Composite fault diagnosis method based on transformer voiceprint

Abstract

The invention discloses a composite fault diagnosis method based on transformer voiceprint, and relates to the technical field of transformer fault diagnosis, and the method comprises the steps: obtaining a transformer voiceprint signal, carrying out the preprocessing of the transformer voiceprint signal, and obtaining a two-dimensional time-frequency fusion feature of the transformer voiceprint signal; constructing a decoupling generative adversarial network, and performing decoupling separation on irrelevant interference factors in the two-dimensional time-frequency fusion features to obtain decoupling fault features; based on the improved conditional diffusion model, generating an unknown compound fault voiceprint sample; constructing a diversified training set based on the real single fault voiceprint sample, the decoupling fault feature and the unknown composite fault voiceprint sample; and performing fusion training by using a multi-label classifier in combination with the diversified training set to obtain a trained fault classifier. The fault classifier obtained by the method can diagnose known single and composite faults with high precision, and has the capability of identifying unknown composite faults, so that the safety monitoring of the transformer is met.

Description

Technical Field

[0001] This invention generally relates to the field of transformer fault diagnosis technology, and specifically to a composite fault diagnosis method based on transformer acoustic signatures. Background Technology

[0002] Voiceprint diagnostics, as a non-contact online monitoring technology, has significant application prospects in transformer mechanical fault identification. Current mainstream methods typically include the following steps: acquiring sound signals through a voiceprint sensor, converting the signals into a two-dimensional time-frequency image using time-frequency transformation (such as short-time Fourier transform or wavelet transform), and then using a deep learning model (such as a convolutional neural network) for classification and identification.

[0003] However, existing technical solutions suffer from several inherent and interconnected flaws that limit their effectiveness in complex real-world scenarios. For example, the time-frequency features extracted directly from the original acoustic signature signal by existing methods are mixed with a large amount of interference information unrelated to the fault, such as load fluctuations and environmental noise, resulting in low feature purity and poor recognition. Secondly, current fault diagnosis models heavily rely on labeled real fault samples for training, making it impossible for the models to comprehensively and accurately diagnose faults under complex transformer operating conditions. This means that the fault diagnosis models lack the ability to warn and diagnose unknown and novel complex faults, failing to meet the core requirements of safe production. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, it is desirable to provide a composite fault diagnosis method based on transformer acoustic signature.

[0005] This invention provides a composite fault diagnosis method based on transformer acoustic signatures, comprising: The transformer acoustic signature signal is acquired and preprocessed to obtain the two-dimensional time-frequency fusion features of the transformer acoustic signature signal. A decoupled generative adversarial network is constructed, and the network is used to decouple and separate irrelevant interference factors in the two-dimensional time-frequency fusion features to obtain decoupled fault features. Based on the improved conditional diffusion model, unknown composite fault acoustic signature samples are generated. Acquire real single-fault acoustic fingerprint samples of transformers, and construct a diversified training set based on the real single-fault acoustic fingerprint samples, the decoupled fault features, and the unknown composite fault acoustic fingerprint samples. A multi-label classifier is used in conjunction with the diversified training set for fusion training to obtain a trained fault classifier, which can be used for comprehensive fault diagnosis of the transformer.

[0006] According to the technical solution provided by the present invention, the acquisition of transformer acoustic signature signal includes: Multiple acoustic sensors are deployed in the key areas of the transformer, which include at least the transformer housing, windings, and core. The transformer acoustic signature signal is collected during the transformer's operation according to the preset acquisition frequency.

[0007] According to the technical solution provided by the present invention, the preprocessing of the transformer acoustic signature signal includes: Discrete wavelet transform is used to decompose the acoustic signature signal of the transformer, and detail coefficients and approximation coefficients are extracted. Based on the detail coefficients and approximation coefficients, a composite fault attribute set containing single faults and composite faults is constructed, and the composite fault attribute set is used to train the multi-label classifier. The transformer acoustic signature signal after discrete wavelet transform is converted into a two-dimensional time-frequency fusion map using continuous wavelet transform, and the two-dimensional time-frequency fusion map is preprocessed and optimized to obtain two-dimensional time-frequency fusion features.

[0008] According to the technical solution provided by the present invention, the decoupled generative adversarial network includes: a generator, a discriminator, and an auxiliary classifier; The generator is used to encode and decode the two-dimensional time-frequency fusion features; The generator includes an encoder and a decoder; the encoder encodes the two-dimensional time-frequency fusion features into a composite feature tensor containing fault features and irrelevant factor features; the decoder reconstructs an extended two-dimensional time-frequency fusion graph based on the composite feature tensor. The discriminator is equipped with multiple classification heads, which are used to simultaneously distinguish the authenticity of the fault and irrelevant interference factors in the extended two-dimensional time-frequency fusion map; The auxiliary classifier is used to classify and verify the fault features obtained after decoupling the extended two-dimensional time-frequency fusion graph.

[0009] According to the technical solution provided by the present invention, the decoupled generative adversarial network is used to decouple and separate irrelevant interference factors in the two-dimensional time-frequency fusion features to obtain decoupled fault features, including: At least two two-dimensional time-frequency fusion maps with different fault types are input into the encoder of the generator to encode the corresponding composite feature tensors; each composite feature tensor contains a fault feature tensor and an irrelevant interference factor tensor. The fault feature tensors in at least two of the composite feature tensors are swapped and recombined with their respective irrelevant interference factor tensors to form a variety of new composite feature tensors; The new composite feature tensor is input into the decoder of the generator for reconstruction to obtain the extended two-dimensional time-frequency fusion map. The extended two-dimensional time-frequency fusion graph is input into the decoupled generative adversarial network, and it is jointly trained with a multi-objective loss function. After the joint training is completed, the fault feature tensor is extracted as the decoupled fault feature.

[0010] According to the technical solution provided by the present invention, based on an improved conditional diffusion model, unknown composite fault acoustic signature samples are generated, including: Based on the real single-fault acoustic print samples, and using the composite fault attribute set as a constraint, Gaussian noise is gradually added to the real single-fault acoustic print samples to generate diffusion data with different noise levels. The improved conditional diffusion model is used to predict and gradually remove noise from the diffusion data. After iterative reconstruction, the unknown composite fault acoustic signature sample is obtained.

[0011] According to the technical solution provided by the present invention, a diversified training set is constructed based on the real single-fault voiceprint samples, the decoupled fault features, and the unknown composite fault voiceprint samples, including: The real single-fault voiceprint samples, the decoupled fault features, and the unknown composite fault voiceprint samples are combined according to a preset ratio to obtain the diversified training set.

[0012] According to the technical solution provided by the present invention, fusion training is performed using a multi-label classifier combined with the diversified training set, including: A multi-label classifier based on residual networks is constructed. The multi-label classifier uses binary encoding to label multiple fault labels of samples and includes network layers using a preset activation function. A fusion training strategy is introduced, which trains multiple specialized classifiers on specific composite faults and single faults; each of the specialized classifiers is configured to focus on distinguishing between a specific type of composite fault and single fault. The multi-label classifier is trained by combining a preset optimizer and a cross-entropy loss function to obtain a fault classifier capable of comprehensive fault diagnosis of the transformer.

[0013] According to the technical solution provided by the present invention, the trained fault classifier includes: Input the test sample set containing known faults and unknown faults into the fault classifier obtained through training, and obtain the fault type prediction result output by the fault classifier; The known fault accuracy, the unknown fault accuracy, and the harmonic mean score of the known fault accuracy and the unknown fault accuracy are used as evaluation indicators to evaluate the fault type prediction results and to validate the fault classifier.

[0014] In summary, this technical solution specifically discloses a composite fault diagnosis method based on transformer acoustic signatures. The method includes: acquiring transformer acoustic signature signals and preprocessing them to obtain two-dimensional time-frequency fusion features; constructing a decoupled generative adversarial network (GAN) to decouple and separate irrelevant interference factors in the two-dimensional time-frequency fusion features, obtaining decoupled fault features; generating acoustic signature samples of unknown composite faults based on an improved conditional diffusion model; acquiring real single-fault acoustic signature samples of the transformer and constructing a diversified training set based on the real single-fault acoustic signature samples, decoupled fault features, and unknown composite fault acoustic signature samples; and performing fusion training using a multi-label classifier combined with the diversified training set to obtain a trained fault classifier for comprehensive transformer fault diagnosis.

[0015] Beneficial Effects: This invention provides an intelligent diagnostic method for complex faults based on transformer acoustic signatures. By constructing a decoupled generative adversarial network and a unique feature exchange training mechanism, it separates the pure fault essence features from mixed features, effectively eliminating irrelevant interference such as load and noise. Addressing the challenge of scarce complex fault samples, the solution employs a conditional diffusion model, using fault attributes as constraints to generate unknown complex fault samples that conform to the true characteristics of transformers. Finally, a training set is constructed by fusing real samples, decoupled features, and generated samples. This set is then trained using a multi-label classifier incorporating a dedicated classifier strategy, enabling the final model to not only diagnose known single and complex faults with high accuracy but also to identify unknown complex faults. Attached Figure Description

[0016] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating a composite fault diagnosis method based on transformer acoustic signatures.

[0017] Figure 2 This is a flowchart illustrating step S100 in the method of the present invention.

[0018] Figure 3 This is a flowchart illustrating step S200 in the method of the present invention.

[0019] Figure 4 This is a flowchart illustrating step S300 in the method of the present invention.

[0020] Figure 5 This is a flowchart illustrating step S500 in the method of the present invention.

[0021] Figure 6 Comparison of intra-class / inter-class similarity before and after decoupling fault features.

[0022] Figure 7 Compare the time-frequency features of generated samples and real samples for conditional diffusion model.

[0023] Figure 8 A comparison of the accuracy of different diagnostic methods. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] Example 1 To make the technical solutions of the embodiments of the present invention clearer and easier to understand, the application background of the embodiments of the present invention will be introduced below.

[0027] As a core hub in power transmission and distribution systems, the long-term stable operation of transformers is crucial for ensuring the reliability of the power grid. Traditional fault monitoring technologies, such as dissolved gas analysis (DGA), partial discharge detection, and vibration monitoring, while widely used, all have inherent limitations: DGA has a large response delay and is a post-event diagnostic method; partial discharge detection is susceptible to electromagnetic interference and its location is complex; vibration monitoring requires the installation of contact sensors, which is inconvenient to implement and may damage the equipment itself. Therefore, in recent years, acoustic monitoring technology, due to its non-invasiveness, flexible installation, and rich information capabilities, has become a cutting-edge direction in transformer online condition assessment.

[0028] Current transformer fault diagnosis technologies based on acoustic monitoring primarily focus on vibration and partial discharge signals, with limited systematic research on acoustic signature signals. Furthermore, in feature extraction, existing techniques often employ single-dimensional analysis, extracting features only from the time or frequency domains of the acoustic signature signal, failing to achieve deep fusion of time and frequency domain information. For the two-dimensional processing of one-dimensional signals, fixed-rule reshaping or simple mapping methods are frequently used, neglecting the collaborative preservation of local details and global correlations of the acoustic signature signal's time-frequency characteristics. Regarding feature processing, existing methods lack effective separation of core fault features from irrelevant factors such as load fluctuations, environmental noise, and equipment operation interference, resulting in a large amount of redundant information in the extracted features. In terms of sample utilization, traditional diagnostic models mainly rely on supervised learning from real-world single-fault samples, failing to address the scarcity of composite fault samples and lacking mechanisms for generating and utilizing unknown composite fault samples. In the classification and diagnosis stage, single classification models are often used to determine fault types, without designing specialized fusion training strategies for the multi-label characteristics of composite faults, making it difficult to accurately distinguish between multiple fault types.

[0029] It is evident that how to decouple high-purity fault essence features from voiceprint signals and overcome the bottleneck of scarce composite fault samples, thereby achieving accurate diagnosis of known and unknown composite faults, has become a technical challenge that urgently needs to be solved in this field.

[0030] In view of this, the present invention proposes a composite fault diagnosis method based on transformer acoustic signatures. The method includes: acquiring transformer acoustic signature signals and preprocessing the transformer acoustic signature signals to obtain two-dimensional time-frequency fusion features of the transformer acoustic signature signals; constructing a decoupled generative adversarial network (GAN) to decouple and separate irrelevant interference factors in the two-dimensional time-frequency fusion features to obtain decoupled fault features; generating acoustic signature samples of unknown composite faults based on an improved conditional diffusion model; acquiring real single-fault acoustic signature samples of transformers and constructing a diversified training set based on real single-fault acoustic signature samples, decoupled fault features, and unknown composite fault acoustic signature samples; and performing fusion training using a multi-label classifier combined with the diversified training set to obtain a trained fault classifier for comprehensive fault diagnosis of transformers.

[0031] In summary, this invention introduces a decoupled generative adversarial network to specifically process the two-dimensional time-frequency fusion features of acoustic signature signals. This effectively separates interference factors (such as load fluctuations and environmental noise) that are unrelated to the essence of the fault, thereby obtaining high-purity and high-identifiability decoupled fault features. An improved conditional diffusion model is used to generate acoustic signature samples of unknown composite faults, effectively solving the problems of difficulty in obtaining composite fault samples and high annotation costs in actual operation and maintenance. A diversified training set is constructed by integrating real single-fault acoustic signature samples, decoupled pure fault features, and generated unknown composite fault acoustic signature samples to obtain a fault classifier for transformer fault diagnosis. This process achieves an organic combination of real data, enhanced features, and synthetic samples. In this way, the diversified training set allows the model to simultaneously learn real distributions, essential features, and unknown patterns, effectively improving the diagnostic system's generalization ability to recognize known and unknown composite faults. This enables effective diagnosis of unknown types of composite faults, thus achieving a more comprehensive and intelligent state assessment and early warning of transformer faults.

[0032] Please refer to the following. Figure 1 The flowchart shown in this embodiment illustrates a composite fault diagnosis method based on transformer acoustic signatures. The executing entity of this embodiment can be a transformer control system or a monitoring equipment terminal; no specific limitations are made here. Next, the steps of this invention will be further explained. The method includes the following steps: S100. Acquire the transformer acoustic signature signal and preprocess the transformer acoustic signature signal to obtain the two-dimensional time-frequency fusion characteristics of the transformer acoustic signature signal. Specifically, transformer acoustic signature refers to the sound signal generated when a transformer is in operation. This sound is produced by the vibration of its internal components, such as windings, core, and insulation, under electromagnetic forces and mechanical stresses, and propagates through the transformer oil and tank into the air. This sound signal contains comprehensive information about the mechanical and electrical status of the equipment and can be used to characterize the health of the transformer's operation.

[0033] In this embodiment of the invention, the process of acquiring transformer acoustic signature signals includes: deploying multiple acoustic signature sensors in the key area of ​​the transformer, and acquiring transformer acoustic signature signals under the transformer's operating state according to a preset acquisition frequency; wherein, the key area includes at least: the transformer housing, windings, and core. In practical applications, multiple high-precision acoustic signature sensor arrays can be deployed at key parts of the transformer (such as the surface of the transformer housing, the outer casing corresponding to the windings, and near the core) to synchronously collect the original acoustic signature signals corresponding to different operating conditions of the transformer at a preset acquisition frequency (e.g., a frequency in the range of 10kHz to 20kHz), forming a time-series acoustic signature signal. (in,B For sample batch size, S Number of sensor channels T (where the time series length is used), thus ensuring that the acoustic signature signal can fully cover the acoustic radiation characteristics corresponding to different fault types.

[0034] After obtaining different transformer acoustic signature signals, preprocessing is required to produce a two-dimensional feature map that preserves time-frequency details while removing irrelevant interference; this is known as a two-dimensional time-frequency fusion map of the transformer acoustic signature signal. For details, see [link to relevant documentation]. Figure 1 The preprocessing of transformer acoustic signature signals includes the following steps: S101. Discrete wavelet transform is used to decompose the transformer acoustic waveform signal and extract detail coefficients and approximation coefficients. Among them, discrete wavelet transform is to extract detail coefficients (reflecting high-frequency transient components, such as partial discharge pulses) and approximation coefficients (reflecting low-frequency trend components, such as the magnetostriction fundamental frequency of iron core) of different frequency bands by performing multi-level discrete wavelet decomposition on the acoustic signal.

[0035] These detail coefficients and approximation coefficients provide the foundation for the subsequent construction of a composite fault attribute set. In practical applications, this embodiment of the invention uses discrete wavelet transform to decompose the acquired acoustic signature signal into 5 scales. The detail coefficients (amplitude range 6±0.2) reflecting local changes in the high-frequency time domain are extracted by convolution operation of wavelet filter and single fault acoustic signature signal. The approximation coefficients (amplitude range 8±0.2) reflecting the trend in the low-frequency domain are extracted by convolution operation of low-frequency filter and single fault acoustic signature signal.

[0036] For example, for the acquired voiceprint signal The decomposition process of the discrete wavelet transform can be represented as follows: (1) The formula for calculating the approximation coefficient is as follows: ; In the formula, Denotes the approximation coefficients of the i-th layer. These are the coefficients of the low-pass filter.

[0037] (2) The formula for calculating detail coefficients is as follows: ; In the formula, Denotes the approximation coefficients of the i-th layer. These are the coefficients of the high-pass filter. j The decomposition level is 5.

[0038] S102. Based on detail coefficients and approximation coefficients, construct a composite fault attribute set containing single faults and composite faults. The composite fault attribute set is used to train a multi-label classifier. Based on the approximation coefficients and detail coefficients of each transformer acoustic signal obtained from the above calculations, and the statistical characteristics reflected by each transformer acoustic signal, multiple single fault attribute matrices representing a single fault mode can be constructed. Then, by combining different single fault attribute matrices, a composite fault attribute set containing single faults and composite faults can be obtained.

[0039] S103. The transformer acoustic waveform signal after discrete wavelet transform is converted into a two-dimensional time-frequency fusion map by continuous wavelet transform, and the two-dimensional time-frequency fusion map is preprocessed and optimized to obtain two-dimensional time-frequency fusion features.

[0040] For example, the transformer acoustic signature signal after discrete wavelet transform processing Select Morlet Wavelets are used as basis functions (which match the oscillation characteristics of transformer acoustic waveforms well in terms of time-frequency focusing) to obtain a fine time-frequency distribution. The specific formula is as follows:

[0041] In the formula, The center frequency is typically set to 6.

[0042] Furthermore, the transformer acoustic signature signal after discrete wavelet transform processing The continuous wavelet transform coefficients are Please refer to the following formula for details:

[0043] In the formula, This is the scale factor, corresponding to the reciprocal of the frequency. This is the translation factor, corresponding to time. Adjustment is made... Generate a set of time-frequency coefficient matrices.

[0044] Finally, the amplitudes of the continuous wavelet transform coefficients are converted into grayscale images, and the one-dimensional voiceprint time series signal (transformer voiceprint signal) is uniformly adjusted to a standard size of 96×96 pixels, thus forming a two-dimensional time-frequency fusion image with a unified format. (in, H , W These represent the height and width of the two-dimensional time-frequency fusion graph, respectively, enabling visualization and structured representation of time-frequency features.

[0045] Next, this embodiment of the invention also employs wavelet threshold denoising to remove irrelevant interference such as environmental noise and power grid electromagnetic interference from the two-dimensional time-frequency fusion graph, thereby preprocessing and optimizing the two-dimensional time-frequency fusion graph to finally obtain the standardized two-dimensional time-frequency fusion features. .

[0046] Specifically, the wavelet coefficients are set as threshold ;in, This is an estimate of the noise standard deviation. The coefficient length is the denoised coefficient. as follows:

[0047] For the denoised data The two-dimensional time-frequency fusion features are obtained by normalization using Z-score, and the specific formula is as follows:

[0048] In the formula, The average value of the voiceprint signal. This represents the standard deviation of the voiceprint signal.

[0049] S200. Construct a decoupled generative adversarial network. The decoupled generative adversarial network will be used to decouple and separate irrelevant interference factors in the two-dimensional time-frequency fusion features to obtain decoupled fault features. In this embodiment of the invention, the decoupled generative adversarial network includes: a generator, a discriminator, and an auxiliary classifier; The generator is used to encode and decode the two-dimensional time-frequency fusion features. The generator includes an encoder and a decoder. The encoder consists of convolutional blocks and residual blocks, and uses the ReLU activation function and instance normalization to encode the preprocessed and optimized two-dimensional time-frequency fusion map into a composite feature tensor containing fault features and irrelevant factor features. The decoder consists of deconvolutional blocks, and the output layer uses the Tanh activation function to reconstruct the extended two-dimensional time-frequency fusion map based on the composite feature tensor. The discriminator has multiple classification heads, actually consisting of a convolutional block and four classification heads, using the LeakyReLU activation function. The classification heads are used for adversarial classification, load classification, fault location size classification, and fault type classification, respectively, thereby achieving accurate discrimination between the authenticity of generated samples and irrelevant factors. Ultimately, it is used to simultaneously discriminate between the authenticity of faults and irrelevant interference factors in the extended two-dimensional time-frequency fusion map. The auxiliary classifier includes a convolutional block and a fully connected layer, using the LeakyReLU activation function, and is used to classify and verify the fault features obtained after decoupling the extended two-dimensional time-frequency fusion map.

[0050] Specifically, see Figure 3Based on the above description of the structure of decoupled generative adversarial networks, the process of "using decoupled generative adversarial networks to decouple and separate irrelevant interference factors in two-dimensional time-frequency fusion features to obtain decoupled fault features" includes the following steps: S201. Input at least two two-dimensional time-frequency fusion diagrams with different fault types into the encoder of the generator, and encode them respectively to obtain the corresponding composite feature tensors; each composite feature tensor contains a fault feature tensor and an irrelevant interference factor tensor. Two two-dimensional time-frequency fusion maps with different fault types are randomly selected and input into the encoder of the generator to obtain a tensor containing fault features. and irrelevant interference factors tensor Composite feature tensor (including interference information such as load and fault location dimensions) The details are as follows:

[0051] S202. Exchange the fault feature tensors in at least two composite feature tensors and recombine them with their respective irrelevant interference factor tensors to form a variety of new composite feature tensors. Below, we input two-dimensional time-frequency fusion diagrams of two different fault types. Taking the example of a two-dimensional time-frequency fusion diagram, we can illustrate this further. The structure is as follows:

[0052]

[0053] The structure of the new composite feature tensor obtained through the above method is shown below:

[0054] Therefore, for each composite feature tensor, we can reorganize it by exchanging the fault feature tensors in at least two composite feature tensors under any different fault types and retaining their respective irrelevant interference factor tensors to obtain a variety of new composite feature tensors.

[0055] S203. The decoder of the new composite feature tensor is input into the generator to reconstruct the extended two-dimensional time-frequency fusion map. S204. Input the extended two-dimensional time-frequency fusion graph into the decoupled generative adversarial network, and train it jointly with a multi-objective loss function. After the joint training is completed, extract the fault feature tensor as the decoupled fault feature.

[0056] Finally, the recombined composite feature tensor is input into the decoder to reconstruct a new two-dimensional time-frequency fusion graph. At the same time, a new two-dimensional time-frequency fusion map will be generated. By re-inputting the generator and repeating the encoding, swapping, and decoding processes to enhance decoupling, a multi-round reconstructed graph can ultimately be obtained. (here) The iteration number represents the extended two-dimensional time-frequency fusion graph. Throughout the training process, the discriminator continuously and synchronously distinguishes between real samples and the extended two-dimensional time-frequency fusion graphs generated by the generator in each round. The goal is to make it impossible to distinguish between genuine and fake generated samples, while accurately identifying the categories of irrelevant interference factors they carry. The generator, through this adversarial game, learns to generate decoupling features that are both realistic (conforming to the real data distribution) and possess the correct fault-interference combination. Specifically, the joint adversarial training in this embodiment of the invention is optimized through a multi-objective loss function that integrates reconstruction loss, adversarial loss, feature decoupling loss, and classification loss. The multi-objective loss function takes the following form:

[0057] In the formula, For multi-objective loss functions; To combat the losses; The classification loss for the task of classifying irrelevant factors on the discriminator; To perceive loss; Loss at the center; For example, in this invention, the weighting coefficients corresponding to the above four types of losses are set as follows:

[0058] In the formula, To counteract the loss weighting coefficient; These are the classification loss weight coefficients; For the perceived loss weighting coefficient; The center loss weight coefficient.

[0059] Finally, after training stabilizes, discard the irrelevant factor tensor and retain the fault feature tensor. As a characteristic of decoupling faults This enables the effective separation of core fault characteristics from irrelevant interference factors.

[0060] S300, Based on the improved conditional diffusion model, generate acoustic signature samples of unknown composite faults; This step leverages the powerful generative capabilities and fitting ability of the conditional diffusion model to achieve high-fidelity generation of unknown composite fault samples that are difficult to obtain in reality, guided by the semantics of the composite fault attribute set.

[0061] Specifically, see Figure 4 The above step S300 includes the following steps: S301. Based on real single-fault acoustic fingerprint samples, and using composite fault attribute sets as constraints, Gaussian noise is gradually added to the real single-fault acoustic fingerprint samples to generate diffusion data with different noise levels. S302. By predicting and gradually removing noise from the diffusion data using an improved conditional diffusion model, and then reconstructing iteratively, unknown composite fault acoustic samples are obtained.

[0062] For example, the above process will be described in detail below: (1) Forward process of conditional diffusion model: For a given real sample Using the composite fault attribute set constructed in the aforementioned steps as a condition, Gaussian noise is gradually added to the forward process within T=100 steps:

[0063] In the formula, This is the noise attenuation coefficient; Standard Gaussian noise; For the first Step-by-step diffusion of data.

[0064] By progressively adding Gaussian noise of 0.1-0.8 in linear increments to real single-fault acoustic signature samples, 100 sets of diffusion data at different noise levels were generated. .

[0065] (2) Forward process of conditional diffusion model: First, a denoising network is trained to predict noise, as follows:

[0066] In the formula, where For the condition of a composite fault attribute set, For model prediction networks.

[0067] Next, in this embodiment of the invention, after the noise is predicted, it is necessary to gradually remove the noise, and after 50 iterations, reconstruct and generate the unknown composite fault voiceprint sample. For each type of unknown composite fault voiceprint sample, 500 samples are generated to ensure sample diversity and authenticity.

[0068] Finally, after obtaining the acoustic signature samples of unknown composite faults, these generated samples need to be optimized to ensure sample quality. The optimization method adopted in this embodiment of the invention is as follows: calculate the time-frequency characteristic amplitude error between the generated acoustic signature samples of unknown composite faults and the real single-fault samples, and iteratively adjust the diffusion step size and noise attenuation coefficient to control the average amplitude error between the generated samples and the real samples within 0.32, thereby obtaining acoustic signature samples of unknown composite faults that conform to the real characteristic patterns. In this way, this step not only solves the problem of data scarcity, but also, by introducing unknown samples, enables the model to learn to recognize unknown fault modes, laying the foundation for subsequent applications of zero-sample diagnosis.

[0069] S400: Obtain real single-fault acoustic fingerprint samples of transformers, and construct a diversified training set based on real single-fault acoustic fingerprint samples, decoupled fault features, and unknown composite fault acoustic fingerprint samples.

[0070] Specifically, step S400 involves combining real single-fault voiceprint samples, decoupled fault features, and unknown composite fault voiceprint samples according to a preset ratio (e.g., 1:1:1) to obtain a diversified training set. Diverse training sets The structure is as follows:

[0071] In the formula, This is a real single-fault voiceprint sample; To decouple fault characteristics; This is a voiceprint sample for an unknown composite fault.

[0072] In practical applications, you can choose to combine 2000 real single-fault voiceprint samples per class, decoupled fault features, and generated unknown composite fault voiceprint samples in a 1:1:1 ratio to obtain a diversified training set. .

[0073] S500 utilizes a multi-label classifier combined with a diverse training set for fusion training to obtain a trained fault classifier, which can be used for comprehensive fault diagnosis of transformers.

[0074] This step is based on the diversified training set constructed by fusing real single-fault voiceprint samples, decoupled fault features and generated unknown composite fault voiceprint samples. It adopts a multi-label classifier based on residual network and a fusion training strategy to achieve accurate identification of known faults and unknown composite faults and model quantization verification.

[0075] Specifically, see Figure 5 The above step S500 includes the following steps: S501. Construct a multi-label classifier based on residual networks. The multi-label classifier uses binary encoding to label the multiple fault labels of the samples and includes network layers using preset activation functions. S502. Introduce a fusion training strategy by constructing multiple dedicated classifiers to train on specific composite faults and single faults; each dedicated classifier is configured to focus on distinguishing between a specific type of composite fault and single fault. S503. The multi-label classifier is trained by combining the preset optimizer and the cross-entropy loss function to obtain a fault classifier that can perform comprehensive fault diagnosis of transformers.

[0076] (1) The multi-label uses binary encoding. The fault types contained in the samples in the diversified training set are marked by binary indicators, where "1" indicates that the current sample contains the corresponding fault type and "0" indicates that it does not contain the fault type. A label matrix is ​​constructed. (here) The total number of samples, (Number of fault types).

[0077] (2) Classifier network structure: Based on the residual network, it includes convolutional blocks, pooling layers, residual blocks and fully connected layers. Local features are extracted through convolutional operations, normalization is performed, SiLU activation function is used to enhance nonlinear expression, pooling operation is used to reduce dimensionality and residual connection is used to avoid gradient vanishing. Finally, the probability of each class is output through fully connected layers and softmax layers. .

[0078] (3) Fusion Training Strategy: For different types of compound faults, three dedicated classifiers are constructed. Each classifier focuses on distinguishing between a compound fault and a single fault (e.g., it is forced to be assigned a specific, easily confused compound fault diagnosis task). During training, corresponding sample subsets are assigned according to the fault type, and the Adam optimizer (learning rate is 100%) is used. The training process is constrained by a batch size of 32 and a cross-entropy loss function to minimize the difference between the fault prediction and the actual value.

[0079] Among them, the cross-entropy loss function The format is as follows:

[0080] Specifically, this embodiment of the invention constructs a multi-label classifier with a ResNet backbone, using the SiLU activation function at the network ends, so that each output neuron independently represents the probability of a certain type of fault. In addition to the main multi-label classifier, this embodiment can train multiple specialized classifiers in parallel, each focusing on distinguishing a specific composite fault from all single faults. These specialized classifiers are trained collaboratively with the main classifier through knowledge distillation or feature sharing mechanisms. During training, a diversified training set is used, and an adaptive optimizer and a binary cross-entropy loss function are employed for end-to-end training of the network, effectively improving the ability to distinguish complex composite patterns. It should be noted that during training, only samples containing the corresponding configured composite faults and their negative examples (single fault samples) are used to calculate the loss of the three specialized classifiers.

[0081] Finally, for the fault classifier obtained after training, the transformer acoustic waveform signal to be diagnosed can be preprocessed in step S100, and then the decoupling features can be extracted by the encoder in step S200 (or the time-frequency diagram can be used directly). The extracted features can be input into the fault classifier to obtain the corresponding multi-label prediction results and complete the comprehensive fault diagnosis.

[0082] In a preferred embodiment, after the trained fault classifier is completed, the following steps are further included: Step 1: Input the test sample set containing known and unknown faults into the trained fault classifier and obtain the fault type prediction result output by the fault classifier. Step 2: Use the accuracy rates of known faults, unknown faults, and the harmonic mean scores of the accuracy rates of known faults and unknown faults as evaluation indicators to evaluate the fault type prediction results and validate the fault classifier.

[0083] In practical applications, test set samples (1500 known and 1500 unknown fault samples per class) are input into the trained fault classifier, which outputs the fault type prediction results. Finally, the accuracy of known faults, the accuracy of unknown faults, and the harmonic mean score (the weighted harmonic mean of the accuracy of known and unknown faults) are used as evaluation indicators to comprehensively verify the diagnostic ability of the fault classifier.

[0084] Among them, (1) the formula for calculating the fault accuracy is as follows:

[0085] (2) The formula for calculating the accuracy of unknown faults is as follows:

[0086] (3) The formula for calculating the harmonic mean fraction of the accuracy rate for known faults and the accuracy rate for unknown faults is as follows:

[0087] Ultimately, the above evaluation indicators can comprehensively assess the effectiveness, generalization ability, and practical value of the fault classifier in this diagnostic method. They can also assist technicians in optimizing the fault classifier and continuously improving the diagnostic capabilities of the fault classifier model.

[0088] Based on the above description, this invention provides a method for diagnosing composite faults in transformers based on zero-sample feature decoupling. This method acquires acoustic signature signals by deploying sensors in key areas such as the transformer tank, windings, and core. It innovatively employs a strategy combining discrete wavelet transform and continuous wavelet transform for signal preprocessing: first, the signal is decomposed using discrete wavelet transform to extract details and approximation coefficients, thereby constructing an attribute set containing single and composite fault modes; then, continuous wavelet transform is used to generate a standardized two-dimensional time-frequency fusion diagram, completing the representation of the signal from a one-dimensional sequence to two-dimensional structured features.

[0089] Next, the core of the solution lies in constructing a generative adversarial network (GAN) specifically for feature decoupling. This network encodes the input features into a composite tensor containing fault features and irrelevant interference factors through its generator. It also designs a method to reconstruct the fault feature tensors after exchanging different samples, and uses a discriminator with multiple classification heads for multi-task adversarial training. Ultimately, it achieves the goal of obtaining high-purity decoupled fault features from the mixed features, effectively solving the interference problem of irrelevant factors such as load fluctuations and environmental noise.

[0090] To address the challenge of scarce composite fault samples, this solution introduces an improved conditional diffusion model. Using the aforementioned composite fault attribute set as a constraint, it generates a large number of realistic unknown composite fault acoustic signature samples by adding noise and iteratively denoising real single fault samples. Subsequently, the solution combines real single fault samples, decoupled clean fault feature samples, and generated unknown composite fault samples in a proportional manner to construct a diversified training set. Finally, a multi-label classifier based on a residual network architecture is employed, and a fusion training strategy is introduced to construct multiple dedicated classifiers focused on distinguishing specific composite faults from single faults. These are trained using the diversified training set to obtain a fault classifier capable of simultaneously outputting multiple fault labels.

[0091] This classifier can not only diagnose known faults with high accuracy, but also has the ability to identify unknown compound faults by learning from the generated unknown samples during training. It is comprehensively verified by indicators such as the accuracy of known faults, the accuracy of unknown faults and their harmonic mean scores, thus realizing comprehensive, accurate and robust intelligent diagnosis of transformers from single faults to compound faults, and from known types to unknown types.

[0092] As can be seen, firstly, this invention uses voiceprint signals as the core data source and achieves simultaneous extraction and fusion of time-domain and frequency-domain features through a combination transformation strategy. This avoids the information limitations of single-dimensional analysis and provides more complete feature support for fault diagnosis. Compared with traditional single-modal feature extraction methods, the feature coverage is significantly improved. Secondly, through a specially designed decoupled generative adversarial network, the core fault features are effectively separated from irrelevant interference factors. Redundant information such as load fluctuations and environmental noise is eliminated, making the extracted fault features more identifiable and providing a foundation for subsequent classification and diagnosis. This solves the problem of weakened effective features in traditional feature extraction. Thirdly, by using an improved conditional diffusion model, unknown composite fault voiceprint samples are generated with fault attribute sets as constraints. This effectively compensates for the industry problem of scarce composite fault samples, and the generated samples have good feature consistency with real samples, enriching the diversity of the training set and greatly improving the model's generalization ability. Fourthly, a diversified training set is constructed by integrating real samples, decoupled feature samples, and generated samples. A multi-label classifier fusion training strategy is employed to specifically improve the ability to distinguish between compound faults and single faults. This not only achieves high-precision identification of known faults but also overcomes the limitations of traditional methods in diagnosing unknown faults. Furthermore, it enhances anti-interference capabilities and adapts to diagnostic needs in complex operating environments. Finally, this invention is specifically designed for transformer acoustic signature signals, expanding the data source types for transformer fault diagnosis. It can simultaneously handle diagnostic scenarios for single faults, known compound faults, and unknown compound faults, meeting the practical needs of power systems for early warning and accurate diagnosis of transformer faults throughout their entire lifecycle, thus enhancing its practicality.

[0093] The effects of this invention in actual experiments will be described below: (1) Figure 6 The similarity curves before and after decoupling cover 5 types of faults. Before decoupling, the average intra-class similarity was 0.40 and the average inter-class similarity was 0.26. After decoupling, the average intra-class similarity increased to 0.86 and the inter-class similarity decreased to 0.11, indicating a significant improvement in intra-class clustering effect and inter-class discriminability.

[0094] (2) Subsequently, an improved conditional diffusion model was loaded, and Gaussian noise with linear increments of 0.1-0.8 was added in the forward direction to generate 100 sets of diffusion data. Then, 50 steps of denoising and reconstruction were performed in the reverse direction to generate 500 unknown composite fault samples per class. Figure 7 The generated samples show good consistency with the real samples in terms of time-frequency characteristics. The real samples have amplitudes of 10 and 5 at 200Hz and 350Hz, respectively, while the generated samples have peak values ​​shifted to 210Hz and 340Hz, with amplitudes of 9.5 and 4.8, respectively. The average amplitude error is only 0.32.

[0095] (3) The test set contains 1500 known and 1500 unknown fault samples per class, and three evaluation indicators are used for verification; Figure 8The accuracy curves of different methods show that the accuracy of the present invention for known faults is 98.2%, the accuracy for unknown faults is 92.5%, and the harmonic mean score of the accuracy of known faults and unknown faults is 95.3%, which is significantly better than the traditional ResNet (89.6%, 65.8%, 76.1%) and the ordinary GAN+ classifier (92.1%, 78.3%, 84.7%), thus comprehensively ensuring the accuracy and robustness of transformer fault diagnosis.

[0096] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A composite fault diagnosis method based on transformer acoustic signature, characterized in that, include: The transformer acoustic signature signal is acquired and preprocessed to obtain the two-dimensional time-frequency fusion features of the transformer acoustic signature signal. A decoupled generative adversarial network is constructed, and the network is used to decouple and separate irrelevant interference factors in the two-dimensional time-frequency fusion features to obtain decoupled fault features. Based on the improved conditional diffusion model, unknown composite fault acoustic signature samples are generated. Acquire real single-fault acoustic fingerprint samples of transformers, and construct a diversified training set based on the real single-fault acoustic fingerprint samples, the decoupled fault features, and the unknown composite fault acoustic fingerprint samples. A multi-label classifier is used in conjunction with the diversified training set for fusion training to obtain a trained fault classifier, which can be used for comprehensive fault diagnosis of the transformer.

2. The composite fault diagnosis method based on transformer acoustic signature according to claim 1, characterized in that, The acquisition of the transformer acoustic signature signal includes: Multiple acoustic sensors are deployed in the key areas of the transformer, which include at least the transformer housing, windings, and core. The transformer acoustic signature signal is collected during the transformer's operation according to the preset acquisition frequency.

3. The composite fault diagnosis method based on transformer acoustic signature according to claim 1, characterized in that, Preprocessing the transformer acoustic signature signal includes: Discrete wavelet transform is used to decompose the acoustic signature signal of the transformer, and detail coefficients and approximation coefficients are extracted. Based on the detail coefficients and approximation coefficients, a composite fault attribute set containing single faults and composite faults is constructed, and the composite fault attribute set is used to train the multi-label classifier. The transformer acoustic signature signal after discrete wavelet transform is converted into a two-dimensional time-frequency fusion map using continuous wavelet transform, and the two-dimensional time-frequency fusion map is preprocessed and optimized to obtain two-dimensional time-frequency fusion features.

4. The composite fault diagnosis method based on transformer acoustic signature according to claim 3, characterized in that, The decoupled generative adversarial network includes: a generator, a discriminator, and an auxiliary classifier; The generator is used to encode and decode the two-dimensional time-frequency fusion features; The generator includes an encoder and a decoder; the encoder encodes the two-dimensional time-frequency fusion features into a composite feature tensor containing fault features and irrelevant factor features; the decoder reconstructs an extended two-dimensional time-frequency fusion graph based on the composite feature tensor. The discriminator is equipped with multiple classification heads, which are used to simultaneously distinguish the authenticity of the fault and irrelevant interference factors in the extended two-dimensional time-frequency fusion map; The auxiliary classifier is used to classify and verify the fault features obtained after decoupling the extended two-dimensional time-frequency fusion graph.

5. The composite fault diagnosis method based on transformer acoustic signature according to claim 4, characterized in that, The decoupled generative adversarial network is used to decouple and separate irrelevant interference factors in the two-dimensional time-frequency fusion features to obtain decoupled fault features, including: At least two two-dimensional time-frequency fusion maps with different fault types are input into the encoder of the generator to encode the corresponding composite feature tensors; each composite feature tensor contains a fault feature tensor and an irrelevant interference factor tensor. The fault feature tensors in at least two of the composite feature tensors are swapped and recombined with their respective irrelevant interference factor tensors to form a variety of new composite feature tensors; The new composite feature tensor is input into the decoder of the generator for reconstruction to obtain the extended two-dimensional time-frequency fusion map. The extended two-dimensional time-frequency fusion graph is input into the decoupled generative adversarial network and jointly trained with a multi-objective loss function. After the joint training is completed, the fault feature tensor is extracted as the decoupled fault feature.

6. The composite fault diagnosis method based on transformer acoustic signature according to claim 3, characterized in that, Based on an improved conditional diffusion model, unknown composite fault acoustic signature samples are generated, including: Based on the real single-fault acoustic print samples, and using the composite fault attribute set as a constraint, Gaussian noise is gradually added to the real single-fault acoustic print samples to generate diffusion data with different noise levels. The improved conditional diffusion model is used to predict and gradually remove noise from the diffusion data. After iterative reconstruction, the unknown composite fault acoustic signature sample is obtained.

7. The composite fault diagnosis method based on transformer acoustic signature according to claim 1, characterized in that, A diversified training set is constructed based on the real single-fault voiceprint samples, the decoupled fault features, and the unknown composite fault voiceprint samples, including: The real single-fault voiceprint samples, the decoupled fault features, and the unknown composite fault voiceprint samples are combined according to a preset ratio to obtain the diversified training set.

8. The composite fault diagnosis method based on transformer acoustic signature according to claim 1, characterized in that, The multi-label classifier is combined with the aforementioned diverse training set for fusion training, including: A multi-label classifier based on residual networks is constructed. The multi-label classifier uses binary encoding to label multiple fault labels of samples and includes network layers using a preset activation function. A fusion training strategy is introduced, which trains multiple specialized classifiers on specific composite faults and single faults; each of the specialized classifiers is configured to focus on distinguishing between a specific type of composite fault and single fault. The multi-label classifier is trained by combining a preset optimizer and a cross-entropy loss function to obtain a fault classifier capable of comprehensive fault diagnosis of the transformer.

9. The composite fault diagnosis method based on transformer acoustic signature according to claim 1, characterized in that, After the fault classifier is trained, it includes: Input the test sample set containing known faults and unknown faults into the fault classifier obtained through training, and obtain the fault type prediction result output by the fault classifier; The known fault accuracy, the unknown fault accuracy, and the harmonic mean score of the known fault accuracy and the unknown fault accuracy are used as evaluation indicators to evaluate the fault type prediction results and to validate the fault classifier.

Images