Method for identifying and suppressing radio frequency switch jamming in complex electromagnetic environment based on spectrum analysis and deep learning

Abstract

The application relates to the technical field of deep learning, and discloses a complex electromagnetic environment radio frequency switch interference identification and suppression method based on spectrum analysis and deep learning. The method establishes a radio frequency switch macro-micro cross-scale geometric model, outputs a physical field response characteristic library, configures a bottom layer synchronous control link to collect high-frequency scattering parameters, outputs a microwave radio frequency real-time data set, constructs a decoupling convolutional neural network model, fuses the physical field response characteristic library and the microwave radio frequency real-time data set, and outputs a high-dimensional fusion feature tensor. The high-dimensional fusion feature tensor is input into a multi-branch output architecture to perform orthogonal decoupling constraint, and continuous interference feature vectors and discrete classification probability vectors are output. The application integrates a nonlinear physical mechanism into a network optimization process to eliminate time-space scale barriers, reduce model feature misjudgment and omission probability, simultaneously drives external hardware to perform adaptive compensation actions based on output vectors, prevents system protection mechanism failure or false triggering, and maintains equipment operation states.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning. Background Technology

[0002] As a core control device in RF front-end modules, RF switches are widely used in communication architectures such as millimeter-wave active antenna systems, responsible for path switching and isolation control of RF signals. In complex electromagnetic environments, the physical structure of RF switches is subjected to multiple effects, including high-frequency electromagnetic pulses, drastic temperature fluctuations, and alternating mechanical stresses, leading to microscopic physical evolution of their contact surfaces. Deep learning technology, with its high-dimensional data feature extraction capabilities, is gradually being introduced into the fields of state monitoring and interference identification for microwave RF devices. By collecting and analyzing spectral data during RF switch operation, establishing a deep neural network model to map the mathematical relationship between environmental stress and device degradation characteristics represents an important technological advancement direction for identifying and suppressing interference in complex electromagnetic environments.

[0003] Current mainstream technologies for identifying radio frequency (RF) switch interference primarily rely on purely data-driven deep convolutional neural network models. These approaches typically use large-scale pre-collected RF scattering parameters or spectral sequences as training samples, employing multi-layer convolutional and pooling structures to extract time-series data features under electromagnetic interference and outputting a classification inference of the interference type. In practice, a common approach is to construct a single-dimensional real-time microwave RF dataset under set standard operating conditions, using backpropagation algorithms to optimize network weights to fit the mapping between input signals and fault labels. Furthermore, some technologies combine basic multiphysics simulation tools to assist in obtaining theoretical data, independently solving for electric or temperature field distributions using simulation software, and feeding single-physics features as auxiliary input parameters into the deep learning network to improve the accuracy of identifying interference in complex electromagnetic environments.

[0004] However, existing data-driven interference identification techniques have limitations when dealing with dynamically evolving and complex electromagnetic environments. Existing convolutional neural network models often struggle to accurately identify hidden physical faults caused by contact degradation and thermal stress accumulation, easily leading to misjudgments or missed detections under dynamic and complex electromagnetic interference. The main reason for this problem is that data-driven models heavily rely on the distribution of historical training data, failing to incorporate the nonlinear physical mechanisms of the internal physical structure of RF switches under extreme environments into the network optimization process. This limits the model's generalization ability when facing unknown and complex electromagnetic environments. Simultaneously, existing technologies struggle to eliminate the spatiotemporal scale barrier between microscopic physical contact degradation and macroscopic RF transmission performance degradation, preventing deep cross-constraint between discrete time-series RF characteristics and physical field response characteristics. Prolonged exposure to these limitations can lead to RF systems failing to accurately perform adaptive compensation when encountering high-dimensional multimodal interference, resulting in the failure or false triggering of RF system protection mechanisms, ultimately reducing the operational reliability and overall lifespan of communication equipment. Therefore, a method for identifying and suppressing RF switch interference in complex electromagnetic environments based on spectrum analysis and deep learning is urgently needed to address these issues. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides a method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] To solve the aforementioned technical problem, the present invention is achieved through the following technical solution: This invention provides a method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning, specifically including: Establish a macro-microscale cross-scale geometric model of the radio frequency switch, perform multi-physics joint solution, and output a physical field response feature library containing theoretical contact conductivity; Configure the underlying synchronous control link to collect the high-frequency scattering parameters of the RF switch according to the preset sampling period and output the microwave RF real-time dataset; A decoupled convolutional neural network model is constructed, and the physical field response feature library is fused with the microwave radio frequency real-time dataset at the feature level to output a high-dimensional fused feature tensor. The high-dimensional fusion feature tensor input includes a regression branch and a classification branch. The multi-branch output architecture is subjected to orthogonal decoupling constraints, and the output is a continuous disturbance feature vector and a discrete classification probability vector.

[0007] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the establishment of the macro-micro cross-scale geometric model of the radio frequency switch includes: Extract the three-dimensional structural dimensional parameters of the RF switch to generate a macroscopic continuous medium solution domain space; Discrete coordinate point clouds are generated using three-dimensional fractal functions to reconstruct micro-rough contact surfaces; At the contact interface, a moving least squares method is introduced to construct a shape function interpolation mapping interface. The radius of the tight support domain is determined based on the average physical spacing of macroscopic grid nodes. The displacement interpolation weight is calculated by combining the weight function to construct a cross-scale displacement transfer matrix. The macroscopic displacement field of the macroscopic continuous medium solution domain is mapped to the microscopic deformation displacement of the microscopic rough contact surface through the cross-scale displacement transfer matrix.

[0008] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the step of performing multi-physics joint solution includes: In the solution domain of macroscopic continuous medium, the wave equation, the unsteady heat transfer equation and the thermoelasticity equation are solved simultaneously to obtain the flattened matrix of mechanical stress-deformation tensor. A dual-loop architecture is adopted in the solution process, and the macroscopic time step of the heat transfer and mechanical solution is set to be a multiple of the microscopic time step of the electromagnetic field solution. The relative norm residuals of the temperature field and the stress field are calculated. When the relative norm residuals are less than the preset convergence tolerance threshold, it is determined that the multiphysics distribution has reached steady-state equilibrium and the macroscopic time variable is advanced until the preset total duration of transient simulation is reached.

[0009] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the output includes a physical field response feature library containing theoretical contact conductivity, comprising: The local normal contact stress at the contact interface is analyzed from the flattened matrix of the mechanical stress deformation tensor. The implicit force balance equation is solved based on the local normal contact stress to determine the equivalent separation distance; The actual microscopic contact area is calculated by combining the probability density function of the height distribution of the micro-protrusions with the equivalent separation distance. The theoretical value of nonlinear contact conductivity is determined based on the actual microscopic contact area. The multi-physics field distribution characteristics are resampled and mapped to the theoretical value of the nonlinear contact conductivity to generate the physical field response feature library.

[0010] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the configuration of the underlying synchronization control link includes: A low-level hard synchronization control link to an external test terminal is established using a field-programmable gate array (FPGA) hardware interface. Connect to an external reference clock to generate a global hardware system clock; A multi-channel trigger signal generation module is constructed based on the global hardware system clock to synchronously send hardware trigger pulses to the external test terminal to drive feature capture.

[0011] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the step of acquiring the high-frequency scattering parameters of the radio frequency switch and outputting a real-time microwave radio frequency dataset includes: The current count value of the synchronous latched high-frequency counter is used to generate an absolute timestamp; Calculate the inherent physical delay of the radio frequency signal in the transmission line; The absolute timestamp is superimposed with the inherent physical delay to generate an effective sampling time, and the complex scattering parameters of the RF switch are captured at the effective sampling time; The amplitude and phase information are extracted and reassembled with the data packets at the effective sampling time. The data is then updated using a sliding window strategy to construct the microwave radio frequency real-time dataset.

[0012] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the step of constructing a decoupled convolutional neural network model and outputting a high-dimensional fused feature tensor includes: The decoupled convolutional neural network model includes a physical field feature extraction branch, a real-time radio frequency feature extraction branch, and a cross-attention fusion module; The physical field response feature library is input into the physical field feature extraction branch to extract spatial features and output a spatial feature tensor. The microwave radio frequency real-time dataset is input into the real-time radio frequency feature extraction branch for time-series feature extraction, and the microwave radio frequency feature tensor is output. The cross-attention fusion module is used to calculate the cross-attention weight matrix between the spatial feature tensor and the microwave radio frequency feature tensor. After feature-level fusion, the high-dimensional fused feature tensor is output.

[0013] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the construction of the decoupled convolutional neural network model further includes iterative training of the decoupled convolutional neural network model: During training, a total loss function including physical mechanism constraints is used to update parameters; an auxiliary reconstruction branch is constructed inside the decoupled convolutional neural network model to output the predicted value of local microscopic contact conductivity. The physical mechanism constraint term uses the theoretical contact conductivity in the physical field response feature library to perform residual constraints on the predicted value of the local microscopic contact conductivity.

[0014] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the high-dimensional fusion feature tensor input includes a multi-branch output architecture with regression and classification branches, which undergoes orthogonal decoupling constraints. The output continuous interference feature vector and discrete classification probability vector include: The high-dimensional fusion feature tensor is reduced in dimensionality and mapped to a global fusion feature vector. The globally fused feature vector is then fed into the regression and classification branches; The regression branch outputs a continuous interference feature vector that characterizes the transmission performance index of the radio frequency channel. The classification branch outputs a discrete classification probability vector that characterizes the probability of mechanical failure.

[0015] As a preferred embodiment of the method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning described in this invention, the method further includes the following after outputting the continuous interference feature vector and the discrete classification probability vector: The transmission performance index in the continuous interference feature vector is analyzed to generate a multi-channel bias control voltage tensor, which is then applied to an external adjustable capacitor array to compensate for the RF channel impedance mismatch. Extract the fault probability variable from the discrete classification probability vector. When the fault probability variable exceeds a preset safety limit threshold, trigger an interrupt request to drive the radio frequency solid-state switch to perform channel switching and apply phase compensation.

[0016] The present invention has the following beneficial effects: 1. This invention establishes a macro-micro cross-scale geometric model of a radio frequency switch, outputs a physical field response feature library containing theoretical contact conductivity, constructs and trains a decoupled convolutional neural network model constrained by nonlinear physical mechanisms, and uses this feature library to perform residual constraints on the network's hidden layer prediction values. This integrates the nonlinear physical mechanisms of the internal physical structure of the radio frequency switch under extreme environments into the network optimization process, changes the dependence of the pure data-driven model on the distribution of historical training data, and enables the model to identify hidden physical faults caused by contact degradation and thermal stress accumulation, reducing the probability of feature misjudgment and missed judgment in complex electromagnetic environments.

[0017] 2. This invention collects real-time microwave radio frequency datasets and uses a cross-attention mechanism to process the physical field response feature library and the real-time microwave radio frequency datasets. It fuses discrete physical field tensors and temporal radio frequency tensors into a high-dimensional fused feature tensor, which can eliminate the spatiotemporal scale barrier between microscopic physical contact degradation and macroscopic radio frequency transmission performance degradation, and enable discrete temporal radio frequency features and physical field response features to form a cross-fusion constraint.

[0018] 3. This invention uses a high-dimensional fusion feature tensor input with a multi-branch output architecture containing regression and classification branches, performs orthogonal decoupling constraints, and outputs a continuous interference feature vector and a discrete classification probability vector. The decoupled continuous interference feature vector and discrete classification probability vector drive external hardware to perform adaptive compensation actions, preventing the RF system from experiencing protection mechanism failures or false triggering under multimodal interference, thus maintaining the operating status and lifespan of the communication equipment.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 The flowchart of this invention provides a method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning.

[0022] Figure 2 This is a schematic diagram of the S3 process provided by the present invention.

[0023] Figure 3 This is a schematic diagram of the S4 process provided by the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 Existing pure data-driven convolutional neural networks often have limited generalization ability when facing unknown and complex electromagnetic environments. They also struggle to capture the nonlinear response of the internal physical structure of RF switches under extreme conditions. This can easily lead to feature misjudgment or omission in the network model under dynamically changing complex electromagnetic interference. Consequently, they cannot promptly identify hidden physical faults caused by contact degradation or thermal stress accumulation, resulting in the failure or false triggering of RF system protection mechanisms and reducing the operational reliability and service life of the equipment.

[0026] To solve the above technical problems, such as Figure 1 As shown, Embodiment 1 of the present invention provides a method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning. Specifically, Embodiment 1 takes the radio frequency front-end module of a 5G millimeter-wave active antenna system as an example. It integrates macro- and micro-scale multi-physics field simulation with a decoupled convolutional neural network model constrained by nonlinear physical mechanisms to form a closed-loop framework for accurate extraction and collaborative suppression of multi-dimensional interference features of radio frequency switches. The specific steps include: S1. Establish a macro-micro cross-scale geometric model of the radio frequency switch, perform joint solution based on multi-timescale synchronization and residual convergence control, and output a physical field response feature library containing theoretical contact conductivity. S2. Configure the underlying synchronous control link, collect the high-frequency scattering parameters of the RF switch in actual operation according to the preset sampling period, execute the acquisition mechanism including multi-parameter joint delay compensation and timing fault tolerance management, and output microwave RF real-time dataset. S3. Construct and train a decoupled convolutional neural network model constrained by nonlinear physical mechanisms, and fuse simulation prior data with real-time test signals at the feature level through a deep learning architecture; S4. Input the high-dimensional fusion feature tensor into a multi-branch output architecture containing regression and classification branches, perform orthogonal decoupling constraints, and output continuous disturbance feature vectors and discrete classification probability vectors.

[0027] In this invention, the physical field response feature library generated in step S1 and the real-time microwave radio frequency dataset acquired in step S2 are passed as parallel inputs to step S3. Step S3 fuses the discrete physical field tensor and the time-series radio frequency tensor into a high-dimensional fused feature tensor through a cross-attention mechanism, and uses the theoretical contact conductivity label calculated in step S1 to apply residual constraints to the network hidden layer prediction values. Step S4 receives the high-dimensional fused feature tensor and splits it into independent continuous interference feature vectors and discrete classification probability vectors through orthogonal transformation. The above steps work together to complete a complete data closed loop from the original physical simulation and measurement to the decoupled control features.

[0028] Furthermore, to better illustrate the technical solution of Embodiment 1 of the present invention, a method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning is described in detail, specifically including the following: First, S1 establishes a macro- and micro-scale multi-scale geometric model of the RF switch, and performs joint solution based on multi-timescale synchronization and residual convergence control, outputting a physical field response feature library containing theoretical contact conductivity. To accurately simulate the physical degradation and micro-contact characteristics of the RF switch under complex electromagnetic environments and eliminate the spatiotemporal scale barriers in multiphysics calculations, step S1 establishes and solves the basic physical feature library through the following sub-steps: S11. Establish a macro-micro cross-scale geometric model of the RF switch and construct a node mapping interface based on the moving least squares method. The specific process is as follows: S111. Extract the three-dimensional structural dimension parameters of the RF switch, and generate a region containing dielectric material through geometric topology reconstruction. With conductor region Solution domain space of macroscopic continuum In the macroscopically overlapping area between the stationary contact and the end of the moving spring, a three-dimensional Weierstrass-Mandelbrot fractal function is used to generate discrete coordinate point clouds to reconstruct the microscopically rough contact surface. The height of this three-dimensional Weierstrass-Mandelbrot fractal surface profile is... The specific mathematical expression is: ; In the formula, For microscopic deformation displacement, the dimensions are... ; The macroscopic cutoff length has the following dimensions: ; Here is the characteristic scale parameter for fractal roughness, with dimensions of ; It is the three-dimensional fractal dimension, which is dimensionless. These are spatial frequency scaling coefficients, dimensionless; and Let be absolute coordinates in space, with dimensions . ; It is the frequency index; This is the cutoff constant used to determine the highest microscopic frequency component.

[0029] Among them, the three-dimensional fractal dimension Characterizing the complexity and spatial density of micro-rough contact surfaces, satisfying The contact materials of radio frequency microelectromechanical system (MEMS) switches mainly use gold, gold-nickel alloys, ruthenium, and platinum. The metal thin film contact surface is prepared by electroplating, and its three-dimensional fractal dimension is... Usually concentrated to Between these, the surface of the sputtered thin film with fine and uniform grains is relatively smooth. Pick The surface structure of the thick electroplated film, exhibiting island-like growth characteristics, is more complex. Pick to Fractal roughness characteristic scale parameters The overall longitudinal amplitude scaling ratio of the micro-protrusion is determined by the RF switch's requirement to maintain extremely high contact surface flatness to increase the actual microscopic contact area in order to reduce the complex contact impedance during high-frequency signal transmission. Under the significant skin effect in the microwave band, the root mean square roughness of the contact surface needs to be controlled at the nanometer to sub-nanometer level, corresponding to the parameters... The range of values ​​is within to Between. Spatial frequency scaling coefficients To ensure that the power spectral density of the generated rough surface conforms to the characteristics of a continuous random process on a real physical surface, it is necessary to avoid the generation of discrete periodic spatial frequency distributions. (Parameters...) Take non-integer values ​​and satisfy In contact mechanics modeling, the value is usually fixed. To ensure the rationality of spatial frequency distribution. Macroscopic cutoff length. Depending on the macroscopic physical overlap characteristics of the RF switch's stationary contact and moving reed, a typical value is within... to Within range; cutoff constant It is necessary to consider the lattice micro-stopping scale of the contact material. The calculation yields the following formula: Taking gold material contacts as an example, its lattice constant Approximately ,Will and Substituting into this relation will determine the result. The discrete integer values ​​enable fractal geometric topological reconstruction to fully cover all spatial frequency components from the macroscopic contact region to the microscopic lattice scale.

[0030] S112. To achieve data interaction between macroscopic entities and microscopic surfaces, a moving least squares method is introduced at the contact interface to construct a shape function interpolation mapping interface. The node coordinate set of the macroscopic continuum mesh at the contact interface is extracted. The nodal coordinate set of the micro-fractal rough surface Define the radius of the tight support region. The parameter is based on the average physical spacing of the macroscopic mesh nodes at the interface. Determined, the selection criteria satisfy In the formula The dimensionless scaling factor for the support domain is configured to have a range of values. This ensures that the local support domain contains the minimum number of nodes that satisfy the nonsingularity condition of the space polynomial basis functions. For any microscopic node... With macro nodes Calculate dimensionless distance Introducing cubic spline weight functions Calculate the displacement interpolation weights using the cubic spline weight function. The piecewise mathematical expression is: ; Establish the cubic spline weight function The continuous spatial integral form constrains the energy distribution of the cross-scale mechanical mapping in the spherically supported domain. The volume integral expression within is: Construct a cross-scale displacement transfer matrix based on the moving least squares method. Define linear polynomial basis vectors for three-dimensional spatial coordinates. Constructing a local weighted residual matrix by traversing nodes. With basis vector weight matrix The interpolation shape function for generating a single micro-node relative to a macro-node is derived. All valid interpolation functions within the contact interface will be used. The displacement transfer matrix is ​​extracted and assembled into the global matrix in order of nodal degrees of freedom, and outputs the cross-scale displacement transfer matrix. Through cross-scale displacement transfer matrix Macroscopic displacement field solved by macroscopic mechanics Down-mapping to microscopic deformation displacement of microscopic nodes Its calculation expression is: ; S113, Solving in the continuous medium solution space The outer boundary is set based on the absorbing boundary condition of the ideal matching layer, and the air-exposed surface is set with the condition satisfying the requirements. The convection heat dissipation boundary conditions are set on the bottom surface of the anchoring structure to satisfy... Fixed support mechanical constraint boundary conditions.

[0031] S12. Perform a multiphysics joint solution based on time discretization algorithm and energy residual control. The specific process is as follows: S121, at the target monitoring angular frequency Within this range, a hybrid solution framework combining the multilevel fast multipole algorithm and the finite element method is used to solve problems in the macroscopic continuous medium solution domain. The system solves the Maxwell-Helmholtz wave equation, the unsteady heat transfer equation, and the thermoelastic equations simultaneously. It employs a separated, sequential, weakly coupled strategy to perform cross-physics field iterations. The unidirectional transfer order of the coupled calculations is set as follows: the electromagnetic field module calculates and generates a spatial heat source, which is then distributed to the thermal module; the thermal module outputs the nodal temperature field, which is mapped to the mechanical module; the mechanical module outputs the boundary deformation displacement and feeds it back to update the spatial mesh of the electromagnetic field module. The frequency domain differential form of the Maxwell-Helmholtz wave equation is: In the formula, Relative permeability The relative permittivity, Let be the free-space wavenumber. The macroscopic time step for the heat transfer and mechanics solutions is set to . The microscopic time step for solving the electromagnetic field is Define an abstract parameter for the number of steps in the sub-loop. The ratio of the characteristic time of the macroscopic thermodynamic response to the physical timescale of the period of the microscopic electromagnetic wave satisfies the following relationship: This eliminates the difference in computational rigidity between the evolution of high-frequency radio frequency signals and the thermal deformation of low-frequency structures.

[0032] S122, at each micro-time step Solve for the electric field distribution matrix. And calculate the Joule heat power density distribution matrix. Entering a single macroscopic time step At that time, the cumulative Joule heat power density after time averaging within the microscopic time step is used as a constant internal heat source to construct the unsteady-state heat transfer equation: In the formula, For material density, For specific heat capacity, Thermal conductivity, For local real-time temperature, The time variable is used. The backward Euler method is employed to discretize the above unsteady-state heat transfer equations in time to solve for the temperature rise vector at the next moment. Thermoelastic equilibrium equations are constructed based on the quasi-static assumption. Introducing the thermoelastic constitutive equation and geometric strain equations Solve the simultaneous equations to obtain the flattened matrix of the mechanical stress-deformation tensor at this moment. In the formula, Let be the elastic stiffness tensor. The coefficient of thermal expansion is It is a second-order unit tensor. In the current iteration step... To the next step During this period, the relative norm residuals of the temperature field and the stress field were calculated. : ; In the formula, and To prevent small regularization constants with a denominator of zero, a convergence tolerance threshold is set. The range of values ​​is to Preferably take When the relative norm residual Three consecutive times the value is less than the set convergence tolerance threshold At this point, it is determined that the multiphysics distribution within the current macroscopic time step has reached steady-state equilibrium. After achieving local convergence in the current macroscopic time step, the system updates the macroscopic time variables to... And determine the global progress status. When the macroscopic time variable Greater than or equal to the preset total duration of transient simulation If the condition is met, the system terminates the entire multiphysics joint solution process; otherwise, it reconstructs the solution domain mesh using the updated mechanical displacement boundaries and returns to perform the next macroscopic time step of the physics loop iteration.

[0033] S13. Construct a physical field response feature library based on the elastoplastic microscopic contact mechanics model and Holm's classical theory. The specific process is as follows: S131, the flattened matrix of mechanical stress deformation tensor obtained in step S12 Local normal contact stress at dynamic contact interface in analytical analysis We introduce the Greenwood-Williamson (GW) microscopic contact mechanics model, assuming the nominal contact area of ​​the interface is... The surface density of the micro-protrusions is The equivalent radius of curvature of the micro-convex body is Local normal contact stress Determines the equivalent separation distance between macroscopic contact surfaces Among them, the equivalent separation distance It is not an independent input variable, but is uniquely determined by solving the following implicit force equilibrium equations based on Hertzian contact theory: In the formula, The comprehensive elastic modulus is calculated using the following expression: , , and , These represent the Young's modulus and Poisson's ratio of the materials in contact, respectively. In practical solutions, the bisection method or the Newton-Raphson iteration method is used to numerically solve the above single-variable nonlinear equations to obtain the equivalent separation distance at the current macroscopic time step. Then, substitute the actual microscopic contact area. The integral expression.

[0034] S132. Assume that the height distribution of the micro-convex body follows a Gaussian distribution, and its probability density function is... The specific mathematical expression is: ; In the formula, The standard deviation of the micro-protrusion height distribution characterizes the surface roughness of the micro-contact surface. The actual micro-contact area is calculated by performing a probability integral based on the Gaussian distribution. : ; For cases where the contact load exceeds the elastic limit, a plasticity index can be introduced into the aforementioned GW micro-contact mechanics model. Make corrections: ; In the formula, The material hardness. When the dimensionless plasticity index... When considering the effect of plastic yielding on the actual contact area, it is necessary to take into account the correction. Specifically, the CEB elastoplastic contact model can be introduced for alternative implementation.

[0035] S133. Based on Holm's classical contact conductivity theory, derive the theoretical value of nonlinear contact conductivity, which includes the temperature rise effect and area regularization term. : In the formula, The intrinsic resistivity of the contact material at the reference temperature. is the temperature coefficient of resistivity of the material. The area regularization constant is used to prevent mathematical singularities in the completely open-circuit state. The dimensions are The typical value range is Preferably, take This value is much smaller than the actual microscopic contact area in the conductive state (typically within the range of...). (on the order of magnitude), ensuring that the theoretical leakage conductivity in the open-circuit state is less than [a certain value]. It has no statistically significant impact on the calculation results under normal conduction conditions.

[0036] An inverse distance weighting algorithm is employed to weight the unstructured mesh based on the spatial distance of the nearest neighbor original data points within the local support domain of the target mesh node, and voxel resampling is performed. The electric field distribution matrix, temperature field distribution data, mechanical stress deformation tensor, and theoretical values ​​of nonlinear contact conductivity under different mesh densities are mapped to a unified three-dimensional regular Cartesian coordinate system. Alignment and stitching of multidimensional tensors along the feature channel dimensions are then performed to generate a physical field response feature library. When the RF switch is off, the equivalent separation distance is... When the lower bound of the integration domain of the Gaussian probability integral approaches positive infinity, the actual microscopic contact area... Strictly converges to zero; operator solution ensures that the theoretical contact conductivity converges to the basic insulation leakage conductivity, guaranteeing consistency with the physical constraints of the micromechanical mathematical model.

[0037] In step S1 of this embodiment, a multi-scale multiphysics model is established by extracting the geometric parameters and micro-fractal features of the RF switch. The theoretical values ​​of the mechanical response and nonlinear contact conductivity are calculated, providing precise physical mechanism constraints for the neural network. Specifically, for example, the RF switch contact material is set to gold, and its Young's modulus... Poisson's ratio Material hardness The comprehensive elastic modulus is calculated based on the material parameters of both contacting materials. Set the standard deviation of the micro-protrusion height distribution. Equivalent radius of curvature of micro-convex body Substitute into the formula to calculate the plasticity index The calculation results show the plasticity index. This indicates significant plastic yielding at the contact interface, necessitating the correction of the actual microscopic contact area using an elastoplastic contact model. Assume that the actual microscopic contact area is obtained through a multiphysics joint solution at a certain macroscopic time step. The intrinsic resistivity of gold at a reference temperature is known. Temperature coefficient of material resistivity Assume the local temperature rise at the current node The area regularization constant to prevent mathematical singularities in a fully open-circuit state. Substituting the above physical quantities into the derivation formula of Holm's classical contact conductivity theory, the theoretical value of nonlinear contact conductivity is calculated. The numerical calculation process verifies the effectiveness and logical rigor of the physical model's output results under normal and extreme open-circuit conditions. The obtained data is resampled and mapped with the multi-physics distribution characteristics and the theoretical value of nonlinear contact conductivity. The generated physical field response feature library can accurately reflect the cross-coupling law under multiple environmental stresses.

[0038] Further, step S2 configures the underlying synchronization control link, collects high-frequency scattering parameters of the RF switch during actual operation according to a preset sampling period, executes a collection mechanism including multi-parameter joint delay compensation and timing fault tolerance management, and outputs a microwave RF real-time dataset. The specific execution flow is as follows: S21. Configure a high-speed field-programmable gate array (FPGA) hardware interface supporting parallel communication protocols to establish a low-level hard synchronization control link between the external vector network analyzer and the RF signal source, eliminating the time delay and communication jitter caused by traditional software protocol stack calls. The specific process is as follows: S211. Deploy a clock management unit within the field-programmable gate array (FPGA) to physically connect to an external reference clock provided by a high-precision cryogenic crystal oscillator. Utilize an internal phase-locked loop (PLL) to multiply and phase-align the reference clock, generating a global hardware system clock.

[0039] S212. Instantiate the intellectual property cores of communication protocols for general-purpose interface bus, gigabit LAN, and general-purpose serial bus within the hardware. Establish data interaction and parameter configuration channels, parse external control commands, and send them to external vector network analyzers and RF signal sources to complete the initial measurement parameter configuration.

[0040] S213. A multi-channel trigger signal generation module, directly driven by the global hardware system clock, is constructed at the hardware logic layer. After initial measurement parameter configuration, the multi-channel trigger signal generation module, based on the global clock tick, synchronously sends hardware trigger pulses with consistent rising edge polarity directly to external test terminals via low-impedance coaxial cables. This pulse bypasses the software operating system of the terminal device, forcing all test terminals to strictly synchronize signal generation and feature acquisition actions on the same hardware clock edge, establishing a low-level hard synchronization control link resistant to delay jitter.

[0041] S22. Execute a timestamp alignment mechanism based on the global hardware system clock and construct a real-time microwave RF dataset. Utilize the underlying logic resources of the field-programmable gate array (FPGA) to eliminate transmission delay errors and handle data overflow and frame loss anomalies in the actual operating environment. The specific process is as follows: S221. Deploy a high-frequency counter, directly driven by the global hardware system clock, within a field-programmable gate array (FPGA) for uninterrupted accumulation and timing. At the same physical rising edge of the hardware trigger pulse output by the multi-channel trigger signal generation module, a dedicated register synchronously latches the current count value of the high-frequency counter, generating an absolute timestamp corresponding to the current sampling period. Microwave network transmission theory is introduced to calculate the... Complex contact resistance of each connector Local group delay caused An equivalent lumped parameter series model of the connector contact interface is established, and the complex contact impedance is... Decomposed into real contact resistance characterizing contact degradation Equivalent inductance characterizing the parasitic effect of geometric gaps The characteristic impedance of the transmission line is constructed as follows: Phase angle function of forward transmission coefficients at time : ; For this phase angle function Along angular frequency Taking the negative derivative yields the local group delay for unidirectional transmission. The parsing expression: ; Based on physical length Relative permittivity function Temperature coefficient of dielectric constant of insulating medium Actual working environment temperature Reference calibration temperature The speed of light in a vacuum and the total number of connectors Calculate the inherent physical delay of radio frequency signals in the transmission line. : ; Add the absolute timestamp to the inherent physical delay. Generate valid sampling times mapped to the RF switch test reference plane. .

[0042] S222: Upon receiving a hardware trigger pulse, the vector network analyzer performs a single discrete frequency scan within the target monitoring angular frequency range, capturing the RF switch at the current valid sampling time. The complex scattering parameters are determined. The system analyzes the returned data stream, extracts the amplitude and phase information at each discrete frequency point, and compares it with the effective sampling time. Data packet reassembly and key-value binding are performed to construct a microwave RF real-time test matrix for a single working cycle.

[0043] S223, the underlying synchronous control link triggers feature acquisition actions cyclically according to a preset sampling period. The system allocates memory at a preset depth of... A continuous circular buffer block, with a preset depth Set to target stacking cycle number The buffer size is an integer multiple of the specified length, reserving timing redundancy for write and read operations to avoid data overwriting conflicts under normal operating conditions. A dynamic buffer management framework based on dual-pointer addressing and status flags is constructed, employing a fixed-length sliding window strategy for data updates. The system extracts the most recent... Microwave and radio frequency real-time test matrices for each complete sampling period are stacked in chronological order to form a microwave and radio frequency real-time dataset. In the formula The total number of frequency sampling points in a single frequency sweep. The number of characteristic channels for the scattering parameters.

[0044] To address the continuous acquisition process, a timing fault-tolerant compensation mechanism based on timestamp verification is introduced to extract the difference between the effective sampling times of adjacent sampling periods. When the difference Greater than the preset timing frame drop tolerance threshold When a data loss fault is detected, the system determines that a data loss fault has occurred. The system performs numerical reconstruction to fill in the missing complex scattering parameters along the time series dimension. For internal missing data points within the sequence that satisfy the condition of at least four consecutive known data points, the system will reconstruct the complex scattering parameters. Orthogonally decoupled along the complex plane, separating into independent orthogonal sequences of real parts. orthogonal imaginary part sequence For each orthogonal real part sequence orthogonal imaginary part sequence Construct a cubic spline interpolation function and set natural boundary conditions, that is, force the second derivative of the first and last points of the known continuous data sequence to be zero. Its mathematical constraint expression is: , In the formula, and These represent the start and end effective sampling times of the effective data window used for interpolation calculations. The system obtains the piecewise cubic polynomial coefficients by solving a tridiagonal linear equation system containing the natural boundary conditions, calculates the real and imaginary predicted values ​​at the reconstruction time, and then recombines them into reconstructed complex scattering parameters.

[0045] For cases where the data sequence has missing boundaries at the beginning and end, or for consecutive missing blocks located in the middle of the sequence with fewer than four valid data points on both sides, cubic spline interpolation degenerates into linear extrapolation or neighbor-filling strategies. When the missing segment is located at the beginning or end of the sequence, zero-order preservative filling is performed using the first or last adjacent valid data point; when it is located in the middle of the sequence with fewer than four valid data points, the order is reduced to linear interpolation. This multi-level interpolation reconstruction strategy ensures that the system can generate valid reconstructed values ​​under any missing mode, maintaining the real-time microwave RF dataset. The tensor dimension is constant, and a binary confidence mask tensor with the same dimension as the dataset is generated simultaneously. The interpolated reconstructed region is marked as a low-confidence state. The confidence mask tensor is used to constrain the gradient backpropagation of the subsequent neural network model, avoiding overfitting of the network to the virtual values ​​of the interpolated reconstructed data.

[0046] In step S2 of this embodiment, by configuring the underlying hard synchronization control link and calculating the inherent delay in conjunction with the physical parameters of the RF transmission line, absolute time alignment is achieved during the high-frequency scattering parameter acquisition process, thereby eliminating sequence misalignment caused by test system delay errors and environmental drift. Specifically, for example, the characteristic impedance of the transmission line where the RF switch test is located is set. At the target monitoring frequency angular frequency below Extract the current total number of connectors. Let the real contact resistance of the local connector in the test link characterize contact degradation. Equivalent inductance characterizing the parasitic effect of geometric gaps Substituting the local group delay into the local group delay analytical expression, calculate the local group delay for unidirectional transmission. Set the physical length of the test link. The relative permittivity of the insulating medium in this frequency band The speed of light in a vacuum Temperature coefficient of dielectric constant Read the current actual operating environment temperature. Reference calibration temperature Substituting the above variables into the inherent delay calculation formula, the inherent physical delay of radio frequency signal transmission can be solved. The calculation process quantifies the specific impact of local contact interface high-frequency impedance variation and global dielectric layer thermal physical drift on RF group delay. The output inherent physical delay is superimposed on the global timestamp to generate the effective sampling time, improving the time alignment accuracy of the feature acquisition process. Based on the data packets reconstructed by this absolute timestamp, a microwave RF real-time dataset is constructed to maintain the physical consistency of the time series data and ensure the application of the physical model in the microwave network transmission and clock synchronization compensation stage.

[0047] Furthermore, such as Figure 2 As shown, step S3 constructs and trains a decoupled convolutional neural network model constrained by nonlinear physical mechanisms. A deep learning architecture is used to fuse simulation prior data with real-time test signals at the feature level. S31. Construct a decoupled convolutional neural network model with a dual-channel feature input structure, a cross-attention fusion module, and a multi-head output layer. The decoupled convolutional neural network model includes a physical field feature extraction branch and a real-time radio frequency feature extraction branch. The specific process is as follows: S311. Transfer the physical field response feature library generated in step S1. The input physics feature extraction branch performs 3D spatial feature extraction, including the height parameter. The value is 64, which is the width parameter. The value is 64, which is the depth parameter. The value is 32, representing the number of physical feature channels. A value of 4 is used to establish the resolution of electromagnetic and mechanical features of the microscopic contact surface and to limit the computational complexity and memory overhead during forward propagation. The physics feature extraction branch integrates a series-configured three-layer 3D convolutional layer, a batch normalization layer, and a linear rectified activation function module. The number of feature channels in the first 3D convolutional layer is determined by… Increased to 16, its 3D spatial convolution kernel size is The stride is 1, and the padding is 1; the number of feature channels in the second 3D convolutional layer increases from 16 to 32, and its 3D spatial convolutional kernel size is... The stride is 2 and the padding is 1, used to perform spatial downsampling; the number of feature channels in the third 3D convolutional layer increases from 32 to 64, and its 3D spatial convolutional kernel size is... With a stride of 1 and padding of 1, each 3D convolutional layer is used to extract thermo-electric field coupling features in a 3D coordinate system. Further downsampling is performed after a max-pooling layer, with the max-pooling layer having a window size of [missing value]. With a step size of 2, it outputs a spatial feature tensor with consistent dimensions through spatial tensor reshaping and a linear mapping layer. In the formula, the length of the spatial feature sequence after downsampling is... The value is 512, representing the dimension of the hidden layer feature vector. The value is 128.

[0048] S312, Transfer the microwave radio frequency real-time dataset generated in step S2. The input real-time radio frequency (RF) feature extraction branch performs one-dimensional temporal causal convolutional feature extraction. This branch integrates four layers of one-dimensional causal convolutional layers with dilation rates. The number of feature channels in each layer is uniformly set to 128, the kernel size is 3, and the stride is 1. The dilation rates of the first to fourth layers are configured as 1, 2, 4, and 8, respectively, achieving an effective receptive field size of 31 to fully cover the dynamic fluctuation features within the temporal sampling period. One-dimensional causal convolution with dilation rates is performed along a sliding window in the time dimension to capture the long-range nonlinear dependence of frequency domain scattering spectral lines on time-varying fluctuations. After passing through a feature tensor linear mapping layer, the output is a microwave RF feature tensor. In the formula, the length of the time-series feature sequence With the number of sampling periods Consistent.

[0049] S32. Deep joint fusion of dual-branch feature tensors is performed using a cross-attention fusion module. (This refers to the physics spatial feature tensor.) With microwave radio frequency characteristic tensor The transformation is performed through parallel linear orthogonal projection matrix mapping layers. Each linear orthogonal projection matrix mapping layer has 128 input neurons and 128 output neurons. The physical field space feature tensor after the mapping transformation is extracted as the key matrix. AND-value matrix The microwave radio frequency feature tensor after mapping transformation is extracted as the query matrix. The cross-attention mechanism uses 8 attention heads, with each attention head having a sub-feature dimension of 16. The query matrix is ​​then calculated. AND key matrix The scaled dot product of the transpose generates a spatiotemporal cross-attention weight matrix that characterizes the interaction between spatial physical degradation and radio frequency time-series signals. The calculation expression is: ; In the formula, To prevent the dot product value from being too large and causing the gradient to vanish during backpropagation, a dimensionless scaling factor is set to 128. This is a row-normalized exponential activation operator. The spatiotemporal cross-attention weight matrix is... Acting on the value matrix The dual-modal feature joint projection is completed, and the output dimension is obtained through the residual connection network and the layer normalization module. The high-dimensional fusion feature tensor. When the microwave RF input terminal encounters strong external electromagnetic pulse interference, causing the tensor data to exhibit a sudden all-zero dead zone state, the denominator term at the bottom of the normalization exponential activation operator forces the sum of the attention weight coefficients in the probability space to be always equal to one, ensuring the continuous extraction of effective features and maintaining the numerical stability of the neural network gradient propagation process.

[0050] S33. Integrating the preprocessed multiphysics simulation feature data with the actual microwave RF synchronization test signal sequence, and dividing it into a model training set, a cross-validation set, and a generalization test set in a 7:1.5:1.5 ratio according to the dynamic hierarchical sampling rule. Before performing gradient backpropagation, a two-way auxiliary reconstruction branch for physics and data is constructed within the decoupled convolutional neural network model. The hidden layer at the end of the physics feature extraction branch is mapped through a spatially fully connected layer and a linear activation layer, outputting a predicted value of local microscopic contact conductivity that corresponds perfectly to the number of grid nodes. The terminal hidden layer of the real-time radio frequency feature extraction branch reconstructs the mapping through one-dimensional deconvolution, outputting a complex scattering parameter prediction tensor. The weight parameter matrix of the neural network nodes is updated via backpropagation using a total loss function that includes constraints based on nonlinear physical mechanisms. Total loss function The direction deflection of the data fitting gradient is balanced by a dynamic weight adaptive allocation mechanism based on homoscedasticity uncertainty. The calculation expression is as follows: ; In the formula, and The penalty parameter is used to balance the uncertainty of the task; The loss scalar is the mean square error of the data fitting for the radio frequency scattering parameters. The loss scalar is the physical mechanism constraint term; A dimensionless decoupling penalty hyperparameter for controlling the orthogonality of decoupling branches; An orthogonal constraint decoupling loss is used to characterize the correlation of features in the two branches. A task uncertainty balance penalty parameter is set. and The initial values ​​are all As a learnable scalar parameter within the network, and The update method is the same as that of the weight parameter matrix inside the network. The partial derivatives are calculated using the chain rule during each backpropagation process. and Acquire gradients and iterate and update them synchronously with the main network structure. When As the number of terms approaches zero, the logarithmic terms As a regularization term, it generates a negative penalty to prevent gradient explosion caused by parameter updates.

[0051] Data fitting mean square error loss Combined with the binary confidence mask tensor generated in step S223 This is used to quantify the approximation error of the feature network reconstructing the scattering feature space. Binary confidence mask tensor. The tensor dimension is also In terms of feature channels and spatiotemporal dimensions, and the prediction tensor of complex scattering parameters and real data tensors To maintain strict consistency, the calculation expression is: ; In the formula, The backpropagation parameter iteration batch size is an abstract variable, which takes the value of ; This is the binary confidence mask tensor for the corresponding batch of samples; For the Hadama product operator; and These are the complex scattering parameter tensors for network forward propagation prediction and real-world synchronous acquisition, respectively.

[0052] Physical mechanism constraint terms The hidden layer feature vector space of a high-dimensional network is forced to approximate the strong coupling laws of real thermodynamic and electrical multiphysics fields. The calculation expression is as follows: ; In the formula, This is an abstract value representing the total number of sampling nodes in the finite element mesh of the space physics field. To extract the local microscopic contact conductivity prediction value of the branch output for physical field feature extraction; The theoretical contact conductivity label value is obtained in step S13 by pre-calculating the model of microscopic contact mechanics and classical contact conductivity theory. This label value is calculated once before network training and remains fixed during training.

[0053] The adaptive moment estimation network with a second-order momentum term is configured to perform gradient descent updates. The initial learning rate for network training is set to... The minimum learning rate is set to Cosine annealing cycle Set as The network learning rate hyperparameter is adjusted every training epoch. The global maximum number of training epochs is set to [value missing]. In this round, an early stopping strategy based on the cross-validation set is introduced: at the end of each training round, the total loss of the model on the cross-validation set is calculated; if the total loss on the cross-validation set falls below the threshold value in consecutive training rounds, an early stopping strategy is implemented. If no numerical decrease occurs within a training epoch, the model is determined to have reached its convergent extremum or begun to overfit. The system terminates the training process early and outputs a network weight matrix with the minimum cross-validation loss. The total loss function is minimized using a gradient algorithm. This enables neural networks to possess both rigorous data and physical interpretability when mining deep implicit features.

[0054] In step S3 of this embodiment, the physical field and microwave / RF dual-modal features are extracted by decoupling the convolutional neural network model. A cross-attention mechanism is then used to jointly map spatial and temporal features. Subsequently, a total loss function with a dynamic weight allocation mechanism is used for backpropagation updates, forcing the network weights to converge to the actual physical degradation law. Specifically, for example, the number of local finite element mesh nodes in the physical field response feature library is selected. Perform physical mechanism constraint calculations. Extract the theoretical contact conductivity label value obtained in step S1. And assume the label value of the adjacent node In the current forward propagation batch, the predicted local microscopic contact conductivity values ​​output by the physics feature extraction branch-assisted reconstruction branch are as follows: and Substituting the physical mechanism constraint terms into the loss scalar formula for calculation. Let the loss scalar be the mean square error of the data fitting of the radio frequency scattering parameters statistically analyzed by the system within the same training round. Orthogonal constraint decoupling loss Set the uncertainty balance penalty parameters for the current task. and Dimensionless decoupling penalty hyperparameter controlling the orthogonality of decoupling branches Substitute the above variables into the total loss function formula to calculate the total loss required for backpropagation. (The internal logarithmic terms are approximated using the natural logarithm): The numerical calculation process quantifies the residual constraint effect of theoretical physical characteristics on the hidden layer parameters of the network. By introducing an uncertainty adaptive allocation mechanism, it effectively balances the data fitting gradient and the physical mechanism constraint gradient. The output high-dimensional fusion feature tensor enables the discrete time-series radio frequency features and physical field response features to form a cross-fusion constraint, maintaining the physical interpretability of the neural network feature extraction process and avoiding feature omissions caused by the model under dynamically changing electromagnetic interference.

[0055] Finally, S4 takes the high-dimensional fusion feature tensor as input, which includes a multi-branch output architecture with regression and classification branches, performs orthogonal decoupling constraints, and outputs a continuous disturbance feature vector and a discrete classification probability vector. For example... Figure 3 As shown, the execution flow is as follows: S41. At the end of the decoupled convolutional neural network model, construct a multi-branch decoupled output architecture containing regression and classification output branches. Perform global average pooling along the time series dimension on the generated high-dimensional fused feature tensor to eliminate the feature dimension misalignment caused by dynamic fluctuations in the real-time RF signal sampling length, compressing and mapping it into a fixed-dimensional global fused feature vector. Globally fuse feature vectors The shared input features are fed in parallel to the regression and classification output branches for multi-dimensional state feature decoding. The regression output branch consists of cascaded multilayer perceptrons, containing alternating linear fully connected layers, Gaussian error linear unit activation modules, and randomly deactivated layers. Globally fused feature vectors are then generated. After the nonlinear mapping of the regression output branch, the output includes a continuous interference feature vector of the quantification index of insertion loss drift and the quantification index of isolation degradation. The calculation expression is: ; In the formula, and These are the hidden layer weight matrix and bias vector; The Gaussian error linear unit activation operator; and This represents the weight matrix and bias vector of the regression output layer. Continuous disturbance eigenvectors. Each internal feature element is mapped to a decibel value characterizing the degree of degradation in the electromagnetic transmission performance of the RF channel. The classification output branch synchronously receives the globally fused feature vector. By constructing fully connected layers with independent weights to linearly map mechanical degradation features, and using a normalized exponential activation function to output a discrete classification probability vector representing the probability of mechanical failure due to micro-deformation of the moving reed, the mechanical degradation features are linearly mapped. The calculation expression is: ; In the formula, and The output layer weight matrix and bias vector are used for classification. This is the normalized exponential activation operator. When the input global fusion feature vector... When the physical link of the RF switch is completely destroyed and degrades to a zero vector, the base operation of the normalized exponential activation function constrains the discrete classification probability vector of the classification output branch. It exhibits a uniform distribution, avoiding the generation of negative or infinitely large singular values, and ensuring that the sum of the internal multidimensional fault probability components is always equal to one.

[0056] S42. The orthogonalization constraint of the feature vectors and the multi-dimensional interference collaborative suppression control are implemented to remove nonlinear entanglement between different feature spaces during network training and to drive external hardware to perform adaptive compensation actions during model deployment and inference. The specific execution flow is as follows: S421. During the model iterative training phase, extract the continuous disturbance feature tensor matrix and the discrete classification probability tensor matrix generated within the same training batch. Perform zero-mean and unit variance normalization along the batch dimension on the continuous disturbance feature tensor matrix and the discrete classification probability tensor matrix to generate a standardized continuous feature matrix. With the standardized discrete characteristic matrix The matrix product of the transpose of the standardized continuous feature matrix and the standardized discrete feature matrix is ​​calculated to generate a cross-modal covariance matrix representing the correlation of the bi-branch features. Orthogonal constraint decoupling loss function is adopted. The penalty applied to the cross-modal covariance matrix is ​​calculated using the following expression: In the formula, The Frobenius norm operator is used. The orthogonal constraint decoupling loss function is minimized through backpropagation, thus forcing the cross-modal covariance matrix. All off-diagonal elements within the range approach zero, ensuring that the regression output branch and the classification output branch maintain orthogonality and independence when extracting multidimensional physical degradation features.

[0057] S422. After the model completes parameter convergence and is deployed on edge computing nodes, output the decoupled continuous disturbance feature vector in real time. With discrete classification probability vector The insertion loss drift contained within the analytical continuous disturbance eigenvector of the external dynamic impedance matching controller. Quantitative indicators of isolation degradation The multi-channel bias control voltage tensor is generated using the built-in impedance compensation mapping equation. The DC tuning voltage components of each control channel in the adjustable capacitor array of the microelectromechanical system are established, and their mathematical expressions are as follows: ; In the formula, For tuning control channel index; For the first The reference intrinsic bias voltage of each control channel under ideal matching conditions; and These are the linear calibration compensation coefficients corresponding to the quantitative indicators of insertion loss drift and isolation degradation, respectively. and These are the weight coefficients of the second-order nonlinear correction terms corresponding to the two degradation indices.

[0058] Among them, the compensation coefficient , , , and During the system's factory calibration phase, the overdetermined equations were solved using the least squares fitting method to determine the following at standard ambient temperature: And under conditions of no external electromagnetic interference, apply to the radio frequency front end A set of known independent discrete combinations of insertion loss drift and isolation degradation excitations ,in and Adjust and record the actual optimal bias voltage for each channel that minimizes the return loss of the RF channel. Construct a known activation matrix with activation combinations and their quadratic terms as elements. and the response vector with the actual optimal bias voltage as an element. The known excitation matrix The row vectors are of the form Solving for the parameter vector using the least squares criterion. The calculation formula is: For mass-produced RF switch modules of the same model, parameter dispersion is controlled within a preset dispersion threshold. When within tolerance, they share the same set of factory calibration parameters; out-of-tolerance devices are individually calibrated and burned into the onboard electrically erasable programmable read-only memory.

[0059] A multi-channel bias control voltage tensor is applied to the drive pins of the adjustable capacitor array in the microelectromechanical system (MEMS) to dynamically compensate for RF channel impedance mismatch. Simultaneously, a hardware fault-tolerant actuator extracts the maximum confidence probability variable representing the mechanical fault of the micro-deformation of the moving reed from the discrete classification probability vector. The mechanical safety limit probability threshold stored in the register. Perform real-time comparisons.

[0060] Among them, the mechanical safety limit probability threshold Based on the two-parameter Weibull distribution function, maximum likelihood estimation and fitting were performed on accelerated life fatigue test data to determine: the cumulative distribution function of mechanical fatigue failure of the RF switch physical prototype was established. The mathematical expression is: ; In the formula, The number of mechanical operation cycles of the dimensionless radio frequency switch; The scaling parameter characterizes the lifetime of a device when it reaches a cumulative failure probability of 63.2%, and its value range is set to [value range missing]. to ; The shape parameter characterizes the evolution rate of wear failure in mechanical structures. For the fatigue wear characteristics of metal contact springs in radio frequency microelectromechanical systems, the value range is set to [value range missing]. to The maximum likelihood estimation method is used to iteratively calculate the actual number of failure cycles of the accelerated life fatigue test samples to obtain... and The optimal statistical estimate. Set the minimum reliability index required for the system to maintain stable communication. According to the correspondence The preset mechanical safety limit probability threshold is calculated. Before a controlled physical material reaches nonlinear rapid plastic yielding, a preset mechanical safety limit probability threshold is established. Typical configuration is in to Interval.

[0061] When the maximum confidence probability variable exceeds the preset mechanical safety limit probability threshold When the hardware fault-tolerant actuator triggers the highest-priority interrupt request at the underlying level, the interrupt response delay of the hardware fault-tolerant actuator is controlled within a certain timeframe during the switching operation. Within this range, the RF solid-state switch completes gate level latching and channel switching, disconnects the currently damaged RF link, and switches the microwave signal transmission path to the backup RF switch branch. The digital phase shifter uses the pre-measured and stored primary / backup branch transmission delay difference in the system memory. Introducing transient phase compensation offset The transient phase compensation offset is applied synchronously along the same hardware clock edge as the RF solid-state switch state transition. To maintain the continuity of signal phase before and after path switching, the deviation of high-frequency scattering phase discontinuity abrupt change is controlled within the preset phase change tolerance. Within.

[0062] In step S4 of this embodiment, orthogonal decoupling is performed on the high-dimensional fused feature tensor through a multi-branch output architecture, simultaneously outputting continuous RF performance degradation indicators and discrete mechanical fault probabilities, thereby driving the underlying microelectromechanical system's adjustable capacitor array to perform voltage compensation and seamless switching of the RF solid-state switch. Specifically, for example, the insertion loss drift is obtained analytically from the continuous interference feature vector output by the regression branch. Quantitative indicators of isolation degradation Extracting the first th variable capacitor from the adjustable capacitor array of a microelectromechanical system. The preset parameters of each tuning control channel, and the reference intrinsic bias voltage under ideal matching conditions. Linear calibration compensation coefficient and Weighting coefficients of the second-order nonlinear correction term and Substituting the above values ​​into the impedance compensation mapping equation, the bias control voltage of this channel is calculated. Simultaneously, the system extracts the discrete classification probability vector output by the classification branch to obtain the maximum confidence probability variable characterizing the mechanical fault of the micro-deformation of the moving spring. Given a preset mechanical safety limit probability threshold. ,because The hardware fault-tolerant actuator determines that the limit threshold has been triggered and executes a channel switch. Set the target monitoring angular frequency. Transmission delay difference between primary and backup branches Substitute into the formula to calculate the transient phase compensation offset. This calculation process verifies the technical effectiveness of the decoupling model in multidimensional feature extraction and intervention control; the generated control voltage It can be precisely applied to the capacitor array to compensate for RF channel impedance mismatch, and the calculated transient phase compensation offset This ensures the continuity of microwave signal phase during switching of the backup branch, preventing the failure of the radio frequency system protection mechanism and the interruption of the communication link.

[0063] In the above embodiment 1, the solution domain space defined in step S11 provides a discrete grid for solving the multi-field partial differential equations in step S12. The electric field distribution matrix calculated in step S12 directly determines the Joule thermal power density distribution matrix, thereby deriving the mechanical stress deformation tensor. The mechanical stress deformation tensor and temperature rise data are passed as input to step S13 to derive the theoretical value of nonlinear contact conductivity, and are finally encapsulated into a physical field response feature library. The control link established in step S21 provides the underlying hardware path for step S22. The high-frequency scattering parameters collected in step S22 are stacked in time series to form a real-time microwave radio frequency dataset. Step S31 involves receiving the physical field response feature library from step S1 in parallel. With the real-time dataset from step S2 After the cross-attention mechanism in step S32, the discrete physical field tensor and the time-series radio frequency tensor are fused into a high-dimensional fused feature tensor. Step S33 uses the theoretical contact conductivity label calculated by simulation to apply residual constraints to the hidden layer prediction values. Step S4 receives this high-dimensional fused feature tensor and splits it into independent continuous interference feature vectors through orthogonal transformation. With discrete classification probability vector This completes the full data loop from original physical simulation and measurement to decoupled control characteristics.

[0064] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning, characterized in that, include: Establish a macro-microscale cross-scale geometric model of the radio frequency switch, perform multi-physics joint solution, and output a physical field response feature library containing theoretical contact conductivity; Configure the underlying synchronous control link to collect the high-frequency scattering parameters of the RF switch according to the preset sampling period and output the microwave RF real-time dataset; A decoupled convolutional neural network model is constructed, and the physical field response feature library is fused with the microwave radio frequency real-time dataset at the feature level to output a high-dimensional fused feature tensor. The high-dimensional fusion feature tensor input includes a regression branch and a classification branch. The multi-branch output architecture is subjected to orthogonal decoupling constraints, and the output is a continuous disturbance feature vector and a discrete classification probability vector.

2. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 1, characterized in that, The establishment of the macro-micro cross-scale geometric model of the radio frequency switch includes: Extract the three-dimensional structural dimensional parameters of the RF switch to generate a macroscopic continuous medium solution domain space; Discrete coordinate point clouds are generated using three-dimensional fractal functions to reconstruct micro-rough contact surfaces; A moving least squares method is introduced at the contact interface to construct a shape function interpolation mapping interface. The radius of the tight support domain is determined based on the average physical spacing of macroscopic grid nodes. The displacement interpolation weight is calculated by combining the weight function to construct a cross-scale displacement transfer matrix. The macroscopic displacement field of the macroscopic continuous medium solution domain is mapped to the microscopic deformation displacement of the microscopic rough contact surface through the cross-scale displacement transfer matrix.

3. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 2, characterized in that, The joint multiphysics solution includes: In the solution domain of macroscopic continuous medium, the wave equation, the unsteady heat transfer equation and the thermoelasticity equation are solved simultaneously to obtain the flattened matrix of mechanical stress-deformation tensor. A dual-loop architecture is adopted in the solution process, and the macroscopic time step of the heat transfer and mechanical solution is set to be a multiple of the microscopic time step of the electromagnetic field solution. The relative norm residuals of the temperature field and stress field are calculated. When the relative norm residuals are less than the preset convergence tolerance threshold, it is determined that the multiphysics distribution has reached steady-state equilibrium and the macroscopic time variable is advanced until the preset total duration of transient simulation is reached.

4. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 3, characterized in that, The output includes a physical field response feature library containing theoretical contact conductivity, including: The local normal contact stress at the contact interface is analyzed from the flattened matrix of the mechanical stress deformation tensor. The implicit force balance equation is solved based on the local normal contact stress to determine the equivalent separation distance; The actual microscopic contact area is calculated by combining the probability density function of the height distribution of the micro-protrusions with the equivalent separation distance. The theoretical value of nonlinear contact conductivity is determined based on the actual microscopic contact area. The multi-physics field distribution characteristics are resampled and mapped to the theoretical value of the nonlinear contact conductivity to generate the physical field response feature library.

5. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 1, characterized in that, The underlying synchronization control link configuration includes: A low-level hard synchronization control link to an external test terminal is established using a field-programmable gate array (FPGA) hardware interface. Connect to an external reference clock to generate a global hardware system clock; A multi-channel trigger signal generation module is constructed based on the global hardware system clock to synchronously send hardware trigger pulses to the external test terminal to drive feature capture.

6. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 5, characterized in that, The method for acquiring high-frequency scattering parameters of the RF switch and outputting a real-time microwave RF dataset includes: The current count value of the synchronous latched high-frequency counter is used to generate an absolute timestamp; Calculate the inherent physical delay of the radio frequency signal in the transmission line; The absolute timestamp is superimposed with the inherent physical delay to generate an effective sampling time, and the complex scattering parameters of the RF switch are captured at the effective sampling time; The amplitude and phase information are extracted and reassembled with the data packets at the effective sampling time. The data is then updated using a sliding window strategy to construct the microwave radio frequency real-time dataset.

7. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 1, characterized in that, The process of constructing a decoupled convolutional neural network model and outputting a high-dimensional fused feature tensor includes: The decoupled convolutional neural network model includes a physical field feature extraction branch, a real-time radio frequency feature extraction branch, and a cross-attention fusion module; The physical field response feature library is input into the physical field feature extraction branch to extract spatial features and output a spatial feature tensor. The microwave radio frequency real-time dataset is input into the real-time radio frequency feature extraction branch for time-series feature extraction, and the microwave radio frequency feature tensor is output. The cross-attention fusion module is used to calculate the cross-attention weight matrix between the spatial feature tensor and the microwave radio frequency feature tensor, and the high-dimensional fused feature tensor is output after feature-level fusion.

8. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 7, characterized in that, The construction of the decoupled convolutional neural network model also includes iterative training of the decoupled convolutional neural network model: During training, a total loss function including physical mechanism constraints is used to update parameters; an auxiliary reconstruction branch is constructed inside the decoupled convolutional neural network model to output the predicted value of local microscopic contact conductivity. The physical mechanism constraint term uses the theoretical contact conductivity in the physical field response feature library to perform residual constraints on the predicted value of the local microscopic contact conductivity.

9. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 1, characterized in that, The high-dimensional fusion feature tensor input includes a regression branch and a classification branch. An orthogonal decoupling constraint is applied to the multi-branch output architecture, resulting in a continuous disturbance feature vector and a discrete classification probability vector, which include: The high-dimensional fusion feature tensor is reduced in dimensionality and mapped to a global fusion feature vector. The globally fused feature vector is then fed into the regression and classification branches; The regression branch outputs a continuous interference feature vector that characterizes the transmission performance index of the radio frequency channel. The classification branch outputs a discrete classification probability vector that characterizes the probability of mechanical failure.

10. The method for identifying and suppressing radio frequency switch interference in complex electromagnetic environments based on spectrum analysis and deep learning according to claim 9, characterized in that, After outputting the continuous disturbance feature vector and the discrete classification probability vector, the following is also included: The transmission performance index in the continuous interference feature vector is analyzed to generate a multi-channel bias control voltage tensor, which is then applied to an external adjustable capacitor array to compensate for the RF channel impedance mismatch. Extract the fault probability variable from the discrete classification probability vector. When the fault probability variable exceeds a preset safety limit threshold, trigger an interrupt request to drive the radio frequency solid-state switch to perform channel switching and apply phase compensation.

Images