GIS partial discharge source positioning method and system based on physical constraint neural network
By using a physical constraint neural network-based approach, combined with time delay consistency and geometric boundary constraints, the problem of insufficient positioning accuracy and physical consistency of local discharge sources in GIS was solved, achieving high-precision and highly interpretable positioning results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING INST OF TECH
- Filing Date
- 2026-01-21
- Publication Date
- 2026-06-05
Smart Images
Figure CN122153832A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power equipment condition monitoring technology, and relates to a GIS local discharge source location method and system based on physical constraint neural network. Background Technology
[0002] Partial discharge localization is a key technology for assessing the insulation status of GIS. Existing localization methods are mainly divided into two categories: traditional algorithms based on physical models and data-driven neural network methods.
[0003] Traditional time difference of location (TDOA) method uses geometric positioning by calculating the time delay of pulse signals arriving at different sensors. However, its accuracy depends heavily on the accuracy of time delay extraction and complex clock synchronization. In the complex GIS cavity, it is susceptible to multiple reflections and noise interference, resulting in large errors.
[0004] In recent years, neural networks have been introduced to improve positioning accuracy. Existing solutions share a common characteristic: first, they use the traditional time-of-flight method to calculate preliminary time delay values or location coordinates; then, they train an independent neural network model to compensate for or correct errors in the preliminary results. While this method improves accuracy to some extent, it essentially treats the neural network as a post-processing module of a traditional physical model, resulting in inherent flaws such as process fragmentation and error accumulation. Furthermore, neural networks do not directly learn the physical laws of electromagnetic wave propagation; their correction behavior lacks clear physical guidance, potentially leading to output results that violate fundamental physical or geometric constraints.
[0005] In addition, some other solutions attempt to use models such as convolutional neural networks to directly regress position coordinates from the signal end-to-end. While these purely data-driven methods are structurally simple, their training relies entirely on a large amount of labeled data, resulting in poor model interpretability. Furthermore, the trained network is a "black box," and its internal decision-making logic cannot be guaranteed to be consistent with the physical laws of electromagnetic wave propagation. When data is insufficient or unseen conditions are encountered, it may output physically unreasonable or even absurd positioning results. Summary of the Invention
[0006] 1. The technical problem to be solved:
[0007] How can we deeply integrate explicit physical laws into data-driven models to construct a positioning method that combines high precision and strong physical consistency, thereby solving the problems of poor physical consistency, limited accuracy, and strong dependence on a large amount of labeled data in existing local discharge source positioning methods?
[0008] 2. Technical Solution:
[0009] To address the above problems, this invention provides a GIS local discharge source location method based on a physical constraint neural network, comprising the following steps:
[0010] Step 1: Collect partial discharge signals received by multiple sensors arranged in the cavity of the electrical equipment, preprocess the signals, and construct an input feature vector based on them. The feature vector is composed of two parts concatenated in sequence. The first part is the energy and peak intensity characteristics of each sensor signal, and the second part is the peak value of the normalized cross-correlation function of the signal waveforms between all possible sensor pairs.
[0011] Step 2: Construct a neural network model for discharge source coordinate regression.
[0012] Step 3: Design a joint loss function for training the neural network model. The joint loss function is composed of a weighted sum of coordinate prediction loss term, time delay consistency physical constraint loss term, and geometric boundary constraint loss term.
[0013] Step 4: Train the neural network model by minimizing the joint loss function, so that the model can simultaneously optimize the coordinate prediction accuracy and embed the physical laws of electromagnetic wave propagation and the spatial geometric constraints of the equipment during the training process.
[0014] Step 5: Input the partial discharge signal to be located into the trained neural network model, and directly output the spatial three-dimensional coordinates of the discharge source.
[0015] In step 1, the specific method for collecting partial discharge signals received by multiple sensors arranged inside the electrical equipment cavity is as follows: A typical coaxial GIS cavity model is established, and within the insulating space of this cavity, M=1000 known discharge points are randomly generated. As sample labels, N=8 sensors were arranged at multiple basin-type insulators of the GIS cavity model, and their precise coordinates were recorded. For each sample discharge point, the signal of its discharge pulse propagating to the above N sensors is simulated and calculated to obtain the time series of electric field amplitude of each sensor. .
[0016] The feature vector Constructed via the following method: signal sequence for each sensor Calculate its total energy With peak amplitude The energy and peak values of all N=8 sensors together constitute a 16-dimensional feature. Subsequently, for each pair of sensors... and ,in The signal is processed by removing the mean to obtain Then, its normalized cross-correlation function is calculated. :
[0017]
[0018] Take the maximum value of the absolute value of the function. As a scalar feature characterizing the overall similarity of the signal waveforms, all 28 pairs of sensors yielded a total of 28-dimensional features.
[0019] The 16-dimensional intensity feature vector and the 28-dimensional cross-correlation feature vector are concatenated in a predefined fixed order to form a 44-dimensional feature vector, resulting in a 44-dimensional feature vector. Dataset consisting of corresponding coordinate labels .
[0020] The number of input layer nodes of the neural network model and the aforementioned 44-dimensional feature vector The dimensions are equal, and the output layer has 3 nodes, directly corresponding to the power supply. The network consists of two hidden layers with 128 and 64 neurons respectively, and uses the ReLU activation function.
[0021] The joint loss function expression mentioned in step 3 is:
[0022]
[0023] in, To predict the mean square error between the actual coordinates and the predicted coordinates, and For the introduced physical constraint terms, and These are the balancing weighting coefficients.
[0024] The delay consistency constraint loss The calculation is as follows: Based on the coordinate positions predicted by the neural network model, calculate the theoretical distances to each sensor, and thus obtain the theoretical signal arrival time difference between any two sensors; compare this theoretical time delay difference with the actual measurement time delay difference extracted from the corresponding sensor signals, and the mean square error is the mean square error. The specific method includes the following steps:
[0025] Step S61: Based on the network's current prediction of the discharge source coordinates Calculate the theoretical Euclidean distance to the two sensors. Thus, the theoretical time delay difference is obtained. ,in Since the speed of light is constant, the theoretical time delay difference must be consistent with the measured time delay difference.
[0026] Step S62: Obtain the actual measured time delay difference :
[0027] ,
[0028] in, It is to calculate the original electric field amplitude sequence of the sensor. and The cross-correlation function between the two.
[0029] Time shift difference Used for physical constraint calculations, the cross-correlation peak value as input feature It is a scalar that characterizes the overall similarity of waveforms.
[0030] Step S63: Delay Consistency Constraint Loss Defined as the mean square error of the difference between the theoretical and measured time delay differences for all sensor pairs:
[0031]
[0032] in This represents the total number of sensor pairs.
[0033] The geometric boundary constraint loss Calculate as follows:
[0034]
[0035] in, The radial distance of the predicted point. The function ensures that a positive penalty is only incurred when the coordinates go out of bounds. The weights for each penalty are determined.
[0036] The parameters of the neural network model are updated using a gradient descent-based optimization algorithm, with the optimization objective being to optimize the joint loss function. Minimize the value.
[0037] The present invention also provides a localization system for a partially discharged power source based on a physically constrained neural network, characterized in that it comprises:
[0038] The signal acquisition and processing unit is used to acquire and preprocess partial discharge signals from multiple sensors to form a feature vector;
[0039] A neural network localization unit is used to load and run the neural network model trained by the GIS local discharge source localization method based on physical constraint neural network according to any one of claims 1-8, so as to receive the feature vector and output the predicted coordinates;
[0040] The result display unit is used to display the predicted coordinates.
[0041] 3. Beneficial effects:
[0042] This invention creatively integrates time-delay consistency physical constraints and equipment geometric boundary constraints into the training process of a neural network in the form of a differentiable loss function. This method abandons the complex traditional process of "physical calculation first, then neural network correction," achieving end-to-end direct coordinate regression. By jointly optimizing data fitting terms and physical constraint terms, this invention forces the neural network to learn while adhering to physical laws, thus ensuring that its output results naturally possess physical rationality. This method not only significantly improves positioning accuracy but also reduces dependence on massive amounts of high-precision labeled data, enhancing the model's generalization ability and robustness in data-scarce or noisy environments, providing a high-precision, highly interpretable intelligent positioning solution for electrical equipment condition monitoring. Attached Figure Description
[0043] Figure 1 This is an overall principle block diagram of a local discharge source localization method based on a physical constraint neural network according to an embodiment of the present invention.
[0044] Figure 2 This is a schematic diagram of a partial discharge power source positioning system according to an embodiment of the present invention.
[0045] Figure 3 This is a schematic diagram of a physical constraint principle according to an embodiment of the present invention; wherein, Figure 3 (a) is a schematic diagram of the delay consistency constraint principle. Figure 3 (b) is a schematic diagram of geometric boundary constraints. Detailed Implementation
[0046] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0047] This embodiment uses a typical gas-insulated switchgear (GIS) coaxial cylindrical cavity as an application scenario to describe the implementation process of the present invention in detail. (Refer to...) Figure 1 The overall principle block diagram of the present invention shown herein includes the following implementation steps in this embodiment:
[0048] Step 1: Collect partial discharge signals received by multiple sensors arranged in the cavity of the electrical equipment, preprocess the signals, and construct an input feature vector based on them. The feature vector is composed of two parts concatenated in sequence. The first part is the energy and peak intensity characteristics of each sensor signal, and the second part is the peak value of the normalized cross-correlation function of the signal waveforms between all possible sensor pairs.
[0049] In one embodiment, the specific method for collecting partial discharge signals received by multiple sensors arranged inside the electrical equipment cavity is as follows: A typical coaxial GIS cavity model is established using electromagnetic simulation software (such as COMSOL Multiphysics), with a length L = 2.5 meters and a three-phase inner conductor radius R.in =0.035 meters, outer conductor radius R out =0.25 meters. Within the insulating space of this cavity, M=1000 discharge points with known locations are randomly generated. As a sample label.
[0050] To ensure comprehensive model training, the sample discharge points should be distributed as evenly as possible within the effective space of the GIS cavity.
[0051] N=8 ultra-high frequency (UHF) sensors are arranged at multiple basin-type insulators in the GIS cavity model, and their precise coordinates are recorded. This is a standard and effective location for detecting partial discharge electromagnetic signals in engineering. For each sample discharge point, the signal of its discharge pulse propagating to the above N sensors is simulated and calculated to obtain the time series of the electric field amplitude of each sensor. Typically, the sampling duration is 20 nanoseconds.
[0052] This embodiment, as a specific implementation of the feature vector, provides a preferred 44-dimensional feature vector construction scheme. Specifically, for each discharge source sample, its input neural network generates a 44-dimensional comprehensive feature vector. Constructed via the following method:
[0053] For each sensor's signal sequence Calculate its total energy With peak amplitude The energy and peak values of all N=8 sensors constitute a 16-dimensional feature. Subsequently, for each pair of sensors... and (in The signal is processed by removing the mean to obtain... Then, its normalized cross-correlation function is calculated. :
[0054]
[0055] Take the maximum value of the absolute value of the function. As a scalar feature characterizing the overall similarity of the signal waveforms, all 28 pairs of sensors yielded a total of 28-dimensional features.
[0056] Finally, the 16-dimensional intensity feature vector and the 28-dimensional cross-correlation feature vector are concatenated in a predefined fixed order to form a 44-dimensional feature vector. Therefore, we obtain a 44-dimensional feature vector. Dataset consisting of corresponding coordinate labels This is used for subsequent network training.
[0057] Step 2: Construct a neural network model for discharge source coordinate regression.
[0058] In one embodiment, the number of input layer nodes of the neural network model is equal to the aforementioned 44-dimensional feature vector. The dimensions are equal, and the output layer has 3 nodes, directly corresponding to the power supply. The network consists of two hidden layers with 128 and 64 neurons respectively, and uses the ReLU activation function.
[0059] Step 3: Design a joint loss function for training the neural network model. The joint loss function is composed of a weighted sum of coordinate prediction loss term, time delay consistency physical constraint loss term, and geometric boundary constraint loss term.
[0060] The joint loss function is expressed as follows:
[0061]
[0062] in, To predict the mean square error between the actual coordinates and the predicted coordinates, and For the introduced physical constraint terms, and These are the balancing weighting coefficients.
[0063] In one embodiment, and Take values of 0.5 and 0.1 respectively.
[0064] Traditional time-of-arrival (TOA) methods suffer from errors in delay calculation due to hardware and algorithm limitations. This invention embeds delay consistency as a soft constraint into the network training process, guiding network learning through a joint loss function. Based on the coordinate positions predicted by the neural network model, the theoretical distances to each sensor are calculated, thus obtaining the theoretical signal arrival time difference between any two sensors. This theoretical delay difference is then compared with the actual measured delay difference extracted from the corresponding sensor signals; the mean square error of this comparison is the mean square error. Specifically, for each pair of sensors and The latency consistency constraint loss is constructed as follows:
[0065] First, based on the network's current predicted discharge source coordinates... Calculate the theoretical Euclidean distance to the two sensors. Thus, the theoretical time delay difference is obtained. ,in For the speed of light, such as Figure 3 As shown in (a), the theoretical time delay difference must be consistent with the measured time delay difference. Secondly, to obtain the actual measured time delay difference... From the simulated sensor original electric field amplitude sequence and In the middle, by calculating the cross-correlation function of the two. And find the time shift that makes the function reach its global maximum, i.e. It is important to emphasize that the time shift value here... Used for physical constraint calculations, and the cross-correlation peak value as an input feature. It is a scalar representing the overall similarity of waveforms, and its physical meaning and role in the network are fundamentally different. Ultimately, the delay consistency constraint loss... Defined as the mean square error of the difference between the theoretical and measured time delay differences for all sensor pairs:
[0066]
[0067] in This represents the total number of sensor pairs.
[0068] like Figure 3 As shown in (b), the predicted coordinates must be located within the cavity's geometric boundaries. To prevent the network from outputting physically impossible locations, this invention explicitly incorporates geometric boundary constraints. The loss function penalizes predictions that violate known cavity dimensions, as shown in the following formula:
[0069]
[0070] in, The radial distance of the predicted point. The function ensures that a positive penalty is only incurred when the coordinates go out of bounds. The weights for each penalty are determined.
[0071] Step 4: Train the neural network model by minimizing the joint loss function, so that the model can simultaneously optimize the coordinate prediction accuracy and embed the physical laws of electromagnetic wave propagation and the spatial geometric constraints of the equipment during the training process.
[0072] The dataset is divided into training, validation, and test sets. The training set data is used to minimize the joint loss function. To achieve this, the Adam optimizer is used to train the network. Performance is monitored on the validation set during training to prevent overfitting.
[0073] Step 5: Input the partial discharge signal to be located into the trained neural network model, and directly output the spatial three-dimensional coordinates of the discharge source.
[0074] To verify the effectiveness of this invention, a program was trained using only... A loss-generating pure data-driven network was used as a control. Evaluation results on the same test set show that the average positioning error of the proposed method is significantly lower than that of the pure data-driven network, and all predicted coordinates satisfy geometric boundary constraints, demonstrating the dual advantages of this method in improving accuracy and ensuring physical plausibility.
[0075] Accordingly, the present invention also provides a partial discharge source positioning system based on the above method. For example... Figure 2 As shown, the system includes:
[0076] The signal acquisition and processing module, corresponding to step 1 above, is used to realize the synchronous acquisition and preprocessing of multiple sensor signals and the construction of the 44-dimensional feature vector.
[0077] The physical constraint neural network localization module, corresponding to steps 2-4 above, internally contains the trained physical constraint neural network model and the joint loss function calculation unit, which is used to receive feature vectors and output predicted coordinates.
[0078] The results display module, corresponding to step 5 above, is used to receive and visualize the positioning results.
[0079] The modules of the system are connected in sequence and work together to achieve the aforementioned high-precision, highly physically consistent end-to-end positioning function.
Claims
1. A GIS local discharge source location method based on a physically constrained neural network, characterized in that: Includes the following steps: Step 1: Collect partial discharge signals received by multiple sensors arranged in the cavity of electrical equipment, preprocess the signals, and construct an input feature vector accordingly. The feature vector is composed of two parts concatenated in sequence: the first part is the energy and peak intensity characteristics of each sensor signal, and the second part is the peak value of the normalized cross-correlation function of the signal waveforms between all possible sensor pairs. Step 2: Construct a neural network model for discharge source coordinate regression; Step 3: Design a joint loss function for training the neural network model. The joint loss function is composed of a weighted sum of a coordinate prediction loss term, a time delay consistency physical constraint loss term, and a geometric boundary constraint loss term. Step 4: Train the neural network model by minimizing the joint loss function, so that the model can simultaneously optimize the coordinate prediction accuracy and embed the physical laws of electromagnetic wave propagation and the spatial geometric constraints of the equipment during the training process; Step 5: Input the partial discharge signal to be located into the trained neural network model, and directly output the spatial three-dimensional coordinates of the discharge source.
2. The GIS local discharge source positioning method based on physical constraint neural network as described in claim 1, characterized in that: In step 1, the specific method for collecting partial discharge signals received by multiple sensors arranged inside the electrical equipment cavity is as follows: A typical coaxial GIS cavity model is established, and within the insulating space of this cavity, M=1000 known discharge points are randomly generated. As sample labels, N=8 sensors were arranged at multiple basin-type insulators of the GIS cavity model, and their precise coordinates were recorded. For each sample discharge point, the signal of its discharge pulse propagating to the above N sensors is simulated and calculated to obtain the time series of electric field amplitude of each sensor. .
3. The GIS local discharge source positioning method based on physical constraint neural network as described in claim 2, characterized in that: The feature vector Constructed via the following method: signal sequence for each sensor Calculate its total energy With peak amplitude The energy and peak values of all N=8 sensors together constitute a 16-dimensional feature. Subsequently, for each pair of sensors... and ,in The signal is processed by removing the mean to obtain Then, its normalized cross-correlation function is calculated. : Take the maximum value of the absolute value of the function. As a scalar feature characterizing the overall similarity of the signal waveforms, all 28 pairs of sensors yielded a total of 28-dimensional features. The 16-dimensional intensity feature vector and the 28-dimensional cross-correlation feature vector are concatenated in a predefined fixed order to form a 44-dimensional feature vector, resulting in a 44-dimensional feature vector. Data set consisting of corresponding coordinate labels .
4. The GIS local discharge source positioning method based on physical constraint neural network as described in claim 3, characterized in that: In step 2, the number of input layer nodes of the neural network model and the aforementioned 44-dimensional feature vector... The dimensions are equal, and the output layer has 3 nodes, directly corresponding to the power supply. The network consists of two hidden layers with 128 and 64 neurons respectively, and uses the ReLU activation function.
5. The GIS local discharge source positioning method based on physical constraint neural network as described in claim 4, characterized in that: The joint loss function expression mentioned in step 3 is: in, To predict the mean square error between the actual coordinates and the predicted coordinates, and For the introduced physical constraint terms, and These are the balancing weighting coefficients.
6. The GIS local discharge source positioning method based on physical constraint neural network as described in claim 5, characterized in that: The delay consistency constraint loss The calculation is as follows: Based on the coordinate positions predicted by the neural network model, calculate the theoretical distances to each sensor, and thus obtain the theoretical signal arrival time difference between any two sensors; compare this theoretical time delay difference with the actual measurement time delay difference extracted from the corresponding sensor signals, and the mean square error is the mean square error. The specific method includes the following steps: Step S61: Based on the network's current prediction of the discharge source coordinates Calculate the theoretical Euclidean distance to the two sensors. Thus, the theoretical time delay difference is obtained. ,in Since the speed of light is constant, the theoretical time delay difference must be consistent with the measured time delay difference. Step S62: To obtain the actual measurement time delay difference , , in, It is to calculate the original electric field amplitude sequence of the sensor. and The cross-correlation function of the two, the time shift difference Used for physical constraint calculations, the cross-correlation peak value as input feature It is a scalar that characterizes the overall similarity of waveforms; Step S63: Delay Consistency Constraint Loss Defined as the mean square error of the difference between the theoretical and measured time delay differences for all sensor pairs: in This represents the total number of sensor pairs.
7. The GIS local discharge source positioning method based on physical constraint neural network as described in claim 6, characterized in that: The geometric boundary constraint loss Calculate as follows: in, The radial distance of the predicted point. The function ensures that a positive penalty is only incurred when the coordinates go out of bounds. The weights for each penalty are determined.
8. The GIS local discharge source positioning method based on physical constraint neural network as described in claim 7, characterized in that: The parameters of the neural network model are updated using a gradient descent-based optimization algorithm, with the optimization objective being to optimize the joint loss function. Minimize the value.
9. A localization system for a partial discharge source based on a physically constrained neural network, characterized in that, include: The signal acquisition and processing unit is used to acquire and preprocess partial discharge signals from multiple sensors to form a feature vector; A neural network localization unit is used to load and run the neural network model trained by the GIS local discharge source localization method based on physical constraint neural network according to any one of claims 1-8, so as to receive the feature vector and output the predicted coordinates; The result display unit is used to display the predicted coordinates.