A fault identification method, system, device and medium for a smoothing reactor

Abstract

The application discloses a kind of flat wave reactors fault identification method, system, equipment and medium, it is related to electric power equipment detection field, the method includes according to the structure of target reactor equipment and constructs geometric model, utilizes predefined mask function on geometric model and identifies material sub-region, the material parameters corresponding to each point in area;Real-time acquisition target reactor equipment of training data, according to material parameters and training data and constructs constraint loss function;Training data are input into integrated temperature branch and heat source branch physical information neural network and are trained, and the temperature distribution and heat source distribution of target reactor equipment are output to obtain;Corresponding mask function is injected into heat source branch and is limited in area, and constraint loss function is introduced to carry out parameter correction;According to temperature distribution and heat source distribution, target reactor equipment is carried out fault identification, to formulate operation and maintenance strategy and execute, improve the accuracy of fault identification.

Description

Technical Field

[0001] This invention relates to the field of power equipment testing, and in particular to a method, system, equipment and medium for identifying faults in smoothing reactors. Background Technology

[0002] Smoothing reactors are key equipment in power systems. Their operating temperature distribution is directly related to the heat source intensity, insulation aging rate, and potential fault risk. Accurate reactor temperature analysis is the core basis for achieving early fault diagnosis and precise condition assessment.

[0003] Traditional analysis methods, such as simulation models, require the construction of complex multiphysics models. The modeling cycle is long and it is difficult to respond quickly to changes in operating conditions. It cannot provide dynamic temperature field data for fault diagnosis in a timely manner, which leads to errors in fault identification and directly affects the reliability of condition assessment.

[0004] Therefore, how to effectively analyze the faults of smoothing reactors and improve the stability of power system operation has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method, system, device, and medium for fault identification of smoothing reactors, which solves the problem of how to divide different geometric regions through masking technology and combine physical information neural network PINN to predict reactor temperature and invert heat sources, thereby improving the accuracy of prediction results.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a method for identifying faults in smoothing reactors, comprising:

[0007] A geometric model is constructed based on the structure of the target reactor device. A predefined mask function is used to identify material sub-regions on the geometric model to generate material parameters corresponding to each point within the material sub-regions.

[0008] Real-time acquisition of training data for the target reactor equipment; and construction of a constrained loss function based on the material parameters and the training data.

[0009] The data to be trained is input into a physical information neural network that integrates temperature and heat source branches for phased training, and the output is the temperature distribution and heat source distribution of the target reactor device.

[0010] During training, the corresponding mask function is injected into the heat source branch to restrict the region, and the constraint loss function is introduced to correct the parameters of the physical information neural network.

[0011] Based on the temperature distribution and heat source distribution, faults in the target reactor equipment are identified in order to formulate and implement operation and maintenance strategies.

[0012] Furthermore, the step of constructing a geometric model based on the structure of the target reactor device and identifying material sub-regions on the geometric model using a predefined mask function includes:

[0013] Based on the winding structure information and encapsulation structure information of the target reactor equipment, a geometric model reflecting the two-dimensional rectangular solution is constructed;

[0014] The coordinates of any point on the geometric model are input into the mask function to obtain the corresponding material sub-region; the material sub-region reflects the aluminum material region or epoxy material region in the target reactor equipment.

[0015] Furthermore, the real-time acquisition of training data for the target reactor equipment, and the construction of a constrained loss function based on the material parameters and the training data, includes:

[0016] Based on the material parameters, establish the transient heat conduction equations for each material sub-region, the interface constraint equations for each material sub-region contact surface, the boundary constraint equations for each material sub-region boundary surface, and the initial constraint equations to generate the corresponding physical constraint losses.

[0017] Using multi-source sensors deployed on the target reactor equipment, temperature data, ambient environmental data, and electrical data are collected in real time to construct the training data;

[0018] The training data is normalized to construct the corresponding data constraint loss.

[0019] The constraint loss function is obtained by integrating the physical constraint loss and the data constraint loss.

[0020] Furthermore, the step of inputting the data to be trained into a physical information neural network integrating temperature and heat source branches for phased training includes:

[0021] The training data is input into the physical information neural network. During the first stage of training, the training parameters corresponding to the heat source branch are frozen, and the temperature branch is trained independently.

[0022] During the second phase of training, the heat source branch is unlocked, and the temperature branch and the heat source branch are trained simultaneously.

[0023] Furthermore, the step of introducing the constraint loss function to correct the parameters of the physical information neural network includes:

[0024] At each iteration, the moving average of each loss term in the constraint loss function is calculated;

[0025] The weights corresponding to each loss term are dynamically updated based on the moving average value to correct the parameters of the physical information neural network.

[0026] Furthermore, the step of identifying faults in the target reactor equipment based on the temperature distribution and the heat source distribution to formulate and execute an operation and maintenance strategy includes:

[0027] Feature extraction is performed on the temperature distribution to obtain the key temperature characteristics of the target reactor equipment;

[0028] The key temperature characteristics are analyzed using a preset abnormal temperature judgment mechanism to determine the operating risk level of the target reactor equipment.

[0029] Key heat source features of the target reactor equipment are extracted from the heat source distribution, and the key heat source features are matched with a pre-established fault mode library to determine the target fault mode corresponding to the target reactor equipment.

[0030] Integrate the work risk level and the target failure mode to formulate and implement corresponding operation and maintenance strategies.

[0031] Furthermore, the step of performing similarity matching between the key heat source features and a pre-established fault mode library to determine the fault information corresponding to the target reactor equipment also includes:

[0032] The similarity between the key heat source features and the samples in the fault mode library is calculated to obtain the initial similarity result;

[0033] When the initial similarity result meets the preset conditions, the heat source distribution and the initial similarity result are input into a predetermined classifier for processing, and the target fault mode is output.

[0034] Another embodiment of the present invention provides a smoothing reactor fault identification system, comprising:

[0035] The parameter determination module is used to construct a geometric model based on the structure of the target reactor device, and to identify material sub-regions on the geometric model using a predefined mask function, so as to generate material parameters corresponding to each point within the material sub-regions.

[0036] The constraint establishment module is used to collect the training data of the target reactor equipment in real time and construct a constraint loss function based on the material parameters and the training data.

[0037] The training module is used to input the data to be trained into a physical information neural network that integrates temperature and heat source branches for phased training, and output the temperature distribution and heat source distribution of the target reactor device.

[0038] The correction module is used to inject the corresponding mask function into the heat source branch for region restriction during training, and to introduce the constraint loss function to correct the parameters of the physical information neural network.

[0039] The fault identification module is used to identify faults in the target reactor equipment based on the temperature distribution and the heat source distribution, so as to formulate and execute operation and maintenance strategies.

[0040] Another embodiment of the present invention provides a computer device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the smoothing reactor fault identification method as described above.

[0041] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the smoothing reactor fault identification method as described above.

[0042] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0043] This invention effectively improves the physical rationality and accuracy of temperature field prediction and heat source inversion by using a mask function to identify material regions and construct physical constraints in the geometric model of the equipment. A dual-branch physical information neural network structure and a two-stage training strategy are designed to overcome the ill-conditioned problem of heat source inversion, enhancing the convergence and stability of the model. The training process is optimized by combining physical and data constraints, improving solution efficiency and the network's generalization ability. Finally, a multi-level fault diagnosis mechanism based on temperature and heat source distribution significantly improves the accuracy of fault identification, providing a reliable guarantee for the safe operation of power equipment. Attached Figure Description

[0044] Figure 1 This is a schematic flowchart of a smoothing reactor fault identification method in one embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the geometric model of the smoothing reactor in one embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the PINN network architecture in one embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of the structure of a smoothing reactor fault identification system in one embodiment of the present invention;

[0048] Figure 5This is a structural block diagram of a preferred embodiment of a computer device provided by the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0050] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0051] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0052] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0053] One embodiment of the present invention provides a method for fault identification of smoothing reactors. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1The diagram shown is a flowchart of a smoothing reactor fault identification method according to one embodiment of the present invention, including the following steps:

[0054] S1. Construct a geometric model based on the structure of the target reactor equipment, and use a predefined mask function to identify material sub-regions on the geometric model in order to generate material parameters corresponding to each point within the material sub-regions.

[0055] In this embodiment, a smoothing reactor is selected as the target reactor device for temperature field modeling to achieve fault diagnosis and subsequent operation and maintenance.

[0056] First, a geometric model of the smoothing reactor needs to be performed. It should be understood that the key structures of the smoothing reactor include the winding and encapsulation structures. Therefore, in this embodiment, a geometric model reflecting the two-dimensional rectangular solution domain is constructed based on the winding structure information and encapsulation structure information of the target reactor equipment.

[0057] For example, the geometric dimensions and material distribution corresponding to the winding structure and the encapsulation structure can be obtained separately to establish a two-dimensional geometric model. Specifically, for example... Figure 2 As shown, this embodiment uses solid conduction and boundary convection to model the temperature field of the smoothing reactor. Geometrically, each conductor of the winding is unfolded into several "concentric cylindrical shells" or 2D axisymmetrically represented by a two-dimensional cross-section. The encapsulation layer is also made into several shells according to its thickness. Then, a geometric model of the composite structure of the smoothing reactor is constructed. Taking a single-layer encapsulation as an example, the modeled geometric model consists of an inner rectangle of aluminum material and an outer layer of epoxy material. The inner rectangular aluminum is wrapped with epoxy material.

[0058] The transient thermal conductivity (PDE) and other characteristics inside a smoothing reactor have the same physical form in different material regions, but the required material parameters differ. Specifically, due to the presence of two different materials in the smoothing reactor—aluminum conductor and epoxy—the thermal conductivity, density, specific heat capacity, and other material parameters differ in different material regions. Furthermore, the heat source is distributed in the aluminum region but is approximately zero in the epoxy region. Based on this, this embodiment uses a mask function / mask matrix to distinguish them in a two-dimensional rectangular solution domain. It should be understood that a mask matrix is ​​a two-dimensional function defined in the computational domain, with values ​​typically between [0,1], used to indicate which material a point belongs to. It is constructed in the geometric model based on the boundary positions. In this embodiment, it is used to distinguish between the aluminum conductor and epoxy insulation regions within a unified coordinate system. Based on this, by further inputting the coordinates of any point on the geometric model into the mask function, the corresponding material sub-region is identified. The material sub-region reflects either the aluminum material region or the epoxy material region in the target reactor device.

[0059] For example, a binary mask can be used:

[0060]

[0061] Or a soft mask:

[0062]

[0063] in, For the Sigmoid function, The boundary description function is determined by the geometric parameters. is the boundary smoothing factor, and (x,y) are the coordinates of the point.

[0064] For example, when calculating the PDE residual, the corresponding material parameters are represented by the mask selection as follows:

[0065]

[0066] in, For any coordinate point in the geometric model region Material density; Density of aluminum material; This represents the density of the epoxy material.

[0067] S2. Real-time acquisition of training data for the target reactor equipment, and construction of a constrained loss function based on material parameters and training data.

[0068] It should be understood that this constraint loss function is a constraint on the training process of the Physical Information Neural Network (PINN). In this embodiment, on the one hand, by designing a physical constraint loss, physical laws (PDEs, boundary conditions, interface conditions, etc.) are imposed as constraints on the prediction results of PINN, so that the temperature field output by PINN is consistent with physical laws. These physical constraints are written into the loss function of PINN. In each training iteration, it will automatically check whether the predicted temperature field satisfies these constraints. If not, the parameters will be adjusted through backpropagation.

[0069] Specifically, this embodiment establishes the transient heat conduction PDE equations for each material sub-region, the interface constraint equations for the contact surfaces of each material sub-region, the boundary constraint equations for the boundary surfaces of each material sub-region, and the initial constraint equations based on the material parameters. The following example details the process of establishing these four constraint equations:

[0070] In this example, the specific parameters of the two-dimensional geometric model are as follows: the width of the inner rectangle made of aluminum is (103.4 - 2 × 2.13) mm, and the height is (3594.3 - 2 × 2.13) mm. Epoxy material is wrapped around the aluminum, with a single-sided thickness of 2.13 mm, forming a two-dimensional rectangular encapsulation structure with an overall size of 103.4 mm in width and 3594.3 mm in height. Other material parameters are shown in Table 1 below.

[0071] Table 1 Examples of Material Parameters

[0072]

[0073] Because for the entire solid region (aluminum winding + cladding), without explicitly characterizing the airflow, the temperature field... If the classic transient heat conduction equation is satisfied, then the transient heat conduction PDE equations are constructed separately for aluminum and epoxy materials. The PDE equation for aluminum is as follows:

[0074]

[0075] The PDE equation for epoxy materials is:

[0076]

[0077] In the formula, , , These are the density, specific heat capacity, and thermal conductivity of aluminum, respectively. , , These are the density, specific heat capacity, and thermal conductivity of epoxy resin, respectively. For temperature, For time, The heat source intensity in aluminum.

[0078] Since the interface between aluminum and epoxy satisfies both temperature continuity and heat flux continuity, the interface constraint equation can be expressed as:

[0079]

[0080]

[0081] in, , These are the interface temperatures of aluminum and epoxy, respectively. This is the interface normal vector.

[0082] On the outer surface boundary where the encapsulation layer is in direct contact with air, this embodiment uses Newton's law of cooling (convection) to characterize the boundary. That is, the contact boundary between the epoxy and air satisfies Newton's law of cooling, and the boundary constraint equation is expressed as:

[0083]

[0084] In the formula, The convective heat transfer coefficient is... This refers to the air temperature.

[0085] The initial constraint equations are expressed across the entire solution domain, with the initial temperature set as follows:

[0086]

[0087] in, The initial temperatures of the aluminum and epoxy materials are given.

[0088] Based on these four constraint equations, the corresponding physical constraint loss is generated, where the PDE loss function is expressed as:

[0089]

[0090] in, and These represent the number of sampling points in the aluminum and epoxy regions, respectively. The residual for the aluminum region is , The residual of the epoxy region is , , , These are the density, specific heat capacity, and thermal conductivity of aluminum, respectively. , , These are the density, specific heat capacity, and thermal conductivity of epoxy resin, respectively. For temperature, For time, The heat source intensity in aluminum.

[0091] The interface conditional loss function is expressed as:

[0092]

[0093] in, The number of interface sampling points. For the continuous temperature residual is , For continuous residual of heat flux ( (for the interface normal vector) .

[0094] The boundary loss function is expressed as:

[0095]

[0096] in, Number of boundary sampling points The residual of Newton's law of cooling ( (the boundary outside normal vector) is , The convective heat transfer coefficient is... This refers to the air temperature.

[0097] The initial loss function is expressed as:

[0098]

[0099] in, Number of initial condition sampling points The initial temperature residual is , The initial temperature.

[0100] Next, a data constraint loss is established. In this embodiment, sensor data is used as the observed value to enter the loss function to constrain the predicted temperature distribution of PINN, so that the prediction result is consistent with the actual measurement.

[0101] Specifically, multi-source sensors deployed on the target reactor device are used to collect temperature data, ambient environmental data, and electrical data in real time to construct training data; in this embodiment, this is used as observation data. For example, temperature sensors can be arranged on the encapsulation surface and at several interface points. The number of sensors depends on engineering conditions; their distribution in key locations ensures inversion reliability, and the more observation points, the better the inversion stability. Simultaneously, the ambient temperature is monitored and recorded. With operating current ,frequency Equal electrical quantities, used for heat sources Prior associations. An example format of the observation data is shown in Table 2 below:

[0102] Table 2 Observation Data

[0103]

[0104] In Table 2, the first column is time, the second column is ambient air temperature, the third column is operating current, and the fourth column is frequency. The fifth column and subsequent columns are sensor data, grouped into sets of three (x-coordinate, y-coordinate, temperature value).

[0105] Then, these observation data are normalized; specifically, the spatial coordinates are normalized. Normalize the geometric dimensions to [0,1] to obtain , means as follows:

[0106]

[0107] In the formula, For geometric models in Maximum dimension in the direction, For geometric models in The maximum dimension in the direction is determined, and the normalized derivative is scaled accordingly in the PDE.

[0108] Regarding time Normalized to [0,1] by the maximum monitoring duration, we get , means as follows:

[0109]

[0110] In the formula, here This is the maximum monitoring duration.

[0111] Regarding temperature ,according to Normalize.

[0112] In the formula, and These are the minimum and maximum temperatures in the observed dataset, respectively.

[0113] Then, based on this normalized data, a corresponding data constraint loss is constructed to ensure that the error between the model's predicted values ​​and the known measurement data is minimized. Finally, the physical constraint loss and the data constraint loss are integrated to obtain the total constraint loss function. , means as follows:

[0114]

[0115] In the formula, , , , , These are the weights of each loss item; This is due to data constraint loss.

[0116] Among them, data constraint loss Specifically, it is expressed as follows:

[0117]

[0118] in, It is the sample size of the temperature measurement data. It is the temperature predicted by the neural network. The measured temperature is known.

[0119] S3. Input the data to be trained into the PINN network that integrates the temperature branch and the heat source branch for phased training, and output the temperature distribution and heat source distribution of the target reactor device.

[0120] In this embodiment, PINN is designed as a fully connected deep neural network, and the network structure diagram is as follows: Figure 3 As shown, the overall structure consists of an input layer, hidden layers, and an output layer; the network has a total of L layers (excluding the input layer, the l-th layer is the first layer). The system consists of hidden layers, a shared backbone with a depth of 3-6 layers and a width of 64-256. Tanh or Swish activation functions are used. The system is then divided into two output branches (Temperature T branch and Heat Source Q branch), each with M layers. The Temperature branch outputs the normalized temperature, and the Heat Source branch outputs the normalized heat source.

[0121] During the training process, since the temperature field of the smoothing reactor is constructed in two parts—prediction of temperature distribution and inversion of heat source—this embodiment designs two training stages to avoid the interference of "noise" heat source values ​​generated by the random initialization of parameters in the early stage of training on the temperature field learning.

[0122] Specifically, the normalized training data is input into the PINN network. Training and test data are also randomly sampled within the geometric model, at its boundaries, and at its interfaces, and used together for PINN training. The training data is used by PINN to calculate the total loss on the sampled training points during training, and backpropagation is used to drive gradient descent. The test data is a set of test points independent of the training process and not involved in backpropagation. PINN calculates the same loss on these independent test points to measure the model's generalization and the accuracy of solving PDEs.

[0123] For example, the total number of sampling points is approximately:

[0124]

[0125] in, This is a time point number, with a quantity of 50; These are the number of internal sampling points, the number of boundary sampling points, and the number of interface sampling points, totaling 2400. Therefore, the total number of sampling points is approximately 120,000.

[0126] During the first stage of training, the training parameters corresponding to the heat source branch are frozen, and the temperature branch is trained independently until the PDE and the physical constraint loss converge, thus learning a robust temperature-physical mapping and reducing the damage of the initial randomness of the heat source branch to the network. During the second stage of training, the heat source branch is unlocked, and the temperature branch and the heat source branch are trained simultaneously, forcing the physical priors of the heat source branch into the network structure or loss.

[0127] For example, the PINN network structure can be configured as a 3-layer shared backbone (64 nodes per layer), 2 layers of temperature branches, and 2 layers of heat source branches, as shown below:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134] in, Normalized spacetime coordinates , For the output of the activation layer, This is the weight matrix. For bias vectors, For activation function, This represents the normalized temperature output by the PINN network. This is the normalized heat source for the final output of the PINN network.

[0135] like Figure 3 As shown, in this embodiment, during the training of PINN, the backbone parameters are denoted as... The temperature branch parameter is denoted as The heat source branch parameters are denoted as Combined and recorded as The training objective is The Adam optimizer is used, combined with cosine annealing scheduling and a hot restart strategy. An appropriate initial learning rate is set, and the learning rate is decayed to 1 / 10 of the initial value every 1000 to 2000 iterations according to the cosine curve. After the decay is completed, the learning rate is reset to 50% of the initial value to avoid getting trapped in local optima.

[0136] S4. During training, the corresponding mask function is injected into the heat source branch to restrict the region, and a constraint loss function is introduced to correct the PINN network parameters.

[0137] It should be understood that in smoothing reactors, heat sources (mainly Joule losses, eddy current losses, and inter-turn circulating losses generated by the current) are only generated inside the aluminum conductor. Epoxy resin is an insulating material, and almost no heat is generated within it. Therefore, to force PINN to obey physical laws when predicting different materials, the output at the heat source branch... Multiplying with the corresponding mask function restricts the effective region. Even if the PINN network's heat source prediction process only applies to the aluminum region, the final heat source output... Represented as:

[0138]

[0139] The calculation process for the output results of the two branches is as follows:

[0140] For any layer Perform linear transformation and activation on the input:

[0141]

[0142]

[0143] in, For the first Weighted input of the layer, For the first Layer weights, For the first Layer bias, For the first Layer activation function, For the first Layer output, From the front Shared backbone layer, calculated .

[0144] Temperature branch:

[0145]

[0146] Final linear output:

[0147]

[0148] in, For the first The weight matrix of the layer, For the first The layer's bias vector, These are intermediate variables after the linear transformation. For the first Layer activation output, For activation function, This represents the number of layers in the temperature branch.

[0149] Heat source branch:

[0150]

[0151] Final linear output:

[0152]

[0153] in, For the heat source branch The weight matrix of the layer, For the heat source branch The layer's bias vector, These are intermediate variables after the linear transformation. For the first Layer activation output, For activation function, This represents the number of layers in the heat source branch.

[0154] In each iteration, the moving average of each loss term in the constrained loss function is calculated, and then the weights corresponding to each loss term are dynamically updated based on the moving average to correct the PINN network parameters.

[0155] For example, each The next iteration calculates the moving average of each loss term. Update weights:

[0156]

[0157] in, For the first The weights of each loss term, For the first The moving average of each loss term, Let be the overall mean of the moving average of all loss terms. To prevent small constants with a denominator of 0, n is the number of loss functions.

[0158] Then to By pruning and limiting the residuals to a certain range, the weights of excessively large residuals are reduced, while the weights of excessively small residuals are increased, thus promoting balance.

[0159] In some embodiments of the present invention, a dynamic adoption strategy is further designed to optimize the training process. In the initial stage, PDE collocation points, interface points and boundary points are uniformly sampled. After each iteration, the PDE residuals on these points are calculated, and regions with high residuals are selected for encrypted sampling. Meanwhile, a high sampling density is maintained for a long time at the interface and encapsulation surface (which are prone to generating high residuals).

[0160] S5. Based on temperature and heat source distribution, identify faults in the target reactor equipment in order to formulate and implement operation and maintenance strategies.

[0161] This step involves fault analysis based on the distribution results. Specifically, feature extraction is performed on the temperature distribution to obtain the key temperature characteristics of the target reactor equipment, including the highest temperature, average temperature rise, and hot spot temperature rise of the smoothing reactor. For example, a numerical search (point-by-point traversal) is performed on the predicted temperature distribution data to obtain the maxima. The average temperature rise is obtained by averaging the entire temperature distribution. ,in, For ambient temperature, The number of sampling points is used to obtain the hot spot temperature rise by measuring the difference between the highest temperature and the ambient temperature.

[0162] These key temperature characteristics are analyzed using a pre-defined abnormal temperature judgment mechanism to determine the corresponding operational risk level of the target reactor equipment. For example, considering a certain threshold and combining the health status classification standard and temperature rise limit standard for smoothing reactors, the operating status of the smoothing reactor is judged. The abnormal temperature judgment mechanism is as follows: when the hot spot temperature rise is less than 90K and the average temperature rise is less than 80K, it is defined as a normal operating state; when the hot spot temperature rise is greater than 90K but less than 115K or the average temperature rise is greater than 80K but less than 100K, it is defined as a warning or dangerous operating state; when the hot spot temperature rise exceeds 115K or the average temperature rise exceeds 100K, it is defined as a severe operating state.

[0163] Next, key heat source features of the target reactor are extracted from the heat source distribution. These key heat source features are then matched with a pre-established fault mode library to determine the target fault mode corresponding to the target reactor. To ensure the accuracy of fault analysis, this embodiment sends the similarity vector to a two-level discrimination system. The first level uses the fault mode library for rapid similarity judgment, while the second level uses a trained classifier for refined judgment.

[0164] Specifically: In the first level, the similarity between the key heat source features and samples in the fault mode library is calculated to obtain initial similarity results. The extracted key heat source features include: heat source peak value, peak location, total energy, and centroid. Among them, the peak value and peak location are obtained by searching for the maximum value and coordinates of the heat source distribution, the total energy is obtained by integrating the overall intensity of the heat source distribution, and the centroid feature is obtained by calculating the geometric center location of the hotspot using the centroid formula.

[0165] When the initial similarity result meets the preset conditions, the second-level discrimination is triggered. The heat source distribution and the initial similarity result are input into a predetermined classifier for processing, and the target fault mode is output.

[0166] Example: Calculate samples in the fault mode library Distance between key heat source characteristics (field examples) Similarity Simultaneously calculate composite similarity The distance calculation process is as follows:

[0167]

[0168]

[0169]

[0170]

[0171] In the formula, For on-site examples and the first in the library Peak dissimilarity (distance) of individual samples. The peak value of the heat source obtained from the inversion of the field sample. For the first in the library The peak heat source of each sample For on-site examples and the first in the library The difference (distance) in the peak position of each sample. and These are the coordinates of the location of the heat source peak in the field example. and For geometric models in , Feature dimensions in the direction, For on-site examples and the first in the library Total energy difference (distance) of heat sources for each sample The total heat source energy of the field sample. For the first in the library The total heat source energy of each sample For on-site examples and the first in the library Distance between the centroids of the heat sources of each sample. and The coordinates of the centroid of the heat source distribution in the field example are shown. and For the first in the library The centroid coordinates of the heat source distribution of each sample.

[0172] In the formula, To map distance to a monotonically decaying similarity function, specifically:

[0173] or

[0174] ,

[0175] in, The similarity index is obtained from distance transformation. For the distance defined above, To adjust the parameters and control the decay rate of the similarity function, As weight, ; The larger the value, the more the heat source distribution resembles Curry's. For each working condition, the maximum value is taken and represented as:

[0176]

[0177] Received For the most similar warehouse operating conditions, The similarity score is calculated.

[0178] Set the similarity threshold to ,like This indicates the current distribution of heat sources and the first [unit / item] in the reservoir. If the operating conditions are highly matched, the fault identification result is directly output. Otherwise, it triggers the second-level classifier for discrimination.

[0179] The second-level classifier uses CKAN, which takes the heat source distribution obtained by PINN inversion and adds the previously calculated composite similarity plus multiple individual similarities as input, classifies it through CKAN, and finally outputs the operating condition type.

[0180] Specifically, it refers to the vector representation of the main input heat source distribution. and one-dimensional similarity feature vector ,in, , , , These are peak similarity, positional similarity, energy similarity, and centroid similarity, respectively. Composite similarity; extracted using CNN features, including Layered convolution and nonlinear transformations, specifically:

[0181]

[0182] in, For convolution kernel weights, For bias terms, This represents the convolution operation. , The activation function is ReLU, and the output is a feature map. After further global pooling, it can be represented as:

[0183]

[0184] in, This is the global feature vector output by the CNN. This is a global average pooling operation.

[0185] The data is then fed into a KAN layer to learn high-dimensional nonlinear combinations of CNN features to better distinguish complex working conditions. The KAN mapping is as follows:

[0186]

[0187] in, The KAN layer outputs a feature vector. This is the KAN mapping function.

[0188] Among them, the first The components are:

[0189]

[0190] in, For univariate nonlinear basis functions, The combined weights are used to control the first The output and the first Linear combination relation of basis functions For projection vectors, As a bias; the KAN output is concatenated with auxiliary features to obtain: .

[0191] Finally, the Softmax classification output is as follows:

[0192] in, For the number of operating condition categories, For classification weights, For bias.

[0193] Softmax is defined as: ,in, The sample was judged as the first The probability of a class.

[0194] The loss function is cross-entropy, which is specifically expressed as: ,in, It's a real label.

[0195] The system integrates work risk levels and target failure modes to generate corresponding operation and maintenance strategies and executes them. During this process, the work risk levels, target failure modes, and operation and maintenance strategies are further summarized into corresponding fault diagnosis reports, which are sent to the operation and maintenance personnel's terminals and saved to the logs.

[0196] It should be understood that the fault mode library is pre-built through simulation (such as COMSOL) and contains standard heat source characteristic vectors under various typical operating conditions (normal, inter-turn short circuit, commutation failure, etc.).

[0197] Example: This embodiment uses typical parameters for a ±800kV smoothing reactor: rated voltage ±800 kV, rated DC current approximately 4000-5000 A, reactance approximately 85-185mH, operating frequency 50 Hz, ambient temperature range -25℃ to +45℃, and maximum allowable temperature rise of winding hot spots not exceeding 120-140℃. The parameters for each operating condition are set as follows:

[0198] Normal operating conditions: The operating current is set to 4500 A, the frequency to 50 Hz, and the ambient temperature to 25℃. The simulation shows that the heat source distribution is uniform, and the highest winding temperature is about 95℃, which does not exceed the allowable value. Maintenance strategy: It is judged to be in a normal state, and it is recommended to conduct regular inspections.

[0199] Inter-turn short-circuit fault condition: The short circuit is set to occur in the 12th layer coil, with a short-circuited turn ratio of 3%, an operating current of 4500A, and an ambient temperature of 25℃. Simulation results show that the local heat source is concentrated at the short-circuit location, and the hot spot temperature rises rapidly to 160℃. The fault is identified as "local inter-turn short circuit". The maintenance strategy is to immediately shut down the system and arrange for the replacement of the faulty winding.

[0200] Commutation failure condition: The operating current was set to 4800 A, the commutation failure duration was 80 ms, and the ambient temperature was 30℃. Simulation results showed a significant increase in the overall heat source, and the overall winding temperature rose, reaching a maximum of 135℃. The fault was identified as "overall commutation anomaly," and the maintenance strategy was to reduce the load and arrange emergency repairs.

[0201] High ambient temperature conditions: The operating current was set at 4500 A, frequency at 50 Hz, and ambient temperature at 45℃. Simulation results showed an overall temperature rise, with the highest temperature approaching 120℃, nearing the insulation safety limit. The fault was identified as "high ambient temperature risk," and the recommended maintenance strategy was to enhance cooling or operate during off-peak hours.

[0202] In some embodiments of the present invention, parameters trained under normal operating conditions can be transferred to fault operating conditions for rapid heat source inversion, and network parameters can be updated online through a small number of iterations after real-time monitoring data input, thereby achieving rapid correction of heat source and temperature field.

[0203] In summary, this embodiment identifies the geometric model constructed from the reactor composite structure using a mask function, determines the material region (aluminum / epoxy), and thus determines the material parameters within different regions. Constraints are designed using both material parameters and sensor data to ensure strict adherence to physical laws during the solution process. A two-stage training process is performed on the PINN network employing temperature and heat source branching structures, while simultaneously introducing constraints to balance multiple losses, thereby improving the model's convergence stability. Furthermore, by combining a working state identification mechanism based on temperature distribution and a two-level fault identification mechanism based on heat source distribution results, accurate fault location and automatic generation of maintenance strategies are achieved.

[0204] This embodiment uses a dual-branch PINN network structure to achieve temperature field prediction and internal heat source inversion, thereby enabling overall temperature rise assessment and local fault diagnosis in the application scenario of smoothing reactors.

[0205] One embodiment of the present invention provides a fault identification system for smoothing reactors. For details, please refer to [link to documentation]. Figure 4 , Figure 4 The diagram shown is a schematic representation of a smoothing reactor fault identification system according to one embodiment of the present invention, comprising:

[0206] The parameter determination module M1 is used to construct a geometric model based on the structure of the target reactor device, and to identify material sub-regions on the geometric model using a predefined mask function to generate material parameters corresponding to each point within the material sub-regions.

[0207] The constraint establishment module M2 is used to collect the training data of the target reactor equipment in real time and construct a constraint loss function based on the material parameters and the training data.

[0208] Training module M3 is used to input the data to be trained into a physical information neural network that integrates temperature branch and heat source branch for phased training, and output the temperature distribution and heat source distribution of the target reactor device.

[0209] The correction module M4 is used to inject the corresponding mask function into the heat source branch for region restriction during training, and to introduce the constraint loss function to correct the parameters of the physical information neural network.

[0210] The fault identification module M5 is used to identify faults in the target reactor equipment based on the temperature distribution and the heat source distribution, so as to formulate and execute operation and maintenance strategies.

[0211] like Figure 5 As shown, this embodiment of the invention also provides a computer device. Figure 5 This is a structural block diagram of a preferred embodiment of a computer device provided by the present invention. The computer device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the method described above.

[0212] Preferably, the computer program can be divided into one or more modules / units (such as computer program 1, computer program 2, ...), and the one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the computer device.

[0213] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor. The processor is the control center of the terminal device, connecting various parts of the terminal device through various interfaces and lines.

[0214] The memory mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard drive, a SmartMedia Card (SMC), a Secure Digital (SD) card, and a Flash Card, or other volatile solid-state storage devices.

[0215] It should be noted that the aforementioned terminal devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 5 The structural block diagram is merely an example of a terminal device and does not constitute a limitation on the terminal device. It may include more or fewer components than shown, or combine certain components, or use different components. Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0216] Accordingly, embodiments of the present invention provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the steps in the method of the above embodiments, for example... Figure 1 Steps S1 to S5 as described above.

[0217] The technical features and effects of the smoothing reactor fault identification system proposed in this embodiment of the invention are the same as those of the smoothing reactor fault identification method proposed in this embodiment of the invention, and will not be repeated here.

[0218] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for fault identification of smoothing reactors, characterized in that, include: A geometric model is constructed based on the structure of the target reactor equipment. A predefined mask function is used to identify material sub-regions on the geometric model. Specifically, based on the winding structure and encapsulation structure information of the target reactor equipment, a geometric model reflecting a two-dimensional rectangular solution is constructed. The coordinates of any point on the geometric model are input into the mask function to obtain the corresponding material sub-region, thereby generating material parameters corresponding to each point within the material sub-region. The material sub-region reflects the aluminum material region or epoxy material region in the target reactor equipment. Real-time acquisition of training data for the target reactor equipment; and construction of a constrained loss function based on the material parameters and the training data. The data to be trained is input into a physical information neural network that integrates temperature and heat source branches for phased training, and the output is the temperature distribution and heat source distribution of the target reactor device. During training, the corresponding mask function is injected into the heat source branch to restrict the region, and the constraint loss function is introduced to correct the parameters of the physical information neural network. Based on the temperature distribution and heat source distribution, faults in the target reactor equipment are identified in order to formulate and implement operation and maintenance strategies.

2. The method for fault identification of smoothing reactors as described in claim 1, characterized in that, The real-time acquisition of training data from the target reactor equipment, and the construction of a constrained loss function based on the material parameters and the training data, include: Based on the material parameters, establish the transient heat conduction equations for each material sub-region, the interface constraint equations for each material sub-region contact surface, the boundary constraint equations for each material sub-region boundary surface, and the initial constraint equations to generate the corresponding physical constraint losses. Using multi-source sensors deployed on the target reactor equipment, temperature data, ambient environmental data, and electrical data are collected in real time to construct the training data; The training data is normalized to construct the corresponding data constraint loss. The constraint loss function is obtained by integrating the physical constraint loss and the data constraint loss.

3. The method for fault identification of smoothing reactors as described in claim 1, characterized in that, The step of inputting the data to be trained into a physical information neural network integrating temperature and heat source branches for phased training includes: The training data is input into the physical information neural network. During the first stage of training, the training parameters corresponding to the heat source branch are frozen, and the temperature branch is trained independently. During the second phase of training, the heat source branch is unlocked, and the temperature branch and the heat source branch are trained simultaneously.

4. The method for fault identification of smoothing reactors as described in claim 1, characterized in that, The process of introducing the constraint loss function to correct the parameters of the physical information neural network includes: At each iteration, the moving average of each loss term in the constraint loss function is calculated; The weights corresponding to each loss term are dynamically updated based on the moving average value to correct the parameters of the physical information neural network.

5. The method for fault identification of smoothing reactors as described in claim 1, characterized in that, The step of identifying faults in the target reactor equipment based on the temperature distribution and the heat source distribution, in order to formulate and execute an operation and maintenance strategy, includes: Feature extraction is performed on the temperature distribution to obtain the key temperature characteristics of the target reactor equipment; The key temperature characteristics are analyzed using a preset abnormal temperature judgment mechanism to determine the operating risk level of the target reactor equipment. Key heat source features of the target reactor equipment are extracted from the heat source distribution, and the key heat source features are matched with a pre-established fault mode library to determine the target fault mode corresponding to the target reactor equipment. Integrate the work risk level and the target failure mode to formulate and implement corresponding operation and maintenance strategies.

6. The method for fault identification of smoothing reactors as described in claim 5, characterized in that, The step of matching the characteristics of each key heat source with a pre-established fault mode library to determine the target fault mode corresponding to the target reactor equipment further includes: The similarity between the key heat source features and the samples in the fault mode library is calculated to obtain the initial similarity result; When the initial similarity result meets the preset conditions, the heat source distribution and the initial similarity result are input into a predetermined classifier for processing, and the target fault mode is output.

7. A fault identification system for smoothing reactors, characterized in that, include: The parameter determination module is used to construct a geometric model based on the structure of the target reactor equipment, and to identify material sub-regions on the geometric model using a predefined mask function. Specifically, based on the winding structure information and encapsulation structure information of the target reactor equipment, the geometric model reflecting the two-dimensional rectangular solution is constructed. The coordinates of any point on the geometric model are input into the mask function to obtain the corresponding material sub-region, thereby generating the material parameters corresponding to each point within the material sub-region. The material sub-region reflects the aluminum material region or epoxy material region in the target reactor equipment. The constraint establishment module is used to collect the training data of the target reactor equipment in real time and construct a constraint loss function based on the material parameters and the training data. The training module is used to input the data to be trained into a physical information neural network that integrates temperature and heat source branches for phased training, and output the temperature distribution and heat source distribution of the target reactor device. The correction module is used to inject the corresponding mask function into the heat source branch for region restriction during training, and to introduce the constraint loss function to correct the parameters of the physical information neural network. The fault identification module is used to identify faults in the target reactor equipment based on the temperature distribution and the heat source distribution, so as to formulate and execute operation and maintenance strategies.

8. A computer device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the smoothing reactor fault identification method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the smoothing reactor fault identification method as described in any one of claims 1 to 6.

Images