A circuit breaker fault diagnosis method and system based on multi-signal fusion features

Abstract

This invention discloses a circuit breaker fault diagnosis method based on multi-signal fusion features, comprising: acquiring vibration signals, coil current signals, and sound signals during the opening and closing process of the circuit breaker, and preprocessing them; using the preprocessed vibration signals, coil current signals, and sound signals to correspond to the brightness values ​​of the three color channels in an RGB image, and synthesizing the three signal data into color image data; training a probabilistic neural network model with the multi-signal fusion features obtained by feature extraction from the color image data to obtain a fault diagnosis model; diagnosing the circuit breaker fault using the fault diagnosis model to determine the fault type; and also disclosing a circuit breaker fault diagnosis system based on this method. This invention fuses data obtained from three different signal acquisition methods into a color image, extracts image features as multi-signal fusion features to train a fault identification model, and performs fault identification on the circuit breaker, thereby improving the accuracy of fault identification.

Description

Technical Field

[0001] This invention relates to the field of circuit breaker fault detection technology, and in particular to a circuit breaker fault diagnosis method and system based on multi-signal fusion features. Background Technology

[0002] High-voltage circuit breakers are crucial switching devices in power systems, and their operating status directly affects the reliability and stability of the power supply. Due to their complex structure and relatively high operating frequency, they are prone to failure. A high-voltage circuit breaker failure can cause not only significant economic losses but also serious safety accidents. According to statistics on high-voltage circuit breaker failure types both domestically and internationally, mechanical failures are the primary type of failure, generally including spring fatigue and latch malfunction. Typically, in the initial stages of a mechanical failure, the high-voltage circuit breaker can function normally. Therefore, timely detection and accurate diagnosis of high-voltage circuit breaker mechanical failures can effectively prevent further escalation and is crucial for maintaining the stable operation of the power system. However, in the actual operation of high-voltage circuit breakers, fault characteristics are complex and diverse, with different fault characteristics manifesting in different characteristic parameters; and existing fault diagnosis methods mostly rely on single characteristic parameters, which can lead to missed faults in some high-voltage circuit breaker systems.

[0003] The "Method for Detecting Mechanical Faults in High-Voltage Circuit Breakers" disclosed in Chinese patent literature, publication number CN112557004B, and publication date June 14, 2022, determines the effective information portion of the sound signal by identifying characteristic events during circuit breaker operation. This reduces the length and number of feature sound signals extracted, and avoids interference from noise signals at the beginning and end of the sound signal during identification. By filtering identifiable feature intervals, the amount of data processing in fault identification is reduced. The feature intervals to be identified are prioritized and detected sequentially according to their weights, resulting in higher detection efficiency. However, this technology only detects faults from a single sound signal perspective, which can easily lead to missed detections for faults that are easily detected by vibration or current signals but extremely difficult to identify by sound signals. Summary of the Invention

[0004] This invention aims to overcome the problem that existing technologies for detecting circuit breaker faults based on a single characteristic parameter are prone to missed faults. It provides a circuit breaker fault diagnosis method and system based on multi-signal fusion features. The method collects vibration signals, coil current signals, and sound signals during the circuit breaker's opening and closing process. Data from these three different signal acquisition methods are fused into a color image. Image features are extracted as multi-signal fusion features containing various signal characteristics to train a fault identification model, thereby improving the accuracy of fault identification.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A circuit breaker fault diagnosis method based on multi-signal fusion features includes:

[0007] Vibration signals, coil current signals, and sound signals are collected during the opening and closing process of the high-voltage circuit breaker and preprocessed.

[0008] The preprocessed vibration signal, coil current signal, and sound signal correspond to the brightness values ​​of the three color channels in the RGB image, respectively, and the three signal data are combined into color image data.

[0009] A fault diagnosis model is obtained by training a probabilistic neural network model based on the multi-signal fusion features obtained from feature extraction of color image data.

[0010] Fault diagnosis models are used to diagnose circuit breakers and determine the type of fault.

[0011] This invention integrates vibration signals, coil current signals, and sound signals for circuit breaker fault diagnosis. Different signal data have their own advantages in diagnosing fault types. Therefore, by increasing the types of signal data, the range of fault diagnosis types is expanded, improving the accuracy of fault diagnosis and avoiding missed detections. For the signal data fusion method, RGB images are used to comprehensively represent the three different signals. The color value of each pixel in the image contains the brightness values ​​of the R, G, and B channels, which can correspond to the three different signal data. Thus, data fusion is completed on the object before feature extraction, so that the subsequently extracted features contain information from the three signals, improving the applicability of fault types for circuit breaker fault detection and avoiding missed detections.

[0012] Preferably, the vibration signal, coil current signal, and sound signal are signal data collected within the same time period, containing several complete opening and closing processes;

[0013] The preprocessing process includes: discretizing the signal data, then normalizing it, and finally mapping the normalized result to the brightness value range [0, 255] of the RGB image.

[0014] In this invention, the vibration, current, and sound signal data collected from a complete opening and closing process contain most of the information reflecting the fault state of the circuit breaker. However, a set of data may have special characteristics. Therefore, data from several complete opening and closing processes can be collected as raw data for preprocessing. In order to generate a color image later, the three types of signal data have the same length. Since the time of a complete process is short, ranging from a few tenths of a second to a few seconds, in order to ensure data accuracy and subsequent image generation, discretization can be performed at one-millisecond intervals to obtain data arranged in chronological order. This data is then normalized to the range of [0,1] and expanded to a value in the range of [0,255] that can represent brightness values.

[0015] Preferably, the process of synthesizing color image data from the three signal data includes:

[0016] For the preprocessed signal data, the values ​​of the vibration signal, coil current signal and sound signal at the same time point are selected and respectively matched with the brightness values ​​of the three color channels of the RGB image to obtain the color corresponding to the time point.

[0017] After performing the same synthesis process on all signal data, a color string is obtained arranged in chronological order. The pixel size of the image is set according to the number of time points, and the color string is mapped to the color of each pixel in the image.

[0018] In this invention, the vibration signal value, coil current signal value, and sound signal value at the time point corresponding to the first discrete point are combined to form the color of the first pixel. This process is repeated to combine the signal data of all discrete points, resulting in a color string of all pixels. The number of pixels is the same as the number of discrete points and they are arranged in the order of the discrete points. For the color string, the pixel size of the image can be determined according to the number of pixels to convert it into a color image. For example, the number of discrete points corresponding to the acquisition time of a complete opening and closing process can be used as the image width (one row), and the number of opening and closing processes acquired at one time can be used as the image height (one column) to generate the image. Alternatively, the image size can be obtained by dividing the number of pixels into the product of the two closest integers.

[0019] Preferably, the process of obtaining multi-signal fusion features based on feature extraction from color image data includes:

[0020] Color feature vectors are obtained by extracting features containing color distribution information from color images.

[0021] After processing a color image into a grayscale image, features containing texture information are extracted to obtain a grayscale feature vector; the color feature vector and the grayscale feature vector are then concatenated to obtain a multi-signal fusion feature.

[0022] In this invention, the color feature information in the color image can reflect the numerical distribution characteristics of the three signal data, serving as part of the multi-signal fusion feature. After the color image is processed to remove color features, the texture information of the grayscale image can represent the changes in texture of various parts of the image, thereby indirectly reflecting the arrangement, repetition, and positional structure of the signal data, serving as another part of the multi-signal fusion feature. The final multi-signal fusion feature is obtained by concatenating the two feature vectors, which can contain different aspects of the three signal data, improving the training accuracy of the recognition model.

[0023] Preferably, the process of obtaining the color feature vector includes:

[0024] The mean of the first moment, the standard deviation of the second moment, and the deviation of the third moment are calculated for the brightness values ​​of the three color channels respectively, resulting in a color feature vector composed of nine parameter features.

[0025] In this invention, the color distribution of a color image can be represented by feature moments. Since color distribution information is mainly concentrated in low-order moments, the first, second, and third moments of color are sufficient to express the color distribution of the image, without the need for feature quantization.

[0026] Preferably, the process of obtaining the grayscale feature vector includes:

[0027] Calculate the probability P(i,j,θ,d) of the occurrence of point O1(x1,y1) with gray value i and point O2(x1+a(θ,d),x2+b(θ,d)) with gray value j in a grayscale image at a distance d in a given θ direction; and arrange all probabilities P into a square matrix with the gray value of point O1 as the row and the gray value of point O2 as the column to obtain the grayscale probability matrix.

[0028] Extract features from the gray-level probability matrix to form a gray-level feature vector.

[0029] In this invention, the joint probability density between two pixels with different positional relationships in a grayscale image is used to reflect the brightness value distribution characteristics in the grayscale image. This can reflect the comprehensive information about direction, adjacent intervals, and variation amplitude in the grayscale image, and reflect the image's arrangement, local features, and positional structure. Furthermore, since the grayscale probability matrix is ​​relatively large, generally N*N, where N is the grayscale level number and is 256, various features representing the grayscale probability matrix can be further extracted to form a grayscale feature vector.

[0030] Preferably, the process of extracting features from the gray-level probability matrix to form a gray-level feature vector includes:

[0031] Calculate sharpness features:

[0032]

[0033] Calculate local variation characteristics:

[0034]

[0035] The energy feature is obtained by summing the squares of each element P(i,j,θ,d) in the gray-level probability matrix; the entropy feature is obtained by summing the product of each element P(i,j,θ,d) in the gray-level probability matrix and the logarithm of that element.

[0036] In this invention, the sharpness feature reflects the clarity and depth of the image texture; the greater the depth difference, the clearer the effect. The local variation feature reflects the magnitude of local changes in the image texture. If the image texture is relatively uniform across different regions and changes slowly, the local variation feature will be larger, and vice versa. The energy feature reflects the uniformity of the image's grayscale distribution and the coarseness of the texture. If each element value is similar, the energy is smaller, indicating a finer texture. If some values ​​are large while others are small, the energy value is larger. The entropy feature is a measure of the randomness of the information contained in the image. The more dispersed the distribution, the greater the information content, and the greater the entropy. It reflects the complexity of the image texture.

[0037] Preferably, the probabilistic neural network model is divided into an input layer, a pattern layer, a summation layer, and an output layer in sequence. The input layer inputs the multi-signal fusion features of the circuit breaker to be detected, the number of neurons in the pattern layer is the same as the total number of training samples, the number of neurons in the summation layer is the same as the number of fault types, and the output layer outputs the fault type with the largest value in the summation layer as the diagnostic result.

[0038] In this invention, the probabilistic neural network is a feedforward neural network, which has a simple structure, good classification effect, and no local optima problem. The number of neurons in the input layer is equal to the dimension of the multi-signal fusion features. The multi-signal fusion features obtained from the circuit breaker to be detected are used as the test samples and input into the pattern layer. The matching degree between the test sample and each training sample in the pattern layer is calculated. In the summation layer, the probability that the test sample belongs to the fault type is calculated based on the matching degree between all training samples corresponding to each fault type and the test sample. In the output layer, the fault type with the highest probability is output as the recognition result.

[0039] A circuit breaker fault diagnosis system based on multi-signal fusion features includes:

[0040] The data acquisition module collects vibration signals, coil current signals, and sound signals during the opening and closing process of the circuit breaker;

[0041] The data processing module preprocesses the three types of data collected and synthesizes them into color image data;

[0042] The feature extraction module extracts features from color and grayscale images to obtain color feature vectors and grayscale feature vectors, and then concatenates them to obtain multi-signal fusion features.

[0043] The fault identification module identifies the fault type of the circuit breaker.

[0044] In this invention, three types of signal data—vibration signal, coil current signal, and sound signal—are collected together for fault diagnosis of circuit breakers. Different signal data have their own advantages in diagnosing fault types. Therefore, by increasing the types of signal data, the range of fault diagnosis types can be expanded, thereby improving the accuracy of fault diagnosis and avoiding missed detections. The data processing module may also include a process of filtering and denoising the various raw signal data collected at the beginning.

[0045] Preferably, the data acquisition module includes several vibration acquisition units, several current acquisition units, and several sound acquisition units; the acquired signal data of the same type are averaged and then preprocessed.

[0046] In this invention, the vibration signals, current signals, and sound signals collected at different locations of the circuit breaker have different numerical intensities due to their relative positional relationship with the fault source. Therefore, the vibration signals collected simultaneously by several vibration acquisition units can be averaged and used as the original vibration signals before preprocessing. This can clearly contain various signal information and will not be ignored due to small values. The operation methods for current signals and sound signals are the same. In addition, the coefficients before different color channels in the grayscale conversion process of the color image can be determined according to the number of vibration signals, current signals, and sound signals.

[0047] The present invention has the following beneficial effects: it collects vibration signals, coil current signals, and sound signals during the opening and closing process of a circuit breaker, fuses the data obtained from three different signal acquisition methods into a color image, and extracts image features as multi-signal fusion features containing various signal features to train a fault identification model, thereby improving the accuracy of fault identification; the multi-signal fusion features include both color feature vectors and grayscale feature vectors, which can reflect the numerical distribution and relative positional relationships of various signal data, so as to improve the accuracy of model training and thus improve the accuracy of fault identification. Attached Figure Description

[0048] Figure 1This is a flowchart of the circuit breaker fault diagnosis method of the present invention;

[0049] Figure 2 This is a schematic diagram illustrating the calculation of the grayscale probability matrix in an embodiment of the present invention. Detailed Implementation

[0050] The present invention will now be further described with reference to the accompanying drawings and specific embodiments.

[0051] like Figure 1 As shown, a circuit breaker fault diagnosis method based on multi-signal fusion features includes:

[0052] Vibration signals, coil current signals, and sound signals are collected during the opening and closing process of the high-voltage circuit breaker and preprocessed.

[0053] Vibration signal, coil current signal and sound signal are signal data collected within the same time period, containing several complete opening and closing processes;

[0054] The preprocessing process includes: discretizing the signal data, then normalizing it, and finally mapping the normalized result to the brightness value range [0, 255] of the RGB image.

[0055] The preprocessed vibration signal, coil current signal, and sound signal correspond to the brightness values ​​of the three color channels in the RGB image, respectively, and the three signal data are combined into color image data.

[0056] The process of synthesizing color image data from three signal data includes:

[0057] For the preprocessed signal data, the values ​​of the vibration signal, coil current signal and sound signal at the same time point are selected and respectively matched with the brightness values ​​of the three color channels of the RGB image to obtain the color corresponding to the time point.

[0058] After performing the same synthesis process on all signal data, a color string is obtained arranged in chronological order. The pixel size of the image is set according to the number of time points, and the color string is mapped to the color of each pixel in the image.

[0059] A fault diagnosis model is obtained by training a probabilistic neural network model with multi-signal fusion features extracted from color image data.

[0060] The process of obtaining multi-signal fusion features from color image data includes:

[0061] Color feature vectors are obtained by extracting features containing color distribution information from color images.

[0062] After processing a color image into a grayscale image, features containing texture information are extracted to obtain a grayscale feature vector; the color feature vector and the grayscale feature vector are then concatenated to obtain a multi-signal fusion feature.

[0063] The process of obtaining the color feature vector includes:

[0064] The mean of the first moment, the standard deviation of the second moment, and the deviation of the third moment are calculated for the brightness values ​​of the three color channels respectively, resulting in a color feature vector composed of nine parameter features.

[0065] The process of obtaining grayscale feature vectors includes:

[0066] Calculate the probability P(i,j,θ,d) of the occurrence of point O1(x1,y1) with gray value i and point O2(x1+a(θ,d),x2+b(θ,d)) with gray value j in a grayscale image at a distance d in a given θ direction; and arrange all probabilities P into a square matrix with the gray value of point O1 as the row and the gray value of point O2 as the column to obtain the grayscale probability matrix.

[0067] Extract features from the gray-level probability matrix to form a gray-level feature vector.

[0068] The process of extracting features from the gray-level probability matrix to form gray-level feature vectors includes:

[0069] Calculate sharpness features:

[0070]

[0071] Calculate local variation characteristics:

[0072]

[0073] The energy feature is obtained by summing the squares of each element P(i,j,θ,d) in the gray-level probability matrix; the entropy feature is obtained by summing the product of each element P(i,j,θ,d) in the gray-level probability matrix and the logarithm of that element.

[0074] The probabilistic neural network model is divided into an input layer, a pattern layer, a summation layer, and an output layer. The input layer inputs the multi-signal fusion features of the circuit breaker to be detected. The number of neurons in the pattern layer is the same as the total number of training samples. The number of neurons in the summation layer is the same as the number of fault types. The output layer outputs the fault type with the largest value in the summation layer as the diagnostic result.

[0075] Fault diagnosis models are used to diagnose circuit breakers and determine the type of fault.

[0076] This invention integrates vibration signals, coil current signals, and sound signals for circuit breaker fault diagnosis. Different signal data have their own advantages in diagnosing fault types. Therefore, by increasing the types of signal data, the range of fault diagnosis types is expanded, improving the accuracy of fault diagnosis and avoiding missed detections. For the signal data fusion method, RGB images are used to comprehensively represent the three different signals. The color value of each pixel in the image contains the brightness values ​​of the R, G, and B channels, which can correspond to the three different signal data. Thus, data fusion is completed on the object before feature extraction, so that the subsequently extracted features contain information from the three signals, improving the applicability of fault types for circuit breaker fault detection and avoiding missed detections.

[0077] In this invention, the vibration, current, and sound signal data collected from a complete opening and closing process contain most of the information reflecting the fault state of the circuit breaker. However, a set of data may have special characteristics. Therefore, data from several complete opening and closing processes can be collected as raw data for preprocessing. In order to generate a color image later, the three types of signal data have the same length. Since the time of a complete process is short, ranging from a few tenths of a second to a few seconds, in order to ensure data accuracy and subsequent image generation, discretization can be performed at one-millisecond intervals to obtain data arranged in chronological order. This data is then normalized to the range of [0,1] and expanded to a value in the range of [0,255] that can represent brightness values.

[0078] In this invention, the vibration signal value, coil current signal value, and sound signal value at the time point corresponding to the first discrete point are combined to form the color of the first pixel. This process is repeated to combine the signal data of all discrete points, resulting in a color string of all pixels. The number of pixels is the same as the number of discrete points and they are arranged in the order of the discrete points. For the color string, the pixel size of the image can be determined according to the number of pixels to convert it into a color image. For example, the number of discrete points corresponding to the acquisition time of a complete opening and closing process can be used as the image width (one row), and the number of opening and closing processes acquired at one time can be used as the image height (one column) to generate the image. Alternatively, the image size can be obtained by dividing the number of pixels into the product of the two closest integers.

[0079] In this invention, the color feature information in the color image can reflect the numerical distribution characteristics of the three signal data, serving as part of the multi-signal fusion feature. After the color image is processed to remove color features, the texture information of the grayscale image can represent the changes in texture of various parts of the image, thereby indirectly reflecting the arrangement, repetition, and positional structure of the signal data, serving as another part of the multi-signal fusion feature. The final multi-signal fusion feature is obtained by concatenating the two feature vectors, which can contain different aspects of the three signal data, improving the training accuracy of the recognition model.

[0080] In this invention, the color distribution of a color image can be represented by feature moments. Since the color distribution information is mainly concentrated in the lower-order moments, the first, second, and third moments of the color are sufficient to express the color distribution of the image, without the need to quantize the features.

[0081] In this invention, the joint probability density between two pixels with different positional relationships in a grayscale image is used to reflect the brightness value distribution characteristics in the grayscale image. This can reflect the comprehensive information about direction, adjacent intervals, and variation amplitude in the grayscale image, thereby reflecting the image's arrangement, local features, and positional structure. Furthermore, since the grayscale probability matrix is ​​relatively large, generally N*N, where N is the grayscale level number and is 256, various features representing the grayscale probability matrix can be further extracted to form a grayscale feature vector.

[0082] In this invention, the sharpness feature reflects the clarity and depth of the image texture; the greater the depth difference, the clearer the effect. The local variation feature reflects the magnitude of local changes in the image texture. If the image texture is relatively uniform across different regions and changes slowly, the local variation feature will be larger, and vice versa. The energy feature reflects the uniformity of the image's grayscale distribution and the coarseness of the texture. If each element value is similar, the energy is smaller, indicating a finer texture. If some values ​​are large while others are small, the energy value is larger. The entropy feature is a measure of the randomness of the information contained in the image. The more dispersed the distribution, the greater the information content, and the greater the entropy. It reflects the complexity of the image texture.

[0083] In this invention, the probabilistic neural network is a feedforward neural network, which has a simple structure, good classification effect, and no local optima problem. The number of neurons in the input layer is equal to the dimension of the multi-signal fusion features. The multi-signal fusion features obtained from the circuit breaker to be detected are used as the test samples and input into the pattern layer. The matching degree between the test sample and each training sample in the pattern layer is calculated. In the summation layer, the probability that the test sample belongs to the fault type is calculated based on the matching degree between all training samples corresponding to each fault type and the test sample. In the output layer, the fault type with the highest probability is output as the recognition result.

[0084] A circuit breaker fault diagnosis system based on multi-signal fusion features includes:

[0085] The data acquisition module collects vibration signals, coil current signals, and sound signals during the opening and closing process of the circuit breaker;

[0086] The data processing module preprocesses the three types of data collected and synthesizes them into color image data;

[0087] The feature extraction module extracts features from color and grayscale images to obtain color feature vectors and grayscale feature vectors, and then concatenates them to obtain multi-signal fusion features.

[0088] The fault identification module identifies the fault type of the circuit breaker.

[0089] The data acquisition module includes several vibration acquisition units, several current acquisition units, and several sound acquisition units; the acquired signal data of the same type are averaged and then preprocessed.

[0090] In this invention, three types of signal data—vibration signal, coil current signal, and sound signal—are collected together for fault diagnosis of circuit breakers. Different signal data have their own advantages in diagnosing fault types. Therefore, by increasing the types of signal data, the range of fault diagnosis types can be expanded, thereby improving the accuracy of fault diagnosis and avoiding missed detections. The data processing module may also include a process of filtering and denoising the various raw signal data collected at the beginning.

[0091] In this invention, the vibration signals, current signals, and sound signals collected at different locations of the circuit breaker have different numerical intensities due to their relative positional relationship with the fault source. Therefore, the vibration signals collected simultaneously by several vibration acquisition units can be averaged and used as the original vibration signals before preprocessing. This can clearly contain various signal information and will not be ignored due to small values. The operation methods for current signals and sound signals are the same. In addition, the coefficients before different color channels in the grayscale conversion process of the color image can be determined according to the number of vibration signals, current signals, and sound signals.

[0092] In an embodiment of the invention, vibration signals, coil current signals, and sound signals, encompassing several complete opening and closing processes, are first acquired using vibration acquisition units, current acquisition units, and sound acquisition units installed on the circuit breaker. The acquisition time periods are identical and of equal length. These signals are then discretized. Since a complete opening and closing process takes between a fraction of a second and several seconds, and the opening or closing operations of the circuit breaker are measured in milliseconds, a discrete time interval of 1 millisecond can be set. That is, the signal is sampled once every 1 millisecond to obtain discrete point signals {x}. n Then, normalization is performed:

[0093]

[0094] The obtained values ​​are in the range of [0,1]. Since the brightness values ​​of the color image are between [0,255], the normalized signal data can be multiplied by 255 to amplify it to the range of brightness values ​​of the image, so that the three signals can be combined into a color image later.

[0095] The vibration signal value, coil current signal value, and sound signal value corresponding to the first discrete point are respectively mapped to the color values ​​of the R, G, and B color channels of the first pixel, thus representing the color of the first pixel. In this way, the signal values ​​of all discrete points are combined to represent the color of the pixel, resulting in a color string arranged in the time order of the discrete points. After setting the pixel size of the image according to the number of pixels, the color string is mapped to a color image.

[0096] For the color feature vector of a color image:

[0097] The mean of the first moment reflects the overall brightness of an image; the larger the value, the brighter the image.

[0098]

[0099] The standard deviation of the second moment reflects the color distribution range of an image; the larger the value, the wider the color distribution range.

[0100]

[0101] The third-order moment deviation reflects the symmetry of the color distribution in the image:

[0102]

[0103] Where c represents the color channel of the image, and in a color image, c = 1, 2, 3 corresponds to R, G, and B respectively; M represents the number of pixels, P cq This represents the color, i.e., the brightness value, of the q-th pixel in the c-th color channel; thus, a color feature vector [E1,σ1,S1,E2,σ2,S2,E3,α3,S3] consisting of nine parameters is obtained.

[0104] Converting a color image to a grayscale image can be achieved using the formula Gray = α. R R+α G G+α B B, where α R α G and α B The coefficients are between 0 and 1, and the sum of the three is 1. When there are a vibration acquisition units, b current acquisition units, and c sound acquisition units in the data acquisition module, these three coefficients can be set to a / (a+b+c), b / (a+b+c), and c / (a+b+c), respectively. The corresponding coefficients are substituted according to the color channel corresponding to the specific signal type.

[0105] For the feature vector of a grayscale image: such as Figure 2As shown, the probability P(i,j,θ,d) of a point O1(x1,y1) with gray value i and a point O2(x1+a(θ,d),y1+b(θ,d)) with gray value j appearing in a grayscale image at a distance d along a given θ-angle direction is calculated. All probabilities P(i,j,θ,d) are then arranged into a square matrix with the gray value of point O1 as the row and the gray value of point O2 as the column, resulting in a grayscale probability matrix. In practice, four angular directions can be selected: 0 degrees, 45 degrees, 90 degrees, and 135 degrees, with an interval of 1, to calculate their respective grayscale probability matrices.

[0106] For the gray-level probability matrix, extract its features:

[0107] Sharpness characteristics:

[0108]

[0109] Local variation characteristics:

[0110]

[0111] Energy characteristics:

[0112]

[0113] Entropy characteristics:

[0114]

[0115] Correlation characteristics:

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] Where μ1, μ2, d1, and d2 are the mean and standard deviation of the gray-level probability matrix in rows and columns, respectively, thus yielding the final gray-level feature vector [CON,H,AE,Ent,Cor,μ1,μ2,d1,d2].

[0122] By concatenating the color feature vector and the grayscale feature vector, we obtain a multi-signal fusion feature [E1,σ1,S1,E2,σ2,S2,E3,σ3,S3,CON,H,AE,Ent,Cor,μ1,μ2,d1,d2] containing three types of signal feature information, for a total of eighteen dimensions.

[0123] In probabilistic neural network models, the multi-signal fusion features of the input layer are first combined with the weighting coefficients ω. i Multiplying them together yields the scalar product Z. i Passed to the pattern layer, select exp((Z) i -1) / σ 2 If ) is used as the activation function, then the probability of the output of the j-th neuron of the i-th fault type in the pattern layer is:

[0124]

[0125] Where p is the dimension of the multi-signal fusion feature, i.e., the dimension of the training samples and the samples to be detected; σ is the smoothing factor; X ij Let (XX) be the vector corresponding to the j-th training sample of the i-th fault type. ij ) T (XX ij The distance () is used to represent the Euclidean distance between the sample to be detected and the faulty sample. In the summation layer, for the i-th fault type, there exists L... i There are Q training samples with a total of Q fault types (including loose connecting pins, loose mechanism connecting bolts, stuck closing trip unit, stuck operating mechanism, insufficient coil voltage, stuck auxiliary coil contacts, and normal state, etc.). The probability f of the i-th fault type is output in the summation layer. i The outputs of neurons belonging to the same category in the pattern layer are averaged to obtain the following:

[0126]

[0127] The output layer selects the fault type with the highest probability value in the summation layer as the final circuit breaker fault type for diagnosis.

[0128] The above embodiments are further elaborations and descriptions of the present invention to facilitate understanding, and are not intended to limit the present invention in any way. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A circuit breaker fault diagnosis method based on multi-signal fusion features, characterized in that, include: Vibration signals, coil current signals, and sound signals during the opening and closing process of high-voltage circuit breakers are collected and preprocessed. The preprocessed vibration signal, coil current signal, and sound signal correspond to the brightness values ​​of the three color channels in the RGB image, respectively, and the three signal data are combined into color image data. Feature extraction is performed on color image data to obtain multi-signal fusion features composed of color feature vectors and grayscale feature vectors. The color image is processed to obtain a grayscale image. The probability P(i,j,θ,d) of the occurrence of pixel pairs with grayscale values ​​i and j at a distance d in the θ direction is calculated. The gray level feature vector includes: a sharpness feature, summing over all pixel pairs (i-j) 2 *P(i,j,θ,d) with all pixel pairs; a local variation feature, summing over all pixel pairs (i-j) 2 +1 as denominator and P(i,j,θ,d) as numerator, summing over all pixel pairs; an energy feature, summing over all pixel pairs (P(i,j,θ,d)) 2 with all pixel pairs; an entropy feature, summing over all pixel pairs P(i,j,θ,d)*logP(i,j,θ,d) A fault diagnosis model is obtained by training a probabilistic neural network model; Fault diagnosis models are used to diagnose circuit breakers and determine the type of fault.

2. The circuit breaker fault diagnosis method based on multi-signal fusion features according to claim 1, characterized in that, The vibration signal, coil current signal, and sound signal are signal data collected within the same time period, containing several complete opening and closing processes; The preprocessing process includes: discretizing the signal data, then normalizing it, and finally mapping the normalized result to the brightness value range [0, 255] of the RGB image.

3. The circuit breaker fault diagnosis method based on multi-signal fusion features according to claim 2, characterized in that, The process of synthesizing color image data from three signal data includes: For the preprocessed signal data, the values ​​of the vibration signal, coil current signal and sound signal at the same time point are selected and respectively matched with the brightness values ​​of the three color channels of the RGB image to obtain the color corresponding to the time point. After performing the same synthesis process on all signal data, a color string is obtained arranged in chronological order. The pixel size of the image is set according to the number of time points, and the color string is mapped to the color of each pixel in the image.

4. A circuit breaker fault diagnosis method based on multi-signal fusion features according to claim 1, 2, or 3, characterized in that, The process of obtaining multi-signal fusion features based on feature extraction from color image data includes: Color feature vectors are obtained by extracting features containing color distribution information from color images. After processing a color image into a grayscale image, features containing texture information are extracted to obtain a grayscale feature vector. The color feature vector and the grayscale feature vector are spliced ​​together to obtain the multi-signal fusion feature.

5. The circuit breaker fault diagnosis method based on multi-signal fusion features according to claim 4, characterized in that, The process of obtaining the color feature vector includes: The mean of the first moment, the standard deviation of the second moment, and the deviation of the third moment are calculated for the brightness values ​​of the three color channels respectively, resulting in a color feature vector composed of nine parameter features.

6. The circuit breaker fault diagnosis method based on multi-signal fusion features according to claim 4, characterized in that, The process of obtaining the grayscale feature vector includes: Calculate the point in a grayscale image that has a grayscale value of i at a distance d in a given θ-angle direction. and the point with gray value j probability of occurrence ; and with points The grayscale value is row, point The grayscale values ​​are arranged into a square matrix by arranging all probabilities P in the column to obtain the grayscale probability matrix; Extract features from the gray-level probability matrix to form a gray-level feature vector.

7. A circuit breaker fault diagnosis method based on multi-signal fusion features according to claim 1, 2, 3, 5, or 6, characterized in that, The probabilistic neural network model is divided into an input layer, a pattern layer, a summation layer, and an output layer. The input layer inputs the multi-signal fusion features of the circuit breaker to be detected. The number of neurons in the pattern layer is the same as the total number of training samples. The number of neurons in the summation layer is the same as the number of fault types. The output layer outputs the fault type with the largest value in the summation layer as the diagnostic result.

8. A circuit breaker fault diagnosis system based on multi-signal fusion features, applicable to the circuit breaker fault diagnosis method as described in any one of claims 1-7, characterized in that, include: The data acquisition module collects vibration signals, coil current signals, and sound signals during the opening and closing process of the circuit breaker; The data processing module preprocesses the three types of data collected and synthesizes them into color image data; The feature extraction module extracts features from color and grayscale images to obtain color feature vectors and grayscale feature vectors, and then concatenates them to obtain multi-signal fusion features. The fault identification module identifies the fault type of the circuit breaker.

9. A circuit breaker fault diagnosis system based on multi-signal fusion features according to claim 8, characterized in that, The data acquisition module includes several vibration acquisition units, several current acquisition units, and several sound acquisition units; the acquired signal data of the same type are averaged and then preprocessed.

Images