CN110006645B - Multi-source fusion high-voltage circuit breaker mechanical fault diagnosis method - Google Patents

Abstract

The invention discloses a multi-source fused high-voltage circuit breaker mechanical fault diagnosis method, and belongs to the field of multi-sensor information fusion diagnosis. The method comprises the steps that firstly, vibration signals of a plurality of positions of the high-voltage circuit breaker are simultaneously collected on the basis of a vibration measuring device, the wavelet energy entropy of the vibration signals is calculated, and vibration characteristic vectors describing the mechanical state of the high-voltage circuit breaker are formed; then, grouping vibration data of each sensor to form a model training set and an evaluation set, designing a Softmax regression diagnosis model according to the model training set, calculating the diagnosis accuracy of each sensor under the Softmax regression diagnosis model by using the evaluation set, and forming confidence weights of mechanical fault diagnosis of the plurality of sensors; and finally, providing an improved D-S evidence fusion method based on the diagnosis weight of each sensor to realize the mechanical fault identification of the high-voltage circuit breaker. The method effectively reduces the one-sided influence of single sensor diagnosis, and greatly improves the accuracy of diagnosis of mechanical defects of the high-voltage circuit breaker.

Description

Multi-source fusion high-voltage circuit breaker mechanical fault diagnosis method

Technical Field

The invention belongs to the field of multi-sensor information fusion diagnosis, and particularly relates to a multi-source fusion high-voltage circuit breaker mechanical fault diagnosis method.

Background

High voltage circuit breakers are one of the important devices in power systems, and assume critical control and protection roles. As the scale of power grids has increased, the usage of high voltage circuit breakers has continued to steadily increase. To cope with the complexity of the power system, the power grid has different requirements on the performance of the circuit breaker under different voltage levels, and the function of the circuit breaker cannot be replaced. In order to prevent the circuit breaker from faults, old or invalid components need to be replaced to maintain the running state of the circuit breaker, and the existing high-voltage circuit breaker is overhauled in a post-overhaul mode, a regular overhaul mode and a state overhaul mode. The influence of high-voltage circuit breaker faults on a power grid cannot be reduced after overhaul, and the damage is large; the regular maintenance has blindness, which causes overlarge maintenance cost and wastes manpower and material resources. Therefore, the operation state of the high-voltage circuit breaker can be known in time, and the fault type of the high-voltage circuit breaker can be identified and predicted, so that the operation reliability of the high-voltage circuit breaker can be improved.

At present, detection and state diagnosis of mechanical characteristics of a high-voltage circuit breaker are effectively reflected, and the detection and state diagnosis mainly depend on the moving contact stroke, the opening and closing operation mechanism current and a mechanical vibration signal of the high-voltage circuit breaker. The contact stroke and coil current detection technology has a lot of inconveniences in the existing application test process, and partial mechanical defects are difficult to identify. The method for extracting vibration signal features and classifying faults still has the defects of low diagnosis accuracy, poor universality and the like in the mechanical vibration detection of the high-voltage circuit breaker. The high-voltage circuit breaker mechanical fault diagnosis accuracy based on vibration information is low, and for example, the vibration signal dispersibility is large, the feature extraction is insufficient, the vibration information at a single position is difficult to independently obtain the comprehensive description of various fault scenes, and the like.

Therefore, the high-efficiency and high-precision circuit breaker mechanical fault diagnosis method based on multi-position vibration information fusion is developed, the identification accuracy can be improved, and the further development of the mechanical state evaluation of the high-voltage circuit breaker is promoted.

Disclosure of Invention

Aiming at the problems, the invention provides a multi-source fused high-voltage circuit breaker mechanical fault diagnosis method, and the accuracy of multi-sensor diagnosis fusion is improved.

The method comprises the following specific steps:

acquiring vibration signals of multiple sensor positions with different defects to form measurement samples, and performing wavelet packet time-frequency conversion on each measurement sample at each sensor position to obtain a time-frequency energy matrix of each measurement sample;

firstly, simultaneously collecting mechanical vibration signals of a high-voltage circuit breaker at n sensor positions with different mechanical defects, wherein the number of samples corresponding to each sensor is m; q measurement sample X for the i sensor_i,qPerforming wavelet packet time-frequency conversion to obtain time-frequency energy matrix Y_i,q,t×f(ii) a Matrix Y_i,q,t×fThe number of rows is t and the number of columns is f.

Step two, respectively equally dividing the time-frequency energy matrix of each measurement sample along a time axis and a frequency axis to form a plurality of respective time-frequency blocks;

time-frequency energy matrix Y_i,q,t×fN is equally divided in the directions of the time axis t and the frequency axis f₁And n₂Part (c) to form n₁×n₂A time-frequency block;

n₁representing the matrix Y in the time direction_i,q,t×fThe number of equal parts of; n is₂Representing the matrix Y in the frequency direction_i,q,t×fThe number of equal parts of; the number of elements in each sub-time period is t/n₁The number of elements of each sub-frequency segment is f/n₂。

Step three, calculating the energy of each time-frequency block, and constructing the vibration information characteristic description of each measurement sample;

first, n is calculated for the qth measurement sample of the ith sensor₁×n₂Energy E of (α) th one of the time-frequency blocks_i,q,α,βThe formula is as follows:

representing the (S) th of the (α) th time-frequency block₁，S₂) An element;

then, according to n₁×n₂Energy of each time-frequency block, calculating wavelet packet energy entropy of the qth measurement sample under the ith sensor, and forming vibration information characteristic description (WPE) of the measurement sample_i,q＝{WPE_i,q,α,f,WPE_i,q,t,β}；

i＝1,2,…,n；q＝1,2,…,m；α＝1,2,…,n₁；β＝1,2,…,n₂。

Dividing an evaluation set and a training subset according to the vibration information characteristic description of each measurement sample, outputting the fault diagnosis accuracy of each sensor by training a Softmax regression diagnosis model, and further calculating the confidence weight of each sensor;

firstly, m measurements are taken for the ith sensorCharacterization of a sample constitutes set A_i；

Set A_i＝{A_i,1,A_i,2,...,A_i,q...,A_i,m}；A_i,qA characterization of the qth measurement sample representing the ith sensor; a. the_i,q＝[x_i,q,1,x_i,q,2,...,x_i,q,w]^T(ii) a w represents the number of features, i.e. the feature space dimension.

Then, the feature description is set A_iAveragely dividing the training data into k subsets, wherein each subset is an evaluation set, and the difference set of each evaluation set and the sample set is a corresponding training subset;

k evaluation sets are { B_i,1,B_i,2...,B_i,k}; training subsets are as { C_i,1,C_i,2...,C_i,k}；C_i,k＝A_i\B_i,k；

Training a Softmax regression diagnosis model according to the k training subsets, inputting the k evaluation sets to output respective corresponding diagnosis accuracy rates, and calculating an average value as a fault diagnosis average accuracy rate P of the ith sensor_i；

Calculating the average accuracy of fault diagnosis of the n sensors in the same way to obtain the confidence weight of the ith sensor;

confidence weights for the ith sensor are:

confidence weight vector ω ═ ω [ ω ] for n sensors₁,ω₂,...,ω_n]。

Step five, aiming at a certain new sample A to be tested, calculating a vibration information feature description vector O of the sample A to be tested under the ith sensor_i；

Step six, describing the vibration information feature vector O_iInput diagnostic model M_iObtaining the probability column vector Q of the s-type faults possibly occurring in the sample A to be tested_i；Q_iIs a vector of s rows and 1 columns.

Diagnostic modelM_iA Softmax regression diagnostic model for the ith sensor; respectively naming the Softmax regression diagnosis models of the sensors as the corresponding diagnosis models;

respectively calculating s-type fault occurrence probability column vector set { Q) of the sample A to be tested under n sensors in the same way₁,Q₂,...Q_i,...,Q_n}；

Step seven, calculating an expected vector Q of the s-type fault occurrence probability by using confidence weights of the n sensors and the s-type fault occurrence probability column vector of each sensor_λ；

Step eight, aiming at the sample A to be tested, calculating s-type fault occurrence probability column vectors and expected vectors Q of all the sensors_λFurther obtaining s-type fault fusion probability column vector mass which is possibly generated by the sample A to be tested;

probability column vector Q_iAnd an expected vector Q_λHas a Euclidean distance d between_iThe calculation is as follows:

defining s-type fault fusion probability column vectors mass which are possibly generated by n sensors for the sample A to be tested by utilizing the Euclidean distance;

and step nine, fusing the column vector mass for n-1 times by utilizing the traditional D-S evidence reasoning, and defining the fault type of the maximum probability in the column as the fault type corresponding to the sample A to be tested to finish the final mechanical fault diagnosis of the high-voltage circuit breaker.

The invention has the following excellent effects:

1. a multisource fused high-voltage circuit breaker mechanical fault diagnosis method is different from a traditional high-voltage circuit breaker fault diagnosis method, the method references relevant knowledge of multisource information fusion, utilizes and improves a D-S evidence theory, forms diagnosis fusion of multi-sensor diagnosis information, avoids blindness of a single sensor to a fault identification result, and improves robustness of fault diagnosis.

2. A multi-source fusion high-voltage circuit breaker mechanical fault diagnosis method includes dividing a training data set of a single sensor into a plurality of subsets, combining the subsets to form a plurality of training subsets and corresponding evaluation sets (difference sets of the training data set and the training subsets), designing a Softmax regression model by utilizing the training subsets, and defining diagnosis accuracy of the sensor according to diagnosis results of the evaluation sets. The diagnosis accuracy of the sensors is normalized, the confidence of each sensor is obtained, and the accuracy of multi-sensor diagnosis fusion is improved.

Drawings

FIG. 1 is a flow chart of a multi-source converged high voltage circuit breaker mechanical fault diagnosis method of the present invention;

FIG. 2 is a schematic diagram of the multi-sensor confidence calculation of the present invention;

FIG. 3 is a diagram of a fault classification scheme for a circuit breaker based on a Softmax regression model according to the present invention;

FIG. 4 is a flow diagram of the multi-sensor information fusion diagnostics of the present invention that improves D-S evidence reasoning;

fig. 5 is an experimental diagram for acquiring vibration information of a high-voltage circuit breaker of a certain model.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

The invention is based on the fault diagnosis model of vibration information of multiple positions of a high-voltage circuit breaker, and forms a mechanical fault diagnosis method of the high-voltage circuit breaker with multi-source fusion, wherein firstly, multi-position vibration data and original wavelet energy entropy characteristic space construction are obtained simultaneously; then, constructing a training subset and an evaluation set and evaluating the confidence of the sensor fault diagnosis result; and finally, fusing the Softmax model of each sensor and the diagnosis confidence to determine a plurality of diagnostic results of Softmax.

The method specifically comprises the following steps: firstly, simultaneously acquiring vibration data samples of mechanical defects of the high-voltage circuit breaker at a plurality of measuring points based on a vibration measuring device, and constructing a traditional wavelet packet energy entropy to the vibration data samples to form a vibration information characteristic space; then, grouping the vibration data of each sensor in an average and non-intersection manner, wherein a vibration characteristic sample set under each measuring point (sensor) is an evaluation set and a corresponding training subset of a difference set with the original collected data; secondly, performing Softmax regression diagnosis modeling on each training subset, calculating the diagnosis accuracy of each sensor under the Softmax regression diagnosis model and the average accuracy of each measuring point by using the corresponding evaluation set, and defining the confidence weight of each sensor according to the diagnosis accuracy and the average accuracy of each measuring point; finally, calculating a diagnosis probability vector expectation according to the confidence and the diagnosis probability of each sensor, and defining various fault occurrence probability vectors of the test sample under the fusion of a plurality of sensors based on the diagnosis probability vectors of each sensor and the expected distance of the obtained probability vector; and on the basis of an improved D-S evidence fusion method, determining a plurality of diagnostic results of Softmax by fusing a Softmax model and diagnostic confidence of each sensor, fusing various fault occurrence probability vectors under the fusion of the plurality of sensors for a plurality of times, defining a fault type corresponding to the maximum probability as the fault occurring in the test sample, and finishing the final diagnosis of the mechanical defect of the high-voltage circuit breaker.

Experimental result analysis shows that the method can effectively reduce the one-sided influence of single sensor diagnosis and greatly improve the accuracy of high-voltage circuit breaker mechanical defect diagnosis.

As shown in fig. 1, the specific steps are as follows:

acquiring vibration signals of multiple sensor positions with different defects to form measurement samples, and performing wavelet packet time-frequency conversion on each measurement sample at each sensor position to obtain a time-frequency energy matrix of each measurement sample;

firstly, simultaneously acquiring n vibrations of the high-voltage circuit breaker in different mechanical defects by using vibration information measuring equipmentThe mechanical vibration signals of the dynamic acceleration sensor position, the number of samples corresponding to each sensor is m; q measurement sample X for the i sensor_i,qPerforming wavelet packet time-frequency conversion, wherein i is 1,2, … n, and i represents a sensor number; q is 1,2, …, m represents the number of measurement sample, and the time-frequency energy matrix Y is obtained_i,q,t×f(ii) a Matrix Y_i,q,t×fThe number of rows is t and the number of columns is f.

Step two, respectively equally dividing the time-frequency energy matrix of each measurement sample along a time axis and a frequency axis to form a plurality of respective time-frequency blocks;

time-frequency energy matrix Y_i,q,t×fN is equally divided in the directions of the time axis t and the frequency axis f₁And n₂Part (c) to form n₁×n₂A time-frequency block;

n₁representing a matrix Y of pairs in the time direction (row direction)_i,q,t×fThe number of equal parts of; n is₂Represents a matrix Y of pairs in the frequency direction (column direction)_i,q,t×fThe number of equal parts of; the number of elements in each sub-time period is t/n₁The number of elements of each sub-frequency segment is f/n₂。

Step three, calculating the energy of each time-frequency block, and constructing the vibration information characteristic description of each measurement sample;

based on the characteristics that the high-voltage circuit breaker is subjected to switching-on process and is represented as an impact free attenuation signal, namely the signal is in sudden change, and the frequency changes along with time, the wavelet packet energy entropy is calculated by utilizing wavelet packet time-frequency domain analysis, and the vibration information characteristic description is measured;

firstly, for the q measurement sample of the ith acceleration sensor of the high-voltage circuit breaker, n is calculated₁×n₂Energy E of (α) th one of the time-frequency blocks_i,q,α,βThe formula is as follows:

wherein E_i,q,α,βA time-frequency matrix Y representing the wavelet packet of the qth measurement sample under the ith acceleration sensor_i,q,t×fTo carry outn₁×n₂The energy of the (α) th time-frequency block after the block,

represents the (S) th time-frequency block in the (α) th time-frequency block in the q th measurement sample under the ith acceleration sensor₁，S₂) An element;

then, according to n₁×n₂Calculating the energy entropy of the wavelet packet of the qth measurement sample under the ith sensor according to the energy of the time-frequency block, and constructing the vibration information feature description WPE of the measurement sample as shown in the formulas (2) and (3)_i,q＝{WPE_i,q,α,f,WPE_i,q,t,β}；

i＝1,2,…,n；q＝1,2,…,m；α＝1,2,…,n₁；β＝1,2,…,n₂。

Dividing an evaluation set and a training subset according to the vibration information characteristic description of each measurement sample, outputting the fault diagnosis accuracy of each sensor by training a Softmax regression diagnosis model, and further calculating the confidence weight of each sensor;

as shown in FIG. 2, first, for the ith sensor, the characteristics of m measurement samples are grouped into a set A_i；

Characterizing the qth measurement sample under the ith sensor by WPE_i,q＝{WPE_i,q,α,f,WPE_i,q,t,βIs defined as A_i,q＝[x_i,q,1,x_i,q,2,...,x_i,q,w]^TWherein A is_i,qThe method comprises the steps of representing a characteristic set of wavelet packet energy entropy of a qth measurement sample under an ith sensor, wherein w represents the number of characteristics, namely a characteristic space dimension; vibration wavelet packet energy entropy feature sample set A without sensor confidence calculation_i＝{A_i,1,A_i,2,...,A_i,q...,A_i,m}；

Then, the feature description is set A_iAveragely dividing the data into k subsets, wherein the intersection among the subsets is a null set, the union set is a whole sample set, the subsets are defined as evaluation sets, and the difference set of each evaluation set and the sample set is a corresponding training subset;

k evaluation sets are { B_i,1,B_i,2...,B_i,kIs satisfied with

Training subsets are as { C_i,1,C_i,2...,C_i,k}；C_i,k＝A_i\B_i,k；

Constructing a Softmax regression diagnosis model from the k training subsets, as shown in FIG. 3, and optimizing the C as the training subset based on the gradient descent method_i,kSoftmax regression diagnostic model CM_i,k. Assuming a total of s fault types, the system equation in the Softmax regression is shown in equation (4).

Wherein p (y)_q＝s|x_i,qTheta) represents the probability of occurrence of the s-th type fault under the parameter theta of the qth measurement sample under the ith sensor, theta represents a matrix of s × w,

calculating gradients

Softmax regression diagnostic model CM_i,kCalculating the kth evaluation set B under the ith sensor_i,kAccuracy P of_i,kInputting k evaluation sets to output the diagnosis accuracy rates corresponding to the k evaluation sets, and calculating the average value after finishing the k evaluation sets as the fault diagnosis average accuracy rate P of the ith sensor_i；

Calculating the average accuracy rate of fault diagnosis of the n sensors in the same way, and defining the confidence weight of each sensor according to the average accuracy rate of each sensor;

confidence weights for the ith sensor are:

confidence weight vector ω ═ ω [ ω ] for n sensors₁,ω₂,...,ω_n]。

Step five, aiming at a certain new sample A to be tested, calculating a vibration information feature description vector O of the sample A to be tested under the ith sensor_i；

O_i＝[x_i,1,x_i,2,...,x_i,w]^T

Step six, describing the vibration information feature vector O_iInput diagnostic model M_iObtaining the probability column vector Q of the s-type faults possibly occurring in the sample A to be tested_i；

Q_iIs a vector of s rows and 1 columns.

Diagnostic model M_iA Softmax regression diagnostic model for the ith sensor; respectively naming the Softmax regression diagnosis models of the sensors as the corresponding diagnosis models;

respectively calculating s-type fault occurrence probability column vector set { Q) of the sample A to be tested under the Softmax diagnostic model of the n sensors in the same way₁,Q₂,...Q_i,...,Q_n}；

Step seven, calculating an expected vector Q of the s-type fault occurrence probability by using confidence weights of the n sensors and the s-type fault occurrence probability column vector of each sensor_λ；

Step eight, aiming at the sample A to be tested, calculating s-type fault occurrence probability column vectors and expected vectors Q of all the sensors_λFurther obtaining the s-type fault fusion probability column direction of the sample A to be tested which is possible to occurMeasuring mass;

probability column vector Q_iAnd an expected vector Q_λHas a Euclidean distance d between_iThe calculation is as follows:

defining s-type fault fusion probability column vectors mass which are possibly generated by n sensors for the sample A to be tested by utilizing the Euclidean distance;

forming initial probability assignments of n sensors to the test sample for possible s-type failures;

step nine, utilizing the traditional D-S evidence to reason the column vector mass_iAnd after fusing for n-1 times, defining the fault type where the maximum probability in the column is positioned as the fault type corresponding to the sample A to be tested, and finally completing the mechanical fault diagnosis of the high-voltage circuit breaker.

The conventional D-S evidence reasoning method is as follows:

defining mass₁And mass₂Is the basic probability assignment of the recognition frame F ═ { F ═ F₁,F₂,...,F_sIs a finite set of s pairwise mutually exclusive elements; f_sIndicating the occurrence of the s-th type fault and simultaneously defining the mass_iRepresenting the ith sensor pair against SoftMax model M_iThe basic probability assignment for s-type faults, then the multi-sensor mass_iThe fusion process is shown in equation (5).

Wherein the mass_c(F_s) Represents the probability of the s-th fault after fusion, and when i is 1, the mass_c(F_s)＝mass₁(F_s)；mass_i+1(F_sj) The probability value of the sj-th fault of the i +1 th sensor to the test sample is shown, which shows that n sensors only need to be fused for n-1 times; k is a radical of_cIndicating that the coefficient of collision is equal to

Defining a fusion probability column vector mass of S-type faults possibly occurring on a test sample by n sensors of the high-voltage circuit breaker as a basic probability assignment of an identification frame F, defining a fault type where the maximum probability is located as a fault type corresponding to the test sample after fusing the mass vector n-1 times by using a formula (5) through a D-S evidence reasoning method, and finishing the mechanical fault diagnosis of the high-voltage circuit breaker for finally improving the D-S evidence reasoning. As shown in fig. 4, a multi-sensor information fusion process for improving D-S evidence reasoning.

For example, as shown in fig. 5, 7 operation conditions, namely, switching-off spring fatigue, switching-on spring fatigue, oil leakage of an oil buffer, transmission shaft damping increase, transmission shaft pin wear and foundation bolt looseness, are set for an experiment platform by a high-voltage circuit breaker of a certain model, switching-on experiments are performed for 350 times in total, vibration information of 3 measurement positions is acquired simultaneously in each operation condition for 50 times, 70% of data of each operation condition is randomly selected, namely, 35 data are trained, and the rest data are used for testing experiment evidence. The training sample number m is 245, and n is equally divided along the time axis in the wavelet packet time-frequency matrix₁12, equally dividing n along the frequency axis₂13, the method is divided into 12 × 13 energy blocks, the original characteristic space dimension w formed by wavelet packet energy entropy is 12+13 and is equal to 25, the accuracy of diagnosis by utilizing a Softmax model at three test positions is 84.76%, 76.19% and 73.33% respectively based on the wavelet packet energy entropy characteristics of 15 test samples of each type of fault, and the diagnosis accuracy of the traditional D-S evidence reasoning and fusion method is 95.24%, after the method is used for reasoning and fusion, the accuracy of diagnosis of mechanical defects of the high-voltage circuit breaker is 97.14%, and the diagnosis result of test data shows that the method effectively reduces the single sensor pair based on vibrationThe fault diagnosis error probability of the high-voltage circuit breaker is obtained through information, an improved D-S multi-source fusion method of confidence weights of a plurality of sensors is established, and the accuracy of fault classification can be further improved.

Finally, it should be noted that: the described embodiments are only some embodiments of the present application and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

Claims (2)

1. A multi-source fused high-voltage circuit breaker mechanical fault diagnosis method is characterized by comprising the following specific steps:

acquiring vibration signals of multiple sensor positions with different defects to form measurement samples, and performing wavelet packet time-frequency conversion on each measurement sample at each sensor position to obtain a time-frequency energy matrix of each measurement sample;

firstly, simultaneously collecting mechanical vibration signals of a high-voltage circuit breaker at n sensor positions with different mechanical defects, wherein the number of samples corresponding to each sensor is m; q measurement sample X for the i sensor_i,qPerforming wavelet packet time-frequency conversion to obtain time-frequency energy matrix Y_i,q,t×f(ii) a Matrix Y_i,q,t×fThe number of rows is t and the number of columns is f;

step two, respectively equally dividing the time-frequency energy matrix of each measurement sample along a time axis and a frequency axis to form a plurality of respective time-frequency blocks;

step three, calculating the energy of each time-frequency block, and constructing the vibration information characteristic description of each measurement sample;

first, n is calculated for the qth measurement sample of the ith sensor₁×n₂Energy E of (α) th one of the time-frequency blocks_i,q,α,βThe formula is as follows:

representing the (S) th of the (α) th time-frequency block₁，S₂) An element;

then, according to n₁×n₂Energy of each time-frequency block, calculating wavelet packet energy entropy of the qth measurement sample under the ith sensor, and forming vibration information characteristic description (WPE) of the measurement sample_i,q＝{WPE_i,q,α,f,WPE_i,q,t,β}；

i＝1,2,…,n；q＝1,2,…,m；α＝1,2,…,n₁；β＝1,2,…,n₂；

Dividing an evaluation set and a training subset according to the vibration information characteristic description of each measurement sample, outputting the fault diagnosis accuracy of each sensor by training a Softmax regression diagnosis model, and further calculating the confidence weight of each sensor;

constructing a Softmax regression diagnosis model according to the k training subsets, and optimizing the model to be C based on the training subsets by using a gradient descent method_i,kSoftmax regression diagnostic model CM_i,k(ii) a Assuming a total of s fault types, the system equation in the Softmax regression is shown in equation (4),

wherein p (y)_q＝s|x_i,qTheta) represents the probability of occurrence of the s-th type fault under the parameter theta of the qth measurement sample under the ith sensor, theta represents a matrix of s × w,

calculating gradients

Softmax regression diagnostic model CM_i,kCalculating the kth evaluation set B under the ith sensor_i,kAccuracy P of_i,kInputting k evaluation sets to output the diagnosis accuracy rates corresponding to the k evaluation sets, and calculating the average value after finishing the k evaluation sets as the fault diagnosis average accuracy rate P of the ith sensor_i；

Step five, aiming at a certain new sample A to be tested, calculating a vibration information feature description vector O of the sample A to be tested under the ith sensor_i；

Step six, describing the vibration information feature vector O_iInput diagnostic model M_iObtaining the probability column vector Q of the s-type faults possibly occurring in the sample A to be tested_i；Q_iA vector of s rows and 1 columns;

diagnostic model M_iA Softmax regression diagnostic model for the ith sensor; respectively naming the Softmax regression diagnosis models of the sensors as the corresponding diagnosis models;

respectively calculating s-type fault occurrence probability column vector set { Q) of the sample A to be tested under n sensors in the same way₁,Q₂,...Q_i,...,Q_n}；

Step seven, calculating an expected vector Q of the s-type fault occurrence probability by using confidence weights of the n sensors and the s-type fault occurrence probability column vector of each sensor_λ；

Step eight, aiming at the sample A to be tested, calculating s-type fault occurrence probability column vectors and expected vectors Q of all the sensors_λFurther obtaining s-type fault fusion probability column vector mass which is possibly generated by the sample A to be tested;

probability column vector Q_iAnd an expected vector Q_λHas a Euclidean distance d between_iComputingThe following were used:

defining s-type fault fusion probability column vectors mass which are possibly generated by n sensors for the sample A to be tested by utilizing the Euclidean distance;

and step nine, fusing the column vector mass for n-1 times by utilizing the traditional D-S evidence reasoning, and defining the fault type of the maximum probability in the column as the fault type corresponding to the sample A to be tested to finish the final mechanical fault diagnosis of the high-voltage circuit breaker.

2. The multi-source fused high-voltage circuit breaker mechanical fault diagnosis method of claim 1, wherein the fourth step is specifically:

firstly, for the ith sensor, the characteristics of m measurement samples are combined into a set A_i；

Set A_i＝{A_i,1,A_i,2,...,A_i,q...,A_i,m}；A_i,qA characterization of the qth measurement sample representing the ith sensor; a. the_i,q＝[x_i,q,1,x_i,q,2,...,x_i,q,w]^T(ii) a w represents the number of features, i.e. the feature space dimension;

then, the feature description is set A_iAveragely dividing the training data into k subsets, wherein each subset is an evaluation set, and the difference set of each evaluation set and the sample set is a corresponding training subset;

k evaluation sets are { B_i,1,B_i,2...,B_i,k}; training subsets are as { C_i,1,C_i,2...,C_i,k}；C_i,k＝A_i\B_i,k；

Training a Softmax regression diagnosis model according to the k training subsets, inputting the k evaluation sets to output respective corresponding diagnosis accuracy rates, and calculating an average value as a fault diagnosis average accuracy rate P of the ith sensor_i；

Calculating the average accuracy of fault diagnosis of the n sensors in the same way to obtain the confidence weight of the ith sensor;

confidence weights for the ith sensor are:

confidence weight vector ω ═ ω [ ω ] for n sensors₁,ω₂,...,ω_n]。

Images