Electromagnetic absorption power density evaluation method and system, electronic device

Abstract

The application relates to the technical field of electromagnetic absorption power density evaluation, and discloses an electromagnetic absorption power density evaluation method, which comprises the following steps: obtaining incident electromagnetic wave parameters of a to-be-evaluated scene, a geometric model of a target object and tissue electromagnetic parameters; inputting the above parameters into an evaluation model to obtain a prediction result of electromagnetic absorption power density distribution and a spatial uncertainty distribution thereof; sampling a real scene to obtain measured data, and a sampling strategy is determined by the spatial uncertainty distribution; optimizing model parameters according to comparison between the measured data and the prediction result; and finally inputting the parameters into the optimized model to re-evaluate. Through the construction of a digital twin and a physical verification closed loop, the application realizes fast prediction and uncertainty quantification, intelligently guides measurement resources by using the uncertainty distribution, calibrates the model by using measured physical data, and further improves evaluation efficiency, precision and reliability. The application also discloses an electromagnetic absorption power density evaluation system and an electronic equipment.

Description

Technical Field

[0001] This application relates to the field of electromagnetic absorption power density assessment technology, specifically to an electromagnetic absorption power density assessment method and system, and electronic equipment. Background Technology

[0002] With the rapid development of wireless communication technology, electromagnetic waves are increasingly used in daily life and industrial applications. Electromagnetic absorbed power density is an important physical quantity for evaluating the interaction between electromagnetic fields and biological tissues or electronic devices. Accurate assessment of electromagnetic absorbed power density is of significant value and importance in fields such as medical device compliance testing, safety assessment of consumer electronics products, and industrial electromagnetic environment monitoring.

[0003] Currently, the assessment of electromagnetic absorption power density typically employs two main methods: numerical simulation and physical measurement. Numerical simulation, based on an idealized geometric model of the target object and its electromagnetic parameters, predicts the internal electromagnetic field distribution by solving Maxwell's equations. This method offers advantages such as low cost, high repeatability, and the ability to obtain full-space data. Physical measurement, on the other hand, relies on devices such as field strength probes and mechanical scanning platforms to perform point-by-point or scanning measurements of the measured area in a real or simulated physical space. The results are considered the gold standard reflecting the actual physical process.

[0004] In practical applications, these two methods are usually treated as two independent steps. First, the evaluator performs numerical simulations based on an idealized model to obtain a prediction result. Then, physical measurements are performed according to a fixed, pre-set sampling path (e.g., equidistant grid scanning) to verify or obtain the final data.

[0005] However, this linear, fragmented workflow has inherent flaws: On the one hand, numerical simulation predictions are based on idealized models and cannot quantify the impact of uncertainties such as material property deviations and boundary condition changes in the real measurement environment, leading to unknown reliability of the simulation results. On the other hand, physical measurements use a fixed uniform sampling pattern for full-domain scanning, which is not only time-consuming but also unable to focus on verifying key areas with low reliability or drastic field strength changes in the simulation results, resulting in low measurement efficiency and information redundancy. Due to the disconnect between simulation and measurement, measurement data cannot effectively feed back and correct systematic biases in the simulation model, limiting the accuracy, efficiency, and reliability of the evaluation results in complex real-world application scenarios.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the embodiments of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0008] This disclosure provides a method, system, and electronic device for evaluating electromagnetic absorption power density, aiming to address the problems of disconnect between simulation and measurement, and insufficient efficiency and reliability in traditional evaluation methods. This method constructs a closed-loop system for digital twin prediction and physical measurement verification, utilizes a deep learning surrogate model to achieve rapid prediction of electromagnetic absorption power density, intelligently guides the sampling strategy through uncertainty quantification, and calibrates and optimizes the model using measured data, resulting in continuous iterative improvement of the evaluation results.

[0009] In some embodiments, an electromagnetic absorption power density evaluation method includes the following steps: Obtain multiple physical field parameters of the scene to be evaluated; among them, the multiple physical field parameters include the wave parameters of the incident electromagnetic wave and the geometric model and organization electromagnetic parameters of the target object; Multiple physical field parameters are input into a model used to evaluate electromagnetic absorption power density to obtain the predicted results of electromagnetic absorption power density distribution and its spatial uncertainty distribution. The spatial electromagnetic field of a real-world scenario is sampled to obtain measured data; the sampling strategy is at least partially determined by the spatial uncertainty distribution. Based on the comparison between measured data and predicted results of electromagnetic absorption power density distribution, the parameters of the model are optimized. Multiple physical field parameters are input into the optimized model for re-evaluation.

[0010] Optionally, the geometric model of the target object is constructed based on a three-dimensional mesh containing nodes and edges, and is enclosed by several nodes and edges into spatial units for physical quantity calculation; wherein, the wave parameters of the incident electromagnetic wave and the organization electromagnetic parameters of the target object are defined as features attached to the nodes or spatial units.

[0011] Optionally, the model used to evaluate electromagnetic absorption power density is built based on a graph neural network and aggregates information from neighboring nodes in a three-dimensional mesh through a message passing mechanism to learn the mapping relationship from the input multi-physics parameters to the electromagnetic absorption power density distribution.

[0012] Optionally, the spatial uncertainty distribution of the prediction results is obtained by: enabling the built-in stochastic unit of the model used to evaluate the electromagnetic absorption power density during the inference process, performing multiple forward propagations on the same set of input parameters to obtain multiple prediction results; calculating the statistical variance of multiple prediction results on each spatial unit, and generating the spatial uncertainty distribution of the prediction results based on the correspondence between each spatial unit and the statistical variance.

[0013] Optionally, the randomness unit is the Dropout layer in the model, and the number of forward propagations N is an integer greater than 50.

[0014] Optionally, the predicted electromagnetic absorption power density distribution output by the model used to evaluate the electromagnetic absorption power density corresponds to each spatial cell, and the predicted result includes the average specific absorption rate of each spatial cell and the corresponding spatial uncertainty distribution.

[0015] Optionally, the average specific absorptivity of each spatial unit covering the target object constitutes the specific absorptivity distribution predicted by the model; the model parameters are optimized based on the comparison between the measured data and the predicted electromagnetic absorption power density distribution, including: calculating the measured specific absorptivity distribution based on the measured data; calculating the difference between the measured specific absorptivity distribution and the model-predicted specific absorptivity distribution as a loss function; when the loss function is greater than a second threshold, the measured specific absorptivity distribution is used as a supervision signal to optimize the model parameters, wherein the second threshold is determined based on the spatial uncertainty distribution, and / or the uncertainty of the measured data, and / or the historical calibration data.

[0016] Optionally, the formula for calculating the measured specific absorptivity distribution based on measured electromagnetic field data is as follows:

[0017] To organize electrical conductivity, For tissue mass density, , , These are the complex electric field components obtained from the measurement in three orthogonal directions.

[0018] Optionally, the model parameters are optimized, including: using the measured specific absorption rate distribution as a supervision signal, and updating the parameters of some layers of the pre-trained model through an optimization algorithm; wherein the optimization algorithm includes a dynamic stochastic gradient descent algorithm or an adaptive variant of dynamic stochastic gradient descent; and some layers include the output layer and the adjacent hidden layers.

[0019] Optionally, the sampling strategy is at least partially determined by the spatial uncertainty distribution, including: identifying the top K regions with the highest statistical variance values ​​based on the spatial uncertainty distribution, and prioritizing measurement of the identified regions, wherein K is determined based on the statistical characteristics of the spatial uncertainty distribution or the physical constraints of the measurement system; or, identifying regions with statistical variance values ​​higher than a first threshold based on the spatial uncertainty distribution, and automatically increasing the spatial density of sampling points within the identified regions, wherein the first threshold is determined based on the statistical characteristics of the spatial uncertainty distribution.

[0020] In some embodiments, an electromagnetic absorption power density assessment system includes: The digital twin prediction subsystem is used to acquire multiple physical field parameters of the scene to be evaluated and input the multiple physical field parameters into the model used to evaluate the electromagnetic absorption power density to obtain the prediction result of the electromagnetic absorption power density distribution and its spatial uncertainty distribution. The multiple physical field parameters include the wave parameters of the incident electromagnetic wave and the geometric model and organization electromagnetic parameters of the target object. The physical measurement subsystem is used to sample the spatial electromagnetic field of a real scene to obtain measured data, wherein the sampling strategy is at least partially determined by the spatial uncertainty distribution. The data fusion and decision-making subsystem is used to optimize the model parameters based on the comparison between measured data and the predicted results of electromagnetic absorption power density distribution, and to input multiple physical field parameters into the optimized model for re-evaluation.

[0021] In some embodiments, an electronic device includes a processor, a memory, and a computer program stored in the memory, wherein the processor, when executing the computer program, implements the aforementioned electromagnetic absorption power density evaluation method.

[0022] The electromagnetic absorption power density evaluation method, system, and electronic equipment provided in this disclosure can achieve the following technical effects: This disclosure improves the performance of electromagnetic absorption power density assessment in three aspects by constructing a closed-loop system of digital twin and physical verification, utilizing a deep learning proxy model to achieve rapid prediction and uncertainty quantification of electromagnetic absorption power density, intelligently guiding measurement resources based on uncertainty distribution, and finally calibrating the model through physical data: increasing efficiency by replacing complex numerical calculations with rapid prediction and avoiding resource waste from uniform scanning through intelligent measurement, thus achieving rapid prediction and accurate measurement; increasing accuracy by correcting systematic biases through physical data calibration, ensuring that the output results highly match physical reality and enabling continuous optimization of the assessment results; and increasing reliability by revealing prediction confidence through uncertainty quantification and avoiding overfitting through intelligent calibration triggering, making the reliability of the assessment results quantifiable and verifiable.

[0023] The above general description and the description below are exemplary and explanatory only, and are not intended to limit the embodiments of this disclosure. Attached Figure Description

[0024] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a schematic diagram of the overall architecture of the electromagnetic absorption power density evaluation system provided in the embodiments of this disclosure; Figure 2 This is a schematic diagram of the structure of the digital twin prediction subsystem provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram of the physical measurement subsystem provided in an embodiment of this disclosure; Figure 4 This is a schematic diagram of the structure of the data fusion and decision-making subsystem provided in the embodiments of this disclosure; Figure 5 This is a schematic diagram of the structure of the electromagnetic field scanning module provided in the embodiments of this disclosure; Figure 6 This is a schematic flowchart of the electromagnetic absorption power density evaluation method provided in the embodiments of this disclosure; Figure 7 This is an interactive schematic diagram of the electromagnetic absorption power density evaluation method provided in the embodiments of this disclosure; Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this disclosure. Detailed Implementation

[0025] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0026] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0027] Unless otherwise stated, the term "multiple" means two or more.

[0028] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0029] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0030] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0031] Currently, the assessment of electromagnetic absorption power density typically employs two main methods: numerical simulation and physical measurement. Numerical simulation, based on an idealized geometric model of the target object and its electromagnetic parameters, predicts the internal electromagnetic field distribution by solving Maxwell's equations. This method offers advantages such as low cost, high repeatability, and the ability to obtain full-space data. Physical measurement, on the other hand, relies on devices such as field strength probes and mechanical scanning platforms to perform point-by-point or scan-by-scan measurements of the measured area in a real or simulated physical space. The results are considered the gold standard reflecting the actual physical process. In practical applications, these two methods are usually treated as two independent steps. First, the assessor performs numerical simulation based on the idealized model to obtain a prediction result. Then, physical measurements are performed according to a fixed, pre-defined sampling path (e.g., equidistant grid scanning) to verify or obtain the final data.

[0032] This disclosure improves the performance of electromagnetic absorption power density assessment in three aspects by constructing a closed-loop system of digital twin and physical verification, utilizing a deep learning proxy model to achieve rapid prediction and uncertainty quantification of electromagnetic absorption power density, intelligently guiding measurement resources based on uncertainty distribution, and finally calibrating the model through physical data: increasing efficiency by replacing complex numerical calculations with rapid prediction and avoiding resource waste from uniform scanning through intelligent measurement, achieving second-level prediction and accurate measurement; increasing accuracy by correcting systematic biases through physical data calibration, ensuring the output results closely match physical reality and enabling continuous optimization of assessment results; and increasing reliability by revealing prediction confidence through uncertainty quantification and avoiding overfitting through intelligent calibration triggering, making the reliability of assessment results quantifiable and verifiable.

[0033] Figure 1 An electromagnetic absorption power density assessment system according to an embodiment of this disclosure is illustrated. The system includes a digital twin prediction subsystem 100, a physical measurement subsystem 200, and a data fusion and decision-making subsystem 300. The digital twin prediction subsystem 100 is used to acquire multiple physical field parameters of the scene to be evaluated, and input the multiple physical field parameters into the model for evaluating the electromagnetic absorption power density to obtain the prediction result of the electromagnetic absorption power density distribution and its spatial uncertainty distribution. The multiple physical field parameters include the wave parameters of the incident electromagnetic wave and the geometric model and organization electromagnetic parameters of the target object. The physical measurement subsystem 200 is used to sample the spatial electromagnetic field of a real scene to obtain measured data. The sampling strategy is at least partially determined by the spatial uncertainty distribution. The data fusion and decision subsystem 300 is used to optimize the model parameters based on the comparison results of measured data and the predicted results of electromagnetic absorption power density distribution, and input multiple physical field parameters into the optimized model for re-evaluation.

[0034] This system, through the collaborative operation of the digital twin prediction subsystem 100, the physical measurement subsystem 200, and the data fusion and decision-making subsystem 300, constructs a closed-loop architecture for digital twin and physical verification, breaking the traditional linear workflow that separates numerical simulation and physical measurement. The digital twin prediction subsystem 100 enables rapid prediction and uncertainty quantification of electromagnetic absorption power density; the physical measurement subsystem 200 achieves intelligent sampling based on uncertainty distribution, avoiding the resource waste of uniform scanning; and the data fusion and decision-making subsystem 300 calibrates model parameters using physical measurement data, achieving continuous model optimization. The synergy of these three subsystems enables the evaluation system to possess high efficiency, accuracy, and reliability, making it adaptable to complex application scenarios such as medical device compliance testing, consumer electronics safety assessment, and industrial electromagnetic environment monitoring.

[0035] Reference Figure 2 As shown, the digital twin prediction subsystem 100 serves as the core prediction unit of the evaluation system. It includes a parameter acquisition module 101, a deep learning proxy model module 102, and an uncertainty quantification module 103. These sub-modules work together to predict the electromagnetic absorption power density distribution and generate the spatial uncertainty distribution.

[0036] In one embodiment of this disclosure, the parameter acquisition module 101 is used to collect and preprocess multi-physics parameters of the scene to be evaluated. The acquired parameters include wave parameters of the incident electromagnetic wave (such as frequency, wave vector, polarization direction, etc.), the geometric model of the target object, and the organizational electromagnetic parameters of the target object (such as dielectric constant, conductivity, etc.). For the geometric model of the target object, the module processes it into a three-dimensional mesh based on nodes and edges. The three-dimensional mesh is composed of several nodes and edges forming spatial units for physical quantity calculation. At the same time, the module defines the wave parameters of the incident electromagnetic wave and the organizational electromagnetic parameters of the target object as features attached to the nodes or spatial units of the three-dimensional mesh. If the feature is attached to a node, it represents the parameter value of that node. If the feature is attached to a spatial unit, it represents the average / representative parameter value within that unit. The preprocessed parameters will be used as input data for the deep learning proxy model module 102 to realize the conversion of physical field parameters into a model-recognizable data format.

[0037] In one embodiment of this disclosure, the deep learning proxy model module 102 is based on an electromagnetic absorption power density assessment model constructed using a graph neural network. The graph neural network aggregates information from adjacent nodes in a three-dimensional mesh through a message passing mechanism, learning the mapping relationship from input multi-physics parameters to the electromagnetic absorption power density distribution. The model's output corresponds to each spatial cell of the target object's three-dimensional mesh, with the output value being the average specific absorption rate of each spatial cell. The average specific absorption rate of all spatial cells constitutes the model's predicted specific absorption rate distribution, which is the predicted result of the electromagnetic absorption power density distribution. Compared to the traditional method of predicting electromagnetic absorption power density by solving Maxwell's equations, the architecture of the graph neural network is naturally compatible with the discretized mesh in electromagnetic computation, enabling it to bypass complex numerical calculations and achieve millisecond-level fast forward prediction. Furthermore, it can effectively capture the near-field coupling and far-field interaction of electromagnetic waves propagating in complex media, improving the prediction accuracy for electromagnetic hotspot regions in complex structures.

[0038] For example, the graph neural network in this embodiment uses a 3-layer graph convolutional neural network (GCN) as its core architecture. The node aggregation method is mean aggregation, assigning lower weights to distant nodes when aggregating features of adjacent nodes, thus aligning with the spatial attenuation characteristics of electromagnetic wave propagation. The ReLU activation function is used to address the gradient vanishing problem during model training. The output layer uses a linear activation function to directly output the average specific absorption rate of each spatial unit. The model's training dataset consists of multi-scenario electromagnetic absorption power density simulation data based on Maxwell's equations, covering different geometric models, organized electromagnetic parameters, and incident electromagnetic wave parameters. The training iterations are set to 2000 rounds, using mean squared error (MSE) as the loss function. The optimization algorithm is a stochastic gradient descent algorithm with momentum (momentum coefficient 0.9), trained until the loss function converges to 1e. -6Stop at the specified time to obtain the pre-trained model.

[0039] In one embodiment of this disclosure, the uncertainty quantification module 103 is linked with the deep learning proxy model module 102 to generate the spatial uncertainty distribution of the prediction results. Specifically, during the inference process of the model used to evaluate the electromagnetic absorption power density, its built-in stochastic unit is enabled. During the inference phase, in the Monte Carlo Dropout manner, that is, the stochastic unit (such as the Dropout layer) in the model is kept in an active state, and multiple forward propagations are performed on the same set of input parameters to obtain multiple prediction results. The statistical variance of the multiple prediction results on each spatial unit is calculated, and the spatial uncertainty distribution of the prediction results is generated based on the correspondence between each spatial unit and the statistical variance.

[0040] In this model, the randomness unit is a Dropout layer built into the model for evaluating electromagnetic absorption power density. The number of forward propagations, N, is an integer greater than 50. During the training phase, the Dropout layer randomly discards some neuron connections as a regularization method. During the inference phase, it maintains activation and can temporarily construct slightly different subnetworks. Multiple forward propagations of the same input are equivalent to Monte Carlo sampling from the model's posterior parameter distribution. When N ≤ 50, the sample size is insufficient, and the statistical variance is easily affected by random disturbances, making it impossible to stably approximate the model's true posterior distribution, resulting in distortion of the spatial uncertainty distribution. In this embodiment, the number of propagations is set to an integer greater than 50, which can meet the requirements of Monte Carlo sampling for statistical significance, making the statistical variance stably and reliably approximate the true model posterior distribution, and ensuring that the spatial uncertainty distribution can truly reflect the prediction confidence of the model at different spatial locations.

[0041] The spatial uncertainty distribution generated by the uncertainty quantification module 103 and the average specific absorption rate output by the deep learning proxy model module 102 are synchronously transmitted to the physical measurement subsystem 200 to provide a basis for the formulation of sampling strategies. At the same time, they are also transmitted to the data fusion and decision-making subsystem 300 to provide a reference for the threshold setting of model parameter optimization.

[0042] Reference Figure 3 As shown, the physical measurement subsystem 200 has a built-in sampling strategy formulation module 201, an electromagnetic field scanning module 202, and a data preprocessing module 203. The sub-modules work together to achieve intelligent sampling and acquisition of measured data based on spatial uncertainty distribution. The electromagnetic field scanning module 202 is equipped with core hardware such as a six-degree-of-freedom robotic arm 2021, a multi-channel vector network analyzer 2022, a millimeter-wave phased array antenna 2023, and a phantom platform for the target under test, so as to achieve high-precision and automated electromagnetic field scanning sampling.

[0043] In one embodiment of this disclosure, the sampling strategy formulation module 201 receives the spatial uncertainty distribution transmitted by the digital twin prediction subsystem 100, and formulates an electromagnetic field sampling strategy for a real-world scenario based on this distribution. The sampling strategy is at least partially determined by the spatial uncertainty distribution, and can be implemented in two ways, which can be used individually or in combination: Method 1: Based on the spatial uncertainty distribution, identify the top K regions with the highest statistical variance values ​​and prioritize measurement of these regions. Here, K is determined based on the statistical characteristics of the spatial uncertainty distribution or the physical constraints of the measurement system. Specifically, calculate the mean μ and standard deviation σ of the uncertainty values ​​for all spatial units, and set a statistical threshold T = μ + ασ, where α is a correction coefficient for the statistical threshold of the spatial uncertainty distribution, ranging from 1.5 to 3.0. If α < 1.5, too many high-uncertainty regions are identified, increasing measurement resource consumption and negating the efficiency advantage of intelligent sampling. If α > 3.0, too few high-uncertainty regions are identified, easily overlooking key areas where model predictions are unreliable, hindering effective calibration. Therefore, α should be set between 1.5 and 3.0. In this way, the optimal balance between measurement efficiency and calibration effectiveness can be achieved. The number of units with uncertainty values ​​greater than T is used as candidate K values, or the units are sorted from high to low uncertainty values, and the top K units are selected so that the proportion of their cumulative uncertainty to the total reaches a preset threshold η (η is 0.7~0.95). Then, the K value is corrected in combination with the physical constraints of the measurement system such as time and spatial resolution. In typical application scenarios, the value of K is an integer between 3 and 50.

[0044] Method 2: Based on the spatial uncertainty distribution, identify regions where the statistical variance value is higher than the first threshold, and automatically increase the spatial density of sampling points within the identified regions. The first threshold is determined based on the statistical characteristics of the spatial uncertainty distribution.

[0045] The sampling strategy formulated by the sampling strategy formulation module 201 abandons the traditional fixed uniform sampling mode and realizes the transformation from "blind uniform scanning" to "intelligent key verification". It directs limited measurement resources to key areas with low model prediction reliability and drastic field strength changes, thereby improving the efficiency and relevance of physical measurements.

[0046] Reference Figure 5 As shown in one embodiment of this disclosure, the electromagnetic field scanning module 202 is the core hardware execution unit of the physical measurement subsystem 200. The core is equipped with a six-degree-of-freedom robotic arm 2021, a multi-channel vector network analyzer 2022, and other hardware such as a millimeter-wave phased array antenna 2023 and a phantom platform for the target under test. According to the sampling strategy instructions output by the sampling strategy formulation module 201, it realizes high-precision and automated electromagnetic field scanning and sampling of the phantom for the target under test in a real scene. The phantom for the target under test is fixed on the platform to ensure the stability of the phantom's position during the sampling process.

[0047] Among them, the six-degree-of-freedom robotic arm 2021 serves as the movement and positioning actuator for the millimeter-wave phased array antenna 2023. It receives instructions such as the spatial coordinates of the sampling points and the scanning path output by the sampling strategy formulation module 201. Through precise movement of six degrees of freedom, it drives the millimeter-wave phased array antenna 2023 mounted at the end of the robotic arm to move in three-dimensional space, realizing fixed-point sampling and continuous scanning of the designated area of ​​the target phantom. Its high-precision positioning capability ensures that the antenna detection position and the coordinates formulated by the strategy are completely matched, with extremely low positioning error, meeting the electromagnetic field sampling requirements of target phantoms with complex geometric structures. For dense sampling areas, the six-degree-of-freedom robotic arm 2021 can automatically adjust the movement step size according to the preset step density to achieve high-density scanning of sampling points. At the same time, it can adjust the antenna detection angle according to the geometric contour of the target phantom to ensure effective coupling between the detection beam and the area to be measured. The Multichannel Vector Network Analyzer 2022, serving as the core hardware for electromagnetic field data acquisition and analysis, establishes a high-frequency signal connection with the millimeter-wave phased array antenna 2023. On one hand, it outputs excitation signals to the millimeter-wave phased array antenna 2023, precisely controlling the antenna to emit detection electromagnetic waves that match the parameters (frequency, polarization, power, etc.) of the incident electromagnetic waves, ensuring a high degree of consistency between the detection scenario and the actual electromagnetic scenario to be evaluated. On the other hand, it receives scattered / reflected electromagnetic wave signals around the target phantom acquired by the millimeter-wave phased array antenna 2023, performs real-time analysis, amplitude and phase demodulation of the vector signals received from multiple channels, accurately acquires and calculates the complex electric field components Ex, Ey, and Ez in three orthogonal directions, and simultaneously records the spatial coordinates, frequency, phase, and other key parameters of each sampling point, providing high-precision raw electromagnetic data for subsequent specific absorption rate calculation. Its multi-channel parallel acquisition characteristics enable simultaneous detection of multiple regions of the target phantom, significantly improving overall scanning efficiency. Millimeter-wave phased array antenna 2023: As the RF front-end for electromagnetic wave transmission and reception, it forms a closed-loop signal interaction with the multi-channel vector network analyzer 2022. Based on the incident electromagnetic wave parameters of the scene to be evaluated, it flexibly adjusts the beam pointing, polarization, transmission power, and beamwidth under the control of the multi-channel vector network analyzer 2022 to achieve focused electromagnetic wave detection of the target phantom area under test, effectively reducing the impact of environmental electromagnetic interference on the sampling data. At the same time, it transmits the electromagnetic scattering / reflection signals of the target phantom received during the detection process back to the multi-channel vector network analyzer 2022 in real time to complete the acquisition and transmission of electromagnetic signals, ensuring that the characteristics of the detected electromagnetic wave are consistent with the incident electromagnetic wave of the scene to be evaluated, and improving the authenticity and matching degree of the measured data.

[0048] In one embodiment of this disclosure, the data preprocessing module 203 performs multi-dimensional preprocessing on the raw electromagnetic field data collected by the electromagnetic field scanning module 202. First, it performs channel calibration on the multi-channel data collected by the multi-channel vector network analyzer 2022, eliminating signal gain and phase deviation between hardware channels through calibration algorithms to ensure the consistency of data from each channel. Second, it performs noise reduction processing on the collected electromagnetic signals, removing invalid signals caused by environmental electromagnetic interference and hardware circuit noise through digital filtering algorithms, while retaining effective electromagnetic feature information. Then, it performs normalization calibration on the spatial coordinates of the sampling points recorded by the six-degree-of-freedom robotic arm 2021, converting the physical sampling coordinates into a three-dimensional grid coordinate system consistent with the digital twin prediction subsystem 100, realizing the coordinate unification between physical space and digital space, and laying the foundation for subsequent data registration. Finally, it performs format standardization and encapsulation on the preprocessed electromagnetic data, extracts the core complex electric field components Ex, Ey, and Ez and their corresponding sampling point coordinates, and synchronously transmits them to the data fusion and decision subsystem 300 to provide high-quality physical measurement data for model parameter optimization.

[0049] Reference Figure 4 As shown, the data fusion and decision subsystem 300 realizes the closed-loop calibration and result output of the evaluation system. It has built-in data registration and comparison module 301, model optimization module 302, re-evaluation module 303, and result output module 304. The sub-modules work together to realize the optimization of model parameters and the final evaluation of electromagnetic absorption power density.

[0050] In one embodiment of this disclosure, the data registration and comparison module 301 receives measured electromagnetic field data transmitted by the physical measurement subsystem 200 and electromagnetic absorption power density distribution prediction results transmitted by the digital twin prediction subsystem 100. First, it calculates the measured specific absorption rate distribution based on the measured electromagnetic field data using the following formula:

[0051] Where σ is the tissue conductivity, ρ is the tissue mass density, and Ex, Ey, and Ez are the complex electric field components in three orthogonal directions obtained from measurement. This formula, based on Joule's law of electromagnetic heat dissipation, transforms directly measurable electric field vector information into a specific absorptivity with clear physical meaning, establishing a rigorous physical bridge from measured physical quantities to evaluation indicators. Subsequently, this module uses spatial interpolation techniques (such as radial basis function interpolation and grid-based linear interpolation) to map sparse measured specific absorptivity data onto dense 3D grid nodes / spatial cells where the model prediction resides, achieving spatial registration between measured and predicted data. After registration, the difference between the measured specific absorptivity distribution and the model-predicted specific absorptivity distribution is calculated, and this difference is used as a loss function. The value of the loss function serves as the criterion for optimizing model parameters.

[0052] In one embodiment of this disclosure, the model optimization module 302 presets a second threshold and compares the loss function calculated by the data registration and comparison module 301 with the second threshold. The second threshold is determined based on the spatial uncertainty distribution, and / or the uncertainty of the measured data, and / or the historical calibration data. When the loss function is greater than the second threshold, it indicates that there is a substantial cognitive bias in the model. At this time, the measured specific absorption rate distribution is used as a supervision signal to optimize the model parameters.

[0053] In one specific embodiment, the second threshold can be determined independently based on the spatial uncertainty distribution. Specifically, a statistical analysis is performed on the spatial uncertainty distribution of the model prediction output, calculating the mean μ and standard deviation σ of the uncertainty values ​​of all spatial units. The second threshold T2 can be set as T2 = μ + β·σ, where β is a preset confidence coefficient, ranging from 1.5 to 3.0. The larger the β value, the higher the tolerance for model prediction deviation, and the more stringent the conditions for triggering model calibration; the smaller the β value, the more sensitive the calibration trigger. This implementation is suitable for scenarios where the uncertainty of measured data is small or historical calibration data is insufficient, and can quickly identify areas with low model prediction confidence and trigger calibration.

[0054] In one specific embodiment, the second threshold can be determined independently based on the uncertainty of the measured data. The uncertainty of the measured data consists of hardware errors of the measurement system, including the positioning error σ1 of the six-degree-of-freedom robotic arm, which is the nominal value of the root mean square error of the three-dimensional spatial positioning calibrated at the factory; the measurement error σ2 of the multi-channel vector network analyzer, which is obtained by converting the amplitude and phase measurement errors into the standard deviation of the electric field component measurement errors; and the overall uncertainty of the measured data is the root value σ of the sum of the squares of each error component (σ1 and σ2). total The second threshold T2 can be set to be the same as σ. total A proportional value, for example, T2 = γ·σ total γ is the amplification factor, ranging from 2 to 5. This implementation method is suitable for scenarios where the accuracy of the measurement system fluctuates greatly or where high measurement reliability is required, ensuring that model calibration is only performed when the measured data is sufficiently reliable.

[0055] In one specific embodiment, the second threshold can be determined solely based on historical calibration data. Historical calibration data includes the loss function values ​​before and after model calibration recorded in previous evaluation cycles for the same or similar scenarios. Statistical analysis is performed on the historical loss function values ​​to calculate their mean μL and the range of successful calibration trigger thresholds. The second threshold T2 can be set as T2 = μL + δ·σL, where σL is the standard deviation of the historical loss function, and δ is an empirical coefficient ranging from 0.5 to 1.5. This implementation is suitable for scenarios with repeated evaluation requirements, enabling dynamic adjustment of calibration trigger conditions based on historical experience to avoid over-calibration or under-calibration.

[0056] In another specific embodiment, the second threshold can also be determined jointly based on the spatial uncertainty distribution and the uncertainty of the measured data, or jointly based on the uncertainty of the measured data and the historical calibration data, or jointly based on the spatial uncertainty distribution and the historical calibration data, for example, by setting the weight coefficients of the corresponding factor values ​​and performing weighted summation.

[0057] Furthermore, the second threshold can also be based on the average variance μσ of the spatial uncertainty distribution and the comprehensive uncertainty σ of the aforementioned measured data. total The loss value μL is determined jointly by the average loss value of historical calibration data, and the calculation formula can be set as T2=k·(μσ+σ total (k is an empirical coefficient, ranging from 1.0 to 1.2). This implementation method integrates comprehensive information from three factors, enabling the most accurate and robust calibration trigger judgment, and is suitable for scenarios with the highest requirements for evaluation accuracy.

[0058] The specific method for optimizing the model parameters is as follows: using the measured specific absorbance distribution as a supervision signal, the parameters of some layers of the pre-trained model are updated through an optimization algorithm. The optimization algorithm includes stochastic gradient descent with dynamics or an adaptive variant of stochastic gradient descent with dynamics (such as the Adam algorithm), and the "some layers" include the output layer and adjacent hidden layers. By adopting a "partial fine-tuning" strategy instead of retraining the model, the model can quickly adapt to the local physical characteristics of a specific scene based on the learned general electromagnetic laws, minimizing computational overhead and the risk of overfitting.

[0059] In one embodiment of this disclosure, after the re-evaluation module 303 completes the optimization of model parameters, it re-inputs multiple physical field parameters of the scene to be evaluated obtained by the digital twin prediction subsystem 100 into the optimized electromagnetic absorption power density evaluation model, starts the inference process of the model, realizes the re-evaluation of electromagnetic absorption power density, and obtains the optimized electromagnetic absorption power density distribution prediction result.

[0060] In one embodiment of this disclosure, the result output module 304 visualizes the re-evaluated electromagnetic absorption power density distribution result, and integrates information such as spatial uncertainty distribution and measured data comparison results to generate a final electromagnetic absorption power density evaluation report, which is then displayed to the user through a human-computer interaction interface, realizing the output and visualization of the evaluation results. If the re-evaluated result still does not meet the preset convergence requirements, the module feeds back the optimized prediction result and the corresponding spatial uncertainty distribution to the physical measurement subsystem 200, initiating a new round of sampling and model calibration until the loss function is less than or equal to the second threshold, thereby achieving continuous iterative optimization of the model.

[0061] Reference Figure 6 , Figure 7 As shown, another embodiment of this disclosure provides a method for evaluating electromagnetic absorption power density, the method comprising the following steps: S1. Obtain multiple physical field parameters of the scene to be evaluated; among which, the multiple physical field parameters include the wave parameters of the incident electromagnetic wave and the geometric model and organization electromagnetic parameters of the target object; S2. Input multiple physical field parameters into the model used to evaluate electromagnetic absorption power density to obtain the predicted results of electromagnetic absorption power density distribution and its spatial uncertainty distribution. S3. Sample the spatial electromagnetic field of the real scene to obtain measured data; wherein the sampling strategy is at least partially determined by the spatial uncertainty distribution. S4. Optimize the model parameters based on the comparison between the measured data and the predicted results of the electromagnetic absorption power density distribution. S5. Input multiple physical field parameters into the optimized model and re-evaluate.

[0062] The method in this embodiment, relying on the collaborative work of the digital twin prediction subsystem 100, the physical measurement subsystem 200, and the data fusion and decision-making subsystem 300, achieves rapid prediction and uncertainty quantification of electromagnetic absorption power density. Intelligent sampling based on uncertainty distribution significantly improves the efficiency of physical measurement. The model is continuously optimized through calibration of physical measurement data. The final output evaluation result is highly consistent with physical reality, and the prediction confidence is quantifiable and verifiable. It effectively solves the problems of simulation and measurement disconnect, low evaluation efficiency, and insufficient accuracy and reliability in traditional evaluation methods.

[0063] The specific implementation of each step of this method will be explained in detail below.

[0064] Step S1: Obtain multiple physical field parameters of the scene to be evaluated. This step S1 is executed by the parameter acquisition module 101 of the digital twin prediction subsystem 100. The multiple physical field parameters of the scene to be evaluated include the wave parameters of the incident electromagnetic wave and the geometric model and organizational electromagnetic parameters of the target object. Among them, the wave parameters of the incident electromagnetic wave include key features such as frequency, wave vector, and polarization direction, and the organizational electromagnetic parameters of the target object include core parameters such as dielectric constant and conductivity. The geometric model refers to the digital geometric representation of the target object in the computer, which is the original geometric data that needs to be discretized before electromagnetic simulation calculation.

[0065] The parameter acquisition module 101 collects the original physical field parameters of the scene to be evaluated through a data acquisition interface, discretizes the geometric model of the target object, and constructs a three-dimensional mesh based on nodes and edges. This three-dimensional mesh consists of several nodes and edges forming spatial units for physical quantity calculation. Simultaneously, the wave parameters of the incident electromagnetic wave and the electromagnetic parameters of the target object's structure are mapped as features attached to the nodes or spatial units of the three-dimensional mesh, completing the parameter preprocessing. Defining the physical field parameters as node / unit features of the three-dimensional mesh allows for compatibility with the input format of the subsequent graph neural network-based evaluation model, realizing the conversion of physical field parameters into model-recognizable data and laying a data foundation for the model's inference and prediction.

[0066] Step S2 involves inputting multiple physical field parameters into a model used to evaluate the electromagnetic absorption power density, thereby obtaining the predicted results of the electromagnetic absorption power density distribution and the spatial uncertainty distribution. This step is executed collaboratively by the deep learning proxy model module 102 and the uncertainty quantization module 103 of the digital twin prediction subsystem 100. The multiple physical field parameters preprocessed in Step 1 are input into the model used to evaluate the electromagnetic absorption power density. First, the deep learning proxy model module 102 outputs the predicted results of the electromagnetic absorption power density distribution, and then the uncertainty quantization module 103 generates the corresponding spatial uncertainty distribution.

[0067] The model for evaluating electromagnetic absorption power density is based on a graph neural network. The graph neural network takes a preprocessed 3D mesh as input and aggregates information from adjacent nodes in the 3D mesh through a message-passing mechanism, learning the mapping relationship from the input multiphysics parameters to the electromagnetic absorption power density distribution. The model's output corresponds to each spatial cell of the target object's 3D mesh, with the output value being the average specific absorption rate of each spatial cell. The average specific absorption rate of all spatial cells constitutes the model's predicted specific absorption rate distribution, i.e., the predicted electromagnetic absorption power density distribution. The message-passing mechanism of the graph neural network can progressively expand the information receptive field of nodes, effectively capturing the spatial correlation of electromagnetic wave propagation in complex media, improving the prediction accuracy of electromagnetic absorption power density distribution for complex structures, and bypassing the cumbersome calculations of solving Maxwell's equations in traditional numerical simulations, achieving rapid prediction.

[0068] Specifically, during the model's inference process, the uncertainty quantification module 103 utilizes the model's built-in Dropout layer as a stochastic unit, performing N forward propagations (N being an integer greater than 50) on the same set of input parameters to obtain N prediction results. For each spatial cell of the 3D mesh, the statistical variance of the N prediction results in that cell is calculated. Based on the correspondence between each spatial cell and the statistical variance, a spatial uncertainty distribution of the prediction results is generated. Using the Dropout layer as a stochastic unit transforms the deterministic neural network into an efficient Bayesian probabilistic model. The Monte Carlo sampling method with multiple forward propagations allows the statistical variance to stably reflect the model's cognitive uncertainty, while the spatial uncertainty distribution quantitatively characterizes the model's prediction confidence at different spatial locations, providing a core basis for formulating sampling strategies for subsequent physical measurements.

[0069] Step S3 involves sampling the spatial electromagnetic field of the real scene to obtain measured data. This step is executed collaboratively by the sampling strategy formulation module 201, the electromagnetic field scanning module 202, and the data preprocessing module 203 of the physical measurement subsystem 200. The sampling strategy is at least partially determined by the spatial uncertainty distribution generated in step S2. The electromagnetic field scanning module 202 achieves high-precision sampling through hardware collaboration of a six-degree-of-freedom robotic arm 2021, a multi-channel vector network analyzer 2022, and a millimeter-wave phased array antenna 2023.

[0070] Specifically, the sampling strategy formulation module 201 receives the spatial uncertainty distribution from step S2 and formulates an intelligent sampling strategy based on this distribution. This can be achieved in two ways: First, it identifies the top K regions with the highest statistical variance values ​​and prioritizes their measurement. K is determined based on the statistical characteristics of the spatial uncertainty distribution or the physical constraints of the measurement system (such as the motion range of a six-degree-of-freedom robotic arm, the sampling efficiency of a multi-channel vector network analyzer, and the detection range of an antenna). Second, it identifies regions with statistical variance values ​​higher than a first threshold and automatically densifies the spatial density of sampling points within these regions. The first threshold is determined based on the statistical characteristics of the spatial uncertainty distribution. This sampling strategy breaks away from the traditional fixed uniform sampling model, precisely targeting measurement resources to key areas with low model prediction reliability, thus achieving optimal allocation of measurement resources.

[0071] The electromagnetic field scanning module 202, according to the sampling strategy, uses a six-degree-of-freedom robotic arm 2021 to drive a millimeter-wave phased array antenna 2023 to achieve high-precision spatial positioning and angle adjustment. The multi-channel vector network analyzer 2022 outputs excitation signals to the antenna and receives the returned electromagnetic signals, calculating the original electromagnetic field data such as the complex electric field components Ex, Ey, and Ez in three orthogonal directions that match the incident electromagnetic wave. The high-precision positioning of the six-degree-of-freedom robotic arm 2021 ensures that the coordinates of the sampling point are completely matched with the digital space, and the multi-channel vector network analyzer 2022 achieves high-precision calculation of electromagnetic signals and multi-channel parallel acquisition, greatly improving the authenticity, accuracy, and acquisition efficiency of the original data.

[0072] The data preprocessing module 203 performs noise reduction, calibration, and coordinate normalization on the raw electromagnetic field data to eliminate the uncertainty of the measured data caused by the positioning error of the six-degree-of-freedom robotic arm 2021, the channel deviation of the multi-channel vector network analyzer 2022, the antenna beam pointing deviation, and environmental electromagnetic interference. After encapsulating the data, the final measured data is obtained and transmitted to the data fusion and decision subsystem 300 to provide high-quality physical data support for model parameter optimization.

[0073] Step S4: Optimize the model parameters based on the comparison results between the measured data and the prediction results. This step is executed collaboratively by the data registration and comparison module 301 and the model optimization module 302 of the data fusion and decision subsystem 300. First, the comparison between the measured data and the prediction results is completed, and then the parameters of the evaluation model are optimized based on the comparison results.

[0074] Specifically, the data registration and comparison module 301 first uses the measured electromagnetic field data from step S3, based on the formula: The measured specific absorptivity distribution is calculated using a formula based on Joule's law of electromagnetic heat dissipation, establishing a physical correlation between the measured electric field and the specific absorptivity, ensuring a unified evaluation benchmark between the digital and physical spaces. Subsequently, spatial interpolation is used to map the measured specific absorptivity data onto the model's predicted 3D grid, achieving spatial registration of the data. After registration, the difference between the measured and model-predicted specific absorptivity distributions is calculated and used as the loss function. A second threshold is preset through the model optimization module 302, determined based on one or more of the following: spatial uncertainty distribution, measured data uncertainty, and historical calibration data. The loss function is compared with the second threshold. If the loss function is greater than the second threshold, it indicates a systematic cognitive bias in the model. In this case, the measured specific absorptivity distribution is used as a monitoring signal, and the output layer and adjacent hidden layers of the pre-trained model are updated using a stochastic gradient descent algorithm with momentum or its adaptive variant. This "partial fine-tuning" parameter optimization method allows the model to quickly adapt to the local physical characteristics of a specific evaluation scenario while retaining the learning results of general electromagnetic laws, improving the model's prediction accuracy while minimizing computational overhead and overfitting risk.

[0075] Step S5 involves inputting multiple physical field parameters into the optimized model for re-evaluation. This step is executed collaboratively by the re-evaluation module 303 and the result output module 304 of the data fusion and decision subsystem 300. The multiple physical field parameters of the scenario to be evaluated obtained in step S1 are re-input into the electromagnetic absorption power density evaluation model optimized in step S4 to achieve a re-evaluation of the electromagnetic absorption power density and output the final evaluation result.

[0076] Specifically, the re-evaluation module 303 initiates the inference process of the optimized model. Based on the updated parameters, the model processes the input multiphysics parameters and outputs the optimized electromagnetic absorption power density distribution prediction result. If the loss function corresponding to this result is still greater than the second threshold, the optimized prediction result and spatial uncertainty distribution are fed back to the physical measurement subsystem 200 to initiate a new round of sampling and model calibration until the loss function is less than or equal to the second threshold. The result output module 304 visualizes the converged prediction result, integrates information such as spatial uncertainty distribution and measured data comparison results, and generates the final SAR evaluation result. This SAR evaluation result includes the specific absorption rate distribution map of each spatial cell, the corresponding confidence interval information, and an overall evaluation report, which is output through a human-computer interaction interface, completing the entire electromagnetic absorption power density evaluation process. This step, through re-evaluation after model optimization, achieves continuous iterative improvement of the evaluation result, making the final evaluation result highly consistent with physical reality, and significantly improving the accuracy and reliability of the evaluation result.

[0077] The method in this embodiment first obtains the physical field parameters (incident electromagnetic wave parameters, target object geometric model, and tissue electromagnetic parameters) of the scene to be evaluated. It then rapidly predicts the electromagnetic absorption power density distribution using a deep learning surrogate model based on a graph neural network. During the inference phase, it performs multiple forward propagations using Monte Carlo Dropout to calculate the statistical variance and generate a spatial uncertainty distribution. Subsequently, it intelligently formulates a sampling strategy based on this uncertainty distribution, prioritizing or intensively sampling high-uncertainty regions to obtain measured data. The measured data is compared with the prediction results, and a loss function is calculated. When the loss function exceeds a second threshold (which can be determined based on the spatial uncertainty distribution, the uncertainty of the measured data, and / or historical calibration data), the measured data is used as a supervision signal to fine-tune the parameters of the model's output layer and adjacent hidden layers. Finally, the optimized model is re-evaluated, forming a closed-loop iterative process of "prediction-sampling-calibration-re-prediction".

[0078] This method replaces traditional numerical simulation with a graph neural network surrogate model to solve Maxwell's equations, achieving rapid prediction. It uses Monte Carlo Dropout to simulate Bayesian inference, quantifying the model's prediction confidence at different spatial locations. Uncertainty distribution is used as a navigation map for measurement resources, concentrating limited physical measurements in the least reliable areas of the model. Finally, the model is "partially fine-tuned" using measured data, enabling it to quickly adapt to local physical characteristics of specific scenarios while preserving general electromagnetic laws. This method breaks the traditional linear, fragmented workflow of "simulation first, measurement later." On one hand, it reveals the credibility of simulation results through uncertainty quantification, overcoming the limitation of existing simulation techniques in quantifying their own uncertainties. On the other hand, it intelligently guides measurement resources through uncertainty, avoiding the inefficiency and information redundancy problems caused by traditional uniform sampling. More importantly, it establishes a closed-loop information feedback between simulation and measurement, enabling measured data to actively calibrate the simulation model and correct systematic biases, thereby simultaneously improving the efficiency (rapid prediction + intelligent measurement), accuracy (physical data calibration and correction of biases), and reliability (quantifiable and verifiable uncertainty) of the evaluation.

[0079] Combination Figure 8 As shown in the illustration, this disclosure also provides an electronic device for evaluating electromagnetic absorption power density, including a processor 800 and a memory 801. Optionally, the device may further include a communication interface 802 and a bus 803. The processor 800, communication interface 802, and memory 801 can communicate with each other via the bus 803. The communication interface 802 can be used for information transmission. The processor 800 can call logical instructions in the memory 801 to execute the electromagnetic absorption power density evaluation method of the above embodiments.

[0080] Furthermore, the logic instructions in the aforementioned memory 801 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0081] The memory 801, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 800 executes functional applications and data processing by running the program instructions / modules stored in the memory 801, that is, it implements the method for evaluating electromagnetic absorption power density in the above embodiments.

[0082] The memory 801 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 801 may include high-speed random access memory and may also include non-volatile memory.

[0083] The foregoing description and accompanying drawings fully illustrate embodiments of the present disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in the embodiments of this disclosure is for descriptive purposes only and is not intended to limit the claims. As used in the description of the embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in the embodiments of this disclosure refers to any and all possible combinations of one or more of the associated listed elements. Additionally, when used in embodiments of this disclosure, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, then the relevant parts can be referred to the description of the method section.

[0084] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

Claims

1. A method for evaluating electromagnetic absorption power density, characterized in that, include: Multiple physical field parameters of the scene to be evaluated are obtained; among them, the multiple physical field parameters include the wave parameters of the incident electromagnetic wave and the geometric model and organization electromagnetic parameters of the target object; the geometric model of the target object is constructed based on a three-dimensional mesh containing nodes and edges, and is enclosed by several nodes and edges to form a spatial unit for physical quantity calculation; among them, the wave parameters of the incident electromagnetic wave and the organization electromagnetic parameters of the target object are defined as features attached to the nodes or spatial units. Multiple physical field parameters are input into a model used to evaluate electromagnetic absorption power density, resulting in a predicted distribution of electromagnetic absorption power density and its spatial uncertainty distribution. The predicted distribution of electromagnetic absorption power density output by the model corresponds to each spatial cell, and the prediction includes the average specific absorption rate of each spatial cell and the corresponding spatial uncertainty distribution. The spatial uncertainty distribution of the prediction results is obtained as follows: during the inference process of the model used to evaluate electromagnetic absorption power density, its built-in stochastic unit is enabled, and the same set of input parameters is forward-propagated multiple times to obtain multiple prediction results. The statistical variance of the multiple prediction results on each spatial cell is calculated, and the spatial uncertainty distribution of the prediction results is generated based on the correspondence between each spatial cell and the statistical variance. The spatial electromagnetic field of a real scene is sampled to obtain measured data. Based on the spatial uncertainty distribution, the top K regions with the highest statistical variance values ​​are identified, and these regions are measured first. Here, K is determined based on the statistical characteristics of the spatial uncertainty distribution or the physical constraints of the measurement system. Alternatively, based on the spatial uncertainty distribution, regions with statistical variance values ​​higher than a first threshold are identified, and the spatial density of sampling points is automatically increased within these regions. Here, the first threshold is determined based on the statistical characteristics of the spatial uncertainty distribution. The difference between the measured specific absorbance distribution and the model-predicted specific absorbance distribution is calculated as the loss function. When the loss function is greater than the second threshold, the measured specific absorbance distribution is used as a supervision signal to optimize the model parameters. The second threshold is determined based on the spatial uncertainty distribution and / or the uncertainty of the measured data and / or the historical calibration data. Multiple physical field parameters are input into the optimized model for re-evaluation.

2. The method according to claim 1, characterized in that, The model used to evaluate electromagnetic absorption power density is built on a graph neural network and aggregates information from neighboring nodes in a 3D mesh through a message passing mechanism to learn the mapping relationship from the input multiphysics parameters to the electromagnetic absorption power density distribution.

3. The method according to claim 1, characterized in that, The randomness unit is the Dropout layer in the model, and the number of forward propagations N is an integer greater than 50.

4. The method according to claim 1, characterized in that, The average specific absorptivity of each spatial unit covering the target object constitutes the specific absorptivity distribution predicted by the model; the measured specific absorptivity distribution is calculated based on the measured data.

5. A system for evaluating electromagnetic absorption power density, characterized in that, include: The digital twin prediction subsystem is used to acquire multiple physical field parameters of the scene to be evaluated and input these parameters into a model for evaluating electromagnetic absorption power density. This yields a prediction of the electromagnetic absorption power density distribution and its spatial uncertainty distribution. The multiple physical field parameters include the wave parameters of the incident electromagnetic wave and the geometric model and organizational electromagnetic parameters of the target object. The geometric model of the target object is constructed based on a three-dimensional mesh containing nodes and edges, and is enclosed by several nodes and edges forming spatial units for physical quantity calculations. The wave parameters of the incident electromagnetic wave and the organizational electromagnetic parameters of the target object are defined as features attached to the nodes or spatial units. The physical measurement subsystem is used to sample the spatial electromagnetic field of a real scene to obtain measured data. Based on the spatial uncertainty distribution, it identifies the top K regions with the highest statistical variance values ​​and prioritizes measurement of these regions. Here, K is determined based on the statistical characteristics of the spatial uncertainty distribution or the physical constraints of the measurement system. Alternatively, based on the spatial uncertainty distribution, it identifies regions with statistical variance values ​​higher than a first threshold and automatically densifies the spatial density of sampling points within these regions. Here, the first threshold is determined based on the statistical characteristics of the spatial uncertainty distribution. The predicted electromagnetic absorption power density distribution output by the model used to evaluate electromagnetic absorption power density corresponds to each spatial cell, and the prediction results include the average specific absorption rate of each spatial cell and the corresponding spatial uncertainty distribution. The spatial uncertainty distribution of the prediction results is obtained as follows: during the inference process of the model used to evaluate electromagnetic absorption power density, its built-in stochastic unit is enabled, and the same set of input parameters is forward-propagated multiple times to obtain multiple prediction results. The statistical variance of the multiple prediction results on each spatial cell is calculated, and the spatial uncertainty distribution of the prediction results is generated based on the correspondence between each spatial cell and the statistical variance. The data fusion and decision subsystem is used to calculate the difference between the measured specific absorbance distribution and the model-predicted specific absorbance distribution as a loss function. When the loss function is greater than a second threshold, the measured specific absorbance distribution is used as a supervision signal to optimize the model parameters. The second threshold is determined based on the spatial uncertainty distribution and / or the uncertainty of the measured data and / or the historical calibration data. Multiple physical field parameters are then input into the optimized model for re-evaluation.

6. An electronic device, characterized in that, include: A processor, a memory, and a computer program stored in the memory, wherein the processor, when executing the computer program, implements the method for evaluating electromagnetic absorption power density as described in any one of claims 1 to 4.

Images