Intelligent gas leakage detection system and method for transformer

Abstract

The invention relates to the technical field of transformer gas detection, and discloses an intelligent gas leakage detection system and method for a transformer, and the method comprises the steps: collecting multi-source operation data, such as gas pressure, oil level and temperature, carrying out the standardization processing, and building a unified feature matrix to eliminate the dimensional difference. Then extracting space-time correlation characteristics of gas leakage, and constructing a supervisible training sample in combination with historical records; a deep learning model is utilized to fuse a convolutional network and a cyclic network structure, local change features and global dynamic features are captured respectively, and common working conditions and extreme working conditions are distinguished through a multi-branch classifier. And angle subareas are divided according to the wind direction of the whole field, a correction curve between historical gas concentration and actual gas concentration is established, and space subarea correction is achieved. And finally, obtaining optimal gas leakage state evaluation through weighted fusion of the branch prediction value and the comprehensive correction value. According to the invention, an intelligent detection principle of data driving, space-time coupling and working condition self-adaption is effectively realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer gas detection, in particular to a transformer gas leakage intelligent detection system and method. BACKGROUND

[0002] With the increasing demand for transformer operation safety in the power system, accurate detection of transformer gas leakage has become a key link to ensure the stability and reliability of the power grid. Since the transformer plays a core role in power transmission and voltage conversion, the integrity of its sealing system is directly related to the oil level, insulation performance and overall operation state of the equipment.

[0003] At present, the traditional gas leakage detection method mainly relies on single parameter monitoring means, such as obtaining local information through gas pressure gauges, oil level gauges or fixed-point gas concentration detectors to determine whether there is leakage. Although these methods have the characteristics of simple structure and easy implementation, their ability shows certain limitations in complex environments. Single sensing parameter is often difficult to fully reflect the actual operating environment of the transformer. Gas leakage behavior is usually affected by temperature changes, environmental humidity, air flow disturbance and dissolved gas characteristics in oil and other factors. When only relying on the change of gas pressure or gas concentration to judge the leakage, false positives or false negatives may occur due to environmental fluctuations, instrument drift or short-term disturbances. In addition, the existing detection devices mostly use fixed threshold logic, lack of data fusion and adaptive algorithm support, limiting the comprehensive analysis and trend identification ability of multi-source information, thereby affecting the timely discovery of early micro-leakage or potential hazards. In addition, the environment of some substations is complex, with frequent changes in external temperature, humidity and wind speed, which puts higher requirements on the long-term stability and accuracy of traditional single-parameter monitoring systems. Due to the lack of intelligent analysis mechanism, the existing system cannot automatically correct the detection model according to historical data and operating characteristics, and cannot fully utilize multi-dimensional perception data to realize intelligent judgment and remote early warning function. Therefore, the current transformer gas leakage detection technology has certain deficiencies in perception dimension, environmental adaptability and intelligent analysis, and cannot fully meet the needs of modern power equipment for safe operation, intelligent diagnosis and predictive maintenance.

[0004] In summary, there is an urgent need for a transformer gas leakage detection system and method with multi-dimensional perception and intelligent analysis capability, which realizes accurate identification and intelligent early warning of leakage state through efficient fusion and deep learning modeling of multi-source data, providing more reliable protection for the safe operation of the power system. SUMMARY

[0005] In view of this, the present application proposes a transformer gas leakage intelligent detection system and method, aiming to solve the problem that the existing transformer gas leakage detection technology in the current technology has single perception dimension, poor environmental adaptability and lacks intelligent analysis capability, leading to the problem that the system is difficult to identify early leakage and potential hazards in time and accurately.

[0006] The present application proposes a transformer gas leakage intelligent detection method, comprising: Collecting multi-source operation data of the transformer sealing system, and preprocessing the multi-source operation data to generate a standardized feature matrix; Extracting the spatio-temporal correlation features of gas leakage based on the standardized feature matrix, and constructing a training sample set combined with historical operation records; Inputting the training sample set into a deep learning model to extract local and global feature representations of gas leakage, and using a multi-branch classifier to predict the gas leakage state to obtain ordinary working condition branch prediction values and extreme working condition branch prediction values; Dividing the training sample set into several angle partitions according to the overall wind direction, and constructing correction curves for each partition according to the relationship between historical gas concentration and actual gas concentration; Substituting the gas concentration prediction values of the target period into the corresponding correction curves of the target partition and its adjacent partitions to obtain multiple partition correction results, and generating a comprehensive correction value based on a weighted fusion strategy; Based on the ordinary working condition branch prediction value, the extreme working condition branch prediction value and the comprehensive correction value, the final gas leakage state is determined.

[0007] Further, when collecting multi-source operation data of the transformer sealing system and generating a standardized feature matrix, it includes: Synchronously collecting transformer internal gas pressure, oil level height, temperature distribution, humidity change, airflow speed and oil dissolved gas concentration data on the time axis; Denoising the time series of each type of data, using wavelet transform to remove high-frequency noise and retain low-frequency trend components; Normalizing the denoised data to ensure that the numerical range of each type of data is uniform to the interval 0 to 1; Splicing the normalized data into a multi-channel feature tensor according to the time step, and establishing a standardized feature matrix.

[0008] Further, when extracting the spatio-temporal correlation features of gas leakage and constructing a training sample set, it includes: Extracting the change rate of gas pressure and oil level height in the time dimension and the distribution difference in the spatial dimension in the standardized feature matrix; Combining the dynamic characteristics of temperature distribution and humidity change, and extracting the correlation features of dynamic characteristics and gas pressure; Samples are constructed based on a sliding time window, with multidimensional sensing data within the window used as input features and the gas leakage status label corresponding to the window backsight step size used as output label. The samples are labeled based on historical operation records. The labeling categories include normal state, minor leakage state, and severe leakage state.

[0009] Furthermore, when extracting local and global feature representations of gas leaks, the following are included: The training sample set is stacked in the time dimension according to a sliding time window, and the multidimensional sensing data is channel-stitched in the feature dimension to form a three-dimensional temporal tensor. Based on the extraction of local features by three-dimensional convolutional neural network, zero padding is used in the first layer to keep the tensor size unchanged, and multi-scale convolution kernels and dilated convolution are combined in the middle layer. The kernel length of the multi-scale convolution kernel is one or more of 3×3×3, 5×5×5, and 7×7×7. Batch normalization and non-linear activation functions are set sequentially after the convolutional layer, and deep degradation is avoided based on residual connections; By inputting local features into a bidirectional long short-term memory network to model global dynamic dependencies, the long-term evolution of gas leakage behavior can be captured. A self-attention mechanism is introduced into the feature sequence output by the bidirectional long short-term memory network to determine the relevance score of the features at each time step, perform weighted aggregation, and generate a context vector.

[0010] Furthermore, when predicting the state of gas leaks using a multi-branch classifier, the following is included: Two independent multilayer perceptron branches are set up, namely the normal operating condition branch and the extreme operating condition branch; Each branch contains at least three fully connected layers, with a non-linear activation function set after each fully connected layer, and a random deactivation mechanism introduced between layers to suppress overfitting; During the training phase, the working condition labels based on the samples only update the parameters of the corresponding branch, while freezing the parameters of the other branch. During the prediction phase, the context vector is fed forward into two branches to generate predicted values ​​for normal operating conditions and predicted values ​​for extreme operating conditions. The classification error is constrained by the cross-entropy loss function, and the contributions of different class samples are balanced by class weights.

[0011] Furthermore, when constructing the calibration curves for each partition, the following steps are included: Using the overall wind direction as the angle variable, the wind direction circle is divided into angular zones of equal width, and the outer extension of the zones is divided into several rings according to the gas concentration level to form a hierarchical structure. Within each angular partition, a monotonic regression fitting is performed with historical gas concentration as the independent variable and actual gas concentration as the dependent variable to generate a calibration curve; A smoothing constraint is introduced during the calibration curve fitting process to ensure that the curve is continuous and differentiable.

[0012] Furthermore, when generating the comprehensive correction value, the following are included: Determine the target angle zone and its adjacent angle zones corresponding to the wind direction of the entire field during the target period, and substitute the predicted gas concentration values ​​of the target period into the corresponding correction curves to obtain multiple zone correction results; The inverse distance weight is calculated based on the angle between the central axis of each zone and the overall wind direction, and the zone correction results are weighted and summed to generate a comprehensive correction value. A weight normalization mechanism is introduced in the weighted summation process to ensure that the sum of all weights is 1.

[0013] Furthermore, when determining the final gas leak status assessment value, the following are included: Performance metrics, including accuracy, recall, and F1 score, are calculated for the branch prediction values ​​under normal operating conditions, the branch prediction values ​​under extreme operating conditions, and the comprehensive correction values. Subjective and objective weights are determined based on performance indicators, and combined weights are generated using weight normalization and non-negativity constraints. The combined weights are then weighted and integrated with the predicted values ​​for normal operating conditions, extreme operating conditions, and the comprehensive correction value to generate the final gas leakage status assessment value.

[0014] Compared with existing technologies, the advantages of this invention are as follows: By collecting and standardizing multi-source operating data of the transformer sealing system (including gas concentration, temperature, pressure, environmental parameters, etc.), it effectively avoids the problems of missed detection and misjudgment caused by single parameter monitoring, making the detection results more comprehensive and robust. Secondly, by extracting spatiotemporal correlation features based on the standardized feature matrix and constructing a training sample set in combination with historical operating records, it can fully capture the patterns of gas leakage in terms of temporal variation and spatial distribution, realizing dynamic modeling and trend identification of the leakage process, and improving the model's sensitivity to leakage behavior in complex environments. Furthermore, by using a deep learning model to extract local and global feature representations of gas leakage, and outputting prediction results under normal and extreme operating conditions through a multi-branch classifier, the model can take into account the feature differences of different operating scenarios, thereby enhancing its generalization ability and stability to abnormal states. In addition, by introducing overall wind direction information to divide the sample set into several angular partitions, and constructing correction curves for each partition based on the relationship between historical and real-time gas concentrations, it can fully consider the impact of environmental wind direction changes on gas diffusion, improving the spatial accuracy of prediction. Finally, by weighted fusion of multiple partition correction results, a comprehensive correction value is generated, which is then combined with multi-branch prediction results to determine the final gas leakage state, thus realizing dynamic correction and adaptive optimization of the prediction results.

[0015] On the other hand, this application also provides an intelligent gas leakage detection system for transformers, comprising: The acquisition module is configured to acquire multi-source operating data of the transformer sealing system and preprocess the multi-source operating data to generate a standardized feature matrix. The feature extraction module is configured to extract the spatiotemporal correlation features of gas leaks based on a standardized feature matrix, and to construct a training sample set by combining historical operation records; The modeling module is configured to input the training sample set into the deep learning model to extract local and global feature representations of gas leaks; The classification module is configured to predict the gas leak status through a multi-branch classifier, and obtain the branch prediction values ​​for normal operating conditions and the branch prediction values ​​for extreme operating conditions. The partitioning module is configured to divide the training sample set into several angular partitions based on the overall wind direction, and to construct the calibration curve for each partition based on the relationship between historical gas concentration and actual gas concentration. The correction module is configured to substitute the predicted gas concentration value for the target time period into the correction curves corresponding to the target partition and its adjacent partitions to obtain multiple partition correction results, and generate a comprehensive correction value through a weighted fusion strategy. The assessment module is configured to calculate the final gas leakage status assessment value based on the normal operating condition branch prediction value, the extreme operating condition branch prediction value, and the comprehensive correction value.

[0016] Furthermore, the acquisition module collects multi-source operating data from the transformer sealing system through multiple sensors, including pressure sensors, oil level sensors, temperature sensors, humidity sensors, airflow velocity sensors, and dissolved gas concentration sensors in the oil.

[0017] It is understood that the intelligent gas leakage detection system and method for transformers in the above embodiments of this application have the same beneficial effects, and will not be described again. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating an intelligent gas leakage detection method for transformers provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating an intelligent gas leakage detection method for transformers provided in an embodiment of the present invention. Figure 3 This is a functional block diagram of an intelligent gas leakage detection system for transformers provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] like Figures 1-2 As shown in some embodiments of this application, this embodiment provides a smart detection method for gas leakage in transformers, including: Step S100: Collect multi-source operating data of the transformer sealing system and preprocess the multi-source operating data to generate a standardized feature matrix.

[0021] Specifically, when collecting multi-source operating data of the transformer sealing system and generating a standardized feature matrix, the process includes: synchronously collecting data on internal transformer gas pressure, oil level, temperature distribution, humidity changes, airflow velocity, and dissolved gas concentration in the oil on the time axis; denoising the time series of each type of data, using wavelet transform to remove high-frequency noise and retain low-frequency trend components; normalizing the denoised data to ensure that the numerical range of each type of data is uniformly within the 0 to 1 range; and concatenating the normalized data into a multi-channel feature tensor according to the time step to establish a standardized feature matrix.

[0022] Understandably, multi-source data fusion and standardized feature modeling provide a high-quality, computable input data foundation for subsequent intelligent gas leak detection. The core idea is to uniformly process the operational data from multiple sensor channels in the transformer sealing system across time and scale to construct a representative feature matrix, thereby more comprehensively reflecting system state changes. First, by simultaneously acquiring multi-dimensional physical parameters such as gas pressure, oil level, temperature distribution, humidity changes, airflow velocity, and dissolved gas concentration in the oil along the time axis, the dynamic characteristics of the transformer sealing system can be characterized from different physical levels. This multi-source data fusion mechanism overcomes the limitations of traditional single-sensor monitoring, enabling the system to simultaneously perceive multiple coupling effects caused by gas leaks, such as the linkage between gas pressure fluctuations and oil level changes. Second, wavelet transform for denoising the time series is a key step in signal preprocessing in this technical solution. Wavelet transform can simultaneously perform analysis in both the time and frequency domains, effectively separating high-frequency noise from low-frequency trends, thereby removing random noise signals such as external environmental interference and instrument jitter, while preserving the slow-changing trend characteristics during the gas leak process. This makes the data relied upon for subsequent feature extraction and model training smoother, more stable, and physically meaningful. Furthermore, by normalizing the denoised data, the numerical ranges of different types of parameters are unified to the 0-1 interval, resolving the differences in dimensions and orders of magnitude between multi-source data. This step not only enhances the comparability of the data and the convergence of model training but also avoids the bias caused by certain high-amplitude signals on the overall feature space. Finally, concatenating the normalized data according to the time step forms a multi-channel feature tensor, essentially establishing a temporally-series high-dimensional feature representation structure. This structure retains the dynamic information of the time series while reflecting the correlation characteristics between different sensing parameters. Through this tensor processing, subsequent deep learning models can extract features with a unified structural input, thereby achieving global perception and efficient identification of gas leakage patterns.

[0023] It can be seen that by synchronously acquiring and standardizing multi-source operational data of the transformer sealing system, the data consistency and information integrity of gas leak detection have been improved. Firstly, by synchronously acquiring multi-dimensional monitoring parameters such as gas pressure, oil level, temperature distribution, humidity changes, airflow velocity, and dissolved gas concentration in the oil along the time axis, the internal operating status and sealing environment characteristics of the transformer can be comprehensively reflected. Compared to traditional detection methods that rely solely on a single gas concentration or pressure parameter, this multi-source fusion method significantly enhances the ability to capture leak signs, avoiding misjudgments or missed detections due to abnormal local parameters. Secondly, by performing wavelet transform denoising on various time series data, high-frequency components such as external electromagnetic interference, environmental fluctuations, and instrument noise are effectively removed, while retaining the slowly changing trend characteristics during gas leakage. This process significantly improves the signal-to-noise ratio of the data, enabling subsequent feature extraction and model training to focus on real signals related to leakage, thereby improving the accuracy and stability of the detection results. Furthermore, by normalizing the denoised multi-source data, different physical quantities are unified into a standardized range of 0 to 1, avoiding feature bias caused by different units and numerical ranges. This not only improves the comparability between different sensor data but also provides balanced input data for deep learning models, helping to accelerate model training convergence and improve generalization performance. Finally, the normalized data are concatenated into a multi-channel feature tensor according to the time step to establish a standardized feature matrix. This not only preserves the dynamic changes of the time series but also reflects the coupling relationship between different monitoring parameters. This structure provides a unified data representation for subsequent spatial-temporal correlation analysis of gas leakage characteristics, enabling the model to capture the comprehensive feature patterns of gas leakage in a multi-dimensional feature space.

[0024] Step S200: Extract the spatiotemporal correlation features of gas leaks based on the standardized feature matrix, and construct a training sample set by combining historical operation records.

[0025] Specifically, when extracting the spatiotemporal correlation features of gas leaks and constructing a training sample set, the process includes: extracting the rate of change of gas pressure and oil level in the time dimension and the distribution differences in the spatial dimension from the standardized feature matrix; combining the dynamic characteristics of temperature distribution and humidity changes, and extracting the correlation features between the dynamic characteristics and gas pressure; constructing samples based on a sliding time window, using the multidimensional sensing data within the window as input features and the gas leak status label corresponding to the window's backward step size as the output label; and labeling the samples based on historical operation records, where the labeling categories include normal state, minor leak state, and severe leak state.

[0026] Understandably, calculating the rate of change in the time dimension reveals the dynamic fluctuation characteristics of the system's internal parameters; calculating the distribution differences in the spatial dimension reflects the non-uniform diffusion characteristics of gas leakage at different sensing locations. The rate of change in time reflects the "speed" information of the leakage trend, while the spatial differences correspond to the "range" information of the leakage impact. Combining the two allows for a two-dimensional dynamic modeling of the leakage process. Secondly, combining the dynamic characteristics of temperature distribution and humidity changes and extracting their correlation characteristics with gas pressure allows for the description of the system's thermal and humidity changes caused by leakage using the coupling relationship between multiple physical parameters. Gas leakage is usually accompanied by oil-gas interface disturbances, temperature gradient changes, and abnormal humidity fluctuations. Therefore, by analyzing the correlation coefficients and time-series response characteristics between these parameters, the abnormal linkage characteristics of the system caused by leakage can be effectively captured, thereby enhancing the model's sensitivity and discrimination ability. Thirdly, constructing training samples through a sliding time window mechanism embodies the idea of ​​time series modeling. The sliding window can not only retain historical information and capture short-term changing trends, but also map the dynamic operating characteristics of the transformer into time-dependent input samples, enabling the model to learn the temporal laws of leakage evolution. Multidimensional features within the window serve as input, while the leakage state corresponding to the window's backsight step size serves as the output label, thus establishing a temporal mapping relationship between input features and result states. Finally, sample labeling based on historical running records is the core step in achieving supervised learning. By dividing historical data into three states—normal, minor leakage, and severe leakage—not only is hierarchical modeling of leakage levels achieved, but reliable training supervision signals are also provided for subsequent deep learning models, enabling them to distinguish and predict under different leakage scenarios.

[0027] As can be seen, by extracting the temporal rate of change and spatial distribution differences of gas pressure and oil level from the standardized feature matrix, this method can simultaneously capture the temporal evolution trend and spatial diffusion characteristics of the leakage process. The temporal rate of change reflects the rate of leakage occurrence and development, while the spatial distribution differences reveal the non-uniformity of the leakage among different monitoring points. This dual-dimensional feature extraction method enables the system to identify early leakage signals and track their propagation paths, thereby improving detection sensitivity and response speed. Secondly, by combining the dynamic characteristics of temperature distribution and humidity changes and extracting their correlation features with gas pressure, the coupling effect between different physical parameters during the leakage process can be reflected. For example, gas leaks are often accompanied by temperature fluctuations, increased humidity, and oil-gas interface disturbances. By analyzing the dynamic correlation between these parameters, the abnormal linkage behavior of the system caused by the leak can be effectively revealed. This not only enhances the ability to identify complex leakage scenarios but also improves the robustness and adaptability of the detection algorithm in multi-physical environments. In addition, by constructing training samples through a sliding time window mechanism, the temporal expression of gas leakage features is realized, enabling the model to learn the temporal evolution of leakage events. This mechanism endows the samples with continuity and dynamism, helping deep learning models identify short-term fluctuations, trend changes, and periodic features, thereby significantly improving the model's prediction accuracy for leakage states at different stages. Finally, by combining historical operation records for sample annotation, leakage states are divided into three categories: normal, minor leakage, and severe leakage. This not only achieves multi-level fault identification but also provides accurate supervision signals for subsequent models. Through this hierarchical annotation mechanism, the model can learn typical characteristics under different leakage intensities during training, thus achieving quantitative identification and risk-stratified early warning in actual detection.

[0028] Step S300: Input the training sample set into the deep learning model to extract the local and global feature representations of gas leakage, and use a multi-branch classifier to predict the gas leakage state to obtain the branch prediction values ​​for normal working conditions and the branch prediction values ​​for extreme working conditions.

[0029] Specifically, the extraction of local and global feature representations of gas leaks includes: stacking the training sample set in a sliding time window along the time dimension, and performing channel stitching on the multidimensional sensing data in the feature dimension to form a three-dimensional temporal tensor; extracting local features based on a three-dimensional convolutional neural network, using zero padding in the first layer to maintain the tensor size, and employing a combination of multi-scale convolutional kernels and dilated convolutions in the intermediate layers, with the kernel length of the multi-scale convolutional kernels taking one or more of 3×3×3, 5×5×5, and 7×7×7; setting batch normalization and nonlinear activation functions sequentially after the convolutional layers, and avoiding deep degradation based on residual connections; inputting the local features into a bidirectional long short-term memory network to model global dynamic dependencies and capture the long-term evolution of gas leak behavior; introducing a self-attention mechanism into the feature sequence output by the bidirectional long short-term memory network to determine the correlation scores of features at each time step and weighted converge them to generate a context vector.

[0030] Specifically, when predicting the gas leak status using a multi-branch classifier, the following steps are taken: Two independent multilayer perceptron branches are set up: a normal operating condition branch and an extreme operating condition branch. Each branch contains at least three fully connected layers, with a non-linear activation function applied after each fully connected layer, and a random deactivation mechanism introduced between layers to suppress overfitting. During the training phase, the operating condition label based on the sample only updates the parameters of the corresponding branch, while freezing the parameters of the other branch. During the prediction phase, the context vector is fed forward into both branches to generate predicted values ​​for the normal operating condition branch and the extreme operating condition branch, respectively. The classification error is constrained based on the cross-entropy loss function, and the contributions of different class samples are balanced by class weights.

[0031] Understandably, in the local feature extraction stage, by stacking training samples in a sliding window along the time dimension and splicing multi-source sensing data along the feature dimension to form a three-dimensional time-series tensor, unified modeling of time, space, and multi-dimensional signal channels is achieved. Based on this three-dimensional tensor, a three-dimensional convolutional neural network (3D-CNN) is used to extract local features of gas leaks. The technical principle is that the convolutional kernel slides in three-dimensional space, enabling simultaneous perception of the changing trend of the time series and the local correlation of spatial features. By combining multi-scale convolutional kernels (3×3×3, 5×5×5, 7×7×7) with dilated convolutions, multi-level features can be extracted under different receptive fields, thereby enhancing the model's ability to perceive the diffusion rate and local anomalies in the leak area. At the same time, residual connections and batch normalization ensure stable training of deep networks, avoiding gradient vanishing and feature degradation. Secondly, in the global feature modeling stage, local convolutional features are input into a bidirectional long short-term memory network (Bi-LSTM) to achieve long-term dynamic dependency modeling of gas leak behavior. Bi-LSTM can simultaneously capture the dependencies between consecutive time steps in a time series, thus more accurately reflecting the temporal evolution of leakage from an initial trace to continuous escalation. Furthermore, a self-attention mechanism is introduced into the feature sequence output by the network. By calculating the correlation scores of features at each time step, important features are weighted and aggregated to generate a context vector. This mechanism enables the model to "pay attention" to key moments (such as leakage mutation points or environmental anomalies), thereby improving the discriminativeness and robustness of the global feature representation. Finally, in the classification and prediction stage, a multi-branch classifier structure consisting of a normal operating condition branch and an extreme operating condition branch is designed. The principle is that two independent multilayer perceptron branches learn the feature distribution under different operating conditions, enabling the model to form a differentiated decision-making mechanism for gas leakage behavior under different environmental conditions. During training, a label-driven selective parameter update strategy is used to optimize parameters only for the corresponding branches, ensuring that the two branches do not interfere with each other and each focuses on feature learning for the target working condition. In the prediction phase, the context vector is input into both branches of the network to output predicted values ​​for normal and extreme working conditions, thus enabling parallel judgment and fusion evaluation of multiple scenarios. By introducing cross-entropy loss constraints and class weight adjustments, the influence of samples with different leakage levels during training is further balanced, improving the model's classification accuracy and generalization ability under imbalanced samples.

[0032] It can be seen that by leveraging multiple mechanisms such as 3D convolution to capture local features, recurrent network modeling of temporal dependencies, self-attention to focus on key features, and multi-branch structure to achieve scene differentiation, deep semantic modeling and multi-condition adaptive identification of gas leakage signals are achieved, providing a highly robust, accurate, and generalizable intelligent analysis framework for transformer gas leakage detection.

[0033] Step S400: Divide the training sample set into several angular partitions according to the overall wind direction, and construct the calibration curve for each partition based on the relationship between historical gas concentration and actual gas concentration.

[0034] Specifically, when constructing the calibration curves for each zone, the following steps are taken: using the overall wind direction as the angle variable, the wind direction circumference is divided into angular zones of equal width, and the outer extension of the zones is divided into several rings according to the gas concentration level to form a hierarchical structure; within each angular zone, a monotonic regression fitting is performed with historical gas concentration as the independent variable and actual gas concentration as the dependent variable to generate the calibration curve; and a smoothing constraint is introduced during the calibration curve fitting process to ensure that the curve is continuous and differentiable.

[0035] Understandably, in the sample partitioning modeling stage, the method uses the overall wind direction as the angular variable, dividing the 360° wind direction circle into several partitions of equal angular width. Within each partition, it further extends outwards to form rings based on gas concentration levels, creating a hierarchical spatial structure. The underlying principle is to utilize the anisotropic characteristics of gas diffusion: when the wind direction changes, the diffusion path, velocity, and concentration distribution of the leaked gas exhibit significant directional differences. By combining wind direction angular partitioning with concentration ring division, a partitioned representation of the leak diffusion spatial field can be achieved, providing a precise spatial positioning basis for subsequent partitioned correction. Secondly, in the partitioned correction curve construction stage, the method uses historical gas concentration as the independent variable and actual gas concentration as the dependent variable, performing monotonic regression fitting within each angular partition to generate local correction curves. The technical principle of monotonic regression is to ensure that the functional relationship between predicted and actual values ​​conforms to physical laws, i.e., gas concentration shows a monotonic upward trend as the degree of leakage increases. By performing this monotonic mapping on historical data, the diffusion patterns under different wind directions and the relationship with observation errors can be reflected, realizing a data-driven regional correction model. Furthermore, a smoothing constraint term is introduced during curve fitting to ensure that the calibration curves for each zone are continuous and differentiable at angular boundaries. The principle behind this design is to prevent abrupt changes or discontinuities between zones, ensuring a smooth transition of the concentration calibration model during wind direction changes, thereby avoiding unreasonable prediction jumps at wind direction shifts or boundary zones. The introduction of the smoothing constraint enables the calibration model to not only possess local adaptability but also maintain global consistency and differentiability, facilitating subsequent weighted fusion or gradient optimization.

[0036] It can be seen that, through multi-dimensional mechanisms such as wind direction angle zoning, hierarchical spatial structure modeling, monotonic regression fitting, and smooth constraint optimization, directional modeling and adaptive correction of gas diffusion characteristics are achieved. This principle enables the system to automatically adjust prediction results under different wind field conditions, effectively compensating for error shifts caused by wind direction changes, thereby significantly improving the spatial accuracy and environmental robustness of gas leak detection.

[0037] Step S500: Substitute the predicted gas concentration values ​​for the target time period into the correction curves corresponding to the target partition and its adjacent partitions to obtain multiple partition correction results, and generate a comprehensive correction value based on a weighted fusion strategy.

[0038] Specifically, the process of generating the comprehensive correction value includes: determining the target angle zone and its adjacent angle zones corresponding to the overall wind direction during the target time period, and substituting the predicted gas concentration values ​​for the target time period into the corresponding correction curves to obtain multiple zone correction results; calculating the inverse distance weight based on the angle between the central axis of each zone and the overall wind direction, and weighting and summing the zone correction results to generate the comprehensive correction value; and introducing a weight normalization mechanism in the weighted summation process to ensure that the sum of all weights is 1.

[0039] It can be understood that, based on the principle of spatial correlation of gas diffusion, it is believed that under certain meteorological conditions (especially under specific wind directions), there is a coupling relationship between the gas concentration changes of the target zone and its adjacent zones. Since gases exhibit directional diffusion characteristics under the dominant influence of wind direction, the degree of influence of different angle zones on the target zone is not consistent. Therefore, it is necessary to establish a calibration curve model for each zone to reflect the deviation between the predicted and actual gas concentration values. Secondly, using the principle of wind direction angle zoning, the target angle zone and its adjacent zones corresponding to the overall wind direction are determined within a specific time period. The calibration curves of these zones are used to perform multi-zone calibration calculations on the predicted values ​​for the target time period. This process reflects the dynamic adaptability of gas concentration distribution to changes with wind direction, enabling the model to flexibly adjust the correction range according to real-time meteorological conditions. Finally, the inverse distance weighted fusion (IDW) principle is used to generate a comprehensive calibration value. Specifically, the system calculates the angle between the overall wind direction and the central axis of each zone, and determines the inverse distance weight based on the angle size. The smaller the angle, the closer the gas propagation direction is to the prevailing wind direction, and the greater the influence on the target zone, thus resulting in a higher weight value. By weighted summation of the correction results from multiple partitions and the introduction of a weight normalization mechanism, the sum of all weights is made equal to 1, thereby obtaining a comprehensive correction value with clear physical meaning and stable numerical value.

[0040] It can be seen that by substituting the predicted gas concentration values ​​into the correction curves of the target zone and its adjacent zones, the spatial correlation information between zones can be fully utilized. Adjacent zones typically exhibit a certain concentration coupling relationship during gas diffusion. This design can comprehensively consider the combined influence of local and surrounding areas, effectively reducing local errors caused by single-region model bias and improving the accuracy of the overall correction results. Secondly, adopting a zone weighting mechanism based on wind direction angle makes the model more sensitive to directional changes in gas diffusion. By calculating the angle between the overall wind direction and the central axis of each zone, and allocating inverse distance weights accordingly, the actual contribution of different zones to the target area can be reflected, making the weighting process more consistent with the physical laws of gas propagation along the prevailing wind direction, thereby enhancing the meteorological adaptability and dynamic response capability of the prediction results. Finally, by introducing a weight normalization mechanism during the weighted summation process, ensuring that the sum of all zone weights is 1, the weight imbalance problem caused by wind direction shifts or abnormal zone data can be avoided. This mechanism guarantees the numerical stability and computational rationality of the fusion process, ensuring that the final comprehensive correction value maintains consistent scale characteristics and comparability under different environmental conditions.

[0041] Step S500: Determine the final gas leakage state based on the branch prediction values ​​of normal operating conditions, the branch prediction values ​​of extreme operating conditions, and the comprehensive correction value.

[0042] Specifically, determining the final gas leak status assessment value includes: calculating performance indicators, including accuracy, recall, and F1 score, for the predicted values ​​of the normal operating condition branch, the predicted values ​​of the extreme operating condition branch, and the comprehensive correction value; determining subjective and objective weights based on the performance indicators, and generating combined weights using weight normalization and non-negativity constraints; and weighting and fusing the combined weights with the predicted values ​​of the normal operating condition branch, the predicted values ​​of the extreme operating condition branch, and the comprehensive correction value to generate the final gas leak status assessment value.

[0043] Understandably, the prediction phase simultaneously considers two prediction models: one for normal operating conditions and one for extreme operating conditions. These are combined with a comprehensive correction value from preceding stages, thus achieving a multi-dimensional information fusion prediction framework. The normal operating condition model reflects the operational characteristics of the gas system under stable conditions, while the extreme operating condition model captures behavioral differences under abnormal or abrupt conditions. Integrating both models with the physically corrected comprehensive value effectively compensates for the bias of a single model under specific environments, resulting in a more comprehensive adaptability to different operating conditions. Secondly, by calculating performance indicators (accuracy, recall, F1 score) from the prediction results of each branch and the comprehensive correction value, the system can quantify the performance differences of different models on historical samples. This indicator-based quantification process forms the basis for adaptive weight allocation, ensuring that weight allocation is not empirically set but derived from data-driven performance evaluation, thereby enhancing the objectivity and scientific rigor of model fusion. Thirdly, a joint optimization mechanism of subjective and objective weights is introduced during weight determination. Subjective weights can reflect human experience or prior preferences in model design, while objective weights are dynamically adjusted based on performance indicators. By normalizing weights and imposing nonnegativity constraints, the stability and interpretability of the final weight vector are ensured, avoiding weight imbalance or negative contributions, thus providing a robust mathematical foundation for the weighted fusion process. Finally, the generated combined weights are weighted and fused with the three input predictions (normal operating condition, extreme operating condition, and comprehensive correction value) to obtain the final gas leakage status assessment value. This fusion is essentially a multi-model ensemble decision-making mechanism that can retain the unique advantages of each branch model while suppressing the influence of single-model errors, thereby achieving higher prediction reliability and judgment accuracy.

[0044] It can be seen that by simultaneously introducing predicted values ​​from normal operating conditions, extreme operating conditions, and a comprehensive correction value, this scheme can integrate multi-source prediction information at the decision-making level. The normal operating condition branch is suitable for stable operating scenarios, the extreme operating condition branch is more sensitive to abnormal states, and the comprehensive correction value compensates for prediction errors at the physical level. The synergistic integration of these three factors ensures that the final result possesses both high accuracy for routine operation and robustness under sudden or extreme conditions, thus comprehensively covering gas leakage characteristics under different operating conditions. Secondly, by calculating performance indicators such as accuracy, recall, and F1 score for each prediction result, the system can dynamically evaluate the performance of each model under the current sample or operating condition. Subjective and objective weights determined based on performance indicators eliminate the reliance on human experience for model weight allocation, allowing for automatic adjustment by a data-driven performance feedback mechanism, thereby achieving adaptive optimization. This mechanism can automatically strengthen high-performing branch predictions and weaken the influence of weaker branches based on environmental changes or data characteristics, ensuring that the evaluation results always maintain optimal responsiveness. Furthermore, by introducing weight normalization and non-negativity constraint mechanisms, the contribution ratio of each prediction result is ensured to be reasonable and stable during the weighted fusion process, avoiding excessive amplification of a certain prediction branch or negative weight interference. This not only improves the numerical stability of the calculation but also makes the final evaluation results more continuous and interpretable under different operating conditions. Finally, the final gas leakage state evaluation value obtained by combined weighted fusion integrates the advantages of multi-dimensional information, effectively reducing biases caused by single-model misjudgment, overfitting, or data noise. Compared with traditional single-branch prediction methods, this scheme exhibits higher anti-interference and stability under complex meteorological conditions, differences in multi-source sensing data, and nonlinear changes in the system.

[0045] In the above embodiments, by collecting and standardizing multi-source operational data of the transformer sealing system (including gas concentration, temperature, pressure, environmental parameters, etc.), the problems of missed detection and misjudgment caused by single parameter monitoring are effectively avoided, making the detection results more comprehensive and robust. Secondly, by extracting spatiotemporal correlation features based on the standardized feature matrix and constructing a training sample set in combination with historical operation records, the patterns of gas leakage in terms of temporal variation and spatial distribution can be fully captured, realizing dynamic modeling and trend identification of the leakage process and improving the model's sensitivity to leakage behavior in complex environments. Furthermore, by using a deep learning model to extract local and global feature representations of gas leakage, and by using a multi-branch classifier to output prediction results under normal and extreme operating conditions, the model can take into account the feature differences of different operating scenarios, thereby enhancing its generalization ability and stability to abnormal states. In addition, by introducing overall wind direction information to divide the sample set into several angular partitions, and constructing correction curves for each partition based on the relationship between historical and real-time gas concentrations, the influence of environmental wind direction changes on gas diffusion can be fully considered, improving the spatial accuracy of the prediction. Finally, by weighted fusion of multiple partition correction results, a comprehensive correction value is generated, which is then combined with multi-branch prediction results to determine the final gas leakage state, thus realizing dynamic correction and adaptive optimization of the prediction results.

[0046] In another preferred embodiment based on the above embodiments, such as Figure 3 As shown, this embodiment provides an intelligent gas leakage detection system for transformers, including: a data acquisition module, a feature extraction module, a modeling module, a classification module, a partitioning module, a calibration module, and an evaluation module.

[0047] Specifically, the acquisition module is configured to acquire multi-source operating data of the transformer sealing system and preprocess the multi-source operating data to generate a standardized feature matrix; the feature extraction module is configured to extract the spatiotemporal correlation features of gas leakage based on the standardized feature matrix and construct a training sample set in combination with historical operating records; the modeling module is configured to input the training sample set into a deep learning model to extract local and global feature representations of gas leakage; the classification module is configured to predict the gas leakage state through a multi-branch classifier to obtain the branch prediction values ​​for normal operating conditions and the branch prediction values ​​for extreme operating conditions; the partitioning module is configured to divide the training sample set into several angular partitions according to the overall wind direction and construct the correction curves for each partition based on the relationship between historical gas concentration and actual gas concentration; the correction module is configured to substitute the predicted gas concentration values ​​for the target time period into the correction curves corresponding to the target partition and its adjacent partitions to obtain multiple partition correction results, and generate a comprehensive correction value through a weighted fusion strategy; the evaluation module is configured to calculate the final gas leakage state evaluation value based on the branch prediction values ​​for normal operating conditions, the branch prediction values ​​for extreme operating conditions, and the comprehensive correction value.

[0048] Specifically, the acquisition module collects multi-source operating data from the transformer sealing system through multiple sensors, including pressure sensors, oil level sensors, temperature sensors, humidity sensors, airflow velocity sensors, and dissolved gas concentration sensors in the oil.

[0049] It is understood that the intelligent gas leakage detection system and method for transformers in the above embodiments of this application have the same beneficial effects, and will not be described again.

[0050] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0051] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0053] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A smart detection method for gas leakage in transformers, characterized in that, include: Collect multi-source operating data of the transformer sealing system and preprocess the multi-source operating data to generate a standardized feature matrix; The spatiotemporal correlation features of gas leaks are extracted based on the standardized feature matrix, and a training sample set is constructed by combining historical operation records. The training sample set is input into the deep learning model to extract local and global feature representations of gas leaks, and a multi-branch classifier is used to predict the gas leak state to obtain the branch prediction values ​​for normal working conditions and the branch prediction values ​​for extreme working conditions. The training sample set is divided into several angular partitions based on the overall wind direction, and a calibration curve for each partition is constructed based on the relationship between historical gas concentration and actual gas concentration. The predicted gas concentration values ​​for the target period are substituted into the correction curves corresponding to the target partition and its adjacent partitions to obtain multiple partition correction results, and a comprehensive correction value is generated based on a weighted fusion strategy. The final gas leakage state is determined based on the branch prediction values ​​for normal operating conditions, the branch prediction values ​​for extreme operating conditions, and the comprehensive correction value.

2. The intelligent gas leakage detection method for transformers as described in claim 1, characterized in that, When collecting multi-source operating data of the transformer sealing system and generating a standardized feature matrix, the following steps are included: Data on internal gas pressure, oil level, temperature distribution, humidity change, airflow velocity, and dissolved gas concentration in the oil are collected synchronously on the time axis. Denoising is performed on the time series of each type of data, and wavelet transform is used to remove high-frequency noise while retaining low-frequency trend components. The denoised data is normalized to ensure that the numerical range of all data is uniformly between 0 and 1. The normalized data is concatenated into a multi-channel feature tensor according to the time step, and a standardized feature matrix is ​​established.

3. The intelligent gas leakage detection method for transformers according to claim 2, characterized in that, When extracting the spatiotemporal correlation features of gas leaks and constructing a training sample set, the following steps are included: Extract the rate of change of gas pressure and oil level in the time dimension and the spatial distribution differences in the standardized feature matrix; By combining the dynamic characteristics of temperature distribution and humidity changes, the correlation features between the dynamic characteristics and gas pressure are extracted. Samples are constructed based on a sliding time window, with multidimensional sensing data within the window used as input features and the gas leakage status label corresponding to the window backsight step size used as output label. The samples are labeled based on historical operation records. The labeling categories include normal state, minor leakage state, and severe leakage state.

4. The intelligent gas leakage detection method for transformers as described in claim 1, characterized in that, When extracting local and global feature representations of gas leaks, the following are included: The training sample set is stacked in the time dimension according to a sliding time window, and the multidimensional sensing data is channel-stitched in the feature dimension to form a three-dimensional temporal tensor. Based on the extraction of local features by three-dimensional convolutional neural network, zero padding is used in the first layer to keep the tensor size unchanged, and multi-scale convolution kernels and dilated convolution are combined in the middle layer. The kernel length of the multi-scale convolution kernel is one or more of 3×3×3, 5×5×5, and 7×7×7. Batch normalization and non-linear activation functions are set sequentially after the convolutional layer, and deep degradation is avoided based on residual connections; By inputting local features into a bidirectional long short-term memory network to model global dynamic dependencies, the long-term evolution of gas leakage behavior can be captured. A self-attention mechanism is introduced into the feature sequence output by the bidirectional long short-term memory network to determine the relevance score of the features at each time step, perform weighted aggregation, and generate a context vector.

5. The intelligent gas leakage detection method for transformers as described in claim 4, characterized in that, When predicting the state of a gas leak using a multi-branch classifier, the following are included: Two independent multilayer perceptron branches are set up, namely the normal operating condition branch and the extreme operating condition branch; Each branch contains at least three fully connected layers, with a non-linear activation function set after each fully connected layer, and a random deactivation mechanism introduced between layers to suppress overfitting; During the training phase, the working condition labels based on the samples only update the parameters of the corresponding branch, while freezing the parameters of the other branch. During the prediction phase, the context vector is fed forward into two branches to generate predicted values ​​for normal operating conditions and predicted values ​​for extreme operating conditions. The classification error is constrained by the cross-entropy loss function, and the contributions of different class samples are balanced by class weights.

6. The intelligent gas leakage detection method for transformers as described in claim 1, characterized in that, When constructing the calibration curves for each partition, the following steps are included: Using the overall wind direction as the angle variable, the wind direction circle is divided into angular zones of equal width, and the outer extension of the zones is divided into several rings according to the gas concentration level to form a hierarchical structure. Within each angular partition, a monotonic regression fitting is performed with historical gas concentration as the independent variable and actual gas concentration as the dependent variable to generate a calibration curve; A smoothing constraint is introduced during the calibration curve fitting process to ensure that the curve is continuous and differentiable.

7. The intelligent gas leakage detection method for transformers as described in claim 1, characterized in that, When generating the overall correction value, the following are included: Determine the target angle zone and its adjacent angle zones corresponding to the wind direction of the entire field during the target period, and substitute the predicted gas concentration values ​​of the target period into the corresponding correction curves to obtain multiple zone correction results; The inverse distance weight is calculated based on the angle between the central axis of each zone and the overall wind direction, and the zone correction results are weighted and summed to generate a comprehensive correction value. A weight normalization mechanism is introduced in the weighted summation process to ensure that the sum of all weights is 1.

8. The intelligent gas leakage detection method for transformers as described in claim 1, characterized in that, When determining the final gas leak status assessment value, the following should be included: Performance metrics, including accuracy, recall, and F1 score, are calculated for the branch prediction values ​​under normal operating conditions, the branch prediction values ​​under extreme operating conditions, and the comprehensive correction values. Subjective and objective weights are determined based on performance indicators, and combined weights are generated using weight normalization and non-negativity constraints. The combined weights are then weighted and integrated with the predicted values ​​for normal operating conditions, extreme operating conditions, and the comprehensive correction value to generate the final gas leakage status assessment value.

9. A smart gas leak detection system for transformers, applicable to the smart gas leak detection method for transformers as described in any one of claims 1-8, characterized in that, include: The acquisition module is configured to acquire multi-source operating data of the transformer sealing system and preprocess the multi-source operating data to generate a standardized feature matrix. The feature extraction module is configured to extract the spatiotemporal correlation features of gas leaks based on a standardized feature matrix, and to construct a training sample set by combining historical operation records; The modeling module is configured to input the training sample set into the deep learning model to extract local and global feature representations of gas leaks; The classification module is configured to predict the gas leak status through a multi-branch classifier, and obtain the branch prediction values ​​for normal operating conditions and the branch prediction values ​​for extreme operating conditions. The partitioning module is configured to divide the training sample set into several angular partitions based on the overall wind direction, and to construct the calibration curve for each partition based on the relationship between historical gas concentration and actual gas concentration. The correction module is configured to substitute the predicted gas concentration value for the target time period into the correction curves corresponding to the target partition and its adjacent partitions to obtain multiple partition correction results, and generate a comprehensive correction value through a weighted fusion strategy. The assessment module is configured to calculate the final gas leakage status assessment value based on the normal operating condition branch prediction value, the extreme operating condition branch prediction value, and the comprehensive correction value.

10. The intelligent gas leakage detection system for transformers as described in claim 9, characterized in that, The acquisition module collects multi-source operating data from the transformer sealing system through multiple sensors, including pressure sensors, oil level sensors, temperature sensors, humidity sensors, airflow velocity sensors, and dissolved gas concentration sensors in the oil.

Images