A low-overhead sampling analysis method for wideband wireless signals

Abstract

The application relates to a low-overhead sampling analysis method for a broadband wireless signal, which comprises the following steps: obtaining an original low-rate IQ sample sequence of a target broadband signal, carrying out multi-dimensional feature embedding to obtain an original feature sequence; carrying out autocorrelation analog sampling processing on the original low-rate IQ sample sequence to obtain an enhanced autocorrelation feature sequence, splicing the enhanced autocorrelation feature sequence with the original feature sequence to form a fusion enhanced feature sequence; inputting the fusion enhanced feature sequence into a spectrum sensing model based on a deep neural network for processing, predicting spectrum occupation prior information in real time, and using the spectrum occupation prior information as a signal reconstruction constraint to reconstruct the original low-rate IQ sample sequence to obtain an effective recovery signal; and using a signal analysis model based on a Transformer to carry out signal analysis on the effective recovery signal to obtain a deep analysis result. Compared with the prior art, the application has the advantages of breaking through the sampling rate limitation in the prior art and greatly simplifying a data processing procedure.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wideband signal processing, and in particular to a low-overhead sampling analysis method for wideband wireless signals. BACKGROUND

[0002] With the rapid development of global wireless communication technology (including terrestrial cellular networks such as 4G / 5G and satellite communication networks), the occupation of the radio frequency spectrum is increasingly dense. In order to adapt to the huge data throughput demand, the bandwidth of the signal to be monitored is continuously expanding, covering from the traditional low frequency band to the high frequency band of tens of gigahertz (GHz), so that the wireless signals in the actual environment present the complex characteristics of large bandwidth, multiple concurrency and non-sparse. Physical layer monitoring, spectrum sensing and data analysis of these wideband signals are of great importance to spectrum resource management, network performance optimization and security protection.

[0003] However, the current mainstream signal analysis methods and systems face the following challenges in actual deployment:

[0004] (1) High sampling rate and huge hardware cost.

[0005] The most commonly used wideband signal analysis method currently relies heavily on the Nyquist sampling rate, i.e. the sampling rate must be greater than the total bandwidth of the monitoring frequency band. For monitoring tasks covering tens of gigahertz bandwidth, this requires the use of high-cost, high-performance analog-to-digital converters and radio frequency front-ends. With the continuous increase of monitoring bandwidth, the resulting hardware cost, power consumption, data storage and real-time transmission bandwidth overheads are huge, which greatly limits the deployment scale and application range of wideband monitoring systems, especially in resource-limited or large-scale deployment scenarios.

[0006] In addition, there are some other problems, such as:

[0007] (2) Limitations of existing low-rate sampling methods.

[0008] Existing low-rate sampling methods cannot meet the needs of wideband signal analysis. For example, a frequency-sweeping spectrum analyzer measures the wideband spectrum by narrowband filtering and point-by-point scanning, which fundamentally lacks the ability to simultaneously obtain the instantaneous and complete information of all signals within the entire monitoring bandwidth. For another example, the premise of sparse sampling to reduce the sampling rate is that the signal must have strict sparsity in a certain transform domain, which is difficult to adapt to the increasingly crowded electromagnetic environment.

[0009] (3) Passive reception analysis process is complex and inefficient.

[0010] As a passive signal receiver, performing physical layer-level deep analysis of broadband signals without prior knowledge of the signal type and parameters is a highly complex process. This process involves a series of interconnected steps, including complex frequency estimation, channel compensation, synchronization, demodulation, and error correction. This traditional approach is computationally intensive and requires extremely high accuracy in parameter estimation at each step. Errors at any step are propagated downwards, resulting in poor system robustness, high complexity, and significant processing latency, making it difficult to meet the real-time analysis needs of massive broadband signals. Summary of the Invention

[0011] The purpose of this invention is to provide a low-overhead sampling analysis method for broadband wireless signals that solves the problem of traditional methods struggling to handle non-sparse broadband signals.

[0012] The objective of this invention can be achieved through the following technical solutions:

[0013] A low-overhead sampling analysis method for broadband wireless signals includes the following steps:

[0014] The original low-rate IQ sample sequence of the target broadband signal is obtained, and multi-dimensional feature embedding is performed to obtain the original feature sequence.

[0015] The original low-rate IQ sample sequence is subjected to autocorrelation simulation sampling processing to obtain an enhanced autocorrelation feature sequence, which is then concatenated with the original feature sequence to form a fused enhanced feature sequence.

[0016] The fused enhanced feature sequence is input into a spectrum sensing model based on a deep neural network for processing, and the spectrum occupancy prior information in the target frequency band is predicted in real time. This information is then used as a signal reconstruction constraint to reconstruct the original low-rate IQ sample sequence, thereby obtaining an effective recovered signal.

[0017] The effective recovered signal was analyzed using a Transformer-based signal analysis model to obtain in-depth analysis results.

[0018] Furthermore, the original low-rate IQ sample sequence is obtained by using multi-coset sampling on the deployed under-Nyquist platform. The steps for obtaining the original low-rate IQ sample sequence include:

[0019] Let the set of sampling delays of each analog-to-digital converter on the under-Nyquist platform be . ,in The number of analog-to-digital converters, For the first Sampling delay of each analog-to-digital converter;

[0020] Each analog-to-digital converter performs multi-channel time-interleaved parallel sampling of the target broadband signal according to its configured sampling delay, resulting in a multi-channel independent, time-interleaved initial raw low-rate IQ sample sequence. The sampling results of each analog-to-digital converter are represented as follows:

[0021] ,

[0022] In the formula, For the first The sampling results of the analog-to-digital converter, For the target broadband signal, Indicates the first One sampling point, The ratio of the sampling time interval for each analog-to-digital converter to that of conventional Nyquist sampling. This indicates the sampling time interval of the Nyquist sampling. The target total bandwidth of the spectrum. For the first Sampling delay of each analog-to-digital converter;

[0023] The initial raw low-rate IQ sample sequence is subjected to time delay calibration, phase synchronization between channels, and noise filtering to obtain the final raw low-rate IQ sample sequence.

[0024] Furthermore, the step of obtaining the original low-rate IQ sample sequence also includes processing the sampling delay set. The optimization includes the following:

[0025] The sampling delay set Perform pairwise subtraction on the elements in the set of sampling delays. Optimize by adjusting The time delay value in the set of time delay differences Maximizing the number of elements in the set completes the optimization process, where the delay difference set is... Represented as:

[0026] ,

[0027] In the formula, For the first Sampling delay of each analog-to-digital converter;

[0028] By adjusting the sampling delay set The delay value in the set of delay differences, so that the delay difference set The optimization process is completed by maximizing the number of elements in the array.

[0029] Furthermore, the step of performing autocorrelation simulation sampling includes:

[0030] (1) Select the autocorrelation function used for the original low-rate IQ sample sequence. for:

[0031] ,

[0032] In the formula, For the target broadband signal, For the time difference, Expressing expectations, Indicates conjugate;

[0033] Then the autocorrelation function Sampling estimation for:

[0034] ,

[0035] ,

[0036] In the formula, To satisfy the time difference equals The number of sample pairs, To represent the time interval between two sampling points, the independent variable corresponds to the autocorrelation function. discrete values, Represents a set The specific sampling time points in the data. For the current moment, It is separated from it Past sampling times, For the time set of multiple coset sampling points of a multiplex analog-to-digital converter, For the first One sampling point, The ratio of the sampling time interval for each analog-to-digital converter to that of conventional Nyquist sampling. for This indicates the sampling time interval of the Nyquist sampling. The target total bandwidth of the spectrum. For the first Sampling delay of each analog-to-digital converter;

[0037] (2) Set the batched original low-rate IQ sample sequence as Calculate parameters , ,in, In the autocorrelation function Scope For the time difference, The number of sampling points. Folding factor This represents the maximum number of anchor points.

[0038] (3) Select anchor points for calculating the autocorrelation function and calculate the anchor point spacing. Construct a set of anchor points The number of anchor points is no greater than The anchor point spacing The calculation expression is:

[0039] ,

[0040] (4) Initialize the autocorrelation function sampling estimate Define a set of time delay differences for a matrix consisting entirely of zeros. The number of elements in the middle is ;

[0041] (5) By index Select elements in ;

[0042] (6) Find all that satisfy Analog-to-digital converter sampling channel number pair ,in For the first Sampling delay of each analog-to-digital converter;

[0043] (7) For each delay difference index Its range is 0 to Initialize an empty list Used to store all that meet the requirements The product result under the time delay difference;

[0044] (8) For each of the sampling channel number pairs If satisfied Then calculate If not satisfied Then calculate and the calculation results Add to Among them For analog-to-digital converters and The time difference near each anchor point is The product of the sampling point pairs, superscript For complex conjugate;

[0045] (9) Each anchor point and each group of sampling channels in the middle Take the average value to obtain the autocorrelation simulation sample value, and assign it to... The latter two dimensions are the anchor point dimension and the channel pair dimension;

[0046] (10) Repeat steps (7)-(9) until the time delay difference index is reached. Traversing from 0 to ;

[0047] (11) Repeat steps (5)-(10) until all iterations are completed. The middle element, the final result express Sampling estimation of the autocorrelation function at the location.

[0048] Furthermore, the step of obtaining the original feature sequence includes:

[0049] The original low-rate IQ sample sequence Convert to Real numbers of the form, and then reshape their last two dimensions into The shape is reshaped , wherein For complex fields, For the real number field, For batch number, The length of the symbol. The number of sampling points. The number of analog-to-digital converters, This represents the number of adjacent sampling points in the stack.

[0050] The reshaped The input is processed in a multilayer perceptron containing three linear layers to obtain... The original feature sequence of the shape, the activation function used by the multilayer perceptron is the Gaussian error linear unit activation function, and a dropout layer is added after the first and second linear layers. The output size of each linear layer is respectively... and After processing each linear layer, a layer normalization operation is performed. This represents the network feature dimension.

[0051] Furthermore, the step of obtaining the fusion-enhanced feature sequence includes:

[0052] The autocorrelation function sampling estimate obtained by autocorrelation simulation sampling processing The result is converted to a real number and then input into the multilayer perceptron containing three linear layers for processing, to obtain... Shape-enhanced autocorrelation feature sequences, In the autocorrelation function Scope;

[0053] The enhanced autocorrelation feature sequence and the original feature sequence are concatenated along the second dimension to obtain... Initial fusion enhancement feature sequence of shape;

[0054] The initial fused enhanced feature sequence is embedded with positional encoding, and a classification code is inserted at the head to obtain the fused enhanced feature sequence.

[0055] Furthermore, the spectrum perception model based on deep neural networks includes two layers of Transformer encoders with multi-head self-attention mechanisms and one layer of linear multilayer perceptron. The prediction step of the spectrum occupancy prior information includes:

[0056] Based on the fused and enhanced feature sequence, the first layer Transformer encoder is input, and different parts of the input are weighted and focused through a multi-head self-attention mechanism to output the feature sequence processed by the first layer Transformer encoder.

[0057] The feature sequence processed by the first Transformer encoder is used as input to the second Transformer encoder. The different parts of the input are weighted and focused through a multi-head self-attention mechanism, and the feature sequence processed by the first Transformer encoder is output as the deep association feature.

[0058] The linear multilayer perceptron is used to perform dimensionality reduction and classification on the deep correlation features, and the output is... The prediction of whether each sub-band of the bandwidth division is occupied serves as prior information for the predicted spectrum occupancy. The target total bandwidth of the spectrum. The ratio of the sampling time interval for each analog-to-digital converter to that of conventional Nyquist sampling.

[0059] Furthermore, the step of obtaining an effective recovery signal includes:

[0060] The sampling results of each analog-to-digital converter in the original low-rate IQ sample sequence Perform discrete Fourier transforms on each sample to obtain the transformed original low-rate IQ sample sequences. Furthermore, underdetermined equations are constructed to describe The relationship between the original low-rate IQ sample sequence and the original low-rate IQ sample sequence is expressed as follows:

[0061] ,

[0062] In the formula, This is the phase compensation matrix. For complex fields, The number of analog-to-digital converters, The number of sampling points. For the first The first channel in the Phase compensation term at each frequency point For the first The sampling delay of an analog-to-digital converter, The ratio of the sampling time interval for each analog-to-digital converter to that of conventional Nyquist sampling. For multi-coset sampling measurement matrix, For the first The analog-to-digital converter in the first... The mixing mode generated by each frequency point The spectrum matrix to be reconstructed Represents frequency The frequency components at each position, with each row corresponding to the spectrum of a sub-band. In the autocorrelation function Scope For the time difference, The target total bandwidth;

[0063] Using the aforementioned spectrum occupancy prior information Reducing the underdetermined equations, we obtain the overdetermined equations, expressed as:

[0064] ,

[0065] In the formula, For the occupied sub-band sequence number The submatrix formed by the columns, For occupied sub-band corresponding The submatrix formed by the rows;

[0066] The overdetermined equations are solved using the least squares method, yielding the following results. This forms the reconstructed spectrum matrix. ,in The expression is:

[0067] ,

[0068] In the formula, It is the conjugate transpose;

[0069] Extract the reconstructed spectrum matrix in units of sub-bands. frequency components Perform an inverse Fourier transform on it to obtain the sub-band signal corresponding to each sub-band, which is used as the effective recovery signal.

[0070] Furthermore, the Transformer-based signal analysis model includes a cascaded shared Transformer encoder, a protocol identification decoder, and a data demodulation decoder, and also includes an independent feature embedding module. The shared Transformer encoder comprises multiple stacked Transformer encoders, and the protocol identification decoder is composed of a single-layer Transformer decoder. The step of obtaining the deep analysis result includes:

[0071] The effective recovered signal is first subjected to shallow feature extraction using the first and second independent feature embedding modules respectively to obtain the first embedded feature and the second embedded feature respectively. The first embedded feature is then input into the shared Transformer encoder to extract the deep features for signal analysis.

[0072] Set query symbols, and use them together with the deep features as input to the protocol identification decoder to extract global features from the effective recovered signal, and output the wireless communication standard and protocol type of the effective recovered signal in real time;

[0073] The deep features extracted by the shared Transformer encoder are fused with the second embedded features to obtain fused features. The fused features and demodulated symbols of a number equal to the number of bits in the corresponding protocol decoding field are used as input to the data demodulation decoder to extract information of the corresponding bits from the effective recovered signal demodulation.

[0074] Furthermore, the Transformer-based signal analysis model is trained using a two-stage training method, the specific training process of which includes:

[0075] During the first phase of training:

[0076] The parameters of the shared Transformer encoder and protocol recognition decoder are trained jointly, while the parameters of the data demodulation decoder and the second independent feature embedding module are frozen.

[0077] In the second phase of training:

[0078] Freeze the parameters of the shared Transformer encoder and protocol identification decoder, and unfreeze and train only the parameters of the data demodulation decoder and the second independent feature embedding module.

[0079] Compared with the prior art, the present invention has the following beneficial effects:

[0080] (1) This invention enhances the features of low-rate samples by autocorrelation simulation sampling and uses the prior information of spectrum occupancy obtained by deep neural network as a constraint for signal reconstruction, thereby achieving high-precision signal recovery and solving the core problem that traditional methods are difficult to handle non-sparse broadband signals.

[0081] (2) The present invention provides a multi-coset under-Nyquist sampling design, which solves the problems of traditional sampling relying on Nyquist rate and high hardware cost.

[0082] (3) This invention overcomes the problem of traditional methods being limited to sparse spectra by using autocorrelation enhancement features and spectrum sensing. Based on optimized time delay configuration, autocorrelation features are constructed and fused with the original samples before being input into the spectrum sensing model. Even in non-sparse scenarios, it can still achieve a spectrum prediction accuracy of over 99% and accurately capture concurrent signals in multiple frequency bands.

[0083] (4) The signal analysis model based on Transformer proposed in this invention adopts a deep neural network shared architecture, which highly integrates protocol identification and data demodulation processes, significantly improves the robustness and analysis efficiency of complex signals, effectively solves the problems of error accumulation and large processing delay in traditional serial processes, and finally meets the needs of real-time and high-accuracy monitoring of massive broadband signals.

[0084] (5) This invention solves the problem of non-sparse signal recovery failure through a two-stage process of "perception-constraint reconstruction". By using the spectrum occupancy prior, the underdetermined equation is transformed into an overdetermined equation, and the signal is recovered with high fidelity by least squares solution. In non-sparse scenarios, the recovered MSE is close to the reference level.

[0085] (6) This invention utilizes a signal analysis model with multiple Transformers deployed in series to simplify the high complexity of traditional physical layer analysis. The shared encoder extracts features, the protocol identification decoder quickly outputs the Wi-Fi type, and the decoder completes the decoding of the protocol public fields end-to-end without complex processing procedures, achieving an identification accuracy of over 99% and near-lossless decoding.

[0086] (7) This invention combines low-cost under-Nyquist sampling and deep neural network intelligent processing to achieve low-overhead acquisition and high-accuracy detection and decoding of multiple concurrent broadband signals. It effectively breaks through the existing monitoring system's dependence on Nyquist sampling rate and expensive hardware, as well as the problems of difficulty in handling non-sparse broadband signals and the complexity and inefficiency of traditional analysis processes. It greatly saves the overhead of data acquisition, storage and transmission and greatly simplifies the complexity of broadband signal analysis system. Attached Figure Description

[0087] Figure 1 This is a flowchart of the method of the present invention;

[0088] Figure 2This is an overview diagram of the overall framework of the present invention;

[0089] Figure 3 This invention relates to a Transformer-based signal analysis model and its two-stage training mechanism.

[0090] Figure 4 This is a schematic diagram illustrating the system spectrum sensing accuracy of the present invention;

[0091] Figure 5 The impact of the autocorrelation sampling feature construction of this invention on existing compressed sensing methods and system model training;

[0092] Figure 6 This is a schematic diagram of the system signal reconstruction performance of the present invention;

[0093] Figure 7 This is a schematic diagram illustrating the system protocol recognition accuracy and protocol disclosure field decoding accuracy of the present invention. Detailed Implementation

[0094] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0095] This embodiment provides a low-overhead sampling analysis method for broadband wireless signals, such as... Figure 2 As shown, this method first utilizes a low-cost under-Nyquist sampling radio front-end to acquire original low-rate IQ sample sequences by performing multi-channel time-interleaved parallel IQ sampling on the target broadband non-sparse wireless signal at a total sampling rate far lower than the Nyquist rate. Next, the original low-rate IQ sample sequences undergo autocorrelation simulation sampling processing, and the processed enhanced autocorrelation features are concatenated with the original sample features and input into a spectrum sensing module based on a deep neural network to predict the prior information of the signal's spectrum occupancy in the target frequency band in real time. Then, a compressed sensing underdetermined equation is constructed based on the original low-rate IQ sample sequences, and the equation is reduced to an overdetermined equation using the prior information of spectrum occupancy obtained in step S2. This overdetermined equation is then solved using least squares to achieve high-fidelity effective signal recovery. Finally, a cascaded signal analysis module performs protocol type identification, data demodulation, and in-depth analysis of physical layer information on the effectively recovered signal.

[0096] Specifically, such as Figure 1 As shown, the method includes the following steps:

[0097] S1, Low-rate signal acquisition: Acquire the original low-rate IQ sample sequence.

[0098] Step S1, low-rate signal acquisition, is primarily based on multi-co-collection sampling technology. Through a low-cost under-Nyquist sampling radio front-end, it achieves multi-channel time-interleaved parallel I / Q sample acquisition of the target broadband non-sparse wireless signal at a total sampling rate far lower than the Nyquist rate, providing a high-quality raw data foundation for subsequent signal analysis. Specifically, it includes:

[0099] S11, Nyquist platform deployment missing:

[0100] In this phase, the first step is to deploy a monitoring front-end with ultra-wideband radio frequency reception capabilities, enabling it to cover monitoring bandwidths exceeding several GHz to meet the needs of broadband radio monitoring. The low-cost under-Nyquist sampling radio front-end has the following technical characteristics:

[0101] (1) It has ultra-wideband radio frequency reception capability, which can cover a wide monitoring bandwidth of several gigahertz (GHz);

[0102] (2) The platform has a broadband analog signal processing front end, the filtering bandwidth of which corresponds to the total bandwidth of the signal to be monitored; it is equipped with multiple analog-to-digital converters, and the actual sampling frequency of each analog-to-digital converter is designed to be much lower than the Nyquist rate of the filtering bandwidth.

[0103] (3) The platform integrates a direct memory access controller, which bypasses the central processing unit and directly transmits the collected raw low-rate IQ sample sequence to the host computer memory used for data processing at high speed.

[0104] Let the total bandwidth of the monitored target spectrum be... After downconversion and filtering by the RF front end, the result is obtained. The baseband signal is within the frequency range. The sampling front end is equipped with... There are one analog-to-digital converter (ADC), and the actual sampling frequency of each ADC is set to... times the Nyquist sampling rate, i.e. The system's total sampling rate is Multiple set sampling requirements This means the total sampling rate is less than the Nyquist sampling rate. During the acquisition process, the platform uses integrated DMA to bypass the CPU and directly transmit the obtained low-rate I / Q sample sequences at high speed into the host computer's memory, ensuring the real-time flow of broadband signals in the undersampled state.

[0105] S12. Fine-grained configuration of the multi-co-set sampling scheme:

[0106] (1) Configure the under-Nyquist sampling radio to use multi-channel parallel multi-co-collection sampling technology;

[0107] (2) The technology sparsely acquires broadband signals at multiple different sampling phases by configuring different, precise time delays (sampling delays) for the analog-to-digital converters of each channel, with each sampling point of the channel corresponding to a coset;

[0108] (3) In order to optimize the subsequent autocorrelation simulation sampling, the time delay between each channel of the multi-coset sampling is precisely set. The setting principle is to maximize the number of elements in the set constructed by subtracting the elements in the delay set, so as to maximize the information acquisition density of the corresponding autocorrelation domain.

[0109] (4) The total data throughput of multiple channels is strictly controlled to be less than twice the total signal bandwidth.

[0110] The specific multi-co-collection sampling scheme in this embodiment is as follows:

[0111] this Multiple ADCs acquire signals in parallel, each ADC is configured with a unique sampling delay. This allows each channel to sparsely acquire the broadband signal at different sampling phases. Let the target broadband signal be denoted as... , Denotes the sampling time interval of the Nyquist sampling, then the first... The expression for the sampling result of each ADC is:

[0112] ,

[0113] in Indicates the first One sampling point.

[0114] Sampling delay set of each ADC The configuration needs to be optimized. The optimization principle is to construct the delay difference set by subtracting each pair of elements in the delay set. The number of elements is maximized to maximize the information acquisition density of the corresponding autocorrelation domain. Furthermore, this constructed time delay set typically allows the subsequently constructed compressed sensing matrix to possess better cross-correlation and constrained isometry, ensuring the accuracy and stability of signal reconstruction. When deploying sampling equipment, the number of low-speed ADCs is considered. and the ratio of the Nyquist sampling rate to the sampling rate of each ADC. Calculate the optimal set of delays in advance. Each ADC is statically configured, and the latency configuration remains unchanged during the acquisition process.

[0115] S13. Acquisition and preprocessing of raw sample sequences:

[0116] This step mainly involves acquiring multiple independent, time-interleaved raw low-rate IQ sample sequences in the digital domain generated by multi-co-set sampling technology; and performing precise time delay calibration, phase synchronization between channels, and noise filtering in the digital domain on the multiple samples to ensure data synchronization and quality.

[0117] The data pipeline performs precise time delay calibration and inter-channel phase synchronization on multiple samples to compensate for sampling deviations caused by differences in analog devices or wiring delays. The calibrated signals are then subjected to noise suppression in the digital domain using low-pass or band-pass filters to improve data quality. Multi-channel I / Q samples acquired by the ADC are first deserialized and verified by the JESD204B receiver within the FPGA, then synchronized and converted from parallel to serial by the data pipeline, forming fixed-size data blocks temporarily stored in a BRAM double buffer (using a ping-pong mechanism to avoid read / write conflicts). When the buffer data volume reaches a threshold (80%), a DMA transfer request is triggered. The FPGA-based DMA controller (AXI DMA IP) reads the BRAM data via the AXI bus in block transfer mode and encapsulates the data into PCIe TLP packets using the PCIe endpoint module. The host computer driver pre-allocates contiguous physical memory, and the DMA directly writes the data to this memory via the PCIe bus, without occupying the CPU. ECC and CRC checks are enabled during transmission, triggering retransmission in case of an anomaly; upon receiving a DMA interrupt, the host computer can read the data for subsequent analysis.

[0118] S2, Autocorrelation Enhanced Feature Extraction and Spectrum Sensing: Real-time Prediction of Spectrum Occupation Prior Information.

[0119] The autocorrelation enhancement feature extraction and spectrum sensing in step S2 are roughly as follows:

[0120] S21. Autocorrelation Simulation Sampling Construction: Based on the original low-rate IQ sample sequence obtained by multi-coset sampling, and combined with the time delay configuration of maximizing the autocorrelation sampling point set set in step S12, the sampling values ​​of the autocorrelation function of the original signal on the autocorrelation sampling point set are simulated and calculated through specific delay multiplication and averaging operations to obtain the enhanced autocorrelation feature sequence. The simulated sampling of the autocorrelation function constructed in this way has a higher equivalent sampling rate, which can improve the accuracy of prior information estimation in the signal reconstruction process.

[0121] The specific details of step S21 are as follows:

[0122] Consider the autocorrelation function of the original signal :

[0123] ,

[0124] in Expressing expectations, This indicates conjugate. Since signals generally possess wide stationarity, their autocorrelation function is related to the time difference. Related. Furthermore, due to the Wiener-Khinchin theorem, the autocorrelation function... The Fourier transform of the original signal is exactly the original signal. The frequency domain power spectral density reflects key prior information about its spectral occupancy.

[0125] The autocorrelation function can be estimated using the following formula:

[0126] ,

[0127] in It is the time set of multiple co-set sampling points of multiple ADCs:

[0128] ,

[0129] Is it that the time difference equals The number of sampling pairs. It can be observed that the simulated sampling of the autocorrelation function constructed using this approach is formally equivalent to multi-coset sampling, and its time delay set is equivalent to the time delay difference set constructed by subtracting each pair of elements in the original multi-coset sampling delay set. Furthermore, it can be proven that the simulated sampling constructed in this way has a higher equivalent sampling rate. For example, let... , ,but The sampling rate is from the original Benyquist rate increased to The Nyquist rate is approximately three times the original sampling rate, while preserving band occupancy information.

[0130] For details on how to construct the autocorrelation function, please refer to the following:

[0131]

[0132] The algorithm's input includes batched multi-coset I / Q sampling. Sampling time delay setting undersampling ratio , folding factor and maximum number of anchor points The algorithm outputs the corresponding batched autocorrelation function sampling estimate. .in , The number of elements in the delay difference set.

[0133] Step 1: Calculate key parameters . . Represents the autocorrelation function The range is smaller here. There are two reasons. First, to reduce the amount of computation; second, the autocorrelation function varies with... Increases while decreases, relatively large The autocorrelation at a given point will be drowned out by noise.

[0134] Step 2: Select anchor points for calculating the autocorrelation function and calculate the anchor point spacing. Then, a set of anchor point locations is constructed. The number of anchor points is no greater than The anchor point will be used as the basis for subsequent steps. Calculate the autocorrelation of each point and take the average;

[0135] Step 3: Initialize the autocorrelation output matrix Given an all-zero matrix, calculate the set of time delay differences according to the definition. ,in ;

[0136] Step 4: By index Extract the time delay difference set one of the elements ;

[0137] Step 5: Find all that satisfy the condition ADC sampling channel number pair ;

[0138] Step Six: For each delay difference index (From 0 to ), initialize an empty list Used to store all that meet the requirements The product result under the time delay difference;

[0139] Step 7: For each sampling channel number pair ,if Then calculate (At this point, the corresponding time delay needs to be compensated for the impact of modulo), otherwise the calculation and the calculation results Add to list middle;

[0140] Step 8: [Regarding...] Each anchor point and each group of sampling channels in the middle Take the average to obtain the final autocorrelation simulation sample value, and assign the result to... ;

[0141] Step 9: Repeat steps 6 through 8 until the delay difference index is reached. Traverse from 0 to ;

[0142] Step 10: Repeat steps 4 through 9 until all have been traversed. The elements in, the final This refers to the sampling of the constructed autocorrelation function. express Estimate the autocorrelation function at the given location.

[0143] With a fixed time delay setting for multiple coset undersampling, the actual index required for computation is calculated for each batch of input. (Anchor point) (Time-delay indexes) are all predetermined and remain unchanged. Therefore, all the indexes required for computation can be pre-computed and cached as tensor indexes. Furthermore, the batch index access function of tensor computation libraries (such as PyTorch and TensorFlow) can transform the originally serial product operations into efficient batch tensor computations, thereby greatly accelerating the construction process of autocorrelation features and improving the system's ability to process real-time data.

[0144] S22. Embedding and splicing of multidimensional features:

[0145] (1) Multidimensional feature embedding is performed on the original low-rate IQ sample sequence to generate the original feature sequence, which retains the original amplitude and phase information of the signal;

[0146] (2) Multidimensional feature embedding is performed on the enhanced autocorrelation feature sequence to generate an autocorrelation feature sequence, which is rich in the frequency and time delay structure information of the signal;

[0147] (3) The two embedded features are spliced ​​together in multiple dimensions to form a fusion-enhanced feature sequence to achieve feature complementarity.

[0148] The specific details of step S22 are as follows:

[0149] After constructing the simulated sampling of the autocorrelation function, the implementation steps follow the process outlined in S22 to perform multidimensional feature embedding to adapt to the input of the spectrum-aware neural network model. The spectrum-aware neural network is based on Transformer, and the feature embedding consists of two parts: the original sampling embedding (also known as the original feature sequence) and the autocorrelation simulated sampling embedding (also known as the enhanced autocorrelation feature sequence).

[0150] The original sampling embedding step first... Turn to The form of a real number, and then its last two dimensions are reshaped into Shape. The reshaped The input is fed into a multilayer perceptron (MLP) with three linear layers. The MLP uses a Gaussian error linear unit activation function and adds a dropout layer after the first and second linear layers. The input feature size of the MLP is... The output sizes of each layer are respectively and (in (For model feature dimensions), layer normalization is performed after each linear layer. The final result is... The shape tensor, in a Transformer network, is the batch tensor. The code length is The code element dimension is Input.

[0151] The embedding steps for autocorrelation sampling are basically the same as those for original sampling, the difference being that no reshaping is performed after realization, and the input dimension of the MLP consisting of three linear layers is... Finally obtained Shape tensors.

[0152] The generated original feature sequence and the enhanced autocorrelation feature sequence are concatenated along the second dimension of the tensor to obtain... The shape tensor is used for positional encoding embedding, and then classification symbols independent of the input are inserted into the head to obtain a fused enhanced feature sequence, which is used to extract frequency domain information to output spectrum-aware prediction results.

[0153] S23. Deep Neural Network Perception Modeling: The fused and enhanced feature sequence is input into a spectrum perception model based on a deep neural network. The model learns the mapping relationship between low-speed original samples, high-resolution autocorrelation information and the real spectrum distribution through its long-range dependency modeling capability, so as to achieve high-precision spectrum perception.

[0154] The specific details of step S23 are as follows:

[0155] Finally, the fused and enhanced feature sequence is input into a spectrum-aware model based on a deep neural network. The model consists of two Transformer encoder layers and one linear MLP layer. During processing, the fused and enhanced feature sequence is input into the first Transformer encoder layer. A multi-head self-attention mechanism is used to weight and focus different parts of the input, outputting the feature sequence processed by the second Transformer encoder layer as deep correlation features. This process leverages the powerful long-range dependency modeling capability of the Transformer encoder's unique multi-head self-attention mechanism, enabling the extraction of deep correlation features hidden in low-speed original samples and high-resolution autocorrelation simulated sampling information across time-domain steps. The model learns the complex nonlinear mapping relationship between low-overhead sampling data and the true broadband spectrum distribution by weighting and focusing different parts of the input sequence through the multi-head self-attention mechanism. Finally, the model's output layer uses a linear MLP to reduce the dimensionality and classify the features, outputting a feature sequence with... The prediction of whether each sub-band of the width division is occupied serves as prior information for the predicted spectrum occupancy. This information will be used as a constraint for signal reconstruction in step S3.

[0156] S3. Constrained signal reconstruction: Transform into overdetermined equations and achieve high-fidelity effective recovery of the signal.

[0157] The constraint signal reconstruction in step S3 specifically includes:

[0158] S31. Construction of underdetermined equations: Based on the signal observation model of multi-coset sampling and the original low-rate IQ sample sequence, the relationship between the unknown components of the original broadband signal in the frequency domain and the aliased frequency components observed in the low-rate samples is modeled as a linear equation. Due to low-rate sampling, the number of unknowns in the equation (the complete frequency components of the signal) is much greater than the number of knowns (low-rate samples), thus forming compressed sensing underdetermined equations.

[0159] The specific steps of step S31 are as follows:

[0160] First, the sampling results of each ADC channel are... Perform discrete Fourier transforms on each to obtain And can obtain Relationship with the original signal:

[0161] ,

[0162] in , ,and The spectrum matrix to be reconstructed Represents frequency The frequency components at each position, with each row corresponding to the spectrum of a sub-band.

[0163] S32. Prior Information Reduction and Equation Transformation:

[0164] (1) Using the spectrum occupancy prior information vector obtained in step S2, determine the precise location and number of non-zero solution components (i.e., the frequency components that actually exist in the signal) in the underdetermined equation.

[0165] (2) By selectively retaining the unknowns in the equation that correspond to the frequency components of the signal, and simultaneously reducing and eliminating the unknowns that correspond to the blank spectrum, the number of effective unknowns in the underdetermined equation is reduced to a number much smaller than the number of knowns.

[0166] (3) The operation effectively transforms the original underdetermined equation into an overdetermined equation, thereby eliminating the uncertainty of the solution.

[0167] The specific details of step S32 are as follows:

[0168] use This indicates the sub-band occupancy status predicted by the neural network in step S2, since the unoccupied sub-bands correspond to... Since the behavior is 0, the signal reconstruction equation is reduced by the result of spectrum sensing in step S2:

[0169] ,

[0170] in For the occupied sub-band sequence number The submatrix formed by the columns, For occupied sub-band corresponding The submatrix formed by rows.

[0171] S33. Precise Least Squares Solution: The least squares estimation algorithm based on decision rules is used to solve the overdetermined equations. The key to the solution process is that it does not rely on the sparsity assumption of the signal itself, but relies entirely on the strong constraint of the spectrum occupying prior information. The algorithm accurately calculates and recovers all physical parameters of the signal, such as the original amplitude, phase and frequency.

[0172] The specific details of step S33 are as follows:

[0173] When the number of channels in multi-co-collection sampling is greater than the number of occupied sub-bands, The number of rows is greater than the number of columns, and due to the pre-set time delay of each channel... It has low cross-correlation and constrained isometry, regarding The linear equation is an overdetermined equation, therefore the individual frequency components of the signal can be accurately solved:

[0174] .

[0175] S34. Sub-band signal separation and output: Based on the frequency position information actually occupied by the signal determined in S32, the reconstructed complete signal is frequency separated in units of sub-bands to separate each independent, non-aliased, high-fidelity sub-band signal as an effective recovered signal. The effective recovered signal can be directly used for subsequent complex physical layer analysis.

[0176] The specific details of step S34 are as follows:

[0177] Based on the frequency position information actually occupied by the signal determined in step S32, the reconstructed complete spectrum matrix is ​​frequency-separated, unit by unit, using sub-bands. Specifically, for each sub-band, the frequency component corresponding to the i-th row of the spectrum matrix is ​​extracted, and an inverse Fourier transform is performed on it to obtain the time-domain signal corresponding to that sub-band. In this way, each independent, non-aliased, high-fidelity sub-band signal is separated as an effective recovered signal. These effective recovered signals retain all the key information of the original signal and can be directly used for subsequent complex physical layer analysis, such as protocol type identification, data demodulation, and physical layer information extraction.

[0178] S4, Intelligent Physical Layer Analysis: Performs in-depth analysis and data extraction on the recovered signal.

[0179] This step uses a Transformer-based signal analysis model deployed in series to perform protocol type identification, data demodulation, and in-depth analysis of physical layer information on the recovered signal reconstructed and separated in step S3, thereby achieving a comprehensive interpretation of the broadband wireless signal.

[0180] Step S4, the intelligent physical layer analysis, specifically includes:

[0181] S41, Encoder-Decoder Shared Architecture: This architecture constructs a cascaded deep neural network consisting of a shared Transformer encoder, a protocol recognition decoder, and a data demodulation decoder (also known as a protocol field decoder). It also includes two independent feature embedding modules for performing feature embedding, collectively building a... Figure 3 The signal analysis model based on Transformer shown here is based on maximizing feature reuse and process integration.

[0182] (1) The first and second independent feature embedding modules are responsible for shallow feature extraction of the effective recovery signal through feature embedding, and obtain the first embedded feature and the second embedded feature respectively;

[0183] (2) The shared Transformer encoder consists of a multi-layer stacked Transformer encoder structure, which is responsible for deep feature extraction from the first embedded features extracted from the effective recovered signal to obtain general deep features;

[0184] (3) The protocol identifier consists of a shared Transformer encoder and a protocol identifier decoder. The protocol identifier decoder is composed of a single-layer Transformer decoder structure and is dedicated to fast and high-accuracy signal type classification.

[0185] (4) The data demodulator consists of a shared Transformer encoder and a data demodulator decoder. The data demodulator decoder consists of a multi-layer stacked Transformer decoder structure and is responsible for extracting physical layer data.

[0186] S42. Deep Protocol Type Identification:

[0187] In the workflow, the shared Transformer encoder receives the effectively recovered signal reconstructed and separated in step S3 and extracts deep features. Subsequently, the protocol identification decoder receives these deep features and uses a specially pre-set, input-independent query symbol as the second part of the input to the protocol identification decoder, thereby extracting the global features of the signal and outputting the wireless communication standard and protocol type of the signal in real time. This Transformer-based deep identification capability can effectively handle signal classification in complex channels and high-concurrency scenarios.

[0188] S43, Encoder-Decoder Adaptive Demodulation and Data Recovery:

[0189] This process employs feature fusion to ensure the robustness of demodulation.

[0190] (1) Feature fusion: The deep features extracted by the shared Transformer encoder are fused with the second embedded features to obtain the fused features;

[0191] (2) Data demodulation: The fusion feature and the demodulation symbols that are specially set in advance and whose number is equal to the number of bits of the corresponding protocol decoding field are used as the input of the data demodulation decoder to extract the information of the corresponding bits by demodulating the effective recovered signal.

[0192] This step, based on the signal identification result (i.e., protocol type), automatically matches and executes the corresponding reverse processing pipeline, and uses the set demodulation symbols to extract the information of the corresponding bits of demodulation, thereby realizing end-to-end decoding from signal features to physical layer public data fields.

[0193] S44. Two-stage training method: To optimize the performance of this complex shared architecture, the training of the deep neural network architecture adopts a partially frozen two-stage training mode, which includes:

[0194] (1) In the first stage of training: the parameters of the Transformer encoder and the protocol recognition decoder are shared during joint training, while the parameters of the data demodulation decoder and the second independent feature embedding module are frozen to ensure that the encoder can learn and converge to high-quality signal features first.

[0195] (2) In the second stage of training: the parameters of the shared Transformer encoder and protocol recognition decoder are frozen, and only the parameters of the data demodulation decoder and the second independent feature embedding module are unfrozen and trained to utilize stable encoder features, specifically optimize demodulation and data recovery capabilities, and ensure the accuracy and robustness of the entire system.

[0196] Experimental verification results:

[0197] The following provides a performance evaluation of the aforementioned Transformer-based low-overhead sampling analysis method for broadband wireless signals. The evaluation dimensions include spectrum sensing accuracy, signal recovery quality, protocol identification accuracy, and protocol public field decoding performance, all of which are based on verification using actual collected datasets.

[0198] Experimental Setup: For hardware, a software-defined radio platform was used to build the experimental environment. The transmitting end included multiple Ettus B210 software-defined radio devices for generating Wi-Fi signals; the receiving end consisted of one high-speed sampling device configured with a sampling rate of 1 GHz. The data processing unit used Ubuntu 22.04 and was equipped with an NVIDIA RTX 4090 GPU. Software implementation was based on Python 3.10.11 and PyTorch 2.0.1.

[0199] The model training used a synthetic dataset containing Wi-Fi signals such as IEEE 802.11g / b and IEEE 802.11n; the total bandwidth of the signals was 2 GHz, simulating a real wireless channel (including Rician / Rayleigh fading), with a sampling rate of 2 GHz (Nyquist rate). Multiple co-set undersampled samples were generated after upsampling in the digital domain (single ADC sampling rate 50 MHz). This dataset contains 300,000 samples. The actual collected dataset consisted of the aforementioned Wi-Fi signals acquired in the 5.0-5.8 GHz frequency band. Multiple co-set undersampled samples were generated after upsampling at 2 GHz and sampling in the digital domain, containing 2,000 test samples.

[0200] The benchmark methods selected for comparison are: the spectrum-sensing Transformer model Wireless Transformer (WT), the traditional compressed sensing method signal reconstruction algorithm, the deep learning compressed sensing algorithm, and the protocol recognition Transformer model T-Prime.

[0201] Spectrum sensing performance: On actual acquired datasets, the spectrum sensing accuracy of the method described in this invention is significantly better than that of the benchmark method. This is evident in the ROC curves of the actual acquired datasets (e.g., Figure 4 As shown in the figure, the method maintains a 99% detection rate and a 1% false alarm rate at a lower sampling rate, while WT requires a higher sampling rate to achieve similar performance. Figure 5 As shown, after replacing the original samples with the constructed autocorrelation-simulated samples, the signal detection accuracy of the traditional compressed sensing method is significantly improved, enabling accurate signal detection even under extremely low sampling rates. Furthermore, while... Figure 5 As can be seen, the training speed of the deep learning spectrum sensing method of this invention is also significantly improved after using the constructed autocorrelation simulated sampling as an additional feature.

[0202] Signal reconstruction quality: such as Figure 6 As shown, the signal reconstruction SNR of the method of the present invention is significantly higher than that of the benchmark method. In actual signal recovery, the signal recovered by the method of the present invention has an SNR improvement of approximately 3 dB compared to existing methods.

[0203] Protocol identification and protocol public field decoding performance: In protocol identification tasks, such as... Figure 7 As shown, the method of this invention outperforms the benchmark method T-Prime across the entire SNR range. This is attributed to the shared features extracted by the Transformer encoder and the feature extraction and fusion capabilities of the decoder. At a relative sampling rate of 1.25 and an SNR of 6 dB, the method achieves an accuracy exceeding 90%, while T-Prime requires an SNR of 8 dB to achieve similar performance. In the open field decoding task, this method achieves near-lossless decoding even with undersampling, with an average bit decoding rate of 99.8% for IEEE 802.11g / b and an MCS and frame length accuracy exceeding 99%, meeting network monitoring requirements.

[0204] The advantages of the analytical method proposed in this embodiment include the following:

[0205] (1) By using a multi-co-collected under-Nyquist sampling design, the problem of traditional sampling relying on Nyquist rate and high hardware cost is solved. Multi-band signals are collected at a total sampling rate much lower than the Nyquist rate (such as reduced to 1 / 40 of the signal bandwidth). Combined with the JESD204B receiver in the FPGA and DMA high-speed transmission, the CPU bottleneck is avoided, and it is suitable for broadband scenarios such as 4G / 5G and Wi-Fi, which greatly reduces hardware and transmission overhead.

[0206] (2) By using autocorrelation enhancement features and Transformer perception, the problem of traditional methods being limited to sparse spectrum is overcome. Autocorrelation features are constructed based on optimized time delay configuration, and after being fused with the original samples, they are input into the Transformer. Even in non-sparse scenarios, the spectrum prediction accuracy can still reach over 99%, accurately capturing multi-band concurrent signals.

[0207] (3) The problem of non-sparse signal recovery failure is solved by a two-stage process of "perception-constraint reconstruction". The underdetermined equation is transformed into an overdetermined equation by using the spectrum occupancy prior. The signal is recovered with high fidelity by solving the least squares solution. The recovered MSE in non-sparse scenarios is close to the reference level.

[0208] (4) By connecting the Transformer analysis module in series, the high complexity of traditional physical layer analysis is simplified. The shared Transformer encoder extracts features, the protocol identification decoder quickly outputs the Wi-Fi (such as IEEE 802.11g / b) type, and the decoding module completes the decoding of the protocol public fields end-to-end without complex processing procedures, achieving an identification accuracy of over 99% and near lossless decoding.

[0209] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0210] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0211] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0212] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0213] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0214] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0215] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A low-overhead sampling analysis method for broadband wireless signals, characterized in that, Includes the following steps: The original low-rate IQ sample sequence of the target broadband signal is obtained, and multi-dimensional feature embedding is performed to obtain the original feature sequence. The original low-rate IQ sample sequence is subjected to autocorrelation simulation sampling processing to obtain an enhanced autocorrelation feature sequence, which is then concatenated with the original feature sequence to form a fused enhanced feature sequence. The fused enhanced feature sequence is input into a spectrum sensing model based on a deep neural network for processing, and the spectrum occupancy prior information in the target frequency band is predicted in real time. This information is then used as a signal reconstruction constraint to reconstruct the original low-rate IQ sample sequence, thereby obtaining an effective recovered signal. The effective recovered signal was analyzed using a Transformer-based signal analysis model to obtain in-depth analysis results.

2. The low-overhead sampling analysis method for broadband wireless signals according to claim 1, characterized in that, The original low-rate IQ sample sequence was obtained by using multi-coset sampling on a deployed under-Nyquist platform. The steps for obtaining the original low-rate IQ sample sequence include: Let the set of sampling delays of each analog-to-digital converter on the under-Nyquist platform be . ,in The number of analog-to-digital converters, For the first Sampling delay of each analog-to-digital converter; Each analog-to-digital converter performs multi-channel time-interleaved parallel sampling of the target broadband signal according to its configured sampling delay, resulting in a multi-channel independent, time-interleaved initial raw low-rate IQ sample sequence. The sampling results of each analog-to-digital converter are represented as follows: ， In the formula, For the first The sampling results of the analog-to-digital converter, For the target broadband signal, Indicates the first One sampling point, The ratio of the sampling time interval for each analog-to-digital converter to that of conventional Nyquist sampling. This indicates the sampling time interval of the Nyquist sampling. The target total bandwidth of the spectrum. For the first Sampling delay of each analog-to-digital converter; The initial raw low-rate IQ sample sequence is subjected to time delay calibration, phase synchronization between channels, and noise filtering to obtain the final raw low-rate IQ sample sequence.

3. The low-overhead sampling analysis method for broadband wireless signals according to claim 2, characterized in that, The step of obtaining the original low-rate IQ sample sequence further includes processing the sampling delay set. Optimization will be carried out, specifically including the following: The sampling delay set Perform pairwise subtraction on the elements in the set of sampling delays. Optimize by adjusting The time delay value in the set of time delay differences Maximizing the number of elements in the set completes the optimization process, where the delay difference set is... Represented as: ， In the formula, For the first Sampling delay of each analog-to-digital converter; By adjusting the sampling delay set The time delay value in the set of time delay differences, so that the time delay difference set The optimization process is completed by maximizing the number of elements in the array.

4. The low-overhead sampling analysis method for broadband wireless signals according to claim 1, characterized in that, The steps for performing autocorrelation simulation sampling include: (1) Select the autocorrelation function used for the original low-rate IQ sample sequence. for: ， In the formula, For the target broadband signal, For the time difference, Expressing expectations, Indicates conjugate; Then the autocorrelation function Sampling estimation for: ， ， In the formula, To satisfy the time difference equals The number of sample pairs, To represent the time interval between two sampling points, the independent variable corresponds to the autocorrelation function. discrete values, Represents a set The specific sampling time points in the data. For the current moment, It is separated from it Past sampling times, For the time set of multiple coset sampling points of a multiplex analog-to-digital converter, For the first One sampling point, The ratio of the sampling time interval for each analog-to-digital converter to that of conventional Nyquist sampling. This indicates the sampling time interval of the Nyquist sampling. The target total bandwidth of the spectrum. For the first Sampling delay of each analog-to-digital converter; (2) Set the batched original low-rate IQ sample sequence as Calculate parameters , ,in, In the autocorrelation function Scope For the time difference, The number of sampling points. Folding factor This represents the maximum number of anchor points. (3) Select anchor points for calculating the autocorrelation function and calculate the anchor point spacing. Construct a set of anchor points The number of anchor points is no greater than The anchor point spacing The calculation expression is: ， (4) Initialize the autocorrelation function sampling estimate Define a set of time delay differences for a matrix consisting entirely of zeros. The number of elements in the middle is ; (5) By index Select elements in ; (6) Find all that satisfy Analog-to-digital converter sampling channel number pair ,in For the first Sampling delay of each analog-to-digital converter; (7) For each delay difference index Its range is 0 to Initialize an empty list Used to store all that meet the requirements The product result under the time delay difference; (8) For each of the sampling channel number pairs If satisfied Then calculate If not satisfied Then calculate and the calculation results Add to Among them For analog-to-digital converters and The time difference near each anchor point is The product of the sampling point pairs, superscript For complex conjugate; (9) Each anchor point and each group of sampling channels in the middle Take the average value to obtain the autocorrelation simulation sample value, and assign it to... ; (10) Repeat steps (7)-(9) until the time delay difference index is reached. Traversing from 0 to ; (11) Repeat steps (5)-(10) until all iterations are completed. The middle element, the final result express Sampling estimation of the autocorrelation function at the location.

5. The low-overhead sampling analysis method for broadband wireless signals according to claim 1, characterized in that, The steps for obtaining the original feature sequence include: The original low-rate IQ sample sequence Convert to Real numbers of the form, and then reshape their last two dimensions into The shape is reshaped , wherein For complex fields, For the real number field, For batch number, The length of the symbol. The number of sampling points. The number of analog-to-digital converters, This represents the number of adjacent sampling points in the stack. The reshaped The input is processed in a multilayer perceptron containing three linear layers to obtain... The original feature sequence of the shape is used in the multilayer perceptron, which employs the Gaussian error linear unit activation function. Dropout layers are added after the first and second linear layers. The output size of each linear layer is... and After processing each linear layer, a layer normalization operation is performed. This represents the network feature dimension.

6. The low-overhead sampling analysis method for broadband wireless signals according to claim 5, characterized in that, The steps for obtaining the fusion-enhanced feature sequence include: The autocorrelation function sampling estimate obtained by autocorrelation simulation sampling processing The result is converted to a real number and then input into the multilayer perceptron containing three linear layers for processing, to obtain... Shape-enhanced autocorrelation feature sequences, In the autocorrelation function Scope; The enhanced autocorrelation feature sequence and the original feature sequence are concatenated along the second dimension to obtain... Initial fusion enhancement feature sequence of shape; The initial fused enhanced feature sequence is embedded with positional encoding, and a classification code is inserted at the head to obtain the fused enhanced feature sequence.

7. The low-overhead sampling analysis method for broadband wireless signals according to claim 1, characterized in that, The spectrum sensing model based on deep neural networks includes two Transformer encoders with multi-head self-attention mechanisms and one linear multilayer perceptron. The prediction step of the spectrum occupancy prior information includes: The fused and enhanced feature sequence is input into the first-layer Transformer encoder. The different parts of the input are weighted and focused through a multi-head self-attention mechanism, and the feature sequence processed by the first-layer Transformer encoder is output. The feature sequence processed by the first Transformer encoder is used as input to the second Transformer encoder. The different parts of the input are weighted and focused through a multi-head self-attention mechanism, and the feature sequence processed by the second Transformer encoder is output as deep association features. The linear multilayer perceptron is used to perform dimensionality reduction and classification on the deep correlation features, and the output is... The prediction of whether each sub-band of the bandwidth division is occupied serves as prior information for the predicted spectrum occupancy. The target total bandwidth of the spectrum. The ratio of the sampling time interval for each analog-to-digital converter to that of conventional Nyquist sampling.

8. The low-overhead sampling analysis method for broadband wireless signals according to claim 1, characterized in that, The steps for obtaining a valid recovered signal include: The sampling results of each analog-to-digital converter in the original low-rate IQ sample sequence Perform discrete Fourier transforms on each sample to obtain the transformed original low-rate IQ sample sequences. Furthermore, underdetermined equations are constructed to describe The relationship between the original low-rate IQ sample sequence and the original low-rate IQ sample sequence is expressed as follows: ， In the formula, This is the phase compensation matrix. For complex fields, The number of analog-to-digital converters, The number of sampling points. For the first The first channel in the Phase compensation term at each frequency point For the first The sampling delay of an analog-to-digital converter, The ratio of the sampling time interval for each analog-to-digital converter to that of conventional Nyquist sampling. For multi-set sampling measurement matrix, For the first The analog-to-digital converter in the first... The mixing mode generated by each frequency point The spectrum matrix to be reconstructed Represents frequency The frequency components at each position, with each row corresponding to the spectrum of a sub-band. In the autocorrelation function Scope For the time difference, The target total bandwidth; Using the aforementioned spectrum occupancy prior information Reducing the underdetermined equations, we obtain the overdetermined equations, expressed as: ， In the formula, For the occupied sub-band sequence number The submatrix formed by the columns, For occupied sub-band corresponding The submatrix formed by the rows; The overdetermined equations are solved using the least squares method, yielding the following results. This forms the reconstructed spectrum matrix. ,in The expression is: ， In the formula, It is the conjugate transpose; Extract the reconstructed spectrum matrix in units of sub-bands. frequency components Perform an inverse Fourier transform on it to obtain the sub-band signal corresponding to each sub-band, which is used as the effective recovery signal.

9. The low-overhead sampling analysis method for broadband wireless signals according to claim 1, characterized in that, The Transformer-based signal analysis model includes a cascaded shared Transformer encoder, a protocol identification decoder, and a data demodulation decoder, and also includes an independent feature embedding module. The shared Transformer encoder comprises multiple stacked Transformer encoders, and the protocol identification decoder consists of a single-layer Transformer decoder. The steps for obtaining the deep analysis results include: The effective recovered signal is first subjected to shallow feature extraction using the first and second independent feature embedding modules respectively to obtain the first embedded feature and the second embedded feature respectively. The first embedded feature is then input into the shared Transformer encoder to extract the deep features for signal analysis. Set query symbols, and use them together with the deep features as input to the protocol identification decoder to extract global features from the effective recovered signal, and output the wireless communication standard and protocol type of the effective recovered signal in real time; The deep features extracted by the shared Transformer encoder are fused with the second embedded features to obtain fused features. The fused features and demodulated symbols of a number equal to the number of bits in the corresponding protocol decoding field are used as input to the data demodulation decoder to extract information of the corresponding bits from the effective recovered signal demodulation.

10. A low-overhead sampling analysis method for broadband wireless signals according to claim 9, characterized in that, The Transformer-based signal analysis model is trained using a two-stage training method. The specific training process includes: During the first phase of training: The parameters of the shared Transformer encoder and protocol recognition decoder are trained jointly, while the parameters of the data demodulation decoder and the second independent feature embedding module are frozen. In the second phase of training: Freeze the parameters of the shared Transformer encoder and protocol identification decoder, and unfreeze and train only the parameters of the data demodulation decoder and the second independent feature embedding module.

Images