CN106353651A - Fault location method of acoustic electric joint partial discharge detection based on BP (Back Propagation) network in GIS (Gas Insulated Switchgear) - Google Patents

Abstract

The invention discloses a fault location method of acoustic electric joint partial discharge detection based on a BP (Back Propagation) network in a GIS (Gas Insulated Switchgear). The method comprises the following steps that: 1, a sensor converts the information acquired on site into an electrical signal, and comprises an ultrasonic sensor and an ultra high frequency sensor; 2, the ultrasonic sensor conducts data preprocessing on the converted electrical signal and fuses the preprocessed data, and the ultra high frequency sensor conducts data preprocessing on the converted electrical signal and fuses the preprocessed data; 3, fusing the data fused by the ultrasonic sensor and the data fused by the ultra high frequency sensor by utilizing the BP neural network, and determining the fault location; 4, outputting the fault location. The method provided by the invention solves the problem that the accuracy rate and the correct rate of the fault location identification of an on-line monitoring system based on a single type sensor are low, and can avoid the false report, missing report and non-report phenomena of the on-line detection system based on the single type sensor.

Description

Acoustoelectric combined detection Fault Locating Method is put based on the gis office of bp network

Technical field

The present invention relates to being related to electronic technology and control field and in particular to a kind of gis office sound reproduction based on bp network is electric Joint-detection Fault Locating Method.

Background technology

Gas insulated combined electrical equipment is all to be enclosed in the plurality of devices such as chopper, disconnecting switch, earthed switch, bus Full of the packet type switch electrical equipment in sulfur hexafluoride gas metal shell, gas insulated combined electrical equipment is in high-voltage testing room Key equipment, once break down it would be possible to cause electrical network major accident occur.Insulation reduction is gas insulated combined electrical equipment The main cause of equipment fault, carries out online office to gas insulated combined electrical equipment (gas insulated switchgear, gis) Gis built-in electrical insulation situation, prevention gis insulation fault tripping operation can be effectively grasped in portion's electric discharge (partial discharge, pd) detection Cause power grid accident.

Gis shelf depreciation can produce sound wave and electromagnetic signal, and particle and the shelf depreciation of beating is two acoustic emission sources, The sound wave propagated in chamber outer wall also has shear wave in addition to compressional wave, and by ultrasound probe, ultrasonic Detection Method detects that pd produces super Detecting pd signal, hyperfrequency method (ultra high frequency, uhf) receives pd by antenna and produces for sound wave and vibration signal Raw 300～3000mhz frequency range uhf electromagnetic wave signal is detecting pd signal.Both approaches are at present in the inspection of gis shelf depreciation More effective method in survey field.Publication number 105807190a discloses a kind of gis local discharge superhigh frequency live detection side Method, uhf sensor receives the electromagnetic pulse signal that gis shelf depreciation produces, and electromagnetic pulse signal is converted into high-frequency electrical After pressure signal, Partial discharge detector is transferred to by shielded cable；Wireless power frequency component generating meanss are examined to shelf depreciation simultaneously Survey instrument transmitting power-frequency voltage signal；Partial discharge detector carries out data to high-frequency voltage signal and power-frequency voltage signal and parses To prpd electric discharge collection of illustrative plates and hyperfrequency Discharge pulse waveform；Partial discharge detector is to prpd electric discharge collection of illustrative plates and hyperfrequency electric discharge Impulse waveform carries out prpd cluster analyses and impulse waveform time frequency analysis respectively, and according to analysis result identification described gis local The electric discharge type of electric discharge.

Supercritical ultrasonics technology is larger by live noise jamming, and ultra-high-frequency detection method can not accurately carry out fault location.Simultaneously The decay to during pop one's head in is very fast in gis internal transmission for two kinds of signals, increased ultrasound wave or uhf sensor electric discharge letter Number collection and the difficulty such as Filtering Analysis, so two kinds of single methods to be accurately positioned abort situation effect unsatisfactory.Inside gis 4 kinds of insulation such as simulation outthrust a class defect, attachment b class defect, insulator air gap c class defect and free microgranule d class defect lack Fall into, carry out fault detect with this two methods, detection atlas analysis are understood: ultrasonic Detection Method is to d type free metal The pd Detection results that grain defect causes are the most obvious, to b class insulator attachment pollutant defect discharge examination and inconspicuous；Hyperfrequency The pd Detection results in detection method, a metalloid outthrust and c class insulator void defects being caused are the most obvious, to d type free Metal particle defect discharge examination effect is worst.

Above-mentioned two kinds of on-line checking based on single type sensor presently, there are different problems: supercritical ultrasonics technology is subject to existing Field noise jamming is larger, and ultra-high-frequency detection method can not accurately carry out fault location.

Content of the invention

In view of this, the purpose of the present invention is for the deficiencies in the prior art, provides and a kind of is put based on the gis office of bp network Acoustoelectric combined detection Fault Locating Method, can be prevented effectively from based on single type sensor on-line measuring device exist wrong report, Fail to report and do not report phenomenon, improve rapidity and the accuracy of fault detect.

For solving above-mentioned technical problem, the technical solution used in the present invention is as follows:

A kind of acoustoelectric combined detection Fault Locating Method is put based on the gis office of bp network, comprise the following steps:

Step 1, the information each collecting at the scene is converted to the signal of telecommunication by sensor；Described sensor includes ultrasound wave Sensor and uhf sensor；

Step 2, the signal of telecommunication after changing is carried out data prediction by described ultrasonic sensor, and by pretreated number According to being merged；The signal of telecommunication after changing is carried out data prediction by described uhf sensor, and by pretreated data Merged；

Step 3, after being merged to the data after ultrasonic sensor fusion and uhf sensor using bp neutral net Data is merged, failure judgement position；

Step 4, exports abort situation.

Preferably, the neuron function of described bp neutral net is using bipolarity s type function: Wherein e is natural constant, and a is variable.

Preferably, the data fusion step of described step 3 is as follows:

1) initialize: setting network each layer weights, the initial value of threshold values are less random number battle array, and cycle-index is maximum ( Big value is 1000)；

By sensor gathered data processed after, as training sample；If input quantity is x=(x¹,x²,…,x^α), correspondence is output as y=(y¹,y²,…,y^p), the desired output of network is d=(d¹,d²,…,d^p)；Being provided with α sample is instruction Practice data, ifExport for p-th sample,WithIt is respectively the output of the first and second hidden layers, ω_uvFor connection weight；Then each node is corresponding defeated Go out for:

$\left\{\begin{array}{c}{g}_j^p=f\left(\underset{i=0}{\overset{no}{\σ}}{\mathrm{\ω}}_{ji}{x}_i^p\right)\\ h_k^p=f\left(\underset{j=0}{\overset{n1}{\σ}}{\mathrm{\ω}}_{kj}{g}_i^p\right)\\ y_j^p=f\left(\underset{k=0}{\overset{n2}{\σ}}{\mathrm{\ω}}_{lk}{x}_k^p\right)\end{array}\right.---\left(4\right)$

In formula: α input sample,For inputting the output result of the first hidden layer after network, J/p/k all represents variable 1,2,3 ... α.Represent the second hidden layer output result after input network,Represent final output result after input network, be p-th sample after input network Reality output；ω_uvFor the connection weight between each neuron；

3) error back propagation；

From output layer, hidden layer to input layer, with the reverse error signal of each layer being calculated based on gradient steepest descent method, obtain To weighed value adjusting pattern, equivalent error δ computing formula is:

$\left\{\begin{array}{c}{\mathrm{\δ}}_j^p=f^{\′}\left({\mathrm{net}}_j^p\right)(t_j^p-y_j^p)\\ {\mathrm{\δ}}_k^p=f^{\′}\left({\mathrm{net}}_k^p\right)(t_k^p-y_k^p)\\ {\mathrm{\δ}}_l^p=f^{\′}\left({\mathrm{net}}_l^p\right)(t_l^p-y_l^p)\end{array}\right.---\left(5\right);$

The equivalent error of each layer neuron is δ, and t is the desired value that each layer neuron sets, and y is the reality of each layer neuron Output valve；

4) each layer connection weight correction, computing formula is:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{lk}(n+1)={\mathrm{\ω}}_{lk}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\σ}}{\mathrm{\δ}}_l^p{h}_k^p\right)\\ {\mathrm{\ω}}_{kj}(n+1)={\mathrm{\ω}}_{kj}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\σ}}{\mathrm{\δ}}_k^p{g}_j^p\right)\\ {\mathrm{\ω}}_{ji}(n+1)={\mathrm{\ω}}_{ji}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\σ}}{\mathrm{\δ}}_j^p{x}_i^p\right)\end{array}\right.---\left(6\right);$

m_cFor factor of momentum, take m_c=0.9, according to additional guide vanes Rule of judgment, now kth step error sum of squares e (k) ＞ e (k-1)；The equivalent error of each layer neuron is α input sample of δ,For after input network first The output result of hidden layer, j/p/k all represents variable 1,2,3 ... α.Represent second after input network Hidden layer output result,Represent final output result after input network, input net It is the reality output of p-th sample after network；ω_uvFor the connection weight between each neuron；

5) according to new connection weight, carry out positive calculating；Judge each learning sample (x^p, t^p) and output layer Whether each neuron meets setting accuracy, if meeting, output result；Otherwise return to step 2) continue training.Judge every One learning sample (x^p, t^p) and each neuron of output layer whether meet setting accuracy, according to operating experience set essence Degree value is 0.001.If meeting, output result；Otherwise return to step 2) continue training.

The invention has the beneficial effects as follows:

The present invention is based on two kinds of dissimilar sensor information fusion, carry out joint on-line checking by two methods, fortune Played a game with multi-sensor information fusion technology and put data fusion decision-making, it is to avoid ultrasound wave, the single online side of hyperfrequency both Method it is impossible to gis apparatus local discharge fault is carried out fully effective positioning it is impossible to meet " State Grid Corporation of China's high-voltage switch gear sets Standby on-line measuring device specification " requirement problem, improve rapidity and the accuracy of fault detect.

The present invention analyzes signal when shelf depreciation produces, in conjunction with the inspection of Current electronic information, control theory subject and electric power Survey technology, can be prevented effectively from the wrong report based on the presence of single type sensor on-line measuring device, fail to report and do not report phenomenon.Acoustic-electric Co-located detecting system adopts reaching time-difference tdoa (time different of arrival, tdoa) method, using bp (back propagation) neutral net carries out ultrasound wave and hyperfrequency method sensor acquisition data are carried out merging differentiating The acoustoelectric combined detection method of abort situation uses multi-information merging technology, will from a certain discharge source ultrasonic with electromagnetism two kinds of different letters In addition intelligence synthesizes breath, can maximize favourable factors and minimize unfavourable ones, resource complementation, compensate for the deficiency of single type detection method.

Brief description

Fig. 1 is the schematic view of the mounting position of ultrasound probe.

Fig. 2 is method of the present invention flow chart.

Fig. 3 is the topology diagram of bp network model of the present invention.

Fig. 4 is the method flow diagram of training network model of the present invention.

Specific embodiment

Below in conjunction with the accompanying drawings inventive technique scheme is further illustrated:

Fig. 1 is the installation site figure of ultrasound probe.In Fig. 1: ultrasound probe 1, point of discharge 2, compressional wave 3, echo 4, Shell 5, bus 6, disc insulator 7, optical cable 8, common n of sensor.One of Partial Discharge Detection main purpose is quick determination Discharge position, ultrasound wave and ultra-high-frequency detection method carry out space orientation using the method for time difference mostly.Ultrasonic Detection Method Detect that the ultrasonic vibration signal that involves that pd produces detects pd signal, hyperfrequency method (ultra high by ultrasound probe 1 Frequency, uhf) pd signal is detected by 300～3000mhz frequency range uhf electromagnetic wave signal that antenna receives pd generation. Sonac receive time of same electric discharge sound wave because the installation site on gis is different difference, sonac With discharge source apart from the relation of s it is: s=t_n×c.Wherein: t_nReach the time of n sensor for electric discharge ultrasonic signal；C is Shear wave transfer rate in the housing.By record is synchronized to the time difference of certain identical signal, amplitude, obtain away from multiple biographies Sensor apart from s₁,s₂,…,s_n, by the combinative analysiss of required distance above are determined with the coordinate of pd.Ultrasonic method is subject to More serious to live noise jamming, the decay to during pop one's head in is very fast in gis internal transmission for ultrasonic signal simultaneously.Hyperfrequency Shelf depreciation positioning principle is similar to said method, but its reception is electromagnetic wave signal, and velocity of electromagnetic wave is approximate with the light velocity, then Plus live electromagnetic interference, the decay to during pop one's head in is very fast in gis internal transmission for electric discharge electromagnetic wave signal.Ultrasonic from increased The difficulty such as ripple or the collection of uhf sensor discharge signal and Filtering Analysis, so two kinds of single methods are accurately positioned fault bit Put effect unsatisfactory.

For the problems referred to above, present invention proposition is a kind of to put acoustoelectric combined detection fault location side based on the gis office of bp network Method, its flow chart is as shown in Figure 2.The present invention is made up of three parts: the 1) set of multiclass sensor, by ultrasound wave, hyperfrequency two Plant dissimilar sensor to constitute；2) sensor acquisition information fusion part, same after data being carried out with pretreatment and processing Prime number evidence is merged；3) pd fault location part, is adopted tdoa method to ultrasound wave and hyperfrequency method, is entered using bp neutral net Row data fusion, Judging fault position.

The present invention specifically includes following steps:

Step 1, the information each collecting at the scene is converted to the signal of telecommunication by sensor；Described sensor includes ultrasound wave Sensor and uhf sensor；

Step 2, the signal of telecommunication after changing is carried out data prediction by described ultrasonic sensor, and by pretreated number According to being merged；The signal of telecommunication after changing is carried out data prediction by described uhf sensor, and by pretreated data Merged；

Step 3, after being merged to the data after ultrasonic sensor fusion and uhf sensor using bp neutral net Data is merged, failure judgement position；

Step 4, exports abort situation.

Embodiment 1:

The present invention adopts the structure of 1 uhf sensor and 6 ultrasonic sensor joint-detection, is known by permutation and combination Knowledge can obtain 20 tdoa positioning subsystems, i.e. 20 positioning targets.Adaptive-learning-rate with momentum using bp neutral net is adjusted Whole algorithm carries out convergence analysis to data, and step is as follows:

1) the neural network model of selector assembly system demand, as shown in Figure 3.

Network topology structure: 12 × 20 × 3；

Input layer: comprise 12 nodes, correspond to three ultrasound senor position coordinate (x of each sample respectively_i,y_i, z_i), and three ultrasonic sensors receive time difference τ of Partial discharge signal with uhf sensor_i.

Hidden layer: 20 nodes, neuron function selection bipolarity s type function:

$y=f\left(a\right)=-\frac{1}{2}+\frac{1}{1+e^{-a}};$

Wherein e is natural constant, and a is variable.

Output layer: 3 nodes, for the final space coordinatess positioning target.

In addition, learning rate is 0.05；Dynamic vector is 0.9；Maximum cycle is 1000；Learning error is 0.001.

2) training network model (method flow diagram is as shown in Figure 4), determines the connection weight between each layer.

Acoustoelectric combined detection method uses multi-information merging technology, is will be ultrasonic different with 2 kinds of electromagnetism from a certain discharge source Information is intelligently synthesized, and can maximize favourable factors and minimize unfavourable ones, resource complementation, compensate for the deficiency of single type detection method.Acoustoelectric combined Position detecting system adopts reaching time-difference tdoa (time different of arrival, tdoa) method, using bp (back Propagation) neutral net carries out ultrasound wave and hyperfrequency method sensor acquisition data are carried out fusion and come Judging fault position Put.

Tdoa algorithm puts, by calculating same office, the time difference that source signal reaches multiple sensors, carries out office and puts source location. Spread speed due to electromagnetic signal is much larger than sound wave, and the electromagnetic wave signal that shelf depreciation produces reaches the moment of uhf sensor Almost there is moment t0, and the ultrasonic signal propagation producing arrives the moment of each ultrasonic sensor, tn and t0 all exist not Same time difference.It is assumed that total n sensor, the space coordinatess that source is put in office are (x, y, z), the space coordinatess of sensor for (xi, Yi, zi), wherein i=0,1 ..., n-1；N is natural number.The coordinate of uhf sensor is (x0, y0, z0)；Represented with ri and put Air line distance between power supply and each sensor；τ_iRepresent that i-th ultrasound wave puts source signal with the ultrasonic sensor office of receiving Time difference；V represents acoustic wave propagation velocity, then have:

r_i=v τ_iI=1,2 ... n-1；

According to space length method for expressing, n-1 equation can be set up:

$r_i=\sqrt{{(x-x_i)}^2+{(y-y_i)}^2+{(z-z_i)}^2}$

Wherein i=1,2 ... n-1；

It is assumed that combined detection system contains 1 uhf sensor and 6 ultrasonic sensors, with uhf sensor it is Datum mark (0,0,0), then obtain according to space length equationThe individual elements of a fix, using suitable Data fusion technique, by this A little elements of a fix values are merged thus the definite coordinate in source is put in the office of being obtained: taking one of solving equations as a example:

Continuing abbreviation is:

$\left\{\begin{array}{c}{\left({\mathrm{v\τ}}_1\right)}^2-{\left({\mathrm{v\τ}}_2\right)}^2=2(x_2-x_1)x+2(y_2-y_1)y+2(z_2-z_1)z+x_1^2+y_1^2+z_1^2-{x_2}^2-{y_2}^2-{z_2}^2\\ {\left({\mathrm{v\τ}}_1\right)}^2-{\left({\mathrm{v\τ}}_3\right)}^2=2(x_3-x_1)x+2(y_3-y_1)y+2(z_3-z_1)z+x_1^2+y_1^2+z_1^2-{x_3}^2-{y_3}^2-{z_3}^2\\ {\left({\mathrm{v\τ}}_2\right)}^2-{\left({\mathrm{v\τ}}_3\right)}^2=2(x_3-x_2)x+2(y_3-y_2)y+2(z_3-z_2)z+{x_2}^2+{y_2}^2+{z_2}^2-{x_3}^2-{y_3}^2-{z_3}^2\end{array}\right.---\left(2\right);$

Coordinates of targets (x, y, z) can be obtained using elimination solution formula (2).Formula (2) both sides differential is obtained:

$\left\{\begin{array}{c}2v{\mathrm{\τ}}_1\·vd{\mathrm{\τ}}_1=2(x-x_1)(dx-d{x}_1)+2(y-y_1)(dy-d{y}_1)+2(z-z_1)(dz-d{z}_1)\\ 2v{\mathrm{\τ}}_2\·vd{\mathrm{\τ}}_2=2(x-x_2)(dx-d{x}_2)+2(y-y_2)(dy-d{y}_2)+2(z-z_2)(dz-d{z}_2)\\ 2v{\mathrm{\τ}}_3\·vd{\mathrm{\τ}}_3=2(x-x_3)(dx-d{x}_3)+2(y-y_3)(dy-d{y}_3)+2(z-z_3)(dz-d{z}_3)\end{array}\right.---\left(3\right);$

It can be seen that, the precision (dx, dy, dz) of the elements of a fix and sensor position error (dx_i,dy_i,dz_i) and reaching time-difference Measurement error d τ_iRelevant.

The functional relationship of bp (back propagation) network neural meta-model is: Wherein: xi is the input of node, n is interstitial content, and y represents the output of node, and f is activation primitive, and w, x are weight matrix and defeated Enter matrix, w=[w1, w2 ..., wn0], x=[x1, x2 ..., xn0], θ are the threshold values of neuron, ω_iFor connection weight.For contracting The Short Training time, neuron of the present invention adopts bipolarity s type function, and expression formula is:

$y=f\left(a\right)=-\frac{1}{2}+\frac{1}{1+e^{-a}};$

$\begin{array}{c}{y}^2={\[f\left(a\right)\]}^2={(-\frac{1}{2}+\frac{1}{1+e^{-a}})}^2\\ =\frac{1}{4}-\frac{1}{1+e^{-a}}+{\left(\frac{1}{1+e^{-a}}\right)}^2\end{array};$

$\begin{array}{c}{y}^{\′}=f^{\′}\left(a\right)=\frac{-e^{-a}}{{(1+e^{-a})}^2}\\ =\frac{1}{4}-\[\frac{1}{4}-\frac{1}{1+e^{-a}}+{\left(\frac{1}{1+e^{-a}}\right)}^2\]\\ =\frac{1}{4}-y^2\end{array};$

Wherein e is natural constant, and a is variable.

Step with bp neural fusion information fusion algorithm is:

1) initialize: setting network each layer weights, the initial value of threshold values are less random number battle array, and cycle-index is maximum.

2) positive input, output relation；

By sensor gathered data carry out after relevant treatment, as training sample；If input quantity is x=(x¹, x²,…,x^α), correspondence is output as y=(y¹,y²,…,y^p), the desired output of network is d=(d¹,d²,…,d^p).It is provided with α sample This is training data, ifExport for p-th sample, WithIt is respectively the output of the first and second hidden layers, ω_uvFor connection weight；Then each node is corresponding defeated Go out for:

$\left\{\begin{array}{c}{g}_j^p=f\left(\underset{i=0}{\overset{no}{\σ}}{\mathrm{\ω}}_{ji}{x}_i^p\right)\\ h_k^p=f\left(\underset{j=0}{\overset{n1}{\σ}}{\mathrm{\ω}}_{kj}{g}_i^p\right)\\ g_j^p=f\left(\underset{k=0}{\overset{n2}{\σ}}{\mathrm{\ω}}_{lk}{x}_k^p\right)\end{array}\right.---\left(4\right);$

3) error back propagation；

From output layer, hidden layer to input layer, with the reverse error signal of each layer being calculated based on gradient steepest descent method, And then obtain weighed value adjusting pattern, equivalent error δ computing formula is:

$\left\{\begin{array}{c}{\mathrm{\δ}}_j^p=f^{\′}\left({\mathrm{net}}_j^p\right)(t_j^p-y_j^p)\\ {\mathrm{\δ}}_k^p=f^{\′}\left({\mathrm{net}}_k^p\right)(t_k^p-y_k^p)\\ {\mathrm{\δ}}_l^p=f^{\′}\left({\mathrm{net}}_l^p\right)(t_l^p-y_l^p)\end{array}\right.---\left(5\right);$

4) each layer connection weight correction, computing formula is:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{lk}(n+1)={\mathrm{\ω}}_{lk}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\σ}}{\mathrm{\δ}}_l^p{h}_k^p\right)\\ {\mathrm{\ω}}_{kj}(n+1)={\mathrm{\ω}}_{kj}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\σ}}{\mathrm{\δ}}_k^p{g}_j^p\right)\\ {\mathrm{\ω}}_{ji}(n+1)={\mathrm{\ω}}_{ji}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\σ}}{\mathrm{\δ}}_j^p{x}_i^p\right)\end{array}\right.---\left(6\right);$

m_cFor factor of momentum, take m_c=0.9, according to additional guide vanes Rule of judgment, now kth step error sum of squares e (k) ＞ e (k-1)；The equivalent error of each layer neuron is α input sample of δ,For after input network first The output result of hidden layer, j/p/k all represents variable 1,2,3 ... α.Represent second after input network Hidden layer output result,Represent final output result after input network, input net It is the reality output of p-th sample after network；ω_uvFor the connection weight between each neuron；

5) judge whether root-mean-square error meets setting accuracy, if meeting, output result；Otherwise return to step 2) continue Training.

The present invention is the multiple sensor information amalgamation method based on ultrasound wave, hyperfrequency two types sensor, solves It is currently based on single type sensor online system failure fixation and recognition accuracy rate and the low problem of accuracy, can effectively keep away Exempt from the wrong report based on the presence of single type sensor on-line measuring device, fail to report and do not report phenomenon.To ultrasound wave and hyperfrequency method Using reaching time-difference tdoa (time different of arrival, tdoa) method, carry out data using bp neutral net Merge, that realizes partial discharges fault is accurately positioned the gis shelf depreciation live line measurement field it is adaptable to conventional.

Finally illustrate, only in order to technical scheme to be described and unrestricted, this area is common for above example Other modifications or equivalent that technical staff is made to technical scheme, without departing from technical solution of the present invention Spirit and scope, all should cover in the middle of scope of the presently claimed invention.

Claims (3)

1. a kind of acoustoelectric combined detection Fault Locating Method is put based on the gis office of bp network it is characterised in that including following walking Rapid:

Step 1, the information each collecting at the scene is converted to the signal of telecommunication by sensor；Described sensor includes supersonic sensing Device and uhf sensor；

Step 2, the signal of telecommunication after changing is carried out data prediction by described ultrasonic sensor, and pretreated data is entered Row merges；The signal of telecommunication after changing is carried out data prediction by described uhf sensor, and pretreated data is carried out Merge；

Step 3, the data after the data after ultrasonic sensor fusion and uhf sensor being merged using bp neutral net Merged, failure judgement position；

Step 4, exports abort situation.

2. according to claim 1 a kind of acoustoelectric combined detection Fault Locating Method is put based on the gis office of bp network, it is special Levy and be: the neuron function of described bp neutral net is using bipolarity s type function:Wherein e For natural constant, a is variable.

3. according to claim 1 a kind of acoustoelectric combined detection Fault Locating Method is put based on the gis office of bp network, it is special Levy and be: the data fusion step of described step 3 is as follows:

1) initialize: setting network each layer weights, the initial value of threshold values are random number battle array, and cycle-index is maximum；

2) positive input, output relation；

By sensor gathered data processed after, as training sample；If input quantity is x=(x¹,x²,…,x^α), Correspondence is output as y=(y¹,y²,…,y^p), the desired output of network is d=(d¹,d²,…,d^p)；Being provided with α sample is training number According to, ifExport for p-th sample,With It is respectively the output of the first and second hidden layers, ω_uvFor connection weight；Then each node is corresponding is output as:

$\left\{\begin{array}{c}{g}_j^p=f\left(\underset{i=0}{\overset{no}{\mathrm{\σ}}}{\mathrm{\ω}}_{ji}{x}_i^p\right)\\ h_k^p=f\left(\underset{j=0}{\overset{n1}{\mathrm{\σ}}}{\mathrm{\ω}}_{kj}{g}_i^p\right)\\ y_j^p=f\left(\underset{k=0}{\overset{n2}{\mathrm{\σ}}}{\mathrm{\ω}}_{lk}{x}_k^p\right)\end{array}\right.---\left(4\right)$

In formula: α input sample,For the output result of the first hidden layer after input network, j/p/ K all represents variable 1,2,3 ... α.Represent the second hidden layer output result after input network,Represent final output result after input network, be p-th sample after input network Reality output；ω_uvFor the connection weight between each neuron；

3) error back propagation；

From output layer, hidden layer to input layer, with the reverse error signal of each layer being calculated based on gradient steepest descent method, weighed Value adjustment modes, equivalent error δ computing formula is:

$\left\{\begin{array}{c}{\mathrm{\δ}}_j^p=f^{\′}\left({\mathrm{net}}_j^p\right)(t_j^p-y_j^p)\\ {\mathrm{\δ}}_k^p=f^{\′}\left({\mathrm{net}}_k^p\right)(t_k^p-y_k^p)\\ {\mathrm{\δ}}_l^p=f^{\′}\left({\mathrm{net}}_l^p\right)(t_l^p-y_l^p)\end{array}\right.---\left(5\right);$

The equivalent error of each layer neuron is δ, and t is the desired value that each layer neuron sets, and y is the reality output of each layer neuron Value；

4) each layer connection weight correction, computing formula is:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{lk}\left(n+1\right)={\mathrm{\ω}}_{lk}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\mathrm{\σ}}}{\mathrm{\δ}}_l^p{h}_k^p\right)\\ {\mathrm{\ω}}_{kj}\left(n+1\right)={\mathrm{\ω}}_{kj}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\mathrm{\σ}}}{\mathrm{\δ}}_k^p{g}_j^p\right)\\ {\mathrm{\ω}}_{ji}\left(n+1\right)={\mathrm{\ω}}_{ji}\left(n\right)+\frac{\mathrm{\η}}{1-m_c}\left(\underset{p=1}{\overset{\mathrm{\α}}{\mathrm{\σ}}}{\mathrm{\δ}}_j^p{x}_i^p\right)\end{array}\right.---\left(6\right);$

m_cFor factor of momentum, take m_c=0.9, according to additional guide vanes Rule of judgment, now kth step error sum of squares e (k) ＞ e (k-1)；The equivalent error of each layer neuron is α input sample of δ,Imply for after input network first The output result of layer, j/p/k all represents variable 1,2,3 ... α.After representing input network, second implies Layer output result,Represent final output result after input network, after input network Reality output for p-th sample；ω_uvFor the connection weight between each neuron；

5) according to new connection weight, carry out positive calculating；Judge each learning sample (x^p, t^p) and output layer is each Whether individual neuron meets setting accuracy, if meeting, output result；Otherwise return to step 2) continue training.