Transformer mechanical fault early diagnosis method based on voiceprint

Abstract

The invention discloses a voiceprint-based early diagnosis method for a mechanical fault of a transformer. The method comprises the following steps: S1, collecting a voiceprint signal during operation of the transformer; s2, processing the voiceprint signal; s3, constructing a comprehensive feature vector; s4, calculating the matching degree between the feature vector and the reference feature in the fault sample library; s5, outputting a diagnosis result; according to the invention, by effectively separating the mechanical fault features of the transformer from environmental interference, the problem that a single-channel signal is easily polluted by noise is solved, a purer data source is provided for subsequent feature extraction, and the identifiability of early fault features is improved; by comprehensively capturing fault feature differences, the characterization capability of comprehensive feature vectors on early mechanical faults is remarkably improved; through hierarchical diagnosis and in combination with an early warning probability, common faults are rapidly matched by using a fault sample library, and special faults are accurately identified by virtue of deep learning, so that the accuracy and generalization ability of early diagnosis of mechanical faults are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of transformer mechanical fault diagnosis technology, specifically to an early diagnosis method for transformer mechanical faults based on acoustic signatures. Background Technology

[0002] In the power industry, transformers serve as crucial hubs for power transmission and distribution. Their stable operation plays a vital role in ensuring the reliability and continuity of power supply. A transformer failure can potentially trigger widespread power outages, severely impacting industrial production and residential lives, and even causing significant economic losses. Early diagnosis of transformer mechanical faults, promptly identifying potential problems such as loose windings and core displacement, is key to preventing fault escalation and reducing maintenance costs. With the application of acoustic signature technology in equipment diagnostics, analyzing faults by collecting acoustic signature signals during transformer operation has become a new research direction. However, traditional methods of analyzing faults using acoustic signature signals still have certain shortcomings:

[0003] First, traditional methods rely heavily on single-channel sensors to acquire signals. During the noise reduction process, the correlation between multi-channel signals is not fully considered, and there is a lack of effective constraints on the frequency consistency of signals from different channels, making it difficult to accurately distinguish mechanical fault characteristics from environmental noise. When decomposing the signal, there is a lack of a spatial consistency-based screening mechanism for the decomposed modal components, which makes it easy for early weak fault characteristics to be masked by noise, and cannot provide a clean signal basis for subsequent analysis.

[0004] Secondly, feature extraction is often limited to a single scale or a single domain, without differentiating the weights based on the importance of different features, making it difficult to fully capture the multidimensional characteristics of mechanical faults. At the same time, the fusion methods for multi-channel features are relatively simple and lack targeted enhancement of fault-sensitive information, resulting in the fused feature vectors failing to effectively represent the subtle differences in early mechanical faults and affecting the accuracy of diagnosis.

[0005] Finally, the diagnostic process relies solely on the fault sample library for feature matching, which limits the ability to identify rare faults outside the sample library; or a single model is used for diagnosis, but the loss function design of the model does not fully consider the complexity of the samples, resulting in insufficient generalization ability; in addition, there is a lack of dynamic quantitative analysis of fault development trends, and it is impossible to combine time factors and fault accumulation effects to give accurate early warning probabilities, which is difficult to meet the needs of early prediction of fault risks in actual operation and maintenance.

[0006] Therefore, it is necessary to design an early diagnosis method for transformer mechanical faults based on acoustic signatures. Summary of the Invention

[0007] The purpose of this invention is to provide an early diagnosis method for transformer mechanical faults based on acoustic signatures. This addresses the shortcomings of traditional methods mentioned in the background, which often rely on single-channel sensors for signal acquisition. These methods fail to adequately consider the correlation between multi-channel signals during noise reduction, lack effective constraints on the frequency consistency of different channel signals, and struggle to accurately distinguish mechanical fault features from environmental noise. Furthermore, during signal decomposition, the methods lack a spatial consistency-based screening mechanism for the decomposed modal components, leading to early, weak fault features being easily masked by noise, failing to provide a clean signal foundation for subsequent analysis. Feature extraction is often limited to a single scale or domain, without differentiated weighting based on the importance of different features, making it difficult to comprehensively capture mechanical fault characteristics. The multi-dimensional characteristics of mechanical failures are a concern. Furthermore, the fusion methods for multi-channel features are relatively simple and lack targeted enhancement of fault-sensitive information. This results in the fused feature vectors failing to effectively represent subtle differences in early mechanical failures, affecting diagnostic accuracy. Additionally, relying solely on a fault sample library for feature matching during diagnosis limits the ability to identify rare faults outside the sample library. Alternatively, using a single model for diagnosis may result in inadequate generalization ability due to the model's loss function design not fully considering sample complexity. Moreover, the lack of dynamic quantitative analysis of fault development trends makes it impossible to combine time factors and the cumulative effect of faults to provide accurate early warning probabilities, failing to meet the needs of early fault risk prediction in actual operation and maintenance.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for early diagnosis of mechanical faults in transformers based on acoustic signatures, comprising the following steps:

[0009] S1: Acoustic signals from the transformer during operation are collected using an array of acoustic sensors to construct an acoustic signal matrix. ,in This refers to the number of sensors, i.e., the number of channels. The number of sampling points for each channel; the array-type acoustic sensors are arranged in a ring, evenly distributed around the center of the transformer core, to ensure coverage of the main vibration sources of the equipment: windings, core, and fasteners; during sampling, a synchronous triggering mechanism is used to control the sampling time deviation of each sensor, ensuring the time consistency of the matrix row vectors and avoiding feature distortion caused by misalignment of multi-channel signals;

[0010] S2: Based on the voiceprint signal matrix The multi-channel correlation is analyzed, and adaptive noise reduction is performed on the acoustic fingerprint signals collected by each sensor in the matrix. The Variational Mode Decomposition (VMD) algorithm is used to decompose the signal of each channel, obtaining multiple modal components. Spatial consistency screening is then used to retain modal components containing mechanical fault characteristics. These retained modal components are then reconstructed into the noise-reduced signals of each channel. Compared to the traditional EMD algorithm, the VMD algorithm has the advantage of anti-modal mixing and can specifically separate periodic vibrations and impact noise in transformer acoustic fingerprints. Spatial consistency screening is based on the principle that mechanical fault characteristics represent global equipment vibration and that multi-channel signals should have frequency consistency, using a threshold... Excluding occasional noise in a single channel, such as environmental interference, only retaining the characteristic modes common to multiple channels; the reconstruction process recovers the temporal characteristics of the signal by superimposing the modal components, providing a reliable data source for subsequent feature extraction;

[0011] S3: Extract multi-scale weighted kurtosis from the denoised signals of each channel after reconstruction. Time-frequency domain entropy, frequency domain peak ratio And short-time energy entropy, through principal component analysis to fuse multi-channel features, construct a comprehensive feature vector. Multi-scale weighted kurtosis captures the impact characteristics of early mechanical failures at different time scales, with adaptive weights. It adaptively adjusts to signal fluctuations, enhancing sensitivity to minor faults; time-frequency domain entropy is based on Hilbert-Huang transform, quantifying signal complexity through marginal spectral entropy values; frequency domain peak ratio focuses on the main vibration frequency range of mechanical faults such as winding loosening and core displacement, highlighting the low-frequency characteristics of early faults through frequency band weighting; principal component analysis introduces fault sensitivity factors, enhancing the distinguishability of fault features by maximizing the inter-class divergence and intra-class divergence ratios.

[0012] S4: Calculate feature vectors using cosine similarity. Reference features in the fault sample library Matching degree; importance weight in cosine similarity calculation. Features obtained through training with the random forest algorithm are given higher weights for those sensitive to mechanical faults; reference features from the fault sample database. The stability of the baseline features is ensured by taking the average value of multiple tests with the same type of fault; the sample library supports incremental learning, and new samples need to be tested three times to verify the consistency of features, so as to avoid errors introduced by abnormal samples.

[0013] S5: If the matching degree reaches the threshold, output the corresponding fault type and early warning probability. If there is no match, a deep belief network (DBN) is used for further identification and diagnostic results are output; early warning probability. Based on the current matching degree and sampling time and historical trends, time decay factor With trend weight Optimized based on extensive experimental data, the deep belief network (DBN) ensures cumulative sensitivity to slowly developing early faults such as gradual loosening of fasteners. It achieves nonlinear feature mapping through three hidden layers, enhances the capture of weak features using the ReLU activation function, and incorporates sample weighting into the loss function. This improves the accuracy of identifying special fault types.

[0014] As a further technical solution of the present invention, in S2, VMD decomposition is performed on the signal of each channel in the voiceprint signal matrix. The variational mode decomposition (VMD) algorithm is as follows:

[0015]

[0016] in, For the first Channel 1 One modal component; For the first Channel 1 The center frequency of each modal component; It is the Dirac function; This is a convolution operation; The imaginary unit; The first in the voiceprint signal matrix The raw signal of the channel; The number of modes; As a penalty factor; The inter-channel frequency consistency penalty coefficient is dynamically optimized using an improved particle swarm optimization algorithm. , , ;

[0017] When combining spatial consistency screening to retain modal components containing mechanical fault characteristics, for each modal component... Calculate the center frequency difference of the same mode in adjacent channels. When the difference is less than a set threshold When the modal component is determined to have spatial consistency, it is retained.

[0018] The retained modal components are reconstructed into the denoised signals of each channel, i.e., the... The noise-reduced signal of the channel ,in The set of indexes for the reserved modal components.

[0019] As a further technical solution of the present invention, the fitness function of the particle swarm optimization algorithm is:

[0020]

[0021] in, , These are the weighting coefficients. ; The signal-to-noise ratio after noise reduction; For the first The effective signal energy of the channel is calculated by reconstructing the first channel using the Variational Mode Decomposition (VMD) algorithm. The total energy of the channel signals is obtained; For the first The total energy of the channel is calculated by the acoustic signal matrix. The Middle The total energy of the original acquired signals from the channel is obtained; This is a channel frequency consistency indicator. For the first Channel 1 The center frequency of each modal component For the first Channel 1 The center frequency of each modal component, exponential function Used to quantify the difference in center frequency of the same mode in adjacent channels.

[0022] As a further technical solution of the present invention, in step S3, the principal component analysis formula is:

[0023]

[0024] in, This forms the comprehensive feature vector, i.e., the feature vector to be diagnosed. It is a multi-channel feature matrix; Let the projection matrix be the objective function. Solve this problem; The scatter matrix within the fault sample class; The inter-class scatter matrix; Projection matrix The Column vector; The dimension is the dimension after dimensionality reduction.

[0025] As a further technical solution of the present invention, in S3, the formula for calculating the peak-to-peak ratio in the frequency domain is:

[0026]

[0027] in, For frequency; For frequency band weighting, 1.2 is used for 100~300Hz and 0.8 is used for 300~500Hz; This refers to the peak amplitude within the corresponding frequency band; The signal amplitude is at a base frequency of 50Hz.

[0028] As a further technical solution of the present invention, in step S3, the multi-scale weighted kurtosis formula is:

[0029]

[0030] in, The scale number; For the first Signal at a specific scale; For the first The mean of the scale signal; For mathematical expectation operators; For the first Standard deviation of the scale signal; For the first The adaptive weights of the scale are inversely proportional to the signal fluctuations; the greater the fluctuation, the higher the weight.

[0031] As a further technical solution of the present invention, in step S4, the cosine similarity calculation formula is:

[0032]

[0033] in, The feature vector to be diagnosed is obtained by extracting multi-scale weighted kurtosis from the denoised signals of each channel. Time-frequency domain entropy, frequency domain peak ratio The multidimensional features of short-time energy entropy are fused through principal component analysis to form a comprehensive feature vector; The reference feature vector in the fault sample library is the transformer acoustic signature signal of known fault types, which is used in conjunction with... A completely consistent feature extraction and fusion process is used, and the results are pre-calculated and stored in the fault sample library as a benchmark template for fault matching. For the first Dimensional features, for The Dimensional features; for Norm; For the first Importance weights of dimensional features; This is the attenuation coefficient.

[0034] As a further technical solution of the present invention, in step S4, the construction rules for the fault sample library are as follows:

[0035] Acoustic signature characteristics of three typical mechanical faults in transformers, including ≥50 samples: loose windings, core displacement, and fastener detachment;

[0036] Reference feature vector for each fault type based on the Euclidean distance of the feature vectors. The following conditions must be met: the standard deviation of features for samples of the same type is ≤0.1;

[0037] Conditions for adding new samples: minimum cosine similarity with existing samples ≤ 0.7, ensuring sample diversity, and feature stability verified by three repeated tests.

[0038] As a further technical solution of the present invention, in step S5, the network structure and loss function of the Deep Belief Network (DBN) are as follows:

[0039] Network structure: Contains: 1 input layer, the number of nodes is the dimension after dimensionality reduction. There are 3 hidden layers with 128, 64, 32, and 1 output layer with the number of nodes equal to the number of fault categories; the activation function for the hidden layers is ReLU, and the activation function for the output layers is Softmax.

[0040] Loss function:

[0041]

[0042] in, The number of samples; For the first The true label of each sample; For the first The predicted probability of each sample; For sample weights, the weights for difficult samples are higher. For the first The predicted probability of the true label of each sample; This is the weight decay coefficient; This is the marginal loss coefficient; The positive and negative sample interval; , are the predicted probabilities of the positive and negative categories corresponding to the i-th sample, respectively.

[0043] As a further technical solution of the present invention, in S5, the early warning probability The calculation formula is:

[0044]

[0045] in, This represents the current matching degree. This represents the current signal acquisition duration. This is the time decay factor; This represents the historical number of samples. For the first Secondary matching degree; The trend weight reflects the cumulative effect of faults.

[0046] Compared with existing technologies, the beneficial effects of this voiceprint-based early diagnosis method for transformer mechanical faults are:

[0047] By constructing a multi-channel acoustic signal matrix The method of introducing a frequency consistency penalty coefficient between channels is adopted. The variational mode decomposition (VMD) algorithm is used for adaptive noise reduction. Combined with spatial consistency screening, the modal components containing mechanical fault features are retained and reconstructed into the noise-reduced signal. This effectively separates the mechanical fault features of the transformer from environmental interference, solves the problem that single-channel signals are easily contaminated by noise, provides a cleaner data source for subsequent feature extraction, and improves the identifiability of early fault features.

[0048] Scale-adaptive weights are extracted from the reconstructed denoised signal. Multiscale weighted kurtosis Time-frequency domain entropy, frequency band weight Peak frequency ratio and short-time energy entropy, and then through the objective function Principal component analysis is used to enhance fault-sensitive information, and multi-channel features are integrated to construct a comprehensive feature vector. Compared with single feature extraction, this method comprehensively captures the differences in fault features, significantly improving the ability of the comprehensive feature vector to represent early mechanical faults.

[0049] By introducing importance weights cosine similarity With optimization loss function Deep belief networks (DBNs) enable hierarchical diagnostics, combined with an integrated time decay factor. and trend weight Early warning probability It utilizes a fault sample library to quickly match common faults and leverages deep learning to accurately identify special faults, significantly improving the accuracy and generalization ability of early diagnosis of mechanical faults. At the same time, it dynamically quantifies fault risks, enhancing the accuracy and practicality of early diagnosis. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0051] 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. 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.

[0052] Please see the appendix Figure 1 The present invention provides an embodiment of an early diagnosis method for transformer mechanical faults based on voiceprints, comprising the following steps:

[0053] S1: Acoustic signals from the transformer during operation are collected using an array of acoustic sensors to construct an acoustic signal matrix. ,in This refers to the number of sensors, i.e., the number of channels. The number of sampling points for each channel; the array-type acoustic sensors are arranged in a ring, evenly distributed around the center of the transformer core, to ensure coverage of the main vibration sources of the equipment: windings, core, and fasteners; during sampling, a synchronous triggering mechanism is used to control the sampling time deviation of each sensor, ensuring the time consistency of the matrix row vectors and avoiding feature distortion caused by misalignment of multi-channel signals;

[0054] S2: Based on the voiceprint signal matrix The multi-channel correlation is analyzed, and adaptive noise reduction is performed on the acoustic fingerprint signals collected by each sensor in the matrix. The Variational Mode Decomposition (VMD) algorithm is used to decompose the signal of each channel, obtaining multiple modal components. Spatial consistency screening is then used to retain modal components containing mechanical fault characteristics. These retained modal components are then reconstructed into the noise-reduced signals of each channel. Compared to the traditional EMD algorithm, the VMD algorithm has the advantage of anti-modal mixing and can specifically separate periodic vibrations and impact noise in transformer acoustic fingerprints. Spatial consistency screening is based on the principle that mechanical fault characteristics represent global equipment vibration and that multi-channel signals should have frequency consistency, using a threshold... Excluding occasional noise in a single channel, such as environmental interference, only retaining the characteristic modes common to multiple channels; the reconstruction process recovers the temporal characteristics of the signal by superimposing the modal components, providing a reliable data source for subsequent feature extraction;

[0055] The variational mode decomposition (VMD) algorithm is as follows:

[0056]

[0057] in, For the first Channel 1 One modal component; For the first Channel 1 The center frequency of each modal component; It is the Dirac function; This is a convolution operation; The imaginary unit; The first in the voiceprint signal matrix The raw signal of the channel; The number of modes; As a penalty factor; The inter-channel frequency consistency penalty coefficient is dynamically optimized using an improved particle swarm optimization algorithm. , , ;

[0058] The fitness function of the particle swarm optimization algorithm is:

[0059]

[0060] in, , These are the weighting coefficients. ; The signal-to-noise ratio after noise reduction; For the first The effective signal energy of the channel is calculated by reconstructing the first channel using the Variational Mode Decomposition (VMD) algorithm. The total energy of the channel signals is obtained; For the first The total energy of the channel is calculated by the acoustic signal matrix. The Middle The total energy of the original acquired signals from the channel is obtained; This is a channel frequency consistency indicator. For the first Channel 1 The center frequency of each modal component For the first Channel 1 The center frequency of each modal component, exponential function Used to quantify the difference in the center frequency of the same mode in adjacent channels;

[0061] When combining spatial consistency screening to retain modal components containing mechanical fault characteristics, for each modal component... Calculate the center frequency difference of the same mode in adjacent channels. When the difference is less than a set threshold When the modal component is determined to have spatial consistency, it is retained.

[0062] The retained modal components are reconstructed into the denoised signals of each channel, i.e., the... The noise-reduced signal of the channel ,in The set of indexes for the retained modal components;

[0063] S3: Extract multi-scale weighted kurtosis from the denoised signals of each channel after reconstruction. Time-frequency domain entropy, frequency domain peak ratio And short-time energy entropy, through principal component analysis to fuse multi-channel features, construct a comprehensive feature vector. Multi-scale weighted kurtosis captures the impact characteristics of early mechanical failures at different time scales, with adaptive weights. It adaptively adjusts to signal fluctuations, enhancing sensitivity to minor faults; time-frequency domain entropy is based on Hilbert-Huang transform, quantifying signal complexity through marginal spectral entropy values; frequency domain peak ratio focuses on the main vibration frequency range of mechanical faults such as winding loosening and core displacement, highlighting the low-frequency characteristics of early faults through frequency band weighting; principal component analysis introduces fault sensitivity factors, enhancing the distinguishability of fault features by maximizing the inter-class divergence and intra-class divergence ratios.

[0064] The principal component analysis formula is:

[0065]

[0066] in, This forms the comprehensive feature vector, i.e., the feature vector to be diagnosed. It is a multi-channel feature matrix; Let the projection matrix be the objective function. Solve this problem; The scatter matrix within the fault sample class; The inter-class scatter matrix; Projection matrix The Column vector; The dimension after dimensionality reduction;

[0067] The formula for calculating the peak-to-peak ratio in the frequency domain is:

[0068]

[0069] in, For frequency; For frequency band weighting, 1.2 is used for 100~300Hz and 0.8 is used for 300~500Hz; This refers to the peak amplitude within the corresponding frequency band; The signal amplitude is at a base frequency of 50Hz.

[0070] The multi-scale weighted kurtosis formula is:

[0071]

[0072] in, The scale number; For the first Signal at a specific scale; For the first The mean of the scale signal; For mathematical expectation operators; For the first Standard deviation of the scale signal; For the first The adaptive weights of the scale are inversely proportional to the signal fluctuations; the greater the fluctuation, the higher the weight.

[0073] S4: Calculate feature vectors using cosine similarity. Reference features in the fault sample library Matching degree; importance weight in cosine similarity calculation. Features obtained through training with the random forest algorithm are given higher weights for those sensitive to mechanical faults; reference features from the fault sample database. The stability of the baseline features is ensured by taking the average value of multiple tests with the same type of fault; the sample library supports incremental learning, and new samples need to be tested three times to verify the consistency of features, so as to avoid errors introduced by abnormal samples.

[0074] The formula for calculating cosine similarity is:

[0075]

[0076] in, The feature vector to be diagnosed is obtained by extracting multi-scale weighted kurtosis from the denoised signals of each channel. Time-frequency domain entropy, frequency domain peak ratio The multidimensional features of short-time energy entropy are fused through principal component analysis to form a comprehensive feature vector; The reference feature vector in the fault sample library is the transformer acoustic signature signal of known fault types, which is used in conjunction with... A completely consistent feature extraction and fusion process is used, and the results are pre-calculated and stored in the fault sample library as a benchmark template for fault matching. For the first Dimensional features, for The Dimensional features; for Norm; For the first Importance weights of dimensional features; The attenuation coefficient;

[0077] The rules for constructing the fault sample library are as follows:

[0078] Acoustic signature characteristics of three typical mechanical faults in transformers, including ≥50 samples: loose windings, core displacement, and fastener detachment;

[0079] Reference feature vector for each fault type based on the Euclidean distance of the feature vectors. The following conditions must be met: the standard deviation of features for samples of the same type is ≤0.1;

[0080] Conditions for adding new samples: the minimum cosine similarity with existing samples is ≤0.7 to ensure sample diversity, and the stability of features is verified by three repeated tests;

[0081] The network structure and loss function of Deep Belief Network (DBN) are as follows:

[0082] Network structure: Contains: 1 input layer, the number of nodes is the dimension after dimensionality reduction. There are 3 hidden layers with 128, 64, 32, and 1 output layer with the number of nodes equal to the number of fault categories; the activation function for the hidden layers is ReLU, and the activation function for the output layers is Softmax.

[0083] Loss function:

[0084]

[0085] in, The number of samples; For the first The true label of each sample; For the first The predicted probability of each sample; For sample weights, the weights for difficult samples are higher. For the first The predicted probability of the true label of each sample; This is the weight decay coefficient; This is the marginal loss coefficient; The positive and negative sample interval; , , respectively, are the predicted probabilities of the positive and negative categories corresponding to the i-th sample;

[0086] S5: If the matching degree reaches the threshold, output the corresponding fault type and early warning probability. If there is no match, a deep belief network (DBN) is used for further identification and diagnostic results are output; early warning probability. Based on the current matching degree and sampling time and historical trends, time decay factor With trend weight Optimized based on extensive experimental data, the deep belief network (DBN) ensures cumulative sensitivity to slowly developing early faults such as gradual loosening of fasteners. It achieves nonlinear feature mapping through three hidden layers, enhances the capture of weak features using the ReLU activation function, and incorporates sample weighting into the loss function. This improves the accuracy of identifying special fault types;

[0087] Early warning probability The calculation formula is:

[0088]

[0089] in, This represents the current matching degree. This represents the current signal acquisition duration. This is the time decay factor; This represents the historical number of samples. For the first Secondary matching degree; The trend weight reflects the cumulative effect of faults.

[0090] In summary, this invention constructs a multi-channel acoustic signature signal matrix. The method of introducing a frequency consistency penalty coefficient between channels is adopted. The variational mode decomposition (VMD) algorithm is used for adaptive noise reduction. Combined with spatial consistency screening, the modal components containing mechanical fault features are retained and reconstructed into the noise-reduced signal. This effectively separates the mechanical fault features of the transformer from environmental interference, solves the problem that single-channel signals are easily contaminated by noise, provides a cleaner data source for subsequent feature extraction, and improves the identifiability of early fault features.

[0091] Scale-adaptive weights are extracted from the reconstructed denoised signal. Multiscale weighted kurtosis Time-frequency domain entropy, frequency band weight Peak frequency ratio and short-time energy entropy, and then through the objective function Principal component analysis is used to enhance fault-sensitive information, and multi-channel features are integrated to construct a comprehensive feature vector. Compared with single feature extraction, this method comprehensively captures the differences in fault features, significantly improving the ability of the comprehensive feature vector to represent early mechanical faults.

[0092] By introducing importance weights cosine similarity With optimization loss function Deep belief networks (DBNs) enable hierarchical diagnostics, combined with an integrated time decay factor. and trend weight Early warning probability It utilizes a fault sample library to quickly match common faults and leverages deep learning to accurately identify special faults, significantly improving the accuracy and generalization ability of early diagnosis of mechanical faults. At the same time, it dynamically quantifies fault risks, enhancing the accuracy and practicality of early diagnosis.

[0093] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for early diagnosis of mechanical faults in transformers based on voiceprints, characterized in that: Includes the following steps: S1: Acoustic signals from the transformer during operation are collected using an array of acoustic sensors to construct an acoustic signal matrix. ,in This refers to the number of sensors, i.e., the number of channels. The number of sampling points for each channel; S2: Based on the voiceprint signal matrix The multi-channel correlation is used to adaptively denoise the acoustic signals collected by each sensor in the matrix. The signal of each channel is decomposed by variational mode decomposition (VMD) algorithm to obtain multiple modal components. The modal components containing mechanical fault characteristics are retained by spatial consistency screening. The retained modal components are reconstructed into the denoised signals of each channel. S3: Extract multi-scale weighted kurtosis from the denoised signals of each channel after reconstruction. Time-frequency domain entropy, frequency domain peak ratio And short-time energy entropy, through principal component analysis to fuse multi-channel features, construct a comprehensive feature vector. ; S4: Calculate feature vectors using cosine similarity. Reference features in the fault sample library The degree of matching; S5: If the matching degree reaches the threshold, output the corresponding fault type and early warning probability. If there is no match, a deep belief network (DBN) is used for further identification and diagnostic results are output.

2. The method for early diagnosis of transformer mechanical faults based on voiceprint as described in claim 1, characterized in that: In step S2, VMD decomposition is performed on the signal of each channel in the voiceprint signal matrix. The variational mode decomposition (VMD) algorithm is as follows: in, For the first Channel 1 One modal component; For the first Channel 1 The center frequency of each modal component; It is the Dirac function; This is a convolution operation; The imaginary unit; The first in the voiceprint signal matrix The raw signal of the channel; The number of modes; As a penalty factor; The inter-channel frequency consistency penalty coefficient is dynamically optimized using an improved particle swarm optimization algorithm. , , ; When combining spatial consistency screening to retain modal components containing mechanical fault characteristics, for each modal component... Calculate the center frequency difference of the same mode in adjacent channels. When the difference is less than a set threshold When the modal component is determined to have spatial consistency, it is retained. The retained modal components are reconstructed into the denoised signals of each channel, i.e., the... The noise-reduced signal of the channel ,in The set of indexes for the reserved modal components.

3. The method for early diagnosis of transformer mechanical faults based on voiceprint according to claim 2, characterized in that: The fitness function of the particle swarm optimization algorithm is: in, , These are the weighting coefficients. ; The signal-to-noise ratio after noise reduction; For the first The effective signal energy of the channel is calculated by reconstructing the first channel using the Variational Mode Decomposition (VMD) algorithm. The total energy of the channel signals is obtained; For the first The total energy of the channel is calculated by the acoustic signal matrix. The Middle The total energy of the original acquired signals from the channel is obtained; This is a channel frequency consistency indicator. For the first Channel 1 The center frequency of each modal component For the first Channel 1 The center frequency of each modal component, exponential function Used to quantify the difference in center frequency of the same mode in adjacent channels.

4. The method for early diagnosis of transformer mechanical faults based on voiceprint as described in claim 1, characterized in that: In S3, the principal component analysis formula is: in, This forms the comprehensive feature vector, i.e., the feature vector to be diagnosed. It is a multi-channel feature matrix; Let the projection matrix be the objective function. Solve this problem; The scatter matrix within the fault sample class; The inter-class scatter matrix; Projection matrix The Column vector; The dimension is the dimension after dimensionality reduction.

5. The method for early diagnosis of transformer mechanical faults based on voiceprint according to claim 1, characterized in that: In S3, the formula for calculating the peak-to-peak ratio in the frequency domain is: in, For frequency; For frequency band weighting, 1.2 is used for 100~300Hz and 0.8 is used for 300~500Hz; This refers to the peak amplitude within the corresponding frequency band; The signal amplitude is at a base frequency of 50Hz.

6. The method for early diagnosis of transformer mechanical faults based on voiceprint according to claim 1, characterized in that: In S3, the multi-scale weighted kurtosis formula is: in, The scale number; For the first Signal at a specific scale; For the first The mean of the scale signal; For mathematical expectation operators; For the first Standard deviation of the scale signal; For the first The adaptive weights of the scale are inversely proportional to the signal fluctuations; the greater the fluctuation, the higher the weight.

7. The method for early diagnosis of transformer mechanical faults based on acoustic signatures according to claim 1, characterized in that: In S4, the formula for calculating the cosine similarity is: in, The feature vector to be diagnosed is obtained by extracting multi-scale weighted kurtosis from the denoised signals of each channel. Time-frequency domain entropy, frequency domain peak ratio The multidimensional features of short-time energy entropy are fused through principal component analysis to form a comprehensive feature vector; The reference feature vector in the fault sample library is the transformer acoustic signature signal of known fault types, which is used in conjunction with... A completely consistent feature extraction and fusion process is used, and the results are pre-calculated and stored in the fault sample library as a benchmark template for fault matching. For the first Dimensional features, for The Dimensional features; for Norm; For the first Importance weights of dimensional features; This is the attenuation coefficient.

8. The method for early diagnosis of transformer mechanical faults based on voiceprint according to claim 1, characterized in that: In S4, the rules for constructing the fault sample library are as follows: Acoustic signature characteristics of three typical mechanical faults in transformers, including ≥50 samples: loose windings, core displacement, and fastener detachment; Reference feature vector for each fault type based on the Euclidean distance of the feature vectors. The following conditions must be met: the standard deviation of features for samples of the same type is ≤0.1; Conditions for adding new samples: minimum cosine similarity with existing samples ≤ 0.7, ensuring sample diversity, and feature stability verified by three repeated tests.

9. The method for early diagnosis of transformer mechanical faults based on voiceprint according to claim 1, characterized in that: In step S5, the network structure and loss function of the Deep Belief Network (DBN) are as follows: Network structure: Contains: 1 input layer, the number of nodes is the dimension after dimensionality reduction. There are 3 hidden layers with 128, 64, 32, and 1 output layer with the number of nodes equal to the number of fault categories; the activation function for the hidden layers is ReLU, and the activation function for the output layers is Softmax. Loss function: in, The number of samples; For the first The true label of each sample; For the first The predicted probability of each sample; For sample weights, the weights for difficult samples are higher. For the first The predicted probability of the true label of each sample; This is the weight decay coefficient; This is the marginal loss coefficient; The positive and negative sample interval; , are the predicted probabilities of the positive and negative categories corresponding to the i-th sample, respectively.

10. The method for early diagnosis of transformer mechanical faults based on voiceprint according to claim 1, characterized in that: In S5, the early warning probability The calculation formula is: in, This represents the current matching degree. This represents the current signal acquisition duration. This is the time decay factor; This represents the historical number of samples. For the first Secondary matching degree; The trend weight reflects the cumulative effect of faults.

Images