CN102087311B - Method for improving measurement accuracy of power mutual inductor - Google Patents

Abstract

The invention discloses a method for improving the measurement accuracy of a power mutual inductor. The method comprises the following steps of: establishing a system model based on a fuzzy neural network; preprocessing a signal acquired by a power mutual inductor; and sending the preprocessed signal into a trained and learned adaptive fuzzy neural network system for calibrating to obtain a calibrated signal of the power mutual inductor. In the invention, a power mutual inductor measurement signal with a higher accuracy can be obtained by adopting an ordinary power mutual inductor, thereby the accuracy of the measurement system including the power mutual inductor is improved, and the cost of the system is reduced.

Description

A kind of method that improves measurement accuracy of power mutual inductor

Technical field

The present invention relates to the electric power mutual-inductor measuring technique of electronic industrial technology, particularly a kind of method that improves measurement accuracy of power mutual inductor.

Background technology

Electric current on the line of electric force and leakage current are of paramount importance parameters in the systems such as electric power monitoring, load prediction and fire hazard monitoring, in existing some electric fire monitoring system, directly compare to judge whether to carry out electric fire alarm with current value or leakage current value and predefined threshold value.Usually electric current and leakage current value collect by electric power mutual-inductor on the line of electric force, and the measured value of electric power mutual-inductor is subjected to the impact of the factors such as itself and the position relationship of line of electric force, on-the-spot electromagnetic environment.In addition, the material that common electric power mutual-inductor uses and manufacturing process etc. also can make its electric current or leakage current measurement value and actual value directly have deviation, and precision is not high, and then cause the performance index of these systems to reduce.The electric current of common electric power mutual-inductor or leakage current measurement value and the direct deviation of actual value are nonlinear, be subjected to simultaneously the impact of installation site, environment temperature and electromagnetic environment etc., can't be described with explicit mathematical formulae, and in application, the site environment parameter is constantly to change, and can not set up the complete data collection for all possible ambient conditions and parameter area.Must consider with nonlinear signal processing technology the signal value of current transformer collection to be carried out the data adjustment, make it near actual value, and then process and application lays the foundation for follow-up data.

Summary of the invention

In order to solve the problems of the technologies described above, the invention provides a kind of method of energy Effective Raise measurement accuracy of power mutual inductor.

The technical scheme that the present invention solves the problems of the technologies described above may further comprise the steps:

1) gathers training sample, common mutual inductor and precision standard mutual inductor are installed in the same tested line of electric force measurement environment, guarantee the position consistency of two mutual inductors and tested line of electric force, and produce temperature, power frequency magnetic field site environment parameter by the environmental baseline generator, the signal that electric power mutual-inductor gathers through A/D conversion and pre-service after as the training sample of system model;

2) set up fuzzy neural network model, the network of fuzzy neural network model determines that by input layer, fuzzy membership function layer, relevance grade computation layer, normalization computation layer and output layer form;

3) set up the fuzzy neural network model structure, utilize the Gradient Descent learning algorithm of self-adaptation momentum decoupling zero to adjust the fuzzy membership function parameter, utilize least square method to adjust the output layer link weight coefficients;

4) signal of electric power mutual-inductor collection carried out pre-service;

5) pretreated signal is sent in the good Adaptive Fuzzy Neural-network system of training study and carried out calibration process, obtain calibrating rear electric power mutual-inductor measuring-signal.

Further, the pretreated method of described step 4) collection signal is any of the in limited time method of average, normalization smoothing method, MMSE method or RLS method.

Technique effect of the present invention is: the present invention utilizes common mutual inductor and the measuring-signal of precision standard mutual inductor under same test environment as the training sample of neural network, set up the Adaptive Fuzzy Neural-network system with this, utilize the measuring-signal of Adaptive Fuzzy Neural-network system calibration electric power mutual-inductor, greatly improve the measuring accuracy of common electric power mutual-inductor, in guaranteed performance, effectively reduced system cost.

Below in conjunction with the drawings and specific embodiments invention is described in further detail.

Description of drawings

Fig. 1 is the electric power mutual-inductor calibrating patterns synoptic diagram among the present invention.

Fig. 2 is the structure of fuzzy neural network figure among the present invention.

Embodiment

As shown in Figure 1, be the common electric power mutual-inductor image data calibrating patterns that the present invention sets up, this model comprises following three partial contents:

1, sets up based on fuzzy neural network model.

Foundation is based on fuzzy neural network model, comprises that mainly data acquisition, model structure, model parameter determine.

Common mutual inductor and precision standard mutual inductor are installed in the same tested line of electric force measurement environment, guarantee simultaneously the position consistency of two mutual inductors and tested line of electric force, produced simultaneously the site environment parameters such as temperature, power frequency magnetic field by the environmental baseline generator, through after A/D conversion and the pre-service as the training sample of fuzzy neural network.

Fuzzy neural network model comprises following five layers, and its structure is referring to Fig. 2: 1. input layer d, each node directly links to each other with pretreated signals of process such as the environment parameter that gathers, common electric power mutual-inductor collection signals, sends into lower one deck.2. fuzzy membership function is determined layer e, each node represents a linguistic variable value, be used for calculating and respectively input the membership function that component belongs to each linguistic variable value fuzzy set, used membership function can be Gauss's membership function, double-flanged end Gauss membership function, bell membership function etc.3. relevance grade computation layer f, each node represents a fuzzy rule, is used for mating the fuzzy rule former piece, calculates the relevance grade of every rule.4. normalization computation layer g realizes that data normalization calculates.5. output layer h realizes that fuzzy sharpening calculates, i.e. solving result.

The model hybrid learning algorithm: learning process is mainly used in adjusting the weighting parameter of connecting line between each node layer of fuzzy neural network.Parameter adjustment is comprised of two parts: 1. fuzzy membership function is determined center and the width of membership function in the layer in the network, and they and output are nonlinear relationships; 2. the link weight coefficients of output layer, they and output are linear relationships.According to parameter and output relation, adopt the learning algorithm of similar BP neural network, adopt respectively the Gradient Descent learning algorithm of self-adaptation momentum decoupling zero to adjust the fuzzy membership function parameter, adjust the output layer link weight coefficients with least square method.The input sample of Training Fuzzy Neural Networks is the signal parameter that input layer connects in the system model, and output sample is the signal parameter that the precision standard mutual inductor gathers.

2, the signal with the electric power mutual-inductor collection carries out pre-service.

The impact that processed by electric power mutual-inductor manufacturing process, neighbourhood noise (thunder and lightning, environment temperature etc.), circuit thermonoise (electron device intrinsic noise, electromagnetic interference (EMI) etc.) and follow-up A/D etc., the twinkling signal saltus step may appear, be that measured signal also can exist vibration behind mutual inductor continuous acquisition signal digitalized when constant and changes, for improving the stability of signal, need to before carrying out follow-up data, carry out pre-service, make signal keep stable, basic skills comprises the in limited time method of average, normalization smoothing method, MMSE method, RLS method etc.

3, pretreated signal is sent in the good Adaptive Fuzzy Neural-network system of training study and carried out calibration process.

Take the collection in worksite signal as input, good Adaptive Fuzzy Neural-network is processed to send into training study, and output is the mutual inductor collection signal after the calibration, can be directly used in the electric power monitoring or the electric fire monitoring system that give threshold value.

The below provides better embodiment of the present invention, and is described in detail, and enables to understand better function of the present invention, characteristics.

The example that is calibrated to the common electric power mutual-inductor that is applied to electric fire monitoring system illustrates sensor calibrating method of the present invention.

Signalization gathers environment, the data that obtain comprise the environment parameter that signal, temperature sensor, the power frequency magnetic field sensor of the collection of common electrical power mutual inductor collect etc., the input layer that is directly connected to fuzzy neural network as shown in Figure 2 after these data process pre-service is as the input sample, and pre-service adopts the mean value of simply asking a continuous acquisition M sample to replace the data of single collection.

The input value corresponding to input layer d of fuzzy neural network is designated as X=[x ₁, x ₂, x ₃] ^T

Fuzzy membership function determines that layer e uses Gaussian function u _Ij(x _i)=exp (((x _i-c _Ij)/σ _Ij) ²) as membership function, c wherein _IjAnd σ _IjThe center and the width that represent respectively membership function.

Relevance grade computation layer f adopts to connect and takes advantage of solving method to calculate relevance grade, i.e. u _j=u _1i(x ₁) u _2i(x ₂) u _3i(x ₃).

Normalization computation layer g realizes normalization calculating, namely

Output layer h calculates Output rusults, namely W wherein _jLink weight coefficients value for output layer.

When learning, adopt fuzzy neural network model gradient descent method to adjust c _IjAnd σ _Ij, error calculation formula is

$E=\frac{1}{2N}\underset{t=1}{\overset{N_p}{\mathrm{\Σ}}}{(y_d^t-y^t)}^2$

Wherein E is the Square-type error function, N _pBe learning sample number, y ^tBe the collection signal data of common electric power mutual-inductor,

Be the collection signal data of precision standard mutual inductor, the parameter adjustment formula is

$c_{\mathrm{ij}}(t+1)=c_{\mathrm{ij}}\left(t\right)-\mathrm{\η}(\∂E/{\∂c}_{\mathrm{ij}})\·{\mathrm{\Δc}}_{\mathrm{ij}}\left(t\right),$ Δc _ij(t)=c _ij(t+1)-c _ij(t)

${\mathrm{\σ}}_{\mathrm{ij}}(t+1)={\mathrm{\σ}}_{\mathrm{ij}}\left(t\right)-\mathrm{\η}(\∂E/{\∂\mathrm{\σ}}_{\mathrm{ij}})\·{\mathrm{\Δ\σ}}_{\mathrm{ij}}\left(t\right),$ Δσ _ij(t)=σ _ij(t+1)-σ _ij(t)

Wherein t is the study iterations, and η is Learning Step.

Adopt least square method adjustment output link weight coefficients, formula is

$W_{t+1}=W_t+S_{t+1}{a}_{t+1}(y_{t+1}^T-a_{t+1}^T{W}_t)$

$S_{t+1}=S_t+\frac{S_t{a}_{t+1}{a}_{t+1}^T{S}_t}{1+a_{t+1}^T{S}_t{a}_{t+1}},t=\mathrm{0,1,2},...,N_p-1$

W wherein _tBe t row vector of link weight coefficients to be adjusted, Be t row vector of relevance grade normalized vector, S _tBe covariance matrix,

K row vector for training output data vector Y.

After model establishes, the signal of common electric power mutual-inductor b collection and the site environment parameter of collection are processed the collection signal data after output is very calibrated through sending into fuzzy neural network model after the pre-service.

Above-described; it only is preferred embodiment of the present invention; be not to limit scope of the present invention, i.e. every simple, equivalence of doing according to claims and the description of the present patent application changes and modifies, and all falls into the claim protection domain of patent of the present invention.

Claims (2)

1. a method that improves measurement accuracy of power mutual inductor is characterized in that, may further comprise the steps:

1) gathers training sample, common mutual inductor and precision standard mutual inductor are installed in the same tested line of electric force measurement environment, guarantee the position consistency of two mutual inductors and tested line of electric force, and produce temperature, power frequency magnetic field site environment parameter by the environmental baseline generator, the signal that electric power mutual-inductor gathers through A/D conversion and pre-service after as the training sample of system model;

2) set up fuzzy neural network model, the network of fuzzy neural network model determines that by input layer, fuzzy membership function layer, relevance grade computation layer, normalization computation layer and output layer form;

3) set up the fuzzy neural network model structure, utilize the Gradient Descent learning algorithm of self-adaptation momentum decoupling zero to adjust the fuzzy membership function parameter, utilize least square method to adjust the output layer link weight coefficients;

4) signal of electric power mutual-inductor collection carried out pre-service;

5) pretreated signal is sent in the good Adaptive Fuzzy Neural-network system of training study and carried out calibration process, obtain calibrating rear electric power mutual-inductor measuring-signal.

2. the method for raising measurement accuracy of power mutual inductor according to claim 1 is characterized in that, the pretreated method of described step 4) collection signal is any of the in limited time method of average, normalization smoothing method, MMSE method or RLS method.

Images