On-load tap-changer fault diagnosis method and system based on multi-source signal fusion

Abstract

The invention discloses an on-load tap-changer fault diagnosis method and system based on multi-source signal fusion. The diagnosis method comprises the following steps: acquiring a detection signal of an on-load tap-changer; the detection signal comprises a voiceprint signal, a vibration signal, a voltage signal and a current signal; fusing the voiceprint signal, the vibration signal, the voltage signal and the current signal by using a symmetric point mode SDP algorithm to generate an SDP image; performing scale invariant feature transformation feature extraction on the SDP image to obtain a feature subset; matching the feature subset with a pre-constructed fault diagnosis dictionary to obtain a diagnosis result; wherein a mapping relation between the feature subset and the fault state is set in the fault diagnosis dictionary. The method has the advantages of simple operation, small calculation amount, high fault diagnosis precision and the like.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of power equipment fault diagnosis, and in particular to an on-load tap changer fault diagnosis method and system based on multi-source signal fusion. Background Art

[0002] On-load Tap Changer (OLTC) is an important device in transformer voltage regulation technology, which can accurately adjust the output voltage during transformer operation to ensure the stable operation and voltage quality of the power system. As the proportion of new energy increases year by year, the large-scale access of new energy and power electronic devices to new power systems with high efficiency and reliability brings about increased harmonics, reduced power quality, random power fluctuations, and reduced system stability. The power system faces huge challenges.

[0003] Based on this background, the number of OLTC operations has increased many times compared to before, and the increase in the number of operations will inevitably lead to an increase in the probability of failure. As the only mechanical moving part in the power transformer, OLTC must not only withstand the contact wear caused by mechanical impact during the gear shifting process, but also withstand the influence of arc erosion, so its operating conditions are very harsh. According to statistics, OLTC failures account for more than 35% of power transformer failures, and often cause short circuits in the transformer voltage regulating windings after failure, leading to serious consequences. In addition, OLTC failures can be divided into mechanical failures and electrical failures, and correspondingly there are typical failures such as loose contacts, stuck mechanisms, spring fatigue, contact erosion / wear, vacuum tube failure, and failure of a branch of the transition circuit. From the failure cases of on-load tap changers, their failures mostly occur during or after the switching process, usually manifested as abnormal switching timing due to incorrect mechanical component action, which further causes failures in the electrical circuit. Therefore, it is very necessary to perform fault diagnosis on on-load tap changers.

[0004] At present, the fault diagnosis of OLTC relies on a single signal. The processing and analysis of a single signal often cannot accurately determine all fault characteristics and the fault diagnosis results are unreliable. For vibration signals, the analysis methods include empirical mode decomposition (EMD), wavelet analysis, short-time Fourier transform, amplitude domain analysis, etc. For voiceprint signals, the voiceprint features include MFCC, Fbank, LPCC, etc. The corresponding processing methods also have their shortcomings. When EMD decomposes the acoustic vibration signal, modal aliasing will occur. Modal aliasing will not only cause distortion of the time-frequency distribution, but also make the physical meaning of the IMF component unclear; the single feature of the voiceprint signal is trained with large storage space and high training complexity. Summary of the invention

[0005] In view of the technical problems existing in the prior art, the present invention provides an on-load tap changer fault diagnosis method and system based on multi-source signal fusion, which is simple to implement, has a small amount of calculation and accurate diagnosis.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0007] A method for diagnosing faults of an on-load tap changer based on multi-source signal fusion comprises the following steps:

[0008] Acquire a detection signal of the on-load tap changer; the detection signal includes a sound print signal, a vibration signal, a voltage signal and a current signal;

[0009] The symmetrical point mode SDP algorithm is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal to generate an SDP image.

[0010] Perform scale-invariant feature transformation on the SDP image to extract features and obtain feature subsets;

[0011] The feature subset is matched with a pre-built fault diagnosis dictionary to obtain a diagnosis result; wherein the fault diagnosis dictionary is provided with a mapping relationship between the feature subset and the fault state.

[0012] Preferably, the process of constructing the fault diagnosis dictionary is as follows:

[0013] Acquire the soundprint signals and vibration signals at multiple positions when the on-load tap-changer is operating in various states, as well as the voltage and current signals during the electromagnetic transient process of the on-load tap-changer;

[0014] The SDP algorithm is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal to generate a polar coordinate image set I = [I1,…,I2,…,IT], and the image set is divided into a selection set and a test set;

[0015] In state t, select R groups of SDP images generated by test data from the image subset It in the selection set, randomly select r groups of SDP images for scale-invariant feature transformation feature extraction, and form the feature subset St in state t. Finally, the total feature subset S corresponding to all states is obtained = [S1, S2, ..., S T ];

[0016] In state t, key point matching is performed between the feature subset and the feature subsets of other states, and the total number of key point matches M of each feature in the feature subset and the randomly selected features of other states is calculated. The data is sorted in descending order according to the total number of matches M, and the first Q groups of data are selected as fault diagnosis reference templates, and finally the initial fault diagnosis dictionary D corresponding to all states is obtained;

[0017] Match the SDP images in the test set with the fault diagnosis reference templates in the fault diagnosis initial dictionary D, count the number of key feature points of the SDP images to be diagnosed in the test set and all the fault diagnosis reference templates in the fault diagnosis initial dictionary D, sort them by size, count the top K groups of dictionary elements with the largest number of matches, and select the fault state with the highest frequency in the K groups as the fault diagnosis result;

[0018] The accuracy of each state is calculated through the test set to determine the rotation rate s of the rotation sampling; u units are randomly selected from the unselected units to replace the last u templates of the corresponding state in the dictionary D until the preset accuracy threshold S or the maximum number of rotations N is reached to obtain the final fault diagnosis dictionary.

[0019] Preferably, the formula for generating the polar coordinate image set by the SDP algorithm is:

[0020]

[0021]

[0022] Where r(i) is the polar coordinate radius; θ(i) is the counterclockwise deflection angle; is the clockwise deflection angle; x i is the i-th sample point of the signal; x max and x min are the maximum and minimum values ​​of the signal respectively; a is the time interval parameter; ε is the angle magnification factor; θ is the rotation angle of the mirror symmetry plane, a=1,2,3……N.

[0023] Preferably, the process of scale-invariant feature transformation feature extraction is:

[0024] Process the SDP image to locate the feature points and their positions of the SDP image;

[0025] Generate multiple feature descriptors based on the feature points and their positions of the SDP image;

[0026] Matching is performed based on the Euclidean distance between the feature descriptor of the SDP image and the feature descriptor of the template image.

[0027] Preferably, the specific process of processing the SDP image and locating the characteristic points and positions of the SDP image is as follows:

[0028] After the SDP image is converted into a grayscale image I(x,y), the two-dimensional Gaussian function G(x,y,σ) is used to obtain the Gaussian Laplace transform image and perform differential processing on it to obtain the Gaussian difference operator D(x,y,σ), which is constructed into a pyramid structure;

[0029] Each pixel point of each layer of the Gaussian difference pyramid image is compared with the surrounding pixels of the same layer and adjacent layer images to find the image extreme points, obtain the feature points and their positions of the SDP image, and use the Hessian matrix H(x,y) to remove the extreme values ​​in the edge area.

[0030] Preferably, the specific process of generating multiple feature descriptors according to the feature points and positions of the SDP image is as follows:

[0031] The histogram is used to count the gradient and amplitude of pixels in the m×m neighborhood centered on each feature point, and the direction with the highest amplitude in the histogram is taken as the main direction;

[0032] The m×m neighborhood is divided into multiple sub-regions, and the gradient directions in multiple directions are counted in each sub-region, eventually forming a SIFT feature descriptor with multiple data.

[0033] Preferably, the specific process of matching according to the Euclidean distance between the feature descriptor in the SDP image and the feature descriptor in the template image is as follows:

[0034] Establish key descriptor sets for the template image and the image to be detected respectively, complete the matching by comparing the key point descriptors in the two point sets, and use Euclidean distance to describe the similarity of the key point descriptors;

[0035] When the distance between the matched key descriptors meets the preset conditions, it is considered that the feature point i in the image to be detected matches the feature point p in the template image successfully;

[0036] The kd tree method is used to complete the feature sub-matching of the image to be diagnosed and the template image, and the feature sub-information of m key points that meet the threshold is obtained. The larger the number of matching points m, the more similar the two images are.

[0037] Preferably, the specific process of rotation sampling is:

[0038] The dictionary library is composed of N sampling units. Randomly sample n samples from the total library to form an initial dictionary. Perform test set diagnosis on the initial dictionary to obtain the initial dictionary diagnosis accuracy α0;

[0039] The initial dictionary is gradually rotated and replaced to determine the optimal rotation amount u. During the rotation process, u dictionary samples with poor representation in the previous sample are discarded, and u samples are selected from the sampling units that were not selected in the past to form a new dictionary.

[0040] Perform test set diagnosis on the new dictionary obtained by rotation sampling to obtain its diagnostic accuracy α k ; If the diagnostic accuracy rate α k If it is greater than the set threshold T or the maximum number of rotations reaches K, the iteration ends and the new dictionary is output.

[0041] Preferably, the specific process of determining the optimal rotation amount u is:

[0042] The sample size is fixed at n, and the sample is rotated. m sampling units are randomly selected from the sample, and their current observation value is recorded as y 11 ,y 12 ,…,y 1m , the sample mean is recorded as:

[0043]

[0044] Randomly select u units from the unselected units, u = nm, and record their current observation value as y 21 ,y 22 ,…,y 2m , the sample mean is recorded as:

[0045]

[0046] According to the dictionary diagnosis accuracy α k Construct homogeneous weighted estimator, 0≤α k ≤1, 0≤β k ≤1 and α k +β k =1, construct the mean weighted estimator:

[0047]

[0048] Weight α k , β k Determined by the accuracy of the previous sample; if α k Take a smaller value, β k If the value is large, it means that the sample judgment rate is low and the representativeness is poor; if α k Take a larger value, β k Taking a smaller value, the sample judgment rate is high, indicating that the sample is representative;

[0049] From the independence of the two parts of sampling, we know that The variance of is described by the following formula, where S 2 Current population variance:

[0050]

[0051] because is a constant. Under the constraint condition of m+u=n, Minimizing the variance is equivalent to finding the values ​​of m and u so that Reach the minimum value;

[0052] Construct the Lagrangian function, where λ is the Lagrangian coefficient, and find the partial derivatives of m and u respectively;

[0053]

[0054] Obtain Substitute the constraint condition m+u=n, and the optimal rotation amount is u=nβ.

[0055] The present invention also discloses an on-load tap changer fault diagnosis system based on multi-source signal fusion, comprising:

[0056] A signal acquisition module, used to acquire a detection signal of the on-load tap changer; the detection signal includes a voiceprint signal, a vibration signal, a voltage signal and a current signal;

[0057] An image generation module is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal using the symmetrical point mode SDP algorithm to generate an SDP image;

[0058] A feature extraction module is used to extract the scale-invariant feature transformation of the SDP image and obtain a feature subset;

[0059] The fault diagnosis module is used to match the feature subset with a pre-built fault diagnosis dictionary to obtain a diagnosis result; wherein the fault diagnosis dictionary is provided with a mapping relationship between the feature subset and the fault state.

[0060] Preferably, the process of constructing the fault diagnosis dictionary is as follows:

[0061] Acquire the soundprint signals and vibration signals at multiple positions when the on-load tap-changer is operating in various states, as well as the voltage and current signals during the electromagnetic transient process of the on-load tap-changer;

[0062] The SDP algorithm is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal to generate a polar coordinate image set I = [I1,…,I2,…,IT], and the image set is divided into a selection set and a test set;

[0063] In state t, select R groups of SDP images generated by test data from the image subset It in the selection set, randomly select r groups of SDP images for scale-invariant feature transformation feature extraction, and form the feature subset St in state t. Finally, the total feature subset S corresponding to all states is obtained = [S1, S2, ..., S T ];

[0064] In state t, key point matching is performed between the feature subset and the feature subsets of other states, and the total number of key point matches M of each feature in the feature subset and the randomly selected features of other states is calculated. The data is sorted in descending order according to the total number of matches M, and the first Q groups of data are selected as fault diagnosis reference templates, and finally the initial fault diagnosis dictionary D corresponding to all states is obtained;

[0065] Match the SDP images in the test set with the fault diagnosis reference templates in the fault diagnosis initial dictionary D, count the number of key feature points of the SDP images to be diagnosed in the test set and all the fault diagnosis reference templates in the fault diagnosis initial dictionary D, sort them by size, count the top K groups of dictionary elements with the largest number of matches, and select the fault state with the highest frequency in the K groups as the fault diagnosis result;

[0066] The accuracy of each state is calculated through the test set to determine the rotation rate s of the rotation sampling; u units are randomly selected from the unselected units to replace the last u templates of the corresponding state in the dictionary D until the preset accuracy threshold S or the maximum number of rotations N is reached to obtain the final fault diagnosis dictionary.

[0067] Preferably, the formula for generating the polar coordinate image set by the SDP algorithm is:

[0068]

[0069]

[0070] Where r(i) is the polar coordinate radius; θ(i) is the counterclockwise deflection angle; is the clockwise deflection angle; x i is the i-th sample point of the signal; x max and x min are the maximum and minimum values ​​of the signal respectively; a is the time interval parameter; ε is the angle magnification factor; θ is the rotation angle of the mirror symmetry plane, a=1,2,3……N.

[0071] Preferably, the process of scale-invariant feature transformation feature extraction is:

[0072] Process the SDP image to locate the feature points and their positions of the SDP image;

[0073] Generate multiple feature descriptors based on the feature points and their positions of the SDP image;

[0074] Matching is performed based on the Euclidean distance between the feature descriptor of the SDP image and the feature descriptor of the template image.

[0075] Preferably, the specific process of processing the SDP image and locating the characteristic points and positions of the SDP image is as follows:

[0076] After the SDP image is converted into a grayscale image I(x,y), the two-dimensional Gaussian function G(x,y,σ) is used to obtain the Gaussian Laplace transform image and perform differential processing on it to obtain the Gaussian difference operator D(x,y,σ), which is constructed into a pyramid structure;

[0077] Each pixel point of each layer of the Gaussian difference pyramid image is compared with the surrounding pixels of the same layer and adjacent layer images to find the image extreme points, obtain the feature points and their positions of the SDP image, and use the Hessian matrix H(x,y) to remove the extreme values ​​in the edge area.

[0078] Preferably, the specific process of generating multiple feature descriptors according to the feature points and positions of the SDP image is as follows:

[0079] The histogram is used to count the gradient and amplitude of pixels in the m×m neighborhood centered on each feature point, and the direction with the highest amplitude in the histogram is taken as the main direction;

[0080] The m×m neighborhood is divided into multiple sub-regions, and the gradient directions in multiple directions are counted in each sub-region, eventually forming a SIFT feature descriptor with multiple data.

[0081] Preferably, the specific process of matching according to the Euclidean distance between the feature descriptor in the SDP image and the feature descriptor in the template image is as follows:

[0082] Establish key descriptor sets for the template image and the image to be detected respectively, complete the matching by comparing the key point descriptors in the two point sets, and use Euclidean distance to describe the similarity of the key point descriptors;

[0083] When the distance between the matched key descriptors meets the preset conditions, it is considered that the feature point i in the image to be detected matches the feature point p in the template image successfully;

[0084] The kd tree method is used to complete the feature sub-matching of the image to be diagnosed and the template image, and the feature sub-information of m key points that meet the threshold is obtained. The larger the number of matching points m, the more similar the two images are.

[0085] Preferably, the specific process of rotation sampling is:

[0086] The dictionary library is composed of N sampling units. Randomly sample n samples from the total library to form an initial dictionary. Perform test set diagnosis on the initial dictionary to obtain the initial dictionary diagnosis accuracy α0;

[0087] The initial dictionary is gradually rotated and replaced to determine the optimal rotation amount u. During the rotation process, u dictionary samples with poor representation in the previous sample are discarded, and u samples are selected from the sampling units that were not selected in the past to form a new dictionary.

[0088] Perform test set diagnosis on the new dictionary obtained by rotation sampling to obtain its diagnostic accuracy α k ; If the diagnostic accuracy rate α k If it is greater than the set threshold T or the maximum number of rotations reaches K, the iteration ends and the new dictionary is output.

[0089] Preferably, the specific process of determining the optimal rotation amount u is:

[0090] The sample size is fixed at n, and the sample is rotated. m sampling units are randomly selected from the sample, and their current observation value is recorded as y 11 ,y 12 ,…,y 1m , the sample mean is recorded as:

[0091]

[0092] Randomly select u units from the unselected units, u = nm, and record their current observation value as y 21 ,y 22 ,…,y 2m , the sample mean is recorded as:

[0093]

[0094] According to the dictionary diagnosis accuracy α k Construct homogeneous weighted estimator, 0≤α k ≤1, 0≤β k ≤1 and α k +β k =1, construct the mean weighted estimator:

[0095]

[0096] Weight α k , β k Determined by the accuracy of the previous sample; if α k Take a smaller value, β k If the value is large, it means that the sample judgment rate is low and the representativeness is poor; if α k Take a larger value, β k Taking a smaller value, the sample judgment rate is high, indicating that the sample is representative;

[0097] From the independence of the two parts of sampling, we know that The variance of is described by the following formula, where S 2 Current population variance:

[0098]

[0099] because is a constant. Under the constraint condition of m+u=n, Minimizing the variance is equivalent to finding the values ​​of m and u so that Reach the minimum value;

[0100] Construct the Lagrangian function, where λ is the Lagrangian coefficient, and find the partial derivatives of m and u respectively;

[0101]

[0102] Obtain Substitute the constraint condition m+u=n, and the optimal rotation amount is u=nβ.

[0103] The present invention also discloses a computer program product, comprising a computer program, wherein the computer program executes the steps of the method described above when executed by a processor.

[0104] The present invention further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described above are executed.

[0105] The present invention also discloses an on-load tap changer fault diagnosis system based on multi-source signal fusion, comprising a memory and a processor connected to each other, wherein a computer program is stored on the memory, and when the computer program is run by the processor, the steps of the method described above are executed.

[0106] Compared with the prior art, the advantages of the present invention are:

[0107] The present invention can express complex fault states by fusing data from different sensors. The symmetrical point pattern algorithm SDP can clearly map complex time series in the form of scattered points in a polar coordinate diagram, which can improve the visualization ability of the original time domain signal in a graphical way. The visualization difference of the graph can more accurately reflect the change of the signal. The implementation is simple, the calculation amount is small, and it is robust to noise. Due to the particularity of polar coordinate images, multi-source signals can be fused in a limited area through the SDP algorithm, which is of great significance to the fault diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Figure 1 The figure is a flow chart of a fault diagnosis method in an embodiment of the present invention.

[0109] Figure 2 This is a schematic diagram of the SDP algorithm of the present invention.

[0110] Figure 3 This is a schematic diagram of the fusion generation of SDP images in the present invention.

[0111] Figure 4 It is the SIFT feature matching graph in the present invention.

[0112] Figure 5 It is a structural schematic diagram of the on-load tap changer in the present invention. DETAILED DESCRIPTION

[0113] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0114] like Figure 1 As shown, the on-load tap changer fault diagnosis method based on multi-source signal fusion of symmetrical dot pattern (Symmetric Dot Pattern, SDP for short) provided in an embodiment of the present invention comprises the following steps:

[0115] Acquire a detection signal of the on-load tap changer; the detection signal includes a sound print signal, a vibration signal, a voltage signal, and a current signal; the above detection signal may be a detection signal of the on-load tap changer in actual operation;

[0116] The symmetrical point mode SDP algorithm is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal to generate an SDP image.

[0117] Perform scale-invariant feature transformation on the SDP image to extract features and obtain feature subsets;

[0118] The feature subset is matched with a pre-built fault diagnosis dictionary to obtain a diagnosis result; wherein the fault diagnosis dictionary is provided with a mapping relationship between the feature subset and the fault state.

[0119] In a specific embodiment, the specific process of pre-building the fault diagnosis dictionary is as follows:

[0120] Step 1: Use soundprint and vibration sensors to collect soundprint and vibration signals at multiple positions when the on-load tap changer is operating in T states (normal and multiple faults);

[0121] Use current and voltage sensors to collect voltage and current signals during the electromagnetic transient process of the on-load tap changer; then pre-process each signal;

[0122] Specifically, inside the on-load tap changer, the soundprint sensor uses an oil-resistant sound sensor, the number is one, and it is installed inside the top cover; the vibration sensor uses an optical fiber vibration sensor, the number is three, and it is outside the on-load tap changer; the vibration sensor uses an acceleration sensor, the number is three, one is installed outside the top cover, and the two are installed on both sides of the energy storage mechanism; the soundprint sensor is connected with a hood, the number is one, and it is installed 0.3m away from the surface of the on-load tap changer; the number of voltage and current sensors are two each, which respectively collect the current of the transition branch of the on-load tap changer and the voltage at both ends of the transition resistor, such as Figure 5 The figure shows the switch schematic diagram and the corresponding voltage and current collection points. The voltage and current signals can identify the timing of the switch action. The signals at multiple positions can ensure comprehensive information collection of the on-load tap changer operating status.

[0123] Specifically, various faults include loose contacts, stuck mechanisms, spring fatigue, contact erosion / wear, vacuum tube failure, or failure of a branch of the transition circuit.

[0124] The preprocessing includes performing wavelet transform to reduce noise on the signal to obtain the noise-reduced signal, and then aligning the noise-reduced voltage, current, voiceprint, and vibration signals, cutting them to the same length, and averaging them.

[0125] Step 2: The pre-processed voiceprints and vibration signals in T states and the voltage and current signals in the transient process are fused using the SDP algorithm to generate a polar coordinate image set I = [I1,…,I2,…,I T ] and divide the image set into a selection set and a test set;

[0126] Figure 2 The figure shows the principle diagram of the SDP algorithm. The mapping relationship of the polar coordinate image generated by the SDP algorithm is:

[0127]

[0128] Where ε is the angle magnification factor, is the clockwise deflection angle, θ(i) is the counterclockwise deflection angle, r(i) is the polar coordinate radius; x i is the i-th sample point of the signal, x max and x min are the maximum and minimum values ​​of the signal, respectively, a is the time interval parameter, ε is the angle magnification factor, θ is the mirror symmetry plane rotation angle, a=1,2,3……N. Specifically, in the SDP fusion algorithm, a is set to 1 and ε is set to 30. Figure 3 The SDP image is a fusion of multiple information. Figure 4 Shown is the SIFT feature matching result.

[0129] Step 3: At state t, the SDP image subset I t The SDP images generated by R groups of test data are included in the image, where the SDP images of r groups of data are randomly selected and scale-invariant feature transform (SIFT) features are extracted to obtain the feature subset S t =[S t 1 ,S t 2 ,…,S t r ].

[0130] After the same processing is performed on the SDP image subsets of other states, the total SIFT feature subset S = [S1, S2, ..., S T ].

[0131] Specifically, the main steps of SIFT feature extraction are:

[0132] 3.1: Feature point detection and positioning

[0133] After the detection image is converted into a grayscale image I(x,y), the two-dimensional Gaussian function G(x,y,σ) is used to obtain the Gaussian Laplace transform image and perform differential processing on it to obtain a more efficient Gaussian difference operator D(x,y,σ), which is constructed into a pyramid structure;

[0134]

[0135] D(x,y,σ)=(G(x,y,kσ)-G(x,y,σ))*I(x,y)

[0136] Where (x, y) represents the position of the image pixel; σ is the scale space factor; k-1 is a constant in all scales, and the error approaches 0 when k approaches 1;

[0137] Compare each pixel of each layer of the Gaussian difference pyramid image with the surrounding pixels of the same layer and adjacent layer images to find the image extreme points, and use the Hessian matrix H(x,y) to remove the extreme points in the edge area to obtain the feature points and their positions;

[0138]

[0139] Where D xx It means that the image of a certain scale in the pyramid is differentiated twice in the x direction.

[0140] 3.2: Generate feature descriptors

[0141] The histogram is used to count the gradient and amplitude of pixels in a 16×16 neighborhood centered on each feature point, and the direction with the highest amplitude in the histogram is taken as the main direction; the extreme value and direction formula of the feature point can be expressed as:

[0142]

[0143]

[0144] Where: m(x,y) is the modulus of the gradient at (x,y); θ(x,y) is the angular direction of the gradient at (x,y); I(x,y) represents the grayscale value of the pixel.

[0145] The 16×16 area is divided into four 4×4 sub-areas, and the gradient directions in eight directions are counted in each area, finally forming a SIFT feature descriptor with 4×4×8=128 data.

[0146] 3.3: Feature point matching

[0147] Establish key descriptor sets for the template image and the image to be detected respectively, and complete the matching by comparing the key point descriptors in the two point sets. Generally, the Euclidean distance is used to describe the similarity of the key point descriptors, which is as follows:

[0148]

[0149] Where d(R i ,S i ) is the Euclidean distance similarity measure between any two descriptors; the key point descriptor R in the template image i =(r i1 ,r i2 ,…,r i128 ), the key point descriptor S in the image to be detected i =(s i1 ,s i2 ,…,s i128 );

[0150] When the distance between the matched key descriptors meets the following conditions, it is considered that the feature point i in the image to be detected matches the feature point p in the template image successfully;

[0151]

[0152] where d1 st min(R i ,S p ) is the distance R in the template image i The nearest point S p Distance; d2 nd min(R i ,S j ) is the distance R in the template image i The second closest point S j The distance; Threshold is the threshold;

[0153] The kd tree method is used to complete the feature sub-matching of the image to be diagnosed and the template image, and the feature sub-information of m key points that meet the threshold is obtained. The larger the number of matching points m, the more similar the two images are.

[0154] Step 4: Make the SIFT feature subset in state t match the key points with the feature subsets obtained in other states, and calculate the total number of key point matches M between each feature sub-set selected in this state and r groups of feature sub-sets randomly selected in other states, and sort the data selected in this state in descending order, and take the first Q groups of data as the reference template for fault diagnosis in this state;

[0155] Perform corresponding processing on other states to obtain the initial dictionary D for fault diagnosis.

[0156] Step 5: Perform SIFT matching on the data in the test set and the templates in the initial dictionary D for fault diagnosis. Count the number of SIFT key feature points between the SDP graph to be diagnosed in the test set and all the templates in the dictionary. Arrange the dictionary order in ascending order by size. Count the top K groups of dictionary elements with the largest number of matches. Select the fault state with the highest frequency in the K groups as the state of the signal to be diagnosed according to the statistical law.

[0157] Step 6: Obtain the accuracy in each state through the test set, determine the rotation rate s of the rotation sampling based on its accuracy, and randomly select u units from the units that have not been selected in the past to replace the last u templates in the dictionary in this state until the preset accuracy threshold S is reached or the maximum number of rotations N is reached, and the final fault diagnosis dictionary is obtained.

[0158] In step 6, the specific process of rotation sampling is:

[0159] 6.1. The dictionary library is composed of N sampling units. Randomly sample n samples from the total library to form an initial dictionary. Perform test set diagnosis on the dictionary to obtain the initial dictionary diagnosis accuracy α0;

[0160] 6.2. Progressively rotate and replace the initial dictionary, determine the optimal rotation amount u, discard u dictionary samples with poor representation in the previous sample during the rotation process, and select u samples from the sampling units that were not selected in the past to form a new dictionary.

[0161] The derivation process of the optimal rotation amount u is as follows:

[0162] The sample size is fixed at n, and the sample is rotated. m sampling units are randomly selected from the sample, and their current observation value is recorded as y 11 ,y 12 ,…,y 1m , sample mean Denoted as:

[0163]

[0164] Randomly select u units from the unselected units, u = nm, and record their current observation value as y 21 ,y 22 ,…,y 2m , sample mean Denoted as:

[0165]

[0166] According to the dictionary diagnosis accuracy α k Construct homogeneous weighted estimator, 0≤α k ≤1, 0≤β k ≤1 and α k +βk =1, construct mean-weighted estimator

[0167]

[0168] Weight α k , β k Determined by the accuracy of the previous sample, if α k Take a smaller value, β k If the value is large, it means that the sample judgment rate is low and the representativeness is poor; if α k Take a larger value, β k A smaller value means a higher sample judgment rate, which indicates that the sample is representative.

[0169] From the independence of the two parts of sampling, we know that The variance of can be described by the following formula, where S 2 Current population variance:

[0170]

[0171] because is a constant. Under the constraint condition of m+u=n, Minimizing the variance is equivalent to finding the values ​​of m and u so that Reach minimum value.

[0172] Construct the Lagrangian function, where λ is the Lagrangian coefficient, and find the partial derivatives of m and u respectively:

[0173]

[0174] Obtain Substitute the constraint condition m+u=n, and the optimal rotation amount is u=nβ.

[0175] 6.3. Perform test set diagnosis on the new dictionary obtained by rotation sampling and obtain its diagnostic accuracy α k ; If the diagnostic accuracy is greater than the set threshold T or the maximum number of rotations reaches K, the iteration ends and the new dictionary is output.

[0176] The present invention can express complex fault states by fusing data from different sensors. The symmetrical point pattern algorithm SDP can clearly map complex time series in the form of scattered points in a polar coordinate diagram, which can improve the visualization ability of the original time domain signal in a graphical way. The visualization difference of the graph can more accurately reflect the change of the signal. The implementation is simple, the calculation amount is small, and it is robust to noise. Due to the particularity of polar coordinate images, multi-source signals can be fused in a limited area through the SDP algorithm, which is of great significance to the fault diagnosis results.

[0177] The on-load tap changer fault diagnosis method based on multi-source signal fusion of symmetrical point pattern (SDP) of the present invention, wherein the multi-source signal contains rich and comprehensive on-load tap changer operating status information, avoids the shortcomings of the previous single signal fault diagnosis method, fuses the multi-source information through the SDP algorithm, and generates a polar coordinate image with high visualization, which can more accurately reflect the change of the signal, and reduces the low accuracy and incompleteness of single signal fault diagnosis.

[0178] The SDP algorithm of the present invention fuses multi-source information, making it more convenient to process multi-channel, multi-sensor and multi-source information, better characterizing data, and having a better fault diagnosis effect on on-load tap changers than a single signal diagnosis effect.

[0179] The features extracted by the SIFT algorithm of the present invention have the characteristics of scale invariance and rotation invariance, which can solve the problems of rotation, scaling, translation and noise masking of the object to be detected, increase the recognition of the object to be detected, and be more robust.

[0180] The fault diagnosis based on dictionary template matching of the present invention establishes the correlation between the on-load tap changer status and the analysis diagram in a manner similar to a "fingerprint", realizes the conversion from data to vision, automatically selects representative dictionary templates, and reduces the dependence on manual experience and professional knowledge.

[0181] The embodiment of the present invention further provides an on-load tap changer fault diagnosis system based on multi-source signal fusion of a symmetrical point mode (SDP), comprising:

[0182] A signal acquisition module is used to acquire detection signals of the on-load tap changer; the detection signals include voiceprint signals, vibration signals, voltage signals and current signals;

[0183] An image generation module is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal using the SDP algorithm to generate an SDP image;

[0184] A feature extraction module is used to extract the scale-invariant feature transformation of the SDP image and obtain a feature subset;

[0185] The fault diagnosis module is used to match the feature subset with a pre-built fault diagnosis dictionary to obtain a diagnosis result; wherein the fault diagnosis dictionary is provided with a mapping relationship between the feature subset and the fault state.

[0186] In a specific embodiment, the specific process of pre-building the fault diagnosis dictionary in the fault diagnosis module is as follows:

[0187] Step 1: Use soundprint and vibration sensors to collect soundprint and vibration signals at multiple positions when the on-load tap changer is operating in T states (normal and multiple faults);

[0188] Use current and voltage sensors to collect voltage and current signals during the electromagnetic transient process of the on-load tap changer; then pre-process each signal;

[0189] Specifically, inside the on-load tap changer, the soundprint sensor uses an oil-resistant sound sensor, the number is one, and it is installed inside the top cover; the vibration sensor uses an optical fiber vibration sensor, the number is three, and it is outside the on-load tap changer; the vibration sensor uses an acceleration sensor, the number is three, one is installed outside the top cover, and the two are installed on both sides of the energy storage mechanism; the soundprint sensor is connected with a hood, the number is one, and it is installed 0.3m away from the surface of the on-load tap changer; the number of voltage and current sensors are two each, which respectively collect the current of the transition branch of the on-load tap changer and the voltage at both ends of the transition resistor, such as Figure 5 The figure shows the switch schematic diagram and the corresponding voltage and current collection points. The voltage and current signals can identify the timing of the switch action. The signals at multiple positions can ensure comprehensive information collection of the on-load tap changer operating status.

[0190] Specifically, various faults include loose contacts, stuck mechanisms, spring fatigue, contact erosion / wear, vacuum tube failure, or failure of a branch of the transition circuit.

[0191] The preprocessing includes performing wavelet transform to reduce noise on the signal to obtain the noise-reduced signal, and then aligning the noise-reduced voltage, current, voiceprint, and vibration signals, cutting them to the same length, and averaging them.

[0192] Step 2: The pre-processed voiceprints and vibration signals in T states and the voltage and current signals in the transient process are fused using the SDP algorithm to generate a polar coordinate image set I = [I1,…,I2,…,I T ] and divide the image set into a selection set and a test set;

[0193] Figure 2 The figure shows the principle diagram of the SDP algorithm. The mapping relationship of the polar coordinate image generated by the SDP algorithm is:

[0194]

[0195] Where ε is the angle magnification factor, is the clockwise deflection angle, θ(i) is the counterclockwise deflection angle, r(i) is the polar coordinate radius; x i is the i-th sample point of the signal, x max and x min are the maximum and minimum values ​​of the signal, respectively, a is the time interval parameter, ε is the angle magnification factor, θ is the mirror symmetry plane rotation angle, a=1,2,3……N. Specifically, in the SDP fusion algorithm, a is set to 1 and ε is set to 30. Figure 3 The SDP image is a fusion of multiple information. Figure 4 Shown is the SIFT feature matching result.

[0196] Step 3: At state t, the SDP image subset I t The SDP images generated by R groups of test data are included in the image, where the SDP images of r groups of data are randomly selected and scale-invariant feature transform (SIFT) features are extracted to obtain the feature subset S t =[S t 1 ,S t 2 ,…,S t r ].

[0197] After the same processing is performed on the SDP image subsets of other states, the total SIFT feature subset S = [S1, S2, ..., S T ].

[0198] Specifically, the main steps of SIFT feature extraction are:

[0199] 3.1: Feature point detection and positioning

[0200] After the detection image is converted into a grayscale image I(x,y), the two-dimensional Gaussian function G(x,y,σ) is used to obtain the Gaussian Laplace transform image and perform differential processing on it to obtain a more efficient Gaussian difference operator D(x,y,σ), which is constructed into a pyramid structure;

[0201]

[0202] D(x,y,σ)=(G(x,y,kσ)-G(x,y,σ))*I(x,y)

[0203] Where (x, y) represents the position of the image pixel; σ is the scale space factor; k-1 is a constant in all scales, and the error approaches 0 when k approaches 1;

[0204] Compare each pixel of each layer of the Gaussian difference pyramid image with the surrounding pixels of the same layer and adjacent layer images to find the image extreme points, and use the Hessian matrix H(x,y) to remove the extreme points in the edge area to obtain the feature points and their positions;

[0205]

[0206] Where D xx It means that the image of a certain scale in the pyramid is differentiated twice in the x direction.

[0207] 3.2: Generate feature descriptors

[0208] The histogram is used to count the gradient and amplitude of pixels in a 16×16 neighborhood centered on each feature point, and the direction with the highest amplitude in the histogram is taken as the main direction; the extreme value and direction formula of the feature point can be expressed as:

[0209]

[0210]

[0211] Where: m(x,y) is the modulus of the gradient at (x,y); θ(x,y) is the angular direction of the gradient at (x,y); I(x,y) represents the grayscale value of the pixel.

[0212] The 16×16 area is divided into four 4×4 sub-areas, and the gradient directions in eight directions are counted in each area, finally forming a SIFT feature descriptor with 4×4×8=128 data.

[0213] 3.3: Feature point matching

[0214] Establish key descriptor sets for the template image and the image to be detected respectively, and complete the matching by comparing the key point descriptors in the two point sets. Generally, the Euclidean distance is used to describe the similarity of the key point descriptors, which is as follows:

[0215]

[0216] Where d(R i ,S i ) is the Euclidean distance similarity measure between any two descriptors; the key point descriptor R in the template image i =(r i1 ,r i2 ,…,r i128 ), the key point descriptor S in the image to be detected i =(s i1 ,s i2 ,…,s i128 );

[0217] When the distance between the matched key descriptors meets the following conditions, it is considered that the feature point i in the image to be detected matches the feature point p in the template image successfully;

[0218]

[0219] where d1 st min(R i ,S p ) is the distance R in the template image i The nearest point S p Distance; d2 nd min(Ri ,S j ) is the distance R in the template image i The second closest point S j The distance; Threshold is the threshold;

[0220] The kd tree method is used to complete the feature sub-matching of the image to be diagnosed and the template image, and the feature sub-information of m key points that meet the threshold is obtained. The larger the number of matching points m, the more similar the two images are.

[0221] Step 4: Make the SIFT feature subset in state t match the key points with the feature subsets obtained in other states, and calculate the total number of key point matches M between each feature sub-set selected in this state and r groups of feature sub-sets randomly selected in other states, and sort the data selected in this state in descending order, and take the first Q groups of data as the reference template for fault diagnosis in this state;

[0222] Perform corresponding processing on other states to obtain the initial dictionary D for fault diagnosis.

[0223] Step 5: Perform SIFT matching on the data in the test set and the templates in the initial dictionary D for fault diagnosis. Count the number of SIFT key feature points between the SDP graph to be diagnosed in the test set and all the templates in the dictionary. Arrange the dictionary order in ascending order by size. Count the top K groups of dictionary elements with the largest number of matches. Select the fault state with the highest frequency in the K groups as the state of the signal to be diagnosed according to the statistical law.

[0224] Step 6: Obtain the accuracy in each state through the test set, determine the rotation rate s of the rotation sampling based on its accuracy, and randomly select u units from the units that have not been selected in the past to replace the last u templates in the dictionary in this state until the preset accuracy threshold S is reached or the maximum number of rotations N is reached, and the final fault diagnosis dictionary is obtained.

[0225] In step 6, the specific process of rotation sampling is:

[0226] 6.1. The dictionary library is composed of N sampling units. Randomly sample n samples from the total library to form an initial dictionary. Perform test set diagnosis on the dictionary to obtain the initial dictionary diagnosis accuracy α0;

[0227] 6.2. Progressively rotate and replace the initial dictionary, determine the optimal rotation amount u, discard u dictionary samples with poor representation in the previous sample during the rotation process, and select u samples from the sampling units that were not selected in the past to form a new dictionary.

[0228] The derivation process of the optimal rotation amount u is as follows:

[0229] The sample size is fixed at n, and the sample is rotated. m sampling units are randomly selected from the sample, and their current observation value is recorded as y 11 ,y 12 ,…,y 1m , sample mean Denoted as:

[0230]

[0231] Randomly select u units from the unselected units, u = nm, and record their current observation value as y 21 ,y 22 ,…,y 2m , sample mean Denoted as:

[0232]

[0233] According to the dictionary diagnosis accuracy α k Construct homogeneous weighted estimator, 0≤α k ≤1, 0≤β k ≤1 and α k +β k =1, construct mean-weighted estimator

[0234]

[0235] Weight α k , β k Determined by the accuracy of the previous sample, if α k Take a smaller value, β k If the value is large, it means that the sample judgment rate is low and the representativeness is poor; if α k Take a larger value, β k A smaller value means a higher sample judgment rate, which indicates that the sample is representative.

[0236] From the independence of the two parts of sampling, we know that The variance of can be described by the following formula, where S 2 Current population variance:

[0237]

[0238] because is a constant. Under the constraint condition of m+u=n, Minimizing the variance is equivalent to finding the values ​​of m and u so that Reach minimum value.

[0239] Construct the Lagrangian function, where λ is the Lagrangian coefficient, and find the partial derivatives of m and u respectively:

[0240]

[0241] Obtain Substitute the constraint condition m+u=n, and the optimal rotation amount is u=nβ.

[0242] 6.3. Perform test set diagnosis on the new dictionary obtained by rotation sampling and obtain its diagnostic accuracy α k ; If the diagnostic accuracy is greater than the set threshold T or the maximum number of rotations reaches K, the iteration ends and the new dictionary is output.

[0243] The present invention also discloses a computer program product, including a computer program, which executes the steps of the above method when executed by a processor. The present invention further discloses a computer-readable storage medium, which stores a computer program, which executes the steps of the above method when executed by a processor. The present invention also discloses an on-load tap changer fault diagnosis system based on multi-source signal fusion, including an interconnected memory and a processor, wherein a computer program is stored on the memory, and wherein the computer program executes the steps of the above method when executed by the processor.

[0244] The product, medium and diagnostic system of the present invention correspond to the above-mentioned diagnostic method, and also have the advantages described in the above-mentioned diagnostic method. The present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiment when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable storage medium includes: any entity or device that can carry computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. The memory is used to store computer programs and / or modules, and the processor realizes various functions by running or executing computer programs and / or modules stored in the memory, and calling data stored in the memory. The memory may include high-speed random access memory and may also include non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0245] The above are only preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A method for fault diagnosis of on-load tap changer based on multi-source signal fusion, characterized in that: Includes steps: Acquire a detection signal of the on-load tap changer; the detection signal includes a sound print signal, a vibration signal, a voltage signal and a current signal; The symmetrical point mode SDP algorithm is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal to generate an SDP image. Perform scale-invariant feature transformation on the SDP image to extract features and obtain feature subsets; Match the feature subset with the pre-built fault diagnosis dictionary to obtain the diagnosis result; The fault diagnosis dictionary contains a mapping relationship between feature subsets and fault states.

2. The on-load tap changer fault diagnosis method based on multi-source signal fusion according to claim 1, characterized in that: The process of constructing a fault diagnosis dictionary is as follows: Acquire the soundprint signals and vibration signals at multiple positions when the on-load tap-changer is operating in various states, as well as the voltage and current signals during the electromagnetic transient process of the on-load tap-changer; The SDP algorithm is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal to generate a polar coordinate image set I = [I1,…,I2,…,IT], and the image set is divided into a selection set and a test set; In state t, select R groups of SDP images generated by test data from the image subset It in the selection set, randomly select r groups of SDP images for scale-invariant feature transformation feature extraction, and form the feature subset St in state t. Finally, the total feature subset S corresponding to all states is obtained = [S1, S2, ..., S T ]; In state t, key point matching is performed between the feature subset and the feature subsets of other states, and the total number of key point matches M of each feature in the feature subset and the randomly selected features of other states is calculated. The data is sorted in descending order according to the total number of matches M, and the first Q groups of data are selected as fault diagnosis reference templates, and finally the initial fault diagnosis dictionary D corresponding to all states is obtained; Match the SDP images in the test set with the fault diagnosis reference templates in the fault diagnosis initial dictionary D, count the number of key feature points of the SDP images to be diagnosed in the test set and all the fault diagnosis reference templates in the fault diagnosis initial dictionary D, sort them by size, count the top K groups of dictionary elements with the largest number of matches, and select the fault state with the highest frequency in the K groups as the fault diagnosis result; Calculate the accuracy of each state through the test set to determine the rotation rate s of the rotation sampling; Randomly select u units from the unselected units and replace the last u templates of the corresponding state in the dictionary D until the preset accuracy threshold S or the maximum number of rotations N is reached to obtain the final fault diagnosis dictionary.

3. The on-load tap changer fault diagnosis method based on multi-source signal fusion according to claim 2 is characterized in that: The formula for generating polar coordinate image sets by the SDP algorithm is: Among them, r(i) is the polar coordinate radius; θ(i) is the counterclockwise deflection angle; is the clockwise deflection angle; x i is the i-th sample point of the signal; x max and x min are the maximum and minimum values ​​of the signal respectively; a is the time interval parameter; ε is the angle magnification factor; θ is the rotation angle of the mirror symmetry plane, a=1,2,3……N.

4. The on-load tap changer fault diagnosis method based on multi-source signal fusion according to claim 2 is characterized in that: The process of scale-invariant feature transformation feature extraction is: Process the SDP image to locate the feature points and their positions of the SDP image; Generate multiple feature descriptors based on the feature points and their positions of the SDP image; Matching is performed based on the Euclidean distance between the feature descriptor of the SDP image and the feature descriptor of the template image.

5. The on-load tap changer fault diagnosis method based on multi-source signal fusion according to claim 4 is characterized in that: The specific process of processing the SDP image and locating the feature points and their positions of the SDP image is as follows: After the SDP image is converted into a grayscale image I(x,y), the two-dimensional Gaussian function G(x,y,σ) is used to obtain the Gaussian Laplace transform image and perform differential processing on it to obtain the Gaussian difference operator D(x,y,σ), which is constructed into a pyramid structure; Each pixel point of each layer of the Gaussian difference pyramid image is compared with the surrounding pixels of the same layer and adjacent layer images to find the image extreme points, obtain the feature points and their positions of the SDP image, and use the Hessian matrix H(x,y) to remove the extreme values ​​in the edge area.

6. The on-load tap changer fault diagnosis method based on multi-source signal fusion according to claim 5, characterized in that: According to the feature points and their positions of the SDP image, the specific process of generating multiple feature descriptors is as follows: The histogram is used to count the gradient and amplitude of pixels in the m×m neighborhood centered on each feature point, and the direction with the highest amplitude in the histogram is taken as the main direction; The m×m neighborhood is divided into multiple sub-regions, and the gradient directions in multiple directions are counted in each sub-region, eventually forming a SIFT feature descriptor with multiple data.

7. The on-load tap changer fault diagnosis method based on multi-source signal fusion according to claim 6, characterized in that: The specific process of matching based on the Euclidean distance between the feature descriptor of the SDP image and the feature descriptor of the template image is as follows: Establish key descriptor sets for the template image and the image to be detected respectively, complete the matching by comparing the key point descriptors in the two point sets, and use Euclidean distance to describe the similarity of the key point descriptors; When the distance between the matched key descriptors meets the preset conditions, it is considered that the feature point i in the image to be detected matches the feature point p in the template image successfully; The kd tree method is used to complete the feature sub-matching of the image to be diagnosed and the template image, and the feature sub-information of m key points that meet the threshold is obtained. The larger the number of matching points m, the more similar the two images are.

8. The on-load tap changer fault diagnosis method based on multi-source signal fusion according to any one of claims 2 to 7, characterized in that: The specific process of rotation sampling is as follows: The dictionary library is composed of N sampling units. Randomly sample n samples from the total library to form an initial dictionary. Perform test set diagnosis on the initial dictionary to obtain the initial dictionary diagnosis accuracy α0; The initial dictionary is gradually rotated and replaced to determine the optimal rotation amount u. During the rotation process, u dictionary samples with poor representation in the previous sample are discarded, and u samples are selected from the sampling units that were not selected in the past to form a new dictionary. Perform test set diagnosis on the new dictionary obtained by rotation sampling to obtain its diagnostic accuracy α k ; If the diagnostic accuracy rate α k If it is greater than the set threshold T or the maximum number of rotations reaches K, the iteration ends and the new dictionary is output.

9. The on-load tap changer fault diagnosis method based on multi-source signal fusion according to claim 8, characterized in that: The specific process of determining the optimal rotation amount u is: The sample size is fixed at n, and the sample is rotated. m sampling units are randomly selected from the sample, and their current observation value is recorded as y 11 ,y 12 ,…,y 1m , the sample mean is recorded as: Randomly select u units from the unselected units, u = nm, and record their current observation value as y 21 ,y 22 ,…,y 2m , the sample mean is recorded as: According to the dictionary diagnosis accuracy α k Construct homogeneous weighted estimator, 0≤α k ≤1, 0≤β k ≤1 and α k +β k =1, construct the mean weighted estimator: Weight α k , β k Determined by the accuracy of the previous sample; if α k Take a smaller value, β k If the value is large, it means that the sample judgment rate is low and the representativeness is poor; if α k Take a larger value, β k Taking a smaller value, the sample judgment rate is high, indicating that the sample is representative; From the independence of the two parts of sampling, we know that The variance of is described by the following formula, where S 2 Current population variance: because is a constant. Under the constraint condition of m+u=n, Minimizing the variance is equivalent to finding the values ​​of m and u so that Reach the minimum value; Construct the Lagrangian function, where λ is the Lagrangian coefficient, and find the partial derivatives of m and u respectively; Obtain Substitute the constraint condition m+u=n, and the optimal rotation amount is u=nβ.

10. A fault diagnosis system for on-load tap changer based on multi-source signal fusion, characterized in that: include: A signal acquisition module, used to acquire a detection signal of the on-load tap changer; the detection signal includes a voiceprint signal, a vibration signal, a voltage signal and a current signal; An image generation module is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal using the symmetrical point mode SDP algorithm to generate an SDP image; A feature extraction module is used to extract the scale-invariant feature transformation of the SDP image and obtain a feature subset; A fault diagnosis module is used to match the feature subset with a pre-built fault diagnosis dictionary to obtain a diagnosis result; The fault diagnosis dictionary contains a mapping relationship between feature subsets and fault states.

11. The on-load tap changer fault diagnosis system based on multi-source signal fusion according to claim 10, characterized in that: The process of constructing a fault diagnosis dictionary is as follows: Acquire the soundprint signals and vibration signals at multiple positions when the on-load tap-changer is operating in various states, as well as the voltage and current signals during the electromagnetic transient process of the on-load tap-changer; The SDP algorithm is used to fuse the voiceprint signal, vibration signal, voltage signal and current signal to generate a polar coordinate image set I = [I1,…,I2,…,IT], and the image set is divided into a selection set and a test set; In state t, select R groups of SDP images generated by test data from the image subset It in the selection set, randomly select r groups of SDP images for scale-invariant feature transformation feature extraction, and form the feature subset St in state t. Finally, the total feature subset S corresponding to all states is obtained = [S1, S2, ..., S T ]; In state t, key point matching is performed between the feature subset and the feature subsets of other states, and the total number of key point matches M of each feature in the feature subset and the randomly selected features of other states is calculated. The data is sorted in descending order according to the total number of matches M, and the first Q groups of data are selected as fault diagnosis reference templates, and finally the initial fault diagnosis dictionary D corresponding to all states is obtained; Match the SDP images in the test set with the fault diagnosis reference templates in the fault diagnosis initial dictionary D, count the number of key feature points of the SDP images to be diagnosed in the test set and all the fault diagnosis reference templates in the fault diagnosis initial dictionary D, sort them by size, count the top K groups of dictionary elements with the largest number of matches, and select the fault state with the highest frequency in the K groups as the fault diagnosis result; Calculate the accuracy of each state through the test set to determine the rotation rate s of the rotation sampling; Randomly select u units from the unselected units and replace the last u templates of the corresponding state in the dictionary D until the preset accuracy threshold S or the maximum number of rotations N is reached to obtain the final fault diagnosis dictionary.

12. The on-load tap changer fault diagnosis system based on multi-source signal fusion according to claim 11, characterized in that: The formula for generating polar coordinate image sets by the SDP algorithm is: Among them, r(i) is the polar coordinate radius; θ(i) is the counterclockwise deflection angle; is the clockwise deflection angle; x i is the i-th sample point of the signal; x max and x min are the maximum and minimum values ​​of the signal respectively; a is the time interval parameter; ε is the angle magnification factor; θ is the rotation angle of the mirror symmetry plane, a=1,2,3……N.

13. The on-load tap changer fault diagnosis system based on multi-source signal fusion according to claim 11, characterized in that: The process of scale-invariant feature transformation feature extraction is: Process the SDP image to locate the feature points and their positions of the SDP image; Generate multiple feature descriptors based on the feature points and positions of the SDP image; Matching is performed based on the Euclidean distance between the feature descriptor of the SDP image and the feature descriptor of the template image.

14. The on-load tap changer fault diagnosis system based on multi-source signal fusion according to claim 13, characterized in that: The specific process of processing the SDP image and locating the feature points and their positions of the SDP image is as follows: After the SDP image is converted into a grayscale image I(x,y), the two-dimensional Gaussian function G(x,y,σ) is used to obtain the Gaussian Laplace transform image and perform differential processing on it to obtain the Gaussian difference operator D(x,y,σ), which is constructed into a pyramid structure; Each pixel point of each layer of the Gaussian difference pyramid image is compared with the surrounding pixels of the same layer and adjacent layer images to find the image extreme points, obtain the feature points and their positions of the SDP image, and use the Hessian matrix H(x,y) to remove the extreme values ​​in the edge area.

15. The on-load tap changer fault diagnosis system based on multi-source signal fusion according to claim 14, characterized in that: According to the feature points and their positions of the SDP image, the specific process of generating multiple feature descriptors is as follows: The histogram is used to count the gradient and amplitude of pixels in the m×m neighborhood centered on each feature point, and the direction with the highest amplitude in the histogram is taken as the main direction; The m×m neighborhood is divided into multiple sub-regions, and the gradient directions in multiple directions are counted in each sub-region, eventually forming a SIFT feature descriptor with multiple data.

16. The on-load tap changer fault diagnosis system based on multi-source signal fusion according to claim 15, characterized in that: The specific process of matching based on the Euclidean distance between the feature descriptor of the SDP image and the feature descriptor of the template image is as follows: Establish key descriptor sets for the template image and the image to be detected respectively, complete the matching by comparing the key point descriptors in the two point sets, and use Euclidean distance to describe the similarity of the key point descriptors; When the distance between the matched key descriptors meets the preset conditions, it is considered that the feature point i in the image to be detected matches the feature point p in the template image successfully; The kd tree method is used to complete the feature sub-matching of the image to be diagnosed and the template image, and the feature sub-information of m key points that meet the threshold is obtained. The larger the number of matching points m, the more similar the two images are.

17. The on-load tap changer fault diagnosis system based on multi-source signal fusion according to any one of claims 11 to 16, characterized in that: The specific process of rotation sampling is as follows: The dictionary library is composed of N sampling units. Randomly sample n samples from the total library to form an initial dictionary. Perform test set diagnosis on the initial dictionary to obtain the initial dictionary diagnosis accuracy α0; The initial dictionary is gradually rotated and replaced to determine the optimal rotation amount u. During the rotation process, u dictionary samples with poor representation in the previous sample are discarded, and u samples are selected from the sampling units that were not selected in the past to form a new dictionary. Perform test set diagnosis on the new dictionary obtained by rotation sampling to obtain its diagnostic accuracy α k ; If the diagnostic accuracy rate α k If it is greater than the set threshold T or the maximum number of rotations reaches K, the iteration ends and the new dictionary is output.

18. The on-load tap changer fault diagnosis system based on multi-source signal fusion according to claim 17, characterized in that: The specific process of determining the optimal rotation amount u is: The sample size is fixed at n, and the sample is rotated. m sampling units are randomly selected from the sample, and their current observation value is recorded as y 11 ,y 12 ,…,y 1m , the sample mean is recorded as: Randomly select u units from the unselected units, u = nm, and record their current observation value as y 21 ,y 22 ,…,y 2m , the sample mean is recorded as: According to the dictionary diagnosis accuracy α k Construct homogeneous weighted estimator, 0≤α k ≤1, 0≤β k ≤1 and α k +β k =1, construct the mean weighted estimator: Weight α k , β k Determined by the accuracy of the previous sample; if α k Take a smaller value, β k If the value is large, it means that the sample judgment rate is low and the representativeness is poor; if α k Take a larger value, β k Taking a smaller value, the sample judgment rate is high, indicating that the sample is representative; From the independence of the two parts of sampling, we know that The variance of is described by the following formula, where S 2 Current population variance: because is a constant. Under the constraint condition of m+u=n, Minimizing the variance is equivalent to finding the values ​​of m and u so that Reach the minimum value; Construct the Lagrangian function, where λ is the Lagrangian coefficient, and find the partial derivatives of m and u respectively; Obtain Substitute the constraint condition m+u=n, and the optimal rotation amount is u=nβ.

19. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are performed.

20. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 9.

21. A fault diagnosis system for on-load tap changer based on multi-source signal fusion, comprising a memory and a processor connected to each other, wherein a computer program is stored in the memory, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 9.

Images