An electromagnetic echo signal processing method based on programmable electromagnetic neural network
By introducing an electromagnetic neural network with a programmable plasmonic transmission line structure, the programmability and integration issues of electromagnetic neural networks in autonomous driving are solved, achieving low-latency and efficient obstacle recognition while reducing hardware costs and power consumption.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2026-05-11
- Publication Date
- 2026-06-05
AI Technical Summary
Existing electromagnetic neural network technology faces challenges in the field of autonomous driving, including difficulties in programmability, low integration, and a lack of real-time deployment capabilities. This results in high obstacle recognition latency, increased hardware costs, and higher power consumption.
An electromagnetic neural network with a programmable plasmonic transmission line structure is used to extract features of electromagnetic echo signals and identify obstacles through electromagnetic neural network encoding training at the transmitting and receiving ends, which simplifies the system structure and reduces latency and power consumption.
It achieves highly integrated electromagnetic echo signal processing, reduces recognition latency and system power consumption, and improves the efficiency and accuracy of obstacle recognition.
Smart Images
Figure CN122151029A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an electromagnetic echo signal processing method, and more particularly to an electromagnetic echo signal processing method for a programmable plasmonic electromagnetic neural network (SPNN, abbreviated as: programmable electromagnetic neural network) for obstacle recognition. Background Technology
[0002] In the fields of autonomous driving and assisted driving, low-latency and high-accuracy obstacle identification by sensor systems is crucial for ensuring vehicle safety. Currently, mainstream sensor solutions include optical imaging, LiDAR, and millimeter-wave radar. Optical imaging is a passive imaging method, relying on ambient light and susceptible to interference in complex lighting conditions. While LiDAR and millimeter-wave radar achieve environmental perception through active narrow-beam scanning, they heavily depend on complex backend algorithms to establish a mapping between data and the real environment, making it difficult to meet the requirements for low-latency response. Furthermore, existing technologies involve extensive analog-to-digital signal conversion in multi-channel acquisition systems, which not only increases hardware costs and power consumption but also significantly impacts the overall system response speed due to the latency introduced during the conversion process.
[0003] Electromagnetic neural networks offer a novel approach to electromagnetic signal processing. Leveraging the near-light-speed propagation of electromagnetic waves, they can perform computations similar to those of digital neural networks natively in the analog domain, thus avoiding complex analog-to-digital conversion and reducing latency. However, existing electromagnetic neural network technologies face challenges such as difficulty in achieving programmability, low integration, and a lack of real-time deployment capabilities, limiting their application in real-world autonomous driving scenarios. Therefore, there is an urgent need for an electromagnetic echo signal processing method that possesses high integration, programmability, and the ability to achieve high-speed and high-precision recognition.
[0004] Traditional vehicle-mounted perception systems (such as optical imaging and radar) rely on backend digital algorithms for obstacle recognition, which presents a contradiction between computing power and low latency requirements. Furthermore, the multi-channel analog-to-digital conversion process further increases system latency. Although existing electromagnetic neural networks possess physical layer signal processing capabilities, they are limited by poor programmability and low integration, making them difficult to apply in practice. Summary of the Invention
[0005] Technical Problem: The purpose of this invention is to provide an electromagnetic echo signal processing method based on a programmable electromagnetic neural network. By introducing a programmable plasmonic transmission line structure, the miniaturization and programmable integration of the electromagnetic neural network are realized, enabling the direct extraction of features from the electromagnetic echo to achieve efficient and low-latency obstacle recognition.
[0006] Technical solution: The electromagnetic echo signal processing method based on a programmable electromagnetic neural network disclosed in this invention includes the following steps: Step 101: Based on the programmable electromagnetic neural network hardware design, the transmitter electromagnetic neural network encoding is implemented to achieve sector scanning of the transmission beam, so as to realize spatial diversity of the acquired data and improve the sensing resolution; Step 102: Install the transmitting and receiving antenna arrays, and collect the echo dataset of the receiving antenna array under various obstacle scenarios to provide preliminary data for the training of the programmable electromagnetic neural network; Step 103: Design the electromagnetic neural network encoding and back-end processing matrix for the receiving end, and perform preliminary training on the programmable electromagnetic neural networks of the transmitting and receiving ends with the goal of classifying obstacle categories, and solidify the weights of the programmable electromagnetic neural network. Step 104: Collect the output dataset of the programmable electromagnetic neural network at the output end under actual working conditions to provide more refined data for the post-training of the electromagnetic echo signal processing system; Step 105: Optimize the backend processing matrix again to further improve the decision performance.
[0007] The processing method employs a system comprising: a transmitter SPNN positioned at the front of the vehicle, incorporating the electromagnetic neural network encoding from step 101; an adjusted radio frequency signal transmitted via the transmitter antenna array, achieving beam scanning; the transmitted electromagnetic signal, after being reflected by obstacles, is received by the receiver antenna array and transmitted via cable to the input port of the receiver module; within the module, the receiver SPNN performs analog neural network calculations in the electromagnetic domain, enabling feature extraction and preliminary training; subsequently, an incoherent detection stage is achieved through an intensity detection module, converting the signal into baseband components, which are then processed by the acquisition and decision module to execute the final simple classifier algorithm and decision.
[0008] The specific steps of step 101 include: The number of spatial scanning beams is determined based on the types of potential obstacles encountered in autonomous driving to implement spatial diversity, and the transmitting electromagnetic neural network is trained. Eleven beams are set, divided into three independent sectors: left, center, and right, with beam counts of 4, 3, and 4 respectively. Based on a programmable electromagnetic neural network (SPNN), the adjustable parameters of the network are optimized using a backpropagation algorithm. This SPNN consists of three cascaded propagation layers and three modulation phase-shifting layers, each with 32 ports. Input ports 5, 9, 13, and 17 are connected to a four-port power divider, which in turn connects to the center frequency signal source. The transmitting SPNN is represented as follows: , in These are the input and output vectors of the transmitting electromagnetic network. For the transmission matrix of the propagation layer, , , These represent the phase modulation matrices of the 1st, 2nd, and 3rd phase-shifting layers of the electromagnetic network at the transmitter; and the input vector of the electromagnetic network based on the power divider excitation mode. Set the specified port to 1 and the rest to 0; to establish the far-field model, the intensity of the 11 beam positions is calculated using the following formula; , Where I is the output port energy vector. This is the free-space far-field propagation matrix from the antenna array connected to the SPNN at the transmitter to the 11 sector corners. Indicates to Square the modulus value; define the beam scanning task loss value L1: , Where L1 is the beam scanning task loss value. Represents the target normalized beam intensity vector; the first beam is Based on formulas (1) to (3), the gradient descent method is used to minimize the beam scanning task loss value L1, thereby obtaining the optimal phase modulation matrix of the transmitter electromagnetic network. , , Eleven phase codes are generated for the eleven target beams, and then converted into control vectors through the phase-voltage mapping relationship of a programmable electromagnetic network for system use.
[0009] The specific steps of step 102 include: An electromagnetic echo detection system consisting of a transmitter SPNN and a receiver SPNN was established. The specific deployment included: connecting the input of the transmitter SPNN to a radio frequency signal source via a power divider, and connecting the output to the transmitter antenna array; mounting the entire system on the hood of the test vehicle, with the main lobe pointing in the direction of travel to detect obstacles; deploying the receiver antenna array at different locations on the vehicle body, with its output connected to the input port of the receiver SPNN; the receiver module integrating the receiver SPNN, intensity detection module, acquisition, and decision-making module; during the data acquisition phase, using a radio frequency matrix switch and a vector network analyzer, traversing 11 transmitting beams and different types of obstacle scenarios, collecting echo intensity data from each element of the receiver antenna array to form a training dataset.
[0010] The specific steps of step 103 include: based on the dataset collected in step 102, jointly training the SPNN parameters of the receiving end and the classifier transformation matrix in the acquisition and decision module, and the SPNN transmission model of the receiving end is shown in formula (4): , in The collected echo data vector, This is the output vector of the electromagnetic neural network at the receiving end. , , These are the phase modulation matrices of the 1st, 2nd, and 3rd phase-shifting layers of the electromagnetic neural network at the receiving end; the detection and linear classifier model is expressed as follows: , in: Output decision vectors for the linear classifier. For the weight matrix of the linear classifier, This is the bias vector for a linear classifier. The electromagnetic neural network outputs an intensity vector; The loss value L2 for the target classification task is defined as the distance between the true label and the output decision vector of the linear classifier, calculated using the cross-entropy function: , Where crossentropy represents the cross-entropy function, T is the true label vector, and softmax is the vector size of the target vector. The normalization function is converted into a normalized probability vector distribution; based on the dataset collected in step 102, the gradient descent method is used to minimize the loss value L2 of the target classification and recognition task to obtain the optimal value. , , With the linear classifier weight matrix and the bias vector of the linear classifier For the left, middle, and right sectors, three independent sets of receiver SPNN encoding and classifier parameters are trained and stored respectively.
[0011] Step 104 includes verification of the system's actual operating conditions and fine-tuning of the backend, specifically as follows: A complete electromagnetic signal identification system is constructed by directly coupling the receiving antenna array to the input of the receiving SPNN. Transmit / receive beam switching and data acquisition are performed according to a preset timing sequence, with timing synchronization provided by the main control module. To eliminate performance degradation caused by physical factors such as cable loss and antenna position deviation in actual deployment, backend adaptive fine-tuning is implemented: Under real-world conditions, based on the receiving SPNN phase parameters obtained in step 103, the intensity data output from the intensity detection module is acquired again according to the scanning process. The corresponding sample labels were then marked for future optimization.
[0012] Step 105 further optimizes the backend processing matrix, including obstacle recognition, and based on the dataset collected in step 104, optimizes the linear classifier weight matrices corresponding to the left, middle, and right sectors. and the bias vector of the linear classifier The system is then retrained and corrected to compensate for physical link errors. Subsequently, the fine-tuned classifier parameters are loaded to complete the final system deployment. Within a single scan cycle, the transmitting SPNN switches 11 sets of beam codes sequentially, while the receiving SPNN switches 3 sets of codes for each of the three sectors. The intensity detection module synchronously acquires 11 echo intensity features, and finally, the decision classifier in the acquisition and decision module performs three matrix multiplication and addition operations to output the obstacle category recognition result in real time.
[0013] Beneficial Effects: The electromagnetic echo signal processing method based on a programmable electromagnetic neural network for obstacle recognition provided by this invention has the following beneficial effects compared with the prior art: 1. By leveraging the analog domain processing capabilities of programmable electromagnetic neural networks, the recognition latency from signal acquisition to final result output is significantly reduced compared to digital computing systems.
[0014] 2. With the feature extraction capability of programmable electromagnetic neural networks, the number of channels at the acquisition end is much smaller than the number of antennas in the receiving array, and coherent detection is not required, greatly simplifying the transceiver system.
[0015] 3. Compared to the high power consumption requirements of digital networks, electromagnetic neural networks rely on the analog domain, with power consumption concentrated in the programmable components at the transceiver end. Simultaneously, the simplified receiver structure reduces the number of RF links, significantly lowering system power consumption. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments are briefly described below.
[0017] Figure 1 This is a flowchart of an electromagnetic echo signal processing method based on a programmable electromagnetic neural network provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the physical system structure of an electromagnetic echo signal processing method based on a programmable electromagnetic neural network provided in an embodiment of the present invention; the figure includes: vehicle 201, obstacle 202, transmitter SPNN 203, transmitter antenna array 204, receiver antenna array 205, receiver module 206, receiver SPNN 207, intensity detection module 208, and acquisition and decision module 209; Figure 3 This is a data processing flowchart of an electromagnetic echo signal processing method based on a programmable electromagnetic neural network according to an embodiment of the present invention. Detailed Implementation
[0018] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims. The invention will now be described in more detail with reference to the accompanying drawings.
[0019] Figure 2 A schematic diagram of the physical system structure is shown. The transmitting SPNN 203 injects the pre-trained beamcode from step 101. The adjusted radio frequency signal is transmitted through the transmitting antenna array 204, achieving beam scanning. The transmitted electromagnetic signal is reflected by obstacle 202 and received by the receiving antenna array 205. It is then transmitted via cable to the input port of the receiving module 206. Within the module, the receiving SPNN 207 performs simulated neural network calculations in the electromagnetic domain, realizing feature extraction and preliminary training. Subsequently, the intensity detection module 208 performs incoherent detection, converting the signal into baseband components, which are then executed by the acquisition and decision module 209 to perform the final simple classifier algorithm and decision.
[0020] Figure 3 The data processing flowchart is shown. The SPNN203 transmitter periodically injects pre-trained beamcoded signals. The system identifies the presence and type of obstacles in three sectors: left, center, and right, with each sector making independent decisions.
[0021] The left sector targets the scene on the left side of the vehicle, and the target recognition includes two categories: unobstructed objects and vehicles. The SPNN207 module at the receiving end is configured accordingly as a binary classification structure. The target recognition in the middle sector includes three categories: unobstructed objects, vehicles, and pedestrians. The SPNN module at the receiving end is configured with a three-class structure. The right sector identifies targets in two categories: unobstructed objects and pedestrians. The SPNN module at the receiver is configured as a binary classification structure.
[0022] Within each sector, different beams share the same receiver SPNN coding strategy, but different sectors employ differentiated coding. Finally, the system concatenates the multi-port output intensity values of the receiver SPNN for each beam into a one-dimensional vector, which is then input to the digital linear classifier to complete the final decision.
[0023] This invention discloses an electromagnetic echo signal processing method based on a programmable electromagnetic neural network, the specific method of which is as follows: Figure 1 As shown, it includes the following steps: Step 101: Based on the potential obstacle types for autonomous driving, determine the number of spatial scanning beams to implement spatial diversity and train the transmitting electromagnetic neural network; set 11 beams, divided into three independent sectors (left, center, and right), with beam numbers distributed as 4, 3, and 4 respectively. Based on the programmable electromagnetic neural network (SPNN), optimize the network's adjustable parameters using the backpropagation algorithm; the SPNN consists of three cascaded propagation layers and three modulation phase-shifting layers, each with 32 ports; input ports 5, 9, 13, and 17 are connected to a four-port power divider, which in turn connects to the center frequency signal source; the transmitting SPNN is represented as the following model: , in These are the input and output vectors of the transmitting electromagnetic network. For the transmission matrix of the propagation layer, , , These represent the phase modulation matrices of the 1st, 2nd, and 3rd phase-shifting layers of the electromagnetic network at the transmitter; and the input vector of the electromagnetic network based on the power divider excitation mode. Set the specified port to 1 and the rest to 0; to establish the far-field model, the intensity of the 11 beam positions is calculated using the following formula; , Where I is the output port energy vector. This is the free-space far-field propagation matrix from the antenna array connected to the SPNN at the transmitter to the 11 sector corners. Indicates to Square the modulus value; define the beam scanning task loss value L1: , Where L1 is the beam scanning task loss value. Represents the target normalized beam intensity vector; the first beam is Based on formulas (1) to (3), the gradient descent method is used to minimize the beam scanning task loss value L1, thereby obtaining the optimal phase modulation matrix of the transmitter electromagnetic network. , , Eleven phase codes are generated for the eleven target beams, and then converted into control vectors through the phase-voltage mapping relationship of a programmable electromagnetic network for system use.
[0024] Step 102: Establish an electromagnetic echo detection system consisting of a transmitter SPNN and a receiver SPNN. The specific deployment includes: connecting the input of the transmitter SPNN203 to a radio frequency signal source via a power divider, and connecting the output to the transmitter antenna array 204. The entire system is installed on the hood of the test vehicle, with the main lobe pointing in the direction of travel to detect obstacles; deploying the receiver antenna array 205 at different locations on the vehicle body, with its output connected to the input port of the receiver SPNN207; the receiver module 206 integrates the receiver SPNN207, the intensity detection module 208, and the acquisition and decision module 209; during the data acquisition phase, using a radio frequency matrix switch and a vector network analyzer, the system traverses 11 transmitting beams and different types of obstacle scenarios, collecting echo intensity data of each element of the receiver antenna array to form a training dataset.
[0025] Step 103: Based on the dataset collected in Step 102, jointly train the parameters of the SPNN207 receiver and the classifier transformation matrix in the acquisition and decision module 209. The SPNN207 transmission model at the receiver is shown in Equation (4): , in The collected echo data vector, This is the output vector of the electromagnetic neural network at the receiving end. , , These are the phase modulation matrices of the 1st, 2nd, and 3rd phase-shifting layers of the electromagnetic neural network at the receiving end; the detection and linear classifier model is expressed as follows: , in: Output decision vectors for the linear classifier. For the weight matrix of the linear classifier, This is the bias vector for a linear classifier. The electromagnetic neural network outputs an intensity vector; The loss value L2 for the target classification task is defined as the distance between the true label and the output decision vector of the linear classifier, calculated using the cross-entropy function: , Where crossentropy represents the cross-entropy function, T is the true label vector, and softmax is the vector size of the target vector. The normalization function is converted into a normalized probability vector distribution; based on the dataset collected in step 102, the gradient descent method is used to minimize the loss value L2 of the target classification and recognition task to obtain the optimal value. , , With the linear classifier weight matrix and the bias vector of the linear classifier For the left, middle, and right sectors, three independent sets of receiver SPNN encoding and classifier parameters are trained and stored respectively.
[0026] Step 104: Build a complete electromagnetic signal identification system. Directly couple the receiving antenna array to the input of the receiving SPNN207. Perform transmit / receive beam switching and data acquisition according to the preset timing sequence. Timing synchronization is provided by the main control module. To eliminate performance degradation caused by physical factors such as cable loss and antenna position deviation in actual deployment, perform backend adaptive fine-tuning: Under real-world conditions, based on the receiving SPNN phase parameters obtained in Step 103, re-acquire the intensity data output by the intensity detection module 208 according to the scanning process. The corresponding sample labels were then marked for future optimization.
[0027] Step 105: Further optimize the backend processing matrix, including obstacle recognition. Based on the dataset collected in Step 104, optimize the linear classifier weight matrices corresponding to the left, middle, and right sectors. and the bias vector of the linear classifier The system is then retrained and corrected to compensate for physical link errors. Subsequently, the fine-tuned classifier parameters are loaded to complete the final system deployment. Within a single scan cycle, the transmitter SPNN203 sequentially switches 11 sets of beam codes, while the receiver SPNN207 switches 3 sets of codes for each of the three sectors. The intensity detection module 208 synchronously acquires 11 echo intensity features, and finally, the decision classifier in the acquisition and decision module 209 performs three matrix multiplication and addition operations to output the obstacle category recognition result in real time.
Claims
1. A method for processing electromagnetic echo signals based on a programmable electromagnetic neural network, characterized in that: Includes the following steps: Step 101: Based on the programmable electromagnetic neural network hardware design, the transmitter electromagnetic neural network encoding is implemented to achieve sector scanning of the transmission beam, so as to realize spatial diversity of the acquired data and improve the sensing resolution; Step 102: Install the transmitting and receiving antenna arrays, and collect the echo dataset of the receiving antenna array under various obstacle scenarios to provide preliminary data for the training of the programmable electromagnetic neural network; Step 103: Design the electromagnetic neural network encoding and back-end processing matrix for the receiving end, and perform preliminary training on the programmable electromagnetic neural networks of the transmitting and receiving ends with the goal of classifying obstacle categories, and solidify the weights of the programmable electromagnetic neural network. Step 104: Collect the output dataset of the programmable electromagnetic neural network at the output end under actual working conditions to provide more refined data for the post-training of the electromagnetic echo signal processing system; Step 105: Optimize the backend processing matrix again to further improve the decision performance.
2. The electromagnetic echo signal processing method based on a programmable electromagnetic neural network according to claim 1, characterized in that: The system used in the processing method includes: a transmitter SPNN (203) is placed at the front end of the vehicle (201), and the electromagnetic neural network encoding in step 101 is injected. The adjusted radio frequency signal is transmitted through the transmitter antenna array (204) and the beam scanning function is realized. The transmitted electromagnetic signal is reflected by the obstacle (202) and received by the receiver antenna array (205), and transmitted to the input port of the receiver module (206) through the cable. In the module, the receiver SPNN (207) realizes the simulated neural network calculation in the electromagnetic domain, realizes the feature extraction and preliminary training process. Subsequently, the incoherent detection link is realized through the intensity detection module (208), which converts it into baseband components. Then, the acquisition and decision module (209) executes the final simple classifier algorithm and decision.
3. The electromagnetic echo signal processing method based on a programmable electromagnetic neural network according to claim 2, characterized in that: The specific steps of step 101 include: The number of spatial scanning beams is determined based on the types of potential obstacles encountered in autonomous driving to implement spatial diversity, and the transmitting electromagnetic neural network is trained. Eleven beams are set, divided into three independent sectors: left, center, and right, with beam counts of 4, 3, and 4 respectively. Based on a programmable electromagnetic neural network (SPNN), the adjustable parameters of the network are optimized using a backpropagation algorithm. This SPNN consists of three cascaded propagation layers and three modulation phase-shifting layers, each with 32 ports. Input ports 5, 9, 13, and 17 are connected to a four-port power divider, which in turn connects to the center frequency signal source. The transmitting SPNN is represented as follows: ， in These are the input and output vectors of the transmitting electromagnetic network. For the transmission matrix of the propagation layer, , , These represent the phase modulation matrices of the 1st, 2nd, and 3rd phase-shifting layers of the electromagnetic network at the transmitter; and the input vector of the electromagnetic network based on the power divider excitation mode. Set the specified port to 1 and the rest to 0; to establish the far-field model, the intensity of the 11 beam positions is calculated using the following formula; ， Where I is the output port energy vector. This is the free-space far-field propagation matrix from the antenna array connected to the SPNN at the transmitter to the 11 sector corners. Indicates to Square the modulus value; define the beam scanning task loss value L1: ， Where L1 is the beam scanning task loss value. Represents the target normalized beam intensity vector; the first beam is Based on formulas (1) to (3), the gradient descent method is used to minimize the beam scanning task loss value L1, thereby obtaining the optimal phase modulation matrix of the transmitter electromagnetic network. , , Eleven phase codes are generated for the eleven target beams, and then converted into control vectors through the phase-voltage mapping relationship of a programmable electromagnetic network for system use.
4. The electromagnetic echo signal processing method based on a programmable electromagnetic neural network according to claim 3, characterized in that: The specific steps of step 102 include: An electromagnetic echo detection system consisting of a transmitter SPNN and a receiver SPNN is established. The specific deployment includes: connecting the input of the transmitter SPNN (203) to a radio frequency signal source via a power divider, and connecting the output to the transmitter antenna array (204). The entire system is installed on the hood of the test vehicle, with the main lobe pointing in the direction of travel to detect obstacles; deploying the receiver antenna array (205) at different locations on the vehicle body, with its output connected to the input port of the receiver SPNN (207); integrating the receiver SPNN (207), intensity detection module (208), and acquisition and decision module (209) into the receiver module (206); during the data acquisition phase, using a radio frequency matrix switch and a vector network analyzer, the system traverses 11 transmitting beams and different types of obstacle scenarios, collecting echo intensity data of each element of the receiver antenna array to form a training dataset.
5. The electromagnetic echo signal processing method based on a programmable electromagnetic neural network according to claim 4, characterized in that: The specific steps of step 103 include: based on the dataset collected in step 102, jointly training the parameters of the receiving SPNN (207) and the classifier transformation matrix in the acquisition and decision module (209), and the transmission model of the receiving SPNN (207) is as shown in formula (4): ， in The collected echo data vector, This is the output vector of the electromagnetic neural network at the receiving end. , , These are the phase modulation matrices of the 1st, 2nd, and 3rd phase-shifting layers of the electromagnetic neural network at the receiving end; the detection and linear classifier model is expressed as follows: ， in: Output decision vectors for the linear classifier. For the weight matrix of the linear classifier, This is the bias vector for a linear classifier. The electromagnetic neural network outputs an intensity vector; The loss value L2 for the target classification task is defined as the distance between the true label and the output decision vector of the linear classifier, calculated using the cross-entropy function: ， Where crossentropy represents the cross-entropy function, T is the true label vector, and softmax is the vector size of the target vector. The normalization function is converted into a normalized probability vector distribution; based on the dataset collected in step 102, the gradient descent method is used to minimize the loss value L2 of the target classification and recognition task to obtain the optimal value. , , With the linear classifier weight matrix and the bias vector of the linear classifier For the left, middle, and right sectors, three independent sets of receiver SPNN encoding and classifier parameters are trained and stored respectively.
6. The electromagnetic echo signal processing method based on a programmable electromagnetic neural network according to claim 5, characterized in that: Step 104 includes verification of the system's actual operating conditions and fine-tuning of the backend, specifically as follows: A complete electromagnetic signal identification system is built by directly coupling the receiving antenna array to the input of the receiving SPNN (207). The transmitting and receiving beam switching and data acquisition are performed according to the preset timing process, and the timing synchronization is provided by the main control module. In order to eliminate the performance degradation caused by physical factors such as cable loss and antenna position deviation in actual deployment, the back-end adaptive fine-tuning is performed: Under real working conditions, based on the phase parameters of the receiving SPNN obtained by optimization in step 103, the intensity data output by the intensity detection module (208) is collected again according to the scanning process. The corresponding sample labels were then marked for future optimization.
7. The electromagnetic echo signal processing method based on a programmable electromagnetic neural network according to claim 6, characterized in that: Step 105 further optimizes the backend processing matrix, including obstacle recognition, and based on the dataset collected in step 104, optimizes the linear classifier weight matrices corresponding to the left, middle, and right sectors. and the bias vector of the linear classifier The system is retrained and corrected to compensate for physical link errors. Then, the fine-tuned classifier parameters are loaded to complete the final deployment of the system. In a single scan cycle, the transmitting SPNN (203) switches 11 sets of beam codes in sequence, and the receiving SPNN (207) switches 3 sets of codes for the three sectors respectively. The intensity detection module (208) collects 11 echo intensity features in a synchronous manner. Finally, the decision classifier in the acquisition and decision module (209) performs three matrix multiplication and addition operations to output the obstacle category recognition result in real time.