A phase-free near-field reconstruction method for high-power pulses based on neural networks
Reconstructing the near-field distribution in high-power pulse measurement through neural network model, solving the problems of slow probe movement speed and long response time in traditional measurement methods, achieving efficient and low-cost pulse field reconstruction, reducing measurement errors and system costs.
Patent Information
- Application Number
- CN202410120379.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-29
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-01-29
AI Technical Summary
High-power pulse measurement has difficulty in testing in near-field measurement, the probe movement speed is limited, the response time is long, and it is impossible to effectively capture narrow-band pulse signals. The system is susceptible to electromagnetic interference, the measurement error is large, and the traditional device is expensive.
Using a neural network-based method, by constructing a neural network model, the pulse field distribution is reconstructed using limited plane amplitude data, reducing data volume and computing resources, improving reconstruction accuracy, and reducing system requirements.
The pulse field distribution is realized through the amplitude data of two actual measurement planes, with fast training speed, less resource occupancy and low cost, and the measurement process is optimized and the working loss of the measurement system is reduced.
Smart Images

Figure CN117952181B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of electromagnetic technology, and in particular to a high-power pulse phase-free near-field reconstruction method based on a neural network. Background Art
[0002] High-power pulse measurement is an important technology in the field of electromagnetic engineering. It is mainly used to evaluate and analyze the performance of various electromagnetic wave sources, such as radar systems, communication equipment, antennas, radio frequency (RF) transmitters and other radio equipment. Accurate field measurements can reflect the actual performance of electromagnetic wave sources, thereby providing improvement directions for the design of wave sources. High-power pulse measurements are usually divided into near-field measurements and far-field measurements. In near-field measurements, the measuring equipment is relatively close to the electromagnetic wave source, which allows for a more detailed understanding of the characteristics of the equipment, such as radiation pattern, radiated power distribution, and polarization. Far-field measurements are used to evaluate the performance of equipment at long distances, usually focusing on aspects such as transmission power, beam width, and radiation pattern.
[0003] The pulse signal used in high-power pulse measurement has instantaneous high peak power and short pulse width, which is difficult to test in traditional near-field measurement: obtaining the pulse field distribution in a certain space requires the probe to scan on a large number of set planes to obtain the field distribution data. Due to the limited moving speed of the test probe, the high-power pulse needs to exist all the time, which may cause problems for system safety and cause damage to the measurement personnel. In addition, the response time of the traditional pulse signal measurement device is relatively long compared to the pulse time, and it cannot effectively capture the information of the narrowband pulse signal, making it difficult to measure the phase data. At the same time, the input power of the system is limited, and the sensitive devices in the measurement system are extremely susceptible to strong electromagnetic interference, which may be damaged during long-term measurement work, increasing the measurement error. If the high-power electromagnetic pulse signal in a certain spatial area is directly measured, the measurement device must have ultra-wideband characteristics and high stability, and the cost of such a system will be very high. Therefore, it is very important to recover the near-field distribution in the spatial area from the limited plane amplitude measurement data by combining certain physical and data-driven models.
[0004] With the development of artificial intelligence related technologies, machine learning and neural networks as a data-driven technology are gradually suitable for application in high-power pulse measurement and reconstruction to improve the efficiency of data processing and analysis, and use prior models for automated data deduction and noise suppression, further improving the prospects for research and practical applications in this field. Due to the fast convergence speed and high accuracy of neural networks, high-speed computing equipment can be used to achieve near-field amplitude reconstruction of high-power pulses through limited planar near-field amplitude data. Summary of the invention
[0005] The purpose of the present invention is to provide a high-power pulse phase-free near-field reconstruction method based on a neural network, which requires a small amount of data, has a fast calculation speed, occupies less computing resources, has low requirements on the measurement system, and has good reconstruction accuracy.
[0006] To achieve the above object, the present invention provides a high-power pulse phase-free near-field reconstruction method based on a neural network, comprising the following steps:
[0007] The following steps are involved:
[0008] Step 1, preprocessing: the measured phase-free amplitude data of surface #1 and surface #2 are used as data sources for resizing and forming a training set;
[0009] Step 2: Network construction: construct a neural network model;
[0010] Step 3, training: input the data into the trainer for training, set the training rounds, and extract the data of the phase modulation part in the network model when the training is stopped, which is the reconstructed phase of the #1 surface;
[0011] Step 4, reconstruction: Combine the measured amplitude and reconstructed phase of surface #1 to obtain the amplitude and phase reconstruction of surface #1, and calculate the full-space reconstructed field through the spatial propagation distance.
[0012] Preferably, the principle of constructing the neural network model in step 2 is: the neural network model is expressed as: f(x)=w T (x·b)
[0013] Where x = |x|e jΨ , |x| is the amplitude of the network input, Ψ is the phase of the network input, j is the imaginary unit, is the spatial propagation factor, b is the phase modulation factor. In this model function, the input is x and the output is f(x). Since the output is the value obtained by the network based on the input operation, it is called network prediction;
[0014] Network input: x = |x|e jΨ , |x| is the amplitude of surface #1, and Ψ is simply set to an initial value of all 0s, constituting the network input x;
[0015] Phase modulation: b is the phase modulation factor, which modulates the network input x and is also used as an optimization parameter in training, and is adjusted after each error backpropagation;
[0016] Spatial propagation: According to the RS formula, a spatial propagation factor value is calculated based on the spatial position of each two neuron connections;
[0017] Network prediction: The network input is the network prediction value obtained by the above model function, which is used for loss calculation with the loss function part;
[0018] Loss function: The loss is obtained according to the following loss function formula, which is used as the basis for adjusting the phase factor b in the trainer.
[0019] Preferably, in step 2, the connection of neurons is calculated according to the RS formula:
[0020]
[0021] Among them, x, y, z are spatial coordinate systems; j is an imaginary unit, z i is the z coordinate of the ith position;
[0022] The loss function is:
[0023]
[0024] Among them, f out is the value predicted by the network; T k is the magnitude of face #2, which is actually used as the training label.
[0025] Preferably, in step 3, when the current round is not greater than the target round, random gradient descent is performed, and the phase modulation, spatial propagation, network prediction and loss function in step 2 are repeated after adjusting the phase; and training is stopped when the current round is greater than the target round.
[0026] Preferably, the full-space reconstruction field in step 4 is to combine the measured amplitude and reconstructed phase of surface #1 to obtain the amplitude and phase reconstruction of surface #1, and to obtain the amplitude and phase distribution of any pulse near-field plane between the two test planes by calculating the spatial propagation distance.
[0027] Therefore, the present invention adopts the above-mentioned high-power pulse phase-free near-field reconstruction method based on neural network, which has the following beneficial effects:
[0028] (1) The pulse field distribution between the two planes can be reconstructed by only using the amplitude data of the two measured planes, without measuring the entire space or the phase.
[0029] (2) The neural network is based on the stochastic gradient descent method, which has a fast training speed, occupies less computing resources, and reconstructs field data with high accuracy.
[0030] (3) Low cost. The measurement system does not need to be measured for a long time, which saves the working loss of the measurement system and optimizes the measurement process.
[0031] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 It is a system block diagram of a reconstruction algorithm of an embodiment of a high-power pulse phase-free near-field reconstruction method based on a neural network of the present invention;
[0033] Figure 2 It is a schematic diagram of a neural network model of an embodiment of a high-power pulse phase-free near-field reconstruction method based on a neural network of the present invention;
[0034] Figure 3 It is an amplitude distribution diagram of reference surface #1 of a calculation example of an embodiment of a high-power pulse phase-free near-field reconstruction method based on a neural network of the present invention;
[0035] Figure 4 It is an amplitude distribution diagram of reference surface #2 of a calculation example of an embodiment of a high-power pulse phase-free near-field reconstruction method based on a neural network of the present invention;
[0036] Figure 5 It is a phase distribution diagram of surface #1 restored by an example of an embodiment of a high-power pulse phase-free near-field reconstruction method based on a neural network of the present invention;
[0037] Figure 6 It is an amplitude distribution diagram of the intermediate surface inferred by a calculation example of an embodiment of a high-power pulse phase-free near-field reconstruction method based on a neural network of the present invention;
[0038] Figure 7 It is a loss function descent curve diagram of an example of an embodiment of a high-power pulse phase-free near-field reconstruction method based on a neural network of the present invention. DETAILED DESCRIPTION
[0039] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.
[0040] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.
[0041] Embodiment 1
[0042] like Figure 1 As shown, the present invention provides a high-power pulse phase-free near-field reconstruction method based on a neural network, introduces a neural network into the field of near-field measurement data recovery, constructs a calculation framework by building a neural network model based on diffraction theory, reconstructs the spatial amplitude and phase distribution of the field through the amplitude distribution of two surfaces combined with the optimization method of the neural network, and then reconstructs the amplitude distribution of any surface in the measurement surface, so as to achieve the purpose of inferring other data with very little measurement data. Specifically, it includes the following steps:
[0043] Step 1: Preprocessing: The measured phase-free amplitude data of surface #1 and surface #2 are used as data sources for resizing and form a training set.
[0044] The phase-free amplitude data of surface #1 and surface #2 are measured by near-field measurement equipment.
[0045] Step 2: Network construction: Build a neural network model.
[0046] The principle of constructing a neural network model is: the neural network model is expressed as: f(x) = w T (x·b)
[0047] Where x = |x|e jΨ , |x| is the amplitude of the network input, Ψ is the phase of the network input, j is the imaginary unit, is the spatial propagation factor, b is the phase modulation factor. In this model function, the input is x and the output is f(x). Since the output is the value obtained by the network based on the input operation, it is called network prediction;
[0048] Network input: x = |x|e jΨ , |x| is the amplitude of surface #1, and Ψ is simply set to an initial value of all 0s, constituting the network input x;
[0049] Phase modulation: b is the phase modulation factor, which modulates the network input x and is also used as an optimization parameter in training, and is adjusted after each error backpropagation;
[0050] Spatial propagation: According to the RS formula, a spatial propagation factor value is calculated based on the spatial position of each two neuron connections;
[0051] Network prediction: The network input is the network prediction value obtained by the above model function, which is used for loss calculation with the loss function part;
[0052] Loss function: The loss is obtained according to the following loss function formula, which is used as the basis for adjusting the phase factor b in the trainer.
[0053] Calculate the connections of neurons according to the RS formula:
[0054]
[0055] Among them, x, y, z are spatial coordinate systems; j is an imaginary unit, z i is the z coordinate of the ith position.
[0056] The loss function is:
[0057]
[0058] Among them, f out is the value predicted by the network; T k is the magnitude of face #2, which is actually used as the training label.
[0059] Step 3, training: input the data into the trainer for training, set the training rounds, and when the training is stopped, extract the data of the phase modulation part in the network model, which is the reconstructed phase of the #1 surface.
[0060] When the current round is not greater than the target round, random gradient descent is performed, and the phase modulation, spatial propagation, network prediction and loss function in step 2 are repeated after adjusting the phase; training is stopped when the current round is greater than the target round.
[0061] Step 4, reconstruction: Combine the measured amplitude and reconstructed phase of surface #1 to obtain the amplitude and phase reconstruction of surface #1, and calculate the full-space reconstructed field through the spatial propagation distance.
[0062] The full-space reconstructed field is obtained by combining the measured amplitude and reconstructed phase of surface #1 to obtain the amplitude and phase reconstruction of surface #1. The amplitude and phase distribution of any pulse near-field plane between the two test planes is calculated by the spatial propagation distance.
[0063] The amplitude distribution diagram of the example reference surface #1 and the amplitude distribution diagram of the example reference surface #2 are extracted from the high-power pulse sample data provided by the electromagnetic simulation software, as shown in Figure 3 and Figure 4 The high-power pulse sample data provided by the electromagnetic simulation software is reconstructed using the phase-free near-field reconstruction method provided by the present invention to obtain the phase distribution diagram of surface #1 restored in the example, as shown in FIG. Figure 5 As shown, from Figure 5 It can be seen that the phase distribution of surface #1 obtained by training is shown in the figure, which verifies that the amplitude phase of surface #1 can be propagated to obtain the amplitude of surface #2, indicating that the phase data can be effectively reconstructed using the high-power pulse phase-free near-field reconstruction method based on neural network.
[0064] The high-power pulse sample data provided by the electromagnetic simulation software is reconstructed using the phase-free near-field reconstruction method to obtain the amplitude distribution diagram of the intermediate surface inferred by the example, as shown in Figure 6 As shown, from Figure 6 It can be seen that the amplitude of the intermediate surface between surface #1 and surface #2 can be inferred based on the trained phase and amplitude distribution of surface #1, and the inferred amplitude distribution basically conforms to the propagation characteristics, indicating that the high-power pulse phase-free near-field reconstruction method based on neural network can effectively reconstruct the full space field.
[0065] The loss function descent curve of the example is obtained by recording the training process executed by programming, such as Figure 7 As shown, from Figure 7It can be seen that after 1000 full trainings, the loss function is minimized and the model converges to sufficient accuracy. During the programming execution, the calculation time of the example is 60s.
[0066] Therefore, the present invention adopts the above-mentioned high-power pulse phase-free near-field reconstruction method based on neural network, which requires less data, has fast calculation speed, occupies less computing resources, has low requirements on the measurement system, and has good reconstruction accuracy.
[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.
Claims
1. A high-power pulse phase-free near-field reconstruction method based on neural network, characterized by: The following steps are involved: Step 1, preprocessing: the measured phase-free amplitude data of surface #1 and surface #2 are used as data sources for resizing and forming a training set; Step 2: Network construction: construct a neural network model; The principle of constructing a neural network model is: the neural network model is expressed as: f(x) = w T (x·b) Where x = |x|e jΨ , |x| is the amplitude of the network input, Ψ is the phase of the network input, j is the imaginary unit, is the spatial propagation factor, b is the phase modulation factor. In this model function, the input is x and the output is f(x). Since the output is the value obtained by the network based on the input operation, it is called network prediction; Network input: x = |x|e jΨ , |x| is the amplitude of surface #1, and Ψ is simply set to an initial value of all 0s, constituting the network input x; Phase modulation: b is the phase modulation factor, which modulates the network input x and is also used as an optimization parameter in training, and is adjusted after each error backpropagation; Spatial propagation: According to the RS formula, a spatial propagation factor value is calculated based on the spatial position of each two neuron connections; Network prediction: The network input is the network prediction value obtained by the above model function, which is used for loss calculation with the loss function part; Loss function: The loss is obtained according to the following loss function formula, which is used as the basis for adjusting the phase factor b in the trainer; Calculate the connections of neurons according to the RS formula: Among them, x, y, z are spatial coordinate systems; j is an imaginary unit, z i is the z coordinate of the ith position; The loss function is: Among them, f out is the value predicted by the network; T k is the magnitude of face #2, which is actually used as the training label; Step 3, training: input the data into the trainer for training, set the training rounds, and extract the data of the phase modulation part in the network model when the training is stopped, which is the reconstructed phase of the #1 surface; Step 4, reconstruction: Combine the measured amplitude and reconstructed phase of surface #1 to obtain the amplitude and phase reconstruction of surface #1, and calculate the full-space reconstructed field through the spatial propagation distance.
2. The high-power pulse phase-free near-field reconstruction method based on a neural network according to claim 1, characterized in that: In step 3, when the current round is not greater than the target round, random gradient descent is performed, and the phase modulation, spatial propagation, network prediction and loss function in step 2 are repeated after adjusting the phase; and training is stopped when the current round is greater than the target round.
3. The high-power pulse phase-free near-field reconstruction method based on a neural network according to claim 1 is characterized in that: The full-space reconstruction field in step 4 is to combine the measured amplitude and reconstructed phase of surface #1 to obtain the amplitude and phase reconstruction of surface #1, and to obtain the amplitude and phase distribution of any pulse near-field plane between the two test planes by calculating the spatial propagation distance.
