Transformer light gas fault detection method and device based on deep learning

Abstract

The invention relates to a transformer light gas fault detection method and device based on deep learning, and belongs to the field of transformer fault diagnosis. An integrated classification model based on an improved whale optimization algorithm and adaptive enhancement is constructed, support vector machine hyper-parameters are optimized through Tent chaotic mapping initialization, a Levy flight strategy and a simulated annealing mechanism, and model stability and classification precision are improved in combination with weighted multi-weak classifier fusion and a regularization technology; the method comprises the following steps: collecting light gas data by adopting multiple types of gas sensors, and carrying out data preprocessing and feature dimension reduction by utilizing standardization, principal component analysis and linear discriminant analysis; real-time fault diagnosis and feedback are realized based on a gas concentration and fault type relation formula and an edge computing node, and the timeliness and accuracy of fault diagnosis are improved; the method has high real-time performance and practicability, and can quickly complete detection and fault identification of light gas fault gas of the transformer.

Description

Technical Field

[0001] The present invention relates to the field of transformer fault diagnosis, and in particular to a transformer light gas fault detection method and device based on deep learning. Background Art

[0002] When an oil-immersed transformer fails, the chemical decomposition of its internal insulation material and the oil medium produces specific characteristic gases. For example, under typical fault conditions such as insulation aging, abnormal electrical stress, or thermal stress overload, the cracking and recombination reactions of hydrocarbons will produce gas products that are fault-indicating. Research has shown that the types and concentration gradients of gases such as hydrogen (H2), carbon monoxide (CO), methane (CH4), and acetylene (C2H2) are significantly correlated with fault types such as partial discharge, arc discharge, or overheating. This makes characteristic gas analysis an important technical means of assessing the health status of transformers. The current mainstream dissolved gas analysis (DGA) method performs fault diagnosis by detecting the composition and concentration changes of dissolved gases in transformer oil.

[0003] Traditional gas protection mechanisms primarily rely on the volumetric action principle of mechanical gas relays, providing fault warnings by monitoring the accumulated gas volume in the transformer oil compartment. When the gas volume reaches a preset threshold, the device triggers a minor gas alarm or a major gas trip, respectively. A growing number of researchers are applying artificial intelligence (AI) technologies, particularly deep learning algorithms, to DGA analysis to improve diagnostic accuracy.

[0004] A new light gas protection device eliminates the risk of manual intervention during gas sampling and addresses the vulnerability of traditional volumetric criteria to environmental interference. It also shifts its research focus to analyzing hydrogen (H2) content and gas volume, constructing a quantitative relationship model between free gas components and discharge energy. This enables the device to accurately identify faults in their early stages and promptly issue alarms or trip signals, effectively reducing the occurrence of false trips. An edge computing device for transformer fault diagnosis connects a photoacoustic spectroscopy sensor and deploys a transformer DGA fault diagnosis algorithm at the edge, enabling real-time monitoring of the transformer's operating status and providing fault warnings. An edge monitoring device based on an ARM11 core combines multi-information fusion technology with edge computing to collect multi-dimensional information such as load current, ambient temperature, top and bottom oil temperatures, and dissolved gas analysis data. Edge computing is performed using a least squares twin support vector regression (LSTSVR) prediction model to efficiently and effectively determine whether the transformer is at risk of overheating in real time.

[0005] Several invention patents have been developed to address the problem of detecting light gas faults in transformers. For example, patent CN112350273A discloses a digital light gas protection method and device for transformers. By obtaining the gas component content extracted from the gas chamber of the transformer gas relay, the severity of the transformer fault is determined. If a serious fault is determined, the transformer is disconnected. This patent can accurately and promptly reflect the problem of transformer faults. However, this patent still has the problem that the setting values of criteria A and B need to be further optimized to improve the sensitivity and accuracy of the criteria. Patent CN118243747A discloses a transformer fault diagnosis method based on online gas monitoring. This method obtains analysis results by quantitatively analyzing the gas composition and, based on the analysis results, determines whether to generate a transformer fault alarm. This patent can collect and analyze capacitance values, explore the dynamic characteristics of capacitance values over time, and thus infer the dynamic characteristics of light gas volume values. This can identify possible abnormal conditions within the transformer, enabling monitoring and fault diagnosis of the transformer's internal conditions. However, in terms of gas online monitoring technology, this patent still has the problem of further optimizing the design and performance of gas sensors and improving the accuracy and stability of monitoring gas concentration and composition.

[0006] The existing technology has the following shortcomings: the traditional light gas protection system is mainly based on gas volume thresholds and mechanical flow rate to trigger alarms, and its detection principle has significant limitations. The gas accumulation speed in the gas box is easily disturbed by non-fault gas retention (such as the gas trap caused by oil flow vortex), resulting in a high false alarm rate; the traditional method ignores the close relationship between light gas and rapidly developing faults, resulting in a significant lag in fault response and an inability to detect and diagnose faults in a timely manner; the existing technology has deficiencies in optimizing the criterion setting value, and the sensitivity and accuracy of the criterion need to be further optimized to improve the accuracy of fault diagnosis; the existing technology still needs to be optimized in terms of the design and performance of gas sensors, and the accuracy and stability of gas concentration and composition monitoring need to be improved. Summary of the Invention

[0007] The existing technology has problems such as the limitations of the detection principle of the traditional light gas protection system, the inability to detect and diagnose faults in a timely manner, insufficient optimization of the judgment setting value, and the need to improve the performance of the gas sensor.

[0008] In light of this, the present invention aims to provide a light gas fault identification model based on deep learning, enabling accurate identification of rapidly developing discharge fault types within transformers. By detecting the gas components of free light gas, a light gas fault identification model based on deep learning is constructed, enabling accurate identification of rapidly developing discharge fault types within transformers. This effectively addresses the inefficiency, proneness to false triggering, and potential hazards associated with traditional DGA technology, as well as maintenance personnel performing manual oil extraction. It also improves detection efficiency, enabling rapid identification of the severity and type of transformer faults.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A method for detecting light gas faults in transformers based on deep learning, comprising the following steps:

[0011] Step 1: Build a light gas fault identification model based on deep learning;

[0012] Step 2: Collect light gas data in the transformer;

[0013] Step 3: Input the collected light gas data into the light gas fault identification model for analysis to determine whether a fault exists. If a fault exists, return to step 2 to continue collecting data. If no fault exists, proceed to step 4.

[0014] Step 4: Output the fault diagnosis results.

[0015] Furthermore, the step 1 includes:

[0016] Step 101: Optimize the hyperparameters of the Support Vector Machine (SVM) using an improved Whale Optimization Algorithm (IWOA) to form an IWOA-SVM weak classifier.

[0017] Step 102: Combining the Adaptive Boosting (AdaBoost) algorithm with the IWOA-SVM weak classifier, a strong classifier is formed by weighted combination of multiple weak classifiers.

[0018] Step 103: Adopt an adaptive learning rate adjustment strategy and regularization technology to improve the stability and accuracy of the model.

[0019] Further, the step 2 includes:

[0020] Step 201: Using a carbon monoxide gas sensor, a methane gas sensor, a hydrogen gas sensor, and an acetylene gas sensor to collect light gas data in the transformer;

[0021] Step 202: perform data preprocessing and feature engineering such as standardization, normalization, principal component analysis, and linear discriminant analysis on the collected light gas data.

[0022] Furthermore, the step 4 includes:

[0023] Step 401: Determine the fault type and severity based on the relationship formula between gas concentration and fault type and the adaptive weight adjustment formula;

[0024] Step 402: Feedback the fault diagnosis results to the operation and maintenance personnel in real time, and provide fault handling suggestions.

[0025] Furthermore, the step 101 is specifically as follows:

[0026] The tent chaotic map is used to initialize the population, adaptive weights are added, and the Levy flight strategy and simulated annealing strategy are introduced to improve the algorithm's global search capability and convergence speed, thereby optimizing the hyperparameters of the SVM and forming the IWOA-SVM weak classifier.

[0027] The step 102 is specifically as follows:

[0028] Integrate multiple IWOA-SVM weak classifiers and use the AdaBoost algorithm to perform weighted combination of the weak classifiers to form a powerful classifier;

[0029] The step 103 is specifically as follows:

[0030] The adaptive learning rate adjustment strategy dynamically adjusts the learning rate to enable the model to converge quickly in the early stages of training, and fine-tune parameters in the later stages of training to avoid overfitting; by adding L1 and L2 regularization terms to the loss function, the complexity of the model is controlled and the generalization ability is improved.

[0031] Furthermore, the step 201 is specifically as follows:

[0032] The carbon monoxide gas sensor model is: JXM-CO, the methane gas sensor model is: JXM-CH4, the hydrogen gas sensor model is: JXM-H2, and the acetylene gas sensor model is: 2M004;

[0033] The step 202 is specifically as follows:

[0034] Standardization and normalization are used to compare different sensor data on the same scale; PCA and LDA are used to reduce feature dimensionality and extract the most representative features, reducing the training time and computational complexity of the model.

[0035] Furthermore, the step 401 is specifically as follows:

[0036] Establish the relationship formula between gas concentration and fault type:

[0037]

[0038] Where C is the comprehensive gas concentration, ω i is the weight of the i-th gas, c i is the concentration of the i-th gas;

[0039] Establish an adaptive weight adjustment formula:

[0040]

[0041] Where k is the adjustment coefficient and c0 is the reference concentration;

[0042] Step 402 specifically includes: using a Raspberry Pi 4B as an edge computing node and deploying a deep learning-based fault recognition model to detect and identify light gas faults.

[0043] Furthermore, step 103 further includes:

[0044] Evaluate and diagnose model performance and calculate the model loss function, which is expressed as the following formula:

[0045]

[0046] Among them, L is the loss function, y i is the actual value, is the predicted value, λ is the regularization coefficient, ω j are model parameters.

[0047] A transformer light gas fault detection device based on deep learning, comprising:

[0048] Power supply, which provides power to the entire device;

[0049] A Raspberry Pi development board, including a circuit board and a core control unit that processes sensor data and detects and identifies light gas faults;

[0050] The sensor module is connected to the Raspberry Pi development board to collect light gas data in the transformer and transmit it to the digital-to-analog conversion module;

[0051] A digital-to-analog conversion module processes the sensor data and transmits it to the Raspberry Pi development board;

[0052] A high-definition multimedia interface device, connected to the Raspberry Pi development board and a display device;

[0053] Display device, which displays the data collected by the sensor and the light gas fault detection results in real time;

[0054] The Raspberry Pi development board uses a Raspberry Pi 4B as an edge computing node and deploys a deep learning-based fault recognition model to identify light gas faults;

[0055] The deep learning-based fault recognition model optimizes the hyperparameters of a support vector machine (SVM) through an improved whale optimization algorithm (IWOA) to form an IWOA-SVM weak classifier. The adaptive boosting algorithm (AdaBoost) is combined with the IWOA-SVM weak classifier to form a strong classifier by weighted combination of multiple weak classifiers. An adaptive learning rate adjustment strategy and regularization technology are used to improve the stability and accuracy of the model.

[0056] Furthermore, the sensor module specifically includes:

[0057] Carbon monoxide gas sensor, methane gas sensor, hydrogen gas sensor, acetylene gas sensor:

[0058] The carbon monoxide gas sensor adopts electrochemical detection, model: JXM-CO;

[0059] The methane gas sensor adopts the detection method of catalytic combustion, model: JXM-CH4;

[0060] The hydrogen gas sensor adopts the detection mode of catalytic combustion, model: JXM-H2;

[0061] The acetylene gas sensor adopts the detection method of semiconductor detection, model: 2M004.

[0062] The beneficial effects of the present invention are:

[0063] By constructing a light gas fault identification model based on deep learning, and using an improved whale optimization algorithm to optimize the hyperparameters of the support vector machine, combined with the AdaBoost algorithm to form a strong classifier, the accurate identification of the rapidly developing discharge fault type inside the transformer is achieved, effectively overcoming the problems of fault response lag and high false alarm rate existing in traditional methods; the adaptive learning rate adjustment strategy and regularization technology, as well as data preprocessing and feature engineering methods such as standardization, normalization, principal component analysis and linear discriminant analysis are adopted to improve the stability, accuracy and generalization ability of the model; a real-time monitoring and feedback mechanism is designed, and a relationship formula between gas concentration and fault type and an adaptive weight adjustment formula are introduced to achieve real-time monitoring of transformer light gas faults, quickly respond to the type and severity of the fault, and improve the timeliness and accuracy of fault diagnosis; data training and model deployment are carried out, and a deep learning-based transformer light gas fault edge detection method is formed by combining four gas sensors with the edge computing device Raspberry Pi. It has strong real-time and practicality and can quickly complete the detection and fault identification of transformer light gas fault gas.

[0064] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0066] Figure 1 This is a hardware diagram of a transformer light gas fault detection device based on deep learning;

[0067] Figure 2 This is a schematic diagram of the accuracy of transformer fault identification using the AdaBoost-IWOA-SVM model according to an embodiment of the present invention;

[0068] Figure 3 This is a comparison chart of the accuracy of transformer fault identification using the AdaBoost-IWOA-SVM model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0070] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0071] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0072] See also Figure 1 , is a hardware diagram of a transformer light gas fault detection device based on deep learning; the device mainly includes:

[0073] Power supply: power adapter and connecting cable to power the entire device;

[0074] Raspberry Pi development board: This includes a circuit board with multiple interfaces (such as HDMI and USB) and a core control unit that processes sensor data and uses the sensor data to detect and identify various types of light gas faults.

[0075] Sensor module: Connects to the Raspberry Pi development board to collect light gas data from the transformer and transmits the light gas data to the digital-to-analog conversion module; the sensor module mainly includes a carbon monoxide (CO) gas sensor; a methane (CH4) gas sensor; a hydrogen (H2) gas sensor; and an acetylene (C2H2) gas sensor;

[0076] Digital-to-analog conversion module: receives and converts sensor data and transmits it to the Raspberry Pi development board;

[0077] High Definition Multimedia Interface (HDMI) device: connects the Raspberry Pi development board and the display device;

[0078] Display device: displays the data collected by the sensor in real time, and displays the detection results of various types of light gas faults.

[0079] Preferably, the carbon monoxide gas sensor adopts the detection method of electrochemical detection, model: JXM-CO; the methane gas sensor adopts the detection method of catalytic combustion, model: JXM-CH4; the hydrogen gas sensor adopts the detection method of catalytic combustion, model: JXM-H2; the acetylene gas sensor adopts the detection method of semiconductor detection, model: 2M004.

[0080] Preferably, the power source is an independent mobile power source.

[0081] In order to overcome the shortcomings of the prior art, the present invention also provides a transformer light gas fault detection method based on deep learning, comprising:

[0082] Step 1: Build a light gas fault identification model based on deep learning;

[0083] Step 2: Collect light gas data in the transformer;

[0084] Step 3: Input the collected light gas data into the light gas fault identification model for analysis to determine whether a fault exists. If a fault exists, return to step 2 to continue collecting data. If no fault exists, proceed to step 4.

[0085] Step 4: Output the fault diagnosis results.

[0086] Example 1:

[0087] The step 1 includes:

[0088] Step 101: Optimize the hyperparameters of the Support Vector Machine (SVM) using an improved Whale Optimization Algorithm (IWOA) to form an IWOA-SVM weak classifier.

[0089] Step 102: Combining the Adaptive Boosting (AdaBoost) algorithm with the IWOA-SVM weak classifier, a strong classifier is formed by weighted combination of multiple weak classifiers.

[0090] Step 103: Adopt an adaptive learning rate adjustment strategy and regularization technology to improve the stability and accuracy of the model.

[0091] The step 2 includes:

[0092] Step 201: Use a carbon monoxide (CO) gas sensor, a methane (CH4) gas sensor, a hydrogen (H2) gas sensor, and an acetylene (C2H2) gas sensor to collect light gas data in the transformer;

[0093] Step 202: perform data preprocessing and feature engineering such as standardization, normalization, principal component analysis (PCA) and linear discriminant analysis (LDA) on the collected light gas data.

[0094] The step 4 includes:

[0095] Step 401: Determine the fault type and severity based on the relationship formula between gas concentration and fault type and the adaptive weight adjustment formula;

[0096] Step 402: Feedback the fault diagnosis results to the operation and maintenance personnel in real time, and provide fault handling suggestions.

[0097] Building a light gas fault identification model based on deep learning also includes:

[0098] Step 101: Optimize the hyperparameters of the support vector machine (SVM) using the improved whale optimization algorithm (IWOA) to form an IWOA-SVM weak classifier. Specifically, the population is initialized using the Tent chaos map, adaptive weights are added, and the Levy flight strategy and simulated annealing strategy are introduced to improve the algorithm's global search capability and convergence speed, thereby optimizing the hyperparameters of the SVM and forming the IWOA-SVM weak classifier.

[0099] Step 102: Combine the AdaBoost algorithm with the IWOA-SVM weak classifier to form a strong classifier by weighted combination of multiple weak classifiers. Specifically, multiple IWOA-SVM weak classifiers are integrated and weighted combination of the weak classifiers is performed using the AdaBoost algorithm to form a strong classifier.

[0100] Step 103: Adaptive learning rate adjustment strategies and regularization techniques are employed to improve the stability and accuracy of the model. Specifically, the adaptive learning rate adjustment strategy dynamically adjusts the learning rate, enabling rapid convergence of the model in the early stages of training and fine-tuning parameters in the later stages of training to avoid overfitting. Regularization effectively controls the complexity of the model and improves generalization by adding L1 and L2 regularization terms to the loss function.

[0101] Step 2: Collect light gas data in the transformer;

[0102] This step includes:

[0103] Step 201: Use a carbon monoxide (CO) gas sensor (model JXM-CO), a methane (CH4) gas sensor (model JXM-CH4), a hydrogen (H2) gas sensor (model JXM-H2), and an acetylene (C2H2) gas sensor (model 2M004) to collect light gas data in the transformer;

[0104] Step 202: The collected light gas data undergoes data preprocessing and feature engineering, including standardization, normalization, principal component analysis (PCA), and linear discriminant analysis (LDA). Specifically, standardization and normalization ensure that data from different sensors are compared on the same scale; PCA and LDA perform feature dimensionality reduction to extract the most representative features, reducing model training time and computational complexity.

[0105] Step 3: Input the collected light gas data into the light gas fault identification model for analysis to determine whether a fault exists. If a fault exists, return to step 2 to continue collecting data. If no fault exists, proceed to step 4.

[0106] Step 4: Output the fault diagnosis results.

[0107] This step includes:

[0108] Step 401: Determine the fault type and severity based on the relationship formula between gas concentration and fault type and the adaptive weight adjustment formula.

[0109] Specifically, the relationship between gas concentration and fault type is as follows:

[0110]

[0111] Where C is the comprehensive gas concentration, ω i is the weight of the i-th gas, c i is the concentration of the i-th gas.

[0112] Adaptive weight adjustment:

[0113]

[0114] Where k is the adjustment coefficient and c0 is the reference concentration.

[0115] Step 402: Feedback the fault diagnosis results to the operation and maintenance personnel in real time, and provide fault handling suggestions.

[0116] The device uses a Raspberry Pi 4B as an edge computing node and successfully deploys a deep learning-based fault identification model, resulting in a deep learning-based method for detecting transformer light gas faults. Gas distribution testing has verified that this method can detect and identify transformer light gas faults in approximately 20 seconds, demonstrating strong real-time and practicality.

[0117] Example 2:

[0118] A transformer light gas fault detection device based on deep learning includes the following steps:

[0119] Step 1: Build a light gas fault identification model based on deep learning;

[0120] This step includes:

[0121] Step 101: Optimize the hyperparameters of the Support Vector Machine (SVM) using the Improved Whale Optimization Algorithm (IWOA) to form an IWOA-SVM weak classifier. Specifically, tent chaos mapping is used to initialize the population, adaptive weights are added, and the Levy Flight Strategy and Simulated Annealing Strategy are introduced to improve the algorithm's global search capability and convergence speed, thereby optimizing the hyperparameters of the SVM and forming a weak classifier based on IWOA-SVM.

[0122] Step 102: Combine the AdaBoost algorithm with the IWOA-SVM weak classifier to form a strong classifier by weighted combination of multiple weak classifiers. Specifically, multiple IWOA-SVM weak classifiers are integrated and weighted combination of the weak classifiers is performed using the AdaBoost algorithm to form a strong classifier.

[0123] Step 103: Adaptive learning rate adjustment strategies and regularization techniques are employed to improve the stability and accuracy of the model. Specifically, the adaptive learning rate adjustment strategy dynamically adjusts the learning rate, enabling rapid convergence of the model in the early stages of training and fine-tuning parameters in the later stages of training to avoid overfitting. Regularization effectively controls the complexity of the model and improves generalization by adding L1 and L2 regularization terms to the loss function.

[0124] To evaluate and diagnose model performance, the model loss function is constructed as:

[0125]

[0126] Among them, L is the loss function, y i is the actual value, is the predicted value, λ is the regularization coefficient, ω j are model parameters.

[0127] Step 2: Collect light gas data in the transformer;

[0128] This step includes:

[0129] Step 201: Use a carbon monoxide (CO) gas sensor (model JXM-CO), a methane (CH4) gas sensor (model JXM-CH4), a hydrogen (H2) gas sensor (model JXM-H2), and an acetylene (C2H2) gas sensor (model 2M004) to collect light gas data in the transformer;

[0130] Step 202: The collected light gas data undergoes data preprocessing and feature engineering, including standardization, normalization, principal component analysis (PCA), and linear discriminant analysis (LDA). Specifically, standardization and normalization ensure that data from different sensors are compared on the same scale; PCA and LDA perform feature dimensionality reduction to extract the most representative features, reducing model training time and computational complexity.

[0131] Step 3: Input the collected light gas data into the light gas fault identification model for analysis to determine whether a fault exists. If a fault exists, return to step 2 to continue collecting data. If no fault exists, proceed to step 4.

[0132] Step 4: Output the fault diagnosis results.

[0133] This step includes:

[0134] Step 401: Determine the fault type and severity based on the relationship formula between gas concentration and fault type and the adaptive weight adjustment formula.

[0135] Step 402: Feedback the fault diagnosis results to the operation and maintenance personnel in real time, and provide fault handling suggestions.

[0136] The device uses a Raspberry Pi 4B as an edge computing node and successfully deploys a deep learning-based fault identification model, resulting in a deep learning-based method for detecting transformer light gas faults. Gas distribution testing has verified that this method can detect and identify transformer light gas faults in approximately 20 seconds, demonstrating strong real-time and practicality.

[0137] See also Figure 2 Figure 2 shows the accuracy of the AdaBoost-IWOA-SVM model in transformer fault identification using an embodiment of the present invention. Specifically, the AdaBoost-IWOA-SVM was used to identify four typical transformer faults (Type 1: Pin-to-Plate Breakdown Discharge Fault, Type 2: Ball-to-Plate Breakdown Discharge Fault, Type 3: Column-to-Plate Breakdown Discharge Fault, and Type 4: Oil-Paper Insulation Breakdown Discharge Fault). Comparing the performance of various intelligent algorithms in discharge fault detection, the AdaBoost-IWOA-SVM hybrid model demonstrated significant advantages. This model maintained a leading position in detecting all five types of insulation breakdown faults, achieving an overall accuracy of 96.8%, a 3.3 percentage point improvement over the second-ranked model, demonstrating its excellent overall performance. Specific test data shows that in the tip discharge (needle-to-plate) scenario, the proposed model significantly outperforms the traditional SVM (83.3%), PSO-BP (83.3%), the improved WOA-SVM (91.7%), and AdaBoost-SVM (91.7%) with a detection rate of 95.8%. In the detection of column-to-plate discharge, while AdaBoost-SVM achieves an accuracy of 83.3%, it is still 12.5 percentage points lower than the hybrid model. In the identification of ball-to-plate discharge, PSO-BP's 91.7% accuracy still lags behind the hybrid model by 4.1 percentage points. In particular, in the detection of oil-paper insulation breakdown, both the hybrid model and PSO-BP achieve full sample correct identification, while AdaBoost-SVM has a 2.1% misclassification rate. Experimental data fully demonstrates the adaptability and stability of the proposed fusion algorithm in identifying different discharge patterns.

[0138] See also Figure 3, which is a comparison chart of the accuracy of the AdaBoost-IWOA-SVM model in identifying transformer faults according to an embodiment of the present invention; in the performance evaluation of the classification model, the confusion matrix is used as a core evaluation tool to effectively quantify the generalization performance of the model in unknown data. The matrix intuitively reflects the discrimination efficiency of the classifier by constructing a cross-correspondence between the predicted category and the actual category. Based on experimental data containing 12 groups of test samples, the present invention focuses on the device's ability to identify different insulation faults. The visual analysis results show that the device achieved 11 correct discriminations (with an accuracy rate of 91.7%) in each of the three types of discharge faults: needle plate, column plate, and ball plate, with 1 case of misjudgment in each. In particular, in the detection of oil-paper insulation faults, the algorithm made accurate judgments on all 12 groups of samples, achieving a perfect recognition effect with zero error. This differentiated detection performance confirms that there are significant differences in the algorithm's ability to capture different physical characteristics of discharges.

[0139] The results show that the use of AdaBoost-IWOA-SVM has significant advantages in identifying four typical transformer faults.

[0140] This paper uses an algorithm for data training and model deployment, combining four sensors (CO, CH4, H2, and C2H2) with a Raspberry Pi for edge detection. This allows for real-time monitoring and rapid reflection of the type and severity of faults.

[0141] This paper proposes a collaborative optimization strategy that combines AdaBoost ensemble learning with an improved whale optimization algorithm (IWOA). By effectively enhancing the diversity of weak classifier error patterns, it significantly improves the recognition accuracy and generalization capability of the support vector machine (SVM) classifier. Verification based on discharge simulation test data shows that the AdaBoost-IWOA-SVM model constructed in this paper exhibits excellent robustness under complex operating conditions, with a comprehensive recognition accuracy of 96.8% for four types of discharge faults, an increase of 8% to 19% compared to traditional models, demonstrating significant performance advantages and practical application potential.

[0142] Using a Raspberry Pi 4B as an edge computing node, a deep learning-based fault identification model was successfully deployed, resulting in a deep learning-based method for detecting transformer light gas faults. Gas distribution testing has proven that this method can detect and identify transformer light gas faults in approximately 20 seconds, demonstrating strong real-time and practicality.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A transformer light gas fault detection method based on deep learning, characterized in that: The following steps are involved: Step 1: Build a light gas fault identification model based on deep learning; Step 2: Collect light gas data in the transformer; Step 3: Input the collected light gas data into the light gas fault identification model for analysis to determine whether a fault exists. If a fault exists, return to step 2 to continue collecting data. If no fault exists, proceed to step 4. Step 4: Output the fault diagnosis results.

2. A transformer light gas fault detection method based on deep learning according to claim 1, characterized in that: The step 1 includes: Step 101: Optimize the hyperparameters of the Support Vector Machine (SVM) using an improved Whale Optimization Algorithm (IWOA) to form an IWOA-SVM weak classifier. Step 102: Combining the Adaptive Boosting (AdaBoost) algorithm with the IWOA-SVM weak classifier, a strong classifier is formed by weighted combination of multiple weak classifiers. Step 103: Adopt an adaptive learning rate adjustment strategy and regularization technology to improve the stability and accuracy of the model.

3. A transformer light gas fault detection method based on deep learning according to claim 1, characterized in that: The step 2 includes: Step 201: Using a carbon monoxide gas sensor, a methane gas sensor, a hydrogen gas sensor, and an acetylene gas sensor to collect light gas data in the transformer; Step 202: perform data preprocessing and feature engineering such as standardization, normalization, principal component analysis, and linear discriminant analysis on the collected light gas data.

4. A transformer light gas fault detection method based on deep learning according to claim 1, characterized in that: The step 4 includes: Step 401: Determine the fault type and severity based on the relationship formula between gas concentration and fault type and the adaptive weight adjustment formula; Step 402: Feedback the fault diagnosis results to the operation and maintenance personnel in real time, and provide fault handling suggestions.

5. A transformer light gas fault detection method based on deep learning according to claim 2, characterized in that: The step 101 is specifically as follows: The tent chaotic map is used to initialize the population, adaptive weights are added, and the Levy flight strategy and simulated annealing strategy are introduced to improve the algorithm's global search capability and convergence speed, thereby optimizing the hyperparameters of the SVM and forming the IWOA-SVM weak classifier. The step 102 is specifically as follows: Integrate multiple IWOA-SVM weak classifiers and use the AdaBoost algorithm to perform weighted combination of the weak classifiers to form a powerful classifier; The step 103 is specifically as follows: The adaptive learning rate adjustment strategy dynamically adjusts the learning rate to enable the model to converge quickly in the early stages of training, and fine-tune parameters in the later stages of training to avoid overfitting; by adding L1 and L2 regularization terms to the loss function, the complexity of the model is controlled and the generalization ability is improved.

6. A transformer light gas fault detection method based on deep learning according to claim 3, characterized in that: The step 201 is specifically as follows: The carbon monoxide gas sensor model is: JXM-CO, the methane gas sensor model is: JXM-CH4, the hydrogen gas sensor model is: JXM-H2, and the acetylene gas sensor model is: 2M004; The step 202 is specifically as follows: Standardization and normalization are used to compare different sensor data on the same scale; PCA and LDA are used to reduce feature dimensionality and extract the most representative features, reducing the training time and computational complexity of the model.

7. A transformer light gas fault detection method based on deep learning according to claim 4, characterized in that: The step 401 is specifically as follows: Establish the relationship formula between gas concentration and fault type: Where C is the comprehensive gas concentration, ω i is the weight of the i-th gas, c i is the concentration of the i-th gas; Establish an adaptive weight adjustment formula: Where k is the adjustment coefficient and c0 is the reference concentration; Step 402 specifically includes: using a Raspberry Pi 4B as an edge computing node and deploying a deep learning-based fault recognition model to detect and identify light gas faults.

8. A transformer light gas fault detection method based on deep learning according to claim 5, characterized in that: The step 103 further includes: Evaluate and diagnose model performance and calculate the model loss function, which is expressed as the following formula: Among them, L is the loss function, y i is the actual value, is the predicted value, λ is the regularization coefficient, ω j are model parameters.

9. A transformer light gas fault detection device based on deep learning, characterized in that: include: Power supply, which provides power to the entire device; A Raspberry Pi development board, including a circuit board and a core control unit that processes sensor data and detects and identifies light gas faults; The sensor module is connected to the Raspberry Pi development board to collect light gas data in the transformer and transmit it to the digital-to-analog conversion module; A digital-to-analog conversion module processes the sensor data and transmits it to the Raspberry Pi development board; A high-definition multimedia interface device, connected to the Raspberry Pi development board and a display device; Display device, which displays the data collected by the sensor and the light gas fault detection results in real time; The Raspberry Pi development board uses a Raspberry Pi 4B as an edge computing node and deploys a deep learning-based fault recognition model to identify light gas faults; The fault recognition model based on deep learning optimizes the hyperparameters of the support vector machine (SVM) through the improved whale optimization algorithm (Improved Whale Optimization Algorithm, IWOA) to form an IWOA-SVM weak classifier; The adaptive boosting algorithm (AdaBoost) is combined with the IWOA-SVM weak classifier to form a strong classifier by weighted combination of multiple weak classifiers. The adaptive learning rate adjustment strategy and regularization technology are used to improve the stability and accuracy of the model.

10. The transformer light gas fault detection device based on deep learning according to claim 9, characterized in that: The sensor module specifically includes: Carbon monoxide gas sensor, methane gas sensor, hydrogen gas sensor, acetylene gas sensor: The carbon monoxide gas sensor adopts electrochemical detection, model: JXM-CO; The methane gas sensor adopts the detection method of catalytic combustion, model: JXM-CH4; The hydrogen gas sensor adopts the detection mode of catalytic combustion, model: JXM-H2; The acetylene gas sensor adopts the detection method of semiconductor detection, model: 2M004.

Images