FNO-based functional link resource autonomous matching test signal generation device and method
By using an FNO-based functional link resource autonomous matching test signal generator, the problems of fidelity and scene adaptability of traditional electromagnetic signal simulation methods have been solved, achieving autonomous learning and efficient electromagnetic signal simulation.
Patent Information
- Application Number
- CN202410994986.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-24
- Publication Date
- 2024-11-19
AI Technical Summary
Traditional electromagnetic signal simulation methods are limited by ideal mathematical formulas and parameter sets, making it impossible to simulate real electromagnetic environment signals with high fidelity. Furthermore, they rely on the experience of testers and are difficult to independently cope with changes in signal environment and multi-scenario matching.
A functional link resource autonomous matching test signal generator based on FNO is adopted, including feature preprocessing, feature generalization learning and signal feature fusion and training modules. It achieves autonomous learning and matching by adaptively processing signal features through Fourier neural operators.
It achieves high-fidelity simulation of complex electromagnetic signals, eliminating the reliance on manual operation, improving simulation efficiency and accuracy, and adapting to various application scenarios.
Smart Images

Figure CN118984198A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of electronic testing, and in particular relates to a device and method for generating a test signal based on an autonomous matching test signal of a functional link resource based on FNO. Background Art
[0002] The main method of generating complex electromagnetic signals is to simulate the style of electromagnetic signals and generate data based on digital domain simulation technology, and then generate electromagnetic signals through digital-to-analog conversion and I / Q modulation. The traditional solution is to set a series of general signal parameters based on the ideal mathematical expression of electromagnetic signals to meet the signal requirements under specific conditions or scenarios, and then generate simulation data under a fixed architecture. The generated signal is an approximation of the real signal under ideal conditions. However, the actual electromagnetic application scenarios are complex and diverse, and various types of signal aliasing exist and signal parameters change in real time. The aforementioned method is limited by ideal mathematical formulas and parameter sets. The generated electromagnetic signals are different from the actual scene signals and cannot meet the needs of high-fidelity simulation of complex electromagnetic environment signals. In addition, the generation of relevant data and the operation of the signal generator are highly dependent on the experience and proficiency of the tester, which limits the ability of the signal generator to quickly perceive environmental changes and simulate signals in real time in different application scenarios. In recent years, signal modeling methods based on deep learning have used signal data and neural networks to construct a framework different from traditional mathematical models. They can adaptively extract parameter characteristics and data characteristics of signals to achieve tasks such as signal classification, recognition, and noise reduction. At the same time, they can also simulate some electromagnetic signals with low complexity. However, their ability to identify and predict unknown complex signals and deduce future trends of signals is relatively weak.
[0003] Complex electromagnetic signal simulation technology is widely used to generate various types of electromagnetic signals and is one of the key technologies in the fields of radar, navigation and communication. The combination of a software simulation system based on a digital complex electromagnetic environment data information database and a complex signal generation hardware platform (such as a signal generator) can generate ideal complex style signals for equipment performance testing. Selecting an appropriate signal style generation model to generate signal data, then performing digital-to-analog conversion and I / Q modulation, and finally using a signal generator to generate electromagnetic signals is the main way to generate complex style electromagnetic signals. The electromagnetic signal generation model based on ideal mathematical expressions and the deep neural network framework based on big data can generate simulated signal data that meets specific conditions or scenarios, which is an approximation of the actual signal under ideal conditions.
[0004] At present, the simulation of electromagnetic signals is usually based on ideal mathematical expressions, and various parameters are preset to generate simulation data. For example, the following sinusoidal expression
[0005]
[0006] Used to generate RF signal, where A i represents the amplitude of the signal component, ω i Represents the frequency and Represents the phase of the signal component. This technology is an approximation of the real signal under ideal assumptions. Taking the simulation of electromagnetic radiation signal as an example, the user generates waveform data through various parameters to simulate and generate the signal. The principle is as follows Figure 1 As shown. Users can set relevant parameters on the control panel according to their needs. The software parses the set data set, compiles and generates electromagnetic radiation signals of specified types and parameters, and stores them in the system storage space in the form of files for users to extract and call. In addition, the generated waveform data can also be directly transmitted to the signal generator through LAN communication to achieve RF output. It can be seen that this scheme is based on the ideal mathematical formula of a specific type of electromagnetic signal to simulate and generate simulation data, which is an approximation of the real signal under ideal assumptions. This also leads to a large error between the simulated electromagnetic radiation signal and the real space and is inconvenient to evaluate, and cannot meet the relevant requirements of electromagnetic signal simulation. The types of parameters set through the panel are limited and fixed. If you want to generate other types of excitation signals, you need to repeat the entire test process, which depends entirely on the tester's calculation and modeling. This method is objectively limited by the fixed mechanism and parameter set of the ideal mathematical formula, and is also largely affected by the knowledge level and test experience of the tester, which brings hidden dangers to the scientificity and efficiency of related testing work.
[0007] The main disadvantages of existing technologies include:
[0008] 1. The traditional electromagnetic signal simulation method based on ideal mathematical formulas is objectively limited by fixed formula mechanisms and parameter sets, and cannot simulate actual electromagnetic environment signals with high fidelity.
[0009] 2. Traditional electromagnetic signal simulation methods are highly dependent on the tester's experience and knowledge and instrument operation level, and are unable to make autonomous corrections to cope with changes in the signal environment, limiting the efficiency of signal simulation.
[0010] 3. Changes in actual application scenarios require the generation of signals with different characteristics. The modeling method based on ideal mathematical expressions and conventional neural networks cannot meet the simulation requirements of the evolution of environmental parameters, and it is difficult to autonomously match different application scenarios. Summary of the invention
[0011] In view of the above-mentioned technical problems existing in the prior art, the present invention proposes a functional link resource autonomous matching test signal generating device and method based on FNO (Fourier Neural Operator), which has a reasonable design, overcomes the shortcomings of the prior art, and has good effects.
[0012] In order to achieve the above object, the present invention adopts the following technical solution:
[0013] A functional link resource autonomous matching test signal generating device based on FNO, including a feature preprocessing module, a feature generalization learning module and a signal feature fusion and training module;
[0014] A feature preprocessing module is configured to process user-defined signal features or factors affecting the generated signal and basic features of the collected signal;
[0015] The feature generalization learning module is configured to process the basic features of the preprocessing stage according to different categories and characteristics to form standardized feature data, and then use the Fourier layer to perform step-by-step generalization to autonomously construct the implicit space of signal features;
[0016] The signal feature fusion and training module is configured to perform affine transformation and scaling on the data in the latent feature space to adapt to the changing target signal.
[0017] In addition, the present invention also mentions a method for generating an autonomous matching test signal for a functional link resource based on FNO. The method adopts the above-mentioned device for generating an autonomous matching test signal for a functional link resource based on FNO, and specifically comprises the following steps:
[0018] Step 1: Process the user-defined signal features or factors affecting the generated signal and the basic features of the collected signal through the feature preprocessing module;
[0019] Step 2: Through the feature generalization learning module, the basic features of the preprocessing stage are processed according to different categories and characteristics to form standardized feature data, and then the Fourier layer is used to generalize step by step to autonomously construct the implicit space of signal features;
[0020] Step 3: Through the signal feature fusion and training module, the data in the latent feature space is affine transformed and scaled, and compared with the characteristics of the target signal to form a loss function, thereby achieving autonomous learning and perfect feature matching.
[0021] Preferably, step 1 specifically includes the following steps:
[0022] Step 1.1: Perform data cleaning on the customized style files and influencing factors;
[0023] Step 1.2: Segment and arrange the remaining characters;
[0024] Step 1.3: Digitally vectorize the divided words and put them into word bags to form part of the signal features; for the collected signals, directly extract their basic features and then import them into the feature library.
[0025] Preferably, step 2 specifically includes the following steps:
[0026] Step 2.1: Use a fully connected artificial neural network to process basic features;
[0027] Step 2.2: Use stacked Fourier layers to extract features step by step and generalize;
[0028] Step 2.3: Use a fully connected network to transform the generalized features into the target space, thereby implicitly constructing the input features.
[0029] Preferably, in step 2, each Fourier layer includes Fourier transform, convolution operation, inverse Fourier transform and a linear operation.
[0030] Beneficial technical effects brought by the present invention:
[0031] The present invention organically combines the broadband RF signal generation mechanism with the deductive operator FNO network algorithm, can adaptively process different input features, can explore and construct the logical relationship and spatial temporal relationship between feature data, and establish a feature-based signal generation mechanism. It is not restricted by the fixed signal generation mechanism and its parameter set based on the ideal mathematical formula, and realizes high-fidelity simulation of complex style electromagnetic signals.
[0032] The present invention realizes autonomous matching of functional link resources in multiple application scenarios, gets rid of the high dependence on manual operation and expert experience in the entire link of the test task, and can autonomously learn and correct according to the signal feature library of the actual environmental signal, thereby improving the simulation efficiency and accuracy of the environmental signal. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is the principle diagram of electromagnetic radiation signal simulation;
[0034] Figure 2 It is a principle block diagram of the device of the present invention;
[0035] Figure 3 Schematic diagram of Fourier-induced deep neural network. DETAILED DESCRIPTION
[0036] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0037] The present invention proposes a test excitation signal generation technology for autonomous matching of functional link resources based on the FNO deductive reasoning network, which organically combines the broadband RF signal generation mechanism with the deductive reasoning neural network algorithm, dynamically identifies and uses the signal parameter set, and realizes autonomous matching and situation deduction of signal generation. This solution breaks through the traditional mathematical formula signal generation architecture, and can not only adaptively extract and utilize signal parameter features and data features, but also process environmental parameters and user-specified signal styles as other input features, and then use Fourier transform and inverse transform operations to extract high-level features layer by layer, detect the logical relationship, spatial connection and temporal sequence between input data, so that the model has deductive reasoning capabilities, and finally autonomously generates matching simulation data according to the test scenario, thereby completing the RF generation of the signal, getting rid of the high dependence on human operation, and realizing high-precision simulation of complex electromagnetic signals and autonomous matching of multiple application scenarios.
[0038] Combining the FNO network for solving parameterized partial differential equations with the classical signal generation mechanism, the present invention proposes a test stimulus signal generation architecture based on the autonomous matching of functional link resources of the FNO deductive network. The model mainly consists of three parts: a feature preprocessing module, a feature generalization learning module with deductive reasoning function, and a signal feature fusion and training module. Figure 2 As shown. The feature preprocessing module is used to process user-defined signal features or factors affecting the generated signal and the basic features of the collected signal. The module first uses traditional technology to clean the customized style files and influencing factors, and then performs word segmentation and arrangement on the remaining characters. Finally, the segmented words are digitally vectorized and put into the word bag to form part of the signal features. In addition, for the collected signal, we can directly extract its basic features and then import them into the feature library. The feature generalization learning module based on Fourier operation processes the basic features of the preprocessing stage according to different categories and characteristics to form standardized feature data, and then uses the Fourier layer to generalize step by step to autonomously construct the implicit space of signal features; the signal feature fusion and training module performs affine stretching or compression according to the implicit space features, so that the signal generated by the signal generation component has different amplitude changes, phase changes and frequency change trends. The data generated by the model passes through the signal generation component to obtain the predicted signal, which is compared with the characteristics of the target signal to form a loss function, thereby realizing the autonomous learning and feature perfect matching of the device.
[0039] The generalization learning module is an operator deductive network with a stacked structure. The module first uses a fully connected artificial neural network to process the basic features, then uses stacked Fourier layers to extract and generalize features step by step, and finally uses a fully connected network to transform the generalized features to the target space, thereby realizing the implicit construction of the input features and providing the required refined input for the signal generation module. Generally speaking, the parameters used to generate signals have great differences in characteristics and scales. Magnitude normalization does not conform to the physical laws of signal generation and affects the performance of the model. At the same time, environmental parameters are also very different from signal parameters. To this end, we classify these parameters according to their properties (such as amplitude, frequency, phase, pulse, and others), and then use multiple simple networks to process these feature groups separately, increase the dimension of the input data, enrich its expression meaning, and expand the search space of the model. Furthermore, we apply the Fourier mapping activation function to the first hidden layer of these simple network models to better capture features with close similarity within the feature group and distinguish features with large similarity, such as Figure 3 shown.
[0040] Basic features with different properties are processed by multiple sub-neural networks to form new features and enter the Fourier layer. Each Fourier layer includes Fourier transform, convolution operation and inverse Fourier transform, as well as a linear operation. The results of the two paths are fused together, then acted upon by the activation function and finally output. Continuous Fourier layers will refine and extract features step by step, enhancing the expression and generalization capabilities of the model. Mathematically, the operator deduction network with an iterative structure is expressed as follows:
[0041]
[0042] in, and These are two shallow neural networks used to increase and reduce feature dimensions. Represents the operator operation. The Fourier layer implements the kernel integral operator, which has the ability to gather information around the point of interest. The relevant concepts and forms are described as follows.
[0043] Definition of nuclear operator: Nuclear integral operator on region D for
[0044]
[0045] in is a continuous kernel function, such as the Green kernel function k(s,t)=k(st) with global aggregation properties. The feature processing of continuous signals is a global operation, so the global kernel function is a good choice. Furthermore, we can use Fourier transform to implement the global convolution kernel.
[0046] Fourier kernel integral operator: Let and represent the Fourier transform and inverse Fourier transform of the continuous function f, respectively. Then for the continuous input variable x and kernel κ, we have the following expression
[0047]
[0048] The signal fusion generation and training module performs affine transformation and scaling on the data in the latent feature space to adapt to the changing target signal, especially the multi-tone composite signal. Finally, these data are played out through the signal generation component.
[0049] For a given training sample point After a series of processing of the above model, the corresponding prediction signal can be obtained Then, we can get the cost function of the model:
[0050]
[0051] in and Represent the network input and true signal label respectively, Represents the signal generated by the network, and Θ represents the set of all weights and bias parameters of the neural network. A regularization term to prevent overfitting of the neural network. β>0 is the weight sparseness of the sparsity factor, which can adjust the sparsity and overfitting state during training.
[0052] The cost function is continuously optimized through the stochastic gradient descent algorithm, so that the network embedding layer of the model can better encode the features of the input data and allow these inputs to exhibit their own unique properties. The Fourier layer processes and aggregates the logical relationships, spatial connections and time sequences between the input data step by step, transforms the features of the user input into the latent feature space, and provides suitable feature inputs for the signal generation component.
[0053] After iterative training, the network model can reasonably process the input features and automatically generate excitation signals for multiple application scenarios. Especially when the feature input is incomplete, the network can automatically construct the input of the signal generation module and heuristically generate the signal style expected by the user. In addition, some features of the complex electromagnetic signals generated by the adaptive network can also be used as the input of the network to further provide training data for the optimization of the network model, so that the network has self-learning ability and gradually improves the robustness and generalization performance of the model.
[0054] In summary, the test stimulus signal generation method based on the autonomous matching of functional link resources of the FNO operator deductive network proposed in the present invention realizes the following Figure 2The iterative optimization process from features to signals shown solves the problems of low simulation fidelity and single application scenarios caused by the high reliance of electromagnetic space construction on traditional ideal mathematical formulas, and achieves the electronic testing goals of "one specialty and multiple capabilities" and "scenario adaptability".
[0055] The invention provides a functional link resource autonomous matching test excitation signal simulation generation framework that organically combines a broadband radio frequency signal generation mechanism with a deductive reasoning network FNO.
[0056] In view of the great differences in signal parameters and environmental parameters, the present invention designs a multi-DNN channel feature embedding submodule for the FNO network module to perform feature scale separation and information smoothing, solve the problems of feature scale imbalance and information oscillation caused by a large number of different parameters, and enhance the ability of FNO to handle complex and changeable signal environment scenarios and signal parameters.
[0057] Aiming at the autonomous matching mechanism of signal generator function link resources, the present invention proposes an adaptive scale scaling strategy to generate signals with different amplitude scales and frequency scales, thereby improving the multi-scenario adaptive matching capability of the proposed network model.
[0058] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A device for generating a test signal for autonomously matching functional link resources based on FNO, characterized in that: It includes feature preprocessing module, feature generalization learning module and signal feature fusion and training module; A feature preprocessing module is configured to process user-defined signal features or factors affecting the generated signal and basic features of the collected signal; The feature generalization learning module is configured to process the basic features of the preprocessing stage according to different categories and characteristics to form standardized feature data, and then use the Fourier layer to perform step-by-step generalization to autonomously construct the implicit space of signal features; The signal feature fusion and training module is configured to perform affine transformation and scaling on the data in the latent feature space to adapt to the changing target signal.
2. A method for generating a functional link resource autonomous matching test signal based on FNO, characterized in that: The FNO-based functional link resource autonomous matching test signal generating device as claimed in claim 1 specifically comprises the following steps: Step 1: Process the user-defined signal features or factors affecting the generated signal and the basic features of the collected signal through the feature preprocessing module; Step 2: Through the feature generalization learning module, the basic features of the preprocessing stage are processed according to different categories and characteristics to form standardized feature data, and then the Fourier layer is used to generalize step by step to autonomously construct the implicit space of signal features; Step 3: Through the signal feature fusion and training module, the data in the latent feature space is affine transformed and scaled, and compared with the characteristics of the target signal to form a loss function, thereby achieving autonomous learning and perfect feature matching.
3. The method for generating a functional link resource autonomous matching test signal based on FNO according to claim 2, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Perform data cleaning on the customized style files and influencing factors; Step 1.2: Segment and arrange the remaining characters; Step 1.3: Digitally vectorize the divided words and put them into word bags to form part of the signal features; for the collected signals, directly extract their basic features and then import them into the feature library.
4. The method for generating a functional link resource autonomous matching test signal based on FNO according to claim 2, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Use a fully connected artificial neural network to process basic features; Step 2.2: Use stacked Fourier layers to extract features and generalize them step by step; Step 2.3: Use a fully connected network to transform the generalized features into the target space, thereby implicitly constructing the input features.
5. The method for generating a functional link resource autonomous matching test signal based on FNO according to claim 2, characterized in that: In step 2, each Fourier layer includes Fourier transform, convolution operation, inverse Fourier transform and a linear operation.
