A transformer oil dissolved gas fault diagnosis method, device, equipment and medium

Abstract

This solution belongs to the field of fault diagnosis technology, and discloses a method, device, and medium for diagnosing dissolved gas faults in transformer oil. The method includes: acquiring raw gas concentration data of various fault-characterizing gases in transformer oil; preprocessing the raw gas data to construct a raw data sample set; inputting the raw data sample set into a smoothing regularized variational framework to generate a synthetic data sample set; inputting the synthetic data sample set and the raw data sample set into a ratio control domain cyclic mechanism to dynamically adjust the training ratio, and extracting multi-domain features to optimize the synthetic data sample set, obtaining an optimized synthetic data sample set; filtering the optimized synthetic data sample set and merging it with the raw data sample set to construct a balanced training dataset; training a fault diagnosis model using the balanced training dataset; and diagnosing the test gas data based on the fault diagnosis model and outputting the fault type. This solution improves the diagnostic accuracy.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a method, apparatus, equipment, and medium for diagnosing dissolved gas faults in transformer oil. Background Technology

[0002] Existing fault diagnosis technologies for dissolved gases in transformer oil mainly consist of two parts: sensor detection and data processing. Regarding sensors, traditional metal-oxide-semiconductor (MOSFET) sensors or sensor arrays are commonly used, with heating primarily relying on constant-temperature heating. Gas concentration detection is achieved by generating electrical signals through the adsorption-desorption reaction between the sensitive material and the gas at a fixed temperature. Some MEMS sensors employ simple support structures without effective suspension and thermal insulation design, and their electrode layouts are often discrete, lacking coplanar integrated structures. In terms of data processing and diagnosis, to address the scarcity of fault data, traditional generative models are frequently used for data synthesis, failing to fully exploit multi-domain features. Data fusion often employs direct mixing or simple filtering methods, lacking multi-dimensional screening of the synthesized data quality. Fault diagnosis models often rely on single-domain feature training, failing to incorporate synergistic features from the spatial, time-frequency, and frequency domains. Furthermore, some technologies improve gas identification capabilities by combining sensor arrays with simple machine learning algorithms, but this increases system complexity and power consumption.

[0003] Regarding sensors: the constant temperature heating mode consumes a lot of power and cannot decouple the effects of temperature, gas diffusion characteristics and reaction kinetics on the sensing signal, resulting in poor gas recognition selectivity; non-suspended MEMS sensors suffer from severe heat loss and slow thermal response, making it difficult to quickly capture gas transient responses; although sensor arrays can improve selectivity, they have problems such as complex structure, high cost and poor adaptability, and are not suitable for long-term monitoring of transformer oil immersion environments.

[0004] In terms of data synthesis: the training process of traditional generative models is unstable and prone to gradient vanishing or pattern collapse, resulting in insufficient structural consistency and realism of synthesized data; it ignores the complementary value of multi-domain features and relies only on single-domain features for modeling, leading to a lack of rich feature expression in synthesized data.

[0005] Regarding data fusion: an effective multi-dimensional quality screening mechanism has not been established. Simple mixing or filtering results in inconsistent quality of fused data, which cannot provide robust support for diagnostic models.

[0006] In terms of fault diagnosis: Due to the scarcity of data and class imbalance, the diagnostic model has weak generalization ability, insufficient accuracy and stability, and is difficult to meet the high-precision diagnostic needs in engineering practice. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a method for diagnosing dissolved gas faults in transformer oil, the method comprising: Obtain raw gas concentration data of various fault characterization gases in transformer oil based on pulsed levitation MEMS sensors; The raw gas data is preprocessed to construct a raw data sample set; The original data sample set is input into a pre-constructed smooth regularized variational framework to generate a synthetic data sample set, wherein the smooth regularized variational framework is constructed based on a variational autoencoder and integrates a smooth activation function and a multi-level continuous regularization strategy. The synthetic data sample set and the original data sample set are input into the ratio control domain loop mechanism so that the ratio control domain loop mechanism dynamically adjusts the training ratio during the discriminator-dominated phase and the synthesizer enhancement phase. After collaboratively extracting spatial domain features, time-frequency domain features and frequency domain features, the synthetic data sample set is optimized based on the extracted multi-domain features to obtain the optimized synthetic data sample set. Based on a multi-index-guided adaptive data fusion strategy, the synthetic data optimization sample set is screened, and the screened data sample set is fused with the original data sample set to construct a balanced training dataset. The fault diagnosis model is trained using the balanced training dataset. The fault diagnosis model is used to diagnose the gas data under test and output the fault type.

[0008] In some embodiments, the pulse-suspended MEMS sensor includes: silicon substrate; Microcavities formed on the silicon substrate; A suspension support layer, which is suspended above the microcavity; The outer serpentine heating electrode is arranged in a serpentine configuration in the outer region of the suspension support layer; An inner sensing electrode is arranged coplanarly with the outer serpentine heating electrode in the inner region of the suspension support layer. The inner sensing electrode surrounds a central region, and a sensitive material is disposed in the central region. The outer serpentine heating electrode is used to receive pulse voltage signals to pulse-heat the sensitive material.

[0009] In some embodiments, preprocessing the raw gas data to construct a raw data sample set includes: The original gas concentration data is normalized to map all concentration values ​​to the [0, 1] interval; The normalized gas concentration time series data is segmented into segments of fixed length. The segmented data is used to construct the original data sample set, wherein the dimensions of the original data sample set are [number of gas components, time series length, number of sensors].

[0010] In some embodiments, the smoothing regularization variational framework includes at least three sequentially connected two-dimensional convolutional layers, with a dropout regularization module connected after each convolutional layer. The step of inputting the original data sample set into the pre-constructed smoothing regularization variational framework to generate a synthetic data sample set includes: The original data sample set is input into a pre-constructed smooth regularized variational framework, so that the smooth regularized variational framework, through the encoder of the variational autoencoder, determines an approximate posterior distribution based on the original data sample set and dropout mask, and uses a reparameterization technique to sample latent variables from the approximate posterior distribution, wherein the approximate posterior distribution is an inferred value that replaces the true posterior distribution. Based on the generator, the latent variables are processed by the at least three two-dimensional convolutional layers and a smooth activation function to obtain the synthetic data sample set. The approximate posterior distribution is: ; The latent variables are: ; In the formula, It is a latent variable; x represents the original data sample set; For dropout mask; Indicates a Gaussian distribution; Represents the approximate posterior distribution of latent variables under the conditions of the original samples and the mask.

[0011] It is the mean of the approximate posterior distribution; The variance is the approximate posterior distribution. The mean is variance is Gaussian distribution; This represents the posterior distribution mean of the encoder output: The standard deviation of the posterior distribution of the encoder output is: (0, I) represents standard Gaussian noise; I is a unit diagonal matrix with the same dimension as the latent variable z, used to construct the covariance matrix of multidimensional independent standard Gaussian noise. This satisfies the gradient differentiability and latent variable space decoupling requirements of the variational autoencoder's reparameterized sampling, ensuring the stability of model training.

[0012] In some embodiments, the step of inputting the synthetic data sample set and the original data sample set into a rate control domain recurrent mechanism, such that the rate control domain recurrent mechanism dynamically adjusts the training ratio during the discriminator-dominated phase and the synthesizer enhancement phase, and collaboratively extracts spatial domain features, time-frequency domain features, and frequency domain features, and then optimizes the synthetic data sample set based on the extracted multi-domain features to obtain an optimized synthetic data sample set, includes: The synthetic data sample set and the original data sample set are input into the rate control domain loop mechanism, so that the rate control domain loop mechanism performs the following operations on the synthetic data sample set and the original data sample set respectively: The number of sensors and the length of the time series are independently encoded and feature fused to extract the spatial domain features corresponding to the synthetic data sample set and the original data sample set, respectively. Multi-scale analysis was performed on the gas concentration time series to extract the time-frequency domain features corresponding to the synthetic data sample set and the original data sample set, respectively. Determine the dynamic spectral loss between the synthetic data sample set and the original data sample set to determine the corresponding frequency domain features of the synthetic data sample set and the original data sample set, respectively. Based on the current number of training iterations, a ratio control factor is applied to adjust the training phase. Specifically, when the current number of training iterations is not greater than the critical number of iterations, the ratio control factor is set to the first value, and the training is in the discriminator-dominated phase. When the current number of training iterations is less than the critical number of iterations, the ratio control factor is set to the second value, and the training is in the synthesizer enhancement phase, where the first value is greater than the second value. By alternating the training with the ratio control through the iterative transfer between spatial domain features, time-frequency domain features, and the time-frequency domain features during the iterative process, data perturbation is introduced in each round of training. The synthetic data sample set is then optimized based on the extracted spatial domain features, time-frequency domain features, and frequency domain features, resulting in an optimized synthetic data sample set that is highly consistent with the original data sample set in the multi-domain feature space.

[0013] In some embodiments, the step of independently encoding and fusing features in the sensor quantity dimension and the time series length dimension, and extracting the spatial domain features corresponding to the synthetic data sample set and the original data sample set respectively, includes: The number of sensors is independently encoded, and a height orientation feature map is extracted. The length dimension of the time series is independently encoded, and the feature map in the width direction is extracted. The height-direction feature map and the width-direction feature map are concatenated along the channel dimension to obtain a concatenated feature map; The concatenated feature maps are then subjected to convolutional fusion processing to obtain intermediate feature maps; The intermediate feature map is subjected to tensor mapping to obtain the mapped spatial domain features.

[0014] In some embodiments, the multi-index-guided adaptive data fusion strategy, which filters the synthetic data optimization sample set and merges the filtered data sample set with the original data sample set to construct a balanced training dataset, includes: Determine the KL divergence value between each synthetic data sample in the synthetic data sample set and the entire original data sample set; All synthetic samples are sorted in first order based on multiple KL divergence values; Based on the sorting results of the first order sorting and combined with the preset first elimination rate, the distribution difference screening is carried out to obtain the first screening sample set after the first stage screening. The first screening sample set includes multiple first screening samples. Determine the maximum mean difference between each selected sample and the original data sample set, whereby the maximum mean difference is used to measure the statistical distance between sample points; Based on multiple maximum mean differences, a second sorting order is performed on multiple primary screening sample sets. Based on the sorting results of the second order sorting and combined with the preset second elimination rate, statistical distance screening is performed to obtain a secondary screening sample set after the second stage screening. The secondary screening sample set includes multiple secondary screening samples. Determine the difference in power spectral density between each secondary screening sample and its corresponding original sample in the original data sample set; Based on the differences in multiple power spectral density values, the multiple secondary screening samples included in the secondary screening sample set are sorted in a third order. Based on the sorting results of the third order, power spectral density difference screening is performed according to the preset third elimination rate to obtain a three-stage screening sample set, which includes multiple three-stage screening samples. Determine the mean square error between each triplet sample and its corresponding original sample in the original data sample set; Based on multiple mean square error values, the sample set of three screenings, including multiple samples of three screenings, is sorted in a fourth order. Based on the fourth order sorting and the preset fourth elimination rate, the mean square error value is used for screening, and the selected sample set is used as the target screening sample set. Based on the target selection sample set and the original data sample set, the balanced training dataset is determined.

[0015] According to another aspect of this application, a fault diagnosis device for dissolved gases in transformer oil is also disclosed, the device comprising: The raw gas concentration data acquisition module is used to acquire the raw gas concentration data of various fault characterization gases in transformer oil based on pulsed suspension MEMS sensors. The raw data sample set construction module is used to preprocess the raw gas data and construct the raw data sample set; A synthetic data sample set determination module is used to input the original data sample set into a pre-constructed smooth regularized variational framework to generate a synthetic data sample set, wherein the smooth regularized variational framework is constructed based on a variational autoencoder and integrates a smooth activation function and a multi-level continuous regularization strategy. The synthetic data optimization sample set determination module is used to input the synthetic data sample set and the original data sample set into the ratio control domain loop mechanism, so that the ratio control domain loop mechanism dynamically adjusts the training ratio in the discriminator-dominated phase and the synthesizer enhancement phase, and after collaboratively extracting spatial domain features, time-frequency domain features and frequency domain features, optimizes the synthetic data sample set based on the extracted multi-domain features to obtain the synthetic data optimization sample set. The balanced training dataset construction module is used to filter the synthetic data optimization sample set based on a multi-index-guided adaptive data fusion strategy, and then merge the filtered data sample set with the original data sample set to construct a balanced training dataset. The fault diagnosis model training module is used to train the fault diagnosis model using the balanced training dataset. The fault type output module is used to diagnose the gas data under test based on the fault diagnosis model and output the fault type.

[0016] According to another aspect of this application, an electronic device is also disclosed, the electronic device including a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform the steps of the method for diagnosing dissolved gas faults in transformer oil as described in any of the preceding claims.

[0017] According to another aspect of this application, a computer-readable storage medium is also disclosed, wherein instructions are stored on the computer-readable storage medium, characterized in that, when executed by a processor, the instructions implement the various steps of the method for diagnosing dissolved gas faults in transformer oil as described in any of the preceding claims.

[0018] The present invention includes, but is not limited to, the following beneficial effects: (1) Significantly improved sensor detection performance: The suspended support structure and coplanar electrode design achieve efficient thermal insulation. The pulse drive mode reduces power consumption while decoupling the influence of temperature, diffusion characteristics and reaction kinetics on the sensing signal. A single sensor can achieve high selectivity identification of various fault gases and is suitable for long-term monitoring of transformer oil immersion environment; (2) Significantly optimized data synthesis quality: The smooth regularized variational framework effectively improves training stability, avoids gradient vanishing and mode collapse, and ensures the structural consistency of the synthesized data; The ratio control domain loop mechanism fully explores the multi-domain collaborative features, improves the authenticity and diversity of the synthesized data, and provides sufficient high-quality data support for subsequent diagnosis; (3) Better data fusion effect: The multi-index guided adaptive screening strategy controls the quality of synthesized data from multiple dimensions such as global distribution, statistical distance, frequency domain features and time domain details. The fused dataset has strong feature representativeness and effectively alleviates the problems of data scarcity and class imbalance; (4) Improved diagnostic accuracy and stability: The diagnostic model trained based on multi-domain collaborative features and high-quality fused data has significantly enhanced generalization ability and more reliable diagnostic results, which can meet the high-precision and stable diagnostic needs of transformer faults in engineering practice. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0020] Figure 1 This is a flowchart of a method for diagnosing dissolved gas faults in transformer oil according to an embodiment of this application; Figure 2 This is a structural block diagram of the transformer oil dissolved gas fault diagnosis device according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application; Figure 4 This is a schematic diagram of the sensor structure according to an embodiment of this application; Figure 5 This is a comparison chart of the effects of adaptive data fusion guided by multiple indicators; Figure 6 This is a comparison chart of the accuracy rates of different methods for transformer fault diagnosis. Detailed Implementation

[0021] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] For ease of understanding, the specific process of the embodiments of the present invention will be described below. Figure 1 This is a flowchart illustrating the fault diagnosis of dissolved gas in transformer oil according to an embodiment of this application. (See attached diagram.) Figure 1 It includes the following steps: S100: Obtain the raw gas concentration data of each fault characterization gas in transformer oil based on pulsed suspension MEMS sensor.

[0023] S101. Preprocess the raw gas data to construct a raw data sample set.

[0024] S102. Input the original data sample set into a pre-constructed smoothed regularized variational framework to generate a synthetic data sample set.

[0025] Among them, the smooth regularization variational framework is built on variational autoencoder and integrates smooth activation function and multi-level continuous regularization strategy; S103. Input the synthetic data sample set and the original data sample set into the ratio control domain loop mechanism so that the ratio control domain loop mechanism dynamically adjusts the training ratio during the discriminator-dominated phase and the synthesizer enhancement phase. After collaboratively extracting spatial domain features, time-frequency domain features and frequency domain features, the synthetic data sample set is optimized based on the extracted multi-domain features to obtain the optimized synthetic data sample set.

[0026] S104. Based on a multi-index-guided adaptive data fusion strategy, the synthetic data optimization sample set is screened, and the screened data sample set is fused with the original data sample set to construct a balanced training dataset.

[0027] S105. Train the fault diagnosis model using a balanced training dataset.

[0028] S106. Based on the fault diagnosis model, diagnose the gas data under test and output the fault type.

[0029] Among them, such as Figure 4 As shown, the pulse-levitated MEMS sensor includes: silicon substrate; Microcavities formed on silicon substrates; Suspension support layer, which is suspended above the microcavity; The outer serpentine heating electrode is arranged in a serpentine configuration on the outer region of the suspension support layer. The inner sensing electrode is arranged coplanarly with the outer serpentine heating electrode in the inner region of the suspension support layer. The inner sensing electrode surrounds a central region, and sensitive material is arranged in the central region. The outer serpentine heating electrode is used to receive pulse voltage signals to pulse-heat the sensitive material.

[0030] Furthermore, the following describes the specific steps involved in this solution. Figure 1 The overall scheme shown is explained step by step: Step 1: Data Acquisition and Preprocessing of Dissolved Gases in Transformer MEMS Oil Data Acquisition: Dissolved gas data in transformer oil are collected using the aforementioned MEMS sensors, covering normal operating conditions and typical fault conditions (partial discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating, high-temperature overheating, etc.). The collected characteristic gases include key fault characterization gases such as H2, CH4, C2H6, C2H4, and C2H2.

[0031] Data preprocessing: The raw gas concentration data is normalized by min-max to map all gas component concentration values ​​to the [0,1] interval, eliminating dimensional differences; the time series data is segmented into segments of a fixed length of 1024 data points / samples to form a sample set with dimensions of [number of gas components × time series length × number of sensors], providing standardized input for subsequent latent variable modeling and multi-domain feature extraction.

[0032] Step 2: Construct a smooth regularized variational framework to guide the synthesis of high-quality data. Innovation: A smooth regularization variational framework is proposed, with explicit modeling of variational autoencoders (VAE) as the core. It integrates smooth activation function mixing and multi-layer continuous regularization strategies to stabilize the latent variable space representation, alleviate gradient vanishing and pattern collapse problems during training, and improve the structural consistency and training stability of synthetic data. 1. Framework Structure Construction: A smooth regularization variational framework is constructed before the generator of the Generative Adversarial Network (GAN). This framework contains three two-dimensional convolutional layers (Conv), each followed by a dropout regularization module. The core components consist of a smooth activation function mixing module and a multi-layer continuous regularization module, which are used to guide the generator to output structured and highly consistent synthetic data.

[0033] 2. Smooth activation function configuration: The first two convolutional layers use the LeakyReLU activation function; the third convolutional layer uses the tanh activation function; 3. Multi-layer continuous regularization implementation: Dropout regularization is applied after each convolutional layer. The dropout mask βi ~ Bernoulli (keeping probability), where the keeping probability is the neuron retention probability (set to 0.8). By randomly deactivating some activation units, the excessive coupling of the latent variable space is reduced, the risk of local overfitting is reduced, and the generalization ability of the generative model is improved.

[0034] 4. Latent variable modeling and sampling: encoder output average and standard deviation Construct an approximate posterior distribution:

[0035] Parameter explanation: x is the dissolved gas sample data in the transformer MEMS oil after preprocessing in step 1; β is the dropout mask; N represents a Gaussian distribution; It is the mean of the approximate posterior distribution; The variance is approximated by the posterior distribution, which is used to achieve structured modeling of data features.

[0036] Generate latent variable z using reparameterization techniques:

[0037] Parameter explanation: (0,I) represents standard Gaussian noise, I is a unit diagonal matrix with the same dimension as the latent variable z, used to construct the covariance matrix of multidimensional independent standard Gaussian noise; z is the latent variable obtained by sampling, which retains the core features of the original data and is used to input the generator for data synthesis.

[0038] 5. Objective Function Optimization: Introduce a variational lower bound objective function with dropout noise, expressed as:

[0039] Parameter explanation: E represents expectation; p(x|z) is the likelihood function of the generative model, measuring the probability that the latent variable z generates the original data x; DKL is the KL divergence, used to measure the approximate posterior distribution from the prior distribution (standard Gaussian distribution). The difference between the latent variable space and the objective function is minimized to achieve stable modeling of the latent variable space, providing reliable feature input for the generator. Step 3: Establish a ratio control domain loop mechanism to enhance multi-domain feature modeling. Innovation: A ratio control domain cyclic mechanism is proposed to dynamically mine multi-domain features in the spatial, time-frequency, and frequency domains and achieve collaborative transfer. By adjusting the training phase (discriminator-dominated phase and synthesis enhancement phase) through the ratio control factor, the problem of traditional methods neglecting the complementarity of multi-domain features is solved, thereby improving the authenticity and diversity of synthetic data. 1. Spatial domain feature extraction (generator side): A direction-aware spatial feature extraction module is introduced to independently encode the input tensor X∈ℝA×B×C (A is the number of sensors, B is the time series length, and C is the number of gas components) in the height A and width B directions, thereby extracting direction-sensitive spatial features:

[0040] Parameter explanation: tcm(m) is the feature value of the c-th gas component at position m; m=a corresponds to the height direction (sensor quantity dimension), m=b corresponds to the width direction (time series length dimension); xc(m,i) is the concentration value of the c-th component at position (m,i); A is the number of sensors, B is the time series length, and this formula is used to extract the gas concentration distribution characteristics of different sensors.

[0041] The vertical and horizontal feature maps are concatenated and then input into the shared convolutional module H1:

[0042] Parameter explanation: δ is the LeakyReLU nonlinear activation function; [·] is the dimension concatenation operation; t a For the height-oriented feature map, t b H1 is the feature map in the width direction; H1 is the shared convolutional module; h is the fused intermediate feature map, realizing cross-directional information unification.

[0043] Maintain channel consistency through 1×1 convolution operators Ha and Hb mapping:

[0044] Parameter explanation: σ is the sigmoid activation function; h m H represents the directional feature tensor after splitting. m For a 1×1 convolution operator; s m For the mapped feature tensor, ensure that the channel dimension is consistent with the original input.

[0045] Generate spatial domain features:

[0046] Parameter explanation: (i,j) represents the output data containing spatial domain features; (i,j) represents the original input concentration value; (i) (j) represents feature tensors in different directions, which are fused using element-wise multiplication to integrate spatial correlation features.

[0047] 2. Time-frequency domain feature extraction (discriminator end): A time-frequency feature extraction module based on continuous wavelet transform is used to reconstruct the one-dimensional gas concentration time series into two dimensions. Then, multi-scale analysis is performed using Mallat-type two-dimensional wavelet transform (Haar wavelet) to generate four frequency bands (one approximate frequency band + three directional detail frequency bands).

[0048]

[0049] Parameter explanation: φ is the scaling function, ψ is the wavelet function; P is the direction parameter (Ho = horizontal direction, Ve = vertical direction, Di = diagonal direction); a and b are the horizontal and vertical coordinates of the two-dimensional data. This transformation converts the dynamic characteristics of gas concentration changing with time into time-frequency domain characteristics, enhancing the discriminator's ability to perceive local time-frequency patterns.

[0050] 3. Frequency domain feature extraction (loss function side): By introducing dynamic spectral loss to capture frequency domain differences, a two-dimensional spectral representation is generated based on Euler's formula:

[0051] Parameter explanation: F(u,v) is the frequency domain representation of the data; u,v are the positions of the data points in the two-dimensional frequency domain; A is the number of sensors, B is the length of the time series; a,b are the positions of the data in the two-dimensional spatial domain; f(a,b) is the spatial domain data value; i is the imaginary unit, which is used to convert the spatial domain data into frequency domain features.

[0052] Dynamic weighting of each frequency:

[0053] Parameter explanation: w(u,v) is the frequency weight; (u,v) is the frequency domain representation of the real data; (u,v) is the frequency domain representation of the synthesized data; α is the scaling factor (set to 1), which allows the model to focus on the frequency components that are difficult to synthesize.

[0054] Calculate the dynamic spectral loss:

[0055] Parameter explanation: The loss is dynamic spectrum loss; A and B are the number of sensors and the time series length, respectively; this loss is used to supplement the GAN loss and enhance the frequency domain feature matching between the synthetic data and the real data.

[0056] 4. Implementation of the ratio control training mechanism: Define the training step count relationship: Let the discriminator training step count be... The generator training steps are The encoder training steps are All three conditions must be met:

[0057] Parameter explanation: This is a ratio control factor used to coordinate the training pace of the discriminator, generator, and encoder. Stage adaptive adjustment:

[0058] Parameter explanation: T is the number of training iterations; The critical iteration number is set to 20000; DP is the discriminator-dominated stage (J=4), which rapidly improves the discriminator's ability to distinguish between real and synthetic data; SP is the synthesis enhancement stage (K=2), which strengthens the encoder's feature encoding ability and the generator's data synthesis ability.

[0059] 5. Domain Cyclic Collaboration and Data Synthesis: The iterative transfer of spatial, time-frequency, and frequency-domain features is achieved through a domain cyclic process. Data perturbation is introduced in each training round, and a shuffling strategy is adopted to ensure data diversity. A large amount of dissolved gas synthesis data in transformer MEMS oil that is highly consistent with the characteristics of real data is generated to make up for the scarcity of real fault data.

[0060] Step 4: Design a multi-indicator-guided adaptive data fusion strategy and construct a balanced dataset. Innovation: A multi-indicator-guided adaptive data fusion strategy is proposed, which combines statistical distribution indicators (MMD, KLD) and time-frequency domain indicators (MSE, PSD) to dynamically screen high-quality synthetic data, solving the problem of low fusion quality caused by direct mixing or simple filtering in traditional methods, and providing robust data support for diagnostic models; 1. Synthetic Data Preprocessing: Divide the synthetic data generated in step 3 into 1024 data points / samples to obtain the synthetic sample set. Where N is the total number of synthetic samples. It is a single synthetic sample (with the same dimensions as the real sample).

[0061] 2. KLD Guiding Phase (Global Distribution Filtering): (1) Calculate the synthetic sample and the real sample set KL divergence (where M is the total number of real samples):

[0062] Parameter explanation: KLD is the KL divergence, which measures the difference in distribution between the synthetic sample and the real sample set; d is the feature dimension (number of gas components). Number of sensors (Time series length) φ(j) is the feature density estimate for the i-th synthetic sample; φ(j) is the feature density estimate for the real sample set.

[0063] (2) Sort by KLD value in ascending order, remove QxN samples (Q is the elimination rate, set to 0.2), and retain the sample set. Where Nr=N, Sort the KLD index and filter out synthetic samples whose distribution is consistent with the real data.

[0064] 3. MMD guidance phase (statistical distance filtering): (1) Calculate the maximum mean difference (MMD) between the remaining synthetic samples and the real sample set:

[0065] Parameter explanation: MMD is the maximum mean difference, which uses a Gaussian kernel with kernel function k to measure the statistical distance between sample sets; , These are real samples; M represents the total number of real samples. This indicator supplements the KL divergence and filters high-quality synthetic data from the perspective of sample point distance.

[0066] (2) Sort by MMD value in ascending order, remove Q×N samples, and retain the sample set. , where π is the MMD sort index.

[0067] 4. Frequency Domain PSD Booting Stage (Frequency Domain Feature Screening): (1) Calculate the power spectral density difference (PSD-Diff) between the remaining samples and the true samples:

[0068] Parameter explanation: PSD-Diff represents the power spectral density difference, which measures the consistency of frequency domain characteristics; f is the frequency. Let f be the power spectral density of the i-th synthetic sample at frequency f; This represents the power spectral density of the corresponding real sample at frequency f.

[0069] (2) Sort the samples in ascending order by PSD-Diff value, remove the remaining Q×N samples, and retain the sample set. This ensures that the frequency domain characteristics of the synthesized samples match those of the real data.

[0070] 5. Time-Domain MSE Bootstrapping Phase (Time-Domain Detail Filtering): (1) Calculate the mean squared error (MSE) between the remaining samples and the true samples:

[0071] Parameter explanation: MSE is the mean squared error, which measures the consistency of time-domain feature details; d is the feature dimension; Let be the j-th eigenvalue of the i-th synthetic sample; Let be the j-th dimension feature value of the corresponding real sample.

[0072] (2) Sort the samples in ascending order of MSE value, remove the remaining Q×N samples, and retain the sample set. We selected synthetic samples whose temporal details were highly consistent with the real data.

[0073] 6. Data Fusion and Reshuffling: The final high-quality synthetic sample set obtained through screening. Compared with the real sample set Merge, and generate a balanced training dataset through a shuffle operation:

[0074] Parameter explanation: This is a balanced dataset after fusion; Shuffle is a shuffling operation to ensure the randomness of data distribution; this dataset solves the class imbalance problem caused by the scarcity of real fault data, and provides sufficient and high-quality training data for diagnostic models.

[0075] Step 5: Transformer Fault Diagnosis Inference Based on Fusion Dataset 1. Diagnostic Model Training: Input the balanced dataset S* obtained in step 4 into a convolutional neural network (CNN) classifier (network structure: 3 convolutional layers, 3 pooling layers, 2 fully connected layers, convolutional kernel size of 3, activation function of RelU, pooling layer stride of 2), and train it using the Adam optimizer (learning rate 0.0002, batch size 8) to enable the model to learn the multi-domain feature patterns (spatial, time-frequency, and frequency domain) corresponding to different fault types of transformers.

[0076] 2. Fault diagnosis output: The dissolved gas data in the MEMS oil of the transformer to be diagnosed is preprocessed and input into the trained CNN classifier. The classifier outputs the fault type (normal, partial discharge, low energy discharge, high energy discharge, low temperature overheating, medium temperature overheating, high temperature overheating, etc.) based on the learned feature patterns, thus completing the transformer fault diagnosis.

[0077] Figure 5This study compared the effectiveness of a multi-index-guided adaptive data fusion strategy at different screening stages. Four indicators—KL divergence (KLD), maximum mean difference (MMD), mean squared error (MSE), and power spectral density difference (PSD-Diff)—were used to quantify the similarity between synthesized and real data. From the "original synthesized data" to the "final fused data," all indicators showed a continuous downward trend. KLD decreased from 0.092 to 0.025, and MSE decreased from 0.12 to 0.028, indicating that the consistency between the distribution characteristics, temporal details, and frequency domain features of the synthesized data and the real data was significantly improved through multi-stage progressive screening. This result validates the hierarchical screening logic of the fusion strategy, effectively eliminating low-quality synthesized samples, providing robust data support for subsequent fault diagnosis, and addressing the quality control deficiencies of traditional fusion methods.

[0078] Figure 6 The accuracy of the method of this invention in transformer fault diagnosis was compared with that of existing mainstream methods. The results show that the accuracy of methods without data synthesis is only 80.16%, and the accuracy of traditional data synthesis methods (VAE, WGAN-GP) is between 87% and 97%. In contrast, the accuracy of the method of this invention reaches 99.81%, with a standard deviation of only 0.22, significantly outperforming other methods. This advantage stems from the triple technological innovation of this invention: the smoothing regularization variational framework ensures the structural consistency of the synthesized data; the ratio control domain cyclic mechanism strengthens multi-domain feature modeling; and the multi-index adaptive fusion strategy improves data quality. The high accuracy and low standard deviation verify the stability and superiority of this invention in data-scarce scenarios, effectively solving the problem of imbalanced transformer fault data categories and meeting the high-precision diagnostic needs in practical engineering.

[0079] Furthermore, Figure 2 This is a structural block diagram of the transformer oil dissolved gas fault diagnosis device according to an embodiment of the application, such as... Figure 2 As shown, the device includes: The raw gas concentration data acquisition module is used to acquire the raw gas concentration data of various fault characterization gases in transformer oil based on pulsed suspension MEMS sensors. The raw data sample set construction module is used to preprocess the raw gas data and construct the raw data sample set; The synthetic data sample set determination module is used to input the original data sample set into a pre-constructed smooth regularization variational framework to generate a synthetic data sample set. The smooth regularization variational framework is built based on a variational autoencoder and integrates a smooth activation function with a multi-level continuous regularization strategy. The synthetic data optimization sample set determination module is used to input the synthetic data sample set and the original data sample set into the ratio control domain loop mechanism, so that the ratio control domain loop mechanism dynamically adjusts the training ratio in the discriminator-dominated stage and the synthesizer enhancement stage, and after collaboratively extracting spatial domain features, time-frequency domain features and frequency domain features, the synthetic data sample set is optimized based on the extracted multi-domain features to obtain the synthetic data optimization sample set. The balanced training dataset construction module is used to filter the synthetic data optimization sample set based on the multi-index guided adaptive data fusion strategy, and then merge the filtered data sample set with the original data sample set to construct the balanced training dataset. The fault diagnosis model training module is used to train the fault diagnosis model using a balanced training dataset. The fault type output module is used to diagnose the gas data under test based on the fault diagnosis model and output the fault type.

[0080] The application of the relevant modules of the device in this example can be referred to the relevant introduction of the method principle above, and will not be repeated here.

[0081] above Figure 2 The CT image motion artifact removal device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The electronic device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0082] Figure 3 This is a schematic diagram of the structure of an electronic device 300 provided in an embodiment of the present invention. The electronic device 300 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the electronic device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the electronic device 300.

[0083] Electronic device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 5 The illustrated electronic device structure does not constitute a limitation on electronic devices and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0084] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of any of the above-described CT image motion artifact removal apparatus methods.

[0085] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0086] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing dissolved gas faults in transformer oil, characterized in that, The method includes: Obtain raw gas concentration data of various fault characterization gases in transformer oil based on pulsed levitation MEMS sensors; The raw gas data is preprocessed to construct a raw data sample set; The original data sample set is input into a pre-constructed smooth regularized variational framework to generate a synthetic data sample set, wherein the smooth regularized variational framework is constructed based on a variational autoencoder and integrates a smooth activation function and a multi-level continuous regularization strategy. The synthetic data sample set and the original data sample set are input into the ratio control domain loop mechanism so that the ratio control domain loop mechanism dynamically adjusts the training ratio during the discriminator-dominated phase and the synthesizer enhancement phase. After collaboratively extracting spatial domain features, time-frequency domain features and frequency domain features, the synthetic data sample set is optimized based on the extracted multi-domain features to obtain the optimized synthetic data sample set. Based on a multi-index-guided adaptive data fusion strategy, the synthetic data optimization sample set is screened, and the screened data sample set is fused with the original data sample set to construct a balanced training dataset. The fault diagnosis model is trained using the balanced training dataset. The fault diagnosis model is used to diagnose the gas data under test and output the fault type.

2. The method for diagnosing dissolved gas faults in transformer oil according to claim 1, characterized in that, The pulse-suspended MEMS sensor includes: silicon substrate; Microcavities formed on the silicon substrate; A suspension support layer, which is suspended above the microcavity; The outer serpentine heating electrode is arranged in a serpentine configuration in the outer region of the suspension support layer; An inner sensing electrode is arranged coplanarly with the outer serpentine heating electrode in the inner region of the suspension support layer. The inner sensing electrode surrounds a central region, and a sensitive material is disposed in the central region. The outer serpentine heating electrode is used to receive pulse voltage signals to pulse-heat the sensitive material.

3. The method for diagnosing dissolved gas faults in transformer oil according to claim 1, characterized in that, The preprocessing of the raw gas data to construct the raw data sample set includes: The original gas concentration data is normalized to map all concentration values ​​to the [0, 1] interval; The normalized gas concentration time series data is segmented into segments of fixed length. The segmented data is used to construct the original data sample set, wherein the dimensions of the original data sample set are [number of gas components, time series length, number of sensors].

4. The method for diagnosing dissolved gas faults in transformer oil according to claim 1, characterized in that, The smoothing regularization variational framework includes at least three sequentially connected two-dimensional convolutional layers, with a dropout regularization module connected after each convolutional layer. The step of inputting the original data sample set into the pre-constructed smoothing regularization variational framework to generate a synthetic data sample set includes: The original data sample set is input into a pre-constructed smooth regularized variational framework, so that the smooth regularized variational framework, through the encoder of the variational autoencoder, determines an approximate posterior distribution based on the original data sample set and dropout mask, and uses a reparameterization technique to sample latent variables from the approximate posterior distribution, wherein the approximate posterior distribution is an inferred value that replaces the true posterior distribution. Based on the generator, the latent variables are processed by the at least three two-dimensional convolutional layers and a smooth activation function to obtain the synthetic data sample set. The approximate posterior distribution is: ； The latent variables are: ； In the formula, It is a latent variable; x represents the original data sample set; For dropout mask; Indicates a Gaussian distribution; Representation: The approximate posterior distribution of latent variables under the conditions of the original samples and the mask; It is the mean of the approximate posterior distribution; The variance is the approximate posterior distribution. The mean is variance is Gaussian distribution; This represents the posterior distribution mean of the encoder output: The standard deviation of the posterior distribution of the encoder output is: (0, I) represents standard Gaussian noise; I is a unit diagonal matrix with the same dimension as the latent variable z, used to construct the covariance matrix of multidimensional independent standard Gaussian noise.

5. The method for diagnosing dissolved gas faults in transformer oil according to claim 3, characterized in that, The process involves inputting the synthetic data sample set and the original data sample set into a rate control domain recurrent mechanism. This mechanism dynamically adjusts the training ratio during the discriminator-dominated phase and the synthesizer enhancement phase. After collaboratively extracting spatial, time-frequency, and frequency domain features, the synthetic data sample set is optimized based on the extracted multi-domain features, resulting in an optimized synthetic data sample set. The synthetic data sample set and the original data sample set are input into the rate control domain loop mechanism, so that the rate control domain loop mechanism performs the following operations on the synthetic data sample set and the original data sample set respectively: The number of sensors and the length of the time series are independently encoded and feature fused to extract the spatial domain features corresponding to the synthetic data sample set and the original data sample set, respectively. Multi-scale analysis was performed on the gas concentration time series to extract the time-frequency domain features corresponding to the synthetic data sample set and the original data sample set, respectively. Determine the dynamic spectral loss between the synthetic data sample set and the original data sample set to determine the corresponding frequency domain features of the synthetic data sample set and the original data sample set, respectively. Based on the current number of training iterations, a ratio control factor is applied to adjust the training phase. Specifically, when the current number of training iterations is not greater than the critical number of iterations, the ratio control factor is set to the first value, and the training is in the discriminator-dominated phase. When the current number of training iterations is less than the critical number of iterations, the ratio control factor is set to the second value, and the training is in the synthesizer enhancement phase, where the first value is greater than the second value. By alternating the training with the ratio control through the iterative transfer between spatial domain features, time-frequency domain features, and the time-frequency domain features during the iterative process, data perturbation is introduced in each round of training. The synthetic data sample set is then optimized based on the extracted spatial domain features, time-frequency domain features, and frequency domain features, resulting in an optimized synthetic data sample set that is highly consistent with the original data sample set in the multi-domain feature space.

6. The method for diagnosing dissolved gas faults in transformer oil according to claim 5, characterized in that, The step of independently encoding and fusing features for the sensor quantity dimension and the time series length dimension, and extracting the spatial domain features corresponding to the synthetic data sample set and the original data sample set respectively, includes: The number of sensors is independently encoded, and a height orientation feature map is extracted. The length dimension of the time series is independently encoded, and the feature map in the width direction is extracted. The height-direction feature map and the width-direction feature map are concatenated along the channel dimension to obtain a concatenated feature map; The concatenated feature maps are then subjected to convolutional fusion processing to obtain intermediate feature maps; The intermediate feature map is subjected to tensor mapping to obtain the mapped spatial domain features.

7. The method for diagnosing dissolved gas faults in transformer oil according to claim 3, characterized in that, The adaptive data fusion strategy based on multiple indicators filters the synthetic data optimization sample set and merges the filtered data sample set with the original data sample set to construct a balanced training dataset, including: Determine the KL divergence value between each synthetic data sample in the synthetic data sample set and the entire original data sample set; Based on multiple KL divergence values, all synthetic samples are sorted in first order. Based on the sorting results of the first order sorting and combined with the preset first elimination rate, the distribution difference screening is carried out to obtain the first screening sample set after the first stage screening. The first screening sample set includes multiple first screening samples. Determine the maximum mean difference between each selected sample and the original data sample set, whereby the maximum mean difference is used to measure the statistical distance between sample points; Based on multiple maximum mean differences, a second sorting order is performed on multiple primary screening sample sets. Based on the sorting results of the second order sorting and combined with the preset second elimination rate, statistical distance screening is performed to obtain a secondary screening sample set after the second stage screening. The secondary screening sample set includes multiple secondary screening samples. Determine the difference in power spectral density between each secondary screening sample and its corresponding original sample in the original data sample set; Based on the differences in multiple power spectral density values, the multiple secondary screening samples included in the secondary screening sample set are sorted in a third order. Based on the sorting results of the third order, power spectral density difference screening is performed according to the preset third elimination rate to obtain a three-stage screening sample set, which includes multiple three-stage screening samples. Determine the mean square error between each triplet sample and its corresponding original sample in the original data sample set; Based on multiple mean squared error values, the sample set of three screenings, including multiple samples of three screenings, is sorted in a fourth order. Based on the fourth order sorting and the preset fourth elimination rate, the mean square error value is used for screening, and the selected sample set is used as the target screening sample set. Based on the target selection sample set and the original data sample set, the balanced training dataset is determined.

8. A fault diagnosis device for dissolved gases in transformer oil, characterized in that, The device includes: The raw gas concentration data acquisition module is used to acquire the raw gas concentration data of various fault characterization gases in transformer oil based on pulsed suspension MEMS sensors. The raw data sample set construction module is used to preprocess the raw gas data and construct the raw data sample set; A synthetic data sample set determination module is used to input the original data sample set into a pre-constructed smooth regularized variational framework to generate a synthetic data sample set, wherein the smooth regularized variational framework is constructed based on a variational autoencoder and integrates a smooth activation function and a multi-level continuous regularization strategy. The synthetic data optimization sample set determination module is used to input the synthetic data sample set and the original data sample set into the ratio control domain loop mechanism, so that the ratio control domain loop mechanism dynamically adjusts the training ratio in the discriminator-dominated phase and the synthesizer enhancement phase, and after collaboratively extracting spatial domain features, time-frequency domain features and frequency domain features, optimizes the synthetic data sample set based on the extracted multi-domain features to obtain the synthetic data optimization sample set. The balanced training dataset construction module is used to filter the synthetic data optimization sample set based on a multi-index-guided adaptive data fusion strategy, and then merge the filtered data sample set with the original data sample set to construct a balanced training dataset. The fault diagnosis model training module is used to train the fault diagnosis model using the balanced training dataset. The fault type output module is used to diagnose the gas data under test based on the fault diagnosis model and output the fault type.

9. An electronic device, characterized in that, The electronic device includes a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform the various steps of the method for diagnosing dissolved gas faults in transformer oil as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the method for diagnosing dissolved gas faults in transformer oil as described in any one of claims 1-7.

Images