A multi-modal voiceprint fault diagnosis method for a converter transformer

Abstract

The present application relates to the technical field of voiceprint fault diagnosis, in particular to a multi-modal voiceprint fault diagnosis method for converter transformers. It comprises a multi-modal signal acquisition module, a multi-modal signal preprocessing module, a feature weighting fusion module, a deep feature extraction module and a fault recognition and classification output module. The mechanical vibration signal and the voiceprint characteristic signal are synchronously acquired by vibration sensors and voiceprint sensors, and the multi-modal feature vectors are constructed after preprocessing. The dynamic weighted fusion of vibration and voiceprint features is realized by using the channel attention mechanism, the channel weight is generated by global average pooling and nonlinear mapping, and finally the voiceprint embedding vector with time sequence distribution characteristics is extracted by attention statistical pooling, and the accurate recognition of fault types is realized by using the Softmax classifier. The present application effectively improves the detection ability of early weak faults of converter transformers through the collaborative optimization of physical signal coupling, algorithm feature fusion and deep representation learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voiceprint fault diagnosis, in particular to a multi-modal voiceprint fault diagnosis method for a converter transformer. BACKGROUND

[0002] In a high-voltage direct-current power transmission system, the mechanical state of the converter transformer as a core device is directly related to the safe and stable operation of the power grid. Traditional monitoring methods mainly rely on offline methods such as oil chromatographic analysis and infrared temperature measurement, which are difficult to achieve real-time early warning of faults. In recent years, voiceprint diagnosis technology has been introduced into this field due to its non-contact and full-cycle monitoring advantages. However, a single sound mode is easily disturbed by background noise on site, and cannot capture the transient characteristics of internal mechanical vibrations. Therefore, existing technologies have begun to use multi-modal fusion methods to improve the complementarity and completeness of fault features by synchronously collecting vibration and voiceprint signals.

[0003] Although the multi-modal fusion method has achieved multi-source signal collaborative monitoring to some extent, it still has obvious deficiencies in the early and accurate diagnosis of mechanical faults of high-voltage direct-current converter transformers, which are manifested as follows: the traditional method has low sensitivity in capturing weak fault features (such as low-frequency energy fluctuations in the early stage of core loosening), relies on manual feature extraction, and is prone to high false alarm rates due to interference from working condition fluctuations. This leads to difficulties in early detection of mechanical faults of converter transformers in high-voltage direct-current power transmission systems, low diagnosis efficiency, and high false alarm rates. SUMMARY

[0004] The present application provides a multi-modal voiceprint fault diagnosis method for a converter transformer, which fuses multi-modal features of vibration signals and voiceprint signals, uses a SE attention mechanism for channel weighting, and combines multi-scale time series modeling and attention statistical pooling processing to solve the problems raised in the background art, i.e.:

[0005] The mechanical faults of the converter transformer in the high-voltage direct-current power transmission system are difficult to detect early, have low diagnosis efficiency, and have high false alarm rates.

[0006] To achieve the above purpose, the multi-modal voiceprint fault diagnosis method includes the following steps:

[0007] S1, collecting vibration signals and voiceprint signals of the converter transformer through a sensor system and preprocessing them to generate a multi-modal feature vector combination;

[0008] S2, concatenating the multi-modal feature vector combination according to the channel dimension to form a joint feature vector;

[0009] S3, performing global average pooling on the joint feature vector to obtain a channel statistical vector;

[0010] S4, dimensionally reduce the channel statistics vector, and generate channel weights through a ReLU activation function and nonlinear mapping;

[0011] S5, multiply the channel weights by the joint feature vector to obtain weighted fusion features, and generate a multi-modal feature sequence after weighted fusion;

[0012] S6, process the multi-modal feature sequence through multi-scale time series modeling and attention statistical pooling to generate a normalized voiceprint embedding vector;

[0013] S7, perform Softmax classification processing on the normalized voiceprint embedding vector to output the fault diagnosis result of the converter transformer.

[0014] In the above technical solution, the cooperative design of the channel dimension splicing of the multi-modal feature vector combination and the SE attention mechanism is a key improvement aimed at the interference between modalities caused by traditional simple feature stacking. If only a single signal modality or simple feature splicing is used, it is difficult to fully exploit the complementary characteristics between vibration and voiceprint signals, and weak but critical fault features are easily lost. By introducing the channel attention mechanism, the system can dynamically adjust the weights of different modal features to avoid important information being overwhelmed by noise. Multi-scale time series modeling breaks through the bottleneck of fixed-scale feature extraction, enabling the model to capture both transient impact and slowly varying fault features. The specially designed attention statistical pooling mechanism not only preserves the concentration trend of the time series features but also takes into account their distribution characteristics, overcoming the defect of traditional pooling methods that are insensitive to intermittent faults.

[0015] On this basis, the deep feature extraction module uses SE-Res2Block modules for multi-scale feature extraction.

[0016] In another technical solution, the deep feature extraction module includes multiple SE-Res2Block modules, each module achieving multi-scale feature extraction through channel division and convolution kernels with different dilation rates.

[0017] This technical solution constructs a hierarchical, multi-scale feature extraction system through the cascaded design of SE-Res2Block modules. Traditional single-kernel convolutional structures have significant limitations when dealing with the complex fault characteristics of converter transformers: the fixed-scale receptive field struggles to simultaneously capture large-scale mechanical deformations such as core loosening and subtle anomalies like partial discharge. By implementing a channel partitioning strategy within each module and employing differentiated dilation rate convolutional kernels, the system achieves parallel extraction and fusion of time-frequency domain features. This design allows shallow networks to focus on local high-frequency vibration features, while deep networks focus on long-range acoustic signature evolution, forming a complete fault characterization capability from micro to macro. Furthermore, the residual connections between modules not only alleviate the gradient vanishing problem but also enable dynamic aggregation of cross-scale features, ensuring that critical fault information is not diluted during deep transmission.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] The core of this invention lies in constructing an adaptive multimodal fault characterization system, whose innovation is reflected in three progressive dimensions: First, through spatiotemporal coupling analysis of vibration and acoustic signature signals, a cross-domain correlation between mechanical state and acoustic signature features is achieved at the physical level, breaking through the path dependence of traditional methods on single signal modes; second, based on an attention-based feature fusion architecture, a dynamic feature selection mechanism is established at the algorithm level, enabling the model to autonomously adjust the contribution weights between modes according to the fault type; finally, through the synergistic effect of multi-scale temporal modeling and statistical pooling, an embedded representation that takes into account both local details and global distribution is formed at the feature expression level, solving the trade-off problem of time-frequency resolution in traditional methods. This enables the system not only to identify typical fault modes but also to capture early composite fault precursor features that are difficult to detect using traditional methods. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0021] Figure 2 This is a schematic diagram of the structure of the present invention;

[0022] Figure 3 This is a schematic diagram of the structure of the present invention.

[0023] The meanings of the labels in the diagram are as follows:

[0024] 100. Multimodal signal acquisition module; 200. Multimodal signal preprocessing module; 300. Feature weighted fusion module; 400. Deep feature extraction module; 500. Fault identification and classification output module. Detailed Implementation

[0025] 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.

[0026] Currently, mechanical faults in converter transformers operating in high-voltage direct current transmission systems are difficult to detect early, resulting in low diagnostic efficiency and high false alarm rates. This invention provides a multi-mode acoustic signature fault diagnosis method for converter transformers, such as... Figure 1 As shown, it includes the following steps:

[0027] S1. The vibration signal and acoustic signature signal of the converter transformer are collected by the sensor system and preprocessed to generate a combination of multimodal feature vectors;

[0028] S2. Combine the multimodal feature vectors and concatenate them along the channel dimension to form a joint feature vector;

[0029] S3. Perform global average pooling on the joint feature vector to obtain the channel statistics vector;

[0030] S4. Dimensionality reduction is performed on the channel statistical vectors, and then channel weights are generated through the ReLU activation function and nonlinear mapping.

[0031] S5. Multiply the channel weights by the joint feature vector to obtain the weighted fusion features, and generate the weighted fusion multimodal feature sequence;

[0032] S6. The multimodal feature sequences are processed by multi-scale temporal modeling and attention statistical pooling to generate normalized voiceprint embedding vectors;

[0033] S7. Perform Softmax classification on the normalized voiceprint embedding vector and output the fault diagnosis results of the converter transformer.

[0034] See Figure 2 As shown, through the coordinated work of the multimodal signal acquisition module 100, the multimodal signal preprocessing module 200, the feature weighted fusion module 300, the deep feature extraction module 400, and the fault identification and classification output module 500, the accuracy of fault diagnosis and early warning capability can be effectively improved, thereby providing intelligent monitoring means for the safe and stable operation of the high voltage DC transmission system.

[0035] The multi-modal signal acquisition module 100 monitors the mechanical vibration and acoustic signal of the operating converter transformer in real time by constructing a complete sensor system. The system uses vibration sensors and acoustic sensors (microphone array) to work together to obtain characteristic information reflecting the internal structural health state of the equipment from different dimensions. The vibration sensors are precisely installed on the oil tank wall, winding support and other key structural parts of the converter transformer to capture the acceleration response signal a(t) generated during the operation of the equipment. The selection of these key installation positions ensures that the sensors effectively obtain the vibration characteristics reflecting the internal mechanical state of the transformer.

[0036] For the collected acceleration response signal a(t), the system uses fast Fourier transform (FFT) for frequency domain feature extraction. The mathematical expression is:

[0037]

[0038] In the formula, A(f) represents the complex representation of the acceleration signal a(t) in the frequency domain, i.e. the spectrum function;

[0039] is the Fourier transform operator, which represents the mathematical operation of converting the time domain signal to the frequency domain;

[0040] e -j2πft is the kernel function in the form of complex exponential, which constitutes the basis function of the Fourier transform;

[0041] j is the imaginary unit, satisfying j 2 =-1;

[0042] f is the frequency variable, with the unit of Hz;

[0043] t is the time variable, with the unit of seconds.

[0044] Through the transformation, the frequency spectrum characteristics of the acceleration signal can be obtained. Among them, the main frequency f main as an important characteristic parameter, corresponds to the maximum amplitude point in the frequency spectrum, and its calculation formula is:

[0045]

[0046] In the formula, argmax is the parameter operator for taking the maximum value, which represents the independent variable value when the target function takes the maximum value;

[0047] f is the frequency variable;

[0048] |A(f)| is the modulus of the complex spectrum A(f), reflecting the energy intensity of the signal at frequency f.

[0049] The main frequency feature can effectively reflect the mechanical state change of the transformer during operation. Meanwhile, the system is also configured with voiceprint sensors (microphone array), which are arranged in the key areas of the sound wave coupling path such as the core near area and the box shell. After processing, the voiceprint signal s(t) collected by the voiceprint sensor can construct a high-dimensional voiceprint feature. In order to effectively represent the energy change feature of the voiceprint signal, the system adopts a short-time energy calculation method:

[0050]

[0051] In the formula, E(t) represents the short-time energy (dimensionless) at time t, reflecting the local energy intensity of the signal;

[0052] s(n) represents a discrete-time acoustic signal, and n is the sample point number;

[0053] N represents the length of the sliding window, which determines the time domain resolution.

[0054] This voiceprint energy change feature can sensitively reflect the changes of mechanical excitation source and structure coupling characteristics, and plays an important role in detecting the initial signs of transformer micro-faults.

[0055] Finally, the multi-modal data collected by the system includes complete information such as the acceleration response a(t), the main frequency f main , the voiceprint signal s(t) and the acoustic energy change E(t) of each detection point. After high-precision analog-to-digital conversion (ADC) processing, these analog signals are uniformly converted into digital signals, providing high-quality data input for subsequent processing modules. Through this high-fidelity acquisition method of multiple sources of information, the system lays a reliable data foundation for subsequent feature analysis and fault diagnosis.

[0056] The multi-modal signal preprocessing module 200 systematically preprocesses the collected structure vibration signal a(t) and voiceprint signal s(t) and their characteristic quantities (main frequency f main , energy E(t)) in order to remove noise interference, enhance time sequence alignment consistency, extract robust features and improve model generalization ability. The preprocessing process includes five core processing links. First, all input signals are normalized to eliminate sensor sensitivity differences. The maximum-minimum normalization method is used to unify the signal amplitude to the range of [0, 1], and the specific calculation formula is:

[0057]

[0058] In the formula, a(t) represents the original vibration signal amplitude at time t;

[0059] s(t) represents the original voiceprint signal amplitude at time t;

[0060] max(a)-min(a) represents the dynamic range of the vibration signal;

[0061] max(s)-min(s) represents the dynamic range of the voiceprint signal.

[0062] The processing significantly improves the convergence stability of subsequent model training while preserving the original waveform characteristics.

[0063] To address the different synchronization problems between the vibration signal and the voiceprint signal, the system uses the cross-correlation function for accurate time alignment processing. By calculating the cross-correlation function of the normalized signal, the optimal time delay is estimated. After determining the optimal time delay, the signal time scale is adjusted based on the new time reference. This processing method effectively solves the timing deviation problem in the multi-source signal collection process and provides an accurate time alignment basis for subsequent feature fusion.

[0064] For high-dimensional time series data, the system uses principal component analysis (PCA) for dimensionality reduction to improve computational efficiency. Let the multi-modal feature combination be a matrix X ∈ R n×d , where n is the number of frames, d is the original feature dimension, and feature compression is achieved through a projection matrix:

[0065] Z = X·W, W ∈ R d×k , k << d.

[0066] In the formula, X represents the original feature matrix;

[0067] W represents the projection transformation matrix;

[0068] Z represents the feature matrix after dimensionality reduction.

[0069] This transformation preserves the main feature components of the signal while significantly reducing the data dimension. In the denoising processing section, a moving average filter is used to smooth the signal. This method can effectively suppress short impulse noise interference while preserving the main trend characteristics of the signal. To enhance the generalization ability of the model, the system also implements a data augmentation strategy. In the time domain, random jitter and amplitude disturbance are used:

[0070]

[0071] In the formula, s aug (t) represents the new sample generated by augmentation;

[0072] α represents the amplitude scaling factor;

[0073] represents the original voiceprint signal after preprocessing;

[0074] where a ~ U(0.9, 1.1) is a random scaling factor, and At ~ U(-5, +5) is a random time shift window (unit: sample points). The noisy perturbation and SpecAugment algorithm are introduced at the frequency domain level:

[0075]

[0076] wherein, represents the frequency domain signal estimation value before enhancement;

[0077] represents the added Gaussian white noise, 0 is the mean value, and represents the zero-centered noise, σ 2 is the variance, and represents the control of noise intensity.

[0078] Through this series of preprocessing steps, the multi-modal feature vector combination with consistent structure, time synchronization, noise suppression and diversity enhancement is finally obtained, and it is passed to the feature weighted fusion module 300, which lays a solid foundation for subsequent feature weighted fusion and deep modeling.

[0079] As Figure 3 shown, after completing the synchronization preprocessing and dimensionality reduction enhancement of the voiceprint signal and the structural vibration signal, the feature weighted fusion module 300 obtains the multi-modal feature vector combination under unified time sequence. Let the processed vibration feature be represented as and the voiceprint feature be represented as wherein dv and ds represent the respective feature dimensions. This module fuses and weights the key modal information through feature concatenation and channel attention mechanism (Squeeze-and-Excitation, abbreviated as SE), and improves the expression ability of fault features in the deep network.

[0080] First, the multi-modal feature vector combination is concatenated according to the channel dimension to form a joint feature vector, which is shown in the formula as follows:

[0081] F concat (t)=F v (t)\mathbin\VertF s (t)∈R d ,d=d v +d s ;

[0082] wherein F concat (t) represents the multi-modal feature vector after concatenation at time t;

[0083] F v (t) represents the vibration modal feature vector at time t;

[0084] \mathbin is a typesetting control command in LaTeX for defining binary operators;

[0085] \Vert represents the vector concatenation operation;

[0086] F s (t) represents the voiceprint modal feature vector at time t;

[0087] R d represents a d-dimensional real space (the space to which the concatenated vector belongs);

[0088] d = d v +d s represents the concatenated feature dimension calculation formula, d v is the original dimension of the vibration feature, d s is the original dimension of the voiceprint feature, and d is the total dimension of the joint feature.

[0089] The joint feature will serve as the input of the SE module. The SE module gives each dimension of the channel feature a dynamically learnable weight by explicitly modeling the inter-channel dependency, thereby enhancing the representation ability of key features. Its processing process mainly includes three steps: global average pooling (Squeeze), full connection weight generation (Excitation), and channel re-labeling (Scale).

[0090] In the Squeeze stage, the concatenated feature is subjected to global average pooling in the time dimension to obtain a channel statistical vector:

[0091]

[0092]

[0093] In the formula, z c represents the global statistics of the cth channel;

[0094] represents the reciprocal of the time step T;

[0095] F concat,c (t) represents the concatenated multi-modal feature vector and the feature value of the cth channel at time step t

[0096] d represents the total number of feature dimensions.

[0097] Finally, the channel statistical vector z is formed. This step obtains the global feature representation of each channel by compressing the time dimension information.

[0098] In the Excitation stage, the channel importance is modeled by a two-layer fully connected network. First, the channel statistic vector z is reduced in dimension, and then the channel weights are generated through a ReLU activation function and a nonlinear mapping:

[0099] s = σ(W2·δ(W1·z)), s ∈ R d ;

[0100] where s represents the channel attention weight vector output by the SE module, with dimension d;

[0101] σ is a Sigmoid activation function that compresses the weights to the interval (0, 1);

[0102] δ() is a ReLU activation function that introduces a nonlinear transformation

[0103] and are learnable parameters;

[0104] z is the input vector;

[0105] R d denotes the d-dimensional real space.

[0106] This step learns the importance of each channel adaptively, highlighting the feature dimensions that contribute more to fault diagnosis.

[0107] In the Scale stage, the generated channel weights s are applied to each frame of the original concatenated features, i.e., the channel weights are multiplied by the joint features to obtain the weighted fusion features, realizing channel weighting. This operation can significantly enhance the vibration and the channel dimensions related to the fault in the voiceprint joint features, while suppressing the dimensions with more redundancy or noise. The multi-modal feature sequence after weighted fusion is finally output, which will be used as the input of the subsequent deep model for time series modeling and embedded feature extraction.

[0108] Through the feature fusion of the SE attention mechanism, the model can adaptively focus on the key features in the vibration and voiceprint signals that are sensitive to faults, thereby improving the accuracy and robustness of diagnosis. For example, in the iron core looseness fault, the high-frequency components of the vibration signal and the specific harmonic components of the voiceprint signal may be given higher weights; while in the winding looseness case, low-frequency vibration features and voiceprint energy changes may be more discriminative. This dynamic weighting mechanism enables the model to automatically adjust the feature importance for different fault types, thereby improving the effect of multi-modal fusion.

[0109] The multi-modal feature dynamic weighting fusion realized by the SE attention mechanism provides more discriminative input features for subsequent deep feature extraction. However, relying solely on shallow feature fusion still cannot fully exploit the deep temporal patterns and cross-modal correlation characteristics of the converter transformer fault. To this end, we introduce a deep feature extraction module 400, which uses an improved ECAPA-TDNN architecture to perform more in-depth representation learning on the weighted fused features through multi-scale temporal modeling and attention statistical pooling. This architecture, designed specifically for the non-stationary nature of converter transformer acoustic vibration signals and multi-scale fault features, uses a progressive expansion convolution and channel attention mechanism to extract deep features with strong discriminability from a joint time-frequency domain perspective, laying the foundation for the final fault recognition and classification.

[0110] The system uses an improved ECAPA-TDNN architecture for deep feature extraction. This model effectively captures fault-sensitive features in converter transformer acoustic vibration signals through multi-scale temporal modeling and hierarchical feature learning. The core of the model is composed of multiple SE-Res2Block modules, each of which uses a unique channel division and dilation convolution strategy to extract features of different time scales in parallel. Specifically, the multi-modal feature sequence is first divided into several channel groups, each of which is processed using convolution kernels with different dilation rates. This design allows the model to capture both local detail features and long-range dependencies. The convolution results of each group are concatenated with the original input to form a residual connection, and then the SE module is used to re-label the channels to ensure that key feature channels are enhanced.

[0111] Between the multiple SE-Res2Block layers, the model introduces a learnable skip connection mechanism to dynamically aggregate deep and shallow features. This design not only alleviates the gradient vanishing problem of deep networks, but also allows the model to adaptively fuse feature representations at different abstraction levels. In particular, for the converter transformer fault diagnosis scenario, this mechanism can effectively coordinate the fusion ratio of high-frequency vibration details and low-frequency acoustic envelope features, improving the representation ability for typical faults such as core looseness and winding deformation.

[0112] After modeling the temporal features, the system uses an attention statistical pooling mechanism to extract global feature representations. This mechanism calculates the attention weights of each frame feature through two fully connected networks, and calculates the weighted mean and standard deviation based on these weights. This statistical pooling method can highlight the information of the fault-sensitive period while suppressing the interference of non-key frames. The final feature vector not only contains the central tendency information of the temporal features, but also retains their distribution characteristics, providing a more comprehensive basis for fault classification.

[0113] To ensure the stability of feature representation, the system normalizes the statistical features and adjusts the feature dimension through a learnable projection matrix. The final output of the normalized voiceprint embedding vector integrates the timing dynamic characteristics, channel attention weight, multi-modal fusion information and multi-scale features, etc. Various discriminative elements form a voiceprint representation with high discriminability. The deep feature extraction module 400 can effectively improve the recognition accuracy of typical mechanical faults while maintaining computational efficiency, especially in complex fault scenarios such as winding looseness.

[0114] The fault recognition and classification output module 500 is based on the normalized voiceprint embedding vector obtained by the deep feature extraction module 400, and realizes intelligent diagnosis of the operating state of the converter transformer through a carefully designed classifier architecture. This module uses a structure combining fully connected layers and a Softmax classifier, where the classifier weight matrix and bias terms jointly construct the mapping relationship from the feature space to the fault category. For each fault class, the system calculates a discriminative score, and these scores combine to form a comprehensive score vector, providing a basis for subsequent probabilistic decision-making.

[0115] In the probability conversion stage, the system converts the original scores into an intuitive probability distribution through the Softmax function:

[0116]

[0117] where P(y = j | e norm ) represents the probability that the sample belongs to the jth fault class given the normalized feature e norm ;

[0118] exp(z j ) represents the exponential operation on the jth original score z j , ensuring that the output is positive;

[0119] represents the exponential sum of all class scores, used for probability normalization;

[0120] z j represents the linear discriminant score calculated by the model for the jth fault class;

[0121] e norm represents the L2 normalized deep feature vector;

[0122] C represents the total number of fault categories (such as core looseness, winding looseness, etc.).

[0123] The probability distribution accurately reflects the matching degree of the input sample and each type of fault. Based on the maximum a posteriori probability criterion, the system automatically selects the class with the highest probability as the final diagnosis result y^ = argmaxjP(y = j | enorm). The identifiable fault types include but are not limited to typical faults such as core loosening, winding loosening, mechanical resonance, etc., which constitute a complete fault set Y, which can be flexibly expanded according to actual engineering needs.

[0124] In terms of model optimization, the system uses a cross-entropy loss function for end-to-end training:

[0125]

[0126] wherein, represents the cross-entropy loss value, which measures the difference between the predicted probability distribution and the true distribution;

[0127] - is a negative sign, used to convert maximizing the log probability into minimizing the loss;

[0128] represents the exponential function δ y,j The cumulative operation of all categories (j = 1 to C);

[0129] log P(y = j | e norm ) represents the conditional probability predicted by the model that the sample belongs to category j;

[0130] e norm represents the normalized feature vector of the input;

[0131] δ y,j 1 when the sample true label y = j, otherwise 0. By minimizing this loss function, the model can continuously improve the accuracy of fault identification. In actual deployment, the system uses a multi-dimensional evaluation system, in which the overall accuracy is calculated as:

[0132]

[0133] wherein, Accuracy represents the overall prediction accuracy of the model, taking a value of 0-1;

[0134] TP represents the true positive, i.e. the number of samples correctly identified as faults;

[0135] TN represents the true negative, i.e. the number of samples correctly identified as normal;

[0136] FP represents the false positive, i.e. the number of normal samples misjudged as faults;

[0137] FN represents the false negative, i.e. the number of missed fault samples.

[0138] The index comprehensively considers the overall performance of the system in all categories. The module effectively promotes the average identification accuracy of typical faults while maintaining high inference speed (single processing time <20 ms). The precision and recall rates are also maintained at industry-leading levels, fully verifying its reliability and practicality in engineering practice.

[0139] Through the above classifier structure and evaluation mechanism, the application can effectively identify typical faults in the operation process of the converter transformer, has high identification accuracy and engineering application value, and is particularly suitable for early warning and state monitoring tasks in the field environment.

[0140] The above shows and describes the basic principles, main features and advantages of the application. Those skilled in the art should understand that the application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the application and are not intended to limit the application. Without departing from the spirit and scope of the application, various changes and improvements can be made to the application, and these changes and improvements all fall within the scope of the claimed application. The scope of protection of the application is defined by the appended claims and their equivalents.

Claims

1. A multi-mode acoustic signature fault diagnosis method for converter transformers, characterized in that, Includes the following steps: S1. The vibration signal and acoustic signature signal of the converter transformer are collected by the sensor system and preprocessed to generate a combination of multimodal feature vectors; S2. Combine the multimodal feature vectors and concatenate them along the channel dimension to form a joint feature vector; S3. Perform global average pooling on the joint feature vector to obtain the channel statistics vector; S4. Dimensionality reduction is performed on the channel statistical vectors, and then channel weights are generated through the ReLU activation function and nonlinear mapping. S5. Multiply the channel weights by the joint feature vector to obtain the weighted fusion features, and generate the weighted fusion multimodal feature sequence; S6. The multimodal feature sequences are processed through multi-scale temporal modeling and attention statistical pooling to generate normalized voiceprint embedding vectors. S7. Perform Softmax classification on the normalized voiceprint embedding vector and output the fault diagnosis results of the converter transformer. The joint feature vectors are fused using the SE attention mechanism through the feature weighted fusion module (300) with channel weighting. Multi-scale feature extraction is performed using the SE-Res2Block module through the deep feature extraction module (400); The deep feature extraction module (400) calculates the weighted mean and standard deviation through attention statistical pooling to highlight information during fault-sensitive periods.

2. The multi-mode acoustic fault diagnosis method for converter transformers according to claim 1, characterized in that: The sensor system includes vibration sensors installed on the tank wall and winding support, as well as acoustic sensors arranged in the near-core area and the outer shell area of ​​the housing.

3. The multi-mode acoustic fault diagnosis method for converter transformers according to claim 1, characterized in that: The vibration signal and acoustic signature signal are normalized, time-aligned and dimension-reduced by the multimodal signal preprocessing module (200) to generate the multimodal feature vector combination.

4. The multi-mode acoustic fault diagnosis method for converter transformers according to claim 1, characterized in that: The feature weighted fusion module (300) learns the nonlinear relationship between channels through a two-layer fully connected network on the channel statistical vectors.

5. The multi-mode acoustic fault diagnosis method for converter transformers according to claim 1, characterized in that: The deep feature extraction module (400) includes multiple SE-Res2Block modules, each of which achieves multi-scale feature extraction through channel partitioning and convolutional kernels with different dilation rates.

6. The multi-mode acoustic fault diagnosis method for converter transformers according to claim 1, characterized in that: The classification results are optimized and trained using the cross-entropy loss function through the fault identification and classification output module (500).

7. The multi-mode acoustic fault diagnosis method for converter transformers according to claim 6, characterized in that: The fault identification and classification output module (500) can identify fault types including core loosening, winding loosening, and mechanical resonance.

Images