Converter transformer multi-granularity fault identification model construction method
Through the generalized minimum entropy deconvolution and coarse-grained grid algorithm combined with the adaptive sparse attention diagnosis model, the noise separation and feature extraction difficulties in converter transformer fault diagnosis are solved, and higher diagnostic accuracy and reliability are achieved.
Patent Information
- Application Number
- CN202510428372.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-07
- Publication Date
- 2025-07-25
AI Technical Summary
The prior art has problems such as noise separation, limitations in model applicability and difficulty in extracting fault characteristics in converter transformer fault diagnosis, resulting in low diagnostic accuracy.
The generalized minimum entropy deconvolution and coarse-grained grid algorithm are used to feature enhance the vibration data. Combined with the adaptive sparse attention diagnosis model, the optimal FIR filter is searched by the standard matrix of the maximum filtering signal, fault characteristics are extracted and redundant sequences are filtered to achieve fault diagnosis.
It effectively reduces the impact of noise, improves the accuracy and stability of fault diagnosis, simplifies the feature extraction process, and improves the reliability of fault diagnosis of converter transformer.
Smart Images
Figure CN120372245A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of transformer fault diagnosis, and specifically to a method for constructing a multi-granularity fault identification model for a converter transformer. Background Art
[0002] As a key device for voltage conversion and power transmission, the safe and reliable operation of a converter transformer (hereinafter referred to as a converter) is an important guarantee for the stable operation of the entire power system. The fault detection of traditional converters requires power-off processing. Due to the complex internal structure of the converter, there are objectively disadvantages such as inflexible periodic maintenance, and it is easy for operators to have deviations and other problems, resulting in low work efficiency, affecting power supply and possibly causing secondary injuries. Therefore, with the continuous increase in power supply demand, it is of great significance to explore an intelligent fault diagnosis method for converters to improve the safety and stability of converters.
[0003] To improve the accuracy of converter fault diagnosis, it is necessary to clarify the converter fault types, extract effective characteristic values for the vibration signals generated under different operating conditions, and conduct in-depth analysis in combination with the vibration mechanism of the converter. Although traditional methods (such as the three-ratio method and Rogers' four-ratio method) are effective in some scenarios, their applicability has limitations, especially for special types of equipment such as converters, and they cannot perform fault diagnosis. In recent years, data-driven methods have received extensive attention in the field of fault diagnosis. This method collects, processes, and analyzes a large amount of operation data, extracts the effective information hidden in the data, and provides theoretical support for fault diagnosis. Commonly used data-driven methods include neural networks, support vector machines, empirical mode decomposition, and particle filtering. These methods can effectively process complex non-linear relationships and high-dimensional data, providing important technical support for the intelligent operation and maintenance of converters and other power equipment.
[0004] However, the existing methods still have the following deficiencies: (1) There are significant deficiencies in noise reduction. Vibration signals often contain mechanical noise, electromagnetic interference, and environmental noise, and these noises overlap with the fault characteristic signals in the frequency domain and time domain, making it difficult for traditional filtering methods to effectively separate them; (2) When the model deals with complex and changeable operating conditions, it is usually difficult to effectively extract potential fault characteristics. Especially when dealing with data of diverse fault types, the performance of a single-characteristic model has significant limitations; (3) Current machine learning and deep learning algorithm models often have difficulty effectively extracting and identifying the key information in complex signals, resulting in a low accuracy rate of fault detection.
[0005] Therefore, there is an urgent need to develop a model that can accurately diagnose the faults of converters, and through continuous verification and optimization, achieve precise diagnosis of converter faults. Summary of the Invention
[0006] To solve the above problems, the object of the present invention is to provide a method for constructing a multi-granularity fault identification model for a converter transformer.
[0007] To achieve the above object, the present invention is realized through the following technical solutions:
[0008] A method for constructing a multi-granularity fault identification model for a converter transformer, first collect the vibration data of the converter transformer under different faults, then use the generalized minimum entropy deconvolution to enhance the features of the collected vibration data, and finally use the coarse-grained grid algorithm to extract fault features and establish an adaptive sparse attention diagnosis model.
[0009] Preferably, a method for constructing a multi-granularity fault identification model for a converter transformer includes the following steps:
[0010] ① Collect the vibration fault data of the converter transformer: Collect the vibration signals of the converter transformer in normal and abnormal states, label the fault types, form a converter transformer vibration fault data set containing multiple fault types, and then divide the training set and the test set;
[0011] ② Enhance the features of the generalized minimum entropy deconvolution signal;
[0012] ③ Extract the coarse-grained grid fault features;
[0013] ④ Construct an adaptive sparse attention diagnosis model: Refine the features of the fault feature map through the attention network, use the soft threshold to filter the attention weight sequence, and ignore the redundant sequence through sparse operations, so as to enhance the attention to the fault features and realize fault diagnosis.
[0014] Preferably, the fault types in step ① include winding looseness, outlet device looseness, and common looseness of the winding and the outlet device.
[0015] Preferably, in step ①, a vibration acceleration sensor is used to collect the vibration signals of the converter transformer in normal and abnormal states, with a sensitivity of 100 mV / g and a sampling frequency of 10 kHz.
[0016] Preferably, when dividing the training set and the test set in step ①, it is randomly divided according to a ratio of 7:3.
[0017] Preferably, the generalized minimum entropy deconvolution is used to search for the optimal finite impulse response filter by maximizing the standard matrix of the filtered signal, reduce the noise information in the signal, and enhance the fault signal of the converter transformer.
[0018] Further preferably, the converter transformer vibration signal x in step ② is composed of three basic components: the fault pulse d, the environmental noise e, and the rotation frequency and its harmonics u;
[0019] x = d * W d + u * Wu +e*W e
[0020] where W d is the transfer function of the system fault pulse, and W e is the transfer function of the system environmental noise, and W u is the transfer function of the system rotation frequency and its harmonics, and * represents convolution;
[0021] The blind deconvolution technique searches for the FIR filter by maximizing the statistical measure of the filtered signal, thereby reducing the influence of environmental noise and enhancing the effective features. The calculation method is as follows:
[0022] y = x * f = (d * W d + u * W u + e * W e ) * f
[0023] where f represents the FIR filter;
[0024] Based on the theory of the blind deconvolution technique, the generalized minimum entropy deconvolution searches for the FIR filter by maximizing the standard matrix of the filtered signal. The calculation process is as follows:
[0025]
[0026] where N represents the signal length, y represents the mean-removed signal filtered by the FIR filter, k is the order of the standard matrix, and is an integer greater than 2;
[0027] By setting the gradient of the standard matrix filter coefficients to zero, thereby maximizing the standard matrix, it is realized by gradual iteration. The calculation process is as follows:
[0028]
[0029] where y i represents the signal after the i-th filtering, and f i+1 represents the filter at the (i + 1)-th time.
[0030] Preferably, perform a frequency-domain transformation on the enhanced vibration signal, and continuously segment the spectrum using the coarse-grained method, perform an amplitude enhancement operation in each section, and convert the data into a coarse-grained grid feature map that enhances the fault features to achieve fault feature extraction;
[0031] More preferably, step ③ specifically includes the following steps:
[0032] First, use the FFT to transform the signal from the time domain to the frequency domain to obtain the spectrum;
[0033] Set the coarse-grained sampling coefficient, divide the spectrum into continuous j segments, and perform amplitude and enhancement operations within each segment. The specific calculation process is as follows:
[0034] j = fix(r / d)
[0035] F cn (j) = sum(F c ((j - 1) × d + 1 : j × d))
[0036] where d is the coarse-grained sampling coefficient, r is the spectrum width, fix() represents a floor function, and F cn represents the corresponding amplitude value;
[0037] Secondly, according to the number of new features, set the feature value range S = j, and correspondingly adjust the size of each feature within the range S;
[0038]
[0039] F cm represents the final feature value, F cnmin is the minimum feature value, and F cnmax is the maximum feature value.
[0040] Preferably, step ④ specifically includes the following steps:
[0041] First, a scale parameter layer and a maximum attention layer are added after the attention layer. The scale parameter layer takes the output z of the attention layer as input and finally generates a scale parameter α:
[0042] α = sigmoid(W α z + b α )
[0043] where W α and b α are the trainable weight matrix and bias term respectively;
[0044] The maximum attention layer takes the output z of the attention layer and the output α of the scale parameter layer as input and finally generates an attention threshold τ. The calculation process is as follows:
[0045] T = α · max(z)
[0046] Secondly, construct a filtering function using the natural exponential function. The calculation process is as follows:
[0047]
[0048] where when z < T, the output result of f(z) tends to 0; when z = T, f(z) equals 0.5; when z > T, f(z) tends to 1;
[0049] Finally, the softmax function is used to reconstruct the outputs of the threshold filter and the attention layer to obtain the attention coefficient newz i ; The calculation process is as follows:
[0050]
[0051] where N seg represents the number of sequences.
[0052] The present invention has the following advantages compared with the prior art:
[0053] For the method for constructing a multi-granularity fault identification model of a converter transformer of the present invention, first, aiming at the problem of strong noise in the vibration signal of the converter transformer, the generalized minimum entropy deconvolution searches for the optimal FIR filter by maximizing the standard matrix of the filtered signal, thereby enhancing the fault signal of the converter transformer, and having the advantages of strong stability and wide applicability; secondly, when extracting signal features, the coarse-grained grid feature method is used to effectively extract fault features, and the feature extraction process is simplified by spectrum segmentation and amplitude enhancement, reducing the computational complexity, reducing noise, and improving the reliability of the results; finally, the adaptive sparse attention network uses a soft threshold to filter the attention weight sequence and ignores redundant sequences through sparse operations, focusing on the corresponding fault features; through the non-linear transformation layer, the adaptive learning attention threshold is used to filter out smaller attention coefficients, improving the accuracy of fault diagnosis. Description of the Drawings
[0054] Figure 1 is a flowchart for collecting vibration fault numbers (data acquisition) of the converter transformer;
[0055] Figure 2 is a schematic diagram of enhancing signal features by using generalized minimum entropy deconvolution and extracting features by using the coarse-grained grid method;
[0056] Figure 3 is a flowchart for constructing the adaptive sparse attention diagnosis model in step ④;
[0057] Figure 4 is a flowchart for validating the fault diagnosis model;
[0058] Figure 5 is a schematic diagram of the vibration data acquisition result of the converter transformer;
[0059] Figure 6 is a schematic diagram of the result of minimum entropy deconvolution noise reduction;
[0060] Figure 7 is a schematic diagram of the result of coarse-grained network feature extraction;
[0061] Figure 8Schematic diagram of the training result of the fault diagnosis model;
[0062] Figure 9 Schematic diagram of the comparison result of the fault diagnosis model. Specific implementation mode
[0063] The purpose of the present invention is to provide a method for constructing a multi-granularity fault identification model of a converter transformer. The following further describes the present invention in conjunction with specific embodiments.
[0064] Embodiment 1
[0065] A method for constructing a multi-granularity fault identification model of a converter transformer includes the following steps:
[0066] ① Collect vibration fault data of the converter transformer: As Figure 1 shown, collect the vibration signals of the converter transformer in normal and abnormal states, label the fault types, form a vibration fault dataset of the converter transformer containing multiple fault types, and then divide it into a training set and a test set;
[0067] Specifically, use vibration sensors to collect the vibration signals of the converter transformer in a power plant in normal and faulty operating states. The fault types include winding looseness, outgoing line device looseness, and common looseness of the winding and outgoing line device. The sensitivity of the vibration acceleration sensor for collecting signals is 100 mV / g, and the sampling frequency is 10 kHz. The collected signals are randomly divided into a training set and a test set according to a ratio of 7:3 to train and evaluate the subsequent fault diagnosis model.
[0068] ② Enhance the features of the generalized minimum entropy deconvolution signal: Use the generalized minimum entropy deconvolution to search for the optimal finite impulse response filter by maximizing the standard matrix of the filtered signal, reduce the noise information in the signal, and enhance the fault signal of the converter transformer;
[0069] The vibration signal x of the converter transformer consists of three basic components: the fault impulse d, the environmental noise e, and the rotational frequency and its harmonics u;
[0070] x = d * W d + u * W u + e * W e
[0071] where W d is the transfer function of the system fault impulse, W e is the transfer function of the system environmental noise, W u is the transfer function of the system rotational frequency and its harmonics, and * represents convolution;
[0072] The blind deconvolution technology searches for the FIR filter by maximizing the statistical measure of the filtered signal, thereby reducing the influence of environmental noise and enhancing the effective features. The calculation method is as follows:
[0073] y = x * f = (d * W d + u * W u + e * W e ) * f
[0074] Where f represents the FIR filter;
[0075] As Figure 2 shown, based on the theory of blind deconvolution technology, the generalized minimum entropy deconvolution searches for the FIR filter by maximizing the criterion matrix of the filtered signal, and the calculation process is as follows:
[0076]
[0077] Where N represents the signal length, y represents the mean-removed signal filtered by the FIR filter, k is the order of the criterion matrix, and is an integer greater than 2;
[0078] By setting the gradient of the criterion matrix filter coefficients to zero, thus maximizing the criterion matrix, and realizing it through step-by-step iteration, the calculation process is as follows:
[0079]
[0080] Where y i represents the signal after the i-th filtering, and f i+1 represents the filter at the (i + 1)-th time.
[0081] ③ Extract the coarse-grained grid fault features: Perform a frequency-domain transformation on the enhanced vibration signal, and use the coarse-grained method to continuously divide the spectrum, perform amplitude enhancement operations in each section, and convert the data into a coarse-grained grid feature map of enhanced fault features to achieve fault feature extraction; Coarse-grained grid feature extraction is a feature enhancement fusion method based on spectral analysis. It can effectively extract fault features under strong background noise, has a simple structure, and high computational efficiency.
[0082] First, as Figure 2 shown, use the FFT to transform the signal from the time domain to the frequency domain to obtain the spectrum;
[0083] Set the coarse-grained sampling coefficient, divide the spectrum into continuous j segments, and perform amplitude and enhancement operations within each segment. The specific calculation process is as follows:
[0084] j = fix(r / d)
[0085] F cn (j) = sum(F c ((j - 1) × d + 1: j × d))
[0086] where d is the coarse-grained sampling coefficient, r is the spectral width, fix() represents a floor function, and F cn represents the corresponding amplitude value;
[0087] Secondly, according to the number of new features, set the feature value range S = j, and correspondingly adjust the size of each feature within the range S;
[0088]
[0089] F cm represents the final feature value, F cnmin is the minimum feature value, and F cnmax is the maximum feature value.
[0090] After the above operations, the signal with enhanced features is subjected to a frequency-domain transformation, and the spectrum is continuously segmented using the coarse-grained method. An amplitude enhancement operation is performed in each section, and the data is converted into a coarse-grained grid feature map with enhanced fault features to achieve fault feature extraction. The extracted feature map is used as the input for the subsequent adaptive sparse attention network to complete the training and classification of the converter transformer fault diagnosis model.
[0091] ④ Construct an adaptive sparse attention diagnosis model: Refine the fault feature map through the attention network, filter the attention weight sequence using a soft threshold, and ignore redundant sequences through sparse operations to enhance the focus on fault features and achieve fault diagnosis.
[0092] The attention mechanism applies a set of attention coefficients to the entire sequence. The attention coefficients applied to more important sequence segments are larger, while those applied to less important sequence segments are smaller, thereby enhancing the classification ability of the network. However, in fault diagnosis applications, the attention coefficients are dispersed over more sequence segments, contributing less to classification, having a high computational cost, and focusing on more redundant information.
[0093] To solve the problems existing in the traditional attention mechanism, the present invention proposes an adaptive sparse attention network. Through a non-linear transformation layer, an adaptive learning attention threshold is used to filter out smaller attention coefficients, thereby achieving the effect of diluting attention.
[0094] As Figure 3 shown, first, a scale parameter layer and a maximum attention layer are added after the attention layer. The scale parameter layer takes the output z of the attention layer as input and finally generates a scale parameter α:
[0095] α = sigmoid(W α z + b α )
[0096] where W α and bα They are a trainable weight matrix and a bias term respectively;
[0097] The maximum attention layer takes the output z of the attention layer and the output α of the scale parameter layer as inputs, and finally generates an attention threshold τ. The calculation process is as follows:
[0098] τ = α · max(z)
[0099] Secondly, a filtering function is constructed using the natural exponential function. The calculation process is as follows:
[0100]
[0101] where, when z < T, the output result of f(z) tends to 0; when z = T, f(z) is equal to 0.5; when z > T, f(z) tends to 1;
[0102] Finally, the softmax function is used to reconstruct the outputs of the threshold filter and the attention layer to obtain the attention coefficient newz i ; The calculation process is as follows:
[0103]
[0104] where, N seg represents the number of sequences.
[0105] After completing the training of the converter transformer fault diagnosis model based on the adaptive sparse attention network, the data of the test set is input into the model and verified multiple times. The results are as Figure 4 shown. The fault types of the converter transformer are predicted and output, the causes of the faults are analyzed, and the faulty converter transformer is repaired in time. By collecting the vibration data of a certain transformer and using classic models such as SVM and CNN for fault diagnosis, and comparing the fault diagnosis results with those of the adaptive sparse attention diagnosis model, and at the same time using the diagnostic accuracy as an evaluation index to compare and analyze the diagnostic results, the advantages of the proposed basic fault diagnosis model are verified.
[0106] Example 2
[0107] The method for constructing a multi-granularity fault identification model of a converter transformer described in Example 1 is adopted for implementation. The specific steps are as follows:
[0108] ① Collect the vibration fault data of the converter transformer: Collect the vibration signals of the converter transformer in normal and abnormal states, label the fault types, form a converter transformer vibration fault data set containing multiple fault types, and then divide the training set and the test set; Collect the trend changes of the vibration signals of the converter transformer as Figure 5 shown.
[0109] ②Enhance the characteristics of the generalized minimum entropy deconvolution signal: The generalized minimum entropy deconvolution is used to search for the optimal finite impulse response filter by maximizing the standard matrix of the filtered signal, reduce the noise information in the signal, and enhance the fault signal of the converter transformer; The results are as Figure 6 shown. It can be seen that the noise signal is suppressed and the vibration signal is enhanced.
[0110] ③Extract the coarse-grained grid fault characteristics: Perform a frequency domain transformation on the enhanced vibration signal, and use the coarse-grained method to continuously divide the spectrum. Perform an amplitude enhancement operation in each section, and convert the data into a coarse-grained grid feature map of enhanced fault characteristics to achieve fault feature extraction; The results are as Figure 7 shown. It can be seen from the frequency diagram that the high-frequency noise signal is suppressed and the low-frequency signal characteristics are enhanced.
[0111] ④Construct an adaptive sparse attention diagnosis model: Refine the fault feature map through the attention network, filter the attention weight sequence using a soft threshold, and ignore the redundant sequence through sparse operations, so as to enhance the attention to the fault characteristics and achieve fault diagnosis; The operation process of the adaptive sparse attention diagnosis model is as Figure 8 shown.
[0112] Compare the multi-granularity fault identification model of the converter transformer of the present invention with the existing models. The results are as Figure 9 shown. It can be seen that the accuracy of the fault diagnosis of the model proposed in this patent is significantly better than that of the support vector machine (Support Vector Machine, SVM) and the convolutional neural network (Convolutional Neural Networks, CNN).
Claims
1. A method for constructing a multi-granularity fault identification model of a converter transformer, characterized in that: First, collect the vibration data of the converter transformer under different faults. Then, use the generalized minimum entropy deconvolution to enhance the features of the collected vibration data. Finally, adopt the coarse-grained grid algorithm to extract the fault features and establish an adaptive sparse attention diagnosis model.
2. A method for constructing a multi-granularity fault identification model of a commutation transformer according to claim 1, characterized in that: It includes the following steps: ① Collect the vibration fault data of the converter transformer: Collect the vibration signals of the converter transformer in normal and abnormal states, label the fault types, form a vibration fault data set of the converter transformer containing multiple fault types, and then divide it into a training set and a test set; ② Enhance the features of the generalized minimum entropy deconvolution signal; ③ Extract the coarse-grained grid fault features; ④ Construct an adaptive sparse attention diagnosis model: Refine the features of the fault feature map through the attention network, filter the attention weight sequence using the soft threshold, and ignore the redundant sequence through sparse operations, so as to enhance the attention to the fault features and achieve fault diagnosis.
3. A method for constructing a multi-granularity fault identification model of a converter transformer according to claim 2, characterized in that: The fault types described in step ① include winding looseness, outgoing line device looseness, and common looseness of the winding and the outgoing line device.
4. A method for constructing a multi-granularity fault identification model of a commutation transformer according to claim 2, characterized in that: In step ①, a vibration acceleration sensor is used to collect the vibration signals of the converter transformer in normal and abnormal states, with a sensitivity of 100 mV / g and a sampling frequency of 10 kHz.
5. A method for constructing a multi-granularity fault identification model of a commutation transformer according to claim 2, characterized in that: In step ①, when dividing the training set and the test set, it is randomly divided according to the ratio of 7:
3.
6. The method for constructing a multi-granularity fault identification model of a commutation transformer according to claim 2, characterized in that: In step ②, the generalized minimum entropy deconvolution is used to search for the optimal finite impulse response filter by maximizing the standard matrix of the filtered signal, reduce the noise information in the signal, and enhance the fault signal of the converter transformer.
7. A method for constructing a multi-granularity fault identification model of a commutation transformer according to claim 6, characterized in that: In step ②, the vibration signal x of the converter transformer consists of three basic components: the fault pulse d, the environmental noise e, and the rotation frequency and its harmonics u; x = d*W d + u*W u + e*W e where W d is the transfer function of the system fault pulse, W e is the transfer function of the system environmental noise, W u is the transfer function of the system rotation frequency and its harmonics, * denotes convolution; The blind deconvolution technology searches for the FIR filter by maximizing the statistical measure of the filtered signal, so as to reduce the influence of environmental noise, enhance the effective features, and the calculation method is as follows: y = x * f = (d * W d + u * W u + e * W e ) * f where, f represents the FIR filter; Based on the theory of blind deconvolution technology, the generalized minimum entropy deconvolution searches for the FIR filter by maximizing the standard matrix of the filtered signal, and the calculation process is as follows: where, N represents the signal length, y represents the mean-removed signal filtered by the FIR filter, k is the order of the standard matrix, and is an integer greater than 2; By setting the gradient of the standard matrix filter coefficient to zero, the standard matrix is maximized and realized by gradual iteration, and the calculation process is as follows: where y i represents the signal after the i-th filtering, and f i+1 represents the filter at the (i + 1)-th time.
8. A method for constructing a multi-granularity fault identification model of a commutation transformer according to claim 2, characterized in that: In step ③, perform a frequency domain transformation on the enhanced vibration signal, and use the coarse-grained method to continuously segment the spectrum, perform an amplitude enhancement operation in each section, and convert the data into a coarse-grained grid feature map that enhances the fault features to achieve fault feature extraction.
9. The method for constructing a multi-granularity fault identification model of a converter transformer according to claim 8, wherein: Step ③ specifically includes the following steps: First, use the FFT to convert the signal from the time domain to the frequency domain to obtain the spectrum; Set the coarse-grained sampling coefficient, divide the spectrum into continuous j segments, and perform amplitude and enhancement operations within each segment. The specific calculation process is as follows: j = fix(r / d) F cn F(j) = sum(F c ((j - 1)×d + 1:j×d)) where d is the coarse-grained sampling coefficient, r is the spectral width, fix() represents a floor function, and F cn represents the corresponding amplitude value; Secondly, according to the number of new features, set the eigenvalue range S = j, and adjust the size of each feature accordingly within the range S; F cm represents the final eigenvalue, and F cnmin is the minimum eigenvalue, and F cnmax is the maximum eigenvalue.
10. A method for constructing a multi-granularity fault identification model of a commutation transformer according to claim 2, characterized in that: Step ④ specifically includes the following steps: First, a scale parameter layer and a maximum attention layer are added after the attention layer. The scale parameter layer takes the output z of the attention layer as input and finally generates a scale parameter α: α = sigmoid(W α z + b α ) Among which W α is a trainable weight matrix, and b α is a trainable bias term; The maximum attention layer takes the output z of the attention layer and the output α of the scale parameter layer as input and finally generates an attention threshold T. The calculation process is as follows: T = α · max(z) Secondly, a filtering function is constructed using the natural exponential function. The calculation process is as follows: Among them, when z < T, the output result of f(z) tends to 0; when z = T, f(z) is equal to 0.5; when z > T, f(z) tends to 1; Finally, the softmax function is used to reconstruct the outputs of the threshold filter and the attention layer to obtain the attention coefficient newz i ; The calculation process is as follows: Where N seg represents the number of sequences.