Rectifier transformer voiceprint fault monitoring method and system based on physical information constraint

Abstract

The present application relates to the technical field of acoustic fault monitoring, in particular to a rectifier transformer voiceprint fault monitoring method and system based on physical information constraint. The method collects multi-channel voiceprint time series data of a cooling fan, and constructs a voiceprint time series sample set through overlapping sliding sampling; a physical information neural network is constructed, the hidden layer includes a built-in adaptive wavelet transform and a long short-term memory time-frequency memory layer with a joint forgetting gate, an attention mechanism layer and a physical mapping layer, the voiceprint time series sample is subjected to time-frequency feature extraction and mapping into a predicted complex sound pressure set; the predicted complex sound pressure set is taken as a constraint object, a total loss function composed of a data-driven loss, a Helmholtz frequency domain wave equation residual loss and a Robin mixed impedance boundary residual loss weighted together is constructed to train the network; the network output feature is input into a fault classification layer, and a cooling fan fault diagnosis result is output. The present application has a high fault recognition accuracy under the conditions of scarce fault samples and sparse measuring points.

Description

Technical Field

[0001] This invention relates to the field of acoustic fault monitoring technology, specifically to a method and system for monitoring acoustic faults in rectifier transformers based on physical information constraints. Background Technology

[0002] As key electrical equipment in industrial rectifier systems, metallurgical electrolysis systems, and DC power supply systems, rectifier transformers operate under high current, nonlinear loads, and complex electromagnetic environments for extended periods. The reliable operation of their cooling fans directly impacts the transformer's heat dissipation capacity and safe, stable operation. Under continuous rotation and complex airflow disturbances, cooling fans are prone to faults such as blade imbalance and foreign object jamming, leading to decreased cooling efficiency and consequently increased localized temperature rise in the transformer. Therefore, online condition monitoring and early fault identification of cooling fans have significant engineering value.

[0003] Acoustic fingerprint signals can directly reflect changes in the operating status of cooling fans, and have the advantages of being non-contact, having a fast response speed, and containing rich information, making them an important technical direction for cooling fan fault monitoring. However, cooling fan acoustic fingerprint signals have obvious non-stationarity and time-frequency coupling characteristics. Blade imbalance faults usually manifest as low-frequency periodic fluctuations, while faults such as foreign object jamming manifest as short-term impulsive high-frequency abrupt changes. Existing methods based on traditional spectrum analysis or ordinary time-series networks are difficult to simultaneously capture these two significantly different types of time-frequency characteristics.

[0004] Furthermore, the number of early failure samples of cooling fans in industrial sites is limited, and the occurrence time of different failure states is short and the repeatability is poor. Purely data-driven deep learning models are prone to overfitting and insufficient generalization ability. Existing physical information neural network methods mostly treat physical constraints as independent sound field solving tools, failing to deeply integrate the extraction of acoustic fingerprint temporal features with physical sound field constraints in the same training process, resulting in insufficient fault identification ability under conditions of sparse measurement points and noise interference. Summary of the Invention

[0005] This invention provides a method and system for monitoring rectifier transformer acoustic faults based on physical information constraints, in order to solve the problem of insufficient fault identification accuracy of existing methods under conditions of scarce fault samples, sparse measurement points, and noise interference.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention relates to a physical information-constrained method for monitoring acoustic fingerprint faults in rectifier transformers, comprising: S100: Collects multi-channel acoustic timing data during the operation of the cooling fan; S200: Synchronously extract and standardize the multi-channel voiceprint time-series data to construct a voiceprint time-series sample set; S300: Construct a physical information neural network. The hidden layer of the physical information neural network includes a long short-term memory time-frequency memory layer, an attention mechanism layer, and a physical mapping layer connected in sequence. The long short-term memory time-frequency memory layer incorporates adaptive wavelet transform during cell state update and sets a joint forgetting gate to extract time-frequency features from the voiceprint time-series sample set, obtaining the time-frequency hidden state of each data point. The attention mechanism layer performs global weighted fusion of the time-frequency hidden states to obtain a time-frequency fusion feature vector. The physical mapping layer maps the time-frequency fusion feature vector to a predicted complex sound pressure set. S400: Using the predicted complex sound pressure set as the constraint object, construct a total loss function composed of data-driven loss, Helmholtz frequency domain wave equation residual loss and Robin hybrid impedance boundary residual loss, and train the physical information neural network under the constraint of the total loss function; S500: Input the features output by the trained physical information neural network into the fault classification layer, and output the fault diagnosis results of the cooling fan.

[0007] As a preferred embodiment of the present invention, in S200, the step of constructing the voiceprint time-series sample set includes: Based on the time period of each operating state, a fixed-length time truncation window is set to synchronously truncate the multi-channel voiceprint time series data with absolute time alignment, and the data of each channel are spliced ​​together according to the channel dimension to construct a multi-channel voiceprint time series matrix. Set a time sliding step size, and continuously overlap and slide the time interception window on the time axis to generate multiple derived voiceprint time sequence sub-matrices that are translated in the time dimension; The multi-channel voiceprint timing matrix and the derived voiceprint timing sub-matrix are combined to construct the voiceprint timing sample set.

[0008] As a preferred embodiment of the present invention, in S300, the combined forgetting gate specifically includes: At each data point, based on the current input signal and the hidden state of the previous data point, the state forget gate, time forget gate, and frequency forget gate are independently calculated and generated in the state dimension, time dimension, and frequency dimension, respectively. Perform tensor outer product operations on the state forget gate, time forget gate, and frequency forget gate to generate a joint control tensor; The joint control tensor is multiplied element-wise with the cell state of the previous data point to obtain the history retention feature.

[0009] As a preferred embodiment of the present invention, in S300, the adaptive wavelet transform specifically includes: Based on the current input signal and the hidden state of the previous data point, the wavelet scale parameter and wavelet translation parameter corresponding to the current data point are dynamically generated through network parameters, and the input modulation signal is calculated. Morlet wavelet basis functions are constructed using the wavelet scaling parameters and wavelet translation parameters to perform dual-domain decomposition of the audioprint time series sample set in both the time and frequency domains. The input modulation signal and the Morlet wavelet basis function are performed by tensor cross product operation and then superimposed on the history preservation feature to complete the cell state update.

[0010] As a preferred embodiment of the present invention, in S300, the step of the attention mechanism layer performing global weighted fusion of the time-frequency hidden state includes: The time-frequency hidden state of each data point is nonlinearly mapped to obtain the internal representation vector of each data point; The inner product similarity between the internal representation vector of each data point and a learnable global fault context vector is calculated, and the attention weight of each data point is obtained after Softmax normalization. The time-frequency hidden states of all data points are weighted and summed according to the attention weights to obtain the time-frequency fusion feature vector.

[0011] As a preferred embodiment of the present invention, in S400, the step of constructing the data-driven loss includes: By using the time-domain acoustic voltage signals collected by each acoustic sensor and combining them with their sound pressure sensitivity coefficients, the real part reference value and imaginary part reference value of the observed sound pressure at the corresponding characteristic frequency of each measuring point are obtained. The mean square error between the real and imaginary parts of each predicted complex sound pressure in the predicted complex sound pressure set and the reference values ​​of the real and imaginary parts of the observed sound pressure is calculated to construct the data-driven loss.

[0012] As a preferred embodiment of the present invention, in S400, the step of constructing the residual loss of the Helmholtz frequency domain wave equation includes: Within the solution domain, unlabeled configuration points are randomly sampled, and the equivalent sound velocity at each configuration point is pre-assigned based on the segmented medium properties of the rectifier transformer insulating oil region, the tank structural component region, and the air region. Substitute the predicted complex sound pressure set into the Helmholtz frequency domain wave equation, decompose the Helmholtz frequency domain wave equation into real part partial differential equation and imaginary part partial differential equation, calculate the second-order partial derivative of the predicted sound pressure with respect to the spatial coordinates by automatic differentiation, calculate the real part residual and imaginary part residual respectively, and construct the residual loss of the Helmholtz frequency domain wave equation.

[0013] As a preferred embodiment of the present invention, in S400, the step of constructing the Robin hybrid impedance boundary residual loss includes: Based on the geometry of the rectifier transformer, boundary configuration points are pre-constructed at the interfaces between the insulating oil and the tank, and between the tank and the air medium, and the equivalent acoustic impedance is pre-determined according to the type of each medium interface and the corresponding characteristic frequency. Based on the impedance relationship between sound pressure and normal particle velocity, the Robin mixed impedance boundary condition is decoupled into real and imaginary residual terms. The normal derivative of the predicted sound pressure along the outer normal direction of the boundary is calculated by automatic differentiation, and the Robin mixed impedance boundary residual loss is constructed.

[0014] As a preferred embodiment of the present invention, in S500, the step of outputting the fault diagnosis result of the cooling fan includes: The parameters of the trained physical information neural network are fixed, and the set of predicted complex sound pressures output by the physical information neural network is used to form a physically constrained sound field feature. The physical constraint acoustic field features are input into the fault classification layer, and the fault classification layer is trained with cross-entropy loss. The fault diagnosis result of the cooling fan is determined according to the category corresponding to the highest probability in the probability distribution output by the fault classification layer.

[0015] This invention also proposes a rectifier transformer acoustic fault monitoring system based on physical information constraints, comprising: Voiceprint acquisition module: Acquires multi-channel voiceprint timing data during the operation of the cooling fan; Sample construction module: synchronously extracts and standardizes the multi-channel voiceprint time-series data to construct a voiceprint time-series sample set; Network construction module: Constructs a physical information neural network. The hidden layer of the physical information neural network includes a long short-term memory (LSTM) time-frequency memory layer, an attention mechanism layer, and a physical mapping layer connected in sequence. The LTM time-frequency memory layer incorporates adaptive wavelet transform during cell state update and sets a joint forgetting gate to extract time-frequency features from the voiceprint time-series sample set, obtaining the time-frequency hidden state of each data point. The attention mechanism layer performs global weighted fusion of the time-frequency hidden states to obtain a time-frequency fusion feature vector. The physical mapping layer maps the time-frequency fusion feature vector to a predicted complex sound pressure set. Physical constraint module: Taking the predicted complex sound pressure set as the constraint object, a total loss function is constructed, which is a weighted sum of data-driven loss, Helmholtz frequency domain wave equation residual loss and Robin hybrid impedance boundary residual loss. The physical information neural network is trained under the constraint of the total loss function. Fault diagnosis module: Input the features output by the trained physical information neural network into the fault classification layer, and output the fault diagnosis results of the cooling fan.

[0016] The beneficial effects of this invention are: 1. This invention deeply integrates the joint forget gate and the built-in adaptive wavelet transform into the cell state update process of the long short-term memory network, enabling the network to collaboratively screen historical voiceprint features in three dimensions: state, time, and frequency. At the same time, it dynamically generates wavelet scale parameters and translation parameters based on the current input, achieving synchronous and accurate extraction of low-frequency periodic features of blade imbalance faults and high-frequency transient impact features of foreign object jamming faults. This overcomes the limitation of ordinary long short-term memory networks that only update memory along the time dimension and cannot take into account the two significantly different fault features.

[0017] 2. This invention directly embeds the residuals of the Helmholtz frequency domain wave equation and the Robin hybrid impedance boundary residuals into the network training process, so that the extraction of audioprint temporal features and physical sound field constraints interact in the same training link. Unlike the existing methods that treat the physical information neural network as an independent sound field solver, this invention can still maintain physical consistency and generalization ability under conditions of sparse measurement points and noise interference.

[0018] 3. This invention uses an overlapping sliding sampling strategy to expand the training samples multiple times while strictly preserving the continuous transient physical characteristics between adjacent samples, effectively alleviating the overfitting problem caused by the scarcity of early failure samples of cooling fans in industrial sites. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the acoustic fingerprint fault monitoring method for rectifier transformers based on physical information constraints according to the present invention. Figure 2 This is a schematic diagram of the overall structure of the physical information neural network of the present invention; Figure 3 This is a schematic diagram of the physical constraint training process and loss calculation of the present invention; Figure 4 This is a schematic diagram of the long short-term memory time-frequency memory unit structure of the present invention; Figure 5 This is a schematic diagram of the structure of the rectifier transformer acoustic fault monitoring system based on physical information constraints according to the present invention; Figure 6 This is a schematic diagram comparing the audioprint timing samples under normal operating conditions and fault conditions of the present invention; Figure 7 This is the two-dimensional visualization result of the physical information neural network feature space t-SNE of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] Example 1: As Figure 1 As shown, the present invention provides a rectifier transformer acoustic fault monitoring method based on physical information constraints, comprising: S100: Collects multi-channel acoustic timing data during the operation of the cooling fan; Specifically, multiple spatial monitoring points are deployed on the rigid support structure surrounding the transformer radiator, with each point avoiding the main airflow channels of the cooling fan's inlet and outlet. Each monitoring point is equipped with at least one acoustic signature sensor, which is mounted on the surface of the rigid support structure using a bracket. A sound-transparent windproof cover is fitted around the sensor to shield it from airflow turbulence noise. All acoustic signature sensors together form a multi-channel acoustic signature acquisition array. Each sensor is connected to the same online monitoring host and synchronously acquires data using a unified sampling frequency and clock to obtain multi-channel acoustic signature time-series data of the cooling fan under various operating conditions. These operating conditions include normal operation and fault conditions caused by blade imbalance or foreign object obstruction.

[0022] In one implementation, six spatial monitoring points are selected on the rigid support structure surrounding the radiator to form a monitoring point set. ,in Indicates the first One monitoring point, Index of monitoring points. Each monitoring point Each device is equipped with two voiceprint sensors in pairs, forming a total of 12 independent physical monitoring channels.

[0023] S200: Synchronously extract and standardize the multi-channel voiceprint time-series data to construct a voiceprint time-series sample set; Specifically, based on historical operation logs, the specific time points of known cooling fan failures and the normal operation time periods are extracted, and the voiceprint data for the corresponding time periods are labeled with operation status categories. The operation status categories include normal operation status and failure status caused by blade imbalance or foreign object jamming.

[0024] Based on the time periods of each operating state, a fixed-length time truncation window is set. Multi-channel acoustic fingerprint time-series data from each monitoring point are synchronously truncated with absolute time alignment to obtain multi-channel discrete time series. The truncated data is then standardized to eliminate amplitude differences between different sensor channels. The data from each channel are concatenated along the channel dimension to construct a multi-channel acoustic fingerprint time-series matrix, defined as the [missing information - likely a specific matrix or term]. One original voiceprint time sequence sample Its mathematical structure is represented as: ; in, This is the index of the original voiceprint time-series samples. This represents the total number of original temporal samples of the voiceprint. The total number of discrete sampled data points contained within the time frame window; For the first The sample at the th A row vector of multi-channel observations at each sampling time. For sampling time index, The original voiceprint time-series sample The dimension is , This represents the total number of sensor channels used for synchronous data acquisition.

[0025] Furthermore, the step of constructing the audioprint time-series sample set includes: addressing the problem of scarce early fault samples of the cooling fan and their short-term self-recovery characteristics by introducing an overlapping sliding sampling strategy. A time sliding step size is set. ,and Strictly less than This preserves the continuous transient physical features between adjacent samples, using the original speaker time series matrix. The starting time is taken as the starting point, and the length is Time capture window by step Perform continuous overlapping sliding cuts on the timeline to generate multiple dimensions that are the same. The derived audioprint time-series submatrix is ​​shifted in the time dimension. The time truncation range of the k-th audioprint time-series submatrix is: ; Where k is the sliding cutoff number, and The matrix structure of the multi-channel long-time series audio data, which does not exceed the end sampling position within the corresponding operating state time period, is represented as follows: ; in, For the first The derived voiceprint temporal submatrix is ​​generated by the k-th overlapping sliding sampling of the original voiceprint temporal samples. To slide and extract the sequence number, and It is a positive integer; The total number of derived audioprint temporal sub-matrices generated from a single original sample through sliding overlap sampling; Indicates the first The original sample with the first A row vector of multi-channel observations extracted from each sampling time point. This represents the row vector of multi-channel observations at the sampling time at the end of the corresponding truncation range.

[0026] All original voiceprint timing matrices With the corresponding derived voiceprint timing submatrix The data is aggregated and then subjected to a global random shuffling operation to construct a temporal sample set of voiceprints. , is represented as: ; in, This is the first original voiceprint time sequence sample. This is the derived voiceprint temporal submatrix generated from the first overlapping sliding sample of the first original voiceprint temporal sample. For the first The derived voiceprint temporal submatrix is ​​generated by the Kth overlapping sliding sampling of the original voiceprint temporal samples. For the Nth original voiceprint time sequence sample, The derived audioprint temporal submatrix generated by the first overlapping sliding sampling of the Nth original audioprint temporal sample. For the first The derived voiceprint temporal submatrix is ​​generated by the Kth overlapping sliding sampling of the original voiceprint temporal samples. This represents the total number of original voiceprint time-series samples.

[0027] S300: Construct a physical information neural network. The hidden layer of the physical information neural network includes a long short-term memory time-frequency memory layer, an attention mechanism layer, and a physical mapping layer connected in sequence. The long short-term memory time-frequency memory layer incorporates adaptive wavelet transform during cell state update and sets a joint forgetting gate to extract time-frequency features from the voiceprint time-series sample set, obtaining the time-frequency hidden state of each data point. The attention mechanism layer performs global weighted fusion of the time-frequency hidden states to obtain a time-frequency fusion feature vector. The physical mapping layer maps the time-frequency fusion feature vector to a predicted complex sound pressure set. Specifically, a physical information neural network is constructed, wherein the hidden layers of the physical information neural network include a long short-term memory time-frequency memory layer, an attention mechanism layer, and a physical mapping layer connected in sequence. The voiceprint time-series sample set is used. The voiceprint time-series samples are input point by point into the long short-term memory time-frequency memory layer. The long short-term memory time-frequency memory layer incorporates adaptive wavelet transform during cell state update and sets a joint forgetting gate to extract time-frequency features from the voiceprint time-series samples, thereby obtaining the time-frequency hidden state of each data point. , For indexing data points, The attention mechanism layer hides the time-frequency state of each data point. Perform global weighted fusion to obtain the time-frequency fusion feature vector. The physical mapping layer fuses time-frequency feature vectors. Mapped to the set of predicted complex sound pressures .

[0028] like Figure 2 As shown, the overall structure of the physical information neural network of this invention, from bottom to top, is as follows: the bottom input layer receives data point 1 to data point 2. The system extracts voiceprint time-series samples; the hidden layer consists of a long short-term memory time-frequency memory layer with built-in wavelet transform, and each memory unit is connected sequentially along the time axis. Time-frequency features are extracted for each data point, and a hidden state sequence is output. The hidden state sequence is simultaneously fed into the attention weight mapping module, which calculates the attention weight for each data point and then performs global weighted fusion on the hidden state sequence to obtain the time-frequency fusion feature vector. Time-frequency fusion feature vector The sound pressure is mapped to a set of predicted complex sound pressures by the physical mapping layer, which is used to map the real physical field to apply physical constraints; finally, the fault diagnosis result is output by the Softmax classification layer.

[0029] like Figure 3 As shown in the diagram, the physical constraint training process of this invention is illustrated from left to right: the LSTM and attention mechanism layer on the left receives audioprint time-series samples and outputs a time-frequency fusion feature vector. The intermediate physical mapping layer will Mapped to the set of predicted complex sound pressures The right-side output layer will By substituting the values ​​into the Robin boundary equation, the Helmholtz equation, and comparing them with measured data, boundary residual losses can be constructed respectively. Physical residual loss and data-driven loss The total loss function is composed of the weighted sum of the three losses. Network parameters are optimized through backpropagation.

[0030] It should be noted that, unlike traditional physical information neural networks which are used only as independent sound field solvers, this invention embeds physical constraints into the entire network training process. The time-frequency features extracted by the long short-term memory time-frequency memory layer are simultaneously constrained by the observed sound pressure data, the residual of the Helmholtz frequency domain wave equation, and the boundary constraints of the Robin hybrid impedance during training, so that the network can maintain physical consistency and generalization ability under the conditions of limited samples and sparse measurement points.

[0031] Furthermore, the joint forget gate specifically includes: At any data point The combined forget gate simultaneously receives the current input signal. Hidden state of the previous data point .

[0032] To enable the Long Short-Term Memory (LSTM) time-frequency memory layer to simultaneously retain the historical features of the voiceprint signal in the state, time, and frequency dimensions, this embodiment defines the cell state of the previous data point as a three-dimensional time-frequency memory tensor: ; in, For the cell state at data point t-1, Represents feature dimension, Represents the state dimension. Indicates the time dimension. Represents the frequency dimension.

[0033] The model utilizes three independent sets of network parameters to and Mapped to the state dimension respectively Time dimension and frequency dimension After processing with the Sigmoid activation function, the state forget gate is calculated independently. Time Oblivion Gate and frequency forget gate Perform a tensor outer product operation on the three forgetting gate vectors to generate a joint control tensor. The calculation formula is: ; in, Forget gate vector, For the time forgetting gate vector, For the frequency forgetting gate vector, This represents the tensor outer product operation. For the joint control tensor. And: ; Specifically, any element in the joint control tensor satisfies: ; in, , , Therefore, the joint control tensor Cell state compared to the previous data point They have the same three-dimensional tensor dimensions.

[0034] Joint control tensor Cell state compared to the previous data point Perform element-wise multiplication to obtain historically preserved features. The calculation formula is: ; in, For the cell state at data point t-1, This represents element-wise multiplication. It preserves historical features and is used to jointly determine the degree of preservation and forgetting of historical voiceprint features in three dimensions: state, time, and frequency.

[0035] Furthermore, the adaptive wavelet transform specifically includes: The input gate is based on the current input signal Hidden state of the previous data point The input modulation signal is calculated. The input modulation signal The weights that represent the importance of the input information for the current data point and , Representing the state dimension. The adaptive wavelet transform unit adjusts according to the input modulation signal. The wavelet scale parameters corresponding to the current data point are dynamically generated through network parameters. With wavelet translation parameters Its dynamic generation formula is: ; ; in, This is the scale weight matrix for adaptive learning of the network. This is the translation weight matrix for adaptive learning of the network. This is the bias vector corresponding to the scale. The offset vector corresponding to the translation. The hyperbolic tangent activation function is used. For wavelet scaling parameters, These are the wavelet translation parameters.

[0036] Using dynamically generated wavelet scaling parameters With wavelet translation parameters Construct Morlet wavelet basis functions ,in , , In terms of time dimension, For frequency dimension. Therefore, Indicates the first The time position and the first All wavelet response coefficients corresponding to each frequency position It constitutes a two-dimensional time-frequency response, whose dimensions correspond to the time dimension and the frequency dimension.

[0037] In the time domain dimension and frequency domain dimension The non-stationary acoustic signature signal is decomposed into two domains. The input modulation signal is then processed. With Morlet wavelet basis functions After performing tensor outer product operations, the data is superimposed onto the history-preserving features. Up, complete the cell state of the current data point. The final update formula is: ; in, To preserve historical characteristics output by the joint forgetting gate, For the input modulation signal, For Morlet wavelet basis functions, This represents the tensor outer product operation. This represents the updated cell state for the current data point.

[0038] like Figure 4 As shown, the long short-term memory (LSM) time-frequency memory unit at any data point Simultaneously receive the current input signal Hidden state of the previous data point The three parallel Sigmoid branches on the left generate state forget gates respectively. Frequency forget gate Time Oblivion Gate The three are multiplied by the tensor outer product. Operations generate joint control tensor , compared with the previous cell state Element-wise multiplication yields historical preservation features Input gates in the intermediate branch The sigmoid-activated input modulated signal is compared with the tanh-activated input modulated signal. Multiply, we get Adaptive wavelet transform unit with To input adaptively learn wavelet parameters, construct Morlet wavelet basis functions. ,and After performing tensor outer product, the components are superimposed. Complete the cell state update The right-side output gate, after being activated by the Sigmoid function, interacts with... Element-wise multiplication outputs the time-frequency hidden state of the current data point. In the picture , , These correspond to the state forgetting gates in the main text. Frequency forget gate Time Oblivion Gate The time subscript is omitted in the illustration. ; This is the intermediate input modulation value. For the input gate control vector, This refers to the input modulation signal in the main text. In the picture This is the Sigmoid activation function.

[0039] After the cell state update is completed, the cell state of the current data point is... This is a three-dimensional time-frequency memory tensor containing state, time, and frequency dimensions. Since the cell state contains complex time-frequency information introduced by the Morlet wavelet response, a complex amplitude extraction operation is first performed on the cell state to obtain the amplitude tensor: ; in, , Let be the magnitude tensor of data point t. For the state dimension, In terms of time dimension, For frequency dimension, For operations that take the real part of a complex number, This refers to the operation of taking the imaginary part of a complex number.

[0040] Subsequently, the amplitude tensor Learnable weighted convergence along the time and frequency dimensions yields a real-valued state representation vector: ; in, Let be the real-valued state representation vector of data point t. Represents the magnitude tensor In the The time position and the A slice of the state dimension at each frequency position. and These are the corresponding reconstruction weights and biases. .

[0041] The output gate is based on the current input signal. Hidden state of the previous data point The output control vector is calculated by including the weight matrix and the Sigmoid activation function. ,in According to the output control vector With the real number state representation vector Generate the time-frequency hidden state of the current data point using the following formula: ; in, This represents element-wise multiplication. This represents the time-frequency hidden state of the current data point.

[0042] Furthermore, the step of the attention mechanism layer performing global weighted fusion of the time-frequency hidden states includes: Time-frequency hidden state of each data point Perform a nonlinear mapping to obtain the internal representation vector of each data point. The calculation formula is: ; in, Here is the weight matrix of the hidden layer. For bias vectors, The hyperbolic tangent activation function is used. For the first The internal representation vector of each data point.

[0043] Calculate the internal representation vector of each data point With global fault context vector The inner product similarity between the data points is normalized using the Softmax function to obtain the attention weights for each data point. The calculation formula is: ; in, Internal representation vector transpose, This is a dynamically updated global fault context vector that propagates back through the network. This represents the total number of data points within the voiceprint time-series sample. For the summation index, , For the first Attention weights for each data point. The global fault context vector. In a physical sense, it acts as a fault feature query probe to measure the contribution of abrupt, non-stationary voiceprint features within each data point to the overall fault determination.

[0044] Based on attention weight Time-frequency hidden state of all data points We perform a weighted summation to obtain the time-frequency fusion feature vector. The calculation formula is: ; in, For the first Attention weights for each data point For the first The time-frequency hidden state of each data point This represents the total number of data points within the voiceprint time-series sample. This is the time-frequency fusion feature vector.

[0045] Time-frequency fusion feature vector The physical mapping layer, constructed via a fully connected network, maps to a set of predicted complex sound pressure levels. , is represented as: ; in, To monitor the total number of monitoring points, For the first The predicted complex sound pressure levels corresponding to each measurement point. Set of predicted complex sound pressure levels. Any predicted complex sound pressure All include the predicted real part of the sound pressure. and predicting the imaginary part of sound pressure , is represented as: ; in, For the corresponding characteristic frequency The predicted sound pressure solid part, For the corresponding characteristic frequency The imaginary part of the predicted sound pressure. The imaginary unit. Predicting the set of complex sound pressure levels. Used for subsequent calculations of physical equation residuals and boundary condition residuals.

[0046] S400: Using the predicted complex sound pressure set as the constraint object, construct a total loss function composed of data-driven loss, Helmholtz frequency domain wave equation residual loss and Robin hybrid impedance boundary residual loss, and train the physical information neural network under the constraint of the total loss function; Specifically, the predicted complex sound pressure set output in step S300 Construct the total loss function for the constraint object. The total loss function is data-driven loss. Helmholtz frequency domain wave equation residual loss And Robin's mixed impedance boundary residual loss Weighted composition, expressed as: ; in, The weighting coefficients for the data-driven loss term. These are the weighting coefficients for the residual loss term in the Helmholtz frequency domain wave equation. This is the weighting coefficient for the Robin mixed impedance boundary residual loss term.

[0047] For the monitoring points set up in step S100, the time-domain acoustic voltage signals collected by each acoustic fingerprint sensor are used, combined with their sound pressure sensitivity coefficients, to obtain the characteristic frequencies of each monitoring point in the characteristic frequency set. The real and imaginary part reference values ​​of the observed sound pressure are determined. At sparse locations with measurement points, the mean square error between the predicted complex sound pressure and the observed sound pressure reference value is calculated, and a data-driven loss function is constructed. , is represented as: ; in: To monitor the total number of monitoring points, For measurement point index, The total number of characteristic frequencies. For characteristic frequency index, For the first Spatial coordinates of each monitoring point For the network in the first Monitoring points, characteristic frequencies The real part of the predicted sound pressure level output. This corresponds to the imaginary part of the predicted sound pressure level. For the first Each monitoring point at the characteristic frequency The reference value of the real part of the observed sound pressure is as follows. This is the reference value for the imaginary part of the corresponding observed sound pressure.

[0048] Furthermore, the steps for constructing the residual loss of the Helmholtz frequency domain wave equation include: Random sampling within the solution domain Each unlabeled configuration point is pre-assigned an equivalent velocity of sound based on the segmented dielectric properties of the rectifier transformer insulating oil region, the tank structural component region, and the air region. ,in For the first Spatial coordinates of an unlabeled configuration point For configuring point indexes, .

[0049] At each unlabeled configuration point, the Helmholtz frequency domain wave equation constrains the prediction of the complex sound pressure field, and its complex wave equation is expressed as: ; in, For the Laplace operator, For the network in the first Unlabeled locative points, feature frequencies The predicted complex sound pressure level is below. The first in the set of characteristic frequencies Each characteristic frequency, For wave number, This is the equivalent sound source term, used to characterize the sound source intensity in the frequency domain of sound source excitation caused by structural faults inside the transformer. Wavenumber satisfy: ; in, For the first Each characteristic frequency, Spatial location The equivalent speed of sound at that location.

[0050] The complex wave equation is decomposed into real and imaginary partial differential equations, and the real part of the predicted sound pressure is calculated using an automatic differentiation method. With the imaginary part The real and imaginary residuals are calculated using the second-order partial derivatives relative to the spatial coordinates, and the residual loss of the Helmholtz frequency domain wave equation is constructed. , is represented as: ; in, To determine the total number of unlabeled configuration points randomly sampled within the solution domain, For the network in the first Unlabeled locative points, feature frequencies The real part of the predicted sound pressure level is shown in the output. This corresponds to the imaginary part of the sound pressure level. For equivalent sound source terms The real part, For equivalent sound source terms The imaginary part of the Laplace operator. The second-order partial derivatives of the real and imaginary parts of the predicted complex sound pressure with respect to spatial coordinates are obtained by automatic differentiation.

[0051] Furthermore, the construction steps of the Robin hybrid impedance boundary residual loss include: Based on the geometry of the rectifier transformer, boundary configuration points are pre-constructed at the interfaces between the insulating oil and the tank, and between the tank and the air dielectric. These boundary configuration points are not obtained by the acoustic signature sensor. The boundary configuration points are determined based on the type of each dielectric interface and its corresponding characteristic frequency. Predetermine the equivalent acoustic impedance ,in For media interface type index, For the first Medium-like interface at characteristic frequency The equivalent acoustic impedance is given below. It should be noted that... Determined by the boundary medium type and characteristic frequency, it is considered a constant at the same medium interface and the same characteristic frequency. It can be obtained by calculation based on the density, sound velocity and damping parameters of insulating oil, enclosure material and air, preset based on material acoustic parameter table or equipment structural parameters, or identified through on-site calibration experiments, impedance tube tests or historical operating acoustic data inversion.

[0052] Based on the impedance relationship between sound pressure and normal particle velocity, a Robin hybrid impedance boundary constraint is constructed at the interfaces of multiple media, such as insulating oil and enclosure, and enclosure and air. Its complex impedance equation is expressed as: ; in, For the first Spatial coordinates of the boundary configuration points Configure point indexes for the boundaries. The direction of the outer normal to the boundary. The imaginary unit, The angular frequency of the sound wave For the first The equivalent medium density corresponding to the medium-like interface, For the first Medium-like interface at characteristic frequency The equivalent acoustic impedance is below. For the network in the first Boundary configuration points, characteristic frequencies The predicted complex sound pressure level.

[0053] Using a local response boundary approximation, under the condition that the boundary medium loss is low and the influence of the boundary phase lag on the sound field distribution of the target frequency band is negligible, the first... Medium-like interface at characteristic frequency Equivalent acoustic impedance under The equivalent acoustic impedance is taken as a real number. It should be noted that the real equivalent acoustic impedance is used to approximately characterize the amplitude constraint relationship between the sound pressure at the boundary and the normal particle velocity, and does not mean that the predicted sound pressure is a real number; the predicted sound pressure in this embodiment is still a complex sound pressure, and the real part and imaginary part of the complex sound pressure are coupled to each other through the imaginary coefficient in the Robin mixed impedance boundary condition.

[0054] The Robin mixed impedance boundary condition is decoupled and decomposed into real residual terms. Residual term with imaginary part , is represented as: ; ; in, For the network in the first Boundary configuration points, characteristic frequencies The real part of the predicted sound pressure level is shown in the output. This corresponds to the imaginary part of the sound pressure level. The direction of the outer normal to the boundary. The normal derivative along the direction of the outer normal of the boundary is calculated from the gradient of the predicted sound pressure relative to the spatial coordinates by automatic differentiation.

[0055] Constructing Robin's hybrid impedance boundary residual loss , is represented as: ; in, Configure the total number of points at the boundary. Configure point indexes for the boundaries. For characteristic frequency index, For the first Each boundary configuration point at characteristic frequency The real residual term below, This corresponds to the imaginary residual term. The Robin hybrid impedance boundary residual loss is used to ensure that the network still satisfies the acoustic energy transmission, reflection, and impedance matching laws at the multi-medium interface even in the absence of boundary measured sensor data.

[0056] In the total loss function Under constraints, the parameters of the physical information neural network are updated through backpropagation, so that the time-frequency hidden features extracted by the network not only meet the sample label classification requirements, but also conform to the propagation of the cooling fan sound field and the boundary impedance constraints.

[0057] S500: Input the features output by the trained physical information neural network into the fault classification layer, and output the fault diagnosis results of the cooling fan.

[0058] Specifically, after completing the physical constraint training in step S400, the parameters of the trained physical information neural network are fixed. The predicted complex sound pressure sets output by the physical mapping layer in the physical information neural network are used to form physical constraint sound field features, which are then input into the fault classification layer. The fault classification layer outputs the probability distribution of the cooling fan in each preset operating state or fault category. According to probability distribution The category corresponding to the highest probability in the middle determines the fault diagnosis result of the cooling fan. The preset operating state or fault category includes the normal operating state and the fault state caused by blade imbalance or foreign object jamming.

[0059] Furthermore, the steps for diagnosing the fault results of the output cooling fan include: With all parameters of the physical information neural network trained in step S400 fixed, the set of predicted complex sound pressures output by the physical information neural network is... The real and imaginary parts of each predicted complex sound pressure level are arranged according to the preset calculation point order and characteristic frequency order to form a physically constrained sound field feature. The physical constraints on the sound field characteristics Composed of multiple spatial locations and multiple characteristic frequencies The predicted complex sound pressure is used to characterize the physical sound field distribution characteristics of the rectifier transformer cooling fan under the current sound pattern time series sample.

[0060] Physically constrained sound field characteristics Input Fault Classification Layer The probability distribution of the cooling fan in each preset operating state or fault category was obtained. , is represented as: ; in, To physically constrain the characteristics of the sound field, For the fault classification layer, For the network parameters of the fault classification layer, This represents the probability distribution corresponding to each preset operating state or fault category.

[0061] While keeping the physical information neural network parameters obtained from step S400 unchanged, the fault classification layer is trained using cross-entropy loss only based on labeled audioprint time-series samples, ensuring that the fault category output by the fault classification layer matches the sample label. In the monitoring application phase, based on the probability distribution... The category corresponding to the highest probability is used to determine the fault diagnosis result of the cooling fan.

[0062] Example 2: A company has multiple large-capacity rectifier transformers that continuously supply high-current DC power. The cooling fans operate under high temperature, heavy load, and complex airflow disturbances, and have historically experienced multiple incidents of blade imbalance and foreign object obstruction leading to reduced heat dissipation capacity. Due to the scarcity of early fault samples, complex background noise, and limited acoustic signature sensor measurement points, traditional pure data-driven methods lack generalization ability. Therefore, the company adopted the following... Figure 5 The present invention, based on the physical information constraint-based rectifier transformer acoustic fault monitoring system, performs online status monitoring of the aforementioned cooling fan.

[0063] The acoustic signature acquisition module deploys a multi-channel acoustic signature sensor array on the rigid support structure surrounding the radiator, simultaneously acquiring acoustic signature time-series data under various operating conditions of the cooling fan. The sample construction module expands early fault samples using an overlapping sliding sampling strategy to construct an acoustic signature time-series sample set. For example... Figure 6 As shown, Figure 6 The left side shows the audioprint time series samples collected under normal operating conditions of the cooling fan. The horizontal axis represents time (unit: s), and the vertical axis represents amplitude. The time range is approximately 0 to 12 seconds. The amplitude generally shows stable periodic fluctuations within the range of ±0.4, with strong signal regularity and no obvious abrupt changes. Figure 6The right side shows the acoustic signature time-series sample collected under fault conditions, with a time range of approximately 0 to 40 seconds. The amplitude fluctuation range is similar to that under normal conditions, but the signal exhibits obvious non-stationary characteristics, with sudden amplitude increases and waveform disorder in local areas, especially with a significant expansion of the waveform envelope in the middle time interval. The above comparison shows that the normal operation and fault conditions of the cooling fan have distinguishable temporal characteristics in the acoustic signature time-series sample, providing a sample basis for subsequent time-frequency feature extraction and fault classification of the physical information neural network. The network construction module constructs the physical information neural network. The long short-term memory time-frequency memory layer extracts the low-frequency periodic features of blade imbalance faults and the high-frequency transient impact features of foreign object jamming faults simultaneously through a joint forgetting gate and built-in adaptive wavelet transform. The attention mechanism layer performs global weighted fusion of the time-frequency hidden states of each data point, and the physical mapping layer maps the time-frequency fused feature vector into a predicted complex sound pressure set. The physical constraint module uses the predicted complex sound pressure set as the constraint object, and constructs a total loss function by weighting and combining the Helmholtz frequency domain wave equation residual loss, the Robin mixed impedance boundary residual loss, and the data-driven loss. This ensures that the time-frequency hidden features extracted by the network simultaneously meet the classification requirements and the physical laws of the sound field, maintaining good generalization ability even under conditions of sparse measurement points and noise interference. The fault diagnosis module inputs the physically constrained sound field features into the fault classification layer and outputs the fault diagnosis results of the cooling fan.

[0064] To verify the technical effectiveness of the system of this invention, a simulation experimental dataset was constructed, using five states as the identification objects: normal operation of the rectifier transformer cooling fan, blade imbalance, foreign object jamming, local friction, and bearing abnormality. 200 sets of acoustic signature time-series samples were constructed for each state, totaling 1000 sets. Each sample contained synchronously acquired data from multiple acoustic signature sensor channels, and background equipment noise and airflow turbulence noise of different amplitudes were superimposed to simulate the actual environment. The dataset was divided into training, validation, and test sets in a ratio of 70%, 15%, and 15%, respectively.

[0065] To quantify the synergistic effect of the various technical features of this invention, the following five comparative schemes were tested under the same training conditions: The first is the complete scheme of this invention; the second is a simplified scheme without the joint forgetting gate, using only the forgetting gate of a regular long short-term memory network for single-dimensional control of historical information; the third is a serial "wavelet transform + long short-term memory network" scheme, which first performs an external wavelet transform on the voiceprint signal and then inputs the result into a regular long short-term memory network, without forming multi-dimensional joint memory control within the network; the fourth is a pure data-driven scheme without physical information neural network constraints, without introducing the residuals of the Helmholtz frequency domain wave equation and the Robin hybrid impedance boundary residuals; the fifth is a two-stage scheme of traditional physical information neural network, which first solves the sound field distribution independently and then inputs the sound field features into the classification layer, with physical constraints not directly participating in the voiceprint temporal feature extraction process. The test results of each scheme are shown in Table 1: Table 1. Comparison of overall recognition performance of each scheme.

[0066] As shown in the table above, the overall recognition accuracy of the complete scheme of this invention reaches 94.7%, which is significantly better than the comparative schemes. Compared with the simplified scheme that removes the joint forgetting gate, it improves by 6.1 percentage points, indicating that the joint forgetting control of the three dimensions of state, time and frequency can more effectively preserve the low-frequency periodic fluctuations and high-frequency transient impact features in the cooling fan soundprint. Compared with the serial "wavelet transform + long short-term memory network" scheme, it improves by 7.8 percentage points, indicating that embedding the wavelet transform mechanism into the long short-term memory network for dynamic time-frequency memory updates is better than simple sequential processing. Compared with the pure data-driven scheme without physical information neural network constraints, it improves by 7.3 percentage points, indicating that the residuals of the Helmholtz frequency domain wave equation and the boundary residuals of the Robin mixed impedance can effectively improve the generalization ability of the model under noise interference and sparse measurement points. Compared with the traditional two-stage scheme of physical information neural network, it improves by 5.6 percentage points, indicating that directly embedding physical constraints into the soundprint temporal feature extraction and network parameter training process is better than the separate processing method of solving the sound field first and then classifying.

[0067] like Figure 7 The figure shows the two-dimensional visualization results of the t-SNE feature space extracted by the physical information neural network. Blue dots represent samples in normal operation, and orange dots represent samples in abnormal (fault) state. It can be seen that the two types of samples form clearly separated clusters in the feature space. Normal samples are concentrated in the right region, while abnormal samples are scattered in the left region, presenting multiple sub-clusters. This indicates that the physical constraint acoustic field features extracted in this invention have strong inter-class discriminative ability, providing an effective feature foundation for accurate identification in the subsequent fault classification layer.

[0068] Further statistical analysis of the identification results for each fault category is shown in Table 2: Table 2 Comparison of Fault Category Identification Accuracy of Different Comparison Schemes

[0069] The identification results for each category show that this invention has a higher recognition rate for short-term impact-type faults such as foreign object jamming. Foreign object jamming faults typically exhibit high-frequency transient characteristics with short duration and significant frequency component abrupt changes in acoustic signature signals, while blade imbalance faults mainly manifest as low-frequency periodic fluctuations. Ordinary long short-term memory networks or serial wavelet processing methods struggle to simultaneously retain these two significantly different types of time-frequency information. This invention uses a joint control tensor to enable the network to collaboratively filter acoustic signature features across the state, time, and frequency dimensions. Simultaneously, by predicting complex sound pressure sets, it introduces constraints from the Helmholtz frequency domain wave equation and Robin's hybrid impedance boundary constraints. This ensures that the hidden layer features learned by the network not only meet the sample label classification requirements but also conform to the sound field propagation laws and boundary impedance relationships of cooling fans. Therefore, it significantly improves fault identification accuracy under conditions of scarce early-stage fault samples, complex on-site noise, and sparse measurement points.

[0070] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring acoustic fingerprint faults in rectifier transformers based on physical information constraints, characterized in that, include: S100: Collects multi-channel acoustic timing data during the operation of the cooling fan; S200: Synchronously extract and standardize the multi-channel voiceprint time-series data to construct a voiceprint time-series sample set; S300: Construct a physical information neural network. The hidden layer of the physical information neural network includes a long short-term memory time-frequency memory layer, an attention mechanism layer, and a physical mapping layer connected in sequence. The long short-term memory time-frequency memory layer incorporates adaptive wavelet transform during cell state update and sets a joint forgetting gate to extract time-frequency features from the voiceprint time-series sample set, obtaining the time-frequency hidden state of each data point. The attention mechanism layer performs global weighted fusion of the time-frequency hidden states to obtain a time-frequency fusion feature vector. The physical mapping layer maps the time-frequency fusion feature vector to a predicted complex sound pressure set. S400: Using the predicted complex sound pressure set as the constraint object, construct a total loss function composed of data-driven loss, Helmholtz frequency domain wave equation residual loss and Robin hybrid impedance boundary residual loss, and train the physical information neural network under the constraint of the total loss function; S500: Input the features output by the trained physical information neural network into the fault classification layer, and output the fault diagnosis results of the cooling fan.

2. The method for monitoring rectifier transformer acoustic faults based on physical information constraints according to claim 1, characterized in that, In S200, the step of constructing the voiceprint time-series sample set includes: Based on the time period of each operating state, a fixed-length time truncation window is set to synchronously truncate the multi-channel voiceprint time series data with absolute time alignment, and the data of each channel are spliced ​​together according to the channel dimension to construct a multi-channel voiceprint time series matrix. Set a time sliding step size, and continuously overlap and slide the time interception window on the time axis to generate multiple derived voiceprint time sequence sub-matrices that are translated in the time dimension; The multi-channel voiceprint timing matrix and the derived voiceprint timing sub-matrix are combined to construct the voiceprint timing sample set.

3. The method for monitoring rectifier transformer acoustic fingerprint faults based on physical information constraints according to claim 1, characterized in that, In S300, the joint forgetting gate specifically includes: At each data point, based on the current input signal and the hidden state of the previous data point, the state forget gate, time forget gate, and frequency forget gate are independently calculated and generated in the state dimension, time dimension, and frequency dimension, respectively. Perform tensor outer product operations on the state forget gate, time forget gate, and frequency forget gate to generate a joint control tensor; The joint control tensor is multiplied element-wise with the cell state of the previous data point to obtain the history retention feature.

4. The method for monitoring rectifier transformer acoustic fingerprint faults based on physical information constraints according to claim 3, characterized in that, In S300, the adaptive wavelet transform specifically includes: Based on the current input signal and the hidden state of the previous data point, the wavelet scale parameter and wavelet translation parameter corresponding to the current data point are dynamically generated through network parameters, and the input modulation signal is calculated. Morlet wavelet basis functions are constructed using the wavelet scaling parameters and wavelet translation parameters to perform dual-domain decomposition of the audioprint time series sample set in both the time and frequency domains. The input modulation signal and the Morlet wavelet basis function are performed by tensor cross product operation and then superimposed on the history preservation feature to complete the cell state update.

5. The method for monitoring rectifier transformer acoustic fingerprint faults based on physical information constraints according to claim 1, characterized in that, In S300, the step of the attention mechanism layer performing global weighted fusion of the time-frequency hidden state includes: The time-frequency hidden state of each data point is nonlinearly mapped to obtain the internal representation vector of each data point; The inner product similarity between the internal representation vector of each data point and a learnable global fault context vector is calculated, and the attention weight of each data point is obtained after Softmax normalization. The time-frequency hidden states of all data points are weighted and summed according to the attention weights to obtain the time-frequency fusion feature vector.

6. The method for monitoring rectifier transformer acoustic fingerprint faults based on physical information constraints according to claim 1, characterized in that, In S400, the steps for constructing the data-driven loss include: By using the time-domain acoustic voltage signals collected by each acoustic sensor and combining them with their sound pressure sensitivity coefficients, the real part reference value and imaginary part reference value of the observed sound pressure at the corresponding characteristic frequency of each measuring point are obtained. The mean square error between the real and imaginary parts of each predicted complex sound pressure in the predicted complex sound pressure set and the reference values ​​of the real and imaginary parts of the observed sound pressure is calculated to construct the data-driven loss.

7. The method for monitoring rectifier transformer acoustic faults based on physical information constraints according to claim 1, characterized in that, In S400, the steps for constructing the residual loss of the Helmholtz frequency domain wave equation include: Within the solution domain, unlabeled configuration points are randomly sampled, and the equivalent sound velocity at each configuration point is pre-assigned based on the segmented medium properties of the rectifier transformer insulating oil region, the tank structural component region, and the air region. Substitute the predicted complex sound pressure set into the Helmholtz frequency domain wave equation, decompose the Helmholtz frequency domain wave equation into real part partial differential equation and imaginary part partial differential equation, calculate the second-order partial derivative of the predicted sound pressure with respect to the spatial coordinates by automatic differentiation, calculate the real part residual and imaginary part residual respectively, and construct the residual loss of the Helmholtz frequency domain wave equation.

8. The method for monitoring rectifier transformer acoustic fingerprint faults based on physical information constraints according to claim 1, characterized in that, In S400, the steps for constructing the Robin hybrid impedance boundary residual loss include: Based on the geometry of the rectifier transformer, boundary configuration points are pre-constructed at the interfaces between the insulating oil and the tank, and between the tank and the air medium, and the equivalent acoustic impedance is pre-determined according to the type of each medium interface and the corresponding characteristic frequency. Based on the impedance relationship between sound pressure and normal particle velocity, the Robin mixed impedance boundary condition is decoupled into real and imaginary residual terms. The normal derivative of the predicted sound pressure along the outer normal direction of the boundary is calculated by automatic differentiation, and the Robin mixed impedance boundary residual loss is constructed.

9. The method for monitoring rectifier transformer acoustic fingerprint faults based on physical information constraints according to claim 1, characterized in that, In S500, the steps for diagnosing the fault results of the output cooling fan include: The parameters of the trained physical information neural network are fixed, and the set of predicted complex sound pressures output by the physical information neural network is used to form a physically constrained sound field feature. The physical constraint acoustic field features are input into the fault classification layer, and the fault classification layer is trained with cross-entropy loss. The fault diagnosis result of the cooling fan is determined according to the category corresponding to the highest probability in the probability distribution output by the fault classification layer.

10. A rectifier transformer acoustic fault monitoring system based on physical information constraints, characterized in that, include: Voiceprint acquisition module: Acquires multi-channel voiceprint timing data during the operation of the cooling fan; Sample construction module: synchronously extracts and standardizes the multi-channel voiceprint time-series data to construct a voiceprint time-series sample set; Network construction module: Constructs a physical information neural network. The hidden layer of the physical information neural network includes a long short-term memory (LSTM) time-frequency memory layer, an attention mechanism layer, and a physical mapping layer connected in sequence. The LTM time-frequency memory layer incorporates adaptive wavelet transform during cell state update and sets a joint forgetting gate to extract time-frequency features from the voiceprint time-series sample set, obtaining the time-frequency hidden state of each data point. The attention mechanism layer performs global weighted fusion of the time-frequency hidden states to obtain a time-frequency fusion feature vector. The physical mapping layer maps the time-frequency fusion feature vector to a predicted complex sound pressure set. Physical constraint module: Taking the predicted complex sound pressure set as the constraint object, a total loss function is constructed, which is a weighted sum of data-driven loss, Helmholtz frequency domain wave equation residual loss and Robin hybrid impedance boundary residual loss. The physical information neural network is trained under the constraint of the total loss function. Fault diagnosis module: Input the features output by the trained physical information neural network into the fault classification layer, and output the fault diagnosis results of the cooling fan.

Images