A substation fault determination method and system based on voiceprint recognition

Abstract

The application discloses a transformer substation fault determination method and system based on voiceprint recognition, and relates to the technical field of voiceprint processing. The method comprises the following steps: collecting transformer original sound signal groups according to a preset period, mapping the pre-processed and optimized sound signal groups to a feature space to generate potential feature vectors, then reconstructing the sound signals and calculating reconstruction errors, subsequently, constructing a weight matrix and a correlation matrix based on the reconstructed sound signals and the optimized sound signals, fusing the difference value and the reconstruction error to obtain a fused difference value, distributing weights to the optimized and reconstructed sound signals according to the fused difference value and reconstructing the sound signals, and finally inputting the reconstructed sound signals into a fault detection model to output fault results. The sound signals of the transformer are periodically collected and pre-processed, the reconstruction error is calculated through feature mapping and reconstruction, the feature correlation between the reconstructed signals and the standard signals is enhanced to reduce the misjudgment rate, the accuracy of fault diagnosis is effectively improved, and the operation and maintenance efficiency of the transformer in the transformer substation is improved.

Description

Technical Field

[0001] This invention belongs to the field of voiceprint processing technology, specifically relating to a substation fault determination method and system based on voiceprint recognition. Background Technology

[0002] With the continuous upgrading of network communication and video surveillance technologies, remote equipment inspection supported by high-definition video systems has become a development trend in the substation operation and maintenance field. Traditional substation video systems mostly focus on security, fire protection, and equipment appearance overview and playback, with cameras concentrated in areas such as passageways and entrances / exits. They have not been optimized for online intelligent inspection and are difficult to adapt to the actual needs of intelligent operation and maintenance.

[0003] Transformers are the core equipment of substations and are of paramount importance. They are mainly responsible for the conversion of voltage levels and are the key hub for transmitting electrical energy from the generation side to the consumption side. They directly determine the power supply capacity and range of the substation. Their stable operation can ensure the quality of power supply and prevent voltage fluctuations from affecting electrical equipment. Once they fail and stop operating, they can easily cause power outages and affect the production and living order of the area.

[0004] The existing publication number is CN117056814A, which discloses a method for diagnosing transformer acoustic vibration faults. The method includes: acquiring acoustic vibration signals of the transformer under normal and fault conditions; classifying these signals to form an acoustic vibration data sample set; extracting digital features from the acquired acoustic vibration signal spectrum; building a network diagnostic model to be trained using a space-time feature extraction model; inputting the acquired transformer acoustic vibration signals into the network diagnostic model for training to obtain the network diagnostic model; and inputting the acquired faulty transformer acoustic vibration signals into the network diagnostic model to determine the fault type. This technique diagnoses faults solely by inputting digital features from the acoustic vibration signal spectrum into the network diagnostic model. However, abnormal features caused by some accidental factors may also be identified as fault features, leading to misjudgments.

[0005] Winding faults are the most common type of transformer fault. When transformer windings deform, they affect the vibration acoustic signature. Current technology collects acoustic signature signals and converts them into acoustic signature maps for transformer fault diagnosis. However, this method tends to ignore the influence of the environment on acoustic signature collection. For example, most substations are built on the outskirts of cities or in remote areas, where large temperature differences between day and night can affect the windings and generate acoustic signature signals with abnormal characteristics. During peak electricity consumption periods, changes in power load can cause large amplitude differences, which can also generate acoustic signature signals with abnormal characteristics. These abnormal characteristics are caused by accidental factors, but they can affect the diagnosis of real faults. If they are not removed, it can easily lead to misjudgment of transformer faults and affect the efficiency of substation operation and maintenance. Summary of the Invention

[0006] The purpose of this invention is to solve the problem that abnormal features may appear in the collected voiceprint signals due to accidental factors, which may lead to misdiagnosis of transformers. Therefore, this invention proposes a substation fault determination method and system based on voiceprint recognition.

[0007] In a first aspect of this invention, a substation fault determination method based on voiceprint recognition is first proposed, the method comprising: The original acoustic signals of the transformer are collected at a preset period and a set of original acoustic signals is formed. The original acoustic signal set is then preprocessed to obtain an optimized acoustic signal set. The optimized acoustic signal is mapped to the feature space to generate a latent feature vector. The optimized acoustic signal is reconstructed based on the latent feature vector to obtain a reconstructed acoustic signal. The reconstruction error is calculated based on the optimized acoustic signal and the reconstructed acoustic signal. The weight matrix and the correlation matrix are respectively constructed using the reconstructed acoustic signal and the optimized acoustic signal; The difference value is calculated based on the correlation matrix and the weight matrix, and the difference value is fused with the reconstruction error to obtain the fusion difference value; The optimized acoustic signal and the reconstructed acoustic signal are weighted according to the fusion difference and reconstructed to obtain a reconstructed acoustic signal. The reconstructed acoustic signal is then input into the fault detection model to output the fault result.

[0008] Optionally, the reconstruction error is calculated based on the optimized acoustic signal and the reconstructed acoustic signal, including: The optimized acoustic signal is processed by a 3-layer convolutional network with an activation function following each layer, which transforms the optimized acoustic signal from the temporal domain to the feature domain and outputs a local feature vector. The process involves mapping local feature vectors to the mean and variance of the latent space, and then generating latent feature vectors through reparameterization. in, Let i be the i-th latent feature vector. This represents the i-th optimized acoustic signal. Let be the mean of the potential space. The variance of the latent space. This is standard normally distributed noise; A 3-layer transposed convolutional network is used to gradually restore the existing feature dimension to be consistent with the optimized acoustic signal dimension, and output the transposed feature vector. The transposed feature vector is mapped to the mean and variance of the original space to obtain the reconstructed acoustic signal; The reconstruction error is calculated using the reconstructed acoustic signal and the optimized acoustic signal.

[0009] Optionally, a weight matrix and a correlation matrix are constructed using the reconstructed acoustic signal and the optimized acoustic signal, respectively, including: Based on the reconstructed acoustic signal, the query matrix and keyword matrix are calculated using two independent linear layer parameters respectively. The weight matrix is ​​obtained by transforming the query matrix and the keyword matrix. Based on the principle that the correlation between acoustic signals at adjacent time points is strong and the correlation weakens with increasing time intervals, an correlation matrix is ​​defined. The process is as follows: in, These are elements in the correlation matrix at time points i and j, where i and j represent time indices. Hyperparameters for controlling the associated decay rate; All calculated Generate an association matrix by matching the positions of row index i and column index j.

[0010] Optionally, the difference value is fused with the reconstruction error to obtain a fused difference value, including: The calculation process using the correlation matrix and the weight matrix is ​​as follows: Where D is the difference between the correlation matrix and the weight matrix. This represents the element in the i-th row and j-th column of the weight matrix. This represents the element in the i-th row and j-th column of the correlation matrix; The smoothed difference value is calculated by using a moving average with a window size of 5 on the difference value, and the smoothed difference value and the reconstruction error are weighted to obtain the fusion difference value.

[0011] Optionally, before inputting the reconstructed acoustic signal to the fault detection model to output the fault result label, the training process of the fault detection model includes: Obtain a training dataset, input the training dataset into a preset model for training to obtain model update parameters, and update the model parameters in the preset model according to the model update parameters to obtain a fault detection model; Obtain a verification dataset, input the verification data into the fault detection model to obtain verification results, and compare the verification results with the actual results; If the verification results are inconsistent with the actual results, the fault detection model is deemed unqualified and iterative training is re-executed until the preset conditions are met; otherwise, the fault detection model is deemed qualified. The verification dataset contains verification results and actual results, and the verification data corresponds one-to-one with the actual results.

[0012] In a second aspect of this invention, a substation fault determination system based on voiceprint recognition is proposed. The system includes an optimization module, a mapping module, a matrix construction module, a difference calculation module, and a fault determination module, wherein: The optimization module is used to collect the original acoustic signals of the transformer at a preset period, form an original acoustic signal set, and preprocess the original acoustic signal set to obtain an optimized acoustic signal set. The mapping module is used to map the optimized acoustic signal to the feature space to generate a potential feature vector, reconstruct the optimized acoustic signal according to the potential feature vector to obtain a reconstructed acoustic signal, and calculate the reconstruction error according to the optimized acoustic signal and the reconstructed acoustic signal. The matrix construction module is used to construct a weight matrix and a correlation matrix from the reconstructed acoustic signal and the optimized acoustic signal, respectively. The difference calculation module is used to calculate the difference value based on the correlation matrix and the weight matrix, and then fuse the difference value with the reconstruction error to obtain a fused difference value. The fault determination module is used to assign weights to the optimized acoustic signal and the reconstructed acoustic signal according to the fusion difference and reconstruct the reconstructed acoustic signal, and input the reconstructed acoustic signal into the fault detection model to output the fault result.

[0013] Optionally, the mapping module includes a first convolution module, a parameter calculation module, a second convolution module, a reconstructed signal module, and a reconstruction error calculation module, wherein: The first convolutional module is used to apply a 3-layer convolutional network to the optimized acoustic signal, with each layer followed by an activation function, to convert the optimized acoustic signal from the temporal domain to the feature domain and output a local feature vector. The parameter calculation module is used to map local feature vectors to the mean and variance of the latent space, and to generate latent feature vectors through reparameterization. The process is as follows: in, Let i be the i-th latent feature vector. This represents the i-th optimized acoustic signal. Let be the mean of the potential space. The variance of the latent space. This is standard normally distributed noise; The second convolution module is used to apply a 3-layer transposed convolutional network to the latent feature vector, gradually restoring the current feature dimension to be consistent with the optimized acoustic signal dimension and outputting the transposed feature. The reconstructed signal module is used to map the transposed features to the mean and variance of the original space and obtain the reconstructed acoustic signal. The reconstruction error calculation is used to calculate the reconstruction error using the reconstructed acoustic signal and the optimized acoustic signal.

[0014] Optionally, the matrix construction module includes a first matrix construction module, a matrix transformation module, a second matrix construction module, and a filling module, wherein: The first matrix construction module is used to calculate the query matrix and the keyword matrix based on the reconstructed acoustic signal using two independent linear layer parameters. The matrix transformation module is used to obtain a weight matrix by calculating and transforming the query matrix and the keyword matrix; The second matrix construction module is used to define an association matrix based on the principle that the association between adjacent time points of the optimized acoustic signal is strong, and the association weakens as the interval increases. The process is as follows: in, These are elements in the correlation matrix at time points i and j, where i and j represent time indices. Hyperparameters for controlling the associated decay rate; The filling module is used to fill all the calculated results. Generate an association matrix by matching the positions of row index i and column index j.

[0015] Optionally, the difference calculation module includes a difference value calculation module and a fusion difference module, wherein: The difference value calculation module is used to calculate the difference using the correlation matrix and the weight matrix, and the calculation process is as follows: Where D is the difference between the correlation matrix and the weight matrix. This represents the element in the i-th row and j-th column of the weight matrix. This represents the element in the i-th row and j-th column of the correlation matrix; The fusion difference module calculates a smoothed difference value by using a moving average with a window size of 5 on the difference value, and then calculates the fusion difference value by weighting the smoothed difference value and the reconstruction error.

[0016] Optionally, the fault determination module includes a model training module and a verification module, wherein: The model training module is used to acquire a training dataset, input the training dataset into a preset model for training to obtain model update parameters, and update the model parameters in the preset model according to the model update parameters to obtain a fault detection model. The verification module is used to acquire a verification dataset, input the verification data into the fault detection model to obtain verification results, and compare the verification results with the actual results. If the verification results are inconsistent with the actual results, the fault detection model is deemed unqualified and iterative training is re-executed until the preset conditions are met; otherwise, the fault detection model is deemed qualified. The verification dataset contains verification results and actual results, and the verification data corresponds one-to-one with the actual results.

[0017] The beneficial effects of this invention are: This invention proposes a substation fault diagnosis method and system based on voiceprint recognition. By improving signal purity through preprocessing, the optimized acoustic signal is mapped to the feature space to generate a potential feature vector. The acoustic signal can be reconstructed and the reconstruction error can be calculated, transforming the complex acoustic signal into an easily analyzable feature dimension. Key information reflecting the transformer's operating status is accurately extracted, the weight matrix highlights important feature components, and the correlation matrix captures the features within the signal. The difference value is calculated by combining the two and fused with the reconstruction error to obtain the fusion difference value. By associating the features within the reconstructed acoustic signal and the optimized acoustic signal, the misjudgment caused by fault features under random factors is reduced, effectively improving the comprehensiveness of fault diagnosis. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 A flowchart illustrating a substation fault determination method based on voiceprint recognition, provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a convolutional network provided in an embodiment of the present invention; Figure 3 This is a framework diagram of a substation fault determination system based on voiceprint recognition, provided for an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and B can represent: A alone, A and B simultaneously, and B alone. Furthermore, descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" can explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0021] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention provides a substation fault determination method based on voiceprint recognition. See also... Figure 1 , Figure 1 A flowchart illustrating a substation fault determination method based on voiceprint recognition, provided in an embodiment of the present invention. The method includes the following steps: S101: Collect the original sound signals of the transformer according to a preset cycle, and form an original sound signal set. Preprocess the original sound signal set to obtain an optimized sound signal set. S102, the optimized acoustic signal is mapped to the feature space to generate a latent feature vector, the optimized acoustic signal is reconstructed according to the latent feature vector to obtain the reconstructed acoustic signal, and the reconstruction error is calculated according to the optimized acoustic signal and the reconstructed acoustic signal. S103, the weight matrix and correlation matrix are constructed by reconstructing the acoustic signal and optimizing the acoustic signal, respectively; S104, calculate the difference value based on the correlation matrix and weight matrix, and fuse the difference value with the reconstruction error to obtain the fusion difference value; S105: Based on the fusion difference, the optimized acoustic signal and the reconstructed acoustic signal are weighted and reconstructed to obtain the reconstructed acoustic signal. The reconstructed acoustic signal is then input into the fault detection model to output the fault result.

[0023] The present invention provides a method and system for substation fault diagnosis based on voiceprint recognition. This method periodically collects and preprocesses the acoustic signals of transformers, calculates the reconstruction error through feature mapping and reconstruction, extracts key features of the acoustic signals, constructs a weight matrix and a correlation matrix using the reconstructed and optimized acoustic signals, enhances the correlation between the features of the two matrices, reduces the risk of misjudgment, obtains a fused acoustic signal by weighting according to the magnitude of the fusion difference, and inputs the fused acoustic signal into a fault detection model to obtain the fault result, thereby achieving transformer fault diagnosis and improving the operation and maintenance efficiency of transformers in substations.

[0024] In one implementation, both the optimized acoustic signal and the reconstructed acoustic signal contain accidental fault features caused by random fault factors and real fault features. The real fault features in the reconstructed acoustic signal are more prominent than those in the optimized acoustic signal, but some real fault features are lost compared to the optimized acoustic signal. By analyzing the fault score by correlating the reconstructed acoustic signal with the optimized acoustic signal, the correlation between the fault features in the reconstructed acoustic signal and the optimized acoustic signal can be effectively guaranteed, reducing the misjudgment of fault diagnosis by accidental fault features and solving the problem of low efficiency in transformer fault detection caused by insufficient correlation of fault features.

[0025] In one implementation, the fault result output by the fault detection module is whether a fault has occurred and the type of fault; the acoustic signal sampling period is set to collect the acoustic signal of the transformer within 4 hours, and the sampling device is a high-precision acoustic wave sensor; the fault result includes whether a fault has occurred and the type of fault.

[0026] In one implementation, the original acoustic signal set is preprocessed to obtain an optimized acoustic signal set, including: A filtering algorithm is used to remove high-frequency interference from the original sound signal in the original sound signal set to obtain the filtered original sound signal; Normalization is performed on each filtered original acoustic signal to output an optimized acoustic signal.

[0027] In one implementation, a wavelet denoising algorithm is used to filter high-frequency interference above 20kHz, while a low-overlap design is used to prevent fault features from breaking across windows. z-score normalization eliminates amplitude differences caused by different scenarios (such as day-night temperature difference and load changes), avoids mistaking amplitude fluctuations as fault features, improves the effectiveness of subsequent feature extraction, reduces redundant calculations, and lowers the risk of misjudgment.

[0028] In one embodiment, the reconstruction error is calculated based on the optimized acoustic signal and the reconstructed acoustic signal, including: A three-layer convolutional network is used to optimize the acoustic signal, with each layer followed by an activation function, to transform the optimized acoustic signal from the temporal domain to the feature domain and output local feature vectors. The process involves mapping local feature vectors to the mean and variance of the latent space, and then generating latent feature vectors through reparameterization. in, Let i be the i-th latent feature vector. This represents the i-th optimized acoustic signal. Let be the mean of the potential space. The variance of the latent space. This is standard normally distributed noise; A 3-layer transposed convolutional network is used to gradually restore the existing feature dimension to be consistent with the optimized acoustic signal dimension, and output the transposed feature vector. The transposed eigenvectors are mapped to the mean and variance of the original space to obtain the reconstructed acoustic signal; The reconstruction error is calculated by reconstructing and optimizing the acoustic signal.

[0029] In one implementation, see [link to implementation details]. Figure 2 , Figure 2The diagram shows the structure of the convolutional network provided in this embodiment of the invention. It uses a 3-layer convolutional network with 3 kernels, 1 channel, and 1 stride to perform convolution. This can capture short-term local correlations (such as instantaneous vibrations of sound signals) without causing feature blurring due to an excessively large receptive field. Each Conv layer is followed by a ReLU activation function to filter negative feature values.

[0030] In one implementation, the reconstruction error is calculated by reconstructing and optimizing the acoustic signal, and is calculated as follows: in, Let i be the reconstruction error. This represents the i-th optimized acoustic signal. Let i be the i-th reconstructed acoustic signal.

[0031] In one implementation, standardized data is mapped to a latent feature space. This process involves stripping redundant information and interference from high-dimensional, noisy transformer acoustic signals, extracting low-dimensional, probabilistic fault-related features, and then enhancing the local feature representation through multi-layer convolution and activation functions.

[0032] In one implementation, the acoustic signal is reconstructed from the latent feature vector to verify the effectiveness of the latent features (if the normal signal can be accurately reconstructed, it indicates that the feature extraction is reliable).

[0033] In one implementation, high-dimensionality reduction is achieved through mapping to improve computational efficiency and highlight the identification of fault features. The reconstructed acoustic signal not only ensures the quality of potential features but also provides a reliable foundation for subsequent weak correlation attention mechanisms and correlation inconsistency quantification, balancing local details and global temporal dependencies and improving the accuracy of fault detection.

[0034] In one embodiment, a weight matrix and a correlation matrix are constructed by reconstructing the acoustic signal and optimizing the acoustic signal, respectively, including: The query matrix and keyword matrix are calculated based on the reconstructed acoustic signal using two independent linear layer parameters. The weight matrix is ​​obtained by transforming the query matrix and keyword matrix. Based on the principle that the correlation between acoustic signals at adjacent time points is strong and the correlation weakens with increasing time intervals, an correlation matrix is ​​defined. The process is as follows: in, These are elements in the correlation matrix at time points i and j, where i and j represent time indices. Hyperparameters for controlling the associated decay rate; All calculated Generate an association matrix by matching the positions of row index i and column index j.

[0035] In one implementation, the weight matrix is ​​obtained by calculating and transforming the query matrix and keyword matrix. The process is as follows: Where Q is the query matrix and K is the keyword matrix, with the query matrix and keyword matrix having the same dimensions. Let A be the transpose of the keyword matrix, and A be the weight matrix. and All are linear layer parameters, where d represents the latent feature dimension of 128.

[0036] In one implementation, the reconstructed acoustic signal is mapped to a query matrix and a keyword matrix through two independent linear layer parameters, and a weight matrix is ​​calculated to highlight important features. At the same time, an association matrix is ​​defined based on the strength of the correlation at time points of the original acoustic signal to capture the time dependence of the acoustic signal, so that the matrix can accurately reflect the inherent correlation of the signal and improve the reliability of subsequent calculations.

[0037] In one embodiment, fusing the difference value with the reconstruction error to obtain a fused difference value includes: The calculation process using the correlation matrix and weight matrix is ​​as follows: Where D is the difference between the correlation matrix and the weight matrix. This represents the element in the i-th row and j-th column of the weight matrix. This represents the element in the i-th row and j-th column of the correlation matrix; The smoothed difference value is calculated by using a moving average with a window size of 5. The smoothed difference value and the reconstruction error are then weighted to obtain the fusion difference value.

[0038] In one implementation, the smoothing difference value and the reconstruction error are weighted to obtain the fusion difference value, which is calculated as follows: in, This represents the smooth difference value of the t-th reconstructed acoustic signal. Let t be the difference value of the reconstructed acoustic signal. This is represented as the difference value of the (t-1)th reconstructed acoustic signal. The smoothing factor was set based on experimental data. This represents the t-th fusion difference. and These are all weighted parameters, set using historical experimental data.

[0039] In one implementation, weights are assigned to the optimized acoustic signal and the reconstructed acoustic signal based on the fusion difference, and the assignment rule is as follows: If the fusion difference is greater than the preset fusion difference, it means that there are more accidental fault features caused by accidental factors in the optimized sound signal and fewer accidental fault features in the reconstructed sound signal. The weight coefficient of the optimized sound signal will be reduced and the weight coefficient of the reconstructed sound signal will be increased. If the fusion difference is not greater than the preset fusion difference, it means that there are fewer accidental fault features caused by accidental factors in the optimized acoustic signal and more accidental fault features in the reconstructed acoustic signal. The optimized acoustic signal weight coefficient will be increased and the reconstructed acoustic signal weight coefficient will be decreased. The reconstructed acoustic signal is obtained by weighted fusion of the optimized acoustic signal and the reconstructed acoustic signal by assigning weights to them. The weighting coefficients are preset in the database and weights are assigned according to rules to realize the influence of accidental fault characteristics on fault judgment, effectively reducing the risk of misjudgment and improving the accuracy of transformer diagnosis.

[0040] In one implementation, the fusion of difference values ​​and reconstruction errors leverages the strong discriminative power of reconstruction errors on fault signals to accurately capture abnormal features caused by signal distortion and constrain the occasional fault characteristics of difference values. The standardized weighted fusion transforms the two types of heterogeneous indicators into a unified comprehensive risk score, which quantifies the severity of the fault and significantly improves the robustness of fault diagnosis.

[0041] In one embodiment, the training process of the fault detection model before inputting the reconstructed acoustic signal to the fault detection model output fault result label includes: Obtain the training dataset, input the training dataset into the preset model to train and obtain the model update parameters, and update the model parameters in the preset model according to the model update parameters to obtain the fault detection model; Obtain the validation dataset, input the validation data into the fault detection model to obtain the validation results, and compare the validation results with the actual results; If the verification results are inconsistent with the actual results, the fault detection model is deemed unqualified and iterative training is re-executed until the preset conditions are met; otherwise, the fault detection model is deemed qualified. The verification dataset contains both verification results and actual results, and the verification data corresponds one-to-one with the actual results.

[0042] In one implementation, the dataset used is acoustic data collected in historical transformer acoustic fault experiments, and the dataset is divided into 80% training dataset and 20% validation dataset, with the preset model being a convolutional neural network model.

[0043] Based on the same inventive concept, this invention also provides a method and system for substation fault determination based on voiceprint recognition. See also Figure 3 , Figure 3 This invention provides a framework diagram of a substation fault determination system based on voiceprint recognition. The system includes an optimization module, a mapping module, a matrix construction module, a difference calculation module, and a fault determination module, wherein: The optimization module is used to collect the original acoustic signals of the transformer at a preset cycle, form an original acoustic signal set, and preprocess the original acoustic signal set to obtain an optimized acoustic signal set. The mapping module is used to map the optimized acoustic signal to the feature space to generate a latent feature vector, reconstruct the optimized acoustic signal based on the latent feature vector to obtain the reconstructed acoustic signal, and calculate the reconstruction error based on the optimized acoustic signal and the reconstructed acoustic signal. The matrix construction module is used to construct the weight matrix and the correlation matrix by reconstructing and optimizing the acoustic signal, respectively. The difference calculation module is used to calculate the difference value based on the correlation matrix and the weight matrix, and then fuse the difference value with the reconstruction error to obtain the fused difference value. The fault determination module is used to assign weights to the optimized acoustic signal and the reconstructed acoustic signal according to the fusion difference and reconstruct the reconstructed acoustic signal. The reconstructed acoustic signal is then input into the fault detection model to output the fault result.

[0044] The present invention provides a method and system for substation fault diagnosis based on voiceprint recognition. This method periodically collects and preprocesses the acoustic signals of transformers, calculates the reconstruction error through feature mapping and reconstruction, extracts key features of the acoustic signals, constructs a weight matrix and a correlation matrix using the reconstructed and optimized acoustic signals, enhances the correlation between the features of the two matrices, reduces the risk of misjudgment, obtains a fused acoustic signal by weighting according to the magnitude of the fusion difference, and inputs the fused acoustic signal into a fault detection model to obtain the fault result, thereby achieving transformer fault diagnosis and improving the operation and maintenance efficiency of transformers in substations.

[0045] In one embodiment, the mapping module includes a first convolution module, a parameter calculation module, a second convolution module, a reconstructed signal module, and a reconstruction error calculation module, wherein: The first convolutional module is used to apply a 3-layer convolutional network to the optimized acoustic signal, with each layer followed by an activation function, to transform the optimized acoustic signal from the temporal domain to the feature domain and output a local feature vector. The parameter calculation module maps local feature vectors to the mean and variance of the latent space, and generates latent feature vectors through reparameterization. The process is as follows: in, Let i be the i-th latent feature vector. This represents the i-th optimized acoustic signal. Let be the mean of the potential space. The variance of the latent space. This is standard normally distributed noise; The second convolutional module is used to apply a 3-layer transposed convolutional network to the latent feature vector, gradually restoring the current feature dimension to be consistent with the optimized acoustic signal dimension and outputting the transposed feature. The reconstructed signal module is used to map the transposed features to the mean and variance of the original space and obtain the reconstructed acoustic signal; Reconstruction error calculation is used to calculate the reconstruction error by reconstructing and optimizing the acoustic signal.

[0046] In one embodiment, the matrix construction module includes a first matrix construction module, a matrix transformation module, a second matrix construction module, and a filling module, wherein: The first matrix construction module is used to calculate the query matrix and the keyword matrix based on the reconstructed acoustic signal through two independent linear layer parameters; The matrix transformation module is used to obtain the weight matrix by calculating and transforming the query matrix and keyword matrix; The second matrix construction module is used to define an association matrix based on the principle that the association between adjacent acoustic signals at different time points is strong, and the association weakens as the interval increases. The process is as follows: in, These are elements in the correlation matrix at time points i and j, where i and j represent time indices. Hyperparameters for controlling the associated decay rate; The fill module is used to fill all the calculated results. Generate an association matrix by matching the positions of row index i and column index j.

[0047] In one embodiment, the difference calculation module includes a difference value calculation module and a fusion difference module, wherein: The difference calculation module is used to calculate the difference using the correlation matrix and the weight matrix. The calculation process is as follows: Where D is the difference between the correlation matrix and the weight matrix. This represents the element in the i-th row and j-th column of the weight matrix. This represents the element in the i-th row and j-th column of the correlation matrix; The fusion difference module calculates a smoothed difference value by using a moving average with a window size of 5 on the difference value, and then calculates the fusion difference value by weighting the smoothed difference value and the reconstruction error.

[0048] In one embodiment, the fault determination module includes a model training module and a verification module, wherein: The model training module is used to obtain the training dataset, input the training dataset into the preset model for training to obtain the model update parameters, and update the model parameters in the preset model according to the model update parameters to obtain the fault detection model. The verification module is used to acquire the verification dataset, input the verification data into the fault detection model to obtain the verification results, and compare the verification results with the actual results. If the verification results are inconsistent with the actual results, the fault detection model is deemed unqualified and iterative training is re-executed until the preset conditions are met; otherwise, the fault detection model is deemed qualified. The verification dataset contains both verification results and actual results, and the verification data corresponds one-to-one with the actual results.

[0049] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A substation fault determination method based on voiceprint recognition, characterized in that, The method includes: The original acoustic signals of the transformer are collected at a preset period and a set of original acoustic signals is formed. The original acoustic signal set is then preprocessed to obtain an optimized acoustic signal set. The optimized acoustic signal is mapped to the feature space to generate a latent feature vector. The optimized acoustic signal is reconstructed based on the latent feature vector to obtain a reconstructed acoustic signal. The reconstruction error is calculated based on the optimized acoustic signal and the reconstructed acoustic signal. The weight matrix and the correlation matrix are respectively constructed using the reconstructed acoustic signal and the optimized acoustic signal; The difference value is calculated based on the correlation matrix and the weight matrix, and the difference value is fused with the reconstruction error to obtain the fusion difference value; The optimized acoustic signal and the reconstructed acoustic signal are weighted according to the fusion difference and reconstructed to obtain a reconstructed acoustic signal. The reconstructed acoustic signal is then input into the fault detection model to output the fault result.

2. The substation fault determination method based on voiceprint recognition according to claim 1, characterized in that, The reconstruction error is calculated based on the optimized acoustic signal and the reconstructed acoustic signal, including: The optimized acoustic signal is processed by a 3-layer convolutional network with an activation function following each layer, which transforms the optimized acoustic signal from the temporal domain to the feature domain and outputs a local feature vector. The process involves mapping local feature vectors to the mean and variance of the latent space, and then generating latent feature vectors through reparameterization. in, Let i be the i-th latent feature vector. This represents the i-th optimized acoustic signal. Let be the mean of the potential space. The variance of the latent space. This is standard normally distributed noise; A 3-layer transposed convolutional network is used to gradually restore the existing feature dimension to be consistent with the optimized acoustic signal dimension, and output the transposed feature vector. The transposed feature vector is mapped to the mean and variance of the original space to obtain the reconstructed acoustic signal; The reconstruction error is calculated using the reconstructed acoustic signal and the optimized acoustic signal.

3. The substation fault determination method based on voiceprint recognition according to claim 1, characterized in that, The weight matrix and correlation matrix are constructed by the reconstructed acoustic signal and the optimized acoustic signal, respectively, including: Based on the reconstructed acoustic signal, the query matrix and keyword matrix are calculated using two independent linear layer parameters respectively. The weight matrix is ​​obtained by transforming the query matrix and the keyword matrix. Based on the principle that the correlation between acoustic signals at adjacent time points is strong and the correlation weakens with increasing time intervals, an correlation matrix is ​​defined. The process is as follows: in, These are elements in the correlation matrix at time points i and j, where i and j represent time indices. Hyperparameters for controlling the associated decay rate; All calculated Generate an association matrix by matching the positions of row index i and column index j.

4. The substation fault determination method based on voiceprint recognition according to claim 1, characterized in that, The fusion difference is obtained by fusing the difference value with the reconstruction error, including: The calculation process using the correlation matrix and the weight matrix is ​​as follows: Where D is the difference between the correlation matrix and the weight matrix. This represents the element in the i-th row and j-th column of the weight matrix. This represents the element in the i-th row and j-th column of the correlation matrix; The smoothed difference value is calculated by using a moving average with a window size of 5 on the difference value, and the smoothed difference value and the reconstruction error are weighted to obtain the fusion difference value.

5. The substation fault determination method based on voiceprint recognition according to claim 1, characterized in that, Before the reconstructed acoustic signal is input to the fault detection model to output the fault result label, the training process of the fault detection model includes: Obtain a training dataset, input the training dataset into a preset model for training to obtain model update parameters, and update the model parameters in the preset model according to the model update parameters to obtain a fault detection model; Obtain a verification dataset, input the verification data into the fault detection model to obtain verification results, and compare the verification results with the actual results; If the verification results are inconsistent with the actual results, the fault detection model is deemed unqualified and iterative training is re-executed until the preset conditions are met; otherwise, the fault detection model is deemed qualified. The verification dataset contains verification results and actual results, and the verification data corresponds one-to-one with the actual results.

6. A substation fault determination system based on voiceprint recognition, characterized in that, The system includes an optimization module, a mapping module, a matrix construction module, a difference calculation module, and a fault determination module, wherein: The optimization module is used to collect the original acoustic signals of the transformer at a preset period, form an original acoustic signal set, and preprocess the original acoustic signal set to obtain an optimized acoustic signal set. The mapping module is used to map the optimized acoustic signal to the feature space to generate a potential feature vector, reconstruct the optimized acoustic signal according to the potential feature vector to obtain a reconstructed acoustic signal, and calculate the reconstruction error according to the optimized acoustic signal and the reconstructed acoustic signal. The matrix construction module is used to construct a weight matrix and a correlation matrix from the reconstructed acoustic signal and the optimized acoustic signal, respectively. The difference calculation module is used to calculate the difference value based on the correlation matrix and the weight matrix, and then fuse the difference value with the reconstruction error to obtain a fused difference value. The fault determination module is used to assign weights to the optimized acoustic signal and the reconstructed acoustic signal according to the fusion difference and reconstruct the reconstructed acoustic signal, and input the reconstructed acoustic signal into the fault detection model to output the fault result.

7. A substation fault determination system based on voiceprint recognition according to claim 6, characterized in that, The mapping module includes a first convolution module, a parameter calculation module, a second convolution module, a reconstructed signal module, and a reconstruction error calculation module, wherein: The first convolutional module is used to apply a 3-layer convolutional network to the optimized acoustic signal, with each layer followed by an activation function, to convert the optimized acoustic signal from the temporal domain to the feature domain and output a local feature vector. The parameter calculation module is used to map local feature vectors to the mean and variance of the latent space, and to generate latent feature vectors through reparameterization. The process is as follows: in, Let i be the i-th latent feature vector. This represents the i-th optimized acoustic signal. Let be the mean of the potential space. The variance of the latent space. This is standard normally distributed noise; The second convolution module is used to apply a 3-layer transposed convolutional network to the latent feature vector, gradually restoring the current feature dimension to be consistent with the optimized acoustic signal dimension and outputting the transposed feature. The reconstructed signal module is used to map the transposed features to the mean and variance of the original space and obtain the reconstructed acoustic signal. The reconstruction error calculation is used to calculate the reconstruction error using the reconstructed acoustic signal and the optimized acoustic signal.

8. A substation fault determination system based on voiceprint recognition according to claim 6, characterized in that, The matrix construction module includes a first matrix construction module, a matrix transformation module, a second matrix construction module, and a filling module, wherein: The first matrix construction module is used to calculate the query matrix and the keyword matrix based on the reconstructed acoustic signal using two independent linear layer parameters. The matrix transformation module is used to obtain a weight matrix by calculating and transforming the query matrix and the keyword matrix; The second matrix construction module is used to define an association matrix based on the principle that the association between adjacent time points of the optimized acoustic signal is strong, and the association weakens as the interval increases. The process is as follows: in, These are elements in the correlation matrix at time points i and j, where i and j represent time indices. Hyperparameters for controlling the associated decay rate; The filling module is used to fill all the calculated results. Generate an association matrix by matching the positions of row index i and column index j.

9. A substation fault determination system based on voiceprint recognition according to claim 6, characterized in that, The difference calculation module includes a difference value calculation module and a fusion difference module, wherein: The difference value calculation module is used to calculate the difference using the correlation matrix and the weight matrix, and the calculation process is as follows: Where D is the difference between the correlation matrix and the weight matrix. This represents the element in the i-th row and j-th column of the weight matrix. This represents the element in the i-th row and j-th column of the correlation matrix; The fusion difference module calculates a smoothed difference value by using a moving average with a window size of 5 on the difference value, and then calculates the fusion difference value by weighting the smoothed difference value and the reconstruction error.

10. A substation fault determination system based on voiceprint recognition according to claim 6, characterized in that, The fault determination module includes a model training module and a verification module, wherein: The model training module is used to acquire a training dataset, input the training dataset into a preset model for training to obtain model update parameters, and update the model parameters in the preset model according to the model update parameters to obtain a fault detection model. The verification module is used to acquire a verification dataset, input the verification data into the fault detection model to obtain verification results, and compare the verification results with the actual results. If the verification results are inconsistent with the actual results, the fault detection model is deemed unqualified and iterative training is re-executed until the preset conditions are met; otherwise, the fault detection model is deemed qualified. The verification dataset contains verification results and actual results, and the verification data corresponds one-to-one with the actual results.

Images