An artificial intelligence oil chromatogram test fault detection method fusing uncertainty perception
By integrating an AI-based oil chromatography fault detection method that incorporates uncertainty perception, the shortcomings of traditional oil chromatography analysis methods in transformer fault diagnosis have been addressed. This method achieves efficient and accurate fault detection, thereby improving the scientific rigor and economic benefits of transformer condition monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANGGANG POWER SUPPLY COMPANY HUBEI ELECTRIC POWER
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-29
AI Technical Summary
Traditional oil chromatography analysis methods are insufficient in accuracy and efficiency in transformer fault diagnosis, making it difficult to detect early faults in a timely manner, leading to equipment damage and economic losses.
An artificial intelligence-based oil chromatography test fault detection method with uncertainty perception is adopted. By noise suppression and BIM grayscale image mapping processing, a Bayesian CNN model is constructed. Combined with real-time data deviation judgment and confidence level decision, the fault probability and uncertainty measurement are realized. The model is optimized by monthly dataset updates and small sample fine-tuning.
It has achieved automated analysis of oil chromatography data, shortened the average diagnosis time to 6.8 seconds, and achieved a diagnosis accuracy rate of 96%, reducing false positives and false negatives, improving the scientificity and accuracy of transformer condition monitoring, and saving manual analysis time and economic losses.
Smart Images
Figure CN122117146A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to electrical equipment condition monitoring, specifically to an artificial intelligence-based oil chromatography test fault detection method that integrates uncertainty perception. Background Technology
[0002] Transformers are core equipment in power systems, widely used for power conversion and transmission. Their reliable operation directly determines the stability of the power grid and the continuity of power supply. Under long-term high-load operation, transformers are susceptible to typical problems such as partial discharge, winding short circuits, core overheating, and tap changer failures due to factors such as insulation aging, electrical stress impact, and abnormal heat dissipation. If these faults are not detected and handled in a timely manner, they may lead to serious equipment damage or even large-scale outages, causing huge economic losses and safety hazards. Therefore, establishing a precise and efficient transformer fault diagnosis mechanism has become a core task of power system operation and maintenance management.
[0003] Oil chromatography, as one of the key technologies for transformer fault diagnosis, can effectively reflect the internal condition of a transformer by analyzing the dissolved and overflowing gas components in transformer oil. It plays an irreplaceable role, especially in detecting early signs of faults, and is one of the important means of current transformer condition monitoring and fault diagnosis.
[0004] However, with the expansion of the power grid and the increasing aging of equipment, traditional oil chromatography analysis methods have gradually revealed their bottlenecks, with room for improvement in accuracy, efficiency, and intelligence. The inventors' team investigated the accuracy rate of transformer oil chromatography tests from 2020 to 2024 and estimated the potential losses incurred in the Huanggang area. The traditional diagnostic method had the highest accuracy rate at 75.4%, indicating significant room for improvement. Summary of the Invention
[0005] The purpose of this invention is to provide an artificial intelligence-based oil chromatography test fault detection method that integrates uncertainty perception. Based on DGA data of seven key dissolved gases from oil-immersed transformers, noise suppression and BIM grayscale image mapping are applied. A balanced and enhanced sample set is constructed through transfer sample fusion, CVAE-GAN generation, and meta-learning supplementation. Relying on a CNN model with a Bayesian fully connected layer, fault probability and uncertainty measurement are achieved through MAML pre-training and DS evidence fusion. The diagnostic results are output by combining real-time data deviation judgment and confidence level decision-making. The model is dynamically optimized through monthly dataset updates and small-sample fine-tuning to accurately detect faults in transformer oil chromatography tests.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] An artificial intelligence-based oil chromatography test fault detection method incorporating uncertainty perception, characterized by comprising the following steps:
[0008] S1: The contents of seven key dissolved gases in the DGA data of oil-immersed transformers are selected as the original data features. The 3σ criterion is used to identify high-noise samples, and noise is eliminated by db4 wavelet basis decomposition. The processed data is then subjected to binary image mapping to obtain the BIM grayscale image.
[0009] S2: Collect DGA data of transformers of the same voltage level and model to form a migration sample set. Merge the target sample set corresponding to the BIM grayscale image obtained in S1 with the migration sample set to form an enhanced sample set. Generate minority class samples that fit the actual fault mechanism through CVAE-GAN. Use meta-learning algorithm to supplement samples for niche faults to ensure that the number of samples of each class is ≥200.
[0010] S3: Construct a Bayesian CNN classification model with a Bayesian fully connected layer. Split the augmented sample set into a meta-training set and a target training set as model inputs. Use the MAML algorithm to perform meta-pre-training on the Bayesian CNN classification model. Through multi-model evidence collection and DS evidence theory fusion, output the fault probability distribution and uncertainty measure, and use multiple indicators to evaluate the model performance.
[0011] S4: Real-time acquisition of DGA data, real-time calculation of its deviation from historical normal data, data with deviations exceeding the set value are preprocessed using the noise suppression method of S1 to obtain BIM grayscale images, and the BIM images are input into the classification model constructed by S3 to obtain fault diagnosis results and uncertainty parameters. Based on confidence level decision-making, the dataset is updated monthly and the classification model of S3 is iteratively optimized.
[0012] Step S1 selects the content of seven key dissolved gases in the DGA data of oil-immersed transformers as the original data features, uses the 3σ criterion to identify high-noise samples, removes the data of high-noise samples, and eliminates noise through db4 wavelet basis decomposition. Specifically, this includes: selecting the content of seven key dissolved gases in the DGA data of oil-immersed transformers as the original data features, including H2, CH4, C2H2, C2H4, C2H6, CO, and CO2; using the 3σ criterion to remove high-noise samples; performing 3-level decomposition using the db4 wavelet basis; and calculating the noise threshold using the Birgé-Massart strategy.
[0013] ;
[0014] In the formula, This is the noise threshold; data exceeding this threshold will be excluded. The standard deviation of noise in the original DGA data; This represents the total number of samples in the original DGA data.
[0015] Data exceeding the noise threshold is excluded, and the denoised data is output. The relative deviations of the denoised samples from the reference mean of similar fault samples are calculated.
[0016] ;
[0017] In the formula, The relative deviation between the denoised sample and the reference mean; This is a single sample data after denoising using db4 wavelet basis decomposition; The reference mean of similar fault samples is obtained by statistical analysis of historical similar fault samples of transformers of the same voltage level and model; if the deviation is >10%, noise suppression is re-executed; if the deviation is ≤10%, the mapping process is entered.
[0018] Step S1 involves binary image mapping of the processed data to obtain a BIM grayscale image. Specifically, the content of the seven gases is divided into eight intervals according to standard thresholds, with each interval corresponding to an 8-bit binary code. The binary sequences of the first five gases are taken as core features. The five binary sequences of each sample are arranged according to the rule of two sequences per row. The remaining empty spaces are filled with a fixed pixel value M=0, ultimately forming a 32×32 pixel two-dimensional grayscale image, thus completing the image dataset. Finally, the deviation normalization method is used to map all pixel values to the [0,1] interval, with the following formula:
[0019] ;
[0020] In the formula, The standardized pixel values are mapped to the [0,1] interval; The original pixel values are obtained after mapping the binary image, where , These are the row and column indices of the pixel in the image, respectively; For all original pixel values The minimum value in; For all original pixel values The maximum value in the range is used to obtain the final BIM grayscale image, which is then used to form the target sample set. .
[0021] Step S2 involves collecting DGA data of transformers of the same voltage level and model to form a transfer sample set. The target sample set corresponding to the BIM grayscale image obtained in S1 is then merged with the transfer sample set to form an enhanced sample set. Specifically, this includes: collecting DGA data of transformers of the same voltage level and model and processing it through step S1 to form a BIM grayscale image, which is then used to form the transfer sample set. Compare it with the target sample set The formula for fusing samples into an enhanced sample set is as follows:
[0022] ;
[0023] In the formula, To enhance the sample set; The weight coefficients for the target sample set. ; The transfer sample set is adjusted by the MMD domain adaptive algorithm. The MMD calculation formula is:
[0024] ;
[0025] In the formula, This represents the maximum mean difference. The sample set is composed of BIM grayscale images of transformers of the same voltage level and model after being processed by the S1 process. The target sample set consists of BIM grayscale images of the transformer to be tested after the DGA data has been processed by the S1 process. For the transfer sample set The number of samples; For the target sample set The number of samples; For the transfer sample set A single sample in; For the target sample set A single sample in; The feature mapping function maps samples to the reproducing kernel Hilbert space. High-dimensional feature vectors in; To regenerate the kernel Hilbert space, high-dimensional mapping and distance calculation of sample distribution are achieved; for In space Norm squared, quantifies the difference between high-dimensional eigenvectors.
[0026] Step S2 generates minority class samples that closely match the actual fault mechanisms using CVAE-GAN. For less common faults, a meta-learning algorithm is used to supplement the samples, ensuring that the number of samples in each class is ≥200. Specifically, the CVAE-GAN consists of a generator G and a discriminator D. The generator input includes random noise. and transfer sample feature vector Output generated samples ;
[0027] Its loss function is designed as follows:
[0028] Generator loss: ;
[0029] The quality control of generated samples was verified using the structural similarity index SSIM.
[0030]
[0031] In the formula, To generate the mean of the sample and the real sample, Standard deviation For covariance, It is a constant, and SSIM ≥ 0.88 is required.
[0032] Minority class samples that closely reflect actual fault mechanisms are generated using CVAE-GAN. For niche faults with fewer than 50 samples, supplementary samples are generated using the MAML algorithm. When generating supplementary samples using the MAML algorithm, a meta-learning task set is first constructed based on the small samples of niche faults in the transfer sample set and the target sample set. For each meta-task, 5-8 samples of the niche fault are randomly selected as the support set and 3-5 samples as the query set, ensuring that the task set covers the core data distribution of the niche faults. The optimization objective is to fine-tune the model parameters using support set data through 2-3 rounds of inner loops, and then update the initial parameters based on the expected loss of the query set. After 5-10 rounds of meta-training, the optimal initial parameters are obtained. Subsequently, these optimal parameters are substituted into the model and combined with the feature vectors of the transfer samples. With random noise Supplementary samples that closely match the actual fault mechanism are generated, with the number controlled between 100-150 sets, ultimately ensuring that the total number of samples for this niche fault is ≥200 sets. Furthermore, the generated samples must pass a relative deviation verification with the reference mean of similar faults to ensure consistency with real fault data. The optimization objective of the MAML algorithm is:
[0033] ;
[0034] In the formula, The optimal initial parameters are obtained after optimizing the MAML algorithm. To find the parameters that minimize the objective function ; This is the initial parameter set for the Bayesian CNN model; Let be the expectation operator, representing the expectation of the task set. All tasks Expectations; The meta-learning task set consists of a small sample dataset of minority faults. for A single meta-learning task in; For the task The corresponding loss function; For the After the cycle update, the model in the task Parameterized prediction function on; For the model in the task Up past Parameters after in-round cyclic training;
[0035] 100-150 supplementary samples are generated through 5-10 rounds of meta-training to ensure that the number of samples of each type is ≥200.
[0036] Step S3 constructs a Bayesian CNN classification model with a Bayesian fully connected layer. The enhanced sample set is split into a meta-training set and a target training set as model input. The MAML algorithm is used to perform meta-pre-training on the Bayesian CNN classification model. Specifically, the Bayesian CNN classification model is based on the original BIM-CNN and optimized by replacing the fully connected layer with a Bayesian fully connected layer. The input is a 32×32×1 BIM grayscale image, and the output is the probability distribution and uncertainty measure of 7 types of faults.
[0037] Its model architecture is as follows:
[0038] The input is a 32×32×1 BIM grayscale image, and the network structure includes:
[0039] Convolutional layer 1: 16 5×5 convolutional kernels, ReLU activation function;
[0040] Pooling layer 1: 2×2 max pooling;
[0041] Convolutional layer 2: 32 5×5 convolutional kernels, ReLU activation function;
[0042] Pooling layer 2: 2×2 max pooling;
[0043] Fully connected layer: Replaced with a Bayesian fully connected layer, whose weights follow a Gaussian distribution. ;
[0044] Output layer: Probability distribution of 7 types of faults With uncertainty measurement.
[0045] Training optimization:
[0046] Optimizer: Adam optimizer, learning rate ;
[0047] Loss function: ,in:
[0048] Cross-entropy loss: In the formula This is a real label;
[0049] KL divergence loss:
[0050] ;
[0051] Weighting coefficient To balance classification loss with uncertainty regularization.
[0052] Training strategy: Train for 150 rounds, using early stopping. Stop if the validation set loss does not decrease for 10 consecutive rounds to avoid overfitting.
[0053] The augmented sample set is split into a meta-training set and a target training set as model inputs. The splitting rule is as follows: the augmented sample set is split into the meta-training set and the target training set in a 6:4 ratio; the target training set is then split into the training set, test set, and validation set in a 7:2:1 ratio. The Bayesian CNN classification model is then pre-trained using the MAML algorithm with an inner loop learning rate of... , outer loop learning rate Specifically, based on the original scheme, an enhanced sample set is split into a 6:4 ratio to construct a task set containing 500-800 meta-tasks. For each meta-task, one type of fault is randomly selected, and 20-30 samples are used as the support set and 10-15 samples as the query set. Training employs a dual-loop strategy: the inner loop uses the SGD optimizer with a learning rate α=1e-3, fine-tuning the support set data for each meta-task for 3 rounds, updating the specific distribution parameters for that task; the outer loop uses the Adam optimizer with a learning rate β=5e-4, optimizing the initial model based on the expected value of the classification cross-entropy loss + KL divergence regularization loss from the query set. The distribution parameters are set, and the total number of training rounds is set to 200-300. The termination condition is that the average F1 score of the meta-validation set is ≥0.85 or the loss decrease is <1e-5. During training, the variance parameter is constrained by the Softplus activation function to avoid losing the ability to measure uncertainty. 32 meta-tasks are used for batch processing to improve efficiency. The gradient norm is clipped to 5.0 to prevent gradient explosion. The target validation set is used for evaluation every 20 rounds. The classification accuracy is required to be ≥85% and the uncertainty calibration error is required to be ≤8%. If the target is not met, the outer loop learning rate is reduced to 0.5 times the original. The final output is the optimized initial distribution parameters.
[0054] Step S3 integrates multi-model evidence collection with DS evidence theory to output fault probability distributions and uncertainty measures, and evaluates model performance using multiple metrics. Specifically, it involves training three Bayesian CNN classification models with different initial weights, using random seeds of 123, 456, and 789 respectively, and initializing the weights using a He normal distribution to obtain three sets of fault probability distributions. Transform the probability distribution into an evidence mass function Where A is a subset of faults; the DS synthesis rule is used to fuse multi-model evidence, and for any subset of faults A, its fusion probability is... The calculation formula is as follows: ;
[0055] In the formula, For the fault subset The mass function value of the fusion evidence, i.e., the fusion probability, is denoted as... , corresponding to the The fusion probability of fault classes; A subset of faults; The conflict coefficient; , , These are the fault subsets output by three Bayesian CNN models with different initialized weights. , , To represent the fault subsets of the three models respectively , , The initial evidence mass function value;
[0056] Among them, the conflict coefficient The formula is: ;
[0057] In the formula, The conflict coefficient has a value range of [0,1]. The closer the value is to 1, the more severe the conflict of evidence among the three models. , , To represent the fault subsets of the three models respectively , , The initial evidence mass function value;
[0058] Uncertainty is measured using confidence entropy:
[0059] ;
[0060] In the formula, Confidence entropy is used to quantify the uncertainty of fault diagnosis results; This is an index for fault types, corresponding to 7 types of faults; For the first The fusion probability of fault classes; This refers to logarithmic operations in information theory.
[0061] Step S3 evaluates the model performance using multiple metrics, specifically accuracy, recall, and F1 score. Each metric must reach 90%. If the requirement is not met, the Bayesian CNN classification model is retrained.
[0062] Step S4 involves real-time acquisition of DGA data and real-time calculation of its deviation from historical normal data. For data with deviations exceeding a set value, the noise suppression method from S1 is used for data preprocessing to obtain a BIM grayscale image. This BIM image is then input into the classification model constructed in S3 to obtain fault diagnosis results and uncertainty parameters. The confidence-based grading decision-making specifically includes: real-time acquisition of DGA data and calculation of the deviation from historical normal data.
[0063] ;
[0064] In the formula, The average relative deviation between real-time DGA data and historical normal data; This is an index of dissolved gas types, corresponding to 7 key gases; For the first Real-time collection of the content of the gas; For the first Historical average normal concentration of the gas;
[0065] If Δ>15%, the noise suppression method of S1 is used for data preprocessing to obtain a BIM grayscale image. This BIM image is then input into the classification model constructed in S3 to obtain the fault diagnosis results and uncertainty parameters.
[0066] Hierarchical decision-making based on confidence level:
[0067] For high confidence: fusion probability and It directly outputs the fault diagnosis results;
[0068] For medium confidence level: and Output the diagnostic results and indicate recommendations for review;
[0069] For low confidence: and Marked as pending confirmation, triggering the manual review process.
[0070] When adding a transformer or a less common fault in step S4, fine-tune the model using 5-10 sets of samples:
[0071] ;
[0072] In the formula, The final parameters after model fine-tuning are for adding a new transformer or a minor fault. These are the pre-trained parameters before model fine-tuning; To fine-tune the learning rate; For parameters The gradient operator; The loss function corresponding to the newly added samples; For parameters The model prediction function at that time;
[0073] After fine-tuning the number of rounds by 5-10 rounds and evaluating it with 2-3 small sample validation sets, if the accuracy is ≥88%, it can be put into use.
[0074] This AI-based oil chromatography fault detection method, which integrates uncertainty perception, revolves around the accurate processing, sample enhancement, intelligent modeling, and dynamic optimization of DGA data from oil-immersed transformers. Its working mechanism is as follows:
[0075] Using the contents of seven key dissolved gases, including H2 and CH4, in the DGA data of oil-immersed transformers as the core input, high-noise samples are first identified and removed using the 3σ criterion. Then, deep noise reduction is achieved by combining db4 wavelet basis 3-level decomposition with Birgé-Massart noise threshold calculation. The reliability of the data is ensured by relative deviation verification. Subsequently, the contents of the seven gases are divided into eight intervals according to standard thresholds and encoded into 8-bit binary sequences. The core feature sequences of the first five gases are selected and arranged into a 32×32 pixel two-dimensional image. After deviation standardization, the image is mapped to the [0,1] interval to form a BIM grayscale image target sample set, completing the transformation from gas values to visual features, which is suitable for the input requirements of subsequent deep learning models.
[0076] To address the scarcity of samples from niche faults, DGA data from transformers of the same voltage level and model were collected and processed using the aforementioned process to form a transfer sample set. The distributional difference between the transfer samples and the target samples was reduced using the MMD domain adaptive algorithm, and the samples were then fused according to weights to form an enhanced sample set. For minority class samples, CVAE-GAN was used to generate supplementary samples that fit the actual fault mechanism. For niche faults with fewer than 50 samples, the MAML meta-learning algorithm was used to generate 100-150 samples through 5-10 rounds of meta-training. Ultimately, this ensured that the number of samples for each class was ≥200, providing sufficient and balanced data support for model training.
[0077] A Bayesian CNN classification model with a fully connected Bayesian layer was constructed. The augmented sample set was split into a meta-training set and a target training set in a 6:4 ratio. The target training set was then split into training / test / validation sets in a 7:2:1 ratio. The MAML algorithm was used for meta-pre-training to give the model the ability to quickly adapt to niche faults. By training three models with different initial weights, the output fault probability distribution was transformed into an evidence mass function. The results of multiple models were fused using DS evidence theory. At the same time, the uncertainty was quantified by confidence entropy, forming a dual output of "fault probability distribution + uncertainty measure". The model performance was then verified with a classification accuracy of ≥90% and an uncertainty validity calibration error of ≤5%.
[0078] During the real-time fault detection phase, DGA data is collected in real time and the average relative deviation from historical normal data is calculated. When the deviation is greater than 15%, the aforementioned noise suppression and image mapping process is used to generate a BIM grayscale image, which is then input into the trained Bayesian CNN model. Based on the fusion probability and confidence entropy, a hierarchical decision is implemented: high confidence results are directly output, medium confidence results are labeled and suggested for review, and low confidence results trigger manual review. At the same time, the dataset is updated monthly to iteratively optimize the model. When new transformers or niche faults are added, 5-10 sets of samples are required for 5-10 rounds of fine-tuning to achieve dynamic adaptation and continuous optimization of the model, taking into account the accuracy of fault diagnosis, the controllability of uncertainty, and the adaptability of the scenario.
[0079] Compared with the prior art, the beneficial effects of the present invention are:
[0080] In terms of technological benefits, the system enables automated analysis and intelligent identification of oil chromatography data, reducing the average diagnostic time to 6.8 seconds and significantly improving work efficiency. Through algorithm optimization and sample balancing, the platform achieves a fault diagnosis accuracy rate of 96%, effectively reducing false positives and false negatives and enhancing the scientific rigor and precision of transformer condition monitoring.
[0081] In terms of economic benefits, the application of the system significantly reduces the time cost of manual analysis and the economic losses caused by misdiagnosis. Based on the current annual oil and chemical test volume of Huanggang Company, the platform can save approximately 480 man-hours per year after it is put into use, which translates to an annual saving of approximately 100,000 yuan based on the average labor cost. At the same time, by improving the timeliness of fault detection, it can avoid the risk of potential equipment damage, resulting in an indirect economic benefit of approximately 250,000 yuan per year.
[0082] In terms of management efficiency, it can standardize oil chromatography test analysis, enabling automatic data archiving and traceability of historical results. The technical management department can monitor the health status of equipment in real time, forming visualized data support, and further improving the standardization and intelligence of operation and maintenance work. Attached Figure Description
[0083] Figure 1 This is a flowchart of an artificial intelligence-based oil chromatography test fault detection method that integrates uncertainty perception according to the present invention.
[0084] Figure 2 This is a CVAE-GAN model structure diagram of an artificial intelligence oil chromatography test fault detection method that integrates uncertainty perception according to the present invention.
[0085] Figure 3 This is a dynamic trend diagram of the training of a classification model for a fault detection method in oil chromatography testing based on artificial intelligence that integrates uncertainty perception, as presented in this invention.
[0086] Figure 4This is a graph showing the evaluation index of the model for an artificial intelligence oil chromatography test fault detection method that integrates uncertainty perception, as described in this invention. Detailed Implementation
[0087] The technical solutions of the present invention will now be described in detail with reference to the accompanying drawings.
[0088] like Figure 1-4 As shown, its specific implementation includes the following steps:
[0089] Step S1 selects the content of seven key dissolved gases in the DGA data of oil-immersed transformers as the original data features, uses the 3σ criterion to identify high-noise samples, removes the data of high-noise samples, and eliminates noise through db4 wavelet basis decomposition. Specifically, this includes: selecting the content of seven key dissolved gases in the DGA data of oil-immersed transformers as the original data features, including H2, CH4, C2H2, C2H4, C2H6, CO, and CO2; using the 3σ criterion to remove high-noise samples; and performing a three-level decomposition using the db4 wavelet basis. Strategy to calculate noise threshold:
[0090] ;
[0091] In the formula, This is the noise threshold; data exceeding this threshold will be excluded. The standard deviation of noise in the original DGA data; This represents the total number of samples in the original DGA data.
[0092] Data exceeding the noise threshold is excluded, and the denoised data is output. The relative deviations of the denoised samples from the reference mean of similar fault samples are calculated.
[0093] ;
[0094] In the formula, The relative deviation between the denoised sample and the reference mean; This is a single sample data after denoising using db4 wavelet basis decomposition; The reference mean of similar fault samples is obtained by statistical analysis of historical similar fault samples of transformers of the same voltage level and model; if the deviation is >10%, noise suppression is re-executed; if the deviation is ≤10%, the mapping process is entered.
[0095] Step S1 involves binary image mapping of the processed data to obtain a BIM grayscale image. Specifically, the content of the seven gases is divided into eight intervals according to standard thresholds, with each interval corresponding to an 8-bit binary code. The binary sequences of the first five gases are taken as core features. The five binary sequences of each sample are arranged according to the rule of two sequences per row. The remaining empty spaces are filled with a fixed pixel value M=0, ultimately forming a 32×32 pixel two-dimensional grayscale image, thus completing the image dataset. Finally, the deviation normalization method is used to map all pixel values to the [0,1] interval, with the following formula:
[0096] ;
[0097] In the formula, The standardized pixel values are mapped to the [0,1] interval; The original pixel values are obtained after mapping the binary image, where , These are the row and column indices of the pixel in the image, respectively; For all original pixel values The minimum value in; For all original pixel values The maximum value in the range is used to obtain the final BIM grayscale image, which is then used to form the target sample set. .
[0098] Step S2 involves collecting DGA data of transformers of the same voltage level and model to form a transfer sample set. The target sample set corresponding to the BIM grayscale image obtained in S1 is then merged with the transfer sample set to form an enhanced sample set. Specifically, this includes: collecting DGA data of transformers of the same voltage level and model and processing it through step S1 to form a BIM grayscale image, which is then used to form the transfer sample set. Compare it with the target sample set The formula for fusing samples into an enhanced sample set is as follows:
[0099] ;
[0100] In the formula, To enhance the sample set; The weight coefficients for the target sample set. ; The transfer sample set is adjusted by the MMD domain adaptive algorithm. The MMD calculation formula is:
[0101] ;
[0102] In the formula, This represents the maximum mean difference. The sample set is composed of BIM grayscale images of transformers of the same voltage level and model after being processed by the S1 process. The target sample set consists of BIM grayscale images of the transformer to be tested after the DGA data has been processed by the S1 process. For the transfer sample set The number of samples; For the target sample set The number of samples; For the transfer sample set A single sample in; For the target sample set A single sample in; The feature mapping function maps samples to the reproducing kernel Hilbert space. High-dimensional feature vectors in; To regenerate the kernel Hilbert space, high-dimensional mapping and distance calculation of sample distribution are achieved; for In space Norm squared, quantifies the difference between high-dimensional eigenvectors.
[0103] like Figure 2 As shown, step S2 generates minority class samples that closely match the actual fault mechanism using CVAE-GAN, and supplements the samples for niche faults using a meta-learning algorithm to ensure that the number of samples for each class is ≥200. Specifically, the CVAE-GAN consists of a generator G and a discriminator D, and the generator input includes random noise. and transfer sample feature vector Output generated samples ;
[0104] Its loss function is designed as follows:
[0105] Generator loss: ;
[0106] The quality control of generated samples was verified using the structural similarity index SSIM.
[0107]
[0108] In the formula, To generate the mean of the sample and the real sample, Standard deviation For covariance, It is a constant, and SSIM ≥ 0.88 is required.
[0109] Minority class samples that closely reflect actual fault mechanisms are generated using CVAE-GAN. For niche faults with fewer than 50 samples, supplementary samples are generated using the MAML algorithm. When generating supplementary samples using the MAML algorithm, a meta-learning task set is first constructed based on the small samples of niche faults in the transfer sample set and the target sample set. For each meta-task, 5-8 samples of the niche fault are randomly selected as the support set and 3-5 samples as the query set, ensuring that the task set covers the core data distribution of the niche faults. The optimization objective is to fine-tune the model parameters using support set data through 2-3 rounds of inner loops, and then update the initial parameters based on the expected loss of the query set. After 5-10 rounds of meta-training, the optimal initial parameters are obtained. Subsequently, these optimal parameters are substituted into the model and combined with the feature vectors of the transfer samples. With random noise Supplementary samples that closely match the actual fault mechanism are generated, with the number controlled between 100-150 sets, ultimately ensuring that the total number of samples for this niche fault is ≥200 sets. Furthermore, the generated samples must pass a relative deviation verification with the reference mean of similar faults to ensure consistency with real fault data. The optimization objective of the MAML algorithm is:
[0110] ;
[0111] In the formula, The optimal initial parameters are obtained after optimizing the MAML algorithm. To find the parameters that minimize the objective function ; This is the initial parameter set for the Bayesian CNN model; Let be the expectation operator, representing the expectation of the task set. All tasks Expectations; The meta-learning task set consists of a small sample dataset of minority faults. for A single meta-learning task in; For the task The corresponding loss function; For the After the cycle update, the model in the task Parameterized prediction function on; For the model in the task Up past Parameters after in-round cyclic training;
[0112] 100-150 supplementary samples are generated through 5-10 rounds of meta-training to ensure that the number of samples of each type is ≥200.
[0113] Step S3 constructs a Bayesian CNN classification model with a Bayesian fully connected layer. The enhanced sample set is split into a meta-training set and a target training set as model input. The MAML algorithm is used to perform meta-pre-training on the Bayesian CNN classification model. Specifically, the Bayesian CNN classification model is based on the original BIM-CNN and optimized by replacing the fully connected layer with a Bayesian fully connected layer. The input is a 32×32×1 BIM grayscale image, and the output is the probability distribution and uncertainty measure of 7 types of faults.
[0114] Its model architecture is as follows:
[0115] The input is a 32×32×1 BIM grayscale image, and the network structure includes:
[0116] Convolutional layer 1: 16 5×5 convolutional kernels, ReLU activation function;
[0117] Pooling layer 1: 2×2 max pooling;
[0118] Convolutional layer 2: 32 5×5 convolutional kernels, ReLU activation function;
[0119] Pooling layer 2: 2×2 max pooling;
[0120] Fully connected layer: Replaced with a Bayesian fully connected layer, whose weights follow a Gaussian distribution. ;
[0121] Output layer: Probability distribution of 7 types of faults With uncertainty measurement.
[0122] Training optimization:
[0123] Optimizer: Adam optimizer, learning rate ;
[0124] Loss function: ,in:
[0125] Cross-entropy loss: In the formula This is a real label;
[0126] KL divergence loss:
[0127] ;
[0128] Weighting coefficient To balance classification loss with uncertainty regularization.
[0129] Training strategy: Training for 150 epochs, using early stopping; stopping is initiated if the validation set loss does not decrease for 10 consecutive epochs to avoid overfitting. The training dynamic trend graph is shown below. Figure 3 As shown;
[0130] The augmented sample set is split into a meta-training set and a target training set as model inputs. The splitting rule is as follows: the augmented sample set is split into the meta-training set and the target training set in a 6:4 ratio; the target training set is then split into the training set, test set, and validation set in a 7:2:1 ratio. The Bayesian CNN classification model is then pre-trained using the MAML algorithm with an inner loop learning rate of... , outer loop learning rate Specifically, based on the original scheme, an enhanced sample set is split into a 6:4 ratio to construct a task set containing 500-800 meta-tasks. For each meta-task, one type of fault is randomly selected, and 20-30 samples are used as the support set and 10-15 samples as the query set. Training employs a dual-loop strategy: the inner loop uses the SGD optimizer with a learning rate α=1e-3, fine-tuning the support set data for each meta-task for 3 rounds, updating the specific distribution parameters for that task; the outer loop uses the Adam optimizer with a learning rate β=5e-4, optimizing the initial model based on the expected value of the classification cross-entropy loss + KL divergence regularization loss from the query set. The distribution parameters are set, and the total number of training rounds is set to 200-300. The termination condition is that the average F1 score of the meta-validation set is ≥0.85 or the loss decrease is <1e-5. During training, the variance parameter is constrained by the Softplus activation function to avoid losing the ability to measure uncertainty. 32 meta-tasks are used for batch processing to improve efficiency. The gradient norm is clipped to 5.0 to prevent gradient explosion. The target validation set is used for evaluation every 20 rounds. The classification accuracy is required to be ≥85% and the uncertainty calibration error is required to be ≤8%. If the target is not met, the outer loop learning rate is reduced to 0.5 times the original. The final output is the optimized initial distribution parameters.
[0131] Step S3 integrates multi-model evidence collection with DS evidence theory to output fault probability distributions and uncertainty measures, and evaluates model performance using multiple metrics. Specifically, it involves training three Bayesian CNN classification models with different initial weights, using random seeds of 123, 456, and 789 respectively, and initializing the weights using a He normal distribution to obtain three sets of fault probability distributions. Transform the probability distribution into an evidence mass function Where A is a subset of faults; the DS synthesis rule is used to fuse multi-model evidence, and for any subset of faults A, its fusion probability is... The calculation formula is as follows: ;
[0132] In the formula, For the fault subset The mass function value of the fusion evidence, i.e., the fusion probability, is denoted as... , corresponding to the The fusion probability of fault classes; A subset of faults; The conflict coefficient; , , These are the fault subsets output by three Bayesian CNN models with different initialized weights. , , To represent the fault subsets of the three models respectively , , The initial evidence mass function value;
[0133] Among them, the conflict coefficient The formula is: ;
[0134] In the formula, The conflict coefficient has a value range of [0,1]. The closer the value is to 1, the more severe the conflict of evidence among the three models. , , To represent the fault subsets of the three models respectively , , The initial evidence mass function value;
[0135] Uncertainty is measured using confidence entropy:
[0136] ;
[0137] In the formula, Confidence entropy is used to quantify the uncertainty of fault diagnosis results; This is an index for fault types, corresponding to 7 types of faults; For the first The fusion probability of fault classes; This refers to logarithmic operations in information theory.
[0138] Step S3 evaluates model performance using multiple metrics, specifically accuracy, recall, and F1 score. Each metric must achieve at least 90%. If the requirement is not met, the Bayesian CNN classification model is retrained. The model evaluation metrics are shown in the graph below. Figure 4 As shown.
[0139] Step S4 involves real-time acquisition of DGA data and real-time calculation of its deviation from historical normal data. For data with deviations exceeding a set value, the noise suppression method from S1 is used for data preprocessing to obtain a BIM grayscale image. This BIM image is then input into the classification model constructed in S3 to obtain fault diagnosis results and uncertainty parameters. The confidence-based grading decision-making specifically includes: real-time acquisition of DGA data and calculation of the deviation from historical normal data.
[0140] ;
[0141] In the formula, The average relative deviation between real-time DGA data and historical normal data; This is an index of dissolved gas types, corresponding to 7 key gases; For the first Real-time collection of the content of the gas; For the first Historical average normal concentration of the gas;
[0142] If Δ>15%, the noise suppression method of S1 is used for data preprocessing to obtain a BIM grayscale image. This BIM image is then input into the classification model constructed in S3 to obtain the fault diagnosis results and uncertainty parameters.
[0143] Hierarchical decision-making based on confidence level:
[0144] For high confidence: fusion probability and It directly outputs the fault diagnosis results;
[0145] For medium confidence level: and Output the diagnostic results and indicate recommendations for review;
[0146] For low confidence: and Marked as pending confirmation, triggering the manual review process.
[0147] When adding a transformer or a less common fault in step S4, fine-tune the model using 5-10 sets of samples:
[0148] ;
[0149] In the formula, The final parameters after model fine-tuning are for adding a new transformer or a minor fault. These are the pre-trained parameters before model fine-tuning; To fine-tune the learning rate; For parameters The gradient operator; The loss function corresponding to the newly added samples; For parameters The model prediction function at that time;
[0150] After fine-tuning the number of rounds by 5-10 rounds and evaluating it with 2-3 small sample validation sets, if the accuracy is ≥88%, it can be put into use.
[0151] Application Examples
[0152] A power supply bureau in a certain district operates and maintains 50 110kV oil-immersed distribution transformers of the same model. Due to the fact that there are only 86 historical samples of some niche faults (such as low-energy partial discharge), and that the on-site DGA data is subject to noise due to environmental interference, the traditional diagnostic method has a high misjudgment rate. Therefore, this method is adopted to carry out operation and maintenance monitoring.
[0153] First, seven key dissolved gases, including hydrogen and methane, were selected as features. The 3σ criterion was used to screen out 12 groups of high-noise abnormal samples. Data fluctuation noise was eliminated through db4 wavelet basis decomposition, and the processed data was mapped to a BIM grayscale image. Subsequently, 500 sets of DGA data from transformers of the same voltage level in the surrounding area were collected as migration samples. After fusing with the target samples, CVAE-GAN was used to generate 114 sets of low-energy partial discharge samples. These were then supplemented and improved using a meta-learning algorithm, ensuring that each type of fault sample reached more than 100 sets.
[0154] Next, a Bayesian CNN classification model was constructed. After MAML meta-pre-training and fusion with DS evidence theory, the model achieved a fault diagnosis accuracy of 94.7% and an uncertainty measurement error of <3%. Finally, real-time DGA data of transformers was collected. In mid-September, the data deviation of a certain transformer exceeded the threshold. After preprocessing, the data was input into the model and diagnosed as high temperature overheating (confidence 92.5%, uncertainty 4.8%). Maintenance personnel carried out timely repairs to prevent the fault from escalating.
Claims
1. A fault detection method for oil chromatography tests using artificial intelligence that integrates uncertainty perception, characterized in that, Includes the following steps: S1: The contents of seven key dissolved gases in the DGA data of oil-immersed transformers are selected as the original data features. The 3σ criterion is used to identify high-noise samples, and noise is eliminated by db4 wavelet basis decomposition. The processed data is then subjected to binary image mapping to obtain the BIM grayscale image. S2: Collect DGA data of transformers of the same voltage level and model to form a migration sample set. Merge the target sample set corresponding to the BIM grayscale image obtained in S1 with the migration sample set to form an enhanced sample set. Generate minority class samples that fit the actual fault mechanism through CVAE-GAN. Use meta-learning algorithm to supplement samples for niche faults to ensure that the number of samples of each class is ≥200. S3: Construct a Bayesian CNN classification model with a Bayesian fully connected layer. Split the augmented sample set into a meta-training set and a target training set as model inputs. Use the MAML algorithm to perform meta-pre-training on the Bayesian CNN classification model. Through multi-model evidence collection and DS evidence theory fusion, output the fault probability distribution and uncertainty measure, and use multiple indicators to evaluate the model performance. S4: Real-time acquisition of DGA data, real-time calculation of its deviation from historical normal data, data with deviations exceeding the set value are preprocessed using the noise suppression method of S1 to obtain BIM grayscale images, and the BIM images are input into the classification model constructed by S3 to obtain fault diagnosis results and uncertainty parameters. Based on confidence level decision-making, the dataset is updated monthly and the classification model of S3 is iteratively optimized.
2. The method for fault detection in oil chromatography tests based on artificial intelligence and incorporating uncertainty perception as described in claim 1, characterized in that, Step S1 selects the content of seven key dissolved gases in the DGA data of oil-immersed transformers as the original data features, uses the 3σ criterion to identify high-noise samples, removes the data of high-noise samples, and eliminates noise through db4 wavelet basis decomposition. Specifically, this includes the following: The contents of seven key dissolved gases in the DGA data of oil-immersed transformers were selected as the original data features, including H2, CH4, C2H2, C2H4, C2H6, CO, and CO2. High-noise samples were removed using the 3σ criterion, and a three-level decomposition was performed using the db4 wavelet basis. The noise threshold was calculated using the Birgé-Massart strategy. ； In the formula, This is the noise threshold; data exceeding this threshold will be excluded. The standard deviation of noise in the original DGA data; The total number of samples in the original DGA data; data exceeding the noise threshold are excluded, and the denoised data is output. The relative deviation of the denoised samples from the reference mean of similar fault samples is calculated: ； In the formula, The relative deviation between the denoised sample and the reference mean; This is a single sample data after denoising using db4 wavelet basis decomposition; The reference mean of similar fault samples is obtained by statistical analysis of historical similar fault samples of transformers of the same voltage level and model; if the deviation is >10%, noise suppression is re-executed; if the deviation is ≤10%, the mapping process is entered.
3. The artificial intelligence oil chromatography test fault detection method integrating uncertainty perception as described in claim 1, characterized in that, Step S1 involves binary image mapping of the processed data to obtain a BIM grayscale image. Specifically, the content of the seven gases is divided into eight intervals according to standard thresholds, with each interval corresponding to an 8-bit binary code. The binary sequences of the first five gases are taken as core features. The five binary sequences of each sample are arranged according to the rule of two sequences per row. The remaining empty spaces are filled with a fixed pixel value M=0, ultimately forming a 32×32 pixel two-dimensional grayscale image, thus completing the image dataset. Finally, the deviation normalization method is used to map all pixel values to the [0,1] interval, with the following formula: ； In the formula, The standardized pixel values are mapped to the [0,1] interval; The original pixel values are obtained after mapping the binary image, where , These are the row and column indices of the pixel in the image, respectively; For all original pixel values The minimum value in; For all original pixel values The maximum value in the range is used to obtain the final BIM grayscale image, which is then used to form the target sample set. .
4. The artificial intelligence oil chromatography test fault detection method integrating uncertainty perception as described in claim 1, characterized in that, Step S2 involves collecting DGA data from transformers of the same voltage level and model to form a transfer sample set. The target sample set corresponding to the BIM grayscale image obtained in S1 is then fused with the transfer sample set to create an enhanced sample set. This process specifically includes the following: The BIM grayscale images obtained after processing DGA data of transformers of the same voltage level and model through step S1 form a transfer sample set. Compare it with the target sample set The formula for fusing samples into an enhanced sample set is as follows: ； In the formula, To enhance the sample set; The weight coefficients for the target sample set. ; The transfer sample set is adjusted by the MMD domain adaptive algorithm. The MMD calculation formula is: ； In the formula, This represents the maximum mean difference. The sample set is composed of BIM grayscale images of transformers of the same voltage level and model after being processed by the S1 process. The target sample set consists of BIM grayscale images of the transformer to be inspected after DGA data has been processed by the S1 process. For the transfer sample set The number of samples; For the target sample set The number of samples; For the transfer sample set A single sample in; For the target sample set A single sample in; The feature mapping function maps samples to the reproducing kernel Hilbert space. High-dimensional feature vectors in; To regenerate the kernel Hilbert space, high-dimensional mapping and distance calculation of sample distribution are achieved; for In space Norm squared, quantifies the difference between high-dimensional eigenvectors.
5. The artificial intelligence oil chromatography test fault detection method integrating uncertainty perception according to claim 1, characterized in that, Step S2 generates minority class samples that closely match the actual fault mechanisms using CVAE-GAN. For less common faults, a meta-learning algorithm is used to supplement the samples, ensuring that the number of samples in each class is ≥200. Specifically, the CVAE-GAN consists of a generator G and a discriminator D. The generator input includes random noise. and transfer sample feature vector Output generated samples The CVAE-GAN algorithm generates minority class samples that closely reflect actual fault mechanisms. For niche faults with fewer than 50 samples, the MAML algorithm is used to generate supplementary samples. The optimization objective of the MAML algorithm is as follows: ； In the formula, The optimal initial parameters are obtained after optimizing the MAML algorithm. To find the parameters that minimize the objective function ; This is the initial parameter set for the Bayesian CNN model; Let be the expectation operator, representing the expectation of the task set. All tasks Expectations; The meta-learning task set consists of a small sample dataset of minority faults. for A single meta-learning task in; For the task The corresponding loss function; For the After the cycle update, the model in the task Parameterized prediction function on; For the model in the task Up past Parameters after in-round cyclic training; 100-150 supplementary samples are generated through 5-10 rounds of meta-training to ensure that the number of samples of each type is ≥200.
6. The artificial intelligence oil chromatography test fault detection method integrating uncertainty perception according to claim 1, characterized in that, Step S3 constructs a Bayesian CNN classification model with a Bayesian fully connected layer. The enhanced sample set is split into a meta-training set and a target training set as model input. The MAML algorithm is used to perform meta-pre-training on the Bayesian CNN classification model. Specifically, the Bayesian CNN classification model is based on the original BIM-CNN and optimized by replacing the fully connected layer with a Bayesian fully connected layer. The input is a 32×32×1 BIM grayscale image, and the output is the probability distribution and uncertainty measure of 7 types of faults. The augmented sample set is split into a meta-training set and a target training set as model inputs. The splitting rule is as follows: the augmented sample set is split into the meta-training set and the target training set in a 6:4 ratio; the target training set is then split into the training set, test set, and validation set in a 7:2:1 ratio. The Bayesian CNN classification model is then pre-trained using the MAML algorithm with an inner loop learning rate of... , outer loop learning rate .
7. The artificial intelligence oil chromatography test fault detection method integrating uncertainty perception as described in claim 1, characterized in that, Step S3 integrates multi-model evidence collection with DS evidence theory to output fault probability distributions and uncertainty measures, and evaluates model performance using multiple metrics. Specifically, it involves training three Bayesian CNN classification models with different initial weights, using random seeds of 123, 456, and 789 respectively, and initializing the weights using a He normal distribution to obtain three sets of fault probability distributions. Transform the probability distribution into an evidence mass function Where A is a subset of faults; the DS synthesis rule is used to fuse multi-model evidence, and for any subset of faults A, its fusion probability is... The calculation formula is as follows: ; In the formula, For the fault subset The mass function value of the fusion evidence, i.e., the fusion probability, is denoted as... , corresponding to the The fusion probability of fault classes; A subset of faults; The conflict coefficient; , , These are the fault subsets output by three Bayesian CNN models with different initialized weights. , , To represent the fault subsets of the three models respectively , , The initial evidence mass function value; Among them, the conflict coefficient The formula is: In the formula, The conflict coefficient has a value range of [0,1]. The closer a value is to 1, the more severe the conflict of evidence among the three models. , , To represent the fault subsets of the three models respectively , , The initial evidence mass function value; Uncertainty is measured using confidence entropy: ； In the formula, Confidence entropy is used to quantify the uncertainty of fault diagnosis results; This is an index for fault types, corresponding to 7 types of faults; For the first The fusion probability of fault classes; This refers to logarithmic operations in information theory.
8. The artificial intelligence oil chromatography test fault detection method integrating uncertainty perception as described in claim 1, characterized in that, Step S3 evaluates the model performance using multiple metrics, specifically accuracy, recall, and F1 score. Each metric must reach 90%. If the requirement is not met, the model is retrained using the Bayesian CNN classification model.
9. The artificial intelligence oil chromatography test fault detection method integrating uncertainty perception according to claim 1, characterized in that, Step S4 involves real-time acquisition of DGA data and real-time calculation of its deviation from historical normal data. For data with deviations exceeding a set value, the noise suppression method from S1 is used for data preprocessing to obtain a BIM grayscale image. This BIM image is then input into the classification model constructed in S3 to obtain fault diagnosis results and uncertainty parameters. The confidence-based grading decision-making specifically includes: real-time acquisition of DGA data and calculation of the deviation from historical normal data. ； In the formula, The average relative deviation between real-time DGA data and historical normal data; This is an index of dissolved gas types, corresponding to 7 key gases; For the first Real-time collection of the content of various gases; For the first Historical average normal concentration of the gas; If Δ>15%, the noise suppression method of S1 is used for data preprocessing to obtain a BIM grayscale image. This BIM image is then input into the classification model constructed in S3 to obtain the fault diagnosis results and uncertainty parameters. Hierarchical decision-making based on confidence level: For high confidence: fusion probability and It directly outputs the fault diagnosis results; For medium confidence level: and Output the diagnostic results and indicate recommendations for review; For low confidence levels: and Marked as pending confirmation, triggering the manual review process; The dataset is updated monthly and the S3 classification model is iteratively optimized.
10. The artificial intelligence oil chromatography test fault detection method integrating uncertainty perception according to claim 1, characterized in that, When adding a transformer or a less common fault in step S4, fine-tune the model using 5-10 sets of samples: ； In the formula, The final parameters after model fine-tuning are for adding a new transformer or a minor fault. These are the pre-trained parameters before model fine-tuning; To fine-tune the learning rate; For parameters The gradient operator; The loss function corresponding to the newly added samples; For parameters The model prediction function at that time; After fine-tuning the number of rounds by 5-10 rounds and evaluating it with 2-3 small sample validation sets, if the accuracy is ≥88%, it can be put into use.