A smart electric energy meter integrated antenna design method

Abstract

This invention relates to the field of smart energy meter antenna design technology, and discloses a method for designing an integrated antenna for a smart energy meter. The method includes: collecting structural parameters of a prototype energy meter; constructing an internal electromagnetic field and installation environment antenna performance dataset through triaxial optical electric field sensing and phase-locked loop amplification measurement; obtaining an electromagnetic field distribution prediction model through encoder-decoder convolutional neural network and structural similarity adaptive transfer training; establishing an antenna performance surrogate model with prediction uncertainties through equivalent magnetic dipole simulation and Gaussian process environment correction; generating multiple sets of candidate antenna parameter schemes by combining hierarchical mixed variable multi-objective optimization and dynamic safety margin constraint iteration; and finally determining the final design scheme through multi-environment robustness scoring and physical prototype verification. This method can accurately characterize the internal electromagnetic environment of the energy meter and achieve rapid prediction and robust optimization design of antenna performance under multiple installation environments.

Description

Technical Field

[0001] This invention relates to the field of smart energy meter antenna design technology, and more specifically, to a method for designing an integrated antenna for a smart energy meter. Background Technology

[0002] Smart meters are core terminal devices in the power Internet of Things (IoT), and the integration of wireless communication modules such as NB-IoT places high demands on antenna design. The internal space of a smart meter is limited, the metal casing provides shielding for antenna radiation, and wideband harmonic interference from the switching power supply can interfere with signals in the antenna's operating frequency band. Furthermore, smart meters face various installation environments in practice, including metal distribution boxes, concrete walls, and densely packed installations, each with significant impacts on antenna performance.

[0003] Current antenna design methods mainly rely on empirical trial and error and iterative full-wave electromagnetic simulation. Designers initially determine the antenna structure based on experience and adjust parameters through repeated simulations until the performance requirements are met. This method has the following shortcomings: the simulation model usually does not include the actual electromagnetic environment inside the energy meter, resulting in a large deviation between the simulation results and the actual performance; each simulation takes several hours, and multiple iterations lead to a design cycle of several months for a single model, making it difficult to meet the engineering requirements of rapid iteration for multiple models.

[0004] Furthermore, existing methods lack systematic modeling of the impact of the installation environment. The designed antennas are often optimized only for specific environments, and their performance degrades significantly in other installation environments, failing to guarantee performance consistency in mass deployment scenarios. Therefore, existing technologies suffer from technical problems such as low antenna design efficiency, inaccurate characterization of the internal electromagnetic environment, and insufficient performance robustness in multiple installation environments. Summary of the Invention

[0005] This invention provides a design method for an integrated antenna in a smart energy meter, which solves the technical problems of low antenna design efficiency, inaccurate characterization of the internal electromagnetic environment, and insufficient performance robustness in multiple installation environments in related technologies.

[0006] This invention provides a design method for an integrated antenna in a smart energy meter, comprising the following steps:

[0007] S1: Collect structural characteristic parameters of the smart energy meter prototype, and measure them using triaxial optical electric field sensing and frequency selective phase-locked amplification to form an internal electromagnetic field distribution dataset and an installation environment antenna performance dataset.

[0008] S2, based on the electromagnetic field distribution dataset and structural feature parameters, through encoder-decoder convolutional neural network training and structural similarity adaptive transfer, outputs the electromagnetic field distribution prediction model and the electromagnetic field distribution prediction result of the current model;

[0009] S3, combining the electromagnetic field distribution prediction results with the antenna performance dataset of the installation environment, and through the equivalent magnetic dipole simulation modeling driven surrogate model training and Gaussian process physical parameterization environment correction, forms an antenna performance surrogate model containing comprehensive prediction uncertainty.

[0010] S4 utilizes the antenna performance proxy model and comprehensive prediction uncertainty, with the statistical characteristics of the electromagnetic environment as a fixed background, and generates multiple sets of candidate antenna parameter schemes through hierarchical mixed variable multi-objective optimization and dynamic safety margin constraint iteration.

[0011] S5. For candidate antenna parameter schemes, the final antenna design scheme and performance report are determined by screening through multi-environment robustness scoring of the environmental correction surrogate model and verification by physical sample preparation and testing. If the measured deviation exceeds the threshold, the source of the error is traced and the corresponding model closed-loop update is triggered.

[0012] In a preferred embodiment, a non-contact optical electric field sensing system based on lithium niobate electro-optic crystal is used to measure the electromagnetic field inside the smart energy meter. The probe is encapsulated with three mutually orthogonally arranged lithium niobate crystals, which correspond to the electric field components in the X, Y, and Z directions, respectively. The three modulated optical signals are synchronously received by three photodetectors and synthesized to obtain the electric field intensity vector at each measurement point.

[0013] Using the periodic transmission signal of the NB-IoT communication module as a reference clock, the lock-in amplifier only integrates and extracts the signal component that is in phase with the reference clock to suppress low-frequency harmonic interference generated by the switching power supply. After completing the three-dimensional scanning measurement, the mean and standard deviation of the field strength amplitude of all grid points in the antenna installation area are calculated. Grid points with field strength amplitude exceeding the mean plus one standard deviation and their surrounding preset range are marked as interference-sensitive areas.

[0014] In a preferred embodiment, the process of forming the installation environment antenna performance dataset includes: selecting four typical installation environments: metal distribution box environment, concrete wall mounting environment, densely arranged environment, and open environment; fabricating a reference antenna and a multi-geometric parameter antenna sample selected according to the Latin hypercube sampling method for each environment; measuring the reflection coefficient and extracting the resonant frequency and VSWR using a vector network analyzer; measuring the radiation pattern and gain using a far-field antenna test system; and measuring the radiation efficiency simultaneously.

[0015] The physical characteristics of each environment are quantitatively characterized by two parameters: metal coverage and equivalent dielectric loss factor. The environmental physical characteristic parameters are associated with the corresponding antenna performance data and stored to form an environmental characteristic-performance mapping dataset.

[0016] In a preferred embodiment, the encoder part of the encoder-decoder convolutional neural network receives structural feature vectors and PCB layout feature maps as inputs. The structural feature vectors are mapped to the latent space through fully connected layers, and the PCB layout feature maps extract spatial features through two two-dimensional convolutional layers. The two are flattened and spliced ​​together and mapped to a high-level semantic feature vector.

[0017] The decoder maps and reshapes the feature vector into a three-dimensional tensor through a fully connected layer. After being progressively upsampled by three three-dimensional deconvolutional layers, the final output is an electromagnetic field distribution prediction tensor with a size of D×H×W×6, where the six channels correspond to the electric field components and magnetic field components in the X, Y, and Z directions, respectively.

[0018] In a preferred embodiment, the loss function of the encoder-decoder convolutional neural network is a weighted composite loss, which consists of an amplitude error term and a polarization direction error term. The amplitude error term adopts a weighted mean square error, and the weights are assigned in three levels according to the region type label output by S1. The non-interference sensitive grid points in the antenna installation area have the highest weight, the grid points in the interference sensitive area have the second highest weight, and the grid points in the other areas have the lowest weight.

[0019] The polarization direction error term calculates the direction cosine of the predicted electric field unit vector and the measured electric field unit vector for each grid point. The direction error is calculated by subtracting the direction cosine from 1. The direction error weight is the highest for grid points in the interference-sensitive area. The total loss function is the sum of the amplitude error term and the polarization direction error term weighted by the balance coefficient.

[0020] In a preferred embodiment, the structural similarity adaptive migration process includes: calculating structural similarity scores for the new model and all existing models respectively. The structural similarity scores are obtained by weighted summation of three sub-items: shell size similarity, PCB layout similarity, and power supply similarity. The base model corresponding to the existing model with the highest score is selected as the migration source.

[0021] When the structural similarity score is higher than the first threshold, all encoder parameters are frozen, and only the last two layers of the decoder are fine-tuned; when the structural similarity score is between the second and first thresholds, the first two layers of the encoder are frozen, and all subsequent layers of the encoder and the decoder are fine-tuned; when the structural similarity score is lower than the second threshold, the model is trained from scratch with the basic model parameters as the initial values.

[0022] In a preferred embodiment, the surrogate model training process driven by the equivalent magnetic dipole simulation modeling includes: performing principal component analysis on the magnetic field distribution of the interference-sensitive region in the electromagnetic field distribution predicted by S2, extracting the dominant magnetic field direction and the corresponding field strength amplitude, and equivalently converting the corresponding region into a magnetic dipole and loading it into the full-wave electromagnetic simulation model in the form of a lumped magnetic dipole.

[0023] The average field strength, field strength variance, and dominant polarization direction of the antenna installation area are extracted from the electromagnetic field distribution predicted by S2 as statistical features. These statistical features are then concatenated with the antenna geometric parameters as the input feature vector of the surrogate model. A fully connected multilayer perceptron is constructed as the antenna performance surrogate model, and the output layer corresponds to the resonant frequency, VSWR, gain, and radiation efficiency.

[0024] In a preferred embodiment, the Gaussian process physical parameterization environment correction process includes: using the spliced ​​vector of four physical parameters—antenna electrical size, relative feed point position, metal coverage, and equivalent dielectric loss factor—as the input to the environment correction model, and using the deviation between the simulation prediction value and the measured value as the dependent variable, and establishing a deviation prediction model using Gaussian process regression with radial basis function kernels.

[0025] The formation process of the comprehensive prediction uncertainty includes: applying a random perturbation that conforms to its prediction error distribution to the electromagnetic field distribution tensor output by S2 and repeatedly sampling it; inputting the electromagnetic field statistical characteristics after each perturbation into the surrogate model to obtain multiple sets of performance prediction results; calculating its standard deviation as the error propagation uncertainty; and merging the error propagation uncertainty with the environmental correction uncertainty output by the Gaussian process correction model to obtain the comprehensive prediction uncertainty.

[0026] In a preferred embodiment, the hierarchical hybrid variable multi-objective optimization process includes: maintaining an independent subpopulation for each antenna type; using simulated binary crossover and polynomial mutation operators to evolve continuous parameters within the subpopulation; performing a unified non-dominated sorting and crowding calculation across all subpopulations after each iteration; and setting a minimum retention number for each antenna type to prevent it from being completely eliminated.

[0027] The dynamic safety margin constraint iteration process includes: for each candidate solution, tightening the constraint threshold by one standard deviation of the prediction uncertainty of the corresponding performance index of the candidate solution based on the nominal value, and setting a preset upper limit for the tightening amount.

[0028] In a preferred embodiment, in step S5, the process of screening the multi-environment robustness score of the environmental correction surrogate model includes: using the environmental correction surrogate model to predict the performance index of each candidate scheme in four typical installation environments, calculating the minimum value and standard deviation of the gain in the four environments, weighted summing the normalized minimum gain score and the normalized stability score to obtain the robustness score, and selecting the scheme with the highest robustness score as the recommended scheme; the closed-loop update process includes:

[0029] If the measured deviation exceeds the preset threshold, the magnetic field distribution of the interference-sensitive area in the measured electromagnetic field data of S1 is re-analyzed as principal component and loaded into the full-wave simulation model in the form of a lumped magnetic dipole. The simulated value is compared with the measured value to determine whether the deviation comes from the electromagnetic field prediction error of S2 or the environmental correction error of S3, and the corresponding model update is triggered.

[0030] The beneficial effects of this invention are as follows:

[0031] The actual electromagnetic field distribution inside the energy meter is obtained by triaxial optical electric field sensing and frequency selective phase-locked amplification measurement. Combined with encoder-decoder convolutional neural network and structural similarity adaptive transfer, rapid prediction from structural feature parameters to internal electromagnetic field distribution is achieved. Then, the training of surrogate model is driven by equivalent magnetic dipole simulation modeling, and the internal electromagnetic environment information is introduced into antenna performance prediction. This solves the problems of existing methods such as the disconnect between simulation model and actual electromagnetic environment and insufficient prediction accuracy, and shortens the antenna design cycle from several months to several days.

[0032] By establishing a deviation prediction model between simulation and actual measurement through Gaussian process physical parameterization environment correction, and combining dynamic safety margin constraints driven by comprehensive prediction uncertainty, the candidate schemes are robustly scored and screened in multiple installation environments. The model is continuously corrected through measured deviation tracing and closed-loop update mechanism, which solves the problems of insufficient modeling of installation environment and poor performance robustness in multiple environments in existing methods, and ensures the performance stability of antennas in batch deployment scenarios. Attached Figure Description

[0033] Figure 1 This is a flowchart of a smart energy meter integrated antenna design method according to the present invention;

[0034] Figure 2 This is a flowchart of a smart energy meter integrated antenna design method according to the present invention. Figure 1 ;

[0035] Figure 3 This is a flowchart of a smart energy meter integrated antenna design method according to the present invention. Figure 2 . Detailed Implementation

[0036] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0037] At least one embodiment of the present invention discloses a design method for an integrated antenna in a smart energy meter, such as... Figures 1 to 3 As shown, it includes the following steps:

[0038] S1: Collect structural characteristic parameters of the smart energy meter prototype, and measure them using triaxial optical electric field sensing and frequency selective phase-locked amplification to form an internal electromagnetic field distribution dataset and an installation environment antenna performance dataset.

[0039] Obtain a prototype smart energy meter with the antenna to be designed. The prototype is fully assembled and includes complete metering circuitry, a power module, a communication module, and a metal casing. Extract the structural parameters of the energy meter, including the length, width, and height dimensions of the casing; the conductivity and permeability of the casing material; the number of layers, dimensions, and layout of the PCB; the types, package sizes, and coordinate positions of the main components on the PCB; the distribution area and thickness of the metal shielding layer; and the location and available space for the reserved antenna mounting area. Place the energy meter prototype in an electromagnetic compatibility anechoic chamber. The chamber is lined with absorbing materials on the walls and top, and has a metal floor to ensure an environment free from external electromagnetic interference.

[0040] The electricity meter casing is a closed metal structure, and the probe cannot be directly inserted inside. Therefore, the prototype needs to be specially modified before measurement. Specifically, on the side wall corresponding to the antenna mounting area of ​​the electricity meter casing, a precision CNC milling machine is used to machine measurement through holes according to the planned positions of the three-dimensional measurement grid. The diameter of the through holes is 3.5mm, matching the outer diameter of the probe's quartz tube. The spacing between adjacent through holes is consistent with the grid spacing, with a default value of 2mm. The inner wall of the through holes is coated with conductive silver paste to maintain the continuity of the casing's electromagnetic shielding. Each through hole is equipped with a conductive rubber sealing ring. When the probe is inserted, the sealing ring fills the gap between the probe and the inner wall of the through hole, maintaining the electromagnetic shielding effect of the casing. When the probe is not inserted, the through holes maintain electrical continuity through the conductive silver paste, ensuring uninterrupted shielding. After measurement, the through holes are sealed with metal plugs of the same material as the casing and welded to restore the integrity of the casing. The function of the three-dimensional displacement platform is to drive the probe to align with the axis of each through hole in sequence and control the insertion depth outside the shell. Each through hole corresponds to a fixed grid measurement point. After the probe is inserted into the through hole axis to the predetermined depth, the measurement of that point is completed. After exiting, it moves to the next through hole and completes the measurement of all grid points one by one.

[0041] A non-contact optical-electric field sensing system based on an electro-optic crystal is used to measure the electromagnetic field inside an energy meter. The core of the optical sensing system is a lithium niobate electro-optic crystal. Under the influence of an applied electric field, the refractive index of the crystal changes, and this change is proportional to the electric field strength. Linearly polarized laser light is incident on the electro-optic crystal after passing through a polarizer. After passing through the crystal, the polarization state of the laser light is changed by the electric field modulation. The polarization state change is then converted into a light intensity change by an analyzer, which is received by a photodetector and converted into an electrical signal. The electric field strength is calculated using a calibration curve. The advantage of this method is that the sensing probe is a non-metallic optical crystal, and its insertion into the measured space does not introduce disturbances from metal conductors, resulting in measurement results that more closely approximate the actual electromagnetic field distribution. To simultaneously acquire the electric field components in the three orthogonal directions (X, Y, and Z) in a single insertion, three orthogonally arranged lithium niobate crystals are encapsulated inside the probe. Each crystal corresponds to an independent optical fiber. The three modulated optical signals are received by three photodetectors, which simultaneously output the electric field components in the three directions, synthesizing them to obtain the electric field strength vector at the measurement point. The triaxial integrated crystal is encapsulated in a quartz tube with a diameter of 5 mm. The end of the quartz tube is connected to an optical fiber bundle. The effective size of the entire probe is less than 6 mm, and it can be inserted into the electricity meter through the measurement through hole for measurement.

[0042] The electricity meter contains both low-frequency harmonic interference (typically ranging from 50kHz to 5MHz, with field strength amplitudes reaching tens of V / m) generated by the switching power supply and high-frequency signals (800MHz / 1800MHz, with induced field strength in the antenna installation area inside the electricity meter typically only on the order of several V / m) generated by the NB-IoT communication module. The strong low-frequency interference will overwhelm the weak high-frequency signal during wide-band measurements, resulting in distorted measurement results for the target frequency band.

[0043] In one embodiment of the present invention, the periodic transmission signal of the NB-IoT communication module is used as a reference clock. The lock-in amplifier only integrates and extracts the signal component that is in phase and frequency with the reference clock, effectively suppressing interference signals at other frequencies, and improving the signal-to-noise ratio by at least 40dB. For the 800MHz band, the default value of the integration time constant of the lock-in amplifier is 100ms; for the 1800MHz band, the default value of the integration time constant is 50ms. This scheme can achieve an electric field measurement resolution better than 0.5V / m in the target frequency band under normal operating conditions of the electricity meter, meeting the data quality requirements for subsequent model training.

[0044] The refractive index of lithium niobate crystals is temperature-sensitive; for every 1°C change in temperature, the refractive index changes by approximately 1 × 10⁻⁻⁻⁶. 4 In the long-term scanning measurement process, the internal temperature of the energy meter increases with the working time (the internal temperature rise is usually 10°C to 20°C due to continuous operation of the switching power supply), which will cause the measurement baseline to drift.

[0045] In one embodiment of the present invention, a miniature platinum resistance temperature sensor is integrated on the outer wall of the quartz tube of the optical probe. The temperature at the probe location is collected in real time throughout the three-dimensional scanning measurement process, and the electric field readings at each measurement point are compensated in real time according to the pre-calibrated refractive index-temperature curve of lithium niobate crystal. The measurement error after temperature compensation does not exceed 2%.

[0046] Before conducting electromagnetic field measurements, the electricity meter is powered and put into standard operating condition: the metering circuit is supplied with rated load current, and the NB-IoT communication module periodically transmits data at standard transmission power (default value is 23dBm) to simulate actual working scenarios. Measurements of all electricity meter models are performed under this unified standard operating condition to ensure consistency of training data. A three-dimensional measurement grid is established within the predetermined antenna installation area of ​​the electricity meter and its surrounding 30mm radius, with grid spacing of 2mm in the X, Y, and Z directions. A three-dimensional displacement platform drives the probe to align with each through-hole sequentially and insert it to the predetermined depth. Synchronous measurement of the three-axis electric field components is performed at each grid point. Simultaneously, a magnetic field probe is used to measure the magnetic field strength vector. The magnetic field probe employs a miniature coil structure; the electromotive force induced by the coil is proportional to the rate of change of the magnetic field, and the magnetic field strength is obtained through an integrator circuit. The measurement frequency covers the antenna's operating frequency band. For the 800MHz band, the measurement frequency range is 700MHz to 900MHz; for the 1800MHz band, the measurement frequency range is 1700MHz to 1900MHz, with a frequency step of 10MHz. At each frequency point, record the electric field intensity vector, magnetic field intensity vector, and phase information of all grid points.

[0047] After completing the 3D scanning measurement, the field strength distribution of the measurement data is analyzed to identify interference-sensitive locations within the antenna installation area. The specific determination method is as follows: calculate the mean and standard deviation of the field strength amplitude at all grid points within the antenna installation area. Grid points with field strength amplitudes exceeding the mean plus one standard deviation are identified as abnormally high field strength points, and these points and their surrounding 10mm radius are marked as interference-sensitive areas. If the field strength amplitude at all grid points within the antenna installation area does not exceed the aforementioned threshold, interference-sensitive areas are not marked, and the antenna position is not subject to additional constraints. Interference-sensitive areas correspond to the near-field region of strong radiation sources inside the electricity meter (usually switching power supply transformers or high-frequency filter inductors). Strong internal interference in this region can significantly affect antenna radiation performance. In the subsequent S4 optimization step, the positions of the antenna feed point and radiating arm are constrained to avoid interference-sensitive areas, thereby actively mitigating the impact of internal interference on antenna performance during the optimization phase. The collected electromagnetic field data is organized into a spatial distribution matrix according to 3D coordinates. Each element of the matrix corresponds to a grid point, containing the point's coordinates, electric field vector, magnetic field vector, phase, and frequency information, along with an additional interference-sensitive area marker.

[0048] After completing the electromagnetic field measurement inside the electricity meter, benchmark data on antenna performance under typical installation environments were further obtained. Four of the most common installation environments in actual power company deployments were selected: metal distribution box environment, concrete wall mounting environment, densely packed environment, and open environment. The metal distribution box environment simulates the electricity meter installed inside a metal-cased distribution box with dimensions of 400mm × 300mm × 200mm. The box material is iron plate with a thickness of 1.5mm. The concrete wall mounting environment simulates the electricity meter installed on a 200mm thick concrete wall of a building with a relative permittivity of 6 and a loss tangent of 0.1. The back of the electricity meter is attached to the wall surface. The densely packed environment simulates multiple electricity meters closely arranged in a centralized meter box, with three meters arranged horizontally and adjacent to each other with a spacing of 50mm. The open environment is free space, with no obstructions within 1 meter of the electricity meter.

[0049] For each environment, two types of antenna samples were fabricated and installed on a prototype energy meter for measurement. The first type served as a reference antenna, using a conventional monopole antenna or inverted-F antenna to establish a performance benchmark for each environment. The second type consisted of multi-geometric parameter samples covering the parameter space. Sample points were selected within the antenna geometric parameter space using the Latin hypercube sampling method. At least 30 antenna samples with different geometric parameters were fabricated for each installation environment, totaling at least 120 samples across the four environments. The measured data from these samples will be used in the subsequent S3 step to train the Gaussian process environment correction model. The setup of the four typical installation environments was completed in one go during the S1 stage. Both types of samples underwent all measurements within the same environment setup, avoiding errors and time costs associated with repeated environment setups. The prototype was placed in each of the four environments, and the reflection coefficient S11 parameter at the antenna port was measured using a vector network analyzer. The measurement frequency range covered the operating frequency band, obtaining a curve showing the reflection coefficient as a function of frequency. The resonant frequency and VSWR of the antenna were extracted from the curve. A far-field antenna test system was used to measure the radiation pattern and gain of the antenna. The test system included a turntable and a receiving antenna. The energy meter prototype was fixed on the turntable, which was rotated 360 degrees. The receiving antenna recorded the radiation intensity at different angles, and a three-dimensional radiation pattern was synthesized. The maximum gain and half-power beamwidth were extracted from the radiation pattern. Simultaneously, the antenna's radiation efficiency was measured. Radiation efficiency is equal to the ratio of radiated power to input power, measured using the Wheeler cap method or the Reverberation Chamber method. Each antenna sample under each environment was measured three times, and the average value was taken as the antenna performance data for that environment.

[0050] To support the construction of the environment correction model in the subsequent S3 step, antenna performance data for four typical installation environments were recorded, and the physical characteristics of each environment were quantitatively characterized. The environmental physical characteristics are described by two measurable parameters: metal coverage and equivalent dielectric loss factor. Metal coverage is defined as the ratio of the projected area of ​​the metal surface to the cross-sectional area of ​​the hemisphere within a 200mm radius centered on the antenna, ranging from 0 to 1. The metal coverage in open environments is close to 0, while in metal distribution box environments it is close to 1. The equivalent dielectric loss factor is defined as the product of the relative permittivity of the main dielectric material surrounding the antenna and the loss tangent. The equivalent dielectric loss factor in open environments is close to 0, while in concrete wall-mounted environments it is the product of the relative permittivity 6 and the loss tangent 0.1, i.e., 0.6. The metal coverage and equivalent dielectric loss factor for the four typical environments are associated and stored with the corresponding antenna performance data to form an environment feature-performance mapping dataset, which is used for training the environment correction model in the subsequent steps.

[0051] The electromagnetic field distribution dataset is associated with the structural parameters of the electricity meter. The dataset is organized as follows: each electricity meter sample contains a structural parameter file and an electromagnetic field data file. The structural parameter file records information such as casing dimensions, PCB layout, and component positions in XML format. The electromagnetic field data file stores the electromagnetic field vector data of the three-dimensional spatial mesh and region type labels in HDF5 format. The region type labels are divided into three categories: non-interference sensitive mesh points within the antenna installation area, interference sensitive mesh points, and other region mesh points. These three types of labels are used for weight allocation in the S2 loss function. Simultaneously, the installation environment performance dataset (including reference antenna data and multi-geometric parameter sample data) is associated with the corresponding environmental physical characteristic parameters for easy retrieval in subsequent steps. Through the above process, an internal electromagnetic field distribution dataset containing multiple electricity meter samples of different models and a measured dataset of antenna performance under typical installation environments are obtained. These real measurement data form the basis for subsequent prediction model construction and environmental correction.

[0052] S2, based on the electromagnetic field distribution dataset and structural feature parameters, through encoder-decoder convolutional neural network training and structural similarity adaptive transfer, outputs the electromagnetic field distribution prediction model and the electromagnetic field distribution prediction result of the current model;

[0053] The system receives the data set of internal electromagnetic field distribution and structural feature parameters of the electricity meter from the first step, and uses deep convolutional neural networks and adaptive transfer learning techniques to obtain a prediction model that can quickly predict the internal electromagnetic field distribution based on the structural features of the electricity meter.

[0054] Key features characterizing the electromagnetic environment are extracted from the structural parameters of the electricity meter. PCB layout features include the number of PCB layers, total area, copper foil area ratio of each layer, and length and routing of key signal lines; component distribution features include the number, location density, and minimum spacing between high-frequency components; metal shielding features include the coverage area ratio of the metal shielding layer and the distance between the shielding layer and the antenna mounting area; power supply features include the operating frequency, power rating, and filter circuit configuration of the switching power supply. These discrete structural features are converted into numerical feature vectors. One-hot encoding is used for discrete categorical features, and normalization is performed for continuous numerical features. For spatial information such as PCB layout, the PCB plane is divided into a grid, and the component types and densities within each grid are statistically analyzed to form a two-dimensional feature map. Finally, all features are combined to form a structural feature vector of length N and a PCB layout feature map of size H×W. The default value for N is 128, and the default value for H×W is 32×32.

[0055] The electromagnetic field spatial distribution data obtained in the first step is reconstructed into a three-dimensional tensor form suitable for convolutional neural network processing. The original electromagnetic field data consists of electric and magnetic field vectors at each point in a three-dimensional spatial grid. The electric field components in the X, Y, and Z directions, as well as the magnetic field components in the X, Y, and Z directions, are each treated as six channels, forming a four-dimensional tensor of size D×H×W×6, where D, H, and W represent the number of grid points in the three directions, respectively. To reduce computational load, the original dense grid with a 2mm spacing is downsampled, and the grid spacing is adjusted to 5mm using trilinear interpolation. The typical tensor size after downsampling is 10×10×6×6.

[0056] A convolutional neural network model based on an encoder-decoder architecture is constructed to learn the mapping relationship from the structural features of an electricity meter to the internal electromagnetic field distribution. The encoder receives structural feature vectors and PCB layout feature maps as input. The structural feature vectors are first mapped to a 128-dimensional latent space through a fully connected layer. The PCB layout feature maps are processed through two 2D convolutional layers to extract spatial features. The convolutional kernel size is 3×3. The first layer outputs 32 feature maps, and the second layer outputs 64 feature maps. Each convolutional layer is followed by a ReLU activation function and a max-pooling layer. The vectors output from the fully connected layers and the feature maps output from the convolutional layers are flattened and concatenated, then mapped to a 256-dimensional high-level semantic feature vector through another fully connected layer. The decoder takes a 256-dimensional feature vector as input, first mapping it to a D×H×W-dimensional flattened vector through a fully connected layer, then reshaping it into a D×H×W×1 three-dimensional tensor. Next, it progressively upsamples and increases the number of channels through three three-dimensional deconvolutional layers: the first layer outputs 32 channels, the second 16, and the third 6, corresponding to the six components of the electromagnetic field. The three-dimensional deconvolutional layers use a 3×3×3 kernel with a stride of 1 and use the same padding method to maintain spatial dimensions. Each deconvolution layer is followed by a ReLU activation function, while the last deconvolution layer does not use an activation function and directly outputs the predicted electromagnetic field distribution tensor with dimensions D×H×W×6.

[0057] In one embodiment of the present invention, the loss function is defined as a weighted composite loss between the predicted electromagnetic field and the measured electromagnetic field, consisting of two parts: an amplitude error term and a polarization direction error term. The amplitude error term uses a weighted mean square error, and the measured electromagnetic field tensor is denoted as... The predicted electromagnetic field tensor is denoted as The weight tensor is denoted as The amplitude error term is:

[0058]

[0059] in, For spatial grid coordinate indexing, This is the index for the field component channels (6 channels in total). The total number of elements in the tensor. For position The amplitude error weighting coefficient at the location is assigned in three levels according to the area type label of S1 output: non-interference sensitive grid points within the antenna installation area. The default value is 3, which is the number of grid points in the interference sensitive area. The default value is 2, and the remaining grid points... The default value is 1. The non-interference-sensitive location within the antenna installation area is the area where the antenna may actually be installed. The prediction accuracy of this area directly affects the training quality of the S3 surrogate model, so it is given the highest weight. The prediction accuracy of the interference-sensitive area affects the strength estimation of the equivalent magnetic dipole, so it is given the second highest weight.

[0060] The polarization direction error term is used to constrain the directional consistency between the predicted and measured electric field vectors. For each grid point, the dot product (i.e., direction cosine) of the predicted and measured electric field unit vectors is calculated. The closer the direction cosine is to 1, the more consistent the directions. Subtracting the direction cosine from 1 gives the direction error at that point. The polarization direction error term is:

[0061]

[0062] in, and Positions The predicted and measured electric field vectors (composed of three components: X, Y, and Z). The directional error weighting coefficient is used for grid points in the interference-sensitive area. The default value is 3 (because the polarization direction of this region directly determines the direction of the equivalent magnetic dipole), and the non-interference sensitive grid points within the antenna installation area. The default value is 2, and the default value for other areas is 1. The total loss function is:

[0063]

[0064] in, These are balancing coefficients used to make the magnitude error term and the direction error term equivalent in numerical order. The default value is set to the square of the mean field strength amplitude of all samples in the training set, estimated with a small number of samples before training begins. The Adam optimizer is used to train the network, with an initial learning rate of 0.001, which decays to 0.9 times after several training rounds. The training dataset contains samples of different models of electricity meters, divided into training and validation sets in an 8:2 ratio. The default batch size is 16, and the default total number of training rounds is 200. Training is stopped early to prevent overfitting when the validation set loss does not decrease for 10 consecutive rounds.

[0065] To improve the model's generalization ability, data augmentation techniques were employed to expand the training samples. Small-amplitude Gaussian noise was applied to the structural feature vectors to simulate measurement errors; random rotation and flipping of the PCB layout feature map were performed to simulate different layout orientations; and spatial translation and scaling transformations were applied to the electromagnetic field distribution tensor to simulate subtle differences in antenna installation positions. Through data augmentation, the number of training samples was increased fivefold, effectively alleviating the deep learning model's dependence on large amounts of training data. After model training, the mean square error, mean absolute error, and correlation coefficient between the predicted and measured electromagnetic fields were calculated on the test set. Prediction errors were separately calculated for interference-sensitive areas to ensure that the prediction accuracy in key areas met requirements. Evaluation results showed that the average relative error between the model's predicted and measured electromagnetic fields was less than 10%, and the correlation coefficient was greater than 0.92, meeting the accuracy requirements for subsequent antenna design.

[0066] The structural differences between different models of energy meters manifest in three dimensions: differences in casing size affect electromagnetic boundary conditions, differences in the number and layout of PCB layers affect near-field distribution patterns, and differences in the power rating of the switching power supply affect internal interference intensity. The varying degrees of these differences determine the feasibility and strategy of transfer learning. If a uniform transfer strategy (such as fixing and freezing the encoder) is directly applied to models with significant structural differences, negative transfer may occur due to the large differences in electromagnetic field distribution patterns between the source and target domains, resulting in the fine-tuned model having lower prediction accuracy than one trained from scratch.

[0067] In one embodiment of the present invention, structural similarity scores are calculated for the new model and all existing models respectively, and the base model corresponding to the existing model with the highest score is selected as the transfer source; if the highest score is lower than 0.5, no transfer is performed, and the model is trained from scratch with the base model parameters as the initial values. The structural similarity score is obtained by weighted summation of three sub-items, namely, shell size similarity (the ratio of the shell volumes of the two models, taking the smaller value and dividing by the larger value), PCB layout similarity (the cosine similarity of the PCB layout feature maps of the two models), and power supply similarity (the ratio of the power levels of the switching power supplies of the two models, taking the smaller value and dividing by the larger value). The weights of the three sub-items are 0.2, 0.6, and 0.2, respectively, with PCB layout similarity having the highest weight because PCB layout has the most direct impact on near-field electromagnetic distribution. When the similarity score is higher than 0.8, all encoder parameters are frozen, and only the last two layers of the decoder are fine-tuned. The default number of new model samples required is 5 to 10. When the similarity score is between 0.5 and 0.8, the first two layers of the encoder are frozen, and all subsequent layers of the encoder and the decoder are fine-tuned. The default number of samples required is 15 to 20. When the similarity score is lower than 0.5, the default number of samples required is more than 30. The learning rate during fine-tuning is set to 0.1 times that of the base training, and the default number of training epochs is 20.

[0068] Through the above process, an electromagnetic environment prediction model was constructed. The input of the model is the structural feature parameters of the electricity meter, and the output is the internal three-dimensional electromagnetic field distribution. The prediction time is in the millisecond range, and it supports rapid adaptation to new models through adaptive transfer learning, providing accurate and efficient electromagnetic environment information for subsequent steps.

[0069] S3, combining the electromagnetic field distribution prediction results with the antenna performance dataset of the installation environment, and through the equivalent magnetic dipole simulation modeling driven surrogate model training and Gaussian process physical parameterization environment correction, forms an antenna performance surrogate model containing comprehensive prediction uncertainty.

[0070] The electromagnetic field distribution and antenna geometric parameters predicted by the electromagnetic environment prediction model output from the second step are received. A deep neural network is used to construct a proxy model for antenna performance. The measured data of the installation environment and the quantified environmental characteristics obtained in the first step are combined to perform physical parameterization environment correction, so as to obtain a proxy model that can quickly and accurately predict the actual performance of the antenna in different electromagnetic environments and installation environments.

[0071] Generate a dataset for training the surrogate model. Antenna geometry parameters are determined based on antenna type: for monopole antennas, the parameters are the length and diameter of the antenna arm; for inverted-F antennas, the parameters are the horizontal arm length, vertical arm height, distance from the feed point to the short-circuit point, and ground plane dimensions; for microstrip antennas, the parameters are the patch length, width, feed point location, substrate dielectric constant, and thickness. Based on the antenna type and the spatial constraints of the energy meter, determine the value range of each geometry parameter. Simultaneously, use the interference-sensitive area identified in step S1 as a positional constraint to exclude parameter combinations where the antenna feed point or radiating arm falls within the interference-sensitive area. A Latin hypercube sampling method is used to select representative sample points in the parameter space. For a space containing 10 parameters, the default number of sample points is 500.

[0072] For each sampled antenna parameter combination, a complete electromagnetic simulation model including the antenna and the internal electromagnetic environment of the energy meter is established by combining the electromagnetic field distribution predicted in the second step. The electromagnetic field predicted by S2 is used in the simulation modeling in two ways: First, principal component analysis is performed on the magnetic field distribution of the interference-sensitive area to extract the dominant magnetic field direction (i.e., the direction with the largest magnetic field variance) and the corresponding field strength amplitude of the area. The area is then equivalent to a magnetic dipole, with the direction of the equivalent magnetic dipole moment being the dominant magnetic field direction. The magnitude of the equivalent magnetic dipole moment is equal to the product of the magnetic field strength amplitude in that direction and the equivalent area of ​​the interference-sensitive area. This is then loaded into the corresponding PCB position in the simulation model in the form of a lumped magnetic dipole, so that the simulation model can reflect the impact of internal interference sources on antenna performance when the energy meter is actually working. Both HFSS and CST support importing equivalent dipole excitation in the form of lumped elements. Second, statistical features of the antenna installation area are extracted from the electromagnetic field distribution predicted by S2, including the average field strength, field strength variance, and dominant polarization direction. These three statistical features are then concatenated with the antenna geometric parameters as the input feature vector of the surrogate model. Simulations were performed using full-wave electromagnetic simulation software such as HFSS or CST, covering the antenna's operating frequency band. The solver was set to adaptive mesh generation, with a convergence accuracy requirement of S-parameter variation less than 0.01. The simulation outputs antenna performance metrics such as reflection coefficient curve, input impedance, gain, radiation efficiency, and radiation pattern. The resonant frequency and minimum VSWR were extracted from the reflection coefficient curve, and the maximum gain and its corresponding direction, as well as the half-power beamwidth, were extracted from the radiation pattern. The antenna geometry and electromagnetic environment statistical characteristics of each sample were used as input features, and the performance metrics were used as output labels to form a training dataset. The simulation time for a single sample is approximately 3 to 5 hours; for 500 samples, parallel computing (using a 20-node computing cluster with dual CPUs per node) can reduce the time to approximately 60 hours.

[0073] A deep neural network is constructed as a proxy model for antenna performance. The network input layer receives antenna geometric parameters and electromagnetic environment statistical features, concatenating them into an input vector with a dimension of approximately 20 to 30. The network adopts a fully connected multilayer perceptron structure, containing five hidden layers with the following neuron counts: 128, 256, 256, 128, and 64, respectively. Each layer is followed by a ReLU activation function and a Dropout layer, with a default Dropout ratio of 0.2 to prevent overfitting. The output layer contains multiple neurons corresponding to several performance metrics to be predicted, including resonant frequency, VSWR, gain, and radiation efficiency. The output layer does not have an activation function and directly outputs continuous values. The loss function is defined as the weighted mean square error between the predicted and simulated performance metrics. Before calculating the loss, each metric is standardized, with the default weights for gain and VSWR being 2, and the default weight for radiation efficiency being 1. The Adam optimizer was used to train the network. The default initial learning rate was 0.001. The learning rate was halved when the validation set loss did not decrease for five consecutive epochs. The training dataset was divided in an 8:2 ratio, with a default batch size of 32 and a default total number of training epochs of 300. An early stopping strategy was used to prevent overfitting. After training, the surrogate model accuracy was evaluated on an independent test set. The mean absolute error of gain prediction should be less than 0.5 dB, the mean absolute error of VSWR prediction should be less than 0.1 dB, the relative error of resonant frequency prediction should be less than 2%, and the single prediction time of the surrogate model should be less than 1 millisecond.

[0074] The training data for the surrogate model comes from electromagnetic simulation. There is a systematic deviation between the simulation and the actual environment, and this deviation changes with the installation environment and cannot be eliminated by a one-time calibration.

[0075] In one embodiment of the present invention, the input to the environmental correction model is a concatenated vector of four physical parameters: antenna electrical dimensions, relative feed point position, metallic coverage, and equivalent dielectric loss factor. The antenna electrical dimensions are the ratio of the total antenna length to the operating wavelength, and the relative feed point position is the ratio of the feed point coordinates to the corresponding directional dimension of the antenna installation area; both are directly calculated from the antenna geometric parameters. At least 30 measured antenna samples with different geometric parameters for each installation environment are used for training, resulting in at least 120 samples across the four environments. Using the four parameters as independent variables and the deviation between the simulated and measured values ​​as the dependent variable, a Gaussian process regression with a radial basis function kernel is used to establish a deviation prediction model. The kernel function parameters are optimized by maximizing the marginal likelihood. For any installation environment other than the four typical environments, the metallic coverage and equivalent dielectric loss factor are obtained through on-site measurement or by consulting building material parameters. These parameters, along with the antenna electrical dimensions and relative feed point position, are input into the correction model to directly predict the performance deviation under that environment, thus achieving correction for any continuous environment. The Gaussian process regression also outputs an uncertainty estimate of the deviation prediction. This uncertainty information is passed down to the optimization step as the basis for calculating the safety margin of the constraint conditions.

[0076] The electromagnetic field distribution predicted by S2 has an average relative error of approximately 10%. This error, as the input uncertainty of the surrogate model S3, propagates downwards and affects the antenna performance prediction results. To quantify this error propagation effect, a Monte Carlo sampling method is used during the surrogate model inference: a random perturbation consistent with its prediction error distribution is applied to the electromagnetic field distribution tensor output by S2, and sampling is repeated several times (the default value is 50 times). The statistical characteristics of the electromagnetic field after each perturbation are input into the surrogate model, resulting in 50 sets of performance prediction results. The mean and standard deviation of these 50 sets of results are calculated. The mean is used as the point estimate of the performance prediction, and the standard deviation is used as the prediction uncertainty caused by the error propagation from S2. This uncertainty is combined with the environmental correction uncertainty output by the Gaussian process correction model to obtain the comprehensive prediction uncertainty, which is then passed to the S4 optimization step for setting the safety margin of the constraints. Through the above process, a physically parameterized environmental correction antenna performance surrogate model is constructed. This model can quickly predict the actual performance of the antenna in specific electromagnetic environments and arbitrary installation environments, and provides quantified prediction uncertainty, providing a reliable performance evaluation tool for subsequent optimization steps.

[0077] S4 utilizes the antenna performance proxy model and comprehensive prediction uncertainty, with the statistical characteristics of the electromagnetic environment as a fixed background, and generates multiple sets of candidate antenna parameter schemes through hierarchical mixed variable multi-objective optimization and dynamic safety margin constraint iteration.

[0078] The system receives the antenna performance proxy model, prediction uncertainty information, and design target requirements output from the third step. It then employs a multi-objective optimization algorithm improved for mixed variable spaces to optimize the system, obtaining multiple sets of candidate antenna parameter schemes that satisfy the constraints. Before entering the optimization iteration, the S2 prediction model is called to extract the electromagnetic environment statistical characteristics of the current energy meter model to be designed, including the average field strength, field strength variance, and dominant polarization direction of the antenna installation area. These three parameters are determined by the energy meter's PCB layout and components, and are independent of the antenna geometric parameters. They remain unchanged throughout the optimization process and are input as fixed constants, concatenated with the antenna geometric parameters each time the S3 proxy model is called.

[0079] Several optimization objectives and constraints for antenna design are clearly defined. Optimization objectives include: maximizing antenna gain (higher gain means longer communication distance); minimizing VSWR (lower VSWR means less reflection loss); maximizing radiation efficiency (higher radiation efficiency means more efficient energy utilization); minimizing antenna size (smaller size means less space occupied inside the energy meter); and minimizing manufacturing cost (lower cost means more economical mass production). Constraints include: antenna size must be smaller than the reserved space; antenna feed point and radiating arm positions must avoid interference-sensitive areas identified by S1; resonant frequency must fall within the operating frequency band; VSWR must be less than a set threshold within the operating frequency band; and gain must not be lower than the minimum required value. For each constraint, a safety margin is dynamically set for each candidate scheme based on the comprehensive prediction uncertainty output by S3: the constraint threshold for each candidate scheme is tightened by one standard deviation of the prediction uncertainty of the corresponding performance index based on the nominal value; the larger the prediction uncertainty, the greater the constraint tightening. To prevent solutions with high uncertainty from being systematically eliminated due to overly strict constraints, an upper limit is set for the tightening amount. The default value is 10% of the nominal constraint threshold. When the upper limit is exceeded, the tightening is performed according to the upper limit, ensuring that the optimization results take into account both performance and prediction reliability.

[0080] Antenna parameter optimization faces the challenge of a mixed variable space: the optimization variables simultaneously include discrete variables (antenna type, such as monopole, inverted-F, microstrip) and continuous variables (arm length, width, feed point position, etc.), and the dimensions of the continuous variables differ for different antenna types (monopole 2D, inverted-F antenna 4D, microstrip antenna 5D). When dealing with this variable-dimensional mixed variable space, the standard NSGA-II cannot directly operate across different types using crossover and mutation operators, leading to insufficient population diversity and uneven Pareto front coverage.

[0081] In one embodiment of the present invention, an independent subpopulation is maintained for each antenna type, with a default size of 30 individuals per subpopulation. Within each subpopulation, the continuous parameters are evolved using the standard NSGA-II simulated binary crossover and polynomial mutation operators. The default crossover probability is 0.9, and the default mutation probability is 1 divided by the parameter dimension of each subpopulation. After each generation iteration, a unified non-dominated sorting and crowding calculation are performed across all subpopulations. When selecting elite individuals, a minimum retention number is set for each antenna type, with a default value of 5, to prevent a certain antenna type from being completely eliminated due to a low target value in the current generation. Under the premise of satisfying the minimum retention number constraint, individuals are selected from the remaining slots according to the principle of low non-dominated level and high crowding, until the size of each subpopulation is restored to the set value.

[0082] The iterative optimization process begins. The default size for each subpopulation is 30 individuals per antenna type. Initial populations are generated randomly within their respective parameter spaces. For each individual in the current generation, its antenna parameters are extracted. The surrogate model from step three is used to predict the mean performance index under a given electromagnetic environment, while simultaneously acquiring the corresponding prediction uncertainty. The values ​​of each optimization objective are calculated based on the mean performance index. Constraints are checked against the individual's prediction uncertainty and dynamic safety margin. Individuals violating constraints are penalized by reducing their fitness. Non-dominated sorting and crowding calculations are performed on each subpopulation. A tournament selection strategy is used to select parent individuals. The default crossover probability is 0.9, and the default mutation probability is 1 divided by the parameter dimension of each subpopulation. After merging the parent and offspring populations, a unified non-dominated sort is performed across all subpopulations. When selecting individuals to retain, a minimum retention number is set for each antenna type, with a default value of 5. This means that at least 5 individuals of each antenna type are retained for the next generation, preventing a certain antenna type from being completely eliminated due to a low target value in the current generation, thus ensuring population diversity. Under the premise of meeting the minimum retention number constraint, individuals are selected from the remaining slots based on the principle of low non-dominated level and high crowding, until the size of each subpopulation recovers to the set value. The default value for the maximum number of iterations is 200 generations, or it can be terminated early if there is no significant improvement in the Pareto front for 20 consecutive generations.

[0083] After iteration, the first layer of non-dominated individuals in the final population, i.e., the Pareto optimal solution set, is extracted, and feasible solutions are selected according to the constraints. Preference weights are set according to actual engineering needs; for power company applications, gain and VSWR have higher weights, cost is secondary, and volume has a relatively lower weight. A weighted comprehensive score is calculated for each feasible solution, and the solutions are sorted from highest to lowest score. The top 10 solutions are selected as candidate antenna parameter schemes. For each candidate scheme, the specific values ​​of its antenna geometric parameters, as well as the mean and uncertainty range of the performance indicators predicted by the surrogate model, are compiled and output. The parameters and performance indicators of the candidate schemes are organized in tabular form for easy evaluation and comparison in subsequent steps.

[0084] S5. For candidate antenna parameter schemes, the final antenna design scheme and performance report are determined by screening the multi-environment robustness score of the environmental correction proxy model and verifying it by physical sample preparation and testing. If the measured deviation exceeds the threshold, the source of the error is traced and the corresponding model closed-loop update is triggered.

[0085] The candidate antenna parameter schemes output from the fourth step are received. Multi-environment performance simulation and robustness evaluation methods are used to obtain the optimal antenna design scheme with stable performance under various installation environments. The reliability of the scheme is verified by actual measurement, and the verification results are fed back to S3 to form a closed-loop iterative mechanism.

[0086] For each set of candidate antenna parameter schemes generated in the fourth step, an environmental robustness assessment was performed. Four typical installation environments were selected: a metal distribution box environment, a concrete wall mounting environment, a densely packed environment, and an open environment. These four environments cover the main scenarios in the actual deployment of smart meters. For each candidate scheme, the corresponding parameters for each environment were input into the environmental correction surrogate model constructed in the third step, according to the four input parameter formats defined in S3 (antenna electrical dimensions, relative feed point position, metal coverage, and equivalent dielectric loss factor). The antenna electrical dimensions and relative feed point position were calculated from the geometric parameters of the candidate scheme according to the transformation method defined in S3. The metal coverage and equivalent dielectric loss factor were taken from the measured values ​​of each typical installation environment. The gain, VSWR, and radiation efficiency of the antenna in the four environments were predicted, and the uncertainty range of each predicted value was obtained. For key candidate schemes, full-wave electromagnetic simulation was used for verification to confirm that the deviation between the predicted values ​​of the surrogate model and the simulated values ​​was within an acceptable range.

[0087] Based on performance data from four different environments, robustness evaluation metrics for each candidate scheme are calculated. The minimum, average, and standard deviation of the gain are calculated for each of the four environments. The minimum value reflects the antenna's performance assurance capability under worst-case conditions, and the standard deviation reflects performance stability. To eliminate the difference in dimensions and numerical range between the minimum gain and the reciprocal of the gain standard deviation, both items are first normalized to minimum and maximum values ​​across all candidate schemes before calculating the robustness score, resulting in a normalized minimum gain score and a normalized stability score, both ranging from 0 to 1. The robustness score is equal to the weighting coefficient. Multiply by the normalized minimum gain score, and add the weighting factor. Multiply by the normalized stability score. The normalized stability score is calculated from the standard deviation of the gain of each candidate scheme. The reciprocal (in dB) is obtained by min-max normalization. When calculating the reciprocal, a small constant is added to the denominator. (Default value is 0.01) Prevents division by zero; The default value is 0.7. The default value is 0.3, and the sum of the two is 1. For power companies deploying in batches, performance assurance under worst-case conditions takes priority over stability. The value is relatively large. A robustness score is calculated for all candidate solutions, and they are sorted from highest to lowest score. The solution with the highest robustness score is selected as the recommended solution.

[0088] To verify the practical feasibility of the recommended scheme, sample fabrication and field testing were conducted. Based on the geometric parameters of the recommended scheme, antenna samples were fabricated using either printed circuit board (PCB) or flexible circuit board (FPCB) technology. These antenna samples were then installed inside a smart meter prototype and soldered to the RF port of the communication module. Four smart meter prototypes equipped with the antenna were fabricated and placed in four typical installation environments for field testing. In a metal distribution box environment, the smart meter prototype was placed inside a 400mm × 300mm × 200mm iron distribution box with a box thickness of 1.5mm. A vector network analyzer was used to measure the antenna port reflection coefficient, and a spectrum analyzer and receiving antenna were used to measure the radiated field strength and calculate the gain. In a concrete wall mounting environment, the back of the smart meter prototype was placed against a 200mm thick concrete wall with a relative permittivity of 6, and the same measurements were performed. In a densely packed environment, three smart meter prototypes were horizontally arranged with an adjacent spacing of 50mm, and the antenna performance of the smart meter in the middle position was measured. In an open environment, the prototype electricity meter was placed in an anechoic chamber with no obstructions within 1 meter around it for benchmark measurements.

[0089] By comparing the measured data with the predicted data from the surrogate model under four different environments, the deviation between the predicted and measured values ​​is calculated. The average and maximum deviations of all performance indicators under the four environments are then statistically analyzed. If the average deviation is less than 15% and the maximum deviation is less than 25%, the prediction model is considered reliable, the recommended scheme is feasible, and this scheme is confirmed as the final design scheme. If the deviation exceeds the threshold, a closed-loop update process is initiated. First, a source analysis is performed on the deviation to determine whether the main source of the deviation is the electromagnetic field prediction error in S2 or the environmental correction error in S3. The specific method for tracing the source is as follows: Perform principal component analysis again on the magnetic field distribution of the interference-sensitive area in the measured electromagnetic field data of S1, extract the dominant magnetic field direction and field strength amplitude, calculate the equivalent magnetic dipole moment, and load it into the full-wave simulation model in the form of a lumped magnetic dipole. Simulate the recommended antenna scheme to obtain the simulation performance value based on the measured electromagnetic field. Compare the simulation value with the measured antenna performance data. If the deviation is within an acceptable range, it indicates that the surrogate model and environmental correction logic of S3 are reliable, and the deviation originates from the inaccurate electromagnetic field prediction of S2, requiring an update to the S2 model. If the deviation still exceeds the threshold, it indicates that the deviation originates from insufficient environmental correction in S3, requiring supplementary measured samples to retrain the Gaussian process correction model.

[0090] In one embodiment of the present invention, the closed-loop update employs an elastic weight consolidation method to complete the incremental model update: when fine-tuning the model by adding new measured data to the training set, the importance of each parameter to the existing model prediction task is first approximated using the diagonal elements of the Fisher information matrix. Regularization constraints proportional to the importance of parameters with high importance are applied, allowing the model to retain its predictive ability for existing models while learning new data. After the model update is completed, optimization and evaluation are re-executed until the measured deviation meets the threshold requirement.

[0091] The final antenna design, validated through field testing, includes complete antenna geometric parameters, manufacturing process requirements, installation location specifications, measured performance data under four different environments, and a performance fluctuation analysis report. This design can be directly applied to the mass production of smart meters. The design process has fully considered the influence of both the internal electromagnetic environment and the external installation environment, ensuring the antenna's performance stability and reliability in practical applications.

[0092] This invention relates to the antenna design of a single-phase smart energy meter, model DTS-001, with a shell size of 180mm×120mm×65mm, an internal reserved antenna installation space of 50mm×30mm×8mm, an operating frequency band of 800MHz, a target gain of not less than -5dBi, and a VSWR of less than 2.

[0093] The electromagnetic field distribution data obtained in the first step is shown in Table 1 below. The field strength amplitude is the highest (31.8V / m) at the measurement point (25,15,5). This location corresponds to the near-field region of the internal switching power supply transformer of the energy meter and is marked as an interference-sensitive area. In the S4 optimization, the antenna feed point and radiating arm are constrained to avoid this region.

[0094] Table 1: Electromagnetic field data at typical measurement points inside the electricity meter

[0095]

[0096] Table 2 below shows examples of typical installation environment test data obtained in the first step, along with the physical characteristic parameters of each environment:

[0097] Table 2: Measured performance and environmental physical characteristics of the reference antenna in typical installation environments.

[0098]

[0099] The optimal antenna design using the method of this invention is an inverted-F antenna with the following geometric parameters: horizontal arm length 42mm, vertical arm height 7mm, distance from feed point to short-circuit point 8mm, floor dimensions 50mm × 30mm, substrate dielectric constant 4.4, thickness 1.6mm, and the feed point location avoids the interference-sensitive area near coordinates (25, 15, 5). The measured gain of this design in four environments are: metal distribution box -5.8dBi, concrete wall -5.2dBi, densely packed -5.6dBi, and open environment -4.1dBi, with a gain fluctuation range of 1.7dB, meeting the design requirement of no more than 2dB. The performance is stable and reliable, meeting the application requirements for mass deployment.

[0100] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A smart meter integrated antenna design method, characterized in that, Includes the following steps: S1: Collect structural characteristic parameters of the smart energy meter prototype, and measure them using triaxial optical electric field sensing and frequency selective phase-locked amplification to form an internal electromagnetic field distribution dataset and an installation environment antenna performance dataset. S2, based on the electromagnetic field distribution dataset and structural feature parameters, through encoder-decoder convolutional neural network training and structural similarity adaptive transfer, outputs the electromagnetic field distribution prediction model and the electromagnetic field distribution prediction result of the current model; S3, combining the electromagnetic field distribution prediction results with the antenna performance dataset of the installation environment, and through the equivalent magnetic dipole simulation modeling driven surrogate model training and Gaussian process physical parameterization environment correction, forms an antenna performance surrogate model containing comprehensive prediction uncertainty. S4 utilizes the antenna performance proxy model and comprehensive prediction uncertainty, with the statistical characteristics of the electromagnetic environment as a fixed background, and generates multiple sets of candidate antenna parameter schemes through hierarchical mixed variable multi-objective optimization and dynamic safety margin constraint iteration. S5. For candidate antenna parameter schemes, the final antenna design scheme and performance report are determined by screening through multi-environment robustness scoring of the environmental correction surrogate model and verification by physical sample preparation and testing. If the measured deviation exceeds the threshold, the source of the error is traced and the corresponding model closed-loop update is triggered.

2. The method of claim 1, wherein, In S1, a non-contact optical electric field sensing system based on lithium niobate electro-optic crystal is used to measure the electromagnetic field inside the smart energy meter. The probe is encapsulated with three mutually orthogonal lithium niobate crystals, which correspond to the electric field components in the X, Y, and Z directions, respectively. The three modulated optical signals are synchronously received by three photodetectors and synthesized to obtain the electric field intensity vector of each measurement point. Using the periodic transmission signal of the NB-IoT communication module as a reference clock, the lock-in amplifier only integrates and extracts the signal component that is in phase with the reference clock, thus suppressing low-frequency harmonic interference generated by the switching power supply. After completing the three-dimensional scanning measurement, the mean and standard deviation of the field strength amplitude of all grid points in the antenna installation area are calculated. Grid points with field strength amplitude exceeding the mean plus one standard deviation and their surrounding preset range are marked as interference-sensitive areas.

3. The method of claim 1, wherein the method further comprises: In S1, the process of forming the antenna performance dataset of the installation environment includes: selecting four typical installation environments: metal distribution box environment, concrete wall mounting environment, densely arranged environment, and open environment; for each environment, fabricating a reference antenna and a multi-geometric parameter antenna sample selected by the Latin hypercube sampling method; using a vector network analyzer to measure the reflection coefficient and extract the resonant frequency and standing wave ratio; using a far-field antenna test system to measure the radiation pattern and gain, and simultaneously measuring the radiation efficiency. The physical characteristics of each environment are quantitatively characterized by two parameters: metal coverage and equivalent dielectric loss factor. The environmental physical characteristic parameters are associated with the corresponding antenna performance data and stored to form an environmental characteristic-performance mapping dataset.

4. The method of claim 1, wherein the method further comprises: In S2, the encoder part of the encoder-decoder convolutional neural network receives structural feature vectors and PCB layout feature maps as inputs. The structural feature vectors are mapped to the latent space through fully connected layers, and the PCB layout feature maps extract spatial features through two two-dimensional convolutional layers. The two are flattened and spliced ​​together and mapped to a high-level semantic feature vector. The decoder maps and reshapes the feature vector into a three-dimensional tensor through a fully connected layer. After being progressively upsampled by three three-dimensional deconvolutional layers, the final output is an electromagnetic field distribution prediction tensor with a size of D×H×W×6, where the six channels correspond to the electric field components and magnetic field components in the X, Y, and Z directions, respectively.

5. The method of claim 1, wherein the method further comprises: In S2, the loss function of the encoder-decoder convolutional neural network is a weighted composite loss, which consists of an amplitude error term and a polarization direction error term. The amplitude error term adopts a weighted mean square error, and the weights are assigned in three levels according to the region type label output by S1. The non-interference sensitive grid points in the antenna installation area have the highest weight, the grid points in the interference sensitive area have the second highest weight, and the grid points in the other areas have the lowest weight. The polarization direction error term calculates the direction cosine of the predicted electric field unit vector and the measured electric field unit vector for each grid point. The direction error is calculated by subtracting the direction cosine from 1. The direction error weight is the highest for grid points in the interference-sensitive area. The total loss function is the sum of the magnitude error term and the polarization direction error term weighted by the balance coefficient.

6. The method of claim 1, wherein the method further comprises: In S2, the process of adaptive structural similarity transfer includes: calculating structural similarity scores for the new model and all existing models respectively. The structural similarity scores are obtained by weighted summation of three sub-items: shell size similarity, PCB layout similarity, and power supply similarity. The basic model corresponding to the existing model with the highest score is selected as the transfer source. When the structural similarity score is higher than the first threshold, all encoder parameters are frozen, and only the last two layers of the decoder are fine-tuned; when the structural similarity score is between the second and first thresholds, the first two layers of the encoder are frozen, and all subsequent layers of the encoder and the decoder are fine-tuned; when the structural similarity score is lower than the second threshold, the model is trained from scratch with the basic model parameters as the initial values.

7. The method of claim 1, wherein the method further comprises: In S3, the proxy model training process driven by the equivalent magnetic dipole simulation modeling includes: performing principal component analysis on the magnetic field distribution of the interference-sensitive region in the electromagnetic field distribution predicted by S2, extracting the dominant magnetic field direction and the corresponding field strength amplitude, and loading the corresponding region into a magnetic dipole in the form of a lumped magnetic dipole into the full-wave electromagnetic simulation model. The average field strength, field strength variance, and dominant polarization direction of the antenna installation area are extracted from the electromagnetic field distribution predicted by S2 as statistical features. These statistical features are then concatenated with the antenna geometric parameters as the input feature vector of the surrogate model. A fully connected multilayer perceptron is constructed as the antenna performance surrogate model, and the output layer corresponds to the resonant frequency, VSWR, gain, and radiation efficiency.

8. The method for designing an integrated antenna for a smart energy meter according to claim 1, characterized in that, In S3, the process of Gaussian process physical parameterization environment correction includes: using the spliced ​​vector of four physical parameters, namely antenna electrical size, feed point relative position, metal coverage and equivalent dielectric loss factor, as the input of the environment correction model, and using the deviation between the simulation prediction value and the measured value as the dependent variable, and establishing a deviation prediction model by using Gaussian process regression with radial basis function kernel. The formation process of the comprehensive prediction uncertainty includes: applying a random perturbation that conforms to its prediction error distribution to the electromagnetic field distribution tensor output by S2 and repeatedly sampling it; inputting the electromagnetic field statistical characteristics after each perturbation into the surrogate model to obtain multiple sets of performance prediction results; calculating its standard deviation as the error propagation uncertainty; and merging the error propagation uncertainty with the environmental correction uncertainty output by the Gaussian process correction model to obtain the comprehensive prediction uncertainty.

9. The design method for an integrated antenna in a smart energy meter according to claim 1, characterized in that, In S4, the hierarchical mixed variable multi-objective optimization process includes: maintaining an independent subpopulation for each antenna type, using simulated binary crossover and polynomial mutation operators to evolve continuous parameters within the subpopulation, performing unified non-dominated sorting and crowding calculation across all subpopulations after each iteration, and setting a minimum retention number for each antenna type to prevent complete elimination. The dynamic safety margin constraint iteration process includes: for each candidate solution, tightening the constraint threshold by one standard deviation of the prediction uncertainty of the corresponding performance index of the candidate solution based on the nominal value, and setting a preset upper limit for the tightening amount.

10. The method of claim 1, wherein the method further comprises: In step S5, the process of screening the multi-environment robustness score of the environmental correction surrogate model includes: using the environmental correction surrogate model to predict the performance index of each candidate scheme in four typical installation environments, calculating the minimum value and standard deviation of the gain in the four environments, weighted summing the normalized minimum gain score and the normalized stability score to obtain the robustness score, and selecting the scheme with the highest robustness score as the recommended scheme; the closed-loop update process includes: If the measured deviation exceeds the preset threshold, the magnetic field distribution of the interference-sensitive area in the measured electromagnetic field data of S1 is re-analyzed as principal component and loaded into the full-wave simulation model in the form of a lumped magnetic dipole. The simulated value is compared with the measured value to determine whether the deviation comes from the electromagnetic field prediction error of S2 or the environmental correction error of S3, and the corresponding model update is triggered.

Images