Oil-immersed current transformer data enhancement method based on transfer learning

Abstract

The invention discloses an oil-immersed current transformer data enhancement method based on transfer learning. The method comprises the following steps: constructing a data preprocessing model, and carrying out standardization processing on oil chromatographic data of a transformer and a current transformer by utilizing a quantile transformation method; constructing an encoder to map original signal data into a potential variable sample space, and proposing a noise enhancement mechanism with multi-distribution fusion; the generative network performs random sampling from known probability distribution to obtain hidden variables, and calculates a preliminary enhanced signal obeying Gaussian distribution; the generative adversarial module adopts a generator and a discriminator, the generator performs incremental generation on the preliminary enhanced signal and outputs a sample conforming to the characteristics of the current transformer, and the discriminator distinguishes the difference between a real current transformer sample and a generated sample according to physical knowledge constraints in the oil chromatography field; and finally, carrying out reverse quantization on the generated sample, and outputting high-quality oil chromatographic data of the current transformer. According to the invention, the problem of rare data of the current transformer is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of current transformer data expansion, and more particularly to a method for oil-immersed current transformer data enhancement based on transfer learning. Background Art

[0002] Oil-immersed current transformers are indispensable equipment for relay protection, automatic control, metering, and measurement in power systems. The safety and stability of their operation directly affect the power supply reliability of the power system.

[0003] Dissolved gas analysis (DGA) is one of the most common and effective methods for diagnosing latent faults within these devices, as different fault types (such as discharge and overheating) produce specific characteristic gas combinations and ratios in the insulating oil. However, traditional DGA diagnostic methods (such as the three-ratio method) have limitations, such as fuzzy coding boundaries and low diagnostic accuracy.

[0004] In recent years, data-driven diagnostic methods based on machine learning and deep learning have been widely studied to improve the automation and accuracy of diagnosis. Deep learning models (such as autoencoders and convolutional networks) can automatically learn complex fault characteristics from data, avoiding the tedious manual feature extraction.

[0005] However, in real-world applications, current transformer failures are inherently low-probability events. This leads to the typical problems of small sample sizes and class imbalance, severely restricting the performance of traditional data-driven approaches. This can easily lead to model overfitting or poor identification of minority faults. Therefore, expanding the scarce current transformer fault data presents a significant challenge. Summary of the Invention

[0006] The purpose of this method is to propose a data enhancement method for oil-immersed current transformer based on transfer learning to solve the above technical problems.

[0007] To achieve the above objectives, the oil-immersed current transformer data enhancement method based on transfer learning is as follows:

[0008] Step 1: Build a data preprocessing module, use the quantile transformation method to standardize the oil chromatogram data of transformers and current transformers, and add gas ratio features;

[0009] Step 2: Construct a variational autoencoder adversarial network, which includes encoder, multi-distribution noise enhancement, decoder, Wasserstein discriminator and other modules;

[0010] Step 3: Build a multi-stage training module, divide the model training into multiple stages, optimize different loss functions in different training stages, and balance the impact of each loss through a dynamic weight adjustment mechanism;

[0011] Step 4: Construct a physical constraint loss design module, introducing oil chromatography domain knowledge and related physical laws as constraints into the training process;

[0012] Step 5: Construct a sample generation module to sample type values based on the fault type distribution of the current transformer, then generate latent vectors from the standard normal distribution, apply multi-distribution noise enhancement, and finally generate samples through the decoder condition.

[0013] According to the above scheme, in step 1, a data preprocessing module is constructed, and the quantile transformation method is used to standardize the oil chromatogram data of the transformer and current transformer, and the gas ratio feature step is added, including:

[0014] The original feature set X is established based on the oil chromatogram data of transformers and current transformers, and the extended feature set X is constructed by combining the gas ratio feature. expanded ;

[0015] According to the above scheme, in step 2, the encoder maps the standardized raw oil chromatogram data to a latent space, generates a mean and logarithmic variance, and samples a latent vector using a reparameterization technique. The multi-distribution noise enhancement module adds noise of multiple distributions to the sampled latent vector to generate an enhanced latent vector. The decoder combines the enhanced latent vector with the fault type condition to generate an oil chromatogram data sample in the target domain. The Wasserstein discriminator distinguishes between real samples and generated samples and improves training stability through a gradient penalty term.

[0016] According to the above scheme, in step 3, the multi-stage training module, the first stage is the autoencoder pre-training, which optimizes the reconstruction loss and KL divergence loss. The specific formula is as follows:

[0017]

[0018] Among them, n is the number of samples, d is the feature dimension, and x i is the original sample, is the reconstruction sample;

[0019]

[0020] Where μ and σ are the mean and standard deviation of each feature in the latent space, respectively;

[0021] The second stage is the discriminator pre-training, which optimizes the Wasserstein distance between the generated samples and the real samples. The specific formula is as follows:

[0022]

[0023] in, is the score of the discriminator on the generated sample, is the score of the discriminator on the real sample, is the interpolation point between the generated sample and the real sample;

[0024] The third stage is the joint training of VAE and GAN to optimize various losses during the training process;

[0025] According to the above solution, in step 4, the oil chromatography field knowledge and related physical laws are introduced to ensure that the generated gas ratio conforms to the physical laws of the specific fault type. Different fault types correspond to different gas concentration ratio ranges. The specific constraint formula is:

[0026]

[0027] Among them, r1(i), r2(i), and r3(i) are the gas ratios of the i-th sample, and is the lower limit and upper limit of the ratio of the i-th item under category t. If an item has no upper or lower limit, it can be omitted;

[0028] At the same time, it is also necessary to capture the correlation between gas concentrations under specific fault types. Certain fault types may lead to stronger positive correlations between certain gases. The specific formula is as follows:

[0029]

[0030] Among them, C i is the total number of samples i, (j,k)∈S i is the relevant gas pair in sample i, x ij and x ik are the concentrations of gas j and gas k respectively;

[0031] According to the above scheme, in step 4, after the sample is generated, it needs to go through the physical constraint filtering link, calculate the physical constraint loss of each sample, and divide the samples into good samples (loss below the 70% percentile), medium samples (loss between 70%-95% percentile) and special value samples (loss above the 95% percentile). The retention strategy is to retain all good samples and randomly select 5% of special value samples to balance sample quality and diversity.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] By utilizing the conditional VAE-WGAN-GP framework combined with physical constraints and complex noise enhancement technology, it is possible to generate a large amount of high-quality, diverse and physically consistent synthetic data based on a small amount of real current transformer DGA data, providing effective data support for solving small sample fault diagnosis problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a workflow diagram of a data enhancement method for an oil-immersed current transformer based on transfer learning in the present invention. DETAILED DESCRIPTION

[0035] Example 1

[0036] The technical solution of the present invention is clearly and completely described below in conjunction with the accompanying drawings and specific embodiments:

[0037] The present invention provides a data enhancement method for an oil-immersed current transformer based on transfer learning, comprising the following steps:

[0038] Step 1: Build a data preprocessing module, use the quantile transformation method to standardize the transformer and current transformer oil chromatogram data, and add gas ratio features:

[0039] S101: Load target data, use the quantile transformation method to fit and transform the transformer and current transformer data, and convert the data into a normal distribution;

[0040] S102: Generate gas ratios C2H2 / C2H4, CH4 / H2, and C2H4 / C2H6 through DGA feature engineering, and include these ratio features in the normalization and model processing;

[0041] Step 2: Construct a variational autoencoder adversarial network, which includes encoder, multi-distribution noise enhancement, decoder, Wasserstein discriminator and other modules:

[0042] S201: Build an encoder using a multi-layer perceptron topology whose core components include fully connected layers, LeakyReLU activation units, one-dimensional batch normalization, and Dropout regularization.

[0043] S202: Construct a decoder, which also uses a multi-layer perceptron structure and includes a category embedding layer. The fault type label is used as a conditional input to achieve conditional generation.

[0044] S203: Construct a discriminator, also using a multi-layer perceptron structure;

[0045] S204: Build a data augmentation module, which is applied to the real data batches input when training the discriminator to increase the robustness of the model. The included techniques include: adding various types of noise, feature scaling, feature swapping, ratio perturbation, feature correlation adjustment, etc.

[0046] Step 3: Build a multi-stage training module, divide the model training into multiple stages, optimize different loss functions in different training stages, and balance the impact of each loss through a dynamic weight adjustment mechanism:

[0047] S301: The first stage is the autoencoder pre-training, which optimizes the reconstruction loss and KL divergence loss. The specific formula is as follows:

[0048]

[0049] Among them, n is the number of samples, d is the feature dimension, and x i is the original sample, is the reconstruction sample;

[0050]

[0051] Where μ and σ are the mean and standard deviation of each feature in the latent space, respectively;

[0052] S302: The second stage is the discriminator pre-training, which optimizes the Wasserstein distance between the generated samples and the real samples. The specific formula is as follows:

[0053]

[0054] in, is the score of the discriminator on the generated sample, is the score of the discriminator on the real sample, is the interpolation point between the generated sample and the real sample;

[0055] S303: The third stage is the joint training of VAE and GAN to optimize various losses during the training process;

[0056] Step 4: Build a physical constraint loss design module, and introduce oil chromatography domain knowledge and related physical laws as constraints into the training process:

[0057] S401: Introduce oil chromatography domain knowledge and related physical laws, that is, ensure that the generated gas ratio conforms to the physical laws of the specific fault type. Different fault types correspond to different gas concentration ratio ranges. The specific constraint formula is:

[0058]

[0059] Among them, r1=C2H2 / C2H4, r1=CH4 / H2, r1=C2H4 / C2H6, and is the lower limit and upper limit of the ratio of the i-th item under category t. If an item has no upper or lower limit, it can be omitted;

[0060] S402: It is necessary to capture the correlation between gas concentrations under specific fault types. Certain fault types may lead to stronger positive correlations between certain gases. The specific formula is as follows:

[0061]

[0062] Among them, C i is the total number of samples i, (j,k)∈S i is the relevant gas pair in sample i, x ij and x ik are the concentrations of gas j and gas k respectively;

[0063] Step 5: Construct a sample generation module to sample type values based on the fault type distribution of the current transformer, then generate latent vectors from the standard normal distribution, apply multi-distribution noise enhancement, and finally generate samples through the decoder condition:

[0064] S501: According to the actual probability of the fault type appearing in the target domain data, randomly extract the required number of fault type labels. If the probability of type k in the target domain is P(type=k), sample N times to obtain the type sequence (t1, t2, ..., t N );

[0065] S502: Randomly sample from a standard multivariate Gaussian distribution to generate a low-dimensional latent vector set. The specific formula is as follows:

[0066] z init ~N(0,I)

[0067] Among them, z init is the initial latent vector, N represents the Gaussian distribution, 0 is the zero-mean vector, and I is the unit covariance matrix;

[0068] S503: Apply a complex random perturbation process to the initial latent vector to significantly increase the diversity of generated samples. Noise vectors are selected and sampled from multiple different probability distributions. The intensity of the noise is adaptively adjusted based on the statistical properties of the initial latent vector. Strong perturbations are then added to 5% of the samples.

[0069] S504: Input the target category label t into a trainable embedding layer to obtain the corresponding conditional embedding vector Embt = EmbeddingLayer(t); enhanced Concatenate with the conditional embedding vector Embt to form a fusion vector z fused =Concat(z fused ,Embt); fuse the vector z fusedInput the decoder neural network, map it back to the high-dimensional feature space through a series of nonlinear mappings, and output the final synthetic sample X generated =Decoder(z fused ).

[0070] Example 2

[0071] We collected 399 sets of transformer oil chromatogram data and 39 sets of current transformer oil chromatogram data. The data was expanded to 400 sets through the above steps. We used various statistical distance metrics to evaluate the generated samples. The values of each metric are shown in Table 1.

[0072] Table 1 Statistical distance metrics

[0073] Feature Name Wasserstein distance Energy distance KS distance KS p-value <![CDATA[H2]]> 35.8 1.85 0.065 0.78 <![CDATA[CH4]]> 15.2 0.95 0.058 0.85 <![CDATA[C2H2]]> 8.5 0.42 0.041 0.97 <![CDATA[C2H4]]> 22.1 1.15 0.061 0.81 <![CDATA[C2H6]]> 11.6 0.68 0.053 0.90

[0074] Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

Claims

1. A data enhancement method for oil-immersed current transformer based on transfer learning, characterized by: Step 1: Build a data preprocessing module, use the quantile transformation method to standardize the oil chromatogram data of transformers and current transformers, and add gas ratio features; Step 2: Construct a variational autoencoder adversarial network, which includes encoder, multi-distribution noise enhancement, decoder, Wasserstein discriminator and other modules; Step 3: Build a multi-stage training module, divide the model training into multiple stages, optimize different loss functions in different training stages, and balance the impact of each loss through a dynamic weight adjustment mechanism; Step 4: Construct a physical constraint loss design module, introducing oil chromatography domain knowledge and related physical laws as constraints into the training process; Step 5: Construct a sample generation module to sample type values based on the fault type distribution of the current transformer, then generate latent vectors from the standard normal distribution, apply multi-distribution noise enhancement, and finally generate samples through the decoder condition.

2. The data enhancement method for oil-immersed current transformer based on transfer learning according to claim 1, characterized in that: In step 1, the data is preprocessed to establish the original feature set X based on the transformer and current transformer oil chromatogram data, and the extended feature set X is constructed by combining the gas ratio feature. expanded , applies quantile transformation normalization to continuous features.

3. The data enhancement method for oil-immersed current transformer based on transfer learning according to claim 1, characterized in that: The encoder described in step 2 maps the standardized raw oil chromatogram data to the latent space, generates the mean and logarithmic variance, and samples the latent vector through the reparameterization technique; the multi-distribution noise enhancement module adds noise of multiple distributions to the sampled latent vector to produce an enhanced latent vector; the decoder combines the enhanced latent vector and the fault type condition to generate the oil chromatogram data sample of the target domain; the Wasserstein discriminator distinguishes between real samples and generated samples, and improves the training stability through the gradient penalty term.

4. The method for data enhancement of an oil-immersed current transformer based on transfer learning according to claim 1, characterized in that: In the multi-stage training module described in step 3, the first stage is the autoencoder pre-training, which optimizes the reconstruction loss and KL divergence loss. The specific formula is as follows: Among them, n is the number of samples, d is the feature dimension, and x i is the original sample, is the reconstruction sample; Where μ and σ are the mean and standard deviation of each feature in the latent space, respectively; The second stage is the discriminator pre-training, which optimizes the Wasserstein distance between the generated samples and the real samples. The specific formula is as follows: in, is the score of the discriminator on the generated sample, is the score of the discriminator on the real sample, is the interpolation point between the generated sample and the real sample; The third stage is the joint training of VAE and GAN to optimize various losses during the training process.

5. The method for data enhancement of an oil-immersed current transformer based on transfer learning according to claim 1, characterized in that: In step 4, we introduce oil chromatography domain knowledge and related physical laws to ensure that the generated gas ratio conforms to the physical laws of the specific fault type. Different fault types correspond to different gas concentration ratio ranges. The specific constraint formula is: Among them, r1(i), r2(i), and r3(i) are the gas ratios of the i-th sample, and is the lower limit and upper limit of the ratio of the i-th item under category t. If an item has no upper or lower limit, it can be omitted; At the same time, it is also necessary to capture the correlation between gas concentrations under specific fault types. Certain fault types may lead to stronger positive correlations between certain gases. The specific formula is as follows: Among them, C i is the total number of samples i, (j,k)∈S i is the relevant gas pair in sample i, x ij and x ik are the concentrations of gas j and gas k, respectively.

6. The method for data enhancement of an oil-immersed current transformer based on transfer learning according to claim 1, characterized in that: The sample generation module described in step 5 needs to go through the physical constraint filtering stage after sample generation, calculate the physical constraint loss of each sample, and classify the samples into good samples (loss below the 70th percentile), medium samples (loss between 70% and 95% percentiles), and outlier samples (loss above the 95th percentile). The retention strategy is to retain all good samples and randomly select 5% of outlier samples to balance sample quality and diversity.

Images