1\left(A_i\right)m_2\left(B_j\right)m_3\left(C_k\right)...+\mathrm{kq}---\left(11\right)$

Wherein, q is the Average Supports of evidence to A, and k is the conflict factor.Present case result of calculation k=0.9557, MASS numerical value is:

Table 5

Transmission line of electricity	1	2	3	4	5	6	7	8
									MASS	0.0010	0.0006	0.0047	0	0.2392	0.0430	0.0	0.0
Transmission line of electricity	9	10	11	12	13	14	15	16
									MASS	0.0025	0.2458	0.0635	0.1546	0.0961	0.006	0.1486	0

From table-5, No. 10 maximum probabilities that break down of transmission line of electricity, secondly No. 5, transmission line of electricity.

In the present embodiment, distributed multi-sensor information fusion structure, fusion method is applied to the measurement of electric power line pole tower monitoring parameter. utilize BP neural network to there is self study, parallel distributed is processed, the feature of Error Tolerance and robustness, by neural network, carry out multi-sensor information fusion, without any prior imformation, the evidence having overcome in evidence theory fusion method is difficult to obtain, the defect that calculated amount is large, this patent has proposed the advantage that two-stage fusion structure takes full advantage of neural network and evidence theory, using the output of neural network as the evidence of evidence theory, this two level fusing methods have solved the calculation combination blast problem based on D-S inference method effectively, in addition the introducing that temporal information merges is distributed in fusion calculation on each node, effectively improve the computing velocity of whole emerging system, and strengthened the robustness of system, this two level fusing methods are applied among transmission line of electricity control measurement, this method has good feasibility.

Claims (3)

1. the Fusion method for early warning that electric power line pole tower is monitored, is characterized in that: described method for early warning comprises the following steps:

1) by installation and measuring sensor on described electric power line pole tower, in order to detect n transmission line parameter, the current transmission line parameter of detecting sensor described in Real-time Obtaining;

2) one-level BP neural network fusion, specific as follows:

2.1) single detecting sensor malfunction monitoring: adopt BP network to set up detecting sensor output sequence forecast model, suppose that sensor t the moment output sample having observed is x (1), x (2), x (3), x (t), prediction t+1 is sensor output value constantly, and detailed process is as follows:

2.1.1) training of BP neural network: sensor collection moment data are as the input of BP neural network, under the effect of initial weight, the predicted value of a monitoring variable of output, predicted value and the actual monitoring variable recording compare, if their error is greater than the global error mean value E of setting, continue input training data, until the value that the error between measured value and network output valve is less than setting with regard to deconditioning or when frequency of training is greater than setting value also deconditioning, so far network training completes, and each weights of BP neural network are determined;

2.1.2) off-line is set up after Network Prediction Model, with the front m step sample x (k-m+1) of the actual output of sensor, x (k-m+2) ..., x (n) prediction sensor n+1 step

sensor n+1 step actual output x (n+1) and prediction output

compare, this value and measured value are carried out to trend analysis, if difference judges that over threshold values this detecting sensor breaks down between measured value and predicted value, these detecting sensor data are invalid data, otherwise described detecting sensor data are valid data;

2.2) the BP neural network of single detecting sensor merges time data: establish detecting sensor buffer memory n wheel data, n value regulates according to demand, definition of T is the matching cycle, it represents the sensor node collection n wheel needed time of data, each node monitoring of environmental periodically in network, and Monitoring Data is stored in nodal cache; After buffer memory is full of by data, node utilizes the Monitoring Data collection (x in buffer memory _j, y _j), j=1,2 ..., n, a BP neural network of structure training, wherein, time x _jas input parameter, the transmission line parameter y relative with this time _jas output parameter, due to the parameter of BP neural network, the transport function between layer, before node is sent out, arrange, detecting sensor only need send to aggregation node by the weights and bias of the BP neural network training, then, detecting sensor empties buffer memory, for next round Data Collection is prepared;

2.3) input signal is normalized, its average of input signal that makes all samples is close to 0 or compare very littlely with its mean square deviation variance, and adopting normalized method is linear transformation, illustrate:

the value that is respectively conversion front and back, max, min are respectively maximal value and the minimum value of sample;

2.4) characteristic layer fusion recognition: BP network classifier is first identified the sample in training set, is written into network topology and parameter setting, and initialization weights, are written into characteristic to be identified, and network carries out forward calculation, provides recognition result;

3) secondary D-S evidence theory fusion, the output of described BP neural network value between 0-1, conduct is that elementary probability is as the evidence of D-S evidence theory after treatment; Specific as follows:

3.1) establish neural network and be output as y ₁, y ₂..., y _n, get

i=1,2 ..., n, by y' _ias elementary probability assignment, after being normalized, the output of BP network can be used as framework of identification;

3.2) adopt the decision-making technique based on elementary probability assignment, as follows:

$m=\left(A_i\right)=\frac{y\left(A_i\right)}{s}---\left(8\right)$

$s=\underset{i=1}{\overset{5}{\mathrm{\&Sigma;}}}y\left(A_i\right)+E_n---\left(9\right)$

M(A _i) be the elementary probability value of i fault mode of sample, the error amount of BP neural network, as the uncertainty m (θ) of D-S evidence theory, has been realized the abridged edition probability assignments of D-S evidence theory, wherein, A _ifor danger classes, i=1,2, ..., n, y (A _i) be the diagnosis output of BP neural network; E _nnetwork sample error.

2. the Fusion method for early warning of electric power line pole tower monitoring as claimed in claim 1, is characterized in that: in described step 1), described transmission of electricity parameter comprises environment temperature, ambient humidity and wind speed, wire tension and shaft tower inclination angle.

3. the Fusion method for early warning that electric power line pole tower as claimed in claim 1 or 2 is monitored, is characterized in that: described step 2.2), single detecting sensor is carried out data-gathering process:

2.2.1) numerical value and threshold values initialization, give each connection weight { w _i, { θ _iand threshold values { θ _i, { q}, gives the random value between [0,1], and neural network propagated forward signal is calculated in given input and output;

2.2.2) input---the output signal of-hidden layer is

$c_i=f[{\mathrm{\&Sigma;}}_{i=1}^m{w}_{i}x-{\mathrm{\&theta;}}_i]---\left(1\right)$

Wherein neuronic transport function is

The output signal of 2.2.3) hidden layer---output layer and input layer is:

$b=\underset{i=1}{\overset{r}{\mathrm{\&Sigma;}}}{v}_i{c}_i-q---\left(2\right)$

2.2.4) revise weights: from output layer, by error signal along connecting path backpropagation, in order to revise weights and bias,

v _i(N+1)＝v _i(N)+adc _i

w _i(N+1)＝w _i(N)+βe _ix

q(N+1)＝q(N)-ad (3)

p _i(N+1)＝p _i(N)-βe _i

Wherein, 0 < α, β < 1 is learning coefficient, general error d=(y-b) b (1-b) of each unit of output layer, the general error e of each unit of hidden layer _i=dv _ic _i(1-c _i),

2.2.5) network carries out learning training until reach error precision requirement,

$E\left(n\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{E}_j=\frac{1}{n}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{(y_j-b_j)}^2---\left(4\right)$

E (n) < ε wherein, 0≤ε≤1, is error precision requirement, comes as required given.

Images