Physical signal sensing method and system approaching data processing inequality limit

Abstract

The invention discloses a physical signal sensing method and system approaching a data processing inequality limit, and belongs to the technical field of intelligent sensing and information processing. According to the method, a data processing inequality in an information theory is taken as a theoretical basis, time, space and frequency multi-dimensional physical waveforms are designed to improve an information upper limit, and a physical calculation model is utilized to carry out end-to-end direct feature extraction and classification identification on the received original physical waveforms. The preprocessing steps of generating images, time-frequency spectrums and other intermediate representations are completely omitted, so that information loss caused by information conversion in a traditional perception normal form is avoided, and the theoretical performance limit specified by a data processing inequality is approached. The physical calculation model can be realized through software or special hardware, and has the advantages of high processing precision, low delay and low power consumption. According to the method, a traditional recognition normal form based on traditional image processing is broken, and an innovative solution is provided for a next-generation high-precision intelligent sensing system.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing and information processing technology, specifically to a physical signal sensing method and system that approximates the limit of data processing inequalities, for realizing intelligent sensing tasks such as high-precision, low-latency radar target recognition, speech recognition, and medical imaging. Background Technology

[0002] Driven by the wave of the intelligent era, intelligent sensing and decision-making, along with their applications, are permeating all key areas of national security, economic development, and social life. From the precise identification of targets by radar detection to the real-time perception and decision-making of complex road conditions by autonomous vehicles, from the natural human-computer interaction based on voice or gestures in smart homes to the early disease diagnosis and surgical navigation through medical imaging in the healthcare field, the performance of intelligent sensing systems, as the direct interaction objects between various intelligent agents and the environment, directly determines the intelligence level and reliability of the entire system. Intelligent sensing systems mainly extract key information about the target's characteristics, state, or identity from the information-carrying physical waveforms through the interaction of physical waves (such as electromagnetic waves, sound waves, light waves, and magnetic signals) with the target or environment, thereby realizing advanced cognitive functions such as identification, positioning, tracking, and diagnosis.

[0003] Current mainstream intelligent sensing systems typically employ preprocessing techniques to extract information from physical waveforms. First, the system receives raw physical waveforms through sensors. These waveforms contain direct mappings of target features in the time, frequency, or spatial domains, representing the richest and most direct target information. However, these raw waveforms are difficult for humans to interpret directly. Therefore, the system preprocesses and transforms the physical waveforms using signal processing algorithms, converting them into intermediate representations that are easier for the human visual system or cognitive habits to understand and process. Examples include: time-spectrum maps obtained from time-domain signals through short-time Fourier transforms; two-dimensional radar images reconstructed from radar echoes using synthetic aperture technology; three-dimensional point clouds obtained through lidar scanning; or medical images acquired by magnetic resonance imaging (MRI). Based on this, the system further abstracts higher-level, task-related features from these intermediate representations using feature extraction algorithms, ultimately using a classifier or decision model to complete tasks such as recognition, segmentation, or regression.

[0004] However, according to information theory, two random variables... and Mutual information between Quantified The contents contained in The amount of information. In classification problems, mutual information directly determines the theoretically achievable minimum error probability. The data processing inequality states that for any Markov chain... The information processing process always satisfies This means that after each layer of processing, the amount of information about the original source that the data provides will only decrease or remain unchanged, not increase. Applying the data processing inequality to the traditional perception paradigm, a clear Markov chain can be constructed: the inherent properties or states of the target The source (i.e., the information source) determines the physical waveform it produces. The system for The intermediate representation is obtained after processing. ,Right now According to the data processing inequality, we must have: This means that in traditional processing methods, preprocessing and other steps inevitably discard or obscure some subtle information in the original waveform that may be crucial to the final decision but is difficult to retain in a specific intermediate representation when converting the signal into an intermediate representation that humans can understand.

[0005] In recent years, an emerging processing paradigm known as physical computing has begun to attract attention. Its core concept is to bypass the step of generating human-interpretable intermediate representations and directly simulate or perform mixed-signal processing on the raw physical waveforms to extract information and complete decision-making tasks in an end-to-end manner. Some preliminary studies have demonstrated the potential for low-power, low-latency implementation of physical computing using acoustic or electromagnetic prototype systems in tasks such as speech recognition and spectrum sensing. These works demonstrate the advantages of physical computing in terms of hardware efficiency.

[0006] However, existing research still has two key theoretical gaps in its understanding of the physical computing paradigm: First, although it is intuitively believed that directly processing waveforms should retain more information, there is a lack of rigorous theoretical proof and quantitative analysis from the perspective of information theory (especially DPI), failing to clearly clarify the performance ceiling that physical computing can achieve and its fundamental advantages over traditional paradigms. Second, most existing work focuses on utilizing a single dimension of waveforms (such as the time domain), failing to systematically propose and verify a theoretical framework and implementation method for actively improving the upper limit of system information capture capability by comprehensively utilizing multiple physical dimensions of physical waveforms, such as time, frequency, and space. Summary of the Invention

[0007] This invention addresses the aforementioned shortcomings of existing technologies, providing a physical computing theory and technical implementation scheme that approximates the limit of the data processing inequality for the field of intelligent sensing and decision-making. This scheme theoretically and experimentally verifies the effectiveness of the data processing inequality, demonstrating that physical computing can retain more information than traditional sensing paradigms. Received physical waveforms are directly processed to extract key information without human-understandable preprocessing, thus reaching the upper limit of the data processing inequality. Furthermore, this scheme comprehensively utilizes multi-dimensional physical waveforms in time, space, and frequency to further enhance the information upper limit defined by the data processing inequality, ultimately achieving performance exceeding that of traditional methods in practical sensing tasks.

[0008] The technical solution of the present invention is as follows: A physical signal sensing method and system based on physical computation to approximate the limit of data processing inequalities, characterized by including: To maximize the received signal Impact response with target Mutual information between To optimize the criteria, a transmit waveform is designed that is jointly modulated in at least two physical dimensions in the time, frequency, and spatial domains. Transmit the optimized transmit waveform and receive the raw physical echo signal generated by target scattering or modulation without any image or time-spectrum conversion; The original physical echo signal is directly input into a pre-trained physical calculation model, which performs end-to-end physical calculation operations on the original physical echo signal to directly extract the target information features. The target information features extracted by the physical calculation model are input into a pre-trained recognition module, which outputs the final target category, thus completing the end-to-end mapping from the original physical waveform to the recognition result.

[0009] The aforementioned design of the spatiotemporal multidimensional transmission waveform is a computer program that implements the optimization algorithm, employing the principle of maximizing mutual information. To measure the target information contained in the physical waveform: In the formula, Represents the target impulse response function. This represents the physical echo containing target information. The covariance matrix representing environmental noise. The covariance matrix represents the target impulse response. This indicates the transmit signal matrix that needs to be designed. It is the identity matrix. This indicates finding the conjugate transpose of a matrix.

[0010] The transmitted waveform, which is jointly modulated by at least two physical dimensions in the time, frequency, and spatial domains of the design, is a complex modulated signal with low autocorrelation sidelobes.

[0011] The complex modulation signal with low autocorrelation sidelobes is a Costas-coded linear frequency modulated signal. By introducing a coded frequency hopping pattern into the time-frequency two-dimensional plane, the sidelobe level of the waveform's autocorrelation function is reduced, thereby achieving the mutual information under a given power constraint. The increase.

[0012] The physical computation model is a trainable deep neural network model, analog computing circuit, or optical computing device capable of directly processing raw waveform data. Its network structure or circuit connections are configured to perform convolution, recursive loops, or encoder-decode operations on the waveform signal to learn a direct mapping function from the raw waveform to abstract features. So that the target information features satisfy .

[0013] The identification module is a classifier, which is implemented as a fully connected neural network, support vector machine or decision tree, and is integrated into a digital processor, analog computing chip or optical computing chip, for high-speed classification decision on the features extracted by the physical computing model.

[0014] The physical signal sensing system is characterized in that the waveform design and transmission module includes a waveform optimization calculation unit, an arbitrary waveform generator, a power amplifier, and a transmitting antenna connected in sequence; the signal receiving and acquisition module includes a receiving antenna, a low-noise amplifier, and an analog-to-digital converter connected in sequence, and the output terminal of the analog-to-digital converter constitutes the output terminal of the signal receiving and acquisition module.

[0015] The physical computing processing module and the intelligent recognition module are physically integrated within the same hardware computing unit. The hardware computing unit is a computer program, an analog computing chip, or an optical computing chip. The physical computing model and the function of the recognition module are executed by a continuous physical process implemented by a computer program, analog circuit, or optical device. This allows the original physical echo signal to be directly processed in the hardware computing unit in the form of a digital signal, an analog signal, or an optical signal. The output signal of the physical computing processing module is directly transmitted to the input of the intelligent recognition module through an internal continuous path.

[0016] The output or intermediate results of the intelligent recognition module are fed back to the waveform design and transmission module. The waveform design and transmission module dynamically adjusts the design parameters of the transmission waveform matrix according to the feedback information to achieve adaptive information acquisition in response to changes in the environment or target.

[0017] A high-precision intelligent sensing and recognition method is achieved by utilizing the aforementioned physical computation theory and technology that approximates the limit of data processing inequalities. This method includes two stages: training and applying a physical waveform information extraction module and a recognition module from physical computation. The steps are as follows: 1) Training phase: The physical waveform information extraction and recognition modules in the physical calculation can only be trained to directly process physical waveforms containing target information under conditions approaching the limit of data processing inequalities, achieving high-precision recognition. Once the spatiotemporal frequency multi-dimensional transmission waveform is designed, the original physical waveforms containing target information are received through the hardware architecture for acquiring various physical waveforms. The stored data is used as the training set for the physical waveform information extraction and recognition modules in the physical calculation. An optimization algorithm is then used to train these modules, enabling them to learn the information features of different targets contained in the physical waveforms, achieving high-precision classification and recognition.

[0018] 2) Application stage: After the spatiotemporal frequency multidimensional waveform is transmitted, the hardware architecture for acquiring various physical waveforms receives the physical waveform containing target information after interaction with the target. The received physical waveform is then input into the physical waveform information extraction and recognition modules in the physical calculation, which have been trained during the training phase. The network outputs a high-accuracy recognition result, achieving high-precision radar target recognition.

[0019] The system is an end-to-end processing system based on a physical computing architecture, including: Waveform design and transmission module, used to maximize received signal. Impact response with target Mutual information between To optimize the criteria, a transmit waveform is designed that is jointly modulated in at least two physical dimensions in the time, frequency, and spatial domains. The signal receiving and acquisition module has its input end coupled to the radiation space of the waveform design and transmission module, and is used to receive the original physical echo signal containing target information. Its output end provides the digitized original physical echo signal. The physical calculation processing module has its input end connected to the output end of the signal receiving and acquisition submodule. It is used to receive the digitized raw physical echo signal, carry and run the physical calculation model to perform end-to-end physical calculation operations on the raw physical echo signal and extract target information features. The intelligent recognition module, whose input end is connected to the output end of the physical computing processing module, is used to receive the target information features, carry and run the recognition module to perform the end-to-end recognition steps and output the target category; The output of the signal receiving and acquisition module is directly connected to the input of the physical calculation and processing module, and the output of the physical calculation and processing submodule is connected to the input of the intelligent recognition module, forming a direct processing link without intermediate representation generation.

[0020] Based on the above technical features, the present invention has the following advantages: 1. The received physical waveform is directly processed using physical calculations. According to the data processing inequality, the direct processing of the physical echo can retain the most target information in the physical waveform, thereby achieving a higher precision perception and recognition function.

[0021] 2. By fully utilizing multiple degrees of freedom such as time, space, and frequency in the physical space to obtain physical waveforms, more target information can be obtained compared to a single dimension, thereby increasing the upper limit of information limited by data processing inequalities and achieving higher precision perception and recognition functions.

[0022] This invention proposes a physical computation-based perception and recognition framework that reaches the upper limit of data processing inequalities. Compared with traditional sensing paradigms, this invention has the advantage of directly extracting information from physical waveforms, thereby enabling efficient information processing based on data processing inequalities. Furthermore, by fully utilizing multiple dimensions of physical waveforms, the total amount of information carried by the waveforms is effectively increased, thereby raising the upper limit of data processing inequalities and achieving high-precision perception and recognition. Attached Figure Description

[0023] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a system flowchart of a physical signal sensing method and system for approximating the limit of data processing inequalities in one embodiment of the present invention.

[0024] Figure 2 This is a spatiotemporal frequency multidimensional transmission waveform autocorrelation function diagram designed in one embodiment of the present invention.

[0025] Figure 3 This is a target mutual information curve diagram carried by the spatiotemporal frequency multidimensional transmission waveform designed in one embodiment of the present invention.

[0026] Figure 4 This is a hardware architecture diagram for obtaining physical waveforms containing target information features in one embodiment of the present invention.

[0027] Figure 5This is a training effect diagram of the physical waveform information extraction module and the recognition module in one embodiment of the present invention. Detailed Implementation

[0028] The following is a detailed description of an embodiment of the physical signal sensing method and system for approximating the limit of data processing inequalities according to the present invention: This embodiment is implemented under the premise of the technical solution of the present invention, and provides detailed implementation methods and specific operation processes, but the scope of protection of the present invention should not be limited to the following embodiments.

[0029] Figure 1 This is a flowchart illustrating a physical signal sensing method for approximating the limit of data processing inequalities, provided in an embodiment of the present invention. The present invention mainly includes the following operations: Based on the principle of maximizing mutual information, a multi-dimensional transmission waveform for spatiotemporal frequency is designed. Obtain the physical waveform containing target information features; A model for implementing physical calculations is provided; in this model, no preprocessing or other operations are required, and the received physical waveform is directly used as the model input. The model extracts the target information contained in the physical waveform by performing physical calculation operations. A recognition module for intelligent perception is provided; the recognition module takes the target information features extracted by physical calculation as input, outputs the target category through a feature classification algorithm, and realizes the perception and recognition function.

[0030] Specifically, the design of the spatiotemporal frequency multi-dimensional transmission waveform mainly adopts the method of maximizing mutual information. To measure the target information contained in the physical waveform: In the formula, Represents the target impulse response function. This represents the physical echo containing target information. The covariance matrix representing environmental noise. The covariance matrix represents the target impulse response. This indicates the transmit signal matrix that needs to be designed. It is the identity matrix. This represents finding the conjugate transpose of a matrix. By solving an optimization problem, we obtain information that allows for mutual information. Largest transmit signal matrix Subsequently, a hardware architecture for acquiring various physical waveforms is constructed, including an arbitrary waveform generator for generating the designed signal with maximum mutual information, an amplifier for increasing signal power in the transmission link, a transmitting antenna for transmitting signals into space, a receiving antenna for receiving signals, and an analog-to-digital converter and memory for storing the acquired physical waveforms. Then, target information is extracted from the physical waveforms by constructing a physical computation algorithm and hardware architecture. In this embodiment, target information extraction is achieved through convolutional feature extraction and pooling feature extraction. Finally, based on the target information features extracted by the physical computation, this embodiment outputs the target category based on a fully connected neural network to achieve classification. From an information theory perspective, this process can be modeled as follows: .because Is it directly from Since the data is learned without undergoing irreversible preprocessing, it can theoretically approach the upper limit specified by the data processing inequality infinitely. ,Right now This theoretically guarantees that the amount of information that physical computation can obtain is no less than, and usually more than, the amount of information obtained by traditional intermediate representation-based paradigms.

[0031] The technical solution provided by the above embodiments of the present invention will be further described in detail below, taking radar target recognition as an example.

[0032] In the design of spatiotemporal multidimensional transmission waveforms, mutual information increases when the autocorrelation function of the transmitted signal approaches zero. Both the target impulse response and noise follow a cyclically symmetric complex Gaussian distribution. The mutual information between the received signal and the target impulse response can be expressed as: in The covariance matrix representing the noise. The convolution matrix representing the transmitted signal. Let represent the covariance matrix of the target impulse response. Assume... (This indicates that the noise is mainly composed of white noise.) (This indicates that each element of the target impulse response is independently distributed). Therefore, mutual information can be re-expressed as: Given a known target and noise, the transmission waveform convolution matrix is ​​modulated. To increase Note: in express The cyclic autocorrelation coefficient, yes Length, Indicates the transmission power.

[0033] Assumption ,in It is the identity matrix. yes The off-diagonal part of the matrix represents the interference terms of the function.

[0034] Assumption , Redefined According to the formula ,in Represents the trace of a matrix, for Perform a Taylor expansion: For sufficiently small The conditions are met. This ensures the convergence of the expansion. We can obtain... therefore: because diagonal matrix The diagonal elements are 0 and conjugate symmetric, which gives us: The quadratic term is expanded as follows: We can obtain: Therefore, in , ( Within the neighborhood of ), as The decrease, Increase, therefore mutual information This will also increase.

[0035] This invention selects the Costas-LFM signal as the transmitted signal, thereby fully utilizing the time and frequency dimensions of the physical waveform to obtain more mutual information. The Costas-LFM signal is obtained by applying Costas coding to the carrier frequency of a linear frequency modulated (LFM) signal; its non-zero autocorrelation function is lower than that of the LFM signal, such as... Figure 2 As shown. When using Costas-LFM as the transmitted signal, the comparison of the mutual information between the obtained electromagnetic signal and the target impulse response is as follows. Figure 3As shown, the autocorrelation functions of various transmitted signals were analyzed through numerical simulation. The Costas-LFM signal's Costas coding sequence is [2, 4, 8, 5, 10, 9, 7, 3, 6, 1]. Compared to the LFM signal used in traditional radar target identification, the Costas-LFM signal exhibits lower autocorrelation sidelobes. Set as Regardless of noise power, the mutual information of the Costas-LFM signal is always higher than that of the LFM signal.

[0036] The physical waveform constructed in this embodiment for acquiring features containing target information is as follows: Figure 4 As shown, the designed transmit signal is generated by an arbitrary waveform generator and processed by a power amplifier. Minimum carrier frequency interval. for The carrier frequency of each part of this signal is , and the bandwidth of each part of the signal is The signal sampling duration is 5 microseconds, and the pulse duration is 4 microseconds. The target is located approximately 3 meters away from the transmitting and receiving antennas. The transmitting and receiving antennas for different channels are arranged closely adjacently, following the multiple-input multiple-output antenna distribution principle, with each group of antennas arranged in an arc at 36 intervals. After receiving the target's physical waveform, a low-noise amplifier is used to improve the signal-to-noise ratio, and then an oscilloscope is used to capture the output signal, which is then input into the physical calculation architecture.

[0037] In the physical computation model, the information extraction module consists of two convolutional layers and two fully connected layers. The kernel parameters for both the first and second convolutional layers are [number of input channels, number of output channels, kernel width] = [4, 4, 3], with a stride of 2 and the activation function ReLU. The number of neurons in the first and second fully connected layers is set to 300 and 4, respectively. The activation functions for these layers are ReLU and Sigmoid, respectively. To obtain the training dataset, 50 physical waveforms were collected for each target category, resulting in a total of 200 waveforms. 100 waveforms were randomly selected for training, and the remaining 100 were used for validation. To verify that Costas-LFM can transmit more mutual information, physical computation was performed using the Costas-LFM physical waveforms to achieve recognition. Figure 5The training curves show a recognition accuracy of 98%, representing a significant improvement in performance (from 89% to 98%) compared to the LFM physical waveform. This indicates that the features of each data category are clearly clustered, and different feature clusters are separated from each other. Therefore, the results demonstrate that transmitting Costas-LFM enhances mutual information compared to the LFM physical waveform. By enabling the received physical waveform to acquire more mutual information, the system can achieve high-precision target recognition.

[0038] To demonstrate that physical computation can approximate the data processing inequality, this invention compares traditional sensing and recognition methods with physical computation: For most radar imaging algorithms, the imaging process inevitably leads to mutual information loss. This embodiment takes the BP algorithm of ISAR imaging as an example to discuss the information loss during the imaging process. The BP algorithm mainly includes three steps: (1) pulse compression for each physical waveform, (2) applying phase compensation to make the waveforms acquired at different times reach zero phase due to the change in distance between each point in the imaging area and the radar, and (3) coherent integration of the phase-compensated waveform, where the position of the scatterer exhibits a peak due to coherent accumulation, thereby forming the target image. According to the data processing inequality, when the processing is irreversible, the inequality does not hold, indicating that there is mutual information loss. First, consider the pulse compression step in BP imaging, which involves performing a matched filtering operation between the waveform and the transmitted signal. This process is reversible because the transmitted signal is known. However, this reversibility assumes that the amplitude and phase in the pulse compression result are preserved. Although most existing high-resolution radar echo recognition studies only use amplitude information, this step leads to information loss. In the phase compensation step, each waveform is multiplied by a compensation factor calculated based on the grid point location and the antenna. This process is reversible as long as these specific compensation factors are preserved during the calculation. However, in the coherent integration step, all phase-compensated waveforms are summed to form the radar image. This constitutes a many-to-one mapping, making it impossible to recover the amplitude and phase of individual waveforms from the integration result; therefore, this step is irreversible. Finally, ISAR image recognition is more akin to an image processing task. Although ISAR images contain both amplitude and phase components, most studies only retain amplitude information. Therefore, both coherent integration and the output of the ISAR image lead to information loss during ISAR imaging.

[0039] Any matters not covered in the above embodiments of the present invention are well-known in the art.

[0040] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A physical signal sensing method that approximates the limit of data processing inequalities, characterized in that, Includes the following steps: Step S1: Maximize the received signal Impact response with target Mutual information between To optimize the criteria, a transmit waveform is designed that is jointly modulated in at least two physical dimensions in the time, frequency, and spatial domains. Step S2: Transmit the optimized transmission waveform and receive the original physical echo signal generated after target scattering or modulation without any image or time-spectrum conversion; Step S3: The original physical echo signal is directly input into a pre-trained physical calculation model, and the physical calculation model performs end-to-end physical calculation operations on the original physical echo signal to directly extract the target information features. Step S4: Input the target information features extracted by the physical calculation model into a pre-trained recognition module, and the recognition module outputs the final target category, completing the end-to-end mapping from the original physical waveform to the recognition result.

2. The physical signal sensing method for approximating the limit of data processing inequalities according to claim 1, characterized in that, Maximizing mutual information in step S1 The criterion is achieved by solving the following optimization problem: In the formula, Represents the target impulse response function. This represents the physical echo containing target information. The covariance matrix representing environmental noise. The covariance matrix represents the target impulse response. This indicates the transmit signal matrix that needs to be designed. It is the identity matrix. This indicates finding the conjugate transpose of a matrix.

3. The physical signal sensing method for approximating the limit of data processing inequalities according to claim 2, characterized in that, The transmit waveform designed in step S1 is a complex modulated signal with low autocorrelation sidelobes.

4. The physical signal sensing method for approximating the limit of data processing inequalities according to claim 3, characterized in that, The complex modulation signal with low autocorrelation sidelobes is a Costas-coded linear frequency modulated signal. By introducing a coded frequency hopping pattern into the time-frequency two-dimensional plane, the sidelobe level of the waveform's autocorrelation function is reduced, thereby achieving the mutual information under a given power constraint. The increase.

5. The physical signal sensing method for approximating the limit of data processing inequalities according to claim 1, characterized in that, The physical computation model described in step S3 is a trainable deep neural network model capable of directly processing raw waveform data, including computer programs, analog computing circuits, or optical computing devices. Its network structure or circuit connections are configured to perform convolution, recursive loops, or encoder-decode operations on waveform signals to learn a direct mapping function from the raw waveform to abstract features. So that the target information features satisfy .

6. The physical signal sensing method for approximating the limit of data processing inequalities according to claim 1, characterized in that, In step S4, the identification module is a classifier, which is implemented as a fully connected neural network, support vector machine or decision tree, and is integrated into a digital processor, analog computing chip or optical computing chip, for high-speed classification decision on the features extracted by the physical computing model.

7. The physical signal sensing method for approximating the limit of data processing inequalities according to any one of claims 1-6, characterized in that, It also includes a model training step, which is performed after S1 and before the actual application of S3, specifically including: Using the optimized transmission waveform, a large number of raw physical echo signals corresponding to targets of known categories are collected to construct a training dataset; Using the original physical echo signal as input and the corresponding target category as the supervision label, the physical calculation model and the recognition module are jointly trained end-to-end. The training objective is to minimize the final recognition error, which essentially maximizes the amount of information extracted and used for decision-making by the entire system, i.e., driving... Approximating the theoretical maximum value determined by the waveform optimization design steps .

8. A physical signal sensing system for implementing the method according to any one of claims 1 to 7, characterized in that, The system is an end-to-end processing system based on a physical computing architecture, including: Waveform design and transmission module, used to maximize received signal. Impact response with target Mutual information between To optimize the criteria, a transmit waveform is designed that is jointly modulated in at least two physical dimensions in the time, frequency, and spatial domains. The signal receiving and acquisition module has its input end coupled to the radiation space of the waveform design and transmission module, and is used to receive the original physical echo signal containing target information. Its output end provides the digitized original physical echo signal. The physical calculation processing module has its input end connected to the output end of the signal receiving and acquisition submodule. It is used to receive the digitized raw physical echo signal, carry and run the physical calculation model to perform end-to-end physical calculation operations on the raw physical echo signal and extract target information features. The intelligent recognition module, whose input end is connected to the output end of the physical computing processing module, is used to receive the target information features, carry and run the recognition module to perform the end-to-end recognition steps and output the target category; The output of the signal receiving and acquisition module is directly connected to the input of the physical calculation and processing module, and the output of the physical calculation and processing submodule is connected to the input of the intelligent recognition module, forming a direct processing link without intermediate representation generation.

9. The physical signal sensing system according to claim 8, characterized in that, The waveform design and transmission module includes a waveform optimization calculation unit, an arbitrary waveform generator, a power amplifier, and a transmitting antenna connected in sequence; the signal receiving and acquisition module includes a receiving antenna, a low-noise amplifier, and an analog-to-digital converter connected in sequence, and the output terminal of the analog-to-digital converter constitutes the output terminal of the signal receiving and acquisition module.

10. The physical signal sensing system according to claim 8, characterized in that, The hardware computing unit is a computer program, an analog computing chip, or an optical computing chip. The physical computing model and the function of the recognition module are executed by a continuous physical process implemented by the computer program, analog circuit, or optical device, so that the original physical echo signal is directly processed in the hardware computing unit in the form of digital signal, analog signal, or optical signal. The output signal of the physical computing processing module is directly transmitted to the input of the intelligent recognition module through an internal continuous path.

11. The physical signal sensing system according to claim 8 or 10, characterized in that, It also includes a feedback connection, wherein the output or intermediate result of the intelligent recognition module is fed back to the waveform design and transmission module, and the waveform design and transmission module dynamically adjusts the design parameters of the transmission waveform matrix according to the feedback information to achieve adaptive information acquisition in response to changes in the environment or target.

Images