Exciting transformer voiceprint library construction method based on denoising reconstruction and two-stage classification

Abstract

The invention relates to the technical field of voiceprint recognition, and discloses an exciting transformer voiceprint library construction method based on denoising reconstruction and two-stage classification, which comprises the following steps: acquiring voiceprint data of an exciting transformer and performing discrete wavelet transform to obtain a low-frequency principal component; variational mode decomposition is carried out on the low-frequency principal component to obtain an intrinsic mode function, the intrinsic mode function of a low-frequency band is selected for reconstruction, and voiceprint features of the excitation transformer body are recovered; acquiring working state data of the exciting transformer, splicing the working state data with voiceprint features of the exciting transformer body to form fusion features, classifying the fusion features by using a voiceprint recognition model based on two-stage multi-stage classification, and determining voiceprint categories; and constructing an exciting transformer voiceprint library according to the voiceprint category, collecting new voiceprint data, performing similarity matching on the new voiceprint data and the exciting transformer voiceprint library, and updating the exciting transformer voiceprint library according to a matching result. And the accuracy of excitation transformer state monitoring and the intelligent level of voiceprint library construction are improved.

Description

Technical Field

[0001] This invention relates to the field of voiceprint recognition technology, specifically to a method for constructing a voiceprint library for excitation transformers based on denoising reconstruction and two-stage classification. Background Technology

[0002] With the continuous development of power systems and the increasing demands for power supply reliability, online monitoring and fault early warning technologies for excitation transformers, as one of the core equipment of the power grid, have become crucial for ensuring the safe and stable operation of the power grid. Among various monitoring methods, acoustic signal analysis methods have shown great application potential in the field of transformer mechanical condition assessment and early fault diagnosis due to their non-invasiveness, cost-effectiveness, and rich information capacity. Specifically, by collecting acoustic signals generated during equipment operation, features that characterize its internal state are extracted, and an acoustic feature library is constructed to achieve accurate identification and anomaly diagnosis of transformer operating conditions. Early research mainly focused on using traditional signal processing methods such as Fourier transform and wavelet transform to extract time-frequency domain features and combining them with machine learning algorithms to build classification models. In recent years, with the evolution of deep learning technology, acoustic recognition technology based on deep models such as convolutional neural networks and recurrent neural networks has been introduced, improving the automation level and recognition accuracy of feature extraction and promoting the improvement of intelligent operation and maintenance of excitation transformers.

[0003] However, existing technologies still face many challenges and limitations in practical applications. First, due to the influence of the working environment, the acoustic signature signal of the excitation transformer is often interfered with by the equipment's own noise (such as core vibration and cooling fan) and external environmental noise (such as vehicle noise, conversations, and weather effects). These noise components often overlap with the transformer's own signal in the frequency domain. Traditional filtering and noise reduction methods are difficult to effectively suppress noise while fully preserving key acoustic signature features, leading to signal distortion and affecting the accuracy of subsequent identification. Second, the acoustic signature features of the excitation transformer under different loads and temperature conditions have subtle but crucial differences. Existing classification models mostly adopt a single-stage end-to-end identification strategy. When processing acoustic signature data with numerous categories and high feature similarity, confusion and misjudgment are easily caused by the small distance between categories, especially when there is a lack of other auxiliary monitoring information for joint analysis, which limits the classification performance of the model. Finally, traditional acoustic signature databases are mostly built and rarely updated, unable to adapt to the aging of the transformer's own condition, component replacement, or changes in the operating environment caused by long-term operation. This results in weak timeliness and specificity of the acoustic signature database, reducing the reliability of long-term monitoring. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification, comprising: Acoustic data of the excitation transformer is collected and discrete wavelet transform is performed on the acoustic data to obtain the low-frequency principal component; The low-frequency principal component is decomposed using a variational mode decomposition based on an intelligent optimization algorithm to obtain the intrinsic mode function corresponding to the number of decomposition layers. The intrinsic mode function of the low-frequency band is selected for reconstruction to restore the acoustic signature characteristics of the excitation transformer body. The operating status data of the excitation transformer is acquired and spliced ​​with the acoustic signature features of the excitation transformer body to form a fused feature. The fused feature is then classified using a two-stage multi-level classification acoustic signature recognition model to determine the acoustic signature category. An excitation transformer voiceprint library is constructed based on the aforementioned voiceprint categories. New voiceprint data is collected and matched with the excitation transformer voiceprint library for similarity. The excitation transformer voiceprint library is then updated based on the matching results, thus realizing the construction of the excitation transformer voiceprint library.

[0006] As a preferred embodiment of the excitation transformer acoustic signature library construction method based on denoising reconstruction and two-stage classification described in this invention, the method for performing variational mode decomposition on the low-frequency principal component based on intelligent optimization algorithm to obtain the intrinsic mode function corresponding to the decomposition level includes using the low-frequency principal component as the input signal for variational mode decomposition, and setting the Dirac distribution function, intrinsic mode function and center frequency of each mode as the basic parameters of the variational mode decomposition model. The excitation transformer acoustic signature data of the intrinsic mode function is calculated by Hilbert transform and converted into baseband. The variational mode is decomposed into a constrained optimization problem according to the L2 norm. The constrained optimization problem is transformed into an unconstrained optimization problem by introducing a penalty factor and Lagrange multipliers. The intrinsic mode function, center frequency and Lagrange multipliers are iteratively updated by the alternating direction multiplier method. Repeat the iterative update steps using the alternating direction multiplier method until the convergence condition is met, and obtain the intrinsic mode function corresponding to the decomposition level.

[0007] As a preferred embodiment of the excitation transformer acoustic signature library construction method based on denoising reconstruction and two-stage classification described in this invention, the method of restoring the acoustic signature features of the excitation transformer body includes: decentering and whitening the intrinsic mode functions, removing the mean of each intrinsic mode function, and solving the whitening matrix. Using the whitened signal as input, the unmixing matrix is ​​calculated using the fast independent principal component analysis algorithm. The excitation transformer body acoustic signature signal is then separated from the whitened signal and extracted as the excitation transformer body acoustic signature feature.

[0008] As a preferred embodiment of the excitation transformer acoustic fingerprint library construction method based on denoising reconstruction and two-stage classification described in this invention, wherein the two-stage multi-level classification includes a pre-classification stage and a refined classification stage; The pre-classification stage includes determining the operating state category of the excitation transformer based on the temperature and current information in the operating state data, merging categories with the same operating state, determining the number of pre-classified categories, and setting a pre-classification category center. For each voiceprint sample, the distance between the fused features of the voiceprint sample and the pre-classified category center is calculated to generate a distribution vector; The pre-classification loss function is calculated based on the distribution vector. The position of the pre-classification category center is adjusted using the pre-classification loss function, and the pre-classification category center is converged through multiple iterations.

[0009] As a preferred embodiment of the excitation transformer acoustic fingerprint library construction method based on denoising reconstruction and two-stage classification described in this invention, the refined classification stage includes constructing a corresponding refined classifier for each pre-classification category, wherein the refined classifier is used to further subdivide the acoustic fingerprint samples classified into the same pre-classification category. The gradient of the refinement classification loss function is calculated using the refinement classifier, and the position of the refinement class center is updated according to the gradient of the refinement classification loss function. Through multiple iterations, the refinement class center is moved towards the optimal position. A two-stage loss function is constructed by weighting the pre-classification loss function and the refined classification loss function. By adjusting the weight coefficients of the pre-classification loss function and the refined classification loss function, a smooth transition from pre-classification to refined classification is achieved.

[0010] As a preferred embodiment of the excitation transformer acoustic signature library construction method based on denoising reconstruction and two-stage classification described in this invention, the method for classifying the fused features using an acoustic signature recognition model based on two-stage multi-level classification includes inputting the fused features into the acoustic signature recognition model based on two-stage multi-level classification for training. The acoustic signature recognition model adopts an ECAPA-TDNN network, and the ECAPA-TDNN network is equipped with an optimized compressed excitation residual module. In the optimized compressed excitation residual module, the attention mechanism is used to concatenate the features after temporal convolution and the features after frequency domain processing of the fused features in the channel dimension. The concatenated features are then input into a weighting function and processed by a sigmoid activation function to generate weighting coefficients. The features after temporal convolution are weighted element-wise using the weighting coefficients, and the features after frequency domain processing are weighted element-wise using the complements of the weighting coefficients. The fused output features are obtained based on the two weighting results.

[0011] As a preferred embodiment of the method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification as described in this invention, the construction of the excitation transformer acoustic signature library according to the acoustic signature category includes labeling each acoustic signature data and organizing and storing the labeled acoustic signature data using an indexing method. Collecting new acoustic signature data and performing similarity matching with the acoustic signature database of excitation transformers, and updating the acoustic signature database of excitation transformers based on the matching results includes collecting new acoustic signature data and performing discrete wavelet transform, variational mode decomposition based on intelligent optimization algorithm, and low-frequency intrinsic mode function reconstruction to obtain new acoustic signature features of the excitation transformer body. The new excitation transformer body acoustic signature features are input into the trained two-stage multi-level classification-based acoustic signature recognition model. After obtaining the predicted output, feature matching is performed with the excitation transformer acoustic signature database to calculate the similarity. Based on similarity and a preset threshold, it is determined whether new voiceprint data belongs to an existing voiceprint category and the excitation transformer voiceprint database is updated accordingly.

[0012] A system for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification, wherein: The voiceprint data processing module collects the voiceprint data of the excitation transformer and performs discrete wavelet transform on the voiceprint data to obtain the low-frequency principal component. The signal reconstruction module performs variational mode decomposition on the low-frequency principal component based on an intelligent optimization algorithm to obtain the intrinsic mode function corresponding to the number of decomposition layers. The intrinsic mode function of the low-frequency band is selected for reconstruction to restore the acoustic signature characteristics of the excitation transformer body. A two-stage multi-level classification module acquires the operating status data of the excitation transformer and splices it with the acoustic signature features of the excitation transformer body to form a fused feature. The fused feature is then classified using a two-stage multi-level classification-based acoustic signature recognition model to determine the acoustic signature category. The voiceprint database construction module constructs an excitation transformer voiceprint database based on the voiceprint category, collects new voiceprint data and performs similarity matching with the excitation transformer voiceprint database, and updates the excitation transformer voiceprint database based on the matching results, thereby realizing the construction of the excitation transformer voiceprint database.

[0013] A computer device includes: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, it implements the steps of the method described in any one of the present invention.

[0014] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the present invention.

[0015] The beneficial effects of this invention are as follows: The method for constructing an excitation transformer acoustic signature database based on denoising reconstruction and two-stage classification provided by this invention combines intelligent optimization algorithms with variational mode decomposition to select parameters for acoustic signature signal decomposition. Then, fast independent principal component analysis is used for blind source separation, enabling accurate extraction of the acoustic signature features of the excitation transformer itself from the acoustic signature signal. By constructing an acoustic signature recognition model based on two-stage multi-level classification, the pre-classification stage combines temperature and current information to quickly classify the operating state categories, and the refined classification stage achieves high-precision recognition by mining subtle differences. The fused features are input into the acoustic signature recognition model using an ECAPA-TDNN network. The attention mechanism in the optimized compressed excitation residual module weights the time-domain and frequency-domain features, enhancing the model's robustness and discriminative power in noisy environments. Simultaneously, an indexing method is used in the acoustic signature database construction to improve retrieval efficiency. By collecting new acoustic signature data in real time and performing similarity matching with the acoustic signature database, the database is updated based on the matching results, adapting to the operating characteristics and environmental changes of different excitation transformers. This technology enables accurate identification of the acoustic signature characteristics of excitation transformers while effectively filtering out noise interference, and adapts to changes in operating conditions, thereby improving the accuracy of excitation transformer condition monitoring and the intelligence level of acoustic signature database construction. Attached Figure Description

[0016] 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.

[0017] Figure 1 The overall flowchart of the excitation transformer acoustic fingerprint library construction method based on denoising reconstruction and two-stage classification provided in the embodiments of the present invention is shown.

[0018] Figure 2 This is an overall flowchart of the excitation transformer acoustic signature reconstruction provided in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of a two-stage, multi-level classification structure provided in an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of the optimized compressed excitation residual module structure provided in an embodiment of the present invention.

[0021] Figure 5 The measured acoustic waveform of the excitation transformer is shown in the embodiment of the present invention.

[0022] Figure 6 The measured acoustic waveform spectrum of the excitation transformer is provided for an embodiment of the present invention.

[0023] Figure 7 The waveform diagram of the excitation transformer acoustic waveform after signal reconstruction is provided in an embodiment of the present invention.

[0024] Figure 8 The above is a spectrum diagram of the acoustic waveform signal of the excitation transformer after signal reconstruction, provided in an embodiment of the present invention.

[0025] Figure 9 The waveform of human speech noise after signal reconstruction is provided in an embodiment of the present invention.

[0026] Figure 10 The noise spectrum of human speech after signal reconstruction is provided in an embodiment of the present invention.

[0027] Figure 11 The image shows the reconstructed sound waveform of the signal after being struck, as provided in an embodiment of the present invention.

[0028] Figure 12 The signal reconstruction and impact sound spectrum is provided in an embodiment of the present invention.

[0029] Figure 13 This is a schematic diagram illustrating the training set accuracy versus results provided in an embodiment of the present invention. Detailed Implementation

[0030] 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.

[0031] Example 1, referring to Figures 1-4 As an embodiment of the present invention, a method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification is provided, including: S1: Collect the acoustic fingerprint data of the excitation transformer and perform discrete wavelet transform on the acoustic fingerprint data to obtain the low-frequency principal component.

[0032] In this embodiment, sensors are used to collect the acoustic signature data generated during the operation of the excitation transformer in real time. The acoustic signature signal of the excitation transformer not only reflects the working status of the transformer, but is also affected by environmental noise and electrical noise. These noises include the equipment's own noise (such as the vibration sound of the iron core winding and the noise of the cooling device) and external interference noise (such as the sound of bird deterrents inside and outside the substation, the engine sound of the operation and maintenance vehicle, and the voice of the staff), which increases the complexity of the acoustic signature signal. In addition, with the change of time and frequency, the noise characteristics generated by the excitation transformer under different operating conditions may also be different, which further increases the difficulty of acoustic signature recognition.

[0033] The acoustic signature signal of the excitation transformer exhibits time-frequency correlation, and its distribution in time and frequency differs under different operating conditions. To address this issue, this embodiment employs wavelet transform for signal analysis. Specifically, the Daubechies wavelet is selected as the basis function, as it possesses compact support and orthogonality, ensuring decomposition accuracy while avoiding frequency aliasing. Furthermore, Discrete Wavelet Transform (DWT) is used to process the acquired noise-contaminated acoustic signature signal, achieving signal denoising and feature extraction at multiple scales, thereby revealing the transformer's state information more clearly.

[0034] Furthermore, the decomposition and reconstruction process of the Daubechies wavelet involves the following specific steps: The original excitation transformer acoustic signature data was decomposed into low-frequency coefficients using the Daubechies wavelet. and high frequency coefficient .

[0035] Since the wavelet coefficients after discrete wavelet transform do not possess the characteristics of the actual signal, reconstruction is required; specifically, firstly, the low-frequency coefficients... The low-frequency component is reconstructed using a low-pass filter. Secondly, the high-frequency coefficients Reconstructed into high-frequency components by a high-frequency filter Finally, through low-frequency components and high frequency components Solve for the reconstructed low-frequency principal components: ; in, For the reconstructed low-frequency principal components; The low-frequency component of the nth layer; For the high-frequency components of the j-th layer; This represents the number of decomposition layers.

[0036] The final voiceprint data can be synthesized using the following formula: ; in, This is the final voiceprint data; For the reconstructed low-frequency principal components; For the reconstructed high-frequency noise components.

[0037] It should be noted that, in response to the problem of the intertwining of equipment noise and external interference noise in the operating environment of excitation transformers, discrete wavelet transform is used to perform multi-scale decomposition processing on the collected acoustic fingerprint data. Utilizing the compact support and orthogonality characteristics of the Daubechies wavelet, frequency aliasing is avoided while ensuring decomposition accuracy. The original acoustic fingerprint signal is decomposed into low-frequency and high-frequency coefficients, which are then reconstructed into low-frequency and high-frequency components respectively through low-pass and high-pass filters. This yields the reconstructed low-frequency principal component, effectively removing high-frequency environmental noise interference while retaining the low-frequency principal component carrying transformer operating status information. This lays a clean signal foundation for subsequent fine frequency domain decomposition, avoiding the impact of noise pollution on feature extraction accuracy and improving the signal-to-noise ratio and analyzability of the acoustic fingerprint signal.

[0038] S2: Perform variational mode decomposition based on intelligent optimization algorithm on the low-frequency principal component to obtain the intrinsic mode function corresponding to the decomposition level. Select the intrinsic mode function of the low-frequency band for reconstruction to restore the acoustic characteristics of the excitation transformer body.

[0039] Furthermore, for the reconstructed low-frequency principal components in step S1 Although the discrete wavelet transform has initially separated the high and low frequency components, the reconstructed low-frequency principal components... The signal still contains mixed signals of multiple frequencies, such as the acoustic signature of the excitation transformer body, noise from the cooling device, or low-frequency interference from the substation environment. In order to achieve accurate feature extraction, a more refined frequency domain decomposition method is needed to perform secondary processing on the low-frequency principal components, effectively separating the signal components of different frequency bands, so as to extract the acoustic signature features of the excitation transformer body from the mixed signal.

[0040] Specifically, the low-frequency principal components are subjected to variational mode decomposition based on an intelligent optimization algorithm to obtain intrinsic mode functions corresponding to the decomposition level, including: The reconstructed low-frequency principal components Input data for Variational Mode Decomposition (VMD) Specifically, setting the input data in the variational mode decomposition model. This is the low-frequency acoustic signature signal of the excitation transformer. For the Dirac distribution function, Let be the intrinsic mode functions (IMFs) of the k-th decomposition. Let the center frequency of the k-th mode be denoted by ; simultaneously, a penalty factor is introduced. and Lagrange multipliers This transforms the constrained variational problem into an unconstrained variational problem.

[0041] The idea behind variational mode decomposition is to decompose a multi-component signal into a series of intrinsic mode functions with specific center frequencies and finite bandwidths. Each mode function represents the oscillation characteristics of the signal in a specific frequency band. Through variational framework and optimization solution, variational mode decomposition can determine the center frequency and bandwidth of each mode, thereby achieving effective separation of the signal in the frequency domain.

[0042] Furthermore, the variational mode decomposition includes the following steps: Step 1: Calculate the excitation transformer acoustic signature data for each Intrinsic Mode Function (IMF) using Hilbert transform and convert it to baseband: ; in, The Dirac distribution function; Let be the intrinsic mode functions (IMFs) of the k-th decomposition. It is a natural constant; The imaginary unit; The center frequency of the k-th mode; For time.

[0043] Step 2: For the bandwidth of each mode, utilize L 2 Solving for the norm decomposes the variational pattern into a constrained optimization problem, with the following constraints: : ; in, This is for gradient calculation.

[0044] Step 3: Based on the constrained optimization problem obtained in Steps 1 and 2 above, a penalty factor is introduced. and Lagrange multipliers Transform the constrained optimization problem into an unconstrained optimization problem: ; in, It is a Lagrange function; The center frequency of the k-th mode; This is for gradient calculation; The Dirac distribution function; The imaginary unit; Let be the intrinsic mode functions (IMFs) of the k-th decomposition. Input data for the VMD model; These are Lagrange multiplier functions; For time.

[0045] It should be noted that step 3 introduces a penalty factor. and Lagrange multipliers Originally, it contained constraints. The optimization problem is transformed into an unconstrained optimization problem. At this point, the optimization problem only involves minimizing the modal function and no longer depends on explicit constraints, making the solution more efficient.

[0046] Step 4: Iteratively update the intrinsic mode functions (IMFs), center frequencies, and Lagrange multipliers using the alternating direction multiplier method (ADMM). : ; Where i and k are modal components in VMD decomposition; n is the number of iterations; The center frequency; The center frequency of the k-th mode; This is the updated value of the center frequency of the k-th mode after the (n+1)th iteration; , , , and They are respectively , , , and Fourier transform; Let be the intrinsic mode functions (IMFs) of the k-th decomposition. Let be the intrinsic mode functions (IMFs) of the i-th decomposition. This represents the updated value of the intrinsic mode functions (IMFs) of the k-th decomposition after the (n+1)-th iteration; Lagrange multiplier function The updated value of the Fourier transform after the (n+1)th iteration; Lagrange multiplier function The Fourier transform of τ represents the updated value after the nth iteration; τ is the update parameter. This represents the number of decomposition layers.

[0047] Step 5: Repeat step 4 until the following constraint is met, such that the signal is decomposed into K intrinsic mode functions (IMFs): ; in, The Fourier transform of the intrinsic mode functions (IMFs) of the k-th decomposition is updated after the (n+1)-th iteration; The Fourier transform of the intrinsic mode functions (IMFs) of the k-th decomposition is updated after the nth iteration; It is a natural constant.

[0048] It should be noted that in variational pattern decomposition, the penalty factor... and decomposition layer number The penalty factor is the main factor that directly affects the decomposition results; firstly, the penalty factor. The penalty factor affects the decomposition accuracy. Too low a value will reduce accuracy; conversely, too high a value will increase program load. Secondly, the number of decomposition layers... Different settings will affect the frequency distribution of each intrinsic mode function (IMF) after variational mode decomposition, ultimately affecting the extraction of model frequency domain features and directly impacting the decomposition results of the excitation transformer acoustic signature signal. Therefore, considering the influence of the penalty factor and the number of decomposition levels on the decomposition results during variational mode decomposition, this embodiment utilizes the Superb Fairy-wren Optimization Algorithm (SFOA) to adjust the penalty factor. and decomposition layer number Optimization is performed to determine the optimal penalty factor and the optimal number of decomposition layers.

[0049] Specifically, the SFOA algorithm used in this embodiment simulates the biological behavioral characteristics of the Magnificent Slender-tailed Warbler, including the growth and exploration of juveniles, the breeding and feeding of adults, and the predator avoidance and migration of the flock, and performs a dynamic balance between global search and local optimization in the parameter space. The SFOA algorithm first randomly initializes multiple candidate solutions (populations) within a preset parameter range, and each candidate solution represents a set of penalty factors. and decomposition layer number The algorithm iteratively updates the parameter combination. During the growth phase, it extensively explores to cover different regions of the parameter space. During the reproduction phase, it performs a local depth search around the optimal solution. During the enemy avoidance phase, it uses the Levy flight mechanism to escape local optima. During the iteration process, the similarity coefficient and signal-to-noise ratio are used as fitness evaluation indicators to continuously screen and update the optimal parameter combination until the convergence condition is met. This effectively avoids the experience dependence of manual parameter tuning and the decomposition instability caused by fixed parameters. Compared with particle swarm optimization (PSO) and genetic algorithm (GA), the SFOA algorithm has stronger global search capability and faster convergence speed when dealing with multi-peak optimization problems, ensuring the adaptability of variational mode decomposition under different voiceprint signal characteristics and the optimality of decomposition results.

[0050] It should be noted that, in addressing the issue of multiple frequency components mixing in the reconstructed low-frequency principal components, such as the acoustic signature of the excitation transformer itself, noise from the cooling device, and low-frequency interference from the substation environment, the magnificent slender-tailed warbler algorithm is introduced to adaptively optimize the penalty factor and the number of decomposition levels in variational mode decomposition. This simulates biological behavior characteristics to achieve a balance between global search and local optimization in the parameter space, avoiding the reliance on experience from manual parameter tuning and solving the decomposition instability problem caused by fixed parameters. This allows variational mode decomposition to automatically adjust according to different acoustic signature signal characteristics, obtaining intrinsic mode functions corresponding to the number of decomposition levels, ensuring effective separation of signal components in different frequency bands, and improving the stability and accuracy of acoustic signature feature extraction.

[0051] Furthermore, considering the frequency characteristics of various sound sources in the operating environment of the excitation transformer, this embodiment uses the aforementioned SFOA algorithm and VMD algorithm to decompose the effective temperature rise data according to the frequency band distribution, selects components below 1000Hz to reconstruct the acoustic signature signal, and filters out high-frequency interference components in the acoustic signature signal, thereby reducing the interference sound sources in the acoustic signature signal; in order to further separate the excitation transformer acoustic signature signal from the noise reduction acoustic signature signal and realize signal reconstruction, the inherent mode function corresponding to the acoustic signature signal reconstructed from the selected components below 1000Hz is processed, thereby recovering the acoustic signature signal of the target excitation transformer body.

[0052] Specifically, the acoustic signature characteristics of the restored excitation transformer body include: For the intrinsic mode function Decentralization and whitening are performed. Decentralization refers to removing the mean of each intrinsic mode function, making the mean of each signal zero. The specific formula is as follows: ; in, The voiceprint signal is the intrinsic mode function (IMF) of the k-th decomposition. ; It is the voiceprint signal after removing the mean; It is a voiceprint signal The mean.

[0053] The signal obtained after decentralization It contains k modes : ; in, The signal obtained after decentralization; The modality represents different frequency bands of the signal.

[0054] It should be noted that the purpose of decentralization is to eliminate the DC bias of the signal, making subsequent statistical analysis more accurate. Next, whitening is performed, and the whitening matrix is ​​calculated using the following formula: ; in, This is the whitening matrix; Let be the covariance matrix of the voiceprint signal; Furthermore, blind source separation is performed on the whitened signal. Using the whitened signal as input, a Fast Independent Principal Component Analysis (FastICA) algorithm is used to calculate the unmixing matrix. The excitation transformer body acoustic signature signal is then separated from the whitened signal and extracted as the excitation transformer body acoustic signature feature. The specific formula for the excitation transformer body acoustic signature signal is as follows: ; in, This refers to the acoustic signature signal of the excitation transformer body. The unmixing matrix is ​​estimated by maximizing the non-Gaussianity of the signal; The signal is obtained after decentralization.

[0055] Specifically, the estimation by maximizing the non-Gaussianity of the signal refers to the Fast Independent Principal Component Analysis (Fast Principal Component Analysis) algorithm extracting independent components by maximizing the negative entropy of the signal. Negative entropy is an important indicator of the non-Gaussianity of a signal. According to the Central Limit Theorem, linear combinations of independent signals tend to a Gaussian distribution; therefore, maximizing negative entropy is equivalent to finding the most independent signal components. The Fast Independent Principal Component Analysis algorithm iteratively optimizes the objective function to obtain the optimal unmixing matrix. for: ; in, For output variables; It is a positive number; The number of independent components that need to be optimized; It is a Gaussian random variable; It is not a quadratic function; Let be the expected value of the function.

[0056] Among them, such as Figure 2The diagram shows the overall flowchart for reconstructing the acoustic signature signal of the excitation transformer. Starting from the original signal data, the low-frequency principal components are first obtained through discrete wavelet transform preprocessing. Then, by initializing the SFOA algorithm parameters and population, and giving the maximum number of iterations, the fitness value of each SFOA member is calculated and the optimal individual is updated. The optimal parameter combination is output, and then VMD decomposition is performed to obtain several IMF components. It is determined whether the IMF principal frequency is lower than 1000Hz. If not, it is discarded; if so, the IMF reconstructed signal is retained, and finally, a low-frequency mixed signal is obtained. Next, the noise-reduced signal is centered and whitened. Then, the separation matrix is ​​obtained through the FastICA algorithm to achieve blind source separation, and finally, the acoustic signature signal of the target excitation transformer body is recovered.

[0057] It should be noted that, to address the issue that the selected low-frequency intrinsic mode functions still contain multiple independent sound source signals, the DC bias is eliminated by decentering the intrinsic mode functions. Then, the signal is preprocessed by solving the whitening matrix through whitening to ensure the independence of each component. Based on this, a fast independent principal component analysis algorithm is used to extract independent components by maximizing the negative entropy of the signal. The most independent signal component is then found according to the central limit theorem. The excitation transformer's acoustic signature signal is separated from the whitened signal. This allows for accurate extraction of the target acoustic signature signal from the mixed signal of multiple sound sources even without a known mixing matrix. It effectively filters out cooling device noise and low-frequency environmental interference, restoring the pure acoustic signature features of the excitation transformer, providing high-quality feature input for subsequent classification and recognition, and improving the reliability of acoustic signature recognition.

[0058] S3: Obtain the operating status data of the excitation transformer and splice it with the voiceprint features of the excitation transformer body to form a fused feature. Use a voiceprint recognition model based on two-stage multi-level classification to classify the fused feature and determine the voiceprint category.

[0059] It should be noted that in large-scale voiceprint recognition tasks, due to the diverse types of voiceprint data and the high complexity of classification models, voiceprints under the same operating conditions are highly similar in the vast feature space, with small inter-class distances, which easily leads to classification errors. Furthermore, labeling voiceprints under different operating conditions with the same tag reduces classification accuracy. Therefore, this invention constructs a two-stage, multi-level classification-based voiceprint recognition model, dividing the classification process into two levels: pre-classification based on the operating state of the excitation transformer and refined classification based on the pre-classification results. First, pre-classification combines other monitoring data of the excitation transformer, including temperature and current information, to classify only operating states with large inter-class distances and easy differentiation, thereby reducing the number of classifications and simplifying the classification task. Subsequently, refined classification, based on the pre-classification results, further subdivides samples of the same category, expanding the inter-class distance by mining subtle differences between similar categories, effectively distinguishing the voiceprint features of the excitation transformer under the same operating conditions, and improving the accuracy of voiceprint recognition.

[0060] Furthermore, acquiring the operating status data of the excitation transformer and splicing it with the acoustic signature features of the excitation transformer body to form a fused feature refers to acquiring the transformer's operating status data in real time through other monitoring systems of the excitation transformer. The operating status data includes temperature information and current information; wherein, the temperature information reflects the thermal state and load level of the transformer, and the current information reflects the electrical load state of the transformer. Splicing the acoustic signature features of the excitation transformer body obtained in step S2 with the operating status data to form a fused feature enables subsequent discrimination not only based on the time-frequency characteristics of the acoustic signature, but also to make comprehensive decisions based on the physical operating status. Especially when the acoustic signature features are similar but the operating status is different, the operating status data can provide key distinguishing information.

[0061] Specifically, such as Figure 3 The diagram illustrates the structure of a two-stage multi-level classification system, which includes a pre-classification stage and a refinement stage. In the pre-classification stage, fused features are input to a pre-classifier. Based on temperature and current information, categories with the same operating state are merged, quickly dividing the system into several coarse categories. The pre-classification category centers are continuously updated through iterative optimization. In the refinement stage, an independent refinement classifier is constructed for each pre-classification category. Voiceprint samples classified into the same pre-classification category are further subdivided, and the refinement category centers are updated through a gradient feedback mechanism, achieving accurate identification of subtle differences. The two-stage loss function achieves a smooth transition from pre-classification to refinement through weight adjustment, ensuring that the model converges quickly and achieves high-precision recognition during training.

[0062] In the pre-classification stage, based on temperature and current information, categories with the same operating state of the excitation transformer are merged into one pre-classification. New voiceprint samples are continuously added during training, and the pre-classified class centers are continuously updated and optimized by analyzing the distance between each sample and the class center. Specifically, the pre-classification stage involves the following steps: First, after reconstructing the acoustic signature signal of the excitation transformer using an optimized support vector machine (SVM) method, the number of pre-classified categories is determined. Set the initial category center for each category in the pre-classification. The initial category center These represent the characteristics of each operating state category of the excitation transformer.

[0063] For each voiceprint sample, the distance between the fused features of the voiceprint sample and the pre-classified category center is calculated. During the distance calculation process, each voiceprint sample generates a distribution vector, reflecting the distribution of the voiceprint sample in each category, which helps to evaluate the similarity between the voiceprint sample and the category center. ; ; in, Let be the distribution vector, representing the predicted sample distribution. Each Both are vectors of size 0×1. The number of categories in the pre-classification; Features of fusion; As the pre-classified category center; This is the distance calculation function.

[0064] The pre-classification loss function is calculated based on the distribution vector to obtain the model's prediction error, which is used to measure the difference between the prediction result and the actual category. ; in, The loss function; The number of categories in the pre-classification; This represents the true sample distribution; Let be the distribution of the i-th predicted sample.

[0065] The update direction of the pre-classified category centers is calculated using the pre-classification loss function: ; in, The direction for updating the pre-classified category centers; The initial category center for each category in the pre-classification; Features of fusion; As the pre-classified category center; For fusion features The distance between the i-th pre-classified category center; This represents the true sample distribution; Let i be the distribution of the i-th predicted sample; This represents the number of categories in the pre-classification.

[0066] The position of the pre-classified category centers is adjusted according to the update direction of the pre-classified category centers. The specific formula is as follows: ; in, For the updated pre-classification category centers; The initial category center for each category in the pre-classification; The learning rate; This indicates the direction for updating the pre-classified category centers.

[0067] Repeat the steps described above, from calculating the distance between the fused features of the voiceprint samples and the pre-classified category centers to adjusting the position of the pre-classified category centers based on their update direction. Through multiple iterations, the pre-classified category centers converge until the pre-classification stage achieves the desired accuracy on the training dataset. This ideal accuracy is adjusted based on factors such as the model's training process, optimization algorithm, and dataset characteristics. Through backpropagation and hyperparameter tuning, the model automatically updates its parameters, gradually approaching the ideal accuracy. Furthermore, the position of the category centers is adjusted in each iteration, thereby continuously improving the classifier's performance.

[0068] Furthermore, based on the pre-classification stage, a more refined classification is carried out, focusing on subtle differences in excitation transformers under the same operating conditions, and a separate refined classifier is designed for each pre-classification.

[0069] Specifically, the detailed classification stage involves the following steps: For each pre-classified category, a corresponding refinement classifier is constructed. This refinement classifier is used to further subdivide voiceprint samples classified into the same pre-classified category. In each classification step, the gradient of the refinement classification loss function is calculated using the refinement classifier, and the refinement category centers are updated based on this gradient. The location of the category center is refined through multiple iterations. Move to the optimal position: ; ; in, For the updated, more refined category center; The learning rate; The loss function; The first The refined category centers after the next iteration This represents the number of iterations.

[0070] It should be noted that the category center is refined. The update is adjusted based on changes in sample features to optimize classification accuracy, where the update magnitude is determined by the learning rate. The position of the refined class center is determined by the gradient of the loss function; it is adjusted after each classification and continuously optimized in multiple iterations. Through the gradient feedback mentioned above, the refined class center gradually moves closer to the optimal position, ultimately achieving accurate classification of minor differences. After multiple iterations, the refined classifier can effectively reduce classification errors and improve the model's classification ability in real-world environments.

[0071] A two-stage loss function is constructed by weighting the pre-classification loss function and the refined classification loss function. By adjusting the weight coefficients of the pre-classification loss function and the refined classification loss function, a smooth transition from pre-classification to refined classification is achieved.

[0072] The specific formula for the two-stage loss function is as follows: ; in, It is a two-stage loss function; This refers to the cross-entropy loss for pre-classification, i.e., the pre-classification loss function; To refine the cross-entropy loss for classification, i.e., to refine the classification loss function; This refers to the current training round. Parameters for controlling the transition speed; This is for transitional training rounds.

[0073] It should be noted that, for the above two-stage loss function When in the pre-classification stage, the corresponding loss function Equal to the cross-entropy loss of pre-classification When in the refinement classification stage, the corresponding loss function is... Equivalent to the cross-entropy loss of refined classification .

[0074] Furthermore, classifying the fused features using a two-stage multi-level classification-based voiceprint recognition model includes the following steps: inputting the fused features into a two-stage multi-level classification-based voiceprint recognition model for training, wherein the voiceprint recognition model employs an ECAPA-TDNN network, and the ECAPA-TDNN network is configured with an optimized squeeze-and-excitation residual block (SE-Res2Block).

[0075] It should be noted that in building the excitation transformer voiceprint recognition model based on two-stage multi-level classification, the ECAPA-TDNN network is used as the core of the model framework based on the above two-stage multi-level classification. The residual summation operation in the original compressed excitation residual module of the TDNN is optimized, and an attention mechanism is introduced to achieve more refined weighted control of the input features, resulting in stronger selectivity and feedback adjustment capabilities; among them, such as Figure 4 The optimized compressed excitation residual module structure is shown. The model input is n-dimensional voiceprint features, where k is the convolution kernel size, d is the expansion spacing, C is the number of channels in the intermediate feature map, T is the time dimension, and S is the number of voiceprint features. The optimized compressed excitation residual module structure enhances the model's feature recognition capability in real noise environments, providing strong support for constructing a highly robust excitation transformer voiceprint recognition method.

[0076] Furthermore, in the optimized compressed excitation residual module, an attention mechanism is used to process the fused features after temporal convolution. Features after frequency domain processing The features are spliced ​​together along the channel dimension and merged into a fused feature. .

[0077] The spliced ​​fusion features Input to weighting function The process involves obtaining initial weighting coefficients, which are then normalized using the sigmoid activation function to generate weighting coefficients within the range [0,1]. .

[0078] By element-wise multiplication The input features are weighted; specifically, the weighting coefficients are used. Features after temporal convolution Perform element-wise weighting, using the complement of the weighting coefficients. Features after frequency domain processing Perform element-wise weighting, and obtain the fused output features based on the two weighting results.

[0079] Specifically, the fused output features are obtained based on the two weighted results. The specific formula is as follows: ; in, The output features after fusion; Features are those processed by temporal convolution, i.e., features processed by temporal convolutional layers and attention mechanisms. These are the features after frequency domain processing.

[0080] It should be noted that, to address the issue of subtle differences in the acoustic signature features of excitation transformers under the same operating conditions, making accurate differentiation difficult, an independent refinement classifier is constructed for each pre-classified category based on the pre-classification. The gradient of the refinement classification loss function is used to update the center position of the refinement category, allowing the refinement category center to move towards the optimal position through multiple iterations. Simultaneously, an ECAPA-TDNN network is used as the model framework, introducing an attention mechanism into the compressed excitation residual module structure. The features after temporal convolution and frequency domain processing are concatenated along the channel dimension, and weighting coefficients are generated using a weighting function and a sigmoid activation function, achieving adaptive weighted fusion of temporal and frequency domain features. This approach can uncover subtle differences between similar categories, increase inter-class distance, improve the model's ability to distinguish different acoustic signatures under the same operating conditions, and enhance the model's robustness and recognition accuracy in real-world noise environments.

[0081] S4: Construct an excitation transformer voiceprint library based on the voiceprint category, collect new voiceprint data and perform similarity matching with the excitation transformer voiceprint library, update the excitation transformer voiceprint library based on the matching results, and realize the construction of the excitation transformer voiceprint library.

[0082] After completing the voiceprint data processing, noise reduction and reconstruction, and classification and recognition in steps S1 to S3, the voiceprint database construction and update stage begins. By establishing a standardized voiceprint data storage mechanism and combining real-time acquisition and similarity matching strategies, the voiceprint database of the excitation transformer is continuously improved and adjusted.

[0083] Furthermore, constructing the excitation transformer voiceprint library based on the voiceprint categories includes labeling each voiceprint data, organizing and storing the labeled voiceprint data using an indexing method. In this embodiment, an efficient index structure is used to manage voiceprint features, ensuring retrieval efficiency during large-scale data queries, and an independent index partition is set for each working status category to facilitate rapid matching in the future.

[0084] The process of collecting new voiceprint data and matching it with the excitation transformer voiceprint database, followed by updating the database based on the matching results, includes: collecting new voiceprint data and performing discrete wavelet transform, variational mode decomposition based on intelligent optimization algorithms, and low-frequency intrinsic mode function reconstruction according to steps S1 and S2 to obtain new voiceprint features of the excitation transformer itself. These new voiceprint features are then input into the two-stage multi-level classification-based voiceprint recognition model trained in step S3. After obtaining the predicted output, feature matching is performed with the excitation transformer voiceprint database to calculate the similarity. Finally, based on the similarity and a preset threshold, it is determined whether the new voiceprint data belongs to an existing voiceprint category, and the excitation transformer voiceprint database is updated accordingly. Through real-time updates, the model is used to supplement and improve the voiceprint database, enabling it to adapt to the operating characteristics and environmental changes of different excitation transformers.

[0085] On the other hand, this embodiment also provides a system for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification, which includes: The voiceprint data processing module collects the voiceprint data of the excitation transformer and performs discrete wavelet transform on the voiceprint data to obtain the low-frequency principal component.

[0086] The signal reconstruction module performs variational mode decomposition on the low-frequency principal component based on an intelligent optimization algorithm to obtain the intrinsic mode function corresponding to the decomposition level. The intrinsic mode function of the low-frequency band is selected for reconstruction to restore the acoustic characteristics of the excitation transformer body.

[0087] The two-stage multi-level classification module acquires the operating status data of the excitation transformer and splices it with the voiceprint features of the excitation transformer body to form a fused feature. The fused feature is then classified using a voiceprint recognition model based on two-stage multi-level classification to determine the voiceprint category.

[0088] The voiceprint database construction module constructs an excitation transformer voiceprint database based on the voiceprint category, collects new voiceprint data and performs similarity matching with the excitation transformer voiceprint database, and updates the excitation transformer voiceprint database based on the matching results, thereby realizing the construction of the excitation transformer voiceprint database.

[0089] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0091] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0092] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0093] Example 2, refer to Figures 5-13 As an embodiment of the present invention, a method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0094] In this embodiment, the sampling frequency is set to 48kHz, and 1200 sets of acoustic fingerprint signals collected from a certain excitation transformer under different operating conditions are divided proportionally, including 706 sets under normal conditions, 266 sets under light load, and 228 sets under heavy load. The training set and test set are divided at an 8:2 ratio. Since the original acoustic fingerprint signal contains noise, the acoustic fingerprint signal data is first processed. In order to filter out high-frequency interference components, the parameters in the SFOA algorithm and VMD algorithm are set. The population size, dimension, maximum number of iterations, and upper and lower bounds are set as shown in Table 1.

[0095] Table 1. Parameter settings for SFOA and VMD algorithms

[0096] The calculations based on the parameters set in Table 1 yielded the penalty factors for the SFOA and VMD algorithms. and decomposition layer number The acoustic waveform of a single-phase excitation transformer was decomposed based on the optimal parameter combination. The time-domain waveform and spectrum waveform before and after signal reconstruction are shown below. Figures 5-12 As shown.

[0097] in Figure 5 and Figure 6 The images show the time-domain waveform and spectrum of the measured acoustic signature signal from the excitation transformer. Figure 5 The horizontal axis represents time, and the vertical axis represents signal amplitude. The graph shows that the original signal contains a large amount of high-frequency noise and low-frequency mixed components, with obvious spikes and irregular oscillations in the waveform. The main features of the signal are masked by noise, making it difficult to directly use for feature extraction and voiceprint recognition. Furthermore... Figure 6The horizontal axis represents frequency, and the vertical axis represents normalized amplitude. The signal shown in the figure is mainly concentrated in the low-frequency band below 1000Hz, but there are obvious noise components in the high-frequency band (above 1000Hz). These high-frequency noise components will interfere with the subsequent feature extraction and classification process.

[0098] Compared to Figure 5 and Figure 6 After signal reconstruction, waveforms and spectrograms of the acoustic signature signal of the excitation transformer body, human speech noise, and knocking sound were obtained, as detailed below. Figures 7-12 As shown. Figure 7 and Figure 8 The images show the time-domain waveform and spectrum of the acoustic signature signal of the excitation transformer body, respectively. Figure 7 The horizontal axis represents time, and the vertical axis represents signal amplitude. As can be seen from the figure, after decomposition using the SFOA and VMD algorithms and blind source separation using FastICA, the acoustic waveform of the excitation transformer exhibits regular periodic oscillations, a smooth waveform, and effective suppression of high-frequency noise. Simultaneously, Figure 8 The horizontal axis represents frequency, and the vertical axis represents normalized amplitude. The acoustic signature signal of the excitation transformer body is mainly concentrated in the low frequency band, with a significant main peak below 200Hz, corresponding to the fundamental frequency component of the excitation transformer core vibration. The spectrum is pure and the peak is prominent, indicating that the acoustic signature features of the excitation transformer body have been effectively extracted. Figure 9 and Figure 10 The images show the time-domain waveform and spectrum of human speech noise, respectively. Figure 9 The horizontal axis represents time, and the vertical axis represents signal amplitude. As can be seen from the graph, human speech noise exhibits a non-periodic, bursty characteristic, with irregular waveform fluctuations; simultaneously, Figure 10 The horizontal axis represents frequency, and the vertical axis represents normalized amplitude. The distribution of human speech noise is relatively concentrated, with discrete peaks in the low-frequency band (below 500Hz) and the mid-frequency band (around 1000Hz), corresponding to the fundamental frequency and harmonic components of human voice. This is significantly different from the frequency characteristics of the excitation transformer body's acoustic signature, and it was successfully separated from the mixed signal through blind source separation. Figure 11 and Figure 12 The images show the time-domain waveform and spectrum of the striking sound, respectively. Figure 11 The horizontal axis represents time, and the vertical axis represents signal amplitude. As can be seen from the graph, the striking sound exhibits transient impact characteristics, with the waveform showing a short, high-amplitude pulse. Within the 0.04 to 0.06 second timeframe, a high-amplitude rectangular pulse lasting approximately 0.02 seconds appears. The signal amplitude before and after this pulse is close to zero, demonstrating the concentrated energy and short duration of the striking sound. Simultaneously, Figure 12The horizontal axis represents frequency, and the vertical axis represents normalized amplitude. The impact sound shown in the figure is distributed in the low-frequency range (below 500Hz), and the spectrum exhibits the broadband characteristics unique to impact signals, which is significantly different from the narrow-band periodic characteristics of the excitation transformer's acoustic signature. Effective separation was achieved through signal reconstruction. Figures 5 to 12 The comparison shows that after decomposition by SFOA and VMD algorithms, reconstruction of frequency bands below 1000Hz, and FastICA blind source separation, the original acoustic signature signal is effectively decomposed into the acoustic signature signal of the excitation transformer body and various noise signals, thus achieving accurate extraction of the target signal.

[0099] Secondly, based on the above separation results, the similarity coefficient (C) and signal-to-noise ratio (SNR) are selected as evaluation indicators, and the specific formulas are as follows: ; ; in, Let be the amplitude of the noisy signal at time t; The amplitude of the denoised signal at time t; The number of time samples of the signal; Let C be the denoised signal at time t. The larger the similarity coefficient C, the higher the similarity between the separated signal and the source signal, and the better the separation performance of the algorithm. The higher the signal-to-noise ratio SNR, the smaller the distortion of the separated signal and the more the features of the source signal are preserved. The above two indicators can be used to judge whether the denoising effect of the algorithm is good.

[0100] Furthermore, in order to verify the superiority of the present invention in voiceprint signal separation, the original voiceprint signal was compared with VDM, SHO-VDM and WOA-VDM respectively. The results of the separation performance comparison are shown in Table 2.

[0101] Table 2. Comparison of Separation Performance

[0102] As shown in Table 2, the signal separation method proposed in this invention achieves excellent results in processing measured acoustic fingerprint signals from excitation transformers, with a similarity coefficient C of 0.9543 and a signal-to-noise ratio (SNR) of 30.76 dB. Compared with traditional unoptimized methods and other optimized methods, this invention demonstrates improvements in both signal recovery quality and noise suppression capability. Specifically, the similarity coefficient is improved by approximately 0.4390 compared to the traditional FASTICA algorithm, while the SNR is improved by approximately 6.49 dB compared to the 24.27 dB of SHO-VDM-FASTICA, indicating that this invention has significant advantages in denoising effect and signal separation accuracy. By introducing an optimization mechanism and more accurate feature extraction, this invention can effectively remove high-frequency noise and recover the low-frequency components of the signal, resulting in higher signal recovery accuracy and stronger denoising capability. This further improves the signal quality in practical applications and provides support for acoustic fingerprint signal identification of excitation transformers.

[0103] Furthermore, this invention uses the Adam optimizer and StepLR learning rate scheduler, and employs AAM-Softmax for training. The training margin is set to 0.2, the Softmax pre-scaling is 30, and the training epochs are 4. The batch size is set to 128. The convolutional frame layer in the two-stage multi-level classification-based voiceprint recognition model is configured with 512 channels, the bottleneck dimension of the attention module is set to 128, and the scale_dimension in Res2Block is set to 8. The separated excitation transformer body voiceprint signal is used as the input feature matrix. To optimize the model, the cross-entropy loss measured by the following formula is used. As the objective function: ; in, The actual label for the i-th class; Let be the Softmax probability of the i-th class; The total number of classes.

[0104] To verify the superiority of the two-stage multi-level classification-based voiceprint recognition model, ResNet34, DCNN, ECAPA-TDNN, and the model proposed in this invention were trained and compared. The accuracy comparison results of the four models on the test set are as follows: Figure 13 As shown, the horizontal axis represents the number of training rounds, and the vertical axis represents the recognition accuracy on the test set. The formula for calculating the accuracy is as follows: ; in, For accuracy; The number of correctly classified samples; This represents the total number of samples.

[0105] from Figure 13 It can be seen that when the iteration reaches 20 times, the accuracy of the DCNN model reaches about 84%, while the accuracy of the ResNet34 model and the ECAPA-TDNN model reaches about 89%. The accuracy of the voiceprint recognition model based on two-stage multi-level classification proposed in this invention reaches about 94%, and is still rising. Compared with the final accuracy of the DCNN model (about 89%) and the ResNet34 model (about 92%), this invention converges quickly during training and finally reaches an accuracy of 97%, which is significantly better than other traditional models. In addition, the accuracy of the ECAPA-TDNN model is about 93%, while this invention improves the accuracy by 4% by designing a two-stage multi-level classification model. It makes full use of the on-site information and effectively improves the recognition ability of the model. This shows that this invention constructs an excitation transformer voiceprint database by optimizing signal denoising and reconstruction technology and establishing a two-stage multi-level classification voiceprint recognition model, and realizes the dynamic updating and adaptive adjustment of the voiceprint database.

[0106] 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 an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification, characterized in that, include: Acoustic data of the excitation transformer is collected and discrete wavelet transform is performed on the acoustic data to obtain the low-frequency principal component; The low-frequency principal component is decomposed using a variational mode decomposition based on an intelligent optimization algorithm to obtain the intrinsic mode function corresponding to the number of decomposition layers. The intrinsic mode function of the low-frequency band is selected for reconstruction to restore the acoustic signature characteristics of the excitation transformer body. The operating status data of the excitation transformer is acquired and spliced ​​with the acoustic signature features of the excitation transformer body to form a fused feature. The fused feature is then classified using a two-stage multi-level classification acoustic signature recognition model to determine the acoustic signature category. An excitation transformer voiceprint library is constructed based on the aforementioned voiceprint categories. New voiceprint data is collected and matched with the excitation transformer voiceprint library for similarity. The excitation transformer voiceprint library is then updated based on the matching results, thus realizing the construction of the excitation transformer voiceprint library.

2. The method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification as described in claim 1, characterized in that: The low-frequency principal component is subjected to variational mode decomposition based on intelligent optimization algorithm to obtain the intrinsic mode function corresponding to the decomposition level. This includes using the low-frequency principal component as the input signal for variational mode decomposition, and setting the Dirac distribution function, intrinsic mode function and center frequency of each mode as the basic parameters of the variational mode decomposition model. The acoustic waveform data of the excitation transformer was calculated using the Hilbert transform and converted to baseband. Based on L... 2 The norm decomposes the variational mode into a constrained optimization problem, introduces a penalty factor and Lagrange multipliers to transform the constrained optimization problem into an unconstrained optimization problem, and uses the alternating direction multiplier method to iteratively update the intrinsic mode function, center frequency and Lagrange multipliers; Repeat the iterative update steps using the alternating direction multiplier method until the convergence condition is met, and obtain the intrinsic mode function corresponding to the decomposition level.

3. The method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification as described in claim 2, characterized in that: The acoustic signature characteristics of the restored excitation transformer body include: The intrinsic mode functions are decentered and whitened by removing the mean of each intrinsic mode function, and the whitening matrix is ​​solved. Using the whitened signal as input, the unmixing matrix is ​​calculated using the fast independent principal component analysis algorithm. The excitation transformer body acoustic signature signal is then separated from the whitened signal and extracted as the excitation transformer body acoustic signature feature.

4. The method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification as described in claim 3, characterized in that: The two-stage multi-level classification includes a pre-classification stage and a detailed classification stage; The pre-classification stage includes determining the operating state category of the excitation transformer based on the temperature and current information in the operating state data, merging categories with the same operating state, determining the number of pre-classified categories, and setting a pre-classification category center. For each voiceprint sample, the distance between the fused features of the voiceprint sample and the pre-classified category center is calculated to generate a distribution vector; The pre-classification loss function is calculated based on the distribution vector. The position of the pre-classification category center is adjusted using the pre-classification loss function, and the pre-classification category center is converged through multiple iterations.

5. The method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification as described in claim 4, characterized in that: The refined classification stage includes constructing a corresponding refined classifier for each pre-classified category. The refined classifier is used to further subdivide the voiceprint samples that are classified into the same pre-classified category. The gradient of the refinement classification loss function is calculated using the refinement classifier, and the position of the refinement class center is updated according to the gradient of the refinement classification loss function. Through multiple iterations, the refinement class center is moved towards the optimal position. A two-stage loss function is constructed by weighting the pre-classification loss function and the refined classification loss function. By adjusting the weight coefficients of the pre-classification loss function and the refined classification loss function, a smooth transition from pre-classification to refined classification is achieved.

6. The method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification as described in claim 5, characterized in that: The classification of the fused features using a two-stage multi-level classification-based voiceprint recognition model includes inputting the fused features into the two-stage multi-level classification-based voiceprint recognition model for training. The voiceprint recognition model uses an ECAPA-TDNN network, and the ECAPA-TDNN network is equipped with an optimized compressed excitation residual module. In the optimized compressed excitation residual module, the attention mechanism is used to concatenate the features after temporal convolution and the features after frequency domain processing of the fused features in the channel dimension. The concatenated features are then input into a weighting function and processed by a sigmoid activation function to generate weighting coefficients. The features after temporal convolution are weighted element-wise using the weighting coefficients, and the features after frequency domain processing are weighted element-wise using the complements of the weighting coefficients. The fused output features are obtained based on the two weighting results.

7. The method for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification as described in claim 6, characterized in that: Constructing an excitation transformer acoustic text library based on the acoustic text categories includes labeling each acoustic text data and organizing and storing the labeled acoustic text data using an indexing method. Collecting new acoustic signature data and performing similarity matching with the acoustic signature database of excitation transformers, and updating the acoustic signature database of excitation transformers based on the matching results includes collecting new acoustic signature data and performing discrete wavelet transform, variational mode decomposition based on intelligent optimization algorithm, and low-frequency intrinsic mode function reconstruction to obtain new acoustic signature features of the excitation transformer body. The new excitation transformer body acoustic signature features are input into the trained two-stage multi-level classification-based acoustic signature recognition model. After obtaining the predicted output, feature matching is performed with the excitation transformer acoustic signature database to calculate the similarity. Based on similarity and a preset threshold, it is determined whether new voiceprint data belongs to an existing voiceprint category and the excitation transformer voiceprint database is updated accordingly.

8. A system for constructing an excitation transformer acoustic signature library based on denoising reconstruction and two-stage classification using the method described in any one of claims 1-7, characterized in that: The voiceprint data processing module collects the voiceprint data of the excitation transformer and performs discrete wavelet transform on the voiceprint data to obtain the low-frequency principal component. The signal reconstruction module performs variational mode decomposition on the low-frequency principal component based on an intelligent optimization algorithm to obtain the intrinsic mode function corresponding to the number of decomposition layers. The intrinsic mode function of the low-frequency band is selected for reconstruction to restore the acoustic signature characteristics of the excitation transformer body. A two-stage multi-level classification module acquires the operating status data of the excitation transformer and splices it with the acoustic signature features of the excitation transformer body to form a fused feature. The fused feature is then classified using a two-stage multi-level classification-based acoustic signature recognition model to determine the acoustic signature category. The voiceprint database construction module constructs an excitation transformer voiceprint database based on the voiceprint category, collects new voiceprint data and performs similarity matching with the excitation transformer voiceprint database, and updates the excitation transformer voiceprint database based on the matching results, thereby realizing the construction of the excitation transformer voiceprint database.

9. A computer device, comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images