Multi-view feature fusion transformer sound signal recognition method and system and medium

Abstract

The application discloses a transformer sound signal recognition method and system based on multi-view feature fusion and a medium. The method comprises the following steps: extracting features of transformer sound signals from multiple dimensions, including time domain, frequency domain and nonlinear domain; generating effective features in four dimensions of monotonicity, robustness, trend and distinguishability according to the extracted features; fusing the generated effective features to obtain a fusion feature vector; and classifying based on the fusion feature vector by using a machine learning model to obtain the running state of the transformer. The application can effectively represent the running state of the transformer by multi-view feature extraction, selection and fusion, and can realize accurate recognition of the running state of the transformer by parameter optimization support vector machine.

Description

Technical Field

[0001] This invention relates to the field of transformer acoustic signal recognition technology, specifically to a transformer acoustic signal recognition method, system, and medium that integrates multi-view feature fusion. Background Technology

[0002] Transformers are key equipment in power systems, and their operating status directly affects the reliability and security of the power grid. However, transformers operate in harsh environments and suffer from complex wiring structures and severe electromagnetic interference, making transformer maintenance difficult. During long-term operation, transformers are subjected to voltage surges, thermal shocks, and mechanical vibrations, leading to frequent failures. Therefore, intelligent diagnosis of transformer operating status is crucial for maintaining power grid safety, reducing operation and maintenance costs, and ensuring people's livelihoods. Acoustic signal analysis is one of the most effective methods for diagnosing transformer operating status. In recent years, scholars both domestically and internationally have conducted extensive research on transformer condition identification based on acoustic signal analysis. However, most existing analysis methods are based on the time / frequency domain characteristics of signals, relying on expert experience and suffering from low diagnostic efficiency and accuracy, making it difficult to meet the needs of intelligent operation and maintenance of power equipment in the era of big data. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method, system and medium for transformer acoustic signal recognition based on multi-view feature fusion, which addresses the above-mentioned problems in the prior art. The present invention extracts, selects and fuses multi-view features to construct a multi-dimensional feature set that can effectively characterize the operating state of the transformer, and achieves accurate identification of the transformer operating state through parameter-optimized support vector machine.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A method for identifying transformer acoustic signals through multi-view feature fusion includes:

[0006] S101 extracts the features of transformer acoustic signals from multiple dimensions, including time domain, frequency domain, and nonlinear domain.

[0007] S102, based on the extracted features, generate effective features in four dimensions: monotonicity, robustness, trend and identifiability;

[0008] S103, fuse the generated effective features to obtain a fused feature vector;

[0009] S104 uses a machine learning model based on fused feature vectors to classify and obtain the operating status of the transformer.

[0010] Optionally, the features of the transformer acoustic signal extracted from the time domain perspective in step S101 include some or all of the peak value, peak-to-peak value, average value, absolute average value, root mean square value, standard deviation, slope, and kurtosis of the transformer acoustic signal; the features of the transformer acoustic signal extracted from the frequency domain perspective include some or all of the average amplitude, center frequency, root mean square frequency, and standard deviation of the frequency of the transformer acoustic signal; the features of the transformer acoustic signal extracted from the nonlinear domain perspective include at least one of the information entropy and fractal dimension of the transformer acoustic signal.

[0011] Optionally, the calculation function expression for the effective features of the monotonicity dimension in step S102 is:

[0012]

[0013] In the above formula, Mon(F) represents the effective feature of the feature sequence F in the monotonicity dimension, K is the sequence length of the feature sequence F, and No{·} represents the counting function.

[0014] Optionally, the calculation function expression for the effective features of the robustness dimension in step S102 is:

[0015]

[0016] In the above formula, Rob(F) represents the effective feature of the feature sequence F in the robustness dimension, K is the sequence length of the feature sequence F, and f(t) k The k-th feature in the sequence F with features ).

[0017] Optionally, the calculation function expression for the effective features of the trend dimension in step S102 is:

[0018]

[0019] In the above formula, Trer(F,T) represents the effective features of the feature sequence F and its corresponding sampling time series T in the trend dimension, and f(t) k ) is the k-th feature in the sequence F, t k The k-th feature f(t) in the sequence F is a feature. k The sampling time of ) Let F be the average of all features in the sequence F. Let T be the average value of all times in the sampled time series T, and K be the sequence length of the feature sequence F.

[0020] Optionally, the calculation function expression for the effective features of the identifiability dimension in step S102 is:

[0021]

[0022] In the above formula, Ide(F,C) represents the effective features of the feature sequence F and its corresponding class label sequence C in the discriminability dimension, f(t) k The k-th feature in the sequence F is denoted as . c(t) is the average of all features in the sequence F. k The k-th feature f(t) in the sequence F with feature ) k Category tags, K is the average of all category labels in the sampled time series T, and K is the sequence length of the feature sequence F.

[0023] Optionally, step S103 includes:

[0024] S201, Input the generated effective feature vector, and denote any i-th effective feature as x. i ∈R n ,i=1,2,…,m, where m is the number of effective features generated, R n The dimension representing the effective feature is a real number n;

[0025] S202, determine the kernel function and the parameters in the kernel function, and obtain any i-th element K in the kernel matrix K. i for:

[0026] K i =k(x i ,x j ), i,j=1,2,…,m,

[0027] In the above formula, k is a defined kernel function, and x i ,x j These are the i-th and j-th valid features, respectively;

[0028] S203, center the kernel matrix K according to the following formula:

[0029] K = K-1 m K-K1 m +1 m K1 m ,

[0030] In the above formula, 1 m This represents a matrix where all elements are 1 / m, and m is the number of valid features generated.

[0031] S204, Solve for the first p eigenvalues ​​of the centered kernel matrix K: λ1≥λ2≥…≥λ p With the first p eigenvectors α 1 ,α 2 ,…,α p And <α i ,αi >=1, i=1,2,…,p, where <,> denotes inner product operation;

[0032] S205, for any test point x∈R n ,have Where Φ(x) represents the fused features for test point x, and k(x) i (x) represents the kernel function pair (x) i The calculation result of ,x).

[0033] Optionally, in step S104, when using a machine learning model to classify based on the fused feature vectors to obtain the transformer's operating state, the machine learning model used is a support vector machine. The use of the support vector machine to classify to obtain the transformer's operating state refers to solving the optimization problem shown in the following equation by selecting a kernel function k(·) and a penalty parameter C:

[0034]

[0035] Furthermore, the classification decision function f(x) used to solve the optimization problem is:

[0036]

[0037] Where N is the number of training samples, α i and α j The eigenvectors x are respectively i and x j The corresponding penalty coefficient, y i and y j The eigenvectors x are respectively i and x j The tag, To and x j The kernel function calculation results, For x i transpose of x i ,x j These are the i-th and j-th valid features, respectively; C is the penalty parameter; sign is the sign function; K(x·x) i For x and x i The kernel function calculation result is given, where x is the feature vector after fusion of the test samples, and b is the bias parameter.

[0038] Furthermore, the present invention also provides a transformer acoustic signal recognition system with multi-view feature fusion, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the transformer acoustic signal recognition method with multi-view feature fusion.

[0039] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program for being programmed or configured by a microprocessor to perform the transformer acoustic signal recognition method of multi-view feature fusion.

[0040] Compared with existing technologies, the present invention has the following advantages: The present invention includes extracting features of transformer acoustic signals from multiple dimensions, including time domain, frequency domain, and nonlinear domain; generating effective features in four dimensions—monotonicity, robustness, trend, and identifiability—based on the extracted features; fusing the generated effective features to obtain a fused feature vector; and using a machine learning model to classify the transformer's operating status based on the fused feature vector. The present invention constructs a multi-dimensional feature set that can effectively characterize the transformer's operating status through multi-perspective feature extraction, selection, and fusion, and achieves accurate identification of the transformer's operating status through parameter-optimized support vector machines. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0042] Figure 2 This is a schematic diagram illustrating the intelligent recognition principle of the method in an embodiment of the present invention.

[0043] Figure 3 This is the time-domain waveform of the acoustic signal collected in an embodiment of the present invention.

[0044] Figure 4 This is the acoustic signal spectrum in an embodiment of the present invention.

[0045] Figure 5 This is the distribution of feature parameters extracted in the embodiments of the present invention.

[0046] Figure 6 This is the result of the support vector machine grid parameter optimization in an embodiment of the present invention.

[0047] Figure 7 This refers to the accuracy of intelligent identification of transformer acoustic signals in this embodiment of the invention. Detailed Implementation

[0048] like Figure 1 As shown, the transformer acoustic signal recognition method based on multi-view feature fusion in this embodiment includes:

[0049] S101 extracts the features of transformer acoustic signals from multiple dimensions, including time domain, frequency domain, and nonlinear domain.

[0050] S102, based on the extracted features, generate effective features in four dimensions: monotonicity, robustness, trend and identifiability;

[0051] S103, fuse the generated effective features to obtain a fused feature vector;

[0052] S104 uses a machine learning model based on fused feature vectors to classify and obtain the operating status of the transformer.

[0053] See Figure 2 In this embodiment, the machine learning model can be divided into a training phase and a usage phase. During the training phase, steps S102 and S103 are executed using transformer acoustic signal samples to extract fused feature vectors. Then, data labeling is performed to attach labels corresponding to the transformer's operating state, and training and testing sets are established. The training set is used to optimize the parameters of the machine learning model (training), and the testing set is used for testing. If the test passes, the trained machine learning model is obtained (if the test fails, training needs to be repeated). Based on this, the transformer acoustic signal under test can be obtained, and steps S102 and S103 are executed to extract the fused feature vectors. Finally, the trained machine learning model is input to obtain the transformer's operating state. It should be noted that... Figure 2 The paper uses an SVM (Support Vector Machine) recognition model as the machine learning model. However, inspired by the clear mapping relationship between input and output in this paper, other machine learning models can also be used as needed to achieve the same mapping relationship between input and output for training and application. Therefore, it will not be described in detail here.

[0054] The features of the transformer acoustic signal extracted from the time domain perspective in step S101 include some or all of the peak value, peak-to-peak value, average value, absolute average value, root mean square value, standard deviation, slope, and kurtosis of the transformer acoustic signal; the features of the transformer acoustic signal extracted from the frequency domain perspective include some or all of the average amplitude, center frequency, root mean square frequency, and standard deviation of the frequency of the transformer acoustic signal; the features of the transformer acoustic signal extracted from the nonlinear domain perspective include at least one of the information entropy and fractal dimension of the transformer acoustic signal.

[0055] In this embodiment, the features of the transformer acoustic signal extracted from the time domain perspective include the peak value, peak-to-peak value, average value, absolute average value, root mean square value, standard deviation, slope, and kurtosis of the transformer acoustic signal. The specific calculation method is as follows: For signal {x} i}, i = 1, 2, ..., N, where N is the length of the time series and its peak value V max This refers to the maximum value of the signal amplitude. It is a time-instability parameter with large fluctuations, and is therefore often used to detect impulse components. Its expression is:

[0056] V max =max(x i ),

[0057] Peak-to-peak value Vp-p It is the difference between the maximum and minimum values ​​of the signal, expressed as:

[0058] V p-p =V max -V min ,

[0059] In the above formula, V min It is the minimum value of the signal (compared to the peak value V). max (corresponding to), its expression is:

[0060] V min =min(x i ).

[0061] It can be observed that the peak-to-peak value V p-p The magnitude of the signal is independent of the DC component of the signal and is mainly used to describe the range of signal changes. It can be used to monitor changes in signal strength.

[0062] signal {x i The average value V c This represents the center value around which the signal revolves, exhibiting steady-state characteristics, and its expression is:

[0063]

[0064] In the above formula, That is, the signal {x} i The average of N signals from a time series {x}, where N is the length of the time series. i The average value V c To reflect the magnitude of data trends and the DC component in the signal, the mean should generally be removed from the data during calculation, leaving the dynamic part that is useful for diagnosis.

[0065] absolute average It refers to the average of the absolute values ​​of the signal, and its expression is:

[0066]

[0067] Root mean square value V rms Reflecting the total energy of the signal, it is not sensitive to early faults but has good stability, and shows a monotonically increasing trend as the fault develops. Its expression is:

[0068]

[0069] The standard deviation σ represents the dispersion of the signal, and its expression is:

[0070]

[0071] The slope α reflects the asymmetry of the amplitude probability density function p(x) of the signal with respect to the ordinate. The greater the asymmetry, the larger the slope. Its expression is:

[0072]

[0073] Kurtosis β represents the degree to which a signal deviates from a normal distribution. As a fault occurs, values ​​such as the root mean square (RMS) and absolute mean increase. However, kurtosis is more sensitive to amplitude fluctuations because it operates on the fourth power of the amplitude, effectively suppressing low amplitudes and highlighting high amplitudes. This makes it highly effective for detecting faults with impulse components. Its expression is:

[0074]

[0075] In this embodiment, the features of the transformer acoustic signal extracted from the frequency domain perspective include the average amplitude, center frequency, root mean square frequency, and standard deviation of the frequency of the transformer acoustic signal. Assuming a time-domain signal x = {x...} i}, i=1,2,…,N, are converted into a frequency domain signal X={X} using Fourier transform. j If j = 1, 2, ..., N, then the calculation method for the above frequency domain characteristics is as follows:

[0076] The expression for the average amplitude F1 is:

[0077]

[0078] The expression for the center frequency F2 is:

[0079]

[0080] In the above formula, f(j) represents the frequency of the j-th point in the spectrum;

[0081] The expression for the root mean square frequency F3 is:

[0082]

[0083] The expression for the frequency standard deviation F4 is:

[0084]

[0085] In the above formula, F2 represents the center frequency.

[0086] In this embodiment, the features of the transformer acoustic signal extracted from the perspective of the nonlinear domain dimension include the information entropy and fractal dimension of the transformer acoustic signal, and the calculation method is as follows:

[0087] Information entropy describes the uncertainty of information. For acoustic signals, the more regular they are, the lower their uncertainty and the smaller their information entropy; conversely, the higher the uncertainty, the larger their information entropy. For a time-domain signal x = {x...} i For each i = 1, 2, ..., N, the expression for calculating its temporal information entropy F5 is:

[0088] And there are

[0089] Where p(x) i ) represents x i The probability of occurrence.

[0090] Fractal dimension can describe complex and irregular geometric objects that traditional Euclidean geometry cannot describe. Specifically, it is defined as follows: Suppose Y is a non-empty subset of the real plane set, and z is a closed set of scale ε in the real plane set. If covering set Y requires N(ε) closed sets z, then the fractal dimension D of set Y can be defined as:

[0091]

[0092] This embodiment selects effective features from the feature set that reflect the transformer's operating state from four dimensions: monotonicity, robustness, trend, and identifiability; assuming a feature value sequence F = [f(t1), f(t2), ..., f(t... k The time series T = [t1, t2, ..., t] is given. k ], where f(t) k ) for t k The feature values ​​at time points, k is the total length of the time series, and the category label sequence is C = [c(t1), c(t2), ..., c(t...]. k Let monotonicity be denoted as Mon(F), robustness as Rob(F), trend as Tre(F,T), and identifiability as Ide(F,C), then we have:

[0093] In this embodiment, the calculation function expression for the effective features of the monotonicity dimension in step S102 is:

[0094]

[0095] In the above formula, Mon(F) represents the effective feature of the feature sequence F in the monotonicity dimension, K is the sequence length of the feature sequence F, and No{·} represents the counting function.

[0096] In this embodiment, the calculation function expression for the effective features of the robustness dimension in step S102 is as follows:

[0097]

[0098] In the above formula, Rob(F) represents the effective feature of the feature sequence F in the robustness dimension, K is the sequence length of the feature sequence F, and f(t) k The k-th feature in the sequence F with features ).

[0099] In this embodiment, the calculation function expression for the effective features of the trend dimension in step S102 is as follows:

[0100]

[0101] In the above formula, Tre(F,T) represents the effective features of the feature sequence F and its corresponding sampling time series T in the trend dimension, and f(t) k ) is the k-th feature in the sequence F, t k The k-th feature f(t) in the sequence F is a feature. k The sampling time of ) Let F be the average of all features in the sequence F. Let T be the average value of all times in the sampled time series T, and K be the sequence length of the feature sequence F.

[0102] In this embodiment, the calculation function expression for the effective features of the identifiability dimension in step S102 is as follows:

[0103]

[0104] In the above formula, Ide(F,C) represents the effective features of the feature sequence F and its corresponding class label sequence C in the discriminability dimension, f(t) k The k-th feature in the sequence F is denoted as . c(t) is the average of all features in the sequence F. k The k-th feature f(t) in the sequence F with feature ) k Category tags, K is the average of all category labels in the sampled time series T, and K is the sequence length of the feature sequence F.

[0105] In this embodiment, step S103 specifically employs kernel principal component analysis to fuse effective features. Step S103 includes:

[0106] S201, Input the generated effective feature vector, and denote any i-th effective feature as x. i ∈R n ,i=1,2,…,m, where m is the number of effective features generated, R n The dimension representing the effective feature is a real number n;

[0107] S202, determine the kernel function and the parameters in the kernel function, and obtain any i-th element K in the kernel matrix K. i for:

[0108] K i =k(x i ,x j ), i,j=1,2,…,m,

[0109] In the above formula, k is a defined kernel function, and x i ,x j These are the i-th and j-th valid features, respectively;

[0110] S203, center the kernel matrix K according to the following formula:

[0111] K = K-1 m K-K1 m +1 m K1 m ,

[0112] In the above formula, 1 m This represents a matrix where all elements are 1 / m, and m is the number of valid features generated.

[0113] S204, Solve for the first p eigenvalues ​​of the centered kernel matrix K: λ1≥λ2≥…≥λ p With the first p eigenvectors α 1 ,α 2 ,…,α p And <α i ,α i >=1, i=1,2,…,p, where <,> denotes inner product operation;

[0114] S205, for any test point x∈R n ,have Where Φ(x) represents the fused features for test point x, and k(x) i (x) represents the kernel function pair (x) i The calculation result of ,x).

[0115] In this embodiment, when using a machine learning model to classify the transformer based on the fused feature vectors to obtain the transformer's operating state in step S104, the machine learning model used is a support vector machine. The classification using the support vector machine to obtain the transformer's operating state refers to solving the optimization problem shown in the following equation by selecting a kernel function k(·) and a penalty parameter C:

[0116]

[0117] Furthermore, the classification decision function f(x) used to solve the optimization problem is:

[0118]

[0119] Where N is the number of training samples, α i and α j The eigenvectors x are respectively i and x j The corresponding penalty coefficient, y i and y j The eigenvectors x are respectively i and x j The tag, To and x j The kernel function calculation results, For x i transpose of x i ,x j These are the i-th and j-th valid features, respectively; C is the penalty parameter; sign is the sign function; K(x·x) i For x and x i The kernel function calculation result is given, where x is the feature vector after fusion of the test samples, and b is the bias parameter.

[0120] Support Vector Machines map the input vector x from the input space to a higher-dimensional Hilbert space. This mapping is typically achieved using a kernel function K(x). i ,x j )=Φ(x i )·Φ(x j The optimal values ​​of the kernel function K(·) and the penalty parameter C are determined. A suitable kernel function K(·) and an appropriate penalty parameter C are selected, and the support vector machine is used to solve the optimization problem. The optimal values ​​of the kernel function K(·) and the penalty parameter C can be searched using a grid parameter optimization algorithm.

[0121] To verify the practical effect of this invention on intelligent recognition of transformer acoustic signals, acoustic signal data of different categories were collected from Changsha 110kV Mawangdui Substation (Category 1), Changsha 110kV Hefeng Substation (Category 2), Changsha 110kV Tiyu Xincheng Field Substation (Category 3), and Loudi Jiulun Substation (Category 4). Their time-domain waveforms are shown below. Figure 3 As shown, the spectrum is as follows Figure 4 The proposed feature extraction results are as follows: Figure 5 As shown. Figure 6 The parameter optimization results of the support vector machine based on grid parameter optimization in this embodiment are given. The optimal penalty parameter C = 256 and the optimal kernel function K(·) parameter g = 0.10882 are obtained. Figure 7The results of transformer acoustic signal recognition are presented, where blue circles represent actual labels, red asterisks represent model recognition results, and the confusion matrix of the recognition results is shown in the figure below. The recognition rate for category 1 reached 100%, category 4 reached 90%, and categories 2 and 3 also achieved recognition rates of over 70%. Experimental results show that the transformer acoustic signal recognition method based on multi-view feature fusion in this embodiment can effectively identify various types of transformer operating states.

[0122] In summary, the method in this embodiment first extracts multi-perspective features of transformer acoustic signals from the time domain, frequency domain, and nonlinear domain. Then, it selects effective features reflecting the transformer's operating state from the feature set based on four dimensions: monotonicity, robustness, trend, and identifiability. These effective features are then fused using kernel principal component analysis. Finally, a parameter-optimized support vector machine is used to learn the mapping relationship between the fused features and fault categories, achieving automatic identification of transformer acoustic signals. This method is characterized by its independence from expert experience and high diagnostic accuracy, and can be applied to the diagnosis of transformer operating conditions in complex noise environments.

[0123] Furthermore, this embodiment also provides a transformer acoustic signal recognition system based on multi-view feature fusion, including a microprocessor and a memory interconnected thereto. The microprocessor is programmed or configured to execute the aforementioned transformer acoustic signal recognition method based on multi-view feature fusion. Additionally, this embodiment also provides a computer-readable storage medium storing a computer program for being programmed or configured by the microprocessor to execute the aforementioned transformer acoustic signal recognition method based on multi-view feature fusion.

[0124] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0125] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying transformer acoustic signals through multi-view feature fusion, characterized in that, include: S101 extracts the features of transformer acoustic signals from multiple dimensions, including time domain, frequency domain, and nonlinear domain. S102, based on the extracted features, generate effective features in four dimensions: monotonicity, robustness, trend and identifiability; S103, fuse the generated effective features to obtain a fused feature vector; S104, based on the fused feature vectors, a machine learning model is used for classification to obtain the operating status of the transformer; The expression for the calculation function of the effective features of the monotonicity dimension in step S102 is: ， In the above formula, The effective features of the sequence F representing the features in the monotonicity dimension. K The sequence length of the characteristic sequence F. Represents a counting function; The expression for the calculation function of the effective features of the robustness dimension in step S102 is: ， In the above formula, The effective features of the sequence F representing the robustness dimension. K The sequence length of the characteristic sequence F. The first sequence F with features k One feature; The calculation function expression for the effective features of the trend dimension in step S102 is as follows: ， In the above formula, The effective features of the sequence F representing the characteristics and its corresponding sampling time series T in the trend dimension. The first sequence F with features k One characteristic, The sequence F with the characteristic of the first k Features Sampling time, Let F be the average of all features in the sequence F. This represents the average value across all times in the sampled time series T. K The sequence length of the characteristic sequence F; The calculation function expression for the effective features of the identifiability dimension in step S102 is as follows: ， In the above formula, The effective features of the feature sequence F and its corresponding class label sequence C in the dimension of discernibility. The first sequence F with features k One characteristic, Let F be the average of all features in the sequence F. The sequence F with the characteristic of the first k Features Category tags, This represents the average of all category labels in the sampled time series T. K The sequence length of the sequence F, which is characterized by .

2. The transformer acoustic signal recognition method based on multi-view feature fusion according to claim 1, characterized in that, The features of the transformer acoustic signal extracted from the time domain perspective in step S101 include some or all of the peak value, peak-to-peak value, average value, absolute average value, root mean square value, standard deviation, slope, and kurtosis of the transformer acoustic signal; the features of the transformer acoustic signal extracted from the frequency domain perspective include some or all of the average amplitude, center frequency, root mean square frequency, and standard deviation of the frequency of the transformer acoustic signal; the features of the transformer acoustic signal extracted from the nonlinear domain perspective include at least one of the information entropy and fractal dimension of the transformer acoustic signal.

3. The transformer acoustic signal recognition method based on multi-view feature fusion according to claim 1, characterized in that, Step S103 includes: S201, Input the generated effective feature vector, let any i-th... i One effective feature is ,in m The number of valid features generated. The dimension of a valid feature is a real number. n ; S202, determine the kernel function and the parameters in the kernel function, and obtain any nth kernel in the kernel matrix K. i element for: ， In the above formula, For a given kernel function, These are the i-th and j-th valid features, respectively; S203, center the kernel matrix K according to the following formula: ， In the above formula, Indicates all elements are The matrix, m The number of valid features generated; S204, Solving for the first part of the centered kernel matrix K p eigenvalues Compared to the past p eigenvectors And there are , ,in Indicates inner product operation; S205, for any test point ,have ,in Indicates the test point x Features after fusion Represents kernel function pairs The calculation results.

4. The transformer acoustic signal recognition method based on multi-view feature fusion according to claim 1, characterized in that, In step S104, when using a machine learning model to classify the transformer based on the fused feature vectors to obtain the transformer's operating status, the machine learning model used is a support vector machine (SVM). The SVM classification to obtain the transformer's operating status refers to selecting a kernel function... Penalty parameter C, solve the optimization problem shown in the following equation: ， Furthermore, the classification decision function used when solving the optimization problem for: ， in The number of training samples. and respectively, feature vectors and The corresponding penalty coefficient, and respectively, feature vectors and The tag, To and The kernel function calculation results, for transpose, These are the i-th and j-th valid features, respectively. Penalty parameters, For symbolic functions, To and The kernel function calculation results, The feature vector after fusion of test samples, This is the bias parameter.

5. A transformer acoustic signal recognition system with multi-view feature fusion, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the transformer acoustic signal recognition method of multi-view feature fusion as described in any one of claims 1 to 4.

6. A computer-readable storage medium storing a computer program, characterized in that, The computer program is used to be programmed or configured by a microprocessor to execute the transformer acoustic signal recognition method of multi-view feature fusion as described in any one of claims 1 to 4.

Images