Method for assessing operational risk of power equipment

Abstract

The application relates to a power equipment operation risk assessment method. The method comprises the following steps: in response to a simulation request of a power equipment in a risk scenario, inputting a first control parameter corresponding to the power equipment in the risk scenario into a pre-trained neural physics sandbox model to obtain a multi-physical simulation rendering result; fusing the multi-physical simulation rendering result with an operation environment of the power equipment to obtain a risk preview simulation picture; and performing evaluation processing on the first control parameter based on the risk preview simulation picture to obtain an operation risk assessment result. The method can reduce cost and repeatability.

Description

Technical Field

[0001] This application relates to the field of power equipment technology, and in particular to a method for assessing operational risks of power equipment. Background Technology

[0002] High-risk operation of power equipment is directly related to power grid safety and personnel safety, and is characterized by high risk, serious consequences, low frequency but great harm.

[0003] Currently, the power industry relies primarily on traditional safety training and on-site physical drills for rehearsals and simulations of such operations. On-site drills are not only costly and inherently risky, but they also struggle to provide safe and repeatable practice for low-probability, high-consequence accident scenarios, thus limiting the effective verification of emergency plans and the continuous improvement of personnel's emergency response capabilities.

[0004] Therefore, existing technical solutions generally suffer from high exercise costs and low repeatability when conducting high-risk operation drills for power equipment. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for operational risk assessment of power equipment that can reduce costs and repetitiveness in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a method for assessing operational risks of power equipment, the method comprising:

[0007] In response to a simulation request for power equipment in a risk scenario, the first control parameters corresponding to the power equipment in the risk scenario are input into a pre-trained neurophysical sandbox model to obtain multi-physics simulation rendering results.

[0008] By integrating the multiphysics simulation rendering results with the operating environment of the power equipment, a risk pre-simulation screen is obtained;

[0009] The first control parameter is evaluated and processed based on the risk pre-simulation screen to obtain the operational risk assessment result.

[0010] Secondly, this application also provides an operational risk assessment device for power equipment, the device comprising:

[0011] The response module is used to respond to simulation requests for power equipment in risk scenarios by inputting the first control parameters corresponding to the power equipment in the risk scenario into a pre-trained neurophysical sandbox model to obtain multi-physics simulation rendering results.

[0012] The fusion module is used to fuse the multi-physics simulation rendering results with the operating environment of the power equipment to obtain a risk pre-simulation screen;

[0013] The evaluation module is used to evaluate the first control parameter based on the risk pre-simulation screen to obtain the operational risk assessment result.

[0014] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0015] In response to a simulation request for power equipment in a risk scenario, the first control parameters corresponding to the power equipment in the risk scenario are input into a pre-trained neurophysical sandbox model to obtain multi-physics simulation rendering results.

[0016] By integrating the multiphysics simulation rendering results with the operating environment of the power equipment, a risk pre-simulation screen is obtained;

[0017] The first control parameter is evaluated and processed based on the risk pre-simulation screen to obtain the operational risk assessment result.

[0018] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0019] In response to a simulation request for power equipment in a risk scenario, the first control parameters corresponding to the power equipment in the risk scenario are input into a pre-trained neurophysical sandbox model to obtain multi-physics simulation rendering results.

[0020] By integrating the multiphysics simulation rendering results with the operating environment of the power equipment, a risk pre-simulation screen is obtained;

[0021] The first control parameter is evaluated and processed based on the risk pre-simulation screen to obtain the operational risk assessment result.

[0022] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0023] In response to a simulation request for power equipment in a risk scenario, the first control parameters corresponding to the power equipment in the risk scenario are input into a pre-trained neurophysical sandbox model to obtain multi-physics simulation rendering results.

[0024] By integrating the multiphysics simulation rendering results with the operating environment of the power equipment, a risk pre-simulation screen is obtained;

[0025] The first control parameter is evaluated and processed based on the risk pre-simulation screen to obtain the operational risk assessment result.

[0026] The aforementioned method for operational risk assessment of power equipment firstly, in response to a simulation request for power equipment under a risk scenario, inputs the first control parameter corresponding to the power equipment under the risk scenario into a pre-trained neurophysical sandbox model to obtain a multiphysics simulation rendering result; secondly, the multiphysics simulation rendering result is fused with the operating environment of the power equipment to obtain a risk pre-simulation screen; finally, the first control parameter is evaluated based on the risk pre-simulation screen to obtain the operational risk assessment result. In this method, the multiphysics simulation rendering result is quickly generated in a virtual environment through a neurophysical sandbox model and fused with the real operating environment to obtain the risk pre-simulation screen, thereby achieving repeatable and low-cost virtual pre-simulation, avoiding the high cost and low repeatability problems of traditional on-site exercises. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is an internal structural diagram of a computer device in one embodiment;

[0029] Figure 2 This is a flowchart illustrating an operational risk assessment method for power equipment in one embodiment;

[0030] Figure 3 This is a flowchart illustrating an operational risk assessment method for power equipment in another embodiment;

[0031] Figure 4 This is a flowchart illustrating an operational risk assessment method for power equipment in another embodiment;

[0032] Figure 5 This is a flowchart illustrating an operational risk assessment method for power equipment in another embodiment;

[0033] Figure 6 This is a flowchart illustrating an operational risk assessment method for power equipment in another embodiment;

[0034] Figure 7 This is a flowchart illustrating an operational risk assessment method for power equipment in another embodiment;

[0035] Figure 8 This is a flowchart illustrating an operational risk assessment method for power equipment in another embodiment;

[0036] Figure 9 This is a flowchart illustrating an operational risk assessment method for power equipment in another embodiment;

[0037] Figure 10 This is a flowchart illustrating an operational risk assessment method for power equipment in another embodiment;

[0038] Figure 11 This is a structural block diagram of an operational risk assessment device for power equipment in one embodiment. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0040] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 1 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores relevant data for the operational risk assessment process of power equipment. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for operational risk assessment of power equipment.

[0041] Those skilled in the art will understand that Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0042] In one exemplary embodiment, such as Figure 2 As shown, a method for operational risk assessment of power equipment is provided, which can be applied to... Figure 1The following steps, 201 to 203, are used as an example of computer equipment.

[0043] Step 201: In response to the simulation request for power equipment in a risk scenario, the first control parameters corresponding to the power equipment in the risk scenario are input into the pre-trained neurophysical sandbox model to obtain the multiphysics simulation rendering results.

[0044] The neurophysics sandbox model is a hybrid simulation model based on deep learning, jointly trained from a neural radiation field model (for 3D scene rendering) and a Fourier neural operator model (for physics field prediction). This model can generate realistic simulation results containing multi-physics data in real time based on the input first control parameters.

[0045] The first control parameter refers to the key variables that can be adjusted or need to be monitored for power equipment under specific operating conditions, such as voltage, current, switch status, operating sequence, and load level. These parameters determine the operating state of the equipment and directly affect the distribution and evolution of physical fields (such as temperature field, stress field, and electromagnetic field).

[0046] Multiphysics simulation rendering results refer to the visualized rendered images or sequences output by the model that integrate multiple physical field data (such as temperature, stress, arc energy, electromagnetic intensity, etc.). This result not only includes visual appearance (color, texture, lighting) but also encodes physical quantity information at each location, enabling it to dynamically reflect changes in equipment status under risk scenarios.

[0047] In this embodiment, in response to a simulation request for power equipment in a risky scenario, the server inputs the first control parameters corresponding to the power equipment in the risky scenario into a pre-trained neurophysical sandbox model to obtain multiphysics simulation rendering results.

[0048] For example, the server receives a simulation request for a circuit breaker in a substation under a "short-circuit tripping" risk scenario. The server extracts the first control parameters corresponding to this scenario, including the peak short-circuit current (31.5kA), tripping time (45ms), and contact pressure (2.8kN), and inputs these parameters into a pre-trained neurophysical sandbox model. The Fourier neural operator module in the model quickly predicts the spatiotemporal distribution of multi-physics fields such as arc energy, contact temperature, and mechanism stress during the tripping process, while the neural radiation field module generates dynamic rendering images with corresponding hot spots, arc flashes, and stress concentration areas based on these physical field data. Finally, the server outputs a simulation video lasting 80ms with a resolution of 1920×1080 as the multi-physics simulation rendering result.

[0049] In another embodiment, during a transformer overload operation risk simulation, the server responds to the simulation request and obtains the current first set of control parameters: load rate (145%), ambient temperature (38°C), and cooling system status (partial fan failure). The server inputs these parameters into a neurophysical sandbox model. The model first predicts the evolution of physical fields such as winding hotspot temperature, oil flow distribution, and leakage magnetic flux density using Fourier neural operators. Then, the neural radiation field renders a 3D model of the transformer based on the prediction results. The winding area displays a color gradient from blue to red according to the temperature, the oil flow path is shown using a semi-transparent particle flow, and the leakage magnetic flux concentration area is represented by a dynamic superposition of magnetic field lines. The rendering results also display key data such as the highest temperature value and maximum stress value of each part in real time through interactive labels. The server finally outputs a multiphysics simulation rendering scene that can be rotated 360 degrees and supports zooming and cross-sectioning to view the internal state.

[0050] Step 202: The multiphysics simulation rendering results are fused with the operating environment of the power equipment to obtain a risk pre-simulation screen.

[0051] The operating environment refers to the actual physical space where the power equipment is located and its surrounding conditions, including the equipment itself, adjacent equipment, building structure, operating corridor, safety distance range, and other real three-dimensional scenes.

[0052] Fusion processing refers to the process of accurately matching and fusing computer-generated virtual simulation rendering results with digital models of the real operating environment through techniques such as spatial alignment, coordinate mapping, and virtual-real overlay.

[0053] Risk simulation visuals refer to augmented reality visuals generated by fusion, which overlay virtual risk simulation effects onto a real-world background. These visuals allow operators to intuitively observe the distribution and dynamic changes of potential risks (such as arc range, high-temperature areas, and stress hazard zones) within a real-world work environment.

[0054] In this embodiment, the server integrates the multiphysics simulation rendering results with the operating environment of the power equipment to obtain a risk pre-simulation screen.

[0055] For example, the server acquires a 3D point cloud map of a switchgear room generated by laser scanning. This map accurately records the spatial locations of physical objects such as switchgear, operating passages, and safety fences. Simultaneously, the server receives the multiphysics simulation rendering result (a video demonstrating the development of an electric arc inside the cabinet) generated in step 201. The server uses a feature point matching algorithm to align the virtual coordinate system of the rendering result with the world coordinate system of the point cloud map, establishing a precise spatial mapping relationship. Then, the server overlays the electric arc simulation video onto the corresponding switchgear area in the point cloud map at the correct position, scale, and orientation. The final generated risk pre-simulation screen shows: in the actual switchgear room, a dynamic electric arc flash effect is overlaid on the door panel of a specific cabinet, and a semi-transparent red area indicates the dangerous range of the electric arc energy, providing operators with an intuitive on-site risk perception.

[0056] Step 203: Evaluate and process the first control parameter based on the risk pre-simulation screen to obtain the operational risk assessment result.

[0057] The operational risk assessment result refers to the conclusive output obtained after assessment processing regarding the safety and risk level of the current primary control parameter under a specific risk scenario. This result can be a quantified risk score, a risk level label (such as high, medium, low), a specific risk description (such as contact temperature exceeding the limit), or improvement suggestions.

[0058] In this embodiment, the server evaluates the first control parameter based on the risk pre-simulation screen to obtain the operational risk assessment result.

[0059] For example, the server extracts key physical data generated from the simulation of the current first control parameter (such as short-circuit current and tripping speed) from the metadata accompanying the risk pre-simulation screen: the maximum arc energy is 28.6kJ, and the maximum temperature rise on the cabinet surface is 145K. The server calls the preset risk cost function: F_risk = max(energy / 30kJ, temperature rise / 150K). The calculated risk cost is max(28.6 / 30, 145 / 150) = 0.967. Based on the preset risk level threshold (high risk > 0.8), the server determines the operation under the current first control parameter to be "high risk". Finally, the server generates an operational risk assessment report that includes the risk level, risk cost value, and the main risk source (arc energy approaching the critical point).

[0060] In the above method, firstly, in response to a simulation request for power equipment in a risk scenario, the first control parameter corresponding to the power equipment in the risk scenario is input into a pre-trained neurophysical sandbox model to obtain a multiphysics simulation rendering result; secondly, the multiphysics simulation rendering result is fused with the operating environment of the power equipment to obtain a risk pre-simulation screen; finally, the first control parameter is evaluated based on the risk pre-simulation screen to obtain an operational risk assessment result. In this method, the multiphysics simulation rendering result is quickly generated in a virtual environment through a neurophysics sandbox model and fused with the real operating environment to obtain a risk pre-simulation screen, thereby achieving repeatable and low-cost virtual pre-simulation, avoiding the high cost and low repeatability problems of traditional on-site exercises.

[0061] In one exemplary embodiment, such as Figure 3 As shown, the aforementioned neurophysical sandbox model is obtained by combining a neural radiation field model and a Fourier neural operator model. Based on this, the training process of the aforementioned "neurophysical sandbox model" includes steps 301 to 304. Wherein:

[0062] Step 301: Acquire multi-view image data and multi-physical characteristic data of the power equipment in the operating environment.

[0063] Multi-view image data refers to a collection of high-resolution images of power equipment and its operating environment acquired from multiple different angles and orientations. These images cover various surfaces and key components of the equipment, providing ample visual information for 3D reconstruction.

[0064] Multi-physical characteristic data refers to the physical state parameters of power equipment under various operating conditions, including multi-dimensional physical field data such as temperature distribution, stress and strain, electromagnetic field strength, arc characteristics, and vibration spectrum.

[0065] In this embodiment of the application, the server acquires multi-view image data and multi-physical characteristic data of the power equipment in the operating environment.

[0066] For example, the server uses eight high-definition industrial cameras deployed at the substation site to simultaneously capture images of the target circuit breaker from different angles, obtaining 128 sets of multi-view images that include details of the equipment's appearance, connecting components, and operating mechanisms. Simultaneously, it acquires data on the circuit breaker's winding temperature, mechanical stress, electromagnetic field strength, and other physical characteristics under different operating conditions such as rated current and short-circuit breaking from the equipment monitoring system, forming a physical field dataset corresponding to the image data.

[0067] In another embodiment, with a more complex technical implementation, the server employs a mobile 3D scanning system to perform panoramic imaging of the large transformer. This is achieved through the fusion of LiDAR and a visible light camera, obtaining point cloud image data containing depth information. Simultaneously, the server connects to the device's digital twin platform to acquire multi-physics simulation data of the transformer under various fault modes, including overload, short circuit, and cooling failure. This data includes refined physical characteristic parameters such as oil flow temperature field distribution, core magnetic flux density, and winding hotspot temperature gradient. The server performs spatiotemporal alignment processing on the acquired image data and physical characteristic data to ensure that each viewpoint corresponds to specific physical state parameters, establishing an accurate data mapping relationship for subsequent training sample construction.

[0068] Step 302: Construct the first training sample set based on multi-view image data and multi-physical characteristic data.

[0069] The first training sample set is used to train the neural radiation field model. This first training sample set is a dedicated training data set for the neural radiation field model, containing the mapping relationship between visual features extracted from multi-view images and their corresponding physical property parameters.

[0070] In this embodiment of the application, the server constructs a first training sample set based on multi-view image data and multi-physical characteristic data.

[0071] For example, the server selects each pixel from multi-view image data, calculates the corresponding camera ray direction, and performs spatial sampling along the ray to obtain the coordinates of multiple sampling points. For each sampling point, the server extracts the RGB color value of that location in the image as a color label, and simultaneously determines whether the point is located inside the device based on the device's 3D model as a density label. The server combines the spatial coordinates of the sampling points, the viewing direction, and the corresponding multi-physical property data (such as the temperature and stress value at that location) into training samples, ultimately constructing a first training sample set containing millions of sample points.

[0072] In other embodiments, in advanced implementations, the server employs an adaptive sampling strategy to construct training samples. For critical areas of power equipment (such as circuit breaker arc-extinguishing chambers and transformer bushing connections), the server increases the sampling density to improve the reconstruction accuracy of these high-risk areas. The server introduces physical consistency constraints to ensure that changes in the physical characteristics of adjacent sampling points conform to physical laws such as heat conduction and stress propagation during sample construction. Furthermore, the server constructs dynamic samples with a time dimension, recording the evolution of the equipment's physical state at different operational stages, enabling the trained neural radiation field model to render the dynamic changes in the equipment's state. The first training sample set ultimately generated by the server not only contains spatial geometric information but also encodes the spatiotemporal evolution laws of the physical field.

[0073] Step 303: Based on the second control parameters and multi-physical characteristic data, construct the second training sample set.

[0074] The second training sample set is used to train the Fourier neural operator model. This second training sample set is specifically designed as the training data set for training the Fourier neural operator model, establishing a dynamic mapping relationship between the first control parameter input and the multiphysics output.

[0075] In this embodiment of the application, the server constructs a second training sample set based on the second control parameters and multi-physical characteristic data.

[0076] For example, the server collects the circuit breaker's operation records under different combinations of first control parameters (such as opening speed, contact pressure, and arc current), as well as the corresponding multi-physical characteristic measurement data (such as arc temperature field, mechanical stress field, and electromagnetic interference field). The server uses each set of first control parameters as input features and the measured multi-physical field data as output labels to construct input-output corresponding training sample pairs. Each sample is labeled with its corresponding risk scenario type (such as normal opening, short circuit breaking, and reignition fault), ultimately forming a second training sample set containing thousands of samples.

[0077] In another embodiment, the server generates training data using a combination of experimental design and numerical simulation. Through orthogonal experimental design, different value levels of the first control parameter are systematically combined to cover the entire parameter space of the device operation. The server calls high-fidelity multiphysics simulation software to perform high-precision simulations on each parameter combination, obtaining refined physical field data including transient arc plasma, thermo-mechanical coupling fields, and electromagnetic-fluid multi-field coupling. The server performs dimensionality reduction and feature extraction on the simulation data, converting the high-dimensional physical field data into a compact representation suitable for neural network training. Simultaneously, the server introduces data augmentation techniques, adding noise to existing samples and perturbing parameters to generate new training samples, improving the model's generalization ability. The final constructed second training sample set can comprehensively reflect the complex nonlinear relationship between changes in the first control parameter and the multiphysics response.

[0078] Step 304: Use the first training sample set and the second training sample set to jointly train the neural radiation field model and the Fourier neural operator model to obtain the neurophysical sandbox model.

[0079] In this embodiment, the server uses a first training sample set and a second training sample set to jointly train the neural radiation field model and the Fourier neural operator model to obtain a neurophysical sandbox model.

[0080] For example, the server initializes the network parameters of the neural radiation field model and the Fourier neural operator model. During training iterations, the server inputs the first training sample set into the neural radiation field model to calculate the difference loss between the rendered color and density and the real labels; simultaneously, it inputs the second training sample set into the Fourier neural operator model to calculate the difference loss between the predicted physical field and the real physical field. The server designs a joint loss function that weights and sums the losses of the two models, and updates the parameters of the two models synchronously through the backpropagation algorithm. After tens of thousands of iterations of training, training stops when the loss function converges to a preset threshold. The two trained models are then cascaded and integrated to obtain the neurophysical sandbox model.

[0081] In some embodiments, the server employs an alternating optimization strategy. In each training round, the parameters of the Fourier neural operator model are first fixed, and the neural radiation field model is optimized using a first training sample set, enabling it to learn scene rendering capabilities under given physical field conditions. Then, the parameters of the neural radiation field model are fixed, and the Fourier neural operator model is optimized using a second training sample set, enabling it to learn accurate prediction capabilities from first control parameters to the physical field. The server designs a cross-model consistency loss to ensure that the physical field predicted by the Fourier neural operator can be correctly rendered through the neural radiation field. The server also introduces a course learning strategy, starting training from simple working condition samples and gradually increasing the difficulty of complex fault scenario samples. In the later stages of training, the server adopts an adversarial training approach, using a discriminator network to evaluate the realism and physical rationality of the generated scenes, further improving the simulation quality of the model. The final neurophysical sandbox model not only has high-precision physical field prediction capabilities but can also generate physically consistent and visually realistic risk pre-simulation scenes, achieving a deep integration of physical simulation and visual rendering.

[0082] In one exemplary embodiment, such as Figure 4 As shown, the above-mentioned "joint training of the neural radiation field model and the Fourier neural operator model using the first training sample set and the second training sample set to obtain the neurophysical sandbox model" includes steps 401 to 404. Wherein:

[0083] Step 401: Input the first training sample set into the neural radiation field model to be trained to obtain the first output; and input the second training sample set into the Fourier neural operator model to be trained to obtain the second output.

[0084] The first output refers to the output result obtained by the neural radiation field model after receiving the first training sample set and performing forward propagation calculation on the six-dimensional conditional vector of each sample. It usually includes the predicted color value, density value and implicit visual feature representation.

[0085] The second output refers to the multi-physics distribution data, such as temperature field matrix, stress field tensor, and electromagnetic field vector, that the Fourier neural operator model obtains after receiving the second training sample set and processes the first control parameters through its network layer.

[0086] The model to be trained refers to the initial model state that has not yet completed parameter optimization. Its network weights and biases are in a state of random initialization or pre-training but have not been fine-tuned for the current task.

[0087] In this embodiment of the application, the server inputs a first training sample set into the neural radiation field model to be trained to obtain a first output; and inputs a second training sample set into the Fourier neural operator model to be trained to obtain a second output.

[0088] For example, the server inputs a first training sample set containing 1 million sampling points into the neural radiation field model to be trained in batches. The model processes the six-dimensional conditional vector of each sample and outputs the predicted RGB color value and volume density value for each sampling point. Simultaneously, the server inputs a second training sample set containing 5,000 sample pairs into the Fourier neural operator model to be trained. The model processes each set of first control parameters and outputs the corresponding multiphysics prediction distribution, such as temperature and stress field data at a resolution of 64×64×64.

[0089] In another embodiment, the server employs a hierarchical forward propagation strategy to process the first training sample set. For each batch of samples from each viewpoint, the neural radiation field model first maps spatial coordinates to a high-dimensional feature space through a position encoding layer, and then extracts geometric features through multiple fully connected layers. The model uses a dual-branch structure to output color and density separately, and performs feature fusion in an intermediate layer to ensure consistency between the rendering result and physical properties. Simultaneously, the server introduces an adaptive frequency selection mechanism in the Fourier neural operator model. Based on the features of the input first control parameter, the model dynamically adjusts the frequency band range processed by the Fourier transform layer, retaining more high-frequency components for drastically changing high-frequency physical fields (such as electric arc plasma), and performing appropriate low-frequency filtering for smoothly changing low-frequency physical fields (such as temperature gradients). The model employs a multi-scale fusion strategy in the output layer, upsampling and stitching feature maps of different depths to generate multi-physics prediction results with adjustable resolution. The server also introduces residual connections and attention mechanisms to enhance the model's ability to capture key physical phenomena (such as local overheating and stress concentration).

[0090] Step 402: Determine the joint loss value based on the first output and the labels corresponding to the first training sample set, and the second output and the labels corresponding to the second training sample set.

[0091] The labels corresponding to the first training sample set refer to the real supervision information associated with each sample point in the first training sample set, including the real color value of the spatial point from a specific perspective, the density label (distinguishing between inside and outside the device), and the real physical characteristic values ​​obtained from physical simulation or measurement.

[0092] The labels corresponding to the second training sample set refer to the real multiphysics distribution data associated with each sample pair in the second training sample set. These data usually come from high-precision numerical simulations or experimental measurements and serve as the standard answer for supervising the learning of the Fourier neural operator model.

[0093] The joint loss value refers to the weighted sum of losses from multiple aspects, including the rendering accuracy of the neural radiation field model, the physics prediction accuracy of the Fourier neural operator model, and the consistency of the outputs of the two models. It is used to guide the co-optimization of the two models. Specifically, the joint loss function utilizes the consistency deviation between the rendering effect of the neural radiation field model and the physics results of the Fourier neural operator model, resulting in:

[0094]

[0095] In the formula, For the joint loss function value, , , All are weighting coefficients. The loss function of the Fourier neural operator model is... Let be the loss function of the neural radiation field model. The loss function is used to determine the consistency deviation between the rendering results of the neural radiation field model and the physical field results of the Fourier neural operator model; where:

[0096]

[0097]

[0098]

[0099] In the formula, The total number of samples, The number of physical field channels. Let be the predicted physical field value of the c-th physical field channel at the i-th sampling point. This represents the actual physical field value of the c-th physical field channel at the i-th sampling point. , , , Let R, G, and B be the predicted colors and densities for the i-th sampling point, respectively. , , , Let R, G, B be the true colors of the R channel, G channel, and B channel, respectively, and let them be the true density of the i-th sampling point. The visual feature dimension is set to 4. Let c be the c-th dimension visual feature output by the neural radiation field model at the i-th sampling point. The c-th dimension visual feature label of the i-th sampling point is transformed by the physical-visual mapping rule from the predicted physical field value obtained by the Fourier neural operator model.

[0100] Among them, the physical-visual mapping rule is to pre-train a mapping network from physical field to visual features. This network takes the physical field data output by the Fourier neural operator model as input, and after multiple layers of convolution and fully connected operations, outputs a feature vector that is aligned with the rendering result of the neural radiation field model in the visual dimension, which is used as the visual feature label.

[0101] Specifically, the mapping network first normalizes the multi-physics distribution data (such as temperature and stress fields) output by the Fourier neural operator model, mapping their numerical range to [0,1]. Then, it extracts local features of the physical field through three convolutional layers (3×3 kernel size, stride 1, padding 1). Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to enhance the network's non-linear expressive power. Next, two fully connected layers (128 and 64 neurons respectively) integrate the local features into global features. Finally, it outputs a 4-dimensional visual feature vector (consistent with the visual feature dimension of the neural radiation field model), serving as the visual feature labels (RGB color and density labels) corresponding to the physical field results of the Fourier neural operator model. This mapping network is pre-trained through supervised learning (using the rendering results of the neural radiation field model as the target) to ensure the accurate and reliable mapping relationship between the physical field data and visual features.

[0102] In this embodiment, the server determines the joint loss value based on the first output and the labels corresponding to the first training sample set, and the second output and the labels corresponding to the second training sample set.

[0103] For example, the server calculates the loss of the neural radiation field model: comparing the color values ​​predicted by the model with the actual RGB labels of the samples, the mean squared error is calculated as the color loss; comparing the predicted density values ​​with the actual binary density labels, the cross-entropy loss is calculated as the geometric loss. Simultaneously, the server calculates the loss of the Fourier neural operator model: comparing the multiphysics distribution predicted by the model with the actual physics data labels, the mean squared error of each physics channel is calculated and summed as the physics prediction loss. The server weights and sums the three loss terms according to preset weights (0.4, 0.2, 0.4) to obtain the joint loss value L_total = 0.4*L_color + 0.2*L_density + 0.4*L_physics.

[0104] Step 403: Based on the joint loss value, the model parameters of the neural radiation field model to be trained and the model parameters of the Fourier neural operator model to be trained are updated synchronously through the backpropagation algorithm.

[0105] In this embodiment of the application, the server synchronously updates the model parameters of the neural radiation field model to be trained and the model parameters of the Fourier neural operator model to be trained using the backpropagation algorithm based on the joint loss value.

[0106] For example, after calculating the joint loss, the server uses automatic differentiation to calculate the gradients of the loss with respect to the parameters of the neural radiation field model and the Fourier neural operator model, respectively. For the neural radiation field model, the server calculates the gradient of the loss with respect to the weights and biases of its eight fully connected layers; for the Fourier neural operator model, the server calculates the gradient of the loss with respect to the parameters of its Fourier transform layer and fully connected layer. The server uses the Adam optimizer to simultaneously update all controllable parameters of both models based on the calculated gradients, with a learning rate set to 0.001 and a batch size of 64.

[0107] Step 404: If the preset convergence condition is met, the trained neural radiation field model and the trained Fourier neural operator model are cascaded to obtain the neurophysical sandbox model.

[0108] Among them, the preset convergence condition refers to the stopping criterion for judging whether the model has reached an acceptable performance level during the training process. The convergence condition includes the joint loss value being lower than the preset threshold (such as 0.01) and the loss value fluctuating by less than 5% within 10 consecutive training cycles, indicating that the model parameters have become stable; or the number of training iterations reaching the preset upper limit (such as 5000 times) to prevent overfitting.

[0109] A fully trained model refers to the final model state after sufficient optimization, where the model parameters have reached a stable state and the model performance on the validation set meets the preset performance requirements.

[0110] Cascaded processing refers to connecting two independently trained but complementary models in a specific order to form an end-to-end unified model architecture, enabling data to flow sequentially through the two models to complete the entire processing flow.

[0111] In this embodiment, upon reaching a preset convergence condition, the server cascades the trained neural radiation field model and the trained Fourier neural operator model to obtain a neurophysics sandbox model. The server monitors the change in the joint loss value during training. When the loss decreases by less than 0.1% for 10 consecutive training cycles, and the rendered PSNR value on the validation set reaches 32dB and the average relative error of physics prediction is less than 5%, the model is deemed to have reached convergence. The server stops training and saves the final weights of the neural radiation field model and the Fourier neural operator model. The server encapsulates the two models into a unified interface: the Fourier neural operator model acts as a front-end processor, receiving the first control parameters and outputting multiphysics predictions; the neural radiation field model acts as a back-end renderer, receiving physics data and viewpoint parameters and outputting rendered images. The two models communicate using a standardized data format to form a complete neurophysics sandbox model.

[0112] In one exemplary embodiment, such as Figure 5 As shown, the above-mentioned "constructing the first training sample set based on multi-view image data and multi-physical characteristic data" includes steps 501 to 504. Wherein:

[0113] Step 501: For each image pixel in the multi-view image data, emit a corresponding ray; perform layered sampling processing on the preset space along the emission direction of the ray to obtain multiple sampling points.

[0114] Among them, ray emission refers to a virtual straight line that starts from the optical center, passes through a specific pixel position in the image plane, and extends into three-dimensional space, used to establish the geometric mapping relationship between two-dimensional pixels and three-dimensional space.

[0115] Layered sampling refers to selecting multiple sampling locations along the ray direction within a set spatial range, according to a certain interval or strategy. Each sampling point represents a 3D spatial point to be queried on the ray path.

[0116] In this embodiment of the application, for each image pixel in the multi-view image data, the server emits a corresponding ray; the preset space is subjected to layered sampling processing along the emission direction of the ray to obtain multiple sampling points.

[0117] For example, after the server acquires multi-view image data of the circuit breaker, it calculates the corresponding camera ray direction for each pixel in an image with a resolution of 1920×1080. The server determines the effective spatial range of the scene as a 2m×2m×3m cube centered on the device, and uniformly selects 64 sampling points along each ray within this range. The sampling interval is adaptively adjusted according to the device size, ultimately resulting in approximately 132 million sampling points.

[0118] Step 502: For each sampling point, obtain the azimuth angle, elevation angle, three-dimensional spatial position, and multi-physical characteristic data of the ray corresponding to the sampling point.

[0119] Among them, azimuth and pitch are spherical coordinate parameters used to describe the direction of the ray. Azimuth represents the rotation angle on the horizontal plane, and pitch represents the tilt angle relative to the horizontal plane.

[0120] In this embodiment of the application, for each sampling point, the server acquires the azimuth angle, elevation angle, three-dimensional spatial position, and multiple physical property data of the ray corresponding to the sampling point.

[0121] For example, for each sampling point, the server calculates the ray direction vector corresponding to that point using the transformation matrix between the camera coordinate system and the world coordinate system, and then converts it into the azimuth angle θ and the elevation angle ϕ. The server queries the three-dimensional coordinates (x, y, z) of the sampling point and obtains the temperature value, stress value, and arc energy density at that location from the pre-processed multiphysics database through trilinear interpolation.

[0122] In another embodiment, the server constructs a hierarchical physical data query system. For fundamental physical fields, the server directly extracts data from a cached high-resolution simulation grid; for derived physical quantities, the server calculates differential components such as gradient, divergence, and curl in real time. The server introduces uncertainty modeling, assigning a confidence score to the physical property data of each sampling point to reflect the reliability of the data source (simulation or measurement) and interpolation errors. The server also implements cross-scale physical coupling: for macroscopic sampling points, it obtains continuous medium mechanical parameters; for mesoscopic sampling points (such as material interfaces), it obtains interface effect parameters; for microscopic sampling points (such as grain boundary positions), it obtains lattice strain parameters. The server establishes the spatiotemporal correlation of physical quantities, ensuring that the changes in physical properties of adjacent sampling points in time and space conform to physical laws. The server also developed a physical quantity normalization and standardization module, uniformly mapping physical parameters of different dimensions and ranges to the [-1,1] interval, facilitating the training and optimization of neural networks.

[0123] Step 503: For each sampling point, determine the color value of the sampling point in the corresponding image pixel as a color label, and determine the location information of the sampling point inside the power equipment as a density label.

[0124] Among them, the color label refers to the true color value of the sampling point under a specific viewpoint. It is usually obtained directly from the RGB value of the corresponding pixel in the original image and serves as the target for the supervised neural radiation field model to learn color rendering.

[0125] Depth labels refer to information describing the geometric relationship between sampling points and power equipment. They are usually represented as binary labels (inside / outside the equipment) or distance values ​​and are used to supervise the model to learn the geometric structure of the scene.

[0126] In this embodiment of the application, for each sampling point, the server determines the color value of the sampling point in the corresponding image pixel as a color label, and determines the location information of the sampling point inside the power equipment as a density label.

[0127] For example, the server uses perspective projection to find the corresponding pixel position in the original image based on the spatial coordinates of the sampling point, and extracts the RGB value of that pixel as a color label. The server then compares the sampling point coordinates with the device's 3D model to determine its position; if the point is inside the model, it is labeled with a density label of 1, otherwise it is labeled with 0.

[0128] Step 504: The azimuth angle, elevation angle, three-dimensional spatial position, multi-physical characteristic data, color label and density label of the sampling points are summarized and processed to obtain the first training sample set.

[0129] In this embodiment of the application, the server summarizes and processes the azimuth angle, elevation angle, three-dimensional spatial position, multi-physical characteristic data, color label and density label of the sampling points to obtain the first training sample set.

[0130] For example, the server will generate a six-dimensional condition vector for each sampling point. The data, including color labels (R, G, B), density labels (0 / 1), timestamps, view IDs, and other metadata, are packaged into a single data record. This record is organized and stored in HDF5 format according to view and ray order, containing approximately 200 million sample points, forming a complete first training sample set.

[0131] In one exemplary embodiment, such as Figure 6 As shown, the above-mentioned "constructing a second training sample set based on the first control parameters and multi-physical characteristic data" includes steps 601 to 602. Wherein:

[0132] Step 601: Combine the second control parameters and multi-physical characteristic data to obtain training samples.

[0133] In this embodiment of the application, the server combines the second control parameter and the multi-physical characteristic data to obtain training samples.

[0134] The server reads a set of initial control parameters for the circuit breaker from the database: opening speed 3.2 m / s, contact pressure 850 N, and arc current 15 kA, forming a 3D input vector. Simultaneously, it reads corresponding physical characteristic data from multiphysics simulation results: maximum temperature field value 1560 K, maximum stress field value 245 MPa, and arc energy density 8.7 kJ / cm³, forming a 3D output vector. The server pairs the input and output vectors and encapsulates them into a training sample.

[0135] Step 602: Label the training samples with the corresponding risk scenario types to obtain the second training sample set.

[0136] In this embodiment of the application, the server labels the training samples with the corresponding risk scenario types to obtain a second training sample set.

[0137] For example, the server labels each training sample with a corresponding risk scenario type based on its first combination of control parameters and physical response characteristics. For instance, a sample with a load rate of 120% and a temperature increase of 25K is labeled as "mild overload"; a sample with a short-circuit current of 31.5kA and an arc energy of 45kJ is labeled as "severe short circuit". After labeling, all labeled samples are divided into a training set, a validation set, and a test set in a 7:2:1 ratio to form a second training sample set.

[0138] In one exemplary embodiment, such as Figure 7 As shown, the above-mentioned "inputting the first control parameters corresponding to the power equipment in the risk scenario into the pre-trained neurophysical sandbox model to obtain the multiphysics simulation rendering result" includes steps 701 to 702. Wherein:

[0139] Step 701: Perform multi-physical property prediction processing on the first control parameter using the Fourier neural operator model to obtain the multi-physical prediction result, and extract the six-dimensional conditional vector of each pixel position in the multi-physical prediction result.

[0140] In this embodiment of the application, the server performs multi-physical property prediction processing on the first control parameter through a Fourier neural operator model to obtain multi-physical prediction results, and extracts the six-dimensional conditional vector of each pixel position in the multi-physical prediction results.

[0141] For example, the server inputs a set of circuit breaker operating parameters (opening speed, contact pressure, arc current) into a trained Fourier neural operator model. After forward propagation, the model outputs a 64×64×64×3 tensor, representing the predicted values ​​of three physical fields—temperature, stress, and arc energy—on a three-dimensional spatial grid. Subsequently, based on preset rendering viewport parameters, the server calculates the corresponding spatial sampling point for each pixel and extracts the three physical field values ​​for each sampling point from the predicted tensor using trilinear interpolation. These values, along with the spatial coordinates and viewpoint direction of the sampling point, form a six-dimensional conditional vector.

[0142] In another embodiment, since the multiphysics prediction results contain multiphysics property data under different spatial locations and viewpoints, each pixel of the rendered image is traversed according to a row-first / column-first rule, with the index denoted as (u,v). For each pixel (u,v), the following operations are performed:

[0143] 1) Calculate the camera ray direction: Convert the pixel coordinates into a ray direction vector in the camera coordinate system based on the camera intrinsic parameters, that is:

[0144]

[0145] In the formula, Let f be the ray direction vector, and f be the focal length. , () is the main point of the camera's intrinsic parameters.

[0146] 2) The ray direction vector is obtained by using the camera extrinsic parameter matrix T. Ray direction d converted to world coordinates world =(d x ,d y ,d z The ray originates at the camera's optical center. world =(o x ,o y ,o z );

[0147] 3) Along the ray direction d in the world coordinate system world Sampling N spatial points (e.g., N=64), the world coordinates of each sampling point are: (x,y,z)=o world +t×d world In the formula, t is the sampling depth.

[0148] The ray direction d in the world coordinate system world The view angle converted to spherical coordinates, i.e., the azimuth angle θ: θ = arctan2(d y ,d x ); Elevation angle ϕ: vertical angle, calculated using the formula ϕ=arccos(dz / ||d world ||)

[0149] Based on the (x, y, z) of the sampling point, corresponding data is extracted from the multi-physical property distribution prediction results output by the Fourier neural operator model. Specifically, the world coordinates (x, y, z) are mapped to the discrete grid index of the multi-physical property distribution prediction results. Since the tensor of the Fourier neural operator model is a discrete grid, trilinear interpolation is performed on the (x, y, z) of non-grid vertices to obtain the multi-physical property data (such as temperature, arc energy, stress, etc., which are integrated into a one-dimensional vector s(t)). The (x, y, z) of each pixel corresponding to the ray sampling point are then calculated. By integrating s(t), a six-dimensional conditional vector for the sampling point is obtained. After traversing all pixels and their corresponding ray sampling points, a full six-dimensional conditional vector is obtained. This vector integrates spatial coordinates and observation direction information, and may also contain time or state variables s(t) to reflect dynamic changes. The extracted six-dimensional conditional vector is then input one by one into a pre-trained neural radiation field model. Based on its internally learned continuous mapping relationship, the model performs high-fidelity rendering of the target power equipment scene, adjusting the output of the color branch so that the rendering result not only includes the geometric structure of the equipment, but also reflects changes in the physical field (such as the reddening phenomenon in high-temperature areas), generating multi-physics simulation rendering results of the target power equipment in a risk scenario.

[0150] Step 702: Input the six-dimensional conditional vector into the neural radiation field model to obtain the multiphysics simulation rendering results.

[0151] In this embodiment, the server inputs a six-dimensional conditional vector into the neural radiation field model to obtain multiphysics simulation rendering results.

[0152] The server inputs a large number of six-dimensional conditional vectors generated in step 701 into the neural radiation field model in batches. The model processes each conditional vector and outputs the predicted RGB color value and volume density value of that sampling point. Subsequently, the server performs volume rendering integration on all sampling points along the same ray to calculate the color of the pixel that the ray finally reaches. After traversing all pixels, a color image with a resolution of 1920×1080 is generated. In the image, the color of the device gradually changes from blue to red according to its temperature, the arc area appears as bright white light, and the stress concentration area is displayed as a semi-transparent contour line overlay, forming a multiphysics simulation rendering result.

[0153] In one exemplary embodiment, such as Figure 8 As shown, the above-mentioned "fusion processing of multi-physics simulation rendering results with the operating environment of power equipment to obtain a risk pre-simulation screen" includes steps 801 to 803. Among them:

[0154] Step 801: Scan the operating environment to obtain a 3D point cloud map of the operating environment.

[0155] In this embodiment of the application, the server scans the operating environment to obtain a three-dimensional point cloud map of the operating environment.

[0156] For example, the server loads a 3D point cloud map and simultaneously loads a virtual 3D model of a switch cabinet from a neurophysical sandbox model. The server manually or automatically selects at least three corresponding feature points (such as cabinet corners or nameplate centers) and determines their coordinates in both coordinate systems. Subsequently, the server uses the least squares method to solve for the optimal rigid body transformation parameters, calculating the rotation matrix R and translation vector T, thereby determining the coordinate system mapping relationship. This relationship is saved as a transformation matrix file.

[0157] Step 802: Align the virtual coordinate system of the 3D point cloud map and the neurophysical sandbox model to obtain the coordinate system mapping relationship.

[0158] In this embodiment, the server aligns the virtual coordinate systems of the 3D point cloud map and the neurophysical sandbox model to obtain a coordinate system mapping relationship.

[0159] The virtual coordinate system of the neurophysical sandbox model is based on the spatial reference frame set during model training. Its origin is usually located at the geometric center of the device or a preset reference point, and the coordinate axis directions are aligned with the main structural directions of the device (e.g., the X-axis corresponds to the length direction of the device, the Y-axis corresponds to the width direction, and the Z-axis corresponds to the height direction).

[0160] To achieve precise alignment between the real scene and the virtual coordinate system, key feature points (such as device edge inflection points and operation panel markers) in the 3D point cloud map are first extracted using feature point matching algorithms (such as SIFT or ORB), and the corresponding feature points are marked in the virtual scene of the neurophysical sandbox model. Then, the ICP (Iterative Closest Point) algorithm is used to calculate the optimal rigid body transformation matrix (containing rotation matrix R and translation vector T) between the two sets of feature points. This matrix describes the transformation relationship from the real scene coordinate system to the virtual coordinate system.

[0161] A coordinate system mapping relationship refers to the mathematical expression of the transformation relationship between two coordinate systems. It is usually represented by a 4x4 homogeneous transformation matrix, which includes a rotation matrix R and a translation vector T, such that any point P in the virtual coordinate system... v It can be done through Pr eal =R*P v +T converts to the real-world coordinate system.

[0162] For example, the server loads a 3D point cloud map of the switch cabinet room and a virtual 3D model of the switch cabinet from a neurophysical sandbox model. The server manually or automatically selects at least three corresponding feature points (such as cabinet corners or nameplate centers) and determines their coordinates in both coordinate systems. Subsequently, the server uses the least squares method to solve for the optimal rigid body transformation parameters, calculating the rotation matrix R and translation vector T, thereby determining the coordinate system mapping relationship. This relationship is saved as a transformation matrix file.

[0163] Step 803: Based on the coordinate system mapping relationship, the digital twin of the multiphysics simulation rendering result is superimposed onto the three-dimensional point cloud map to obtain the risk pre-simulation screen.

[0164] In this embodiment of the application, the server overlays the digital twin of the multiphysics simulation rendering result onto the three-dimensional point cloud map based on the coordinate system mapping relationship to obtain the risk pre-simulation screen.

[0165] Optionally, the multiphysics simulation rendering results are converted into the format required by the MR device, and the multiphysics simulation rendering results are input into the MR device. The MR device then overlays the digital twin of the multiphysics simulation rendering results onto the three-dimensional point cloud map of the real target power equipment based on the coordinate system mapping relationship, thereby generating a risk pre-simulation screen containing multiphysics characteristic data.

[0166] The MR (Multi-Physical Reduction) device precisely overlays the converted multi-physics simulation rendering results onto a 3D point cloud map of the real scene based on the coordinate system mapping relationship, achieving a fusion of virtual and reality. During this process, the MR device performs corresponding spatial transformations on the multi-physics simulation rendering results according to the rotation matrix R and translation vector T in the coordinate system mapping relationship, ensuring that its position, orientation, and scale in the real scene are consistent with those in the virtual coordinate system. Simultaneously, the MR device utilizes its built-in rendering engine to fuse the multi-physics simulation rendering results with environmental factors such as lighting and shadows in the real scene, generating a risk pre-simulation screen containing multi-physics characteristic data. This screen not only displays the geometric structure of the target power equipment but also reflects changes in the physical field (such as the distribution of the temperature field and areas of stress concentration) through visual elements such as color and texture, providing operators with intuitive and comprehensive risk pre-simulation information.

[0167] For example, the server acquires the multiphysics simulation rendering results of a short-circuit fault in the switchgear (a dynamic rendering video showing the development of an electric arc and the rise in temperature inside the cabinet) and the coordinate system mapping relationship calculated in step 802. Based on the mapping relationship, the server calculates the correct position and orientation of each frame in the simulation video in the coordinate system of the real point cloud map. Then, the server renders the point cloud map as the background and renders each frame of the simulation video as the foreground, overlaying them according to the calculated pose. The final generated image displays a dynamically superimposed flickering electric arc and a temperature field that gradually changes from blue to red within the actual point cloud outline of the switchgear, thus forming a risk pre-simulation image.

[0168] In one exemplary embodiment, such as Figure 9 As shown, the above-mentioned "evaluating and processing the first control parameter based on the risk pre-simulation screen to obtain the operational risk assessment result" includes steps 901 to 903. Wherein:

[0169] Step 901: Obtain the physical characteristic data corresponding to the first control parameters after simulation rendering from the risk pre-simulation screen.

[0170] In this embodiment of the application, the server obtains the physical characteristic data corresponding to the first control parameters after simulation rendering from the risk pre-simulation screen.

[0171] The server parses the metadata file accompanying the risk simulation screen. This file records the physical quantity values ​​of each key area (such as the circuit breaker arc-extinguishing chamber and transformer winding hot spots) during the rendering process. The server extracts the key physical characteristic data corresponding to this simulation from it: maximum temperature T. max =1560K, maximum stress σ max =245MPa, maximum arc energy E arc =48kJ.

[0172] Step 902: Based on the physical characteristic data, determine the risk cost of each physical characteristic data in conjunction with a preset risk cost function.

[0173] The risk cost function aims to maximize each physical characteristic data point. Its design must comprehensively consider the impact of different physical characteristics on equipment safety and operational stability. The risk cost function can be defined as maximizing the value of each physical characteristic data point, i.e.: J(s(t)) = max(Simulate(s(t)); where Simulate(s(t)) represents the multi-physics characteristic data determined by the multi-physics simulation rendering results.

[0174] In this embodiment, the server determines the risk cost of each physical characteristic data based on the data and a preset risk cost function; wherein the risk cost function aims to maximize the risk cost of each physical characteristic data.

[0175] The server retrieves three physical characteristic data points: temperature exceedance ratio R. T =T actual / T safe =1560 / 1200=1.3, stress exceedance ratio R σ =σ actual / σ_ safe =245 / 200=1.225, arc energy exceeding the standard ratio R E =E actual / E safe =48 / 40=1.2. The preset risk cost function is to take the maximum value of the three: J=max(R T ,R σ ,R E )=max(1.3,1.225,1.2)=1.3. Therefore, the operational risk cost under the current first control parameter is determined to be 1.3.

[0176] Step 903: Normalize the risk costs of each physical characteristic data, and perform weighted calculations on the normalized physical characteristic data to obtain the total operational risk cost of the power equipment under risk scenarios; determine the operational risk assessment result based on the total operational risk cost.

[0177] In this embodiment of the application, the server normalizes the risk cost of each physical characteristic data and performs weighted calculation on the normalized physical characteristic data to obtain the total operational risk cost of the power equipment in the risk scenario; based on the total operational risk cost, the operational risk assessment result is determined.

[0178] In an exemplary embodiment, after obtaining the operational risk assessment results, the above method further includes:

[0179] Sensitivity analysis is performed on the first control parameter based on the operational risk assessment results in order to identify the weak first control parameter.

[0180] Among them, the weak first control parameters are used to identify the first control parameters that need to be monitored or optimized in risk scenarios. These are the first control parameters that have the most significant impact on the operational risk assessment results (risk costs), as identified through sensitivity analysis. Small changes in these parameters may lead to large fluctuations in the risk level, thus making them a key aspect of risk management and operational optimization.

[0181] In this embodiment of the application, the server performs sensitivity analysis on the first control parameter based on the operational risk assessment results in order to identify the weak first control parameter from the first control parameter.

[0182] For example, the server obtains the operational risk assessment results under the current first set of control parameters (load rate 145%, ambient temperature 38°C, cooling status 'poor'), showing a total risk cost of 1.18 (high risk). To perform sensitivity analysis, the server sequentially applies a small perturbation (such as ±5% or ± a fixed value) to each first control parameter, re-invokes the neurophysical sandbox model for simulation and evaluation, and calculates the change in risk cost (gradient). Hypothesis analysis reveals that the risk cost has the largest gradient with respect to the "load rate" perturbation (0.25 / %), a moderate gradient with respect to "ambient temperature" (0.1 / °C), and the largest risk jump caused by the switching of "cooling status" (a binary variable). Based on this, the server identifies "load rate" and "cooling status" as the weak first control parameters under the current risk scenario.

[0183] In one exemplary embodiment, such as Figure 10 As shown, the above-mentioned "performing sensitivity analysis on the first control parameter based on the operational risk assessment results to identify the weak first control parameter from the first control parameters" includes steps 1001 to 1003. Wherein:

[0184] Step 1001: Randomly generate parameter disturbance quantities corresponding to each first control parameter, and superimpose each first control parameter with its corresponding parameter disturbance quantity to obtain multiple superimposed parameters.

[0185] The parameter disturbance refers to a small change added to the original value of the first control parameter. It is used to simulate potential fluctuations or uncertainties in the parameter during actual operation, or to explore local changes in the parameter space using numerical methods. The update process for the parameter disturbance is as follows:

[0186]

[0187] In the formula, , These represent the parameter perturbations before and after the update. The step size is typically 0.01 to 0.1.

[0188] Superimposed parameters refer to the new parameter values ​​obtained by algebraically adding the original first control parameter with its corresponding disturbance, which are used to construct a set of neighboring parameter combinations around the original operating point.

[0189] In this embodiment of the application, the server randomly generates parameter disturbance quantities corresponding to each first control parameter, and superimposes each first control parameter with its corresponding parameter disturbance quantity to obtain multiple superimposed parameters.

[0190] For example, the server takes the following circuit breaker operating parameters as an example: opening speed v0 = 3.5 m / s, contact pressure F0 = 800 N, and arc current I0 = 12 kA. The server independently samples a disturbance for each parameter from a normal distribution with a mean of 0 and a standard deviation of 5% of the parameter value. For example, this might yield Δv = +0.15 m / s, ΔF = -30 N, and ΔI = +0.5 kA. The server then superimposes the disturbance onto the original parameters, resulting in a new set of superimposed parameters: v1 = 3.65 m / s, F1 = 770 N, and I1 = 12.5 kA. Repeating this process multiple times yields multiple sets of superimposed parameters.

[0191] Step 1002: Input multiple superimposed parameters into the neurophysical sandbox model to obtain the multi-physical property distribution results corresponding to each superimposed parameter.

[0192] In this embodiment, the server inputs multiple superimposed parameters into the neurophysical sandbox model to obtain the multi-physical property distribution results corresponding to each superimposed parameter.

[0193] For example, the server submits a batch of superimposed parameters (v1=3.65m / s, F1=770N, I1=12.5kA) generated in step 1001 to the neurophysical sandbox model. The Fourier neural operator module in the model processes this input, predicting the distribution of the temperature field, stress field, and arc energy field on a three-dimensional spatial grid during circuit breaker operation under these parameters. The server records this distribution data as the multi-physics property distribution result corresponding to this set of superimposed parameters. This process is repeated for each set of superimposed parameters.

[0194] Step 1003: Based on the distribution results of each multi-physical characteristic, determine the risk cost of each superimposed parameter in conjunction with the risk cost function.

[0195] Here, risk cost refers to the scalar output value calculated by inputting the multi-physical property distribution results into the risk cost function. The larger this value, the higher the risk of operation under the corresponding superposition parameters.

[0196] In this embodiment, the server determines the risk cost of each superimposed parameter based on the distribution results of each multi-physical characteristic and the risk cost function.

[0197] For example, the server obtains the distribution results of multiple physical properties corresponding to a set of superimposed parameters, and extracts the extreme values ​​of key physical quantities from them: the highest temperature T. max =1420K, maximum stress σmax =210MPa. The server calls the preset risk cost function J=max(T) max / T crit ,σ max / σ crit ), where T crit =1300K, σ crit =200MPa. Substituting into the calculation, we get J=max(1420 / 1300,210 / 200)=max(1.092,1.05)=1.092. This value of 1.092 is the risk cost of this set of superimposed parameters.

[0198] Step 1004: Perform partial differential calculations on the corresponding superposition parameters according to each risk cost to obtain the risk cost gradient corresponding to each first control parameter.

[0199] Here, the risk cost gradient is a vector, where each component represents the partial derivative of the risk cost with respect to a specific first control parameter. It indicates which parameter's direction, near the original parameter point, causes the risk cost to rise or fall most rapidly. Since the risk cost is partially differentiable with respect to each superimposed first control parameter, calculating the risk cost gradient for each superimposed first control parameter using gradient descent clarifies the direction and extent of each parameter's influence on the overall risk cost.

[0200]

[0201] In the formula, As a risk cost gradient, For risk costs, To superimpose the first control parameter.

[0202] In this embodiment, the server performs partial differential calculations on the corresponding superposition parameters based on each risk cost to obtain the risk cost gradient corresponding to each first control parameter.

[0203] For example, the server has obtained the risk cost J0 of the original parameter point (baseline point), and the risk costs J1, J2, ... of multiple sets of parameter points (superimposed parameters) after slight perturbations near the baseline point. For a certain first control parameter (such as the tripping speed v), the server selects samples of superimposed parameters that have only perturbed v while keeping other parameters unchanged, and calculates the ratio (ΔJ / Δv) of the change in risk cost ΔJ to the change in parameter Δv. By averaging multiple samples or fitting using the least squares method, the server obtains the approximate partial derivative of the risk cost J with respect to the tripping speed v, i.e., the component of the gradient in that direction. This process is repeated for all first control parameters to obtain the complete risk cost gradient vector.

[0204] Step 1005: Based on the risk cost gradient, determine the parameters in the first control parameters whose risk cost gradient is greater than the preset gradient threshold as weak first control parameters.

[0205] In this embodiment, the server determines the parameters in the first control parameters whose risk cost gradient is greater than a preset gradient threshold as weak first control parameters based on the risk cost gradient.

[0206] For example, the server calculates the gradient components of the risk cost relative to the three first control parameters, namely: the gradient G with respect to the tripping speed. v =0.25 (unit risk cost change / unit velocity change), gradient G with respect to contact pressure F =-0.05, the gradient G with respect to the arc current I =0.30. The preset gradient threshold is 0.10. Comparison shows that G... v and G I The absolute value of G is greater than 0.10, while G F The absolute value is less than 0.10. Therefore, the server identifies "shut-off speed" and "arc current" as the weakest first control parameters in the current risk scenario.

[0207] In some embodiments, in order to reverse-engineer the optimal first control parameters and meet the requirements of high-risk operation of power equipment for accurate and real-time simulation, the method further includes:

[0208] Given that the operational risk assessment indicates the existence of operational risk, a first control parameter optimization model is constructed with multiple optimization objectives, namely minimizing operational risk, maximizing operational efficiency, and minimizing operational cost, and the constraints of these multiple optimization objectives are determined.

[0209] Among these, the multiple optimization objectives—minimizing operational risk, maximizing operational efficiency, and minimizing operational cost—require a comprehensive consideration of various risk factors, efficiency indicators, and cost components during the operation of power equipment. When determining constraints, the physical characteristics of the equipment, operating procedures, and safety standards must be taken into account. Specifically, the objective function corresponding to these multiple optimization objectives is:

[0210]

[0211] In the formula, Let x be the objective function value, and let x be the first control parameter. To account for operational risk costs, For operational efficiency, For operating costs, , , All are weighted values; weighted values , , It can be set based on experience and weight values. , , The sum of these is 1, where:

[0212]

[0213] In the formula, , , These are all weighting coefficients; similarly, weighting coefficients... , , It can be set based on experience and weight values. , , The sum of them is 1. This represents the highest temperature output by the neurophysical sandbox model at time t. This refers to the critical temperature value (e.g., 80℃). This represents the highest arc energy output by the neurophysical sandbox model at time t. This is the critical value of electric arc energy (e.g., 15 kJ). The maximum stress output by the neurophysical sandbox model at time t is the simulation result. This is the critical stress value (e.g., 300 MPa). The total operation time is denoted as ; where all items in the operation risk cost are dimensionless parameters.

[0214]

[0215] in, The maximum allowable operation time is defined as follows: the total operation time is determined based on the ratio of the operation path length to the operation speed, while the maximum allowable operation time is set based on the actual application scenario.

[0216]

[0217] In the formula, Cost per unit of time operation The maximum budget cost is set based on the actual application scenario; the unit time operation cost is a fixed value determined comprehensively based on the equipment type, operation complexity and energy consumption level. For example, the unit time cost of transformer maintenance includes sub-item accounting values ​​such as labor costs, equipment depreciation and energy consumption.

[0218] The constraints include temperature boundary constraints, arc energy boundary constraints, stress boundary constraints, first control parameter boundary constraints, equipment load boundary constraints, and emergency operation time boundary constraints.

[0219] The temperature boundary constraints are as follows:

[0220] ≤

[0221] The boundary constraints for electric arc energy are:

[0222] ≤

[0223] The stress boundary constraints are:

[0224] ≤

[0225] The first control parameter boundary constraint is:

[0226]

[0227] In the formula, For the i-th first control parameter, , These are the minimum and maximum values ​​of the i-th first control parameter, respectively.

[0228] The equipment load boundary constraints are:

[0229]

[0230] In the formula, For equipment load, This represents the maximum device load.

[0231] Emergency operation time boundary constraints are:

[0232] ≥

[0233] In the formula, For a safe response time, such as 10 seconds.

[0234] Based on the Monte Carlo sampling method, multiple sets of candidate first control parameters of the target power equipment under risk scenarios are randomly selected, and the selected multiple sets of candidate first control parameters are input into the neurophysical sandbox model to obtain the multi-physical characteristic distribution results corresponding to the multiple sets of candidate first control parameters.

[0235] Monte Carlo sampling is a probability-based random sampling technique that simulates the random behavior of complex systems by generating a large number of random samples that conform to a specific distribution.

[0236] Multiple sets of candidate first control parameters are extracted for the target power equipment under risk scenarios, including extreme weather conditions (such as high temperature, low temperature, heavy rain, and strong wind), equipment aging failure, and sudden short circuits. For each risk scenario, first control parameters are set based on historical equipment operating data and expert experience. Multiple sets of candidate first control parameters are generated through improved strategies such as Latin hypercube sampling and uniform distribution sampling to ensure that the samples cover the key areas of the parameter space.

[0237] The candidate first control parameter is input into the neurophysical sandbox model, and the multi-physics property distribution result corresponding to the candidate first control parameter can be output through the Fourier neural operator model, or the multi-physics property distribution result can be obtained through the multi-physics simulation rendering result output by the neurophysical sandbox model.

[0238] Based on multiple sets of candidate first control parameters and the distribution results of multiple physical characteristics corresponding to each candidate first control parameter, the optimization model of the first control parameter is optimized and solved. Based on the optimal solution, the optimal first control parameter of the target power equipment under the risk scenario is determined from multiple sets of candidate first control parameters.

[0239] In this study, a non-dominated sorting genetic algorithm (NSGA-II) was used to perform multi-objective optimization on the first control parameter optimization model. This algorithm searches for the optimal set of solutions that satisfy multiple optimization objectives in the parameter space by simulating the selection, crossover, and mutation mechanisms in biological evolution.

[0240] In practice, multiple candidate first control parameters are first encoded into chromosomes, with each gene locus corresponding to a first control parameter. The population is divided into multiple frontiers through fast non-dominated sorting, and non-dominated solutions are preferentially retained to maintain the diversity of the solution set. The crossover operation adopts a simulated binary crossover method to generate offspring parameters near the parent parameters. The mutation operation adjusts the parameter values ​​through polynomial mutation to avoid local optimum traps.

[0241] During the iteration process, the fitness value of each individual is calculated based on the multiphysics simulation results of multiple sets of candidate first control parameters and the objective function, and the elite solution library is dynamically maintained to finally obtain the Pareto front solution set.

[0242] The solution with the best overall performance is selected from the solution set as the optimal first control parameter for the target power equipment under risk scenarios. This parameter, while satisfying physical constraints such as temperature, stress, and arc energy, as well as equipment constraints, simultaneously achieves the optimization effect of reducing operational risk costs by 15%-20%, improving operational efficiency by 10%-15%, and reducing operational costs by 8%-12%.

[0243] For example, in transformer maintenance scenarios, the optimized first control parameter can keep the maximum equipment temperature below 75℃, the peak arc energy not exceeding 12kJ, and the single operation time reduced to 85% of the standard working time, significantly improving operational safety and economy.

[0244] In one exemplary embodiment, the method further includes:

[0245] Step 1: Acquire multi-view image data and multi-physical characteristic data of power equipment under operating conditions.

[0246] Step 2: For each image pixel in the multi-view image data, emit a corresponding ray.

[0247] Step 3: Perform layered sampling processing on the preset space along the emission direction of the ray to obtain multiple sampling points.

[0248] Step 4: For each sampling point, obtain the azimuth angle, elevation angle, three-dimensional spatial position, and multi-physical property data of the ray corresponding to the sampling point.

[0249] Step 5: For each sampling point, determine the color value of the sampling point in the corresponding image pixel as the color label, and determine the location information of the sampling point inside the power equipment as the density label.

[0250] Step 6: Summarize and process the azimuth, elevation, three-dimensional spatial position, multi-physical characteristic data, color labels, and density labels of the sampling points to obtain the first training sample set.

[0251] Step 7: Combine the second control parameters and multi-physical characteristic data to obtain training samples.

[0252] Step 8: Label the training samples with the corresponding risk scenario types to obtain the second training sample set.

[0253] Step 9: Input the first training sample set into the neural radiation field model to be trained to obtain the first output; and input the second training sample set into the Fourier neural operator model to be trained to obtain the second output.

[0254] Step 10: Determine the joint loss value based on the first output and the labels corresponding to the first training sample set, and the second output and the labels corresponding to the second training sample set.

[0255] Step 11: Based on the joint loss value, synchronously update the model parameters of the neural radiation field model to be trained and the model parameters of the Fourier neural operator model to be trained using the backpropagation algorithm.

[0256] Step 12: Under the condition of reaching the preset convergence, the trained neural radiation field model and the trained Fourier neural operator model are cascaded to obtain the neurophysical sandbox model.

[0257] Step 13: In response to the simulation request for power equipment under risk scenarios, the first control parameter is processed by multi-physical characteristic prediction using a Fourier neural operator model to obtain multi-physical prediction results, and the six-dimensional conditional vector of each pixel position in the multi-physical prediction results is extracted.

[0258] Step 14: Input the six-dimensional conditional vector into the neural radiation field model to obtain the multiphysics simulation rendering results.

[0259] Step 15: Scan the operating environment to obtain a 3D point cloud map of the operating environment.

[0260] Step 16: Align the virtual coordinate system of the 3D point cloud map and the neurophysical sandbox model to obtain the coordinate system mapping relationship.

[0261] Step 17: Based on the coordinate system mapping relationship, the digital twin of the multiphysics simulation rendering result is superimposed onto the three-dimensional point cloud map to obtain the risk pre-simulation screen.

[0262] Step 18: Obtain the physical characteristic data corresponding to the first control parameters after simulation rendering from the risk pre-simulation screen.

[0263] Step 19: Based on the physical characteristic data, determine the risk cost of each physical characteristic data in conjunction with a preset risk cost function; wherein, the risk cost function aims to maximize the risk cost of each physical characteristic data.

[0264] Step 20: Normalize the risk costs of each physical characteristic data, and perform weighted calculations on the normalized physical characteristic data to obtain the total operational risk cost of the power equipment under risk scenarios.

[0265] Step 21: Determine the operational risk assessment results based on the total cost of operational risk.

[0266] Step 22: Randomly generate parameter disturbance quantities corresponding to each first control parameter, and superimpose each first control parameter with its corresponding parameter disturbance quantity to obtain multiple superimposed parameters.

[0267] Step 23: Input multiple superimposed parameters into the neurophysical sandbox model to obtain the multi-physical property distribution results corresponding to each superimposed parameter.

[0268] Step 24: Based on the distribution results of each physical property, determine the risk cost of each superimposed parameter in conjunction with the risk cost function.

[0269] Step 25: Perform partial differential calculations on the corresponding superposition parameters according to each risk cost to obtain the risk cost gradient corresponding to each first control parameter.

[0270] Step 26: Based on the risk cost gradient, determine the parameters in the first control parameters whose risk cost gradient is greater than the preset gradient threshold as weak first control parameters; wherein, weak first control parameters are used to identify the first control parameters that need to be monitored or optimized in risk scenarios.

[0271] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0272] Based on the same inventive concept, this application also provides an operational risk assessment device for power equipment to implement the above-described operational risk assessment method for power equipment. The solution provided by this device is similar to the implementation described in the above-described method. Therefore, the specific limitations of one or more operational risk assessment device embodiments for power equipment provided below can be found in the limitations of the operational risk assessment method for power equipment described above, and will not be repeated here.

[0273] In one exemplary embodiment, such as Figure 11 As shown, an operational risk assessment device for power equipment is provided, comprising: a response module 1101, a fusion module 1102, and an assessment module 1103, wherein:

[0274] The response module 1101 is used to respond to the simulation request for power equipment in a risk scenario, input the first control parameter corresponding to the power equipment in the risk scenario into the pre-trained neurophysical sandbox model, and obtain the multi-physics simulation rendering result.

[0275] The fusion module 1102 is used to fuse the multi-physics simulation rendering results with the operating environment of the power equipment to obtain a risk pre-simulation screen;

[0276] The evaluation module 1103 is used to evaluate the first control parameter based on the risk pre-simulation screen to obtain the operational risk assessment result.

[0277] It should be noted that each module in the above-mentioned operational risk assessment device for power equipment can execute the above-described method embodiments, and their implementation principles and technical effects are similar, so they will not be repeated here.

[0278] Each module in the aforementioned operational risk assessment device for power equipment can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0279] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0280] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0281] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0282] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0283] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0284] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for assessing operational risks of power equipment, characterized in that, include: In response to a simulation request for power equipment in a risk scenario, the first control parameters of the power equipment in the risk scenario are input into a pre-trained neurophysical sandbox model to obtain multiphysics simulation rendering results. The multiphysics simulation rendering results are fused with the operating environment of the power equipment to obtain a risk pre-simulation screen; The first control parameter is evaluated and processed based on the risk pre-simulation screen to obtain the operational risk assessment result.

2. The method according to claim 1, characterized in that, The neurophysical sandbox model is obtained by combining a neural radiation field model and a Fourier neural operator model. The training process of the neurophysical sandbox model includes: Acquire multi-view image data and multi-physical characteristic data of the power equipment under the operating environment; Based on the multi-view image data and the multi-physical property data, a first training sample set is constructed; wherein, the first training sample set is used to train the neural radiation field model; Based on the second control parameters and the multi-physical characteristic data, a second training sample set is constructed; wherein, the second training sample set is used to train the Fourier neural operator model; The neural radiation field model and the Fourier neural operator model are jointly trained using the first training sample set and the second training sample set to obtain the neurophysical sandbox model.

3. The method according to claim 2, characterized in that, The step of jointly training the neural radiation field model and the Fourier neural operator model using the first training sample set and the second training sample set to obtain the neurophysical sandbox model includes: The first training sample set is input into the neural radiation field model to be trained to obtain the first output; and the second training sample set is input into the Fourier neural operator model to be trained to obtain the second output. Based on the first output and the labels corresponding to the first training sample set, the second output and the labels corresponding to the second training sample set, the joint loss value is determined; Based on the joint loss value, the model parameters of the neural radiation field model to be trained and the model parameters of the Fourier neural operator model to be trained are synchronously updated through the backpropagation algorithm. Once the preset convergence condition is met, the trained neural radiation field model and the trained Fourier neural operator model are cascaded to obtain the neurophysical sandbox model.

4. The method according to claim 2, characterized in that, The construction of the first training sample set based on the multi-view image data and the multi-physical characteristic data includes: For each image pixel in the multi-view image data, a corresponding ray is emitted; The preset space is sampled in layers along the emission direction of the ray to obtain multiple sampling points; For each sampling point, the azimuth angle, elevation angle, three-dimensional spatial position, and multiple physical property data of the ray corresponding to the sampling point are obtained; For each sampling point, the color value of the sampling point in the corresponding image pixel is determined as a color label, and the position information of the sampling point inside the power equipment is determined as a density label; The azimuth angle, elevation angle, three-dimensional spatial position, multi-physical property data, color label, and density label of the sampling points are summarized and processed to obtain the first training sample set.

5. The method according to claim 2, characterized in that, The construction of a second training sample set based on the first control parameters and the multi-physical characteristic data includes: The second control parameter and the multi-physical characteristic data are combined and processed to obtain training samples; The training samples are labeled with the corresponding risk scenario types to obtain the second training sample set.

6. The method according to claim 2, characterized in that, The step of inputting the first control parameter corresponding to the power equipment under the risk scenario into a pre-trained neurophysical sandbox model to obtain multiphysics simulation rendering results includes: The first control parameter is subjected to multi-physical property prediction processing through the Fourier neural operator model to obtain multi-physical prediction results, and the six-dimensional conditional vector of each pixel position in the multi-physical prediction results is extracted. The six-dimensional conditional vector is input into the neural radiation field model to obtain the multiphysics simulation rendering result.

7. The method according to claim 1, characterized in that, The process of fusing the multiphysics simulation rendering results with the operating environment of the power equipment to obtain a risk pre-simulation screen includes: The operating environment is scanned to obtain a three-dimensional point cloud map of the operating environment; Align the virtual coordinate system of the 3D point cloud map with that of the neurophysical sandbox model to obtain the coordinate system mapping relationship; Based on the coordinate system mapping relationship, the digital twin of the multiphysics simulation rendering result is superimposed onto the three-dimensional point cloud map to obtain the risk pre-simulation screen.

8. The method according to claim 1, characterized in that, The step of evaluating the first control parameter based on the risk pre-simulation screen to obtain the operational risk assessment result includes: From the risk pre-simulation screen, obtain the physical characteristic data corresponding to the first control parameters after simulation rendering; Based on the aforementioned physical characteristic data, the risk cost of each physical characteristic data is determined by combining it with a preset risk cost function; wherein, the risk cost function aims to maximize the risk cost of each physical characteristic data. The risk costs of each physical characteristic data are normalized, and the normalized physical characteristic data are weighted and calculated to obtain the total operational risk cost of the power equipment under the risk scenario. The operational risk assessment result is determined based on the total cost of the operational risk.

9. The method according to claim 1, characterized in that, After obtaining the operational risk assessment results, the method further includes: Sensitivity analysis is performed on the first control parameter based on the operational risk assessment results to identify the weak first control parameter from the first control parameter; wherein, the weak first control parameter is used to identify the first control parameter that needs to be monitored or optimized in the risk scenario.

10. The method according to claim 9, characterized in that, The process of performing sensitivity analysis on the first control parameter based on the operational risk assessment results to identify the weak first control parameter includes: Randomly generate parameter disturbance values ​​corresponding to each of the first control parameters, and superimpose each of the first control parameters with the corresponding parameter disturbance values ​​to obtain multiple superimposed parameters; The multiple superimposed parameters are respectively input into the neurophysical sandbox model to obtain the multi-physical characteristic distribution results corresponding to each superimposed parameter; Based on the distribution results of each of the multiple physical characteristics, the risk cost of each of the superimposed parameters is determined in conjunction with the risk cost function. Based on each of the aforementioned risk costs, partial differential calculations are performed on the corresponding superposition parameters to obtain the risk cost gradient corresponding to each of the first control parameters. Based on the risk cost gradient, the parameters in the first control parameters whose risk cost gradient is greater than a preset gradient threshold are determined as the weak first control parameters.

Images