Artificial intelligence diagnosis method for partial discharge fault of power equipment
The TEV probe and photoelectric sensor jointly detect the local discharge signal of the power equipment, generate the PRPD map data set, and use the multi-task element learning model to solve the problem of small samples in the local discharge fault diagnosis of power equipment, and achieve high accuracy and robust fault identification.
Patent Information
- Application Number
- CN202510383284.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-07-29
AI Technical Summary
In the prior art, local discharge fault diagnosis of power equipment has problems such as the risk of overfitting caused by small sample size, increased difficulty in model training, limited classification performance, category imbalance and low recognition accuracy, especially the lack of phase information and insufficient noise robustness under DC voltage.
The TEV probe and photoelectric sensor are used to jointly detect the local discharge signals of the power equipment, and the PRPD map data set is generated through feature extraction and data complementarity. The multi-task meta-learning model is used for training, including the input layer, task layer, basic learner, meta feature extraction layer, meta learner layer and meta optimization layer to realize signal feature sharing and model optimization.
It improves the accuracy of local discharge pattern recognition and the generalization ability of the model, can quickly adapt to new tasks with few samples, and improves the accuracy and robustness of fault recognition.
Smart Images
Figure CN120385890A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for diagnosing faults in power equipment, specifically an artificial intelligence diagnosis method for partial discharge faults in power equipment. Background Art
[0002] As a core component of modern power systems, the stability and reliability of power equipment are directly related to the safe supply of electricity and the stable operation of the system. The existing fault diagnosis methods usually rely on manual experience and manual detection, which have deficiencies and mostly rely on manual judgment, often being limited by the experience and skill levels of operators, resulting in a decrease in the accuracy of diagnosis results. In the case of complex power equipment or multiple coexisting faults, empirical judgment may be incorrect.
[0003] To solve the problems existing in manual diagnosis, the prior art has adopted artificial intelligence technology to achieve fault diagnosis of power equipment. However, when the prior art artificial intelligence technology diagnoses faults in power equipment, the problem is that the number of samples is small, which will lead to many key problems in partial discharge pattern recognition:
[0004] Increased risk of overfitting: Due to the limited number of samples, the neural network model is easily overfitted to these few samples, that is, the model performs well on the training set but has poor generalization ability on unseen data.
[0005] Increased difficulty in model training: A small number of samples means that the features learned by the model may not be comprehensive enough, resulting in an increase in the difficulty of model training, especially in the case where a deep learning model needs to learn complex features.
[0006] Limited classification performance: Insufficient samples may make it difficult for the classifier to accurately judge the type of partial discharge because the model does not have enough data to learn the features for distinguishing different discharge patterns.
[0007] Class imbalance: The partial discharge samples often have an imbalanced class ratio, and some types of samples may be very scarce, which further increases the difficulty of training an effective classifier.
[0008] Influence on recognition accuracy: Due to the scarcity of samples and class imbalance, traditional classification methods may be difficult to achieve high recognition accuracy, especially in partial discharge under DC voltage, where the lack of phase information makes the problem more complex.
[0009] Robustness to noise: The partial discharge signals collected on-site contain a large amount of noise. In the case of a small number of samples, the model may be insufficiently effective in recognizing partial discharge signals under different noise intensities. Summary of the Invention
[0010] The present invention provides an artificial intelligence diagnosis method for partial discharge faults of power equipment to solve the problem of difficult identification caused by a small number of samples in the prior art.
[0011] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0012] An artificial intelligence diagnosis method for partial discharge faults of power equipment includes the following steps:
[0013] Step 1: Obtain the transient earth voltage signal generated during partial discharge of the power equipment and the electrical signal corresponding to the discharge light signal generated during partial discharge of the power equipment;
[0014] Step 2: Perform feature extraction and data complementation on the transient earth voltage signal and the electrical signal corresponding to the discharge light signal obtained in Step 1 to obtain a PRPD pattern dataset;
[0015] Step 3: Divide the PRPD pattern dataset obtained in Step 2 into a training set and a test set, use the training set to train a multi-task meta-learning model, use the trained multi-task meta-learning model as a fault identification model, and use the test set to test the performance of the fault identification model;
[0016] Step 4: Input at least one of the transient earth voltage signal and the electrical signal corresponding to the discharge light signal generated during partial discharge of the power equipment to be detected into the fault identification model obtained in Step 3, and the fault identification model outputs a fault identification result.
[0017] Further, during feature extraction in Step 2, the total discharge times, regional discharge times, and discharge average value of each cycle of the transient earth voltage signal and the electrical signal corresponding to the discharge light signal after complementation are respectively extracted as features.
[0018] Further, during data complementation in Step 2, a multi-task meta-learning network is used to perform meta-learning training on the features extracted from the transient earth voltage signal and the electrical signal corresponding to the discharge light signal to share the unique discharge features of these two signals, achieving a mutual assistance effect; finally, the features after data complementation form a PRPD pattern dataset.
[0019] Further, the multi-task meta-learning model in Step 3 includes an input layer, multiple task layers, a base learner, a meta-feature extraction layer, a meta-learner layer, a meta-optimizer layer, and an output layer;
[0020] During training, the training set data is input into the input layer for preprocessing, and the preprocessed data is divided into multiple task data; the divided task data is respectively input into the corresponding task layer, and each task layer is trained by a basic learner, so that each task layer obtains a prediction result for the corresponding task; the prediction results of each task layer are then input into the meta-feature extraction layer, and the meta-features of each task are extracted through the meta-feature extraction layer; the meta-features of each task are then input into the meta-learner layer to learn the patterns between tasks, so as to optimize and adjust the basic learner, and then the basic learner is fine-tuned on new tasks through the meta-optimization layer; finally, the optimized and adjusted basic learner outputs the final prediction result through the output layer.
[0021] Further, during training, a meta-training method is used. The meta-training method includes an inner loop and an outer loop. In the inner loop, a fault identification model based on a multi-task learning model is trained using the learning tasks constructed; in the outer loop, the features extracted from the transient electrical wave signals are used as the original samples, and the features extracted from the electrical signals corresponding to the discharge optical signals are used as the new samples to verify the trained multi-task meta-learning model for the identification task, so as to provide feedback on the generalization of the model, thereby updating the initial parameters.
[0022] Compared with the prior art, the advantages of the present invention are as follows:
[0023] 1. It solves various problems in partial discharge pattern recognition caused by a small number of samples, such as inaccurate recognition, overly complex models, etc.
[0024] 2. Through the discharge signals collected by two different means, the data complementarity of both sides is realized, the accuracy of model recognition is improved, and when there is only data obtained by one collection means, it is possible to quickly generalize this type of data and cooperate with the multi-task meta-learning network to learn and train tasks of new data types. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 is a flowchart of the method in an embodiment of the present invention.
[0026] Figure 2 is a schematic structural diagram of the detection device in an embodiment of the present invention
[0027] Figure 3 is a schematic diagram of a partial discharge defect constructed in an embodiment of the present invention.
[0028] Figure 4 is a schematic diagram of the meta-training method used in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0029] The present invention will be further described below with reference to the drawings and embodiments.
[0030] As Figure 1As shown, this embodiment discloses an artificial intelligence diagnosis method for partial discharge faults of power equipment, including the following steps:
[0031] Step 1: Obtain the transient earth voltage (TEV) signal generated during partial discharge of power equipment, and the electrical signal corresponding to the discharge light signal generated during partial discharge of power equipment.
[0032] In this embodiment, the TEV signal generated during partial discharge of power equipment is obtained through a TEV probe, and the discharge light signal generated during partial discharge of power equipment is obtained through a photoelectric sensor and converted into a corresponding electrical signal. As Figure 2 Shown is the schematic diagram of the signal acquisition system of this embodiment. The transient earth voltage signal collected by the TEV probe and the electrical signal corresponding to the discharge light signal collected by the photoelectric sensor are obtained through an oscilloscope.
[0033] The combined detection of TEV and photoelectric sensors adopted in this embodiment is a comprehensive partial discharge detection technology, which is widely used in the partial discharge diagnosis of power equipment such as switch cabinets and cable plugs. The detection principles and detection frequency bands of each method are different. The combined use can provide more comprehensive detection results, helping to improve the positioning accuracy and diagnosis effect of partial discharge.
[0034] Photoelectric sensor detection mainly uses photoelectric elements to receive the light signals reflected or transmitted by the object to be measured and convert them into electrical signals for transmission to the corresponding receiving devices. These devices are usually powered by a DC power supply, and the voltage range is generally from 12V to 24V. There are also some devices that support rechargeable battery power supply, which is suitable for on-site portable detection. TEV detection equipment usually relies on external power supply to ensure the continuity and high precision of signal acquisition. Its equipment can be powered by AC or DC. The AC voltage is usually 220V, while the DC voltage is mostly 12V or 24V.
[0035] In this embodiment, an oscilloscope is used to display the signals obtained through detection. The TEV signal is usually in the frequency range of dozens of MHz. Therefore, the bandwidth of the oscilloscope can be selected between 50MHz and 200MHz. The photoelectric sensor usually detects light signals, and its frequency range is generally between kHz and MHz. Therefore, the bandwidth requirement for the oscilloscope is relatively low, and the range of dozens of kHz to hundreds of kHz is usually sufficient. For some high-speed applications, the bandwidth may need to reach the MHz level to ensure the complete capture and accurate measurement of the signal. The bandwidth range of the 5-series oscilloscope usually ranges from 350MHz to 2GHz, and the bandwidth range provided by this series of oscilloscopes can meet the detection requirements of TEV and photoelectric sensor signals. Therefore, the 5-series oscilloscope of this embodiment is used to receive the signals output by the TEV and photoelectric sensors.
[0036] As Figure 3As shown in the figure, in this embodiment, four typical partial discharge defects are obtained by constructing the partial discharge defects of the target power equipment: tip discharge, surface discharge, air gap discharge, and suspended particle discharge. The transient earth voltage data and the electrical signals corresponding to the discharge optical signals of each typical partial discharge defect of the target power equipment are obtained through a TEV probe and a photodetector respectively.
[0037] Step 2: Perform data complementation and feature extraction on the transient earth voltage signals and the electrical signals corresponding to the discharge optical signals obtained in Step 1, thereby obtaining a PRPD pattern dataset.
[0038] In this embodiment, during feature extraction, the total discharge times, regional discharge times, and discharge average values of each cycle of the transient earth voltage signals and the electrical signals corresponding to the discharge optical signals after complementation are respectively extracted as features. Among them, the regional discharge times refer to dividing each cycle into 360 regions on average and counting the discharge times of each region. The peak method is used for the extraction of the discharge times. After feature extraction, a dataset in the true sense can be obtained. This dataset consists of the above-mentioned three features: the total discharge times per cycle, the regional discharge times, and the discharge average value.
[0039] In this embodiment, data complementation is performed using a multi-task meta-learning network. The transient earth voltage signals and the electrical signals corresponding to the discharge optical signals are subjected to meta-learning training, which are respectively used for diagnosis and positioning. The unique discharge features of these two signals, namely the transient earth voltage signals and the electrical signals corresponding to the discharge optical signals, are shared through the feature sharing layer of the multi-task meta-learning network, achieving a mutually assisting effect.
[0040] Finally, the PRPD pattern dataset is composed of the features after data complementation.
[0041] Step 3: Divide the data in the PRPD pattern dataset obtained in Step 2 into a training set and a test set according to a ratio of 7:3.
[0042] Generate a multi-task meta-learning model. The multi-task meta-learning model includes an input layer, multiple task layers, a base learner, a meta-feature extraction layer, a meta-learner layer, a meta-optimizer layer, and an output layer. The training set is used to train the multi-task meta-learning model. The data processing process during training is as follows:
[0043] The training set data enters through the input layer and undergoes preprocessing through the input layer, including standardization, normalization, and denoising, to ensure that the data is suitable for subsequent processing.
[0044] The preprocessed data is then divided into multiple task data, and each task data has its independent data subset. The division of these tasks helps the model learn shared features and can effectively perform knowledge transfer between tasks.
[0045] The divided task data enters their respective corresponding task layers, and each task is trained by a basic learner. The core of the task layer is to use the shared network part to capture the common features between tasks, while the specific part of each task layer is responsible for generating prediction results for the corresponding task. These prediction results are initially learned through the shared part of the basic model based on their respective task requirements.
[0046] After the task-level prediction results are generated for each task layer, the prediction results obtained by each task layer enter the meta-feature extraction layer, and the meta-features of each task are extracted through the meta-feature extraction layer, such as the learning rate of the task, the similarity between tasks, and the difficulty of the task. These meta-features reflect the relationships between tasks and provide important context information for subsequent optimization steps.
[0047] The meta-features extracted by the meta-feature extraction layer are passed to the meta-learner layer, and the meta-learner layer is responsible for learning the patterns and relationships between tasks at this stage. The meta-learner layer identifies which tasks have similar learning rules, thereby helping to optimize and adjust the basic learner. Through these associations between tasks, the meta-learner layer adjusts the basic learner to make it more efficient in handling multiple tasks.
[0048] The learning results of the meta-learner layer then enter the meta-optimization layer, and the meta-optimization layer fine-tunes the basic learner based on the learning results of the meta-learner. This fine-tuning process enables the basic learner to quickly adapt to new tasks and perform rapid learning using a small amount of new data without having to retrain the entire network. Through the adjustment of the meta-optimization layer, the basic learner can achieve higher learning efficiency and accuracy on new tasks.
[0049] Finally, the optimized and fine-tuned basic learner generates the final prediction results through the output layer. Thus, the trained multi-task meta-learning model is obtained as a fault identification model.
[0050] After training, the test set is input into the trained fault identification model to evaluate the performance of the fault identification model. Evaluation metrics include accuracy, F1 score, etc., which help to judge the performance and generalization ability of the fault identification model on multiple tasks, thereby verifying its effectiveness.
[0051] Through this complete process, the multi-task network used in this embodiment can not only learn the shared knowledge between multiple tasks, but also efficiently optimize and adapt to new tasks, achieving better performance and fast task transfer ability.
[0052] In this embodiment, the meta-training method is used to optimize the initial parameters during training, as Figure 4 shown, there are a total of two optimizations, one is the inner loop and the other is the outer loop.
[0053] Specifically, in the inner loop, a fault identification model based on a multi-task learning model is trained using the constructed learning tasks. The constructed learning tasks refer to the tasks created by the multi-task meta-learning network from different types of data, including the diagnosis and location tasks of partial discharge. Specifically, the tasks include data complementarity and feature extraction through the signals collected by TEV and optoelectronic sensors, and then using these data to train the model for the diagnosis and location of partial discharge. The parameter update process is shown in the following formula:
[0054]
[0055] where: θ is the optimal parameter updated on this task; α is the learning rate of the inner loop network; f θ is the inner loop network structure of parameter θ; is the loss on a specific task; is the loss gradient.
[0056] In the outer loop, the dataset composed of the features extracted from the transient earth voltage (TEV) signals collected by TEV is used as the original sample, and the dataset composed of the features extracted from the electrical signals corresponding to the discharge optical signals collected by the optoelectronic sensor is used as the new sample. The trained multi-task meta-learning model, i.e., the fault identification model, is verified for the identification task to provide feedback on the generalization of the model, thereby updating the initial parameters.
[0057] In the outer loop, it is required that the optimal parameter θ has the minimum loss error for different tasks, and the specific form is shown in the following formula:
[0058]
[0059] where: p(T) is the distribution of the task set. The MAML algorithm can complete the update of parameter θ through the stochastic gradient descent method, and its process is as follows:
[0060]
[0061] where: β is the learning rate of the meta-learner.
[0062] The loss function for training is l1 as shown in the following formula:
[0063]
[0064] where: y i and p i represent the expectation and predicted probability respectively; N is the number of training samples.
[0065] Step 4: Input at least one of the transient ground wave signals generated during the partial discharge of the power equipment to be detected and the electrical signals corresponding to the discharge optical signals into the fault identification model obtained in Step 3, and output the fault identification result by the fault identification model.
[0066] The preferred embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Among the various specific technical features described in the above specific embodiments, they can be combined in any suitable manner without contradiction. As long as such a combination does not violate the idea of the present invention, it should also be regarded as the content disclosed in this disclosure. To avoid unnecessary repetition, the present invention will not separately describe various possible combination methods.
[0067] The present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention and without departing from the design idea of the present invention, various modifications and improvements made by those skilled in the art to the technical solutions of the present invention should fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.
Claims
1. An artificial intelligence diagnosis method for partial discharge faults of power equipment, characterized in that, The method includes the following steps: Step 1: Obtain the transient ground wave signals generated during partial discharge of power equipment, and the electrical signals corresponding to the discharge optical signals generated during partial discharge of power equipment; Step 2: Extract features and complement data for the transient ground wave signals and the electrical signals corresponding to the discharge optical signals obtained in Step 1, thereby obtaining a PRPD pattern dataset; Step 3: Divide the PRPD pattern dataset obtained in Step 2 into a training set and a test set, use the training set to train a multi-task meta-learning model, use the trained multi-task meta-learning model as a fault identification model, and use the test set to test the performance of the fault identification model; Step 4: Input at least one of the transient ground wave signals and the electrical signals corresponding to the discharge optical signals generated during partial discharge of the power equipment to be detected into the fault identification model obtained in Step 3, and the fault identification model outputs a fault identification result.
2. The artificial intelligence diagnosis method for partial discharge faults of a power device according to claim 1, wherein During feature extraction in Step 2, the total discharge times, regional discharge times, and discharge average values of each period of the transient ground wave signals and the electrical signals corresponding to the discharge optical signals after complementation are respectively extracted as features.
3. The artificial intelligence diagnosis method for partial discharge faults of a power device according to claim 2, wherein, During data complementation in Step 2, a multi-task meta-learning network is used to perform meta-learning training on the features extracted from the transient ground wave signals and the electrical signals corresponding to the discharge optical signals, so as to share the unique discharge features of these two signals of the transient ground wave signals and the electrical signals corresponding to the discharge optical signals, achieving a mutual assistance effect; finally, the features after data complementation constitute the PRPD pattern dataset.
4. An artificial intelligence diagnosis method for partial discharge faults of power equipment according to claim 1, characterized in that The multi-task meta-learning model described in Step 3 includes an input layer, multiple task layers, a base learner, a meta-feature extraction layer, a meta-learner layer, a meta-optimization layer, and an output layer; During training, the training set data is input into the input layer for preprocessing, and the preprocessed data is divided into multiple task data; The divided task data is respectively input into the corresponding task layers, and each task layer is trained by the base learner, so that each task layer obtains a prediction result for the corresponding task; the prediction results of each task layer are then input into the meta-feature extraction layer, and the meta-features of each task are extracted through the meta-feature extraction layer; the meta-features of each task are then input into the meta-learner layer to learn the patterns between tasks, so as to optimize and adjust the base learner, and then the base learner is fine-tuned on new tasks through the meta-optimization layer; finally, the optimized and adjusted base learner outputs the final prediction result through the output layer.
5. An artificial intelligence diagnosis method for partial discharge faults of power equipment according to claim 4, characterized in that, During training, a meta-training method is used. The meta-training method includes an inner loop and an outer loop. In the inner loop, a fault identification model based on a multi-task learning model is trained using the learning tasks constructed; in the outer loop, the features extracted from the transient ground wave signals are used as the original samples, and the features extracted from the electrical signals corresponding to the discharge optical signals are used as new samples to verify the trained multi-task meta-learning model for the identification task, so as to provide feedback on the generalization of the model, thereby updating the initial parameters.