Transformer insulating oil chromatographic data abnormity identification method

Abstract

The invention discloses a transformer insulating oil chromatographic data abnormity identification method, and relates to the technical field of data identification, and the method comprises the following steps: carrying out dissolved gas analysis on an insulating oil sample through a gas chromatograph to obtain a chromatographic original response curve; performing baseline correction and de-noising processing on the chromatographic original response curve, positioning each gas peak based on a first-order derivative method and a second-order derivative method, and extracting chromatographic peak height and chromatographic peak area parameters; according to the extracted chromatographic peak height and chromatographic peak area, three Duval fault ratios are calculated, and the ratio of the total soluble gas concentration to each gas concentration threshold value is calculated; combining into a multi-dimensional diagnosis feature vector; according to the method, the chromatographic peak shape of the insulating oil sample is subjected to multi-stage baseline correction and de-noising treatment, and the gas peak position is accurately positioned based on the first-order and second-order derivative methods, so that the high precision of chromatographic peak height and chromatographic peak area extraction is ensured, the accuracy of dissolved gas concentration calculation is improved, and the misjudgment risk caused by noise or baseline drift is reduced.

Description

Technical Field

[0001] This invention relates to the field of data recognition technology, specifically to a method for identifying anomalies in transformer insulating oil chromatographic data. Background Technology

[0002] Chromatographic analysis, due to its high resolution and sensitivity, is widely used in environmental monitoring, drug screening, and food safety testing. However, in practical applications, chromatographic data is often affected by various factors such as baseline drift, system noise, uneven sample pretreatment, peak tailing, and peak overlap, leading to shifts or distortions in retention time and peak shape characteristics, resulting in quantitative errors and qualitative misjudgments. Traditional anomaly detection methods are mostly based on fixed thresholds or simple statistical models, making it difficult to balance time dependence and peak shape diversity, and exhibiting poor robustness to sudden noise and system fluctuations. Furthermore, in online monitoring scenarios, changes in mobile phase composition and batch-to-batch differences in samples further exacerbate the complexity of anomaly detection. A method capable of adaptively handling multiple... Anomaly identification methods for chromatographic data, characterized by high sensitivity and low false alarm rate, are developed to ensure the reliability and stability of chromatographic analysis results. Dissolved gas components in insulating oil provide crucial information for diagnosing internal transformer faults. Gas chromatography (GC), an internationally recognized dissolved gas analysis technique, can highly sensitively separate and quantify various gaseous components in oil samples, including hydrogen, methane, ethylene, acetylene, and carbon dioxide. The peak height and peak area parameters reflected in the original chromatographic response curve are not only positively correlated with gas concentration but can also reflect various fault modes such as partial discharge, thermal breakdown, and arc discharge through multi-dimensional indicators such as the Duval fault ratio method and the ratio of total soluble gas concentration to the safety threshold. In existing chromatographic analysis techniques, the calculation of chromatographic peak height and chromatographic peak area simply relies on the sharp peak assumption, lacks accurate boundary identification, and often results in deviation of the integration range, leading to an increase in the error of dissolved gas concentration calculation; To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for identifying anomalies in transformer insulating oil chromatographic data.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying anomalies in transformer insulating oil chromatographic data, comprising: S1. Dissolved gas analysis of insulating oil sample is performed by gas chromatography to obtain raw chromatographic response curve; baseline correction and noise reduction are performed on the raw chromatographic response curve, and each gas peak is located based on the first and second derivative methods, and chromatographic peak height and chromatographic peak area parameters are extracted. S2. Based on the extracted chromatographic peak height and chromatographic peak area, calculate the three Duval failure ratios, calculate the ratio of total soluble gas concentration to the concentration threshold of each gas, and combine them into a multidimensional diagnostic feature vector. S3. Input the multidimensional diagnostic feature vector into a One-Class SVM model, and calculate the anomaly metric by the distance between the sample and the training boundary in the high-dimensional feature space. S4. Weighted fusion of the anomaly metrics output by a One-Class SVM model to obtain a comprehensive anomaly confidence score; and output alarm levels according to preset anomaly confidence score thresholds.

[0005] Dissolved gas analysis was performed on the insulating oil sample using gas chromatography to obtain the original chromatographic response curve. The baseline of the original chromatographic response curve was corrected by adaptive cubic polynomial to remove baseline drift. On the baseline-corrected curve, Symlet8 wavelet transform was first used for multi-scale noise suppression, and then 0.2s moving average filtering was used to eliminate high and low frequency interference to obtain a smooth signal y(t).

[0006] For the smoothed signal y(t), calculate the first derivative y'(t). By identifying the point of maximum rate of increase from positive to negative and the point of maximum rate of decrease from negative to positive, determine the peak start point tstart and peak end point tend. Then, calculate the second derivative y''(t) within this interval. Use its minimum value (the point of maximum curvature) as the candidate peak extremum point tpeak2. In the same interval, use the zero crossover point of the first derivative tpeak1 as another candidate peak extremum point. If |tpeak1-tpeak2|≤0.05 s, take the average of the two as the final peak extremum point; otherwise, prioritize the second derivative for location. At the final peak extreme point, the chromatographic peak height is obtained by the difference between the signal intensity and the corresponding baseline value, and the chromatographic peak area is obtained by performing a compound Simpson integral on the difference between the signal and the baseline over the interval from the peak start to the peak end.

[0007] Step two includes: S21. The area of ​​each gas chromatographic peak is converted using a pre-calibrated response factor, and the area of ​​each chromatographic peak is multiplied by the corresponding response factor to obtain the concentration of each dissolved gas. S22. Calculate the three Duval fault ratios: the ratio of acetylene to ethylene concentration, the ratio of methane to hydrogen concentration, and the ratio of ethylene to hydrogen concentration, and linearly normalize the three Duval fault ratios to between zero and one. S23. The total soluble gas concentration is obtained by summing up all gas concentrations, and the ratio of each gas concentration to its preset safety threshold is calculated as the gas concentration ratio. The largest of the concentration ratios is taken as the maximum concentration ratio. S24. The three Duval fault ratios, the concentrations of each dissolved gas, the maximum concentration ratio, and the concentration ratios of each gas are concatenated in a fixed order to form a multidimensional diagnostic feature vector.

[0008] Using multidimensional diagnostic feature vectors as the training subset sample set, a one-class support vector machine (SVM) model is constructed, including: mapping the standardized training samples to a high-dimensional feature space, establishing the sample inner product kernel matrix based on the radial basis kernel function; introducing Lagrange multipliers and bias terms to form a dual optimization objective, wherein the kernel width parameter is determined according to the sample feature dimension, and the tolerance anomaly ratio parameter ν is set to 0.05.

[0009] The dual optimization problem is solved by the Sequence Minimum Optimization (SMO) algorithm: all Lagrange multipliers are initialized to the same reasonable value, and the bias term is initialized to zero; during the iteration, two multipliers are selected in turn, and their values ​​are adjusted synchronously using the closed-loop update rule within their allowable range, and the bias term is updated to maintain the optimality condition; this is repeated until all samples satisfy the KKT condition or the maximum number of iterations is reached.

[0010] Once the SMO iteration converges, all multipliers are traversed, and the samples corresponding to the multipliers located in the interval between zero and the upper limit are identified as support vectors, and the bias term is determined by their decision function value; only support vectors and their multipliers are retained.

[0011] Package the standardized parameters, kernel parameters, support vector indexes, corresponding multipliers, and bias terms into a model file; test using an independent validation set, and the proportion of normal samples being judged as abnormal should be approximately equal to ν. If the deviation exceeds the predetermined range, adjust the kernel width or ν value and retrain; after confirming that the model performance meets the design specifications, deploy the model for dissolved gas anomaly detection. The multidimensional diagnostic feature vector is input into the One-Class SVM model. The model calculates the feature offset of the sample from the training boundary in the high-dimensional feature space, and calculates and outputs the anomaly metric based on the feature offset.

[0012] The anomaly metrics of each gas output by the One-Class SVM model are multiplied by a pre-set weighting coefficient and then summed to obtain a weighted sum. The weighted sum is then mapped to an anomaly confidence level of zero to one using a mapping function calibrated for normal and anomaly samples. The anomaly confidence level is compared with a preset warning level threshold and a severity level threshold. A warning level alarm is triggered when the anomaly confidence level is greater than or equal to the warning level threshold and less than the severity level threshold. A severity level alarm is triggered when the anomaly confidence level is greater than or equal to the severity level threshold.

[0013] This invention provides a method for identifying anomalies in the chromatographic data of transformer insulating oil. Compared with existing technologies, it has the following advantages: This invention performs multi-level baseline correction and noise reduction on the chromatographic peak shape of insulating oil samples, and accurately locates the gas peak position based on the first and second derivative methods, ensuring high precision in the extraction of chromatographic peak height and area, thereby improving the accuracy of dissolved gas concentration calculation and reducing the risk of misjudgment caused by noise or baseline drift. This invention constructs a multidimensional diagnostic feature vector, integrates the Duval fault ratio, the ratio of total soluble gas concentration to concentration threshold, and inputs it into a support vector machine model for single-class learning and boundary distance determination, thereby distinguishing between normal and abnormal dissolved gas patterns and significantly improving the detection rate and diagnostic reliability of fault samples. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the principle framework of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 This application provides a method for identifying anomalies in transformer insulating oil chromatographic data, including: S1. Dissolved gas analysis of insulating oil sample is performed by gas chromatography to obtain raw chromatographic response curve; baseline correction and noise reduction are performed on the raw chromatographic response curve, and each gas peak is located based on the first and second derivative methods, and chromatographic peak height and chromatographic peak area parameters are extracted. S2. Based on the extracted chromatographic peak height and chromatographic peak area, calculate the three Duval failure ratios, calculate the ratio of total soluble gas concentration to the concentration threshold of each gas, and combine them into a multidimensional diagnostic feature vector. S3. Input the multidimensional diagnostic feature vector into a One-Class SVM model, and calculate the anomaly metric by the distance between the sample and the training boundary in the high-dimensional feature space. S4. Weighted fusion of the anomaly metrics output by a One-Class SVM model to obtain a comprehensive anomaly confidence score; and output alarm levels according to preset anomaly confidence score thresholds.

[0017] Dissolved gas analysis was performed on the insulating oil sample using gas chromatography to obtain the original chromatographic response curve. The baseline of the original chromatographic response curve was corrected by adaptive cubic polynomial to remove baseline drift. On the baseline-corrected curve, Symlet8 wavelet transform was first used for multi-scale noise suppression, and then 0.2s moving average filtering was used to eliminate high and low frequency interference to obtain a smooth signal y(t).

[0018] It is worth noting that in step one, the insulating oil sample is placed in a headspace vial and injected into a GC-FID chromatograph under preset headspace equilibrium conditions. The current signal of the dissolved gas is recorded at a sampling frequency higher than 200Hz, and a raw chromatographic response curve is generated. Then, non-peak regions are automatically identified on the curve, and adaptive cubic polynomial least squares fitting is used to remove baseline drift. If necessary, the fitting window and order are iteratively adjusted to eliminate baseline residual spurious peaks. Next, the baseline-corrected curve is subjected to five-level Symlet8 wavelet decomposition, and soft threshold denoising is performed on each detail coefficient based on the general threshold method before reconstruction to remove high-frequency random noise. Finally, with a sliding window length of 0.2s, moving average filtering is performed on the reconstructed signal, and boundary extension processing is performed on the beginning and end of the sequence to eliminate low-frequency interference, resulting in a signal y(t) that retains peak characteristics and is highly smooth.

[0019] For the smoothed signal y(t), calculate the first derivative y'(t). By identifying the point of maximum rate of increase from positive to negative and the point of maximum rate of decrease from negative to positive, determine the peak start point tstart and peak end point tend. Then, calculate the second derivative y''(t) within this interval. Use its minimum value (the point of maximum curvature) as the candidate peak extremum point tpeak2. In the same interval, use the zero crossover point of the first derivative tpeak1 as another candidate peak extremum point. If |tpeak1-tpeak2|≤0.05 s, take the average of the two as the final peak extremum point; otherwise, prioritize the second derivative for location. At the final peak extreme point, the chromatographic peak height is obtained by the difference between the signal intensity and the corresponding baseline value, and the chromatographic peak area is obtained by performing a compound Simpson integral on the difference between the signal and the baseline over the interval from the peak start to the peak end.

[0020] Suppose that in a certain chromatographic injection, the discrete values ​​of the smoothed signal y(t) at t={0,1,2,3,4,5,6}s sampled at 1Hz are {0,1,3,7,5,2,0}, then firstly, the first derivative y′(t)≈[y(t+1)-y(t-1)] / 2 is calculated using the central difference method. This yields the maximum rate of increase from positive to negative at t=2s, thus determining the peak start point tstart=2s; and the maximum rate of decrease from negative to positive at t=5s, thus determining the peak end point tend=5s. Subsequently, the second derivative y″(t)≈y(t+1)-2y(t)+y(t-1) was calculated using the central difference method within the interval [2s, 5s]. Its minimum value occurred at t=4s, which was taken as the candidate peak extreme point tpeak2. The zero crossover point of the first derivative tpeak1≈3.5s, and the difference between it and tpeak2 is 0.5s, which is greater than 0.05s. Therefore, the second derivative was preferred for positioning, and the final peak extreme point tpeak=4s was determined. The chromatographic peak height is equal to the difference between y(4s) and the corresponding baseline value. The chromatographic peak area was calculated by performing a compound Simpson integral on the baseline-delimited signal value within the interval from t=2s to t=5s.

[0021] Step two includes: S21. The area of ​​each gas chromatographic peak is converted using a pre-calibrated response factor, and the area of ​​each chromatographic peak is multiplied by the corresponding response factor to obtain the concentration of each dissolved gas. S22. Calculate the three Duval fault ratios: the ratio of acetylene to ethylene concentration, the ratio of methane to hydrogen concentration, and the ratio of ethylene to hydrogen concentration, and linearly normalize the three Duval fault ratios to between zero and one. S23. The total soluble gas concentration is obtained by summing up all gas concentrations, and the ratio of each gas concentration to its preset safety threshold is calculated as the gas concentration ratio. The largest of the concentration ratios is taken as the maximum concentration ratio. S24. The three Duval fault ratios, the concentrations of each dissolved gas, the maximum concentration ratio, and the concentration ratios of each gas are concatenated in a fixed order to form a multidimensional diagnostic feature vector.

[0022] It is worth noting that the peak areas of each gas are converted, and the concentration of each dissolved gas Ci is obtained by multiplying the peak area by the corresponding response factor according to Ci=Ai×ri. Concentration values ​​exceeding the instrument's range are marked as overflow alarms. Based on this concentration result, three Duval fault ratios (D1=C_acetylene / C_ethylene, D2=C_methane / C_hydrogen, D3=C_ethylene / C_hydrogen) are calculated. Then, using the preset minimum value Dk_min and maximum value Dk_max as boundaries, each ratio is defined by Dk_norm=(Dk–Dk_min) / (Dk_max–Dk_min). The system maps the values ​​to [0,1] and then prunes values ​​that exceed the boundary, setting negative values ​​to zero and values ​​exceeding one to one. Next, the system accumulates all Ci values ​​to calculate the total soluble gas concentration C_total and calculates the safety ratio by comparing it with the safety threshold Ti ratio Ri=Ci / Ti for each gas. The maximum ratio R_max is taken as the concentration over-limit criterion (when R_max≥1, a concentration over-limit alarm is triggered). Finally, the three normalized Duval fault ratios, the concentrations of each dissolved gas Ci, the maximum concentration ratio R_max, and the safety ratios of each single gas are concatenated in a fixed order to form a multidimensional diagnostic feature vector X.

[0023] This invention performs multi-level baseline correction and noise reduction on the chromatographic peak shape of insulating oil samples, and accurately locates the gas peak position based on the first and second derivative methods, ensuring high precision in the extraction of chromatographic peak height and area, thereby improving the accuracy of dissolved gas concentration calculation and reducing the risk of misjudgment caused by noise or baseline drift.

[0024] Using multidimensional diagnostic feature vectors as the training subset sample set, a one-class support vector machine (SVM) model is constructed, including: mapping the standardized training samples to a high-dimensional feature space, establishing the sample inner product kernel matrix based on the radial basis kernel function; introducing Lagrange multipliers and bias terms to form a dual optimization objective, wherein the kernel width parameter is determined according to the sample feature dimension, and the tolerance anomaly ratio parameter ν is set to 0.05.

[0025] It is worth noting that after using the multidimensional diagnostic feature vectors concatenated in step S2 as the training subset sample set, the training sample set is first preprocessed with zero mean and unit variance standardization. Then, the radial basis function (RBF) is used to generate the sample inner product kernel matrix in the high-dimensional feature space. The kernel width parameter is determined and fine-tuned based on the sample feature dimension and variance distribution using empirical methods. Subsequently, Lagrange multipliers and bias terms are introduced into the dual space, and the multiplier constraint condition is set according to the tolerance anomaly ratio parameter ν = 0.05. Then, the Sequential Minimal Optimization (SMO) algorithm is called to iteratively solve the above dual optimization problem until all samples satisfy the KKT optimality condition or reach the preset maximum number of iterations. Finally, the converged Lagrange multipliers and bias terms are packaged and persisted together with the support vector index to form a One-Class SVM model file, which is used for the subsequent online calculation of dissolved gas anomaly measurement values.

[0026] The dual optimization problem is solved by the Sequence Minimum Optimization (SMO) algorithm: all Lagrange multipliers are initialized to the same reasonable value, and the bias term is initialized to zero; during the iteration, two multipliers are selected in turn, and their values ​​are adjusted synchronously using the closed-loop update rule within their allowable range, and the bias term is updated to maintain the optimality condition; this is repeated until all samples satisfy the KKT condition or the maximum number of iterations is reached.

[0027] Once the SMO iteration converges, all multipliers are traversed, and the samples corresponding to the multipliers located in the interval between zero and the upper limit are identified as support vectors, and the bias term is determined by their decision function value; only support vectors and their multipliers are retained.

[0028] Package the standardized parameters, kernel parameters, support vector indexes, corresponding multipliers, and bias terms into a model file; test using an independent validation set, and the proportion of normal samples being judged as abnormal should be approximately equal to ν. If the deviation exceeds the predetermined range, adjust the kernel width or ν value and retrain; after confirming that the model performance meets the design specifications, deploy the model for dissolved gas anomaly detection. The multidimensional diagnostic feature vector is input into the One-Class SVM model. The model calculates the feature offset of the sample from the training boundary in the high-dimensional feature space, and calculates and outputs the anomaly metric based on the feature offset.

[0029] It is worth noting that in the construction of the One-Class SVM model, multidimensional diagnostic feature vectors are used as a training subset. First, the mean and standard deviation of each feature dimension of all samples are calculated, and normalization is performed by subtracting the mean from the current value and then dividing by the standard deviation. Then, a radial basis function is used to map the normalized samples to a high-dimensional feature space, where the kernel width parameter is set to twice the reciprocal of the sample feature dimension, and the inner product kernel matrix of the samples is constructed accordingly. Next, Lagrange multipliers and bias terms are introduced to form the dual optimization problem of One-Class SVM, and the tolerance anomaly ratio parameter ν is set to 0.05 to allow about 5% of the training samples to be outside the boundary. This dual problem is solved by the Sequence Minimum Optimization (SMO) algorithm, which first initializes all multipliers to the same reasonable value and the bias term to zero. Then, in each iteration, two multipliers are selected and their values ​​are synchronously adjusted within their constraints using a closed-loop update formula. Simultaneously, the bias term is corrected to maintain the KKT conditions. This iterative process continues until all samples satisfy optimality or the maximum number of iterations is reached. After convergence, all multipliers are traversed, and samples corresponding to non-zero multipliers between zero and the upper limit are identified as support vectors. The bias term is determined by the decision function value of the support vectors, and only the support vectors and their multipliers are retained to simplify the model. Finally, the normalization parameters, kernel width parameters, support vector indices, corresponding multipliers, and bias terms are packaged and stored as a model file. The model is then tested using an independent validation set. The proportion of normal samples judged as abnormal should be approximately equal to ν. If the deviation exceeds a predetermined range, the kernel width or ν value is adjusted, and the model is retrained until its performance meets the design specifications. Finally, the model is deployed in the online dissolved gas anomaly detection module.

[0030] The anomaly metrics of each gas output by the One-Class SVM model are multiplied by a pre-set weighting coefficient and then summed to obtain a weighted sum. The weighted sum is then mapped to an anomaly confidence level of zero to one using a mapping function calibrated for normal and anomaly samples. The anomaly confidence level is compared with a preset warning level threshold and a severity level threshold. A warning level alarm is triggered when the anomaly confidence level is greater than or equal to the warning level threshold and less than the severity level threshold. A severity level alarm is triggered when the anomaly confidence level is greater than or equal to the severity level threshold.

[0031] First, the anomaly measurement acquisition unit receives anomaly measurement values ​​for each dissolved gas from a One-Class SVM model. Then, in the weight setting unit, these measurement values ​​are multiplied by pre-set weight coefficients, and the product results are summed in the accumulation unit to obtain a weighted sum. Next, the mapping calibration unit inputs this weighted sum into a mapping function obtained by calibrating normal and abnormal samples. This function maps the weighted sum to an anomaly confidence value between 0 and 1. Subsequently, the alarm determination unit compares the anomaly confidence value with preset warning level thresholds and severity level thresholds. When the anomaly confidence value is greater than or equal to the warning level threshold and less than the severity level threshold, the system triggers a warning level alarm and indicates the affected gas and its corresponding confidence value in the alarm information. When the anomaly confidence value is greater than or equal to the severity level threshold, a severity level alarm is triggered, and the emergency response module is activated simultaneously to perform subsequent maintenance measures.

[0032] Furthermore, this invention constructs a multidimensional diagnostic feature vector, integrates the Duval fault ratio, the ratio of total soluble gas concentration to concentration threshold, and inputs it into a support vector machine model for single-class learning and boundary distance determination, thereby achieving the distinction between normal and abnormal dissolved gas patterns, and thus significantly improving the detection rate and diagnostic reliability of fault samples.

[0033] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0034] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0035] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for identifying anomalies in transformer insulating oil chromatographic data, characterized in that, Executed by a computer, including: S1. Dissolved gas analysis of insulating oil sample is performed by gas chromatography to obtain raw chromatographic response curve; baseline correction and noise reduction are performed on the raw chromatographic response curve, and each gas peak is located based on the first and second derivative methods, and chromatographic peak height and chromatographic peak area parameters are extracted. S2. Based on the extracted chromatographic peak height and chromatographic peak area, calculate the three Duval failure ratios, calculate the ratio of total soluble gas concentration to the concentration threshold of each gas, and combine them into a multidimensional diagnostic feature vector. S3. Input the multidimensional diagnostic feature vector into a One-Class SVM model, and calculate the anomaly metric by the distance between the sample and the training boundary in the high-dimensional feature space. S4. Weighted fusion of the anomaly metrics output by a One-Class SVM model to obtain a comprehensive anomaly confidence score; and output alarm levels according to preset anomaly confidence score thresholds.

2. The method for identifying anomalies in transformer insulating oil chromatographic data according to claim 1, characterized in that, Dissolved gas analysis was performed on the insulating oil sample using gas chromatography to obtain the original chromatographic response curve. The baseline of the original chromatographic response curve was corrected by adaptive cubic polynomial to remove baseline drift. On the baseline-corrected curve, Symlet8 wavelet transform was first used for multi-scale noise suppression, and then 0.2s moving average filtering was used to eliminate high and low frequency interference to obtain a smooth signal y(t).

3. The method for identifying anomalies in transformer insulating oil chromatographic data according to claim 1, characterized in that, Calculate the first derivative y'(t) of the smoothed signal y(t), and determine the peak start point tstart and peak end point by identifying the point of maximum rise rate from positive to negative and the point of maximum fall rate from negative to positive. Then, calculate the second derivative y''(t) within the interval, and use its minimum value (the point of maximum curvature) as the candidate peak extreme point tpeak2. In the same interval, use the zero crossover point of the first derivative tpeak1 as another candidate peak extreme point. If |tpeak1-tpeak2|≤0.05 s, then take the average of the two as the final peak extreme point; otherwise, the second derivative is preferred for location. At the final peak extreme point, the chromatographic peak height is obtained by the difference between the signal intensity and the corresponding baseline value, and the chromatographic peak area is obtained by performing a compound Simpson integral on the difference between the signal and the baseline over the interval from the peak start to the peak end.

4. The method for identifying anomalies in transformer insulating oil chromatographic data according to claim 1, characterized in that, Step two includes: S21. The area of ​​each gas chromatographic peak is converted using a pre-calibrated response factor, and the area of ​​each chromatographic peak is multiplied by the corresponding response factor to obtain the concentration of each dissolved gas. S22. Calculate the three Duval fault ratios: the ratio of acetylene to ethylene concentration, the ratio of methane to hydrogen concentration, and the ratio of ethylene to hydrogen concentration, and linearly normalize the three Duval fault ratios to between zero and one. S23. The total soluble gas concentration is obtained by summing up all gas concentrations, and the ratio of each gas concentration to its preset safety threshold is calculated as the gas concentration ratio. The largest of the concentration ratios is taken as the maximum concentration ratio. S24. The three Duval fault ratios, the concentrations of each dissolved gas, the maximum concentration ratio, and the concentration ratios of each gas are concatenated in a fixed order to form a multidimensional diagnostic feature vector.

5. The method for identifying anomalies in transformer insulating oil chromatographic data according to claim 1, characterized in that, Using multidimensional diagnostic feature vectors as the training subset sample set, a one-class support vector machine (SVM) model is constructed, including: mapping the standardized training samples to a high-dimensional feature space, establishing the sample inner product kernel matrix based on the radial basis kernel function; introducing Lagrange multipliers and bias terms to form a dual optimization objective, wherein the kernel width parameter is determined according to the sample feature dimension, and the tolerance anomaly ratio parameter ν is set to 0.

05.

6. The method for identifying anomalies in transformer insulating oil chromatographic data according to claim 5, characterized in that, The dual optimization problem is solved by the Sequence Minimum Optimization (SMO) algorithm: all Lagrange multipliers are initialized to the same reasonable value, and the bias term is initialized to zero; during the iteration process, two multipliers are selected in turn, and their values ​​are adjusted synchronously using the closed-loop update rule within their allowable range, and the bias term is updated to maintain the optimality condition; this is repeated until all samples satisfy the KKT condition or the maximum number of iterations is reached.

7. The method for identifying anomalies in transformer insulating oil chromatographic data according to claim 5, characterized in that, After the SMO iteration converges, all multipliers are traversed, and the samples corresponding to the multipliers located in the interval between zero and the upper limit are identified as support vectors, and the bias term is determined by their decision function value. Only support vectors and their multipliers are retained.

8. The method for identifying anomalies in transformer insulating oil chromatographic data according to claim 5, characterized in that, Package the standardized parameters, kernel parameters, support vector indexes, corresponding multipliers, and bias terms into a model file; test using an independent validation set, and the proportion of normal samples being judged as abnormal should be approximately equal to ν. If the deviation exceeds the predetermined range, adjust the kernel width or ν value and retrain; after confirming that the model performance meets the design specifications, deploy the model for dissolved gas anomaly detection. The multidimensional diagnostic feature vector is input into the One-Class SVM model. The model calculates the feature offset of the sample from the training boundary in the high-dimensional feature space, and calculates and outputs the anomaly metric based on the feature offset.

9. The method for identifying anomalies in transformer insulating oil chromatographic data according to claim 1, characterized in that, The anomaly metrics of each gas output by the One-Class SVM model are multiplied by a pre-set weighting coefficient and then summed to obtain a weighted sum. The weighted sum is then mapped to an anomaly confidence level of zero to one using a mapping function calibrated for normal and anomaly samples. The anomaly confidence level is compared with a preset warning level threshold and a severity level threshold. A warning level alarm is triggered when the anomaly confidence level is greater than or equal to the warning level threshold and less than the severity level threshold. A severity level alarm is triggered when the anomaly confidence level is greater than or equal to the severity level threshold.

Images