Transformer abnormal sound source positioning method and system

Abstract

The invention relates to a transformer abnormal sound source positioning method and system, and belongs to the technical field of power equipment state evaluation, and the method comprises the steps: constructing a Bayesian neural network embedded with a voiceprint physical mechanism, and enabling a sound wave propagation equation to serve as a physical constraint to be embedded into the Bayesian neural network; reconstructing the voiceprint signal by adopting a compressed sensing technology to obtain a reconstructed voiceprint field; designing a multi-task objective function including data fitting, physical constraint and positioning loss, and optimizing data fitting, physical constraint and positioning precision to obtain a trained Bayesian neural network; based on a gradient sound source inversion positioning algorithm and the trained Bayesian neural network, sparse regularization is combined to obtain a prediction result of accurate positioning; and a prediction result is visualized to a three-dimensional model of the transformer, and the position of an abnormal sound source is visually displayed. According to the method, the limitation of a traditional method in a complex environment is overcome, and high-precision and high-robustness transformer abnormal sound source positioning is realized.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition assessment technology, specifically to a method and system for locating abnormal sound sources in transformers. Background Technology

[0002] In power systems, the safe operation of transformers is crucial. Faults such as partial discharge inside transformers generate sound waves, which can be detected for early warning and location of faults. However, traditional sound source localization methods are ineffective inside the complex structures of transformers and are computationally expensive. In recent years, Physical Information Neural Networks (PINNs) have shown excellent performance in solving partial differential equation (PDE) problems, while Bayesian Physical Information Neural Networks (B-PINNs) further incorporate a Bayesian framework, enabling them to handle noisy data and quantify uncertainty.

[0003] The main problems are as follows: 1. The complex internal structure of transformers and the diverse propagation paths of sound waves lead to insufficient accuracy of traditional positioning methods. 2. High noise interference in the transformer operating environment affects positioning accuracy. 3. Traditional methods involve large computational loads, making real-time monitoring difficult. 4. The lack of quantification of the uncertainty of positioning results affects the reliability of decision-making.

[0004] These problems mainly stem from the fact that traditional methods cannot effectively combine prior physical knowledge with data-driven approaches, making it difficult to accurately capture the propagation characteristics of sound waves in complex environments, and also lacking consideration for model uncertainties. Summary of the Invention

[0005] This invention provides a method and system for locating abnormal sound sources in transformers. This method effectively combines prior physical knowledge, Bayesian uncertainty quantification, and data-driven deep learning, overcoming the limitations of traditional methods in complex environments and achieving high-precision and robust localization of abnormal sound sources in transformers.

[0006] The technical solution adopted by this invention to solve its technical problem is: In a first aspect, the present invention provides a method for locating abnormal sound sources in a transformer, comprising: A Bayesian neural network with embedded voiceprint physical mechanism is constructed, and the sound wave propagation equation is embedded into the Bayesian neural network as a physical constraint. Based on Bayesian neural networks, compressed sensing technology is used to reconstruct the voiceprint signal to obtain the reconstructed voiceprint field. Design a multi-task objective function that includes data fitting, physical constraints, and localization loss, and optimize the data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network. Based on the reconstructed voiceprint field, a gradient-based sound source inversion localization algorithm and a trained Bayesian neural network are combined with sparse regularization to obtain accurate localization prediction results. The prediction results are visualized onto a 3D model of the transformer, which intuitively shows the location of the abnormal sound source.

[0007] As a further improvement to this application, the construction of a Bayesian neural network embedding the physical mechanism of voiceprints, which embeds the sound wave propagation equation as a physical constraint into the Bayesian neural network, includes: Acquire the acoustic signature signal of the transformer as the raw signal; Construct a compressed sensing framework, assuming the original signal Through linear measurement matrix Compressed into low-dimensional observations y=Φx+ Where y is the low-dimensional observation signal, This indicates that the measurement matrix Φ is an M-row N-column matrix; The noise is a vector with the same dimension as y; Next, physical information is embedded, and the sound wave propagation equation followed by signal x is embedded into the Bayesian neural network in the form of a partial differential equation. The Bayesian neural network structure is as follows: Design a Bayesian neural network Map the compressed observation y to the reconstructed signal Where y is the low-dimensional observation signal mentioned earlier, These are the network parameters, including all learnable weights and biases in a Bayesian neural network. The signal is reconstructed using a Bayesian neural network; The physical loss term is the residual of the partial differential equation. Add a loss function to ensure the reconstructed signal satisfies physical constraints: Lphysics represents the physical loss term; λ is a balancing parameter used to adjust the weight of the physical loss term in the total loss function; Ephysics y This represents the expectation of the low-dimensional observation signal y; R[f(y;θ)] is the PDE residual, which is the deviation value calculated by substituting the output of the Bayesian neural network into the physical equation; The parameter inference of a Bayesian neural network satisfies: the assumed parameters Follows prior distribution Through variational inference or Markov chain Monte Carlo approximation of the posterior distribution Quantification parameter uncertainty; among which Before observation data is available, the parameters are... The probability distribution assumption; Given the observed data y and the original signal x, the parameters... The posterior probability distribution.

[0008] As a further improvement to this application, the sound wave propagation equation is expressed as:

[0009] In this equation, is the Laplace operator, representing the second-order partial derivative operation with respect to spatial coordinates; p(x,t) represents sound pressure, which is a function of spatial coordinates x=(x,y,z) and time t; c is the speed of sound, which is a constant. This represents the second-order partial derivative of sound pressure p with respect to time t; the residual of the equation is calculated by automatic differentiation and used as part of the loss function.

[0010] As a further improvement to this application, the acquisition of the transformer's acoustic signature signal includes: A microphone array is placed on the transformer casing to collect multi-channel sound pressure signals, which are then processed to obtain sparse voiceprint signals.

[0011] As a further improvement to this application, the method of reconstructing the voiceprint signal based on a Bayesian neural network using compressed sensing technology to obtain a reconstructed voiceprint field includes: Bayesian uncertainty quantification and physical equations are embedded into a Bayesian neural network. The network structure is designed as follows: input is compressed observation data and spatiotemporal coordinates, output is sound pressure / sound source intensity, and the residuals of the equations are calculated through automatic differentiation. Physical constraint integration includes: the residuals of the wave equation as part of the loss function; Bayesian parameter inference: the uncertainty of the model is quantified through variational inference or MCMC approximation of the posterior distribution of the parameters. Calculate and minimize the error between the reconstructed signal and the true signal:

[0012] In the formula, Ldata is the data-driven loss term; Ey,x represents the expectation of the low-dimensional observed signal y and the original signal x. The square of the L2 norm of the difference between the reconstructed signal f(y;θ) and the original true signal x; Construct a total loss function for the joint optimization objective, which is a weighted sum of data fitting and physical constraints:

[0013] In the formula, L `total` is the total loss function, which is driven by the data loss term. L data and physical loss items L Physics, when added together, yields the result; Compressed sensing reconstruction training is performed based on the total loss function to obtain the probability distribution of the parameters; Bayesian posterior Generate multiple network samples Calculate the mean and variance of the reconstructed signal to quantify model uncertainty; It is a parametric posterior distribution, in which multiple different parameters are sampled. Let i = 1, 2, ..., and then substitute the parameters into the Bayesian neural network. In the process, multiple reconstructed signals are obtained. The uncertainty of the model prediction is measured by calculating the mean and variance of the reconstructed signal; Finally, by utilizing sparsity priors, the complete voiceprint signal is reconstructed from a small number of measurements; by combining physical constraints and data-driven loss, the reconstruction quality is optimized to obtain the reconstructed voiceprint field.

[0014] As a further improvement to this application, the design includes a multi-task objective function encompassing data fitting, physical constraints, and localization loss, and optimizes data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network; including: Based on the sound wave propagation equation, the sound field inside the transformer is regarded as a three-dimensional sound wave propagation problem. The physical process is described by the wave equation, and the physical information is embedded into a Bayesian neural network. The continuity equation is discretized into spatiotemporal grid points, and the spatial and temporal derivatives are approximated using finite difference or spectral methods to achieve spatiotemporal discretization. The boundary conditions consider the reflection boundary conditions of the transformer enclosure, and the normal derivative is set to zero as the Neumann boundary condition; Multi-task joint training strategies include: Loss function design:

[0015] Among them, the data fitting term It is the mean square error between the observed sound pressure and the predicted sound pressure, and a physical constraint term. It is the MSE of the wave equation residuals, and the location loss term. α is the predicted Euclidean distance between the sound source location and the actual location; L is the total loss value; α, β, and γ are the data fitting terms L and L, respectively. data Physical constraint term L physics Location loss item L localization Weighting coefficients; The optimization algorithm for the multi-task joint training strategy uses the Adam or L-BFGS optimizer to dynamically adjust the learning rate to balance the multi-task objectives, and introduces dropout or weights based on regularization methods to obtain a trained Bayesian neural network.

[0016] As a further improvement to this application, the network architecture of the Bayesian neural network includes: Input layer: compressed observation vector y and spatiotemporal coordinates (x,t); y carries the voiceprint feature information after compressed sensing processing, and the spatiotemporal coordinates (x,t) are used to provide the network with contextual information of location and time; Hidden layers: Fully connected layers or convolutional layers are used to extract non-linear features; Fully connected layers connect all input and output nodes using a weight matrix, while convolutional layers use convolutional kernels to extract local features. Output layer: Predicted sound pressure level or sound source intensity , It is the sound pressure level predicted by the network. It is the predicted sound source intensity, corresponding to the actual sound pressure p(x,t) and sound source intensity q(x,t); The residuals of the wave equation are calculated by automatic differentiation, and the squared residuals are added to the loss function.

[0017] As a further improvement to this application, the method of obtaining accurate localization prediction results by combining a gradient-based sound source inversion localization algorithm and a trained Bayesian neural network with sparse regularization based on the reconstructed speaker field includes: Input compressed observation data and predict the sound field distribution using a trained PINN. Using a trained Bayesian neural network, compressed observation data is used as input to obtain the predicted spatial and temporal distribution of sound pressure. ; To predict sound source terms As an initial guess, the objective function is minimized using the conjugate gradient method: ; in, Let s be the location of the sound source, and x(s) be the sound field signal at the corresponding sound source location s. This represents the observation vector y and the vector reconstructed from the sound source location s. The squared L2 norm between the observed data and the data reconstructed from the sound source location measures the difference between them; λ is a weighting coefficient used to balance the data difference term and the physical constraint term R. physics By continuously adjusting the sound source position s using the conjugate gradient method, the objective function is minimized, thereby finding the sound source position that best matches the observation data and physical laws. By utilizing the l1 norm to constrain the sparsity of the sound source terms, we can focus on localizing a single sound source; the l1 norm is... ,in q i It is a sound source item q The value of (x,t) at discrete points; The confidence interval of the sound source location is estimated by using Bayesian PINN or Monte Carlo dropout.

[0018] As a further improvement to this application, the visualization of the prediction results onto a three-dimensional transformer model to intuitively display the location of the abnormal sound source includes: The predicted sound pressure is mapped onto a 3D model of the transformer to generate a dynamic sound field cloud map; by mapping the predicted sound pressure values... Corresponding to the three-dimensional spatial model of the transformer, the changes in sound pressure at different locations and over time are displayed in the form of cloud maps, intuitively presenting the sound field inside the transformer; The location of the abnormal sound source is displayed as a point cloud or marked points and superimposed on the transformer structural diagram. This allows for a direct view of the specific location of the sound source inside the transformer.

[0019] Secondly, the present invention provides a transformer abnormal sound source localization system, comprising: The module is used to build a Bayesian neural network that embeds the physical mechanism of voiceprints, and embeds the sound wave propagation equation as a physical constraint into the Bayesian neural network. The reconstruction module is used to reconstruct the voiceprint signal based on Bayesian neural network and compressed sensing technology to obtain the reconstructed voiceprint field. The training module is used to design a multi-task objective function that includes data fitting, physical constraints, and localization loss, and to optimize the data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network. The localization module is used to obtain accurate localization prediction results based on the reconstructed voiceprint field, the gradient-based sound source inversion localization algorithm, and the trained Bayesian neural network, combined with sparse regularization. The display module is used to visualize the prediction results onto the 3D model of the transformer, intuitively showing the location of the abnormal sound source.

[0020] As a further improvement to this application, the building module is used for: Acquire the acoustic signature signal of the transformer as the raw signal; Construct a compressed sensing framework, assuming the original signal Through linear measurement matrix Compressed into low-dimensional observations y=Φx+ Where y is the low-dimensional observation signal, This indicates that the measurement matrix Φ is an M-row N-column matrix; The noise is a vector with the same dimension as y; Next, physical information is embedded, and the sound wave propagation equation followed by signal x is embedded into the Bayesian neural network in the form of a partial differential equation. The Bayesian neural network structure is as follows: Design a Bayesian neural network Map the compressed observation y to the reconstructed signal Where y is the low-dimensional observation signal mentioned earlier, These are the network parameters, including all learnable weights and biases in a Bayesian neural network. The signal is reconstructed using a Bayesian neural network; The physical loss term is the residual of the partial differential equation. Add a loss function to ensure the reconstructed signal satisfies physical constraints: Lphysics represents the physical loss term; λ is a balancing parameter used to adjust the weight of the physical loss term in the total loss function; Ephysics y This represents the expectation of the low-dimensional observation signal y; R[f(y;θ)] is the PDE residual, which is the deviation value calculated by substituting the output of the Bayesian neural network into the physical equation; The parameter inference of a Bayesian neural network satisfies: the assumed parameters Follows prior distribution Through variational inference or Markov chain Monte Carlo approximation of the posterior distribution Quantification parameter uncertainty; among which Before observation data is available, the parameters are... The probability distribution assumption; Given the observed data y and the original signal x, the parameters... The posterior probability distribution.

[0021] As a further improvement to this application, the sound wave propagation equation is expressed as:

[0022] In this equation, is the Laplace operator, representing the second-order partial derivative operation with respect to spatial coordinates; p(x,t) represents sound pressure, which is a function of spatial coordinates x=(x,y,z) and time t; c is the speed of sound, which is a constant. This represents the second-order partial derivative of sound pressure p with respect to time t; the residual of the equation is calculated by automatic differentiation and used as part of the loss function.

[0023] As a further improvement to this application, the reconstruction module is used for: A microphone array is placed on the transformer casing to collect multi-channel sound pressure signals, which are then processed to obtain sparse voiceprint signals.

[0024] As a further improvement to this application, the method of reconstructing the voiceprint signal based on a Bayesian neural network using compressed sensing technology to obtain a reconstructed voiceprint field includes: Bayesian uncertainty quantification and physical equations are embedded into a Bayesian neural network. The network structure is designed as follows: input is compressed observation data and spatiotemporal coordinates, output is sound pressure / sound source intensity, and the residuals of the equations are calculated through automatic differentiation. Physical constraint integration includes: the residuals of the wave equation as part of the loss function; Bayesian parameter inference: the uncertainty of the model is quantified through variational inference or MCMC approximation of the posterior distribution of the parameters. Calculate and minimize the error between the reconstructed signal and the true signal:

[0025] In the formula, Ldata is the data-driven loss term; Ey,x represents the expectation of the low-dimensional observed signal y and the original signal x. The square of the L2 norm of the difference between the reconstructed signal f(y;θ) and the original true signal x; Construct a total loss function for the joint optimization objective, which is a weighted sum of data fitting and physical constraints:

[0026] In the formula, L `total` is the total loss function, which is driven by the data loss term. L data and physical loss items L Physics, when added together, yields the result; Compressed sensing reconstruction training is performed based on the total loss function to obtain the probability distribution of the parameters; Bayesian posterior Generate multiple network samples Calculate the mean and variance of the reconstructed signal to quantify model uncertainty; It is a parametric posterior distribution, in which multiple different parameters are sampled. Let i = 1, 2, ..., and then substitute the parameters into the Bayesian neural network. In the process, multiple reconstructed signals are obtained. The uncertainty of the model prediction is measured by calculating the mean and variance of the reconstructed signal; Finally, by utilizing sparsity priors, the complete voiceprint signal is reconstructed from a small number of measurements; by combining physical constraints and data-driven loss, the reconstruction quality is optimized to obtain the reconstructed voiceprint field.

[0027] As a further improvement to this application, the training module is used for: Based on the sound wave propagation equation, the sound field inside the transformer is regarded as a three-dimensional sound wave propagation problem. The physical process is described by the wave equation, and the physical information is embedded into a Bayesian neural network. The continuity equation is discretized into spatiotemporal grid points, and the spatial and temporal derivatives are approximated using finite difference or spectral methods to achieve spatiotemporal discretization. The boundary conditions consider the reflection boundary conditions of the transformer enclosure, and the normal derivative is set to zero as the Neumann boundary condition; Multi-task joint training strategies include: Loss function design:

[0028] Among them, the data fitting term It is the mean square error between the observed sound pressure and the predicted sound pressure, and a physical constraint term. It is the MSE of the wave equation residuals, and the location loss term. α is the predicted Euclidean distance between the sound source location and the actual location; L is the total loss value; α, β, and γ are the data fitting terms L and L, respectively. data Physical constraint term L physics Location loss item L localization Weighting coefficients; The optimization algorithm for the multi-task joint training strategy uses the Adam or L-BFGS optimizer to dynamically adjust the learning rate to balance the multi-task objectives, and introduces dropout or weights based on regularization methods to obtain a trained Bayesian neural network.

[0029] As a further improvement to this application, the network architecture of the Bayesian neural network includes: Input layer: compressed observation vector y and spatiotemporal coordinates (x,t); y carries the voiceprint feature information after compressed sensing processing, and the spatiotemporal coordinates (x,t) are used to provide the network with contextual information of location and time; Hidden layers: Fully connected layers or convolutional layers are used to extract non-linear features; Fully connected layers connect all input and output nodes using a weight matrix, while convolutional layers use convolutional kernels to extract local features. Output layer: Predicted sound pressure level or sound source intensity , It is the sound pressure level predicted by the network. It is the predicted sound source intensity, corresponding to the actual sound pressure p(x,t) and sound source intensity q(x,t); The residuals of the wave equation are calculated by automatic differentiation, and the squared residuals are added to the loss function.

[0030] As a further improvement to this application, the positioning module is used for: Input compressed observation data and predict the sound field distribution using a trained PINN. Using a trained Bayesian neural network, compressed observation data is used as input to obtain the predicted spatial and temporal distribution of sound pressure. ; To predict sound source terms As an initial guess, the objective function is minimized using the conjugate gradient method: ; in, Let s be the location of the sound source, and x(s) be the sound field signal at the corresponding sound source location s. This represents the observation vector y and the vector reconstructed from the sound source location s. The squared L2 norm between the observed data and the data reconstructed from the sound source location measures the difference between them; λ is a weighting coefficient used to balance the data difference term and the physical constraint term R. physics By continuously adjusting the sound source position s using the conjugate gradient method, the objective function is minimized, thereby finding the sound source position that best matches the observation data and physical laws. By utilizing the l1 norm to constrain the sparsity of the sound source terms, we can focus on localizing a single sound source; the l1 norm is... ,in q i It is a sound source item q The value of (x,t) at discrete points; The confidence interval of the sound source location is estimated by using Bayesian PINN or Monte Carlo dropout.

[0031] As a further improvement to this application, the display module is used for: The predicted sound pressure is mapped onto a 3D model of the transformer to generate a dynamic sound field cloud map; by mapping the predicted sound pressure values... Corresponding to the three-dimensional spatial model of the transformer, the changes in sound pressure at different locations and over time are displayed in the form of cloud maps, intuitively presenting the sound field inside the transformer; The location of the abnormal sound source is displayed as a point cloud or marked points and superimposed on the transformer structural diagram. This allows for a direct view of the specific location of the sound source inside the transformer.

[0032] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the transformer abnormal sound source localization method.

[0033] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the transformer abnormal sound source localization method.

[0034] Fifthly, the present invention provides a computer program product, the computer program product including computer instructions, the computer instructions instructing a computer to execute the transformer abnormal sound source localization method.

[0035] The beneficial effects of this invention are as follows: This invention proposes a transformer abnormal sound source localization method. First, a Bayesian neural network embedding the physical mechanism of acoustic signatures is constructed, with the sound wave propagation equation as a physical constraint embedded in the network. Compressed sensing technology is used to reconstruct sparse acoustic signature signals, improving data utilization efficiency. A multi-task joint training strategy is designed to simultaneously optimize data fitting, physical constraints, and localization accuracy. The Bayesian framework is used to quantify model uncertainty, providing reliable confidence interval estimation. A gradient-based sound source inversion localization algorithm, combined with sparse regularization, achieves accurate localization. The prediction results are visualized on a 3D transformer model, intuitively showing the location of the abnormal sound source. Specifically, the sound wave propagation equation is embedded as a physical constraint in the neural network to ensure that the reconstructed signal conforms to physical laws. The residual of the equation is calculated by automatic differentiation and used as part of the loss function, forcing the network output to satisfy the physical constraints. The neural network is designed to map compressed observations to the reconstructed signal, while introducing a Bayesian framework to quantify parameter uncertainty. The posterior distribution of parameters is approximated through variational inference or Markov chain Monte Carlo methods, utilizing sparsity priors to reconstruct the complete acoustic signature signal from a small number of measurements. The reconstruction quality is optimized by combining physical constraints and data-driven loss. The design incorporates a multi-task objective function that integrates data fitting, physical constraints, and localization loss to balance different task objectives and improve the overall model performance. Based on a pre-trained physical information neural network, the inverse problem is solved using gradient optimization, combined with sparse regularization to achieve accurate localization. Simultaneously, a Bayesian framework is used to estimate the confidence interval of the localization results. The predicted sound pressure and localization results are mapped onto a 3D transformer model, generating dynamic sound field cloud maps and abnormal sound source markers to visually display the localization results. In this way, the proposed method effectively combines prior physical knowledge, Bayesian uncertainty quantification, and data-driven deep learning, overcoming the limitations of traditional methods in complex environments and achieving high-precision, highly robust transformer abnormal sound source localization. Attached Figure Description

[0036] Figure 1 A flowchart of a transformer abnormal sound source localization method provided by the present invention; Figure 2 A compressed sensing reconstruction method that enhances physical information neural networks; Figure 3This is an abnormal sound source localization method based on physical information neural networks; Figure 4 This invention provides a transformer abnormal sound source localization system; Figure 5 This is a schematic diagram of an electronic device provided by the present invention. Detailed Implementation

[0037] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] like Figure 1 As shown, this invention constructs a method for locating abnormal sound sources in transformers. It uses a Bayesian Physical Information Bayesian Neural Network (B-PINN) with embedded sound wave propagation equations to deduce the sound wave propagation process inside the transformer, and combines compressed sensing technology to achieve efficient inversion of the sound source location. This method of compressed sensing reconstruction using a Bayesian Physical Information Bayesian Neural Network combines prior physical knowledge, Bayesian uncertainty quantification, and data-driven deep learning, effectively improving the accuracy and robustness of compressed sensing in complex signal reconstruction.

[0039] This paper proposes a method to quantify Bayesian uncertainty and embed physical equations into a Bayesian neural network. Inputting compressed observation data and spatiotemporal coordinates, the network outputs sound pressure intensity, while simultaneously calculating the residuals of the equations through automatic differentiation. The residuals of the sound pressure wave equation are used as part of the loss function, forcing the network output to conform to physical laws. Variational inference of the posterior distribution of approximate parameters quantifies model uncertainty. Combined with compressed sensing algorithms, the relatively sparse acoustic signature signals are reconstructed to obtain denser spatial acoustic signature signals. A loss function constructed using the acoustic signature propagation equation is embedded in the Bayesian neural network, and physical constraints suppress noise and multipath reflection interference, improving localization robustness. Physical constraints require the network output to conform to the wave equation, reducing reliance on large-scale labeled data. This allows the network to accurately learn features consistent with physical mechanisms even with less labeled data, enabling effective operation in complex transformer industrial scenarios.

[0040] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0041] Example 1 like Figure 1 As shown, this invention proposes a method for locating abnormal sound sources in transformers, and the main implementation route is as follows: Figure 2 and Figure 3 As shown, the technical approach is as follows: First, a Bayesian neural network embedding the physical mechanism of voiceprints is constructed. Network parameters are inferred by combining Bayesian parameters. The relatively sparse voiceprint signals are reconstructed by comprehensively considering data fitting and physical constraints. Second, confidence interval analysis is performed using Bayesian PINN and Monte Carlo dropout to estimate the sound source location. Finally, the predicted sound pressure is mapped onto a 3D transformer model, displaying the location of abnormal sound sources in the form of point clouds or marked points. The main implementation steps include: Step 1: Construct a Bayesian neural network that embeds the physical mechanism of voiceprints, and embed the sound wave propagation equation as a physical constraint into the Bayesian neural network. 1.1 Voiceprint Physical Constraint Embedding: First, a microphone array (e.g., 8-16 sensors) is placed on the transformer casing. Multi-channel sound pressure signals are collected through the microphone array and processed to obtain relatively sparse voiceprint signals.

[0042] Construct a compressed sensing framework, assuming the original signal Through linear measurement matrix Compressed into low-dimensional observations y=Φx+ Where y is the low-dimensional observation signal, This indicates that the measurement matrix Φ is an M-row N-column matrix; The noise is a vector with the same dimension as y; Next, physical information is embedded, whereby the physical laws governing the signal x (such as the sound wave propagation equation) are embedded into the Bayesian neural network in the form of partial differential equations (PDEs). For example, the sound wave propagation equation can be expressed as:

[0043] In this equation, is the Laplace operator, representing the second-order partial derivative operation with respect to spatial coordinates; p(x,t) represents sound pressure, which is a function of spatial coordinates x=(x,y,z) (three-dimensional spatial coordinates) and time t; c is the speed of sound, which is a constant; Let p represent the second-order partial derivative of the sound pressure p with respect to time t. The residual of the equation is calculated using automatic differentiation and used as part of the loss function.

[0044] 1.2 Bayesian Physical Information Bayesian Neural Network Structure Network Architecture: Designing Bayesian Neural Networks Map the compressed observation y to the reconstructed signal Where y is the low-dimensional observation signal mentioned earlier, These are the network parameters, which include all learnable weights and biases in a Bayesian neural network. The signal is obtained by reconstructing it through a Bayesian neural network.

[0045] Physical loss term: This refers to the PDE residuals. Incorporate a loss function to ensure the reconstructed signal satisfies physical constraints: Here, Lphysics represents the physical loss term; λ is the balance parameter used to adjust the weight of the physical loss term in the total loss function; Ey represents the expectation of the low-dimensional observation signal y; R[f(y;θ)] is the PDE residual, which is the bias value calculated by substituting the output of the Bayesian neural network into the physical equation.

[0046] Bayesian parameter inference: Hypothesis parameters Follows prior distribution Approximate the posterior distribution through variational inference (VI) or Markov chain Monte Carlo (MCMC). The uncertainty of quantification parameters. Before observation data is available, the parameters are... The probability distribution assumption; Given the observed data y and the original signal x, the parameters... The posterior probability distribution.

[0047] Step 2: Based on Bayesian neural networks, compressed sensing technology is used to reconstruct the voiceprint signal to obtain the reconstructed voiceprint field. 2.1 Compressed sensing reconstruction training includes: Data-driven loss: Minimizes the error (e.g., mean square error) between the reconstructed signal and the true signal.

[0048] L_data represents the data-driven loss term; E_y,x represents the expectation of the low-dimensional observed signal y and the original signal x. The mean square error is the square of the L2 norm of the difference between the reconstructed signal f(y;θ) and the original true signal x.

[0049] Joint optimization objective: The total loss function is a weighted sum of data fitting and physical constraints.

[0050] L `total` is the total loss function, which is driven by the data loss term. L data and physical loss items L The physics terms are obtained by adding them together, and joint optimization is achieved by adjusting the weights of the two terms (the weights of the physics loss terms are controlled by λ).

[0051] Bayesian training: variational inference is performed by maximizing the lower bound of evidence (ELBO), or posterior sampling is performed using a Bayesian deep learning framework (such as Pyro) to obtain the probability distribution of the parameters.

[0052] 2.2 Quantification and Reconstruction of Uncertainty: Predicting Uncertainty: Through Bayesian Posterior Generate multiple network samples The mean and variance of the reconstructed signal are calculated to quantify the model uncertainty. Here... It is the posterior distribution of the parameters mentioned earlier, from which multiple different parameters are sampled. (i=1,2,…), and then these parameters are substituted into the Bayesian neural network. In the process, multiple reconstructed signals are obtained. The uncertainty of the model prediction is measured by calculating the mean and variance of these reconstructed signals.

[0053] Enhanced physical consistency: By utilizing physical constraints, reconstruction results that do not conform to physical laws are suppressed, thereby improving the reconstruction quality of sparse signals or complex scenarios.

[0054] Compressed measurement: Acoustic fingerprint signals are acquired using a small number of distributed probes. These probes, located at different positions on the transformer, acquire the acoustic fingerprint signals generated during transformer operation; these signals will subsequently serve as the raw data for compressed sensing.

[0055] Physical constraints: The sound wave propagation equation is embedded into the network to ensure that the reconstructed signal conforms to physical laws. Inside the transformer, the propagation of the acoustic signature signal follows the sound wave propagation equation. This equation is embedded into the training process of the Bayesian neural network, so that the reconstructed acoustic signature signal is physically reasonable, such as the propagation speed and attenuation characteristics of sound pressure conforming to actual physical conditions.

[0056] Step 3: Design a multi-task objective function that includes data fitting, physical constraints, and localization loss, and optimize the data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network; 3.1 Sound Field Physical Modeling and Equation Discretization: Similar to step one, the following sound wave propagation equation is used to treat the sound field inside the transformer as a three-dimensional sound wave propagation problem, and the physical process is described by the wave equation.

[0057]

[0058] 3.2 Spatiotemporal Discretization: The continuous equation is discretized into spatiotemporal grid points, and the spatial and temporal derivatives are approximated using finite difference or spectral methods. This step is to transform the continuous wave equation into a discrete form that can be processed by a computer. The finite difference method approximates the derivative by using differences at the grid points, while the spectral method approximates the derivative based on the orthogonal expansion of the function. 3.3 Boundary Conditions: Considering the reflection boundary conditions of the transformer enclosure, Neumann boundary conditions (normal derivative is zero) are set. This means that the normal variation rate of sound pressure is zero at the boundary of the transformer enclosure, reflecting the reflection characteristics of sound waves at the enclosure boundary.

[0059] 3.4 Constructing a Bayesian Neural Network for Physical Information Network architecture design: Input layer: compressed observation vector y and spatiotemporal coordinates (x,t). y carries the voiceprint feature information after compressed sensing processing, and the spatiotemporal coordinates (x,t) are used to provide the network with location and time context information.

[0060] Hidden layers: Fully connected layers or convolutional layers (such as ResNet structure) are used to extract non-linear features.

[0061] Fully connected layers connect all input and output nodes through weight matrices, while convolutional layers use convolutional kernels to extract local features. The ResNet structure solves the gradient vanishing problem in the training of deep Bayesian neural networks through residual connections, effectively extracting complex nonlinear features.

[0062] Output layer: Predicted sound pressure level or sound source intensity . It is the sound pressure level predicted by the network. These are the predicted sound source intensities, which correspond to the actual sound pressure p(x,t) and sound source intensity q(x,t), and are used for subsequent loss calculations and result evaluation.

[0063] Physical constraint embedding: The residuals of the wave equation are calculated using automatic differentiation: By adding the squared residuals to the loss function, the network output is forced to conform to physical laws.

[0064] Where R represents the residual of the wave equation. For predicting sound pressure Perform Laplace operator operations. It is the product of the second-order partial derivative of the predicted sound pressure with respect to time and the reciprocal of the square of the sound speed. The residual represents the predicted sound source intensity. This residual reflects the degree of deviation between the network prediction and the actual physical wave equation.

[0065] By incorporating the squared residuals into the loss function, the network output is forced to conform to physical laws. In this way, the Bayesian neural network not only fits the observed data during training but also follows the physical wave equation, improving the accuracy and reasonableness of predictions.

[0066] 3.5 The multi-task joint training strategy is as follows: Loss function design:

[0067] Data fitting term Mean square error (MSE) between observed and predicted sound pressure.

[0068] Physical constraints : MSE of the residuals of the wave equation.

[0069] Location loss item : Predict the Euclidean distance between the sound source location and the actual location.

[0070] In this loss function, L is the total loss value. α, β, and γ are the weighting coefficients for the data fitting term L_data, the physical constraint term L_physics, and the localization loss term L_localization, respectively, used to adjust the relative importance of these three terms in the total loss.

[0071] Data fitting term L data The mean square error (MSE) between the observed and predicted sound pressure levels. , where pi is the actual observed sound pressure value, p^i is the sound pressure value predicted by the network, n is the number of data samples, and MSE is used to measure the closeness between the network-predicted sound pressure and the actual observed sound pressure.

[0072] Physical constraints L physics The MSE of the wave equation residuals. That is... ,in It is the first i The residual of the wave equation corresponding to each sample is minimized to make the network output more consistent with physical laws.

[0073] Location loss item L localization : The predicted Euclidean distance between the sound source location and the actual location. Let the predicted sound source location be... If the actual sound source location is (x, y, z), then It is used to measure the accuracy of sound source localization.

[0074] Optimization Algorithms: The Adam or L-BFGS optimizer is used to dynamically adjust the learning rate to balance multiple task objectives. The Adam optimizer combines the advantages of Adagrad and RMSProp and can adaptively adjust the learning rate; L-BFGS is a quasi-Newton method suitable for large-scale optimization problems. By dynamically adjusting the learning rate, it enables the network to better balance the three task objectives of data fitting, physical constraints, and localization accuracy at different stages.

[0075] Regularization techniques: Introducing dropout or weight decay prevents overfitting and enhances the model's generalization ability. Dropout randomly discards some neurons during training to avoid excessive dependence between neurons; weight decay adds a penalty term to the weight parameters to prevent the weights from becoming too large. Both help improve the model's generalization ability on new data.

[0076] Step 4: Based on the reconstructed speaker field, a gradient-based sound source inversion localization algorithm and a trained Bayesian neural network, combined with sparse regularization, are used to obtain accurate localization prediction results; including: 4.1 Sound Source Inversion and Localization Algorithm Forward propagation prediction: Input compressed observation data and predict the sound field distribution using a trained PINN. Using a trained Bayesian neural network, compressed observation data is used as input to obtain the predicted spatial and temporal distribution of sound pressure. This provides a foundation for subsequent sound source localization.

[0077] Solving the inverse problem: Gradient-based optimization: to predict sound source terms As an initial guess, the objective function is minimized using the conjugate gradient method:

[0078] in Let s be the location of the sound source, and x(s) be the sound field signal at the corresponding sound source location s. This represents the observation vector y and the vector reconstructed from the sound source location s. The squared L2 norm between the observed data and the data reconstructed from the sound source location measures the difference between them; λ is a weighting coefficient used to balance the data difference term and the physical constraint term R. physics By continuously adjusting the sound source position s using the conjugate gradient method, the objective function is minimized, thereby finding the sound source position that best matches the observation data and physical laws.

[0079] Sparsity regularization: Utilizing the l1 norm to constrain the sparsity of the sound source terms, focusing on the localization of a single sound source. The l1 norm is... ,in q i It is a sound source itemq The value of (x,t) at discrete points, by applying the l1 norm constraint to the sound source term, makes it more likely to identify the main single sound source in complex situations with multiple sound sources, thus improving the accuracy of localization.

[0080] 4.2 Uncertainty Quantification: The confidence interval of the sound source location is estimated using Bayesian PINN or Monte Carlo dropout. Bayesian PINN considers the uncertainty of parameters by probabilistically modeling the network parameters; Monte Carlo dropout estimates the confidence interval of the sound source location by randomly dropping neurons multiple times and statistically analyzing the variance of the prediction results, thus providing a reliability assessment for the localization results.

[0081] Step 5: Visualize the prediction results onto the 3D model of the transformer to intuitively show the location of the abnormal sound source.

[0082] 3D sound field reconstruction: Mapping the predicted sound pressure level to a 3D transformer model generates a dynamic sound field cloud map. This is achieved by reconstructing the predicted sound pressure level. The changes in sound pressure at different locations over time are displayed in the form of cloud maps on the three-dimensional spatial model of the transformer, providing an intuitive view of the sound field inside the transformer. Sound source localization output: Displays the location of abnormal sound sources as point clouds or marked points, overlaid on the transformer structural diagram. The located abnormal sound source locations are displayed on the transformer structural drawings as point clouds (represented by multiple discrete points) or marked points (represented by a single prominent point), facilitating a direct and intuitive view of the specific location of the sound source inside the transformer.

[0083] For example: Transformer partial discharge location: The discharge point is located using high-frequency acoustic signature signals (100 kHz - 1 MHz). Partial discharge generates acoustic signature signals within a specific frequency range. By analyzing and processing these high-frequency acoustic signatures using this technology, the location of the discharge point can be accurately found, allowing for the timely detection of potential internal faults in the transformer.

[0084] Mechanical Fault Diagnosis: Identifying abnormal sound sources such as loose core and winding deformation. Different mechanical faults can cause transformers to produce abnormal sound sources with different characteristics. By analyzing the collected acoustic signals and combining the feature extraction and localization algorithms in this technology, it is possible to determine which type of mechanical fault, such as loose core or winding deformation, is present and pinpoint the location of the fault.

[0085] Concealed sound source detection: Locating fault points in enclosed enclosures that are difficult for traditional sensors to reach. As transformers are enclosed equipment, some fault points may be difficult to detect directly with traditional sensors. However, this technology utilizes the propagation characteristics of acoustic signals to penetrate the enclosure and locate these concealed fault points.

[0086] This technology provides a non-intrusive, high-precision solution for transformer condition monitoring, significantly reducing operation and maintenance costs and improving power system reliability. This invention optimizes the model by analyzing and comparing verification results, addressing its advantages and disadvantages, and selecting the optimal training method to meet the comprehensive requirements of practical applications for high accuracy, strong robustness, and good adaptability.

[0087] In online monitoring systems of large power transformers, anomaly source localization methods based on Bayesian physical information neural networks and compressed sensing reconstruction can be effectively applied. This method can accurately locate faults such as partial discharge in the complex internal structure of transformers, providing real-time monitoring and early warning. However, in practical applications, changes in transformer oil temperature significantly affect the propagation speed and attenuation characteristics of sound waves. Due to uneven temperature distribution, the sound wave propagation characteristics differ in different regions inside the transformer. This temperature gradient effect causes the sound wave propagation path to bend, thus affecting the localization accuracy. Traditional methods typically assume a constant sound velocity, which is difficult to adapt to such complex temperature distribution environments, and therefore may produce large localization errors in practical applications. To address the impact of temperature gradients on sound source localization accuracy, an adaptive temperature compensation sound source localization optimization method is proposed. This method mainly includes the following steps: 1. Temperature field modeling: Multiple temperature sensors are placed inside the transformer to collect temperature data at different locations in real time. An interpolation algorithm (such as Kriging interpolation) is used to construct a three-dimensional temperature field model of the entire transformer interior.

[0088] 2. Sound velocity correction: Based on the known temperature-sound velocity relationship, the temperature field model is converted into a sound velocity distribution model. This allows us to obtain the local sound velocity at each point inside the transformer.

[0089] 3. Ray Tracing Algorithm: Ray tracing technology is used to simulate the propagation path of sound waves in a non-uniform medium. Considering the bending effect of sound waves under temperature gradients, the Runge-Kutta method is used to solve the ray equation to obtain the actual propagation path and time from the sound source to each sensor.

[0090] 4. Physical Constraint Update: The modified sound wave propagation model is embedded as a new physical constraint into the Bayesian physical information neural network. Specifically, the sound wave propagation equation in the network is updated to take into account the spatial variation of the speed of sound.

[0091] 5. Adaptive Training Strategy: An adaptive training strategy is designed to dynamically adjust network parameters based on the real-time temperature field. Before each localization, the network is quickly fine-tuned using the latest temperature data to adapt to the current temperature distribution.

[0092] 6. Iterative Optimization Localization: During the sound source inversion and localization process, iterative optimization is performed by incorporating temperature field information. Each iteration updates the sound source location estimate and recalculates the sound wave propagation path based on the latest location until convergence.

[0093] This adaptive temperature compensation method effectively overcomes the influence of temperature gradients on sound source localization. It not only considers the spatial variation of sound velocity but also dynamically adjusts the model using real-time temperature data, making the localization results more accurate and reliable. This optimization scheme fully utilizes existing temperature monitoring systems, requires no additional complex equipment, and is highly practical and feasible.

[0094] Example 2 like Figure 4 As shown, a transformer abnormal sound source localization system includes: Module 100 is used to construct a Bayesian neural network that embeds the physical mechanism of voiceprints, and embeds the sound wave propagation equation as a physical constraint into the Bayesian neural network. The reconstruction module 200 is used to reconstruct the voiceprint signal based on Bayesian neural network and compressed sensing technology to obtain the reconstructed voiceprint field. Training module 300 is used to design a multi-task objective function that includes data fitting, physical constraints and localization loss, and to optimize data fitting, physical constraints and localization accuracy to obtain a trained Bayesian neural network. The localization module 400 is used to obtain accurate localization prediction results based on the gradient-based sound source inversion localization algorithm and the trained Bayesian neural network, combined with sparse regularization, on the basis of the reconstructed voiceprint field. The display module 500 is used to visualize the prediction results onto the 3D model of the transformer, intuitively showing the location of the abnormal sound source.

[0095] This system combines compressed sensing dimensionality reduction with prior physical knowledge to improve the accuracy of sparse signal reconstruction. Compressed sensing preprocessing: STFT frequency domain feature extraction and random Gaussian matrix dimensionality reduction. Physically guided regularization: Wave equation constraints suppress non-physical reconstruction results, overcoming the performance bottleneck of traditional compressed sensing; Bayesian uncertainty quantization and physical equations are embedded into a Bayesian neural network to achieve compressed sensing reconstruction. Network structure design: Input compressed observation data and spatiotemporal coordinates, output sound pressure / sound source intensity, and simultaneously calculate the equation residuals through automatic differentiation. Physical constraint integration: The wave equation residuals are used as part of the loss function, forcing the network output to conform to physical laws. Bayesian parameter inference: Model uncertainty is quantified through variational inference or MCMC approximation of the posterior distribution of parameters. Innovation of the gradient-based sound source localization inverse problem algorithm: Combining conjugate gradient optimization and L1 sparse regularization to achieve accurate localization of a single sound source. This includes the design of the inverse problem objective function. Sparse constraint strength The dynamic adjustment strategy.

[0096] Example 3 like Figure 5 As shown, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the following method for locating transformer abnormal sound sources: A Bayesian neural network with embedded voiceprint physical mechanism is constructed, and the sound wave propagation equation is embedded into the Bayesian neural network as a physical constraint. Based on Bayesian neural networks, compressed sensing technology is used to reconstruct the voiceprint signal to obtain the reconstructed voiceprint field. Design a multi-task objective function that includes data fitting, physical constraints, and localization loss, and optimize the data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network. Based on the reconstructed voiceprint field, a gradient-based sound source inversion localization algorithm and a trained Bayesian neural network are combined with sparse regularization to obtain accurate localization prediction results. The prediction results are visualized onto a 3D model of the transformer, which intuitively shows the location of the abnormal sound source.

[0097] Example 4 A fourth objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned transformer abnormal sound source localization method. The aforementioned transformer abnormal sound source localization method includes: A Bayesian neural network with embedded voiceprint physical mechanism is constructed, and the sound wave propagation equation is embedded into the Bayesian neural network as a physical constraint. Based on Bayesian neural networks, compressed sensing technology is used to reconstruct the voiceprint signal to obtain the reconstructed voiceprint field. Design a multi-task objective function that includes data fitting, physical constraints, and localization loss, and optimize the data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network. Based on the reconstructed voiceprint field, a gradient-based sound source inversion localization algorithm and a trained Bayesian neural network are combined with sparse regularization to obtain accurate localization prediction results. The prediction results are visualized onto a 3D model of the transformer, which intuitively shows the location of the abnormal sound source.

[0098] Example 5 A fifth objective of this invention is to provide a computer program product comprising computer instructions that instruct a computer to execute the aforementioned transformer abnormal sound source localization method. The aforementioned transformer abnormal sound source localization method includes: A Bayesian neural network with embedded voiceprint physical mechanism is constructed, and the sound wave propagation equation is embedded into the Bayesian neural network as a physical constraint. Based on Bayesian neural networks, compressed sensing technology is used to reconstruct the voiceprint signal to obtain the reconstructed voiceprint field. Design a multi-task objective function that includes data fitting, physical constraints, and localization loss, and optimize the data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network. Based on the reconstructed voiceprint field, a gradient-based sound source inversion localization algorithm and a trained Bayesian neural network are combined with sparse regularization to obtain accurate localization prediction results. The prediction results are visualized onto a 3D model of the transformer, which intuitively shows the location of the abnormal sound source.

[0099] These 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 function specified in one or more boxes.

[0100] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment 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.

[0101] This invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, readable storage media, optical storage, etc.) containing computer-usable program code.

[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for locating abnormal sound sources in a transformer, characterized in that, include: A Bayesian neural network with embedded voiceprint physical mechanism is constructed, and the sound wave propagation equation is embedded into the Bayesian neural network as a physical constraint. Based on Bayesian neural networks, compressed sensing technology is used to reconstruct the voiceprint signal to obtain the reconstructed voiceprint field. Design a multi-task objective function that includes data fitting, physical constraints, and localization loss, and optimize the data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network. Based on the reconstructed voiceprint field, a gradient-based sound source inversion localization algorithm and a trained Bayesian neural network are combined with sparse regularization to obtain accurate localization prediction results. The prediction results are visualized onto a 3D model of the transformer, which intuitively shows the location of the abnormal sound source.

2. The method for locating abnormal sound sources in a transformer according to claim 1, characterized in that, The construction of a Bayesian neural network embedding the physical mechanism of voiceprints, using the sound wave propagation equation as a physical constraint embedded in the Bayesian neural network, includes: Acquire the acoustic signature signal of the transformer as the raw signal; Construct a compressed sensing framework, assuming the original signal Through linear measurement matrix Compressed into low-dimensional observations y=Φx+ Where y is the low-dimensional observation signal, This indicates that the measurement matrix Φ is an M-row N-column matrix; The noise is a vector with the same dimension as y; Next, physical information is embedded, and the sound wave propagation equation followed by signal x is embedded into the Bayesian neural network in the form of a partial differential equation. The Bayesian neural network structure is as follows: Design a Bayesian neural network Map the compressed observation y to the reconstructed signal Where y is the low-dimensional observation signal mentioned earlier, These are the network parameters, including all learnable weights and biases in a Bayesian neural network. The signal is reconstructed using a Bayesian neural network; The physical loss term is the residual of the partial differential equation. Add a loss function to ensure the reconstructed signal satisfies physical constraints: Lphysics represents the physical loss term; λ is a balancing parameter used to adjust the weight of the physical loss term in the total loss function; Ephysics y This represents the expectation of the low-dimensional observation signal y; R[f(y;θ)] is the PDE residual, which is the deviation value calculated by substituting the output of the Bayesian neural network into the physical equation; The parameter inference of a Bayesian neural network satisfies: the assumed parameters Follows prior distribution Through variational inference or Markov chain Monte Carlo approximation of the posterior distribution Quantification parameter uncertainty; among which Before observation data is available, the parameters are... The probability distribution assumption; Given the observed data y and the original signal x, the parameters... The posterior probability distribution.

3. The method for locating abnormal sound sources in a transformer according to claim 2, characterized in that, The sound wave propagation equation is expressed as: In this equation, is the Laplace operator, representing the second-order partial derivative operation with respect to spatial coordinates; p(x,t) represents sound pressure, which is a function of spatial coordinates x=(x,y,z) and time t; c is the speed of sound, which is a constant. This represents the second-order partial derivative of sound pressure p with respect to time t; the residual of the equation is calculated by automatic differentiation and used as part of the loss function.

4. The method for locating abnormal sound sources in a transformer according to claim 2, characterized in that, The acquisition of the transformer's acoustic signature signal includes: A microphone array is placed on the transformer casing to collect multi-channel sound pressure signals, which are then processed to obtain sparse voiceprint signals.

5. The method for locating abnormal sound sources in a transformer according to claim 2, characterized in that, The method of reconstructing the voiceprint signal using compressed sensing technology based on a Bayesian neural network to obtain the reconstructed voiceprint field includes: Bayesian uncertainty quantification and physical equations are embedded into a Bayesian neural network. The network structure is designed as follows: input is compressed observation data and spatiotemporal coordinates, output is sound pressure / sound source intensity, and the residuals of the equations are calculated through automatic differentiation. Physical constraint integration includes: the residuals of the wave equation as part of the loss function; Bayesian parameter inference: the uncertainty of the model is quantified through variational inference or MCMC approximation of the posterior distribution of the parameters. Calculate and minimize the error between the reconstructed signal and the true signal: In the formula, Ldata is the data-driven loss term; Ey,x represents the expectation of the low-dimensional observed signal y and the original signal x. The square of the L2 norm of the difference between the reconstructed signal f(y;θ) and the original true signal x; Construct a total loss function for the joint optimization objective, which is a weighted sum of data fitting and physical constraints: In the formula, L `total` is the total loss function, which is driven by the data loss term. L data and physical loss items L Physics, when added together, yields the result; Compressed sensing reconstruction training is performed based on the total loss function to obtain the probability distribution of the parameters; Bayesian posterior Generate multiple network samples Calculate the mean and variance of the reconstructed signal to quantify model uncertainty; It is a parametric posterior distribution, in which multiple different parameters are sampled. Let i = 1, 2, ..., and then substitute the parameters into the Bayesian neural network. In the process, multiple reconstructed signals are obtained. The uncertainty of the model prediction is measured by calculating the mean and variance of the reconstructed signal; Finally, by utilizing sparsity priors, the complete voiceprint signal is reconstructed from a small number of measurements; by combining physical constraints and data-driven loss, the reconstruction quality is optimized to obtain the reconstructed voiceprint field.

6. The method for locating abnormal sound sources in a transformer according to claim 1, characterized in that, The design includes a multi-task objective function encompassing data fitting, physical constraints, and localization loss, and optimizes data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network; including: Based on the sound wave propagation equation, the sound field inside the transformer is regarded as a three-dimensional sound wave propagation problem. The physical process is described by the wave equation, and the physical information is embedded into a Bayesian neural network. The continuity equation is discretized into spatiotemporal grid points, and the spatial and temporal derivatives are approximated using finite difference or spectral methods to achieve spatiotemporal discretization. The boundary conditions consider the reflection boundary conditions of the transformer enclosure, and the normal derivative is set to zero as the Neumann boundary condition; Multi-task joint training strategies include: Loss function design: Among them, the data fitting term It is the mean square error between the observed sound pressure and the predicted sound pressure, and a physical constraint term. It is the MSE of the wave equation residuals, and the location loss term. α is the predicted Euclidean distance between the sound source location and the actual location; L is the total loss value; α, β, and γ are the data fitting terms L and L, respectively. data Physical constraint term L physics Location loss item L localization Weighting coefficients; The optimization algorithm for the multi-task joint training strategy uses the Adam or L-BFGS optimizer to dynamically adjust the learning rate to balance the multi-task objectives, and introduces dropout or weights based on regularization methods to obtain a trained Bayesian neural network.

7. The transformer abnormal sound source localization method according to claim 6, characterized in that, The network architecture of the Bayesian neural network includes: Input layer: compressed observation vector y and spatiotemporal coordinates (x,t); y carries the voiceprint feature information after compressed sensing processing, and the spatiotemporal coordinates (x,t) are used to provide the network with contextual information of location and time; Hidden layers: Fully connected layers or convolutional layers are used to extract non-linear features; Fully connected layers connect all input and output nodes using a weight matrix, while convolutional layers use convolutional kernels to extract local features. Output layer: Predicted sound pressure level or sound source intensity , It is the sound pressure level predicted by the network. It is the predicted sound source intensity, corresponding to the actual sound pressure p(x,t) and sound source intensity q(x,t); The residuals of the wave equation are calculated by automatic differentiation, and the squared residuals are added to the loss function.

8. The method for locating abnormal sound sources in a transformer according to claim 1, characterized in that, Based on the reconstructed voiceprint field, a gradient-based sound source inversion localization algorithm and a trained Bayesian neural network are combined with sparse regularization to obtain accurate localization prediction results. include: Input compressed observation data and predict the sound field distribution using a trained PINN. Using a trained Bayesian neural network, compressed observation data is used as input to obtain the predicted spatial and temporal distribution of sound pressure. ; To predict sound source terms As an initial guess, the objective function is minimized using the conjugate gradient method: ； in, Let s be the location of the sound source, and x(s) be the sound field signal at the corresponding sound source location s. This represents the observation vector y and the vector reconstructed from the sound source location s. The squared L2 norm between the observed data and the data reconstructed from the sound source location measures the difference between the observed data and the data reconstructed from the sound source location; λ is a weighting coefficient used to balance the data difference term and the physical constraint term R. physics By continuously adjusting the sound source position s using the conjugate gradient method, the objective function is minimized, thereby finding the sound source position that best matches the observation data and physical laws. By utilizing the l1 norm to constrain the sparsity of the sound source terms, we can focus on localizing a single sound source; the l1 norm is... ,in q i It is a sound source item q The value of (x,t) at discrete points; The confidence interval of the sound source location is estimated by using Bayesian PINN or Monte Carlo dropout.

9. The method for locating abnormal sound sources in a transformer according to claim 1, characterized in that, The process of visualizing the prediction results onto a 3D model of the transformer to intuitively display the location of the abnormal sound source includes: The predicted sound pressure is mapped onto a 3D model of the transformer to generate a dynamic sound field cloud map; by mapping the predicted sound pressure values... Corresponding to the three-dimensional spatial model of the transformer, the changes in sound pressure at different locations and over time are displayed in the form of cloud maps, intuitively presenting the sound field inside the transformer; The location of the abnormal sound source is displayed as a point cloud or marked points and superimposed on the transformer structural diagram. This allows for a direct view of the specific location of the sound source inside the transformer.

10. A transformer abnormal sound source localization system, characterized in that, include: The module is used to build a Bayesian neural network that embeds the physical mechanism of voiceprints, and embeds the sound wave propagation equation as a physical constraint into the Bayesian neural network. The reconstruction module is used to reconstruct the voiceprint signal based on Bayesian neural network and compressed sensing technology to obtain the reconstructed voiceprint field. The training module is used to design a multi-task objective function that includes data fitting, physical constraints, and localization loss, and to optimize the data fitting, physical constraints, and localization accuracy to obtain a trained Bayesian neural network. The localization module is used to obtain accurate localization prediction results based on the reconstructed voiceprint field, the gradient-based sound source inversion localization algorithm, and the trained Bayesian neural network, combined with sparse regularization. The display module is used to visualize the prediction results onto the 3D model of the transformer, intuitively showing the location of the abnormal sound source.

11. The transformer abnormal sound source localization system according to claim 10, characterized in that, The building module is used for: Acquire the acoustic signature signal of the transformer as the raw signal; Construct a compressed sensing framework, assuming the original signal Through linear measurement matrix Compressed into low-dimensional observations y=Φx+ Where y is the low-dimensional observation signal, This indicates that the measurement matrix Φ is an M-row N-column matrix; The noise is a vector with the same dimension as y; Next, physical information is embedded, and the sound wave propagation equation followed by signal x is embedded into the Bayesian neural network in the form of a partial differential equation. The Bayesian neural network structure is as follows: Design a Bayesian neural network Map the compressed observation y to the reconstructed signal Where y is the low-dimensional observation signal mentioned earlier, These are the network parameters, including all learnable weights and biases in a Bayesian neural network. The signal is reconstructed using a Bayesian neural network; The physical loss term is the residual of the partial differential equation. Add a loss function to ensure the reconstructed signal satisfies physical constraints: Lphysics represents the physical loss term; λ is a balancing parameter used to adjust the weight of the physical loss term in the total loss function; Ephysics y This represents the expectation of the low-dimensional observation signal y; R[f(y;θ)] is the PDE residual, which is the deviation value calculated by substituting the output of the Bayesian neural network into the physical equation; The parameter inference of a Bayesian neural network satisfies: the assumed parameters Follows prior distribution Through variational inference or Markov chain Monte Carlo approximation of the posterior distribution Quantification parameter uncertainty; among which Before observation data is available, the parameters are... The probability distribution assumption; Given the observed data y and the original signal x, the parameters... The posterior probability distribution.

12. The transformer abnormal sound source localization system according to claim 11, characterized in that, The sound wave propagation equation is expressed as: In this equation, is the Laplace operator, representing the second-order partial derivative operation with respect to spatial coordinates; p(x,t) represents sound pressure, which is a function of spatial coordinates x=(x,y,z) and time t; c is the speed of sound, which is a constant. This represents the second-order partial derivative of sound pressure p with respect to time t; the residual of the equation is calculated by automatic differentiation and used as part of the loss function.

13. The transformer abnormal sound source localization system according to claim 11, characterized in that, The reconstruction module is used for: A microphone array is placed on the transformer casing to collect multi-channel sound pressure signals, which are then processed to obtain sparse voiceprint signals.

14. The transformer abnormal sound source localization system according to claim 10, characterized in that, The method of reconstructing the voiceprint signal using compressed sensing technology based on a Bayesian neural network to obtain the reconstructed voiceprint field includes: Bayesian uncertainty quantification and physical equations are embedded into a Bayesian neural network. The network structure is designed as follows: input is compressed observation data and spatiotemporal coordinates, output is sound pressure / sound source intensity, and the residuals of the equations are calculated through automatic differentiation. Physical constraint integration includes: the residuals of the wave equation as part of the loss function; Bayesian parameter inference: the uncertainty of the model is quantified through variational inference or MCMC approximation of the posterior distribution of the parameters. Calculate and minimize the error between the reconstructed signal and the true signal: In the formula, Ldata is the data-driven loss term; Ey,x represents the expectation of the low-dimensional observed signal y and the original signal x. The square of the L2 norm of the difference between the reconstructed signal f(y;θ) and the original true signal x; Construct a total loss function for the joint optimization objective, which is a weighted sum of data fitting and physical constraints: In the formula, L `total` is the total loss function, which is driven by the data loss term. L data and physical loss items L Physics, when added together, yields the result; Compressed sensing reconstruction training is performed based on the total loss function to obtain the probability distribution of the parameters; Bayesian posterior Generate multiple network samples Calculate the mean and variance of the reconstructed signal to quantify model uncertainty; It is a parametric posterior distribution, in which multiple different parameters are sampled. Let i = 1, 2, ..., and then substitute the parameters into the Bayesian neural network. In the process, multiple reconstructed signals are obtained. The uncertainty of the model prediction is measured by calculating the mean and variance of the reconstructed signal; Finally, by utilizing sparsity priors, the complete voiceprint signal is reconstructed from a small number of measurements; by combining physical constraints and data-driven loss, the reconstruction quality is optimized to obtain the reconstructed voiceprint field.

15. The transformer abnormal sound source localization system according to claim 10, characterized in that, The training module is used for: Based on the sound wave propagation equation, the sound field inside the transformer is regarded as a three-dimensional sound wave propagation problem. The physical process is described by the wave equation, and the physical information is embedded into a Bayesian neural network. The continuity equation is discretized into spatiotemporal grid points, and the spatial and temporal derivatives are approximated using finite difference or spectral methods to achieve spatiotemporal discretization. The boundary conditions consider the reflection boundary conditions of the transformer enclosure, and the normal derivative is set to zero as the Neumann boundary condition; Multi-task joint training strategies include: Loss function design: Among them, the data fitting term It is the mean square error between the observed sound pressure and the predicted sound pressure, and a physical constraint term. It is the MSE of the wave equation residuals, and the location loss term. α is the predicted Euclidean distance between the sound source location and the actual location; L is the total loss value; α, β, and γ are the data fitting terms L and L, respectively. data Physical constraint term L physics Location loss item L localization Weighting coefficients; The optimization algorithm for the multi-task joint training strategy uses the Adam or L-BFGS optimizer to dynamically adjust the learning rate to balance the multi-task objectives, and introduces dropout or weights based on regularization methods to obtain a trained Bayesian neural network.

16. The transformer abnormal sound source localization system according to claim 15, characterized in that, The network architecture of the Bayesian neural network includes: Input layer: compressed observation vector y and spatiotemporal coordinates (x,t); y carries the voiceprint feature information after compressed sensing processing, and the spatiotemporal coordinates (x,t) are used to provide the network with contextual information of location and time; Hidden layers: Fully connected layers or convolutional layers are used to extract non-linear features; Fully connected layers connect all input and output nodes using a weight matrix, while convolutional layers use convolutional kernels to extract local features. Output layer: Predicted sound pressure level or sound source intensity , It is the sound pressure level predicted by the network. It is the predicted sound source intensity, corresponding to the actual sound pressure p(x,t) and sound source intensity q(x,t); The residuals of the wave equation are calculated by automatic differentiation, and the squared residuals are added to the loss function.

17. The transformer abnormal sound source localization system according to claim 10, characterized in that, The positioning module is used for: Input compressed observation data and predict the sound field distribution using a trained PINN. Using a trained Bayesian neural network, compressed observation data is used as input to obtain the predicted spatial and temporal distribution of sound pressure. ; To predict sound source terms As an initial guess, the objective function is minimized using the conjugate gradient method: ； in, Let s be the location of the sound source, and x(s) be the sound field signal at the corresponding sound source location s. This represents the observation vector y and the vector reconstructed from the sound source location s. The squared L2 norm between the observed data and the data reconstructed from the sound source location measures the difference between the observed data and the data reconstructed from the sound source location; λ is a weighting coefficient used to balance the data difference term and the physical constraint term R. physics By continuously adjusting the sound source position s using the conjugate gradient method, the objective function is minimized, thereby finding the sound source position that best matches the observation data and physical laws. By utilizing the l1 norm to constrain the sparsity of the sound source terms, we can focus on localizing a single sound source; the l1 norm is... ,in q i It is a sound source item q The value of (x,t) at discrete points; The confidence interval of the sound source location is estimated by using Bayesian PINN or Monte Carlo dropout.

18. The transformer abnormal sound source localization system according to claim 10, characterized in that, The display module is used for: The predicted sound pressure is mapped onto a 3D model of the transformer to generate a dynamic sound field cloud map; by mapping the predicted sound pressure values... Corresponding to the three-dimensional spatial model of the transformer, the changes in sound pressure at different locations and over time are displayed in the form of cloud maps, intuitively presenting the sound field inside the transformer; The location of the abnormal sound source is displayed as a point cloud or marked points and superimposed on the transformer structural diagram. This allows for a direct view of the specific location of the sound source inside the transformer.

19. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the transformer abnormal sound source localization method according to any one of claims 1-9.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the transformer abnormal sound source localization method according to any one of claims 1-9.

21. A computer program product, the computer program product comprising computer instructions, characterized in that, The computer instructions instruct the computer to execute the transformer abnormal sound source localization method according to any one of claims 1-9.

Images