A method, system, device and medium for constructing a feature space of multi-source heterogeneous data of an excitation transformer

Abstract

The present application relates to the technical field of multi-source heterogeneous signal data fusion, and discloses a method, system, device and medium for constructing a feature space of multi-source heterogeneous data of an excitation transformer, comprising: converting time-frequency domain information into a two-dimensional graph using an improved symmetric point pattern method on data in a vibration signal window, extracting graph features using an image feature descriptor, and forming a vibration signal feature vector; extracting time domain statistics of data in a thermoelectric signal window, evaluating feature importance of the time domain statistics using a random forest algorithm, correcting feature scores by introducing a stable weight, and then performing feature sorting and dimension optimization, screening an optimal feature subset, and forming a thermoelectric signal feature vector; and splicing the two feature vectors to construct a multi-source heterogeneous data feature space of the excitation transformer. The present application constructs an excitation transformer feature space with better discrimination and stability, effectively improving the accuracy of equipment operation state evaluation and fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of multi-source heterogeneous signal data fusion technology, and in particular to a method, system, device and medium for constructing the feature space of multi-source heterogeneous data of an excitation transformer. Background Technology

[0002] Excitation transformers, as key power supply equipment in the excitation systems of generators and synchronous motors, are widely used in new energy power generation scenarios such as hydropower plants, playing a vital role in maintaining voltage regulation and grid stability. A failure in the excitation transformer will directly affect the safe and stable operation of the power system.

[0003] Currently, fault diagnosis methods based on multi-source information fusion are considered an effective way to improve the accuracy and robustness of excitation transformer condition monitoring. Vibration signals can effectively reflect mechanical faults such as core magnetostriction and coil loosening; winding temperature, operating voltage, and current signals are closely related to electrical faults such as inter-turn short circuits. By fusing multi-source information from different sensors, comprehensive perception and early warning of equipment status can be achieved. However, existing multi-source information fusion methods still face significant challenges, especially when processing various heterogeneous signals such as vibration, temperature, and electrical signals. Different signals have significant differences in sampling rate, dynamic characteristics, and physical meaning. Vibration signals have transient and non-stationary characteristics, requiring short-time analysis to capture the impact components of faults; while temperature and electrical signals change slowly, requiring longer observation windows to reflect their heat accumulation and trend changes. Existing methods mostly focus on feature extraction and fusion of similar data. For various signals with large differences in sampling rate and time-varying characteristics, an effective unified feature representation and fusion framework has not yet been formed, making it difficult to construct a multi-source feature space with strong discriminative power and high stability, thus restricting further improvement in the accuracy and reliability of fault diagnosis.

[0004] Therefore, there is an urgent need for a feature space construction method that can effectively integrate multi-source heterogeneous data from excitation transformers and overcome differences in signal characteristics, so as to provide a technical foundation for accurate equipment status assessment and fault diagnosis. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a method, system, device, and medium for constructing the feature space of multi-source heterogeneous data from excitation transformers, addressing the challenges faced by multi-source heterogeneous signal data fusion in the prior art, particularly how to effectively address the differences in signal characteristics under different sampling rates, the complexity of feature space construction, and the problems of insufficient stability and discrimination ability in the feature selection process.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for constructing the feature space of multi-source heterogeneous data of an excitation transformer, comprising:

[0009] Collect the operating parameter signals of the excitation transformer and divide the sampling window;

[0010] The improved symmetric point pattern method is used to convert the time-frequency domain information into a two-dimensional spectrum of the data within the vibration signal window, and image feature descriptors are used to extract the spectrum features to form a vibration signal feature vector.

[0011] The time-domain statistics of the data within the thermoelectric signal window are extracted, and the random forest algorithm is used to evaluate the feature importance of the time-domain statistics. Stable weights are introduced to correct the feature scores.

[0012] Based on the corrected feature scores, feature sorting and dimensionality optimization are performed to select the optimal feature subset and form the thermoelectric signal feature vector.

[0013] The vibration signal feature vector and the thermoelectric signal feature vector are concatenated to construct a multi-source heterogeneous data feature space for the excitation transformer.

[0014] As a preferred embodiment of the method for constructing the feature space of multi-source heterogeneous data of excitation transformers according to the present invention, the step of converting time-frequency domain information into a two-dimensional spectrum using the improved symmetric point pattern method includes:

[0015] The three-axis vibration signal matrix within the vibration signal window is mapped to the polar coordinate system. The polar coordinate radius and rotation angle of each axis vibration signal are calculated respectively, and a symmetrical point pattern that fuses the three-axis time-frequency information is generated.

[0016] The petal-shaped features of the symmetrical point pattern characterize the changes in harmonic frequency bands and amplitudes in the vibration signal.

[0017] As a preferred embodiment of the method for constructing a multi-source heterogeneous data feature space for an excitation transformer according to the present invention, the step of extracting spectral features using image feature descriptors to form a vibration signal feature vector includes:

[0018] Local feature points are identified from the two-dimensional map using a corner detection algorithm;

[0019] A direction is assigned to each identified local feature point. The direction is determined based on the local gradient histogram by calculating the gradient direction of the neighborhood region of the local feature point to determine the main direction.

[0020] A binary feature descriptor for each local feature point is generated using a feature descriptor extraction algorithm.

[0021] The binary feature descriptors of all local feature points in the two-dimensional spectrum are combined to form a feature vector representing the time-frequency characteristics of the vibration signal, thus completing the feature extraction of the excitation transformer vibration signal.

[0022] As a preferred embodiment of the method for constructing the feature space of multi-source heterogeneous data of an excitation transformer according to the present invention, the time-domain statistics of the extracted thermoelectric signal window include:

[0023] The temperature, voltage, and current signals within the window are preprocessed and normalized respectively.

[0024] Calculate the time-domain statistics of each signal within the window time, and use the calculated time-domain statistics as the initial feature set of the thermoelectric signal.

[0025] As a preferred embodiment of the method for constructing a feature space for multi-source heterogeneous data of an excitation transformer according to the present invention, the step of evaluating the feature importance of the time-domain statistics using the random forest algorithm includes:

[0026] A labeled training dataset is constructed based on the initial feature set;

[0027] Set the number of decision trees, the number of samples, and the number of features. Randomly select K samples with replacement from the training dataset to form a training subset. Use the training subset to train the decision trees and retain the out-of-bag sample set.

[0028] The model is evaluated using the out-of-bag sample set, and the classification accuracy is calculated for each iteration.

[0029] Based on the classification accuracy, each feature in the feature set is perturbed sequentially in each iteration, and the contribution of each feature in the feature set to the model performance is calculated by the change in accuracy before and after the perturbation, so as to obtain the importance score of each feature.

[0030] The features are ranked according to their importance scores, and the features that have the greatest impact on classification accuracy are selected. Then, the feature subset is obtained by progressively filtering based on the importance of the features that have the greatest impact on classification accuracy.

[0031] After evaluation and screening, the final importance score of each feature is output.

[0032] As a preferred embodiment of the method for constructing a multi-source heterogeneous data feature space for an excitation transformer according to the present invention, the method for forming a thermoelectric signal feature vector includes:

[0033] The coefficient of variation is introduced to characterize the relative volatility of features under different time windows;

[0034] The stable weight of each feature is calculated based on the coefficient of variation. The final importance score of each feature is multiplied by the stable weight of each feature to obtain the final feature score.

[0035] The features are ranked by importance according to the final feature score. Based on the importance ranking, a forward incremental search is performed on the feature subset, gradually increasing the feature dimension. A classifier is used to calculate the classification error under each dimension.

[0036] The feature dimension corresponding to the minimum classification error is selected as the optimal feature dimension, and the optimal feature is selected from the sorted features to form the thermoelectric signal feature vector.

[0037] As a preferred embodiment of the method for constructing the feature space of multi-source heterogeneous data of an excitation transformer according to the present invention, the method for dividing the sampling window includes:

[0038] The vibration signal in the operating parameter signal of the excitation transformer is used to extract short-term transient time-domain and frequency-domain information with the first interval time as the window length. The temperature and electrical signal in the operating parameter signal of the excitation transformer is used to extract the signal change trend and time-domain information over a long period of time with the second interval time as the window length.

[0039] By setting the same window sliding step size, synchronous alignment of multi-source heterogeneous signals on the time axis can be achieved.

[0040] Secondly, the present invention provides a system for constructing the feature space of multi-source heterogeneous data of excitation transformers, comprising:

[0041] The data acquisition module is used to collect the operating parameter signals of the excitation transformer and divide the sampling window;

[0042] The vibration signal processing module is used to convert the time-frequency domain information into a two-dimensional spectrum of the data within the vibration signal window using the improved symmetric point pattern method, and to extract the spectrum features using image feature descriptors to form a vibration signal feature vector.

[0043] The thermoelectric signal processing module is used to extract the time-domain statistics of the data within the thermoelectric signal window, evaluate the feature importance of the time-domain statistics using the random forest algorithm, and introduce stable weights to correct the feature scores; based on the corrected feature scores, the features are sorted and the dimensions are optimized, and the optimal feature subset is selected to form the thermoelectric signal feature vector.

[0044] The feature space construction module is used to concatenate the vibration signal feature vector with the thermoelectric signal feature vector to construct the multi-source heterogeneous data feature space of the excitation transformer.

[0045] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor, when executing the computer-executable instructions, implements the steps of a method for constructing a multi-source heterogeneous data feature space of an excitation transformer.

[0046] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of a method for constructing a multi-source heterogeneous data feature space for an excitation transformer.

[0047] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a method, system, device, and medium for constructing a multi-source heterogeneous data feature space for excitation transformers. By fusing features of multi-source heterogeneous signals, it constructs a more discriminative and stable feature space for excitation transformers, effectively improving the accuracy of equipment operation status assessment and fault diagnosis. Firstly, this invention employs differentiated feature extraction strategies for different types of signals, fully mining the key information of various signals. For vibration signals, an improved symmetric point pattern method is used to transform their complex time-frequency information into a two-dimensional spectrum, enabling the extraction of their triaxial features. For low-frequency signals such as temperature, voltage, and current, a time-domain statistical analysis combined with random forest sorting and stability correction strategies is used to achieve the screening and dimensional optimization of key features. Secondly, this invention adopts a refined time window division strategy, taking into account both the dynamic characteristics of signal changes and data processing efficiency, which is beneficial for constructing a high-timeliness, high-resolution monitoring model. This not only enhances the sensitivity and robustness of fault identification but also provides a universal technical solution for multi-source heterogeneous signal fusion and feature optimization. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of the overall process logic of a method for constructing the feature space of multi-source heterogeneous data of an excitation transformer, provided in an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the sampling window division for vibration signals in a method for constructing the feature space of multi-source heterogeneous data of an excitation transformer according to an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the sampling window division of the thermoelectric signal in the method for constructing the feature space of multi-source heterogeneous data of excitation transformers according to an embodiment of the present invention.

[0052] Figure 4 The diagram illustrates the principle of a method for extracting vibration signals from excitation transformers based on spatial mode fusion, which is an embodiment of the present invention for constructing a feature space for multi-source heterogeneous data of excitation transformers.

[0053] Figure 5 This is a schematic diagram illustrating the excitation transformer vibration signal feature extraction process based on spatial mode fusion, which is part of the method for constructing a feature space of multi-source heterogeneous data of excitation transformers according to an embodiment of the present invention.

[0054] Figure 6 This is a schematic diagram of the vibration spectrum of the excitation transformer under different operating conditions, which is a method for constructing the feature space of multi-source heterogeneous data of the excitation transformer according to an embodiment of the present invention.

[0055] Figure 7 This is a schematic diagram of the importance of time-domain statistics evaluated by the random forest algorithm in a method for constructing the feature space of multi-source heterogeneous data of an excitation transformer according to an embodiment of the present invention.

[0056] Figure 8 This is a schematic diagram illustrating the importance of time-domain statistics after introducing stable weights in the method for constructing the feature space of multi-source heterogeneous data of excitation transformers according to an embodiment of the present invention.

[0057] Figure 9 The graph shows the correlation between feature dimension and classification error in the feature space construction method for multi-source heterogeneous data of excitation transformers provided in an embodiment of the present invention.

[0058] Figure 10 This is a schematic diagram of the confusion matrix of the classification result of the method for constructing the feature space of multi-source heterogeneous data of excitation transformers according to an embodiment of the present invention. Detailed Implementation

[0059] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0060] Example 1, referring to Figures 1-5 As one embodiment of the present invention, a method for constructing the feature space of multi-source heterogeneous data of an excitation transformer is provided, such as... Figure 1 The specific steps shown are as follows:

[0061] S100: Acquires operating parameter signals of the excitation transformer and divides the sampling window;

[0062] S200: The improved symmetric point pattern method is used to convert the time-frequency domain information into a two-dimensional spectrum for the data within the vibration signal window, and image feature descriptors are used to extract spectrum features to form a vibration signal feature vector;

[0063] S300: Extract time-domain statistics of data within the thermoelectric signal window, use the random forest algorithm to evaluate the feature importance of the time-domain statistics, and introduce stable weights to correct the feature scores; perform feature sorting and dimension optimization based on the corrected feature scores, select the optimal feature subset, and form the thermoelectric signal feature vector.

[0064] S400: The vibration signal feature vector and the thermoelectric signal feature vector are concatenated to construct a multi-source heterogeneous data feature space for the excitation transformer.

[0065] It should be noted that, to address the challenges of multi-source heterogeneous signal data fusion in existing technologies, particularly how to effectively handle the differences in signal characteristics under different sampling rates, the complexity of feature space construction, and the insufficient stability and discriminative ability during feature selection, steps S100-S400 above, by fusing multi-source heterogeneous signal features, construct a more discriminative and stable excitation transformer feature space, effectively improving the accuracy of equipment operation status assessment and fault diagnosis. This invention employs differentiated feature extraction strategies for different types of signals, fully mining the key information of various signals. For vibration signals, an improved symmetric point pattern method is used to transform their complex time-frequency information into a two-dimensional spectrum, enabling the extraction of their triaxial features. For low-frequency signals such as temperature, voltage, and current, a time-domain statistical analysis combined with random forest sorting and stability correction strategies is used to achieve the screening and dimensional optimization of key features. A refined time window division strategy is adopted, taking into account both the dynamic characteristics of signal changes and data processing efficiency, which is beneficial for constructing a high-timeliness, high-resolution monitoring model. This not only enhances the sensitivity and robustness of fault identification but also provides a universal technical solution for multi-source heterogeneous signal fusion and feature optimization.

[0066] In this embodiment of the invention, step S100, which involves acquiring the operating parameter signals of the excitation transformer and dividing the sampling window, includes:

[0067] It should be noted that the temperature change of the excitation transformer over time is essentially a process of heat accumulation, which can be described by the following differential equation:

[0068] ,

[0069] Where C is the heat capacity, T(t) is the temperature, and P(t) is the heating power. The active power of the excitation transformer is P(t) = I. 2Let I(t)*R, where I(t) is the current, R is the load resistance, and Q(t) is the heat dissipation power. When heat dissipation is linear, we have Q(t) = T(t) / R. Substituting these values, we get:

[0070] ,

[0071] in, Let be the thermal time constant. It can be seen that the temperature response depends on the difference between the power and the current temperature, and the thermal time constant... Temperature changes are typically slow, lasting from tens of seconds to several minutes. Single-point values ​​and short-window sampling ignore this cumulative characteristic, so temperature requires sampling over a long window.

[0072] It should be noted that the electrical signals of the excitation transformer, such as power and current, are closely related to temperature, and their values ​​can often aid in the diagnosis of thermal faults in the transformer. Although the instantaneous fluctuations of the electrical signals are quite noticeable, the long-term average power is what determines the temperature rise, as shown in the following formula:

[0073] ,

[0074] in, Average power, The time interval is denoted by . As can be seen from the formula, current and power signals are the driving factors for temperature rise; therefore, it is also necessary to statistically analyze their average value or trend over a period of time.

[0075] It should be noted that the vibration of the excitation transformer can reflect its mechanical faults. Its vibration acceleration signal is generated by the Lorentz force and the magnetostrictive force of the iron core, both of which are correlated with twice the power frequency signal in the frequency band. Therefore, by judging its vibration acceleration spectrum, potential mechanical faults in the excitation transformer can be identified. The mechanical force generated by electromagnetic excitation produces vibration through the structure, and its spatial distribution characteristics are shown in Table 1. Therefore, for the vibration signal, it is necessary to extract its time-frequency domain information along the x, y, and z axes to construct its feature space. Vibration signals have significant transient and non-stationary characteristics. Too long a time window will lead to feature averaging, and abnormal information will be buried. Therefore, it is necessary to shorten the time window of the vibration signal as much as possible for feature space construction.

[0076] Table 1: Spatial distribution characteristics of excitation transformer vibration.

[0077]

[0078] In this embodiment of the invention, vibration, winding temperature, operating voltage, and current signals of the excitation transformer are first acquired. After obtaining the characteristics of the temperature, electrical, and vibration signals of the excitation transformer, a sliding window constructed based on these characteristics is set, such as... Figure 2 and Figure 3 As shown, the vibration signal uses a first time interval as the window length to extract short-term transient time-domain and frequency-domain information, while the thermoelectric signal uses a second time interval as the window length to extract signal change trends and time-domain information over a longer period. By setting the same window sliding step size, the synchronous alignment of multi-source heterogeneous signals on the time axis is achieved.

[0079] Specifically, in this embodiment, the first interval time is 1 second and the second interval time is 10 minutes; the window sliding step size is 1 second to achieve synchronization of multi-source heterogeneous signals.

[0080] It should be noted that step S100 above, by reasonably dividing the sampling window and using windows of different time lengths and a unified sliding step size for vibration signals and thermoelectric signals respectively, effectively combines the transient characteristics of vibration signals and the slow-changing characteristics of thermoelectric signals, ensuring the synchronization and comparability of multi-source signals on the time scale, and improving the timeliness of feature construction and the ability to respond to different operating states.

[0081] In this embodiment of the invention, step S200 includes the following sub-steps B1 and B2:

[0082] In B1: The improved symmetric point pattern method is used to convert the time-frequency domain information into a two-dimensional spectrum for the data within the vibration signal window; the specific steps include:

[0083] The three-axis vibration signal matrix within the vibration signal window is mapped to the polar coordinate system. The polar coordinate radius and rotation angle of each axis vibration signal are calculated respectively, and a symmetrical point pattern that fuses the three-axis time-frequency information is generated.

[0084] The petal-shaped features of symmetrical point patterns are used to characterize the changes in harmonic frequency bands and amplitudes in vibration signals.

[0085] In this embodiment of the invention, taking the triaxial time series matrix V of the vibration signal as an example, it can be expressed as:

[0086] ,

[0087] Among them, v x v y v z These are the vibration time-domain signals on the x, y, and z axes, respectively, where n is the length of the vibration signal;

[0088] Furthermore, the triaxial vibration data matrix V can be mapped to polar coordinates, as expressed by the formula:

[0089] ,

[0090] ,

[0091] ,

[0092] Where r(i,j) is the polar coordinate radius of the i-th vibration signal on the j-th axis, V ij Let V be the i-th sampling point of the vibration signal on the j-th axis, where j=[1,2,3]. jmax Let V be the maximum amplitude of the j-th axis of matrix V. jmin Let be the minimum amplitude of the j-th axis of matrix V; δ(i,j) is the counterclockwise rotation angle of the mirror symmetry plane of the j-th column of matrix V, where k represents the rotation direction, k∈{0,1}, V (i+t)j Let η be the value of the i-th vibration signal on the j-th axis of matrix V after being changed by the time delay offset t, and let η be the angle gain factor. max V represents the maximum amplitude in matrix V. min Let λ be the minimum amplitude in matrix V, and λ(i,j) be the clockwise rotation angle of the mirror symmetry plane of the j-th column of matrix V.

[0093] Furthermore, such as Figure 4 The diagram illustrates the principle of a feature extraction method for excitation transformer vibration signals based on spatial mode fusion. Given a sampling frequency, the frequency domain information of the signal can be obtained by observing the petal shape characteristics. When multiple harmonics are introduced into the signal, the spectrum no longer presents a standard petal structure, and the internal features of the petals also change significantly. The larger the harmonic component, the greater its impact on the internal structural features of the petals, and the more severe the petal distortion. Changes in harmonic frequencies also affect the petal shape. Therefore, the frequency bands and amplitude changes of harmonics in the signal can be determined by examining the converted spectrum.

[0094] In B2: Image feature descriptors are used to extract spectral features and form a vibration signal feature vector; specifically, this includes the following sub-steps B21~B24:

[0095] In B21: Local feature points are identified from a two-dimensional map using a corner detection algorithm;

[0096] It should be noted that corner detection algorithms are based on the gray-level differences of pixels within a local area to quickly detect corners in an image. These corners are usually areas with a large amount of information in the image, representing points with important features in the time-frequency spectrum.

[0097] In B22: an orientation is assigned to each identified local feature point. The orientation is determined based on the local gradient histogram, and the main orientation is determined by calculating the gradient orientation of the neighborhood region of the local feature point.

[0098] Specifically, let N be the neighborhood region of a certain local feature point. k Its gradient direction can be expressed as:

[0099] ,

[0100] Where, θ k It is the direction of the local feature point k, θ p It is the gradient direction of the neighboring pixel p, N k It is the pixel region surrounding the local feature point k, and θ represents the principal direction angle.

[0101] In B23: A binary feature descriptor is generated for each local feature point using a feature descriptor extraction algorithm;

[0102] It should be noted that the feature descriptor extraction algorithm selects pixel pairs in the neighborhood of a local feature point and generates a binary string based on their grayscale differences to represent the local features of that local feature point.

[0103] In B24: such as Figure 5 As shown, the binary feature descriptors of all local feature points in the two-dimensional spectrum are combined to form a feature vector representing the time-frequency characteristics of the vibration signal, thus completing the feature extraction of the excitation transformer vibration signal.

[0104] It should be noted that the above step S200 can fully preserve the multidimensional characteristics of the vibration signal in the spatial and frequency domains, effectively overcome the feature extraction difficulties caused by the non-stationary and transient characteristics of the vibration signal, and significantly enhance the expressive power and distinguishability of mechanical fault features.

[0105] In this embodiment of the invention, step S300 includes the following sub-steps C1 to C3:

[0106] In C1: Extract the time-domain statistics of the data within the thermoelectric signal window;

[0107] Specifically, after windowing and signal preprocessing, various time-domain statistics are extracted within the window time of the thermoelectric signal, as shown in Table 2.

[0108] Table 2: Types of statistical quantities of time-domain characteristics of thermoelectric signals.

[0109]

[0110] Specifically, after completing the data preprocessing of the thermoelectric signal within the window time, the normalized initial feature set D={(x1,y1),(x2,y2),…,(x N ,y N )}, where x N Let y be a time-domain statistical value. N This is a label for the operating status of the excitation transformer.

[0111] In C2: The random forest algorithm is used to evaluate the feature importance of time-domain statistics, and stable weights are introduced to correct the feature scores; the specific steps include:

[0112] Construct a labeled training dataset based on the initial feature set;

[0113] Set the number of decision trees, the number of samples, and the number of features. Randomly select K samples with replacement from the training dataset to form a training subset. Use the training subset to train the decision trees and retain the out-of-bag sample set.

[0114] The model is evaluated using an out-of-bag sample set, and the classification accuracy is calculated for each iteration.

[0115] Based on classification accuracy, each feature in the feature set is perturbed sequentially in each iteration. The contribution of each feature in the feature set to the model performance is calculated by the change in accuracy before and after the perturbation, so as to obtain the importance score of each feature.

[0116] The features are ranked according to their importance scores. The features that have the greatest impact on classification accuracy are selected, and then further filtered according to their importance to obtain a feature subset.

[0117] After evaluation and screening, the final importance score of each feature is output.

[0118] It should be noted that the evaluation of the importance of time-domain statistics for different signals based on the random forest method is accomplished by ranking the importance of these statistics. The out-of-bag (OOB) error method is used as the evaluation tool for the importance of random forest features. During training, the samples not selected for the training set are called OOB samples. After the random forest classification tree model is trained, the OOB samples are used as test samples, and the classification accuracy of the OOB samples is evaluated based on the test results. In each iteration, perturbations are applied to different features in the OOB samples, and the difference in classification accuracy before and after the perturbation is compared to rank the feature importance. The importance is calculated as follows:

[0119] ,

[0120] Among them, S Fj For feature F j Importance, where K is the number of training samples, P k OOB P k,j OOB These represent the classification accuracy before and after applying the perturbation, respectively.

[0121] It should be noted that after the random forest classification tree model is trained, OOB samples are used as test samples and input into each trained decision tree for prediction. The prediction results of all trees are summarized and the final classification label of each OOB sample is determined by voting or averaging. Then, it is compared with the true label, and the proportion of correctly classified samples to the total number of OOB samples is counted. In this way, the overall classification accuracy of OOB samples is calculated and evaluated.

[0122] In C3: Based on the corrected feature scores, feature sorting and dimensionality optimization are performed to select the optimal feature subset and form the thermoelectric signal feature vector;

[0123] Specifically, the coefficient of variation is introduced to characterize the relative volatility of features across different time windows; the coefficient of variation (CV) j The following formula can be used for calculation:

[0124] ,

[0125] Where, σ j For feature F j The standard deviation, μ, under different window times j The value is the corresponding average; the smaller the coefficient of variation, the more stable the feature is over a long period of time.

[0126] Furthermore, the stable weight K for each feature is calculated based on the coefficient of variation. j :

[0127] ,

[0128] Furthermore, the final importance score of each feature is multiplied by the stable weight of each feature to obtain the final feature score I. Fj for:

[0129] ,

[0130] Specifically, features are ranked by importance according to their final feature scores. Based on the importance ranking, a forward incremental search is performed on the feature subsets, gradually increasing the feature dimensions. A classifier is used to calculate the classification error under each dimension. The feature dimension corresponding to the minimum classification error is selected as the optimal feature dimension, and the optimal feature is selected from the ranked features to form the thermoelectric signal feature vector.

[0131] It should be noted that step S300 effectively improves the representativeness and stability of thermoelectric signal features, effectively suppresses the interference of feature fluctuations on model performance, selects key features with strong discriminative power and high stability, and improves the efficiency and generalization ability of subsequent state recognition models.

[0132] In this embodiment of the invention, step S400, which involves concatenating the vibration signal feature vector with the thermoelectric signal feature vector to construct the multi-source heterogeneous data feature space of the excitation transformer, includes:

[0133] The vibration signal feature vector and the thermoelectric signal feature vector are concatenated sequentially to form a high-dimensional feature vector. This concatenated feature vector contains multimodal information from both the vibration and electrical signals, representing the multidimensional state of the excitation transformer. The formula is as follows:

[0134] ,

[0135] Among them, F v F represents the vibration signal features formed after image feature descriptor processing. d The feature vector F represents the thermoelectric signal features formed by optimizing the feature dimensions; the final concatenated feature vector is F. final It contains all the information of the vibration signal and temperature signal after feature extraction.

[0136] It should be noted that step S400 above concatenates the feature vectors of vibration signals and thermoelectric signals to construct a unified multi-source heterogeneous data feature space. This enables feature-level fusion of signals with different physical meanings and sampling rates, fully leveraging the complementary advantages of multi-source information to form a more comprehensive and discriminative state representation. This provides a reliable data foundation for the comprehensive state assessment and accurate fault diagnosis of excitation transformers.

[0137] Example 2, refer to Figures 6-10 Based on the previous embodiment, this embodiment provides an application example of the method for constructing the feature space of multi-source heterogeneous data of excitation transformers, in order to verify and illustrate the technical effects adopted in this method.

[0138] The vibration spectrum characteristics constructed in this embodiment under different operating conditions are as follows: Figure 6 As shown, NO, WL, CL, CG, CS, and WE represent six operating states: normal operation, loose winding, loose core, multi-point grounding of the core, inter-turn short circuit, and winding eccentricity, respectively; S1-S5 represent five operating conditions: no-load rated voltage, no-load 90% rated voltage, no-load 80% rated voltage, load rate 78% rated voltage, and load rate 82% rated voltage, respectively.

[0139] This embodiment uses random forest to evaluate feature importance, and the results are as follows: Figure 7 As shown, the feature importance score after introducing stable weights is as follows: Figure 8 As shown; after sorting the time-domain statistics according to feature importance, a support vector machine classifier is used to optimize the feature dimension, resulting in the correlation curve between feature dimension and classification error, as shown. Figure 9As shown, the time step with the lowest classification error is selected as the optimal feature dimension F. e In this embodiment, the feature dimension is 54, and the classification error reaches a minimum of 11.3%. When the feature dimension is less than 54, the classification error gradually increases due to the neglect of effective features; while when the feature dimension is greater than 54, feature redundancy leads to a decrease in model performance, and the classification error also increases slightly. After obtaining the vibration signal feature F... v With thermoelectric signal characteristics F e The two are then concatenated to obtain the fused feature vector F=F v ⊕F e This means completing the construction of the feature space of multi-source heterogeneous data of the excitation transformer.

[0140] Using support vector machines as the classification model, the fault diagnosis capability of the feature vectors of the proposed method was tested. The test set data included data on six operating states: normal operation, loose winding, inter-turn short circuit, loose core, winding eccentricity, and multi-point grounding of the core. The classification confusion matrix results are as follows: Figure 10 As shown, the proposed feature space construction method can effectively identify the above-mentioned operating states, with an average accuracy of 93.8%.

[0141] Therefore, the method provided by this invention, by fusing features of multi-source heterogeneous signals, constructs a more discriminative and stable feature space for excitation transformers, effectively improving the accuracy of equipment operation status assessment and fault diagnosis. First, differentiated feature extraction strategies are adopted for different types of signals to fully explore the key information of various signals. For vibration signals, an improved symmetric point pattern method is used to transform their complex time-frequency information into a two-dimensional spectrum, enabling the extraction of their triaxial features. For low-frequency signals such as temperature, voltage, and current, time-domain statistical analysis combined with random forest sorting and stability correction strategies is used to achieve the screening and dimensional optimization of key features. Second, a refined time window division strategy is adopted, taking into account both the dynamic characteristics of signal changes and data processing efficiency, which is conducive to building a high-timeliness, high-resolution monitoring model. This not only enhances the sensitivity and robustness of fault identification but also provides a universal technical solution for multi-source heterogeneous signal fusion and feature optimization.

[0142] Example 3: This example provides a system for constructing the feature space of multi-source heterogeneous data of excitation transformers, including:

[0143] The data acquisition module is used to collect the operating parameter signals of the excitation transformer and divide the sampling window;

[0144] The vibration signal processing module is used to convert the time-frequency domain information into a two-dimensional spectrum of the data within the vibration signal window using the improved symmetric point pattern method, and to extract the spectrum features using image feature descriptors to form a vibration signal feature vector.

[0145] The thermoelectric signal processing module is used to extract the time-domain statistics of the data within the thermoelectric signal window, use the random forest algorithm to evaluate the feature importance of the time-domain statistics, and introduce stable weights to correct the feature scores; based on the corrected feature scores, the features are sorted and the dimensions are optimized, and the optimal feature subset is selected to form the thermoelectric signal feature vector.

[0146] The feature space construction module is used to concatenate the feature vectors of vibration signals and thermoelectric signals to construct the feature space of multi-source heterogeneous data of the excitation transformer.

[0147] It should be noted that the technical solution of the excitation transformer multi-source heterogeneous data feature space construction system is based on the same concept as the above-mentioned excitation transformer multi-source heterogeneous data feature space construction method. For details not described in detail in the technical solution of the excitation transformer multi-source heterogeneous data feature space construction system in this embodiment, please refer to the description of the above-mentioned excitation transformer multi-source heterogeneous data feature space construction method.

[0148] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0149] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for constructing a multi-source heterogeneous data feature space for an excitation transformer. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0150] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0151] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0152] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory, random access memory, flash memory, hard disk, or optical disk, and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0153] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing a feature space of a multi-source heterogeneous data of an excitation transformer, characterized in that, include: Collect the operating parameter signals of the excitation transformer and divide the sampling window; The improved symmetric point pattern method is used to convert the time-frequency domain information into a two-dimensional spectrum of the data within the vibration signal window, and image feature descriptors are used to extract the spectrum features to form a vibration signal feature vector. The time-domain statistics of the data within the thermoelectric signal window are extracted, and the random forest algorithm is used to evaluate the feature importance of the time-domain statistics. Stable weights are introduced to correct the feature scores. Based on the corrected feature scores, feature sorting and dimensionality optimization are performed to select the optimal feature subset and form the thermoelectric signal feature vector. The vibration signal feature vector and the thermoelectric signal feature vector are concatenated to construct a multi-source heterogeneous data feature space for the excitation transformer; The method of converting time-frequency domain information into a two-dimensional spectrum using the improved symmetric point pattern method includes: The three-axis vibration signal matrix within the vibration signal window is mapped to the polar coordinate system. The polar coordinate radius and rotation angle of each axis vibration signal are calculated respectively, and a symmetrical point pattern that fuses the three-axis time-frequency information is generated. The petal shape characteristics of the symmetrical point pattern are used to characterize the changes in harmonic frequency bands and amplitudes in the vibration signal; The step of extracting spectral features using image feature descriptors to form a vibration signal feature vector includes: Local feature points are identified from the two-dimensional map using a corner detection algorithm; A direction is assigned to each identified local feature point. The direction is determined based on the local gradient histogram by calculating the gradient direction of the neighborhood region of the local feature point to determine the main direction. A binary feature descriptor for each local feature point is generated using a feature descriptor extraction algorithm. The binary feature descriptors of all local feature points in the two-dimensional spectrum are combined to form a feature vector representing the time-frequency characteristics of the vibration signal, thus completing the feature extraction of the excitation transformer vibration signal. The formation of the thermoelectric signal feature vector includes: The coefficient of variation is introduced to characterize the relative volatility of features under different time windows; The stable weight of each feature is calculated based on the coefficient of variation. The final importance score of each feature is multiplied by the stable weight of each feature to obtain the final feature score. The features are ranked by importance according to the final feature score. Based on the importance ranking, a forward incremental search is performed on the feature subset, gradually increasing the feature dimension. A classifier is used to calculate the classification error under each dimension. The feature dimension corresponding to the minimum classification error is selected as the optimal feature dimension, and the optimal feature is selected from the sorted features to form the thermoelectric signal feature vector. The segmentation sampling window includes: The vibration signal in the operating parameter signal of the excitation transformer is used to extract short-term transient time-domain and frequency-domain information with the first interval time as the window length. The temperature and electrical signal in the operating parameter signal of the excitation transformer is used to extract the signal change trend and time-domain information over a long period of time with the second interval time as the window length. By setting the same window sliding step size, synchronous alignment of multi-source heterogeneous signals on the time axis can be achieved.

2. The method of claim 1, wherein the method comprises: The time-domain statistics of the data extracted within the thermoelectric signal window include: The temperature, voltage, and current signals within the window are preprocessed and normalized respectively. Calculate the time-domain statistics of each signal within the window time, and use the calculated time-domain statistics as the initial feature set of the thermoelectric signal.

3. The method of claim 2, wherein, The evaluation of the feature importance of the time-domain statistics using the random forest algorithm includes: A labeled training dataset is constructed based on the initial feature set; Set the number of decision trees, the number of samples, and the number of features. Randomly select K samples with replacement from the training dataset to form a training subset. Use the training subset to train the decision trees and retain the out-of-bag sample set. The model is evaluated using the out-of-bag sample set, and the classification accuracy is calculated for each iteration. Based on the classification accuracy, each feature in the feature set is perturbed sequentially in each iteration, and the contribution of each feature in the feature set to the model performance is calculated by the change in accuracy before and after the perturbation, so as to obtain the importance score of each feature. The features are ranked according to their importance scores, and the features that have the greatest impact on classification accuracy are selected. Then, the feature subset is obtained by progressively filtering based on the importance of the features that have the greatest impact on classification accuracy. After evaluation and screening, the final importance score of each feature is output.

4. A system for constructing a field transformer multi-source heterogeneous data feature space, applying the method for constructing a field transformer multi-source heterogeneous data feature space according to any one of claims 1-3, characterized in that, include: The data acquisition module is used to collect the operating parameter signals of the excitation transformer and divide the sampling window; The vibration signal processing module is used to convert the time-frequency domain information into a two-dimensional spectrum of the data within the vibration signal window using the improved symmetric point pattern method, and to extract the spectrum features using image feature descriptors to form a vibration signal feature vector. The thermoelectric signal processing module is used to extract the time-domain statistics of the data within the thermoelectric signal window, evaluate the feature importance of the time-domain statistics using the random forest algorithm, and introduce stable weights to correct the feature scores; based on the corrected feature scores, the features are sorted and the dimensions are optimized, and the optimal feature subset is selected to form the thermoelectric signal feature vector. The feature space construction module is used to concatenate the vibration signal feature vector with the thermoelectric signal feature vector to construct the multi-source heterogeneous data feature space of the excitation transformer. 5.An electronic device comprising a memory and a processor, the electronic device characterized by: The memory is used to store computer-executable instructions, and when the processor executes the computer-executable instructions, it implements the steps of the method for constructing the feature space of multi-source heterogeneous data of an excitation transformer as described in any one of claims 1 to 3.

6. A computer-readable storage medium having stored thereon computer- executable instructions, the computer-executable instructions comprising instructions for: When the computer-executable instructions are executed by the processor, they implement the steps of the method for constructing the feature space of multi-source heterogeneous data of an excitation transformer as described in any one of claims 1 to 3.

Images