A multi-agent cooperative configuration method for wireless signal identification

Abstract

This invention relates to the field of wireless signal recognition technology, specifically to a multi-agent collaborative configuration method for wireless signal recognition. The method involves: acquiring the wireless signal task state and constructing a hierarchical configuration search tree; determining the stages to be optimized through planning agents and controlling corresponding dedicated agents to generate atomic actions; generating sub-configurations when constraints are met based on a total judgment of operational domain, tensor shape, processing order, label retention, and device constraints; obtaining pairing benefits, interval benefit vectors, and resource changes through single-factor counterfactual pairing evaluation; updating the experience reliability through a task-conditionalized experience base and iteratively searching for target pipeline configurations. This invention reduces unexecutable configuration evaluations, improves the efficiency of wireless signal recognition configuration search, and enhances recognition robustness in complex signal-to-noise ratio environments.

Description

Technical Field

[0001] This invention relates to the field of wireless signal recognition technology, and more specifically to a multi-agent collaborative configuration method for wireless signal recognition. Background Technology

[0002] Wireless signal identification is a crucial supporting technology for intelligent communication, spectrum sharing, and wireless security, providing key capabilities for tasks such as spectrum management, device authentication, and anomaly detection. With the development of deep learning, data-driven wireless signal identification methods have achieved good results. However, their performance is highly dependent on a complete processing pipeline configuration, typically including data preprocessing, representation selection, data augmentation, and network architecture design. Different tasks and datasets often require different pipeline combinations. Manually designing such pipelines not only heavily relies on domain expert experience but also incurs high parameter tuning costs and long iteration cycles, often making it difficult to achieve an ideal balance between accuracy and efficiency.

[0003] Existing automated design methods mostly employ traditional search paradigms, such as evolutionary algorithms, reinforcement learning, or gradient-based search. These methods typically rely on a predefined search space. While they can optimize within a certain range, their search boundaries are relatively fixed, making it difficult to fully utilize domain knowledge and providing a clear explanation of the final solution generation process. Meanwhile, although agent-based methods using large language models have begun to be applied to automated design in recent years, existing research has focused more on parameter-level optimization of existing processing flows or design tasks in other domains. Cooperative search for end-to-end pipelines in wireless signal recognition remains significantly insufficient. Particularly for discrete time-series data like wireless signals, the selection of input representations is often crucial, and this step is still not adequately automated in many existing methods.

[0004] Therefore, there is an urgent need for a method that can actually apply state information such as the signal-to-noise ratio distribution, frequency offset, phase perturbation, sampling rate, occupied bandwidth, and deployment constraints of wireless signals to the candidate pipeline generation and search process. At the same time, it is necessary to solve the problems of difficulty in attributing benefits when multiple pipeline stages change simultaneously, insufficient reliability of directly reusing historical experience across tasks, and the tendency of general Monte Carlo tree search to generate data domains or tensor shapes with incompatible configurations, so as to improve the proportion of effective candidates and robustness against low signal-to-noise ratio within a limited training and evaluation budget. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a multi-agent collaborative configuration method for wireless signal recognition. This method effectively solves the problems in existing technologies, such as reliance on human experience in wireless signal recognition pipeline design, difficulty in ensuring physical compatibility between candidate configurations, inability to accurately attribute benefits when multiple modules are changed simultaneously, and high invalid search ratios due to unutilized wireless signal states in general tree search.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a multi-agent cooperative configuration method for wireless signal identification, comprising at least:

[0008] Extract the wireless signal task state and construct a hierarchical configuration search tree. The search nodes in the hierarchical configuration search tree store the complete configuration as well as the intermediate states of the preprocessing, characterization, enhancement and network architecture stages.

[0009] The planning agent receives wireless signal task status, accuracy of each signal-to-noise ratio interval of the current node, resource consumption, and historical edge gains. The selected dedicated agent outputs atomic actions, which change the configuration of one stage.

[0010] Read the intermediate state and atomic actions of the parent node, and perform a total compatibility determination based on the operation domain, tensor shape, processing order, label retention and device constraints. If the total compatibility determination passes, perform configuration extension to generate sub-configurations.

[0011] By performing single-factor counterfactual pairing evaluation on the sub-configuration based on compatibility determination, the parent and sub-configurations use the same training set, validation set, batch size, optimizer, learning rate, number of training rounds, and random seed to obtain pairing benefits, interval benefit vectors, and resource changes.

[0012] The task status, atomic actions, compatibility results, interval benefits, and resource changes are encoded as experience items. A task-conditional experience base is constructed and the experience reliability is updated. The target pipeline is obtained by iterative search based on the updated experience reliability.

[0013] Furthermore, the process of the planning agent identifying the phase to be optimized and the selected specialized agent outputting atomic actions includes:

[0014] The plan includes the task status of the intelligent agent receiving wireless signals, the accuracy of the current node in each signal-to-noise ratio range, resource consumption, and historical edge rewards.

[0015] When the accuracy in the low signal-to-noise ratio (SNR) interval is lower than that in other SNR intervals or does not contribute sufficiently to the improvement of robust performance in SNR stratification, the planning agent should prioritize sending robust improvement targets to the preprocessing agent, the characterization agent, or the enhancement agent.

[0016] When resource constraints are violated, the planning agent prioritizes sending the structure compression target to the network architecture agent;

[0017] The selected specialized intelligent agent outputs atomic actions based on the robustness improvement objective or the structural compression objective.

[0018] Furthermore, the process by which an atomic action changes a stage configuration includes:

[0019] Atomic actions include stage number, operation type, and operation parameters;

[0020] Determine one of the stages—preprocessing, characterization, enhancement, and network architecture—based on the stage number, and change the configuration of the corresponding stage according to the operation type and operation parameters.

[0021] The phase configuration corresponding to the atomic action changes, while the other three phase configurations remain unchanged, so that there is only one phase configuration difference between the parent and child configurations, i.e., the number of phase differences is 1.

[0022] Furthermore, the process of performing the overall compatibility determination includes:

[0023] Read the intermediate state and atomic actions of the parent node, and obtain the judgment results of five sub-decisions: operation domain, tensor shape, processing order, label preservation, and device constraints.

[0024] The results of the five sub-decisions are either 0 or 1. The results of the five sub-decisions are multiplied together to obtain the overall compatibility determination result.

[0025] If the result of any sub-decision is 0, the result of the overall compatibility determination is 0, and the configuration extension is not executed.

[0026] When the results of the five sub-decisions are all 1, the result of the overall compatibility determination is 1, and the configuration extension is executed.

[0027] Furthermore, the determination process for the operational domain, tensor shape, and label preservation includes:

[0028] When using a random frequency domain mask, the random frequency domain mask is applied to the spectrum or time-frequency representation;

[0029] When the input representation is changed from IQ dual-channel to short-time Fourier time-frequency diagram, the network input layer synchronously meets the two-dimensional time-frequency tensor size requirement;

[0030] When the enhancement operation involves frequency offset injection, phase perturbation, sampling rate deviation, or IQ imbalance, the parameter range of the enhancement operation is determined based on the corresponding statistics in the wireless signal mission state, and the tag retention conditions before and after enhancement are verified.

[0031] If the compatibility check fails, model training will not be started, and the failure type will be recorded.

[0032] Furthermore, the process of performing a single-factor counterfactual paired evaluation on the sub-configurations that pass the compatibility determination includes:

[0033] Model building and training are performed on the sub-configurations that pass the compatibility test;

[0034] The parent and child configurations use the same training set, validation set, batch size, optimizer, learning rate, number of training epochs, and random seed.

[0035] Calculate the accuracy, FLOPs, number of parameters, and latency of the parent and child configurations in each signal-to-noise ratio range.

[0036] The paired revenue is obtained based on the comprehensive evaluation function value of the parent and child configurations, the interval revenue vector is saved, and the resource changes are obtained based on FLOPs, parameter quantity, and latency.

[0037] Furthermore, the process of obtaining the paired payoff and interval payoff vectors includes:

[0038] Subtract the comprehensive evaluation function value of the parent configuration from the comprehensive evaluation function value of the child configuration to obtain the paired benefit caused by the atomic action;

[0039] Subtract the accuracy of the parent configuration in the first SNR interval from the accuracy of the sub-configuration in the first SNR interval, and so on, until the accuracy of the sub-configuration in the B SNR interval is subtracted from the accuracy of the parent configuration in the B SNR interval. Then, form a sub-interval profit vector in the order of the first to the B SNR intervals.

[0040] Furthermore, the process of constructing a task-conditional experience base includes:

[0041] Encode task status, atomic actions, compatibility results, interval benefits, and resource changes into experience items;

[0042] Context clusters are determined using the distance between standardized task states;

[0043] The number of successful actions within the same context cluster is counted, and the empirical reliability is updated based on historical samples within the same context cluster.

[0044] Furthermore, the process of updating empirical reliability based on historical samples within the same context cluster includes:

[0045] Using the minimum effective return threshold as a comparison benchmark, an indicator function is used to determine whether the pairing return of each historical sample is greater than the minimum effective return threshold. The results of the indicator function for pairing returns greater than the minimum effective return threshold are accumulated into the prior parameters of success.

[0046] The indicator function is used to determine whether the pairing payoff of each historical sample is less than or equal to the minimum effective payoff threshold. The results of the indicator function when the pairing payoff is less than or equal to the minimum effective payoff threshold are added to the prior parameters of the failure.

[0047] The Beta posterior is determined based on the successful and unsuccessful posterior parameters, and the empirical lower confidence bounds of the Beta posterior at quantiles are calculated.

[0048] Furthermore, the method for obtaining the target pipeline through iterative search based on the updated empirical reliability is as follows: the compatible tree edges are scored and selected according to the updated empirical reliability, historical average pairing revenue and expected evaluation cost, and the pairing revenue is back-propagated along the search path to update the search state, and the target pipeline is obtained through iterative search.

[0049] The technical solution provided by this invention has the following advantages compared with the known prior art:

[0050] This invention constrains the search tree expansion by using the wireless signal task state and pipeline intermediate state constraints. Before training candidate configurations, it eliminates actions that are mismatched in data domain, tensor dimension, processing order, label preservation, or deployment resources, reducing unexecutable configurations and invalid training. This invention limits tree edges to single-module atomic modifications and performs counterfactual pairing evaluation of parent-child configurations under the same training conditions. This allows performance changes to be attributed to specific preprocessing, characterization, enhancement, or architectural modifications, reducing empirical pollution caused by simultaneous changes to multiple modules.

[0051] This invention constructs a robust evaluation function based on the recognition accuracy and dispersion of different signal-to-noise ratio (SNR) intervals, suppressing candidate configurations that only perform well in high SNR intervals and improving the stability of search results under complex channel and low SNR conditions. A Beta posterior distribution is established for both task context and atomic actions, and a posterior lower confidence bound is used to measure the reliability of experience. This distinguishes between a small number of accidental successful experiences and stable experiences that have been verified multiple times, avoiding the repeated influence of low-quality historical knowledge on the search.

[0052] This invention combines average pairing rewards, exploration rewards, conditionalized experience reliability, and projected evaluation costs in the selection of parent node-atomic action combinations, changing the approach of solely using standard UCBs to select general parent nodes and improving the effective search ratio under limited evaluation budgets. By transferring structured intermediate states and constraint causes between the planning agent, the dedicated generating agent, and the reflective agent, the input-output relationships of each stage configuration are clarified, enhancing the executability, traceability, and interpretability of the automatic generation pipeline. This invention normalizes computational load, parameter count, and inference latency using reference values ​​and jointly optimizes them with hierarchical robustness performance based on signal-to-noise ratio, ensuring that the output configuration simultaneously meets the requirements of recognition performance and resource-constrained deployment. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0054] Figure 1 This is a schematic diagram of the overall method of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0056] Example 1 (see Figure 1 A multi-agent cooperative configuration method for wireless signal recognition, comprising at least:

[0057] Step 1: Divide the dataset and establish unified evaluation criteria.

[0058] The RadioML 2016.04c dataset for the wireless signal recognition task was divided into training and training sets in a 6:2:2 ratio. Validation set and test set The training set is used for model training in each pipeline configuration, the validation set is used for performance verification of the trained model in each pipeline configuration, and the test set is used to test the final recognition performance of the model obtained in different pipeline configurations.

[0059] In this embodiment, the training set, validation set, and test set are obtained in a 6:2:2 ratio, respectively; all candidate configurations use the same partition, the same random seed set, the same number of training rounds, and the same timing device.

[0060] Step 2: Extract the wireless signal task status and build a hierarchical configuration search tree.

[0061] The configuration c for each pipeline of the wireless signal identification task is represented as:

[0062]

[0063] In this embodiment, , , and These are instantiated by the preprocessing operations, representation methods, enhancement operations, and executable network structures listed below. This indicates the data preprocessing configuration. Indicates the input representation configuration. Indicates data enhancement configuration, This indicates the network architecture configuration.

[0064] The data preprocessing configuration determines the data cleaning and correction operations to be performed on the raw wireless signal, including the selected preprocessing operations, the execution order of each preprocessing operation, and the corresponding processing parameters. For example, the data preprocessing configuration can be represented as:

[0065]

[0066] The filtering operation can include filter type, cutoff frequency, and filter order; the amplitude normalization operation can include normalization method and normalization range; and the frequency offset compensation operation can include frequency offset estimation method and compensation frequency value. Data preprocessing configuration can also include operations such as phase compensation and resampling, and different operations can be selected, combined, and ordered according to the target task and input data characteristics.

[0067] The input representation configuration determines the model input form to which the preprocessed wireless signal is converted. For example, the input representation configuration can select the original in-phase quadrature IQ sequence, or it can select an amplitude-phase sequence, spectrum, time-frequency plot, cyclic spectrum, higher-order cumulants, or entropy features calculated from the IQ sequence. The input representation configuration also includes parameters such as window length, number of transform points, step size, scale, and frequency resolution used when generating the corresponding representation.

[0068] Data augmentation configuration is used to determine the perturbation operations, operation order, execution probability, and perturbation strength applied to wireless signal samples during training. For example, a data augmentation configuration can be represented as:

[0069]

[0070] Among them, Gaussian white noise operation can include target signal-to-noise ratio or noise power; frequency offset injection operation can include frequency offset range; random time masking operation can include mask length and execution probability.

[0071] The network architecture configuration describes the neural network structure that performs wireless signal feature extraction and category prediction, including but not limited to the number of network branches, feature extraction module type, number of network layers, convolutional kernel size, number of channels, feature fusion method, classification head structure, and their connection relationships. In one implementation, the network architecture configuration is not directly selected from a pre-fixed finite set of structures, but is generated and adjusted by the network architecture design agent based on the wireless signal recognition task description, dataset statistical characteristics, and historical search results.

[0072] The above definition maps the originally fragmented pipeline design problem into a searchable configuration space C. The aim is to transform the complex end-to-end design problem into an iteratively optimizeable structured decision problem, thereby providing a unified interface for subsequent collaborative generation and search.

[0073] In one implementation, the data preprocessing search space includes one or more of the following six operations:

[0074] (1) Filtering. Obtain the original complex IQ sequence, and determine the passband and stopband parameters of the filter based on the target signal bandwidth or spectral distribution; process the IQ sequence using a low-pass filter, band-pass filter, median filter, or other digital filter to suppress out-of-band noise and interference; output the filtered IQ sequence to the next processing step. Filter type, filter order, cutoff frequency, and bandwidth can all be used as search parameters.

[0075] (2) Remove DC. Calculate the mean of the I-channel sample and the Q-channel sample in a segment of IQ signal respectively, and subtract the corresponding mean from each sample; or estimate the DC component of the complex IQ sequence and subtract it from the original sequence to reduce the impact of receiver DC bias on signal representation.

[0076] (3) Amplitude normalization. Calculate the maximum amplitude, average power, root mean square value, or other amplitude statistics of the input IQ sequence; scale the IQ samples according to the statistics so that their amplitude, power, or energy are within a predetermined range. For example, the average power of the signal can be normalized to 1, or the sample amplitude can be mapped to a preset interval.

[0077] (4) Frequency offset compensation. The carrier frequency offset is estimated based on the phase change, spectral peak, cyclostationary characteristics, or pilot information of the IQ sequence; a complex compensation sequence is generated based on the estimated frequency offset; the input IQ sequence is multiplied by the complex compensation sequence to correct the carrier frequency offset. The frequency offset estimation method, compensation range, and estimation window length can be used as search parameters.

[0078] (5) Phase compensation. Estimate the phase deviation based on the average phase of the IQ sequence, known pilots, constellation distribution, or phase statistical characteristics; generate a complex rotation factor opposite to the direction of the phase deviation; multiply the IQ sequence by the rotation factor to reduce the influence of fixed phase offset or phase rotation on the recognition result.

[0079] (6) Resampling. Determine the resampling ratio based on the original sampling rate and the target sampling rate; change the sampling rate of the IQ sequence by interpolation, decimation or polyphase filtering to ensure that signals under different acquisition conditions have a uniform number of sampling points, symbol sampling rate or time resolution.

[0080] The input representation selection search space includes one or more of the following twenty-one representations:

[0081] (1) Raw IQ data. The complex baseband signal is split into I and Q channels and input into the neural network in the form of two channels or complex sequences.

[0082] (2) Amplitude-phase representation. For each complex IQ sample, its amplitude and phase are calculated separately, and the amplitude sequence and phase sequence are combined into a dual-channel input.

[0083] (3) Fourier transform representation. Perform discrete Fourier transform or fast Fourier transform on the IQ sequence to obtain the complex coefficients in the frequency domain, and further extract the amplitude spectrum, power spectrum, phase spectrum or real and imaginary parts as input representation.

[0084] (4) Inverse Fourier Transform Representation. Perform an inverse Fourier transform on the signal after frequency domain processing or frequency domain truncation to obtain the corresponding time-domain reconstructed sequence, and use the reconstructed result as the model input.

[0085] (5) Power spectrum characterization. Calculate the square of the magnitude of the Fourier transform of the IQ signal, or obtain the power spectral density using the periodogram method or piecewise averaging method, and normalize or logarithmically transform the result.

[0086] (6) Fractional Fourier Transform Representation. Perform a fractional Fourier transform on the IQ signal according to the candidate fractional order to obtain the energy distribution of the signal under different rotation angles between the time and frequency domains; the fractional order can be used as a search parameter.

[0087] (7) Discrete wavelet transform representation. Select the wavelet basis and the number of decomposition levels, perform multi-scale decomposition on the IQ sequence, obtain the approximation coefficients and detail coefficients at different scales, and combine the coefficients as model input.

[0088] (8) Continuous wavelet transform representation. Select the wavelet basis and scale range, perform continuous wavelet transform on the IQ signal to generate a two-dimensional coefficient graph that varies with time and scale, and convert the coefficient amplitude or power into a time-frequency graph.

[0089] (9) Short-time Fourier transform representation. The IQ signal is divided into frames according to the set window function, window length and sliding step size, and a Fourier transform is performed on each frame to obtain a two-dimensional time spectrum.

[0090] (10) Characterization of Wigner-Ville distribution. The Wigner-Ville time-frequency distribution is calculated based on the instantaneous autocorrelation function of the IQ signal; alternatively, smoothing or pseudo-Wigner-Ville processing is used to reduce cross-term interference.

[0091] (11) Multi-synchronous squeezing transformation characterization. First, time-frequency transformation is performed on the signal, and then the dispersed time-frequency coefficients are redistributed or squeezed multiple times synchronously based on the instantaneous frequency estimation results to form a more concentrated time-frequency characterization.

[0092] (12) Representation of Cyclic Autocorrelation Function. The cyclic autocorrelation value of the IQ signal is calculated for candidate cyclic frequencies and time delays, and the calculation results under different cyclic frequencies and time delays are combined to form a cyclic autocorrelation feature matrix.

[0093] (13) Characterization of spectral correlation density. Perform a Fourier transform on the cyclic autocorrelation function along the time delay dimension to obtain the cyclic frequency-frequency two-dimensional spectral correlation density, and use its amplitude or power as input.

[0094] (14) Bispectral characterization. The bispectral or bicoherent spectrum on the dual-frequency plane is calculated based on the third-order statistics of the IQ signal to characterize the non-Gaussianity and frequency coupling relationship of the signal.

[0095] (15) Characterization of second-order cumulants. Calculate the second-order cumulants of the IQ sequence under one or more time delays, and form them into an eigenvector or eigenmatrix.

[0096] (16) Characterization of third-order cumulants. Calculate the third-order cumulants of the IQ sequence under multiple time delay combinations to extract higher-order statistical features from the signal.

[0097] (17) Characterization of fourth-order cumulants. Calculate the fourth moment of the IQ sequence and subtract the terms formed by the combination of lower-order moments to obtain the fourth-order cumulant characteristics; one or more typical fourth-order cumulants can be selected as input.

[0098] (18) Higuchi fractal dimension characterization. Subsequences of IQ amplitude sequences or phase sequences are constructed according to multiple step sizes. The curve lengths under different step sizes are calculated, and the fractal dimension is obtained by logarithmic fitting.

[0099] (19) Permutation entropy representation. The IQ amplitude, phase or other one-dimensional sequence is reconstructed according to the set embedding dimension and time delay. The permutation pattern of the local sequence is statistically analyzed, and the entropy value is calculated according to the probability of the permutation pattern.

[0100] (20) Approximate entropy representation. Construct subsequences according to the set embedding dimension and similarity threshold, count the occurrence probability of similar subsequences, and calculate the sequence complexity.

[0101] (21) Sample entropy representation. Construct a fixed-length subsequence, count the number of matching subsequences under a given distance threshold, and calculate the sample entropy based on the matching probability before and after the length is increased.

[0102] In one implementation, the data augmentation search space includes one or more of the following twenty-one operations:

[0103] (1) Gaussian white noise. Generate complex Gaussian noise based on the set signal-to-noise ratio or noise power, and then superimpose it onto the original IQ sequence.

[0104] (2) Rayleigh channel simulation. Generate complex channel coefficients that satisfy the Rayleigh distribution, and convolve the input IQ signal with the channel coefficients or multiply them sample by sample to simulate a pathless fading channel.

[0105] (3) Frequency offset injection. A complex exponential sequence is generated based on a randomly selected or preset frequency offset, and the complex exponential sequence is multiplied by the IQ signal.

[0106] (4) Phase noise. Generate a random phase perturbation sequence or a phase sequence determined by a phase noise power spectrum mask, and apply a time-varying phase rotation to the IQ signal.

[0107] (5) Sampling rate deviation. The signal is interpolated and resampled according to a randomly selected sampling rate deviation ratio to simulate the inconsistency between the sampling clocks of the transmitting and receiving ends.

[0108] (6) IQ imbalance. Change the amplitude gain of the I and Q paths respectively, and introduce a phase error between the two signals to simulate the non-ideal quadrature demodulation of the receiver.

[0109] (7) Symbol sampling offset. The sampling position is offset by an integer or non-integer relative to the ideal symbol center, and the offset IQ sequence is obtained by interpolation.

[0110] (8) Multipath simulation. Generate a channel impulse response containing multiple path delays, attenuation coefficients and phases, and convolve the impulse response with the IQ signal.

[0111] (9) Doppler frequency shift. A complex rotation factor that varies with time is generated based on the set Doppler frequency and applied to the IQ signal.

[0112] (10) Energy scaling. The amplitude of the IQ signal is scaled according to a random or preset ratio to simulate different received power and automatic gain control states.

[0113] (11) Time flipping. The IQ sequence is reversed along the time dimension to obtain the time-flipped enhanced sample.

[0114] (12) Cyclic Time Shift. The IQ sequence is cyclically shifted by a random offset, and the shifted samples are added to the other end of the sequence.

[0115] (13) Time domain compression. The IQ signal is resampled according to the set compression ratio to compress the signal on the time axis, and truncation or padding is performed according to the target input length.

[0116] (14) Selective frequency domain attenuation. The IQ signal is converted to the frequency domain, one or more frequency intervals are randomly selected and the amplitude of the corresponding frequency coefficients is reduced, and then the signal is restored to the time domain through inverse transformation.

[0117] (15) Mixup. Randomly select two training samples and their labels, and perform a weighted combination of the two samples and their labels based on the mixing coefficients obtained from the preset distribution.

[0118] (16) Sample mixing. Select two or more signal samples and splice, superimpose or partially replace them in the time domain, frequency domain or channel dimension to generate new training samples.

[0119] (17) Random time mask. Randomly select a continuous interval in the time dimension of the IQ sequence or time-frequency representation, and set the corresponding interval to zero, mean or random value.

[0120] (18) Random frequency domain mask. Randomly select a continuous frequency band in the frequency dimension of the spectrum or time-frequency representation, and set the frequency band to zero or attenuate it.

[0121] (19) Gaussian kernel processing. Use a Gaussian kernel with a preset or random standard deviation to smooth the time-domain sequence, spectrum or time-frequency plot to simulate resolution changes or local blurring.

[0122] (20) Random amplitude transformation. Random amplitude coefficients are generated for the entire signal or different time segments, and the coefficients are applied to the IQ signal to form a global or local amplitude perturbation.

[0123] (21) Random phase shift. Randomly generate phase shift angles and multiply all IQ samples by the corresponding complex rotation factor; or generate a slowly changing phase shift sequence for different time positions.

[0124] For the network architecture search space, the network architecture design agent first receives the target wireless signal identification task, input representation dimension, sample length, number of categories, data scale, and signal-to-noise ratio range. Then, based on this information, it determines the overall network structure and generates an architecture description including network layer types, inter-layer connections, and network parameters. Next, it constructs an executable neural network model based on the architecture description. The generated network architecture can include single-branch or multi-branch networks. Different branches are used to process the original IQ sequence, frequency domain features, time-frequency plots, or statistical features. Each branch can use convolutional neural networks, recurrent neural networks, Transformer networks, attention modules, or combinations thereof for feature extraction. Features from different branches are combined through concatenation, summation, weighted fusion, attention fusion, or other fusion methods, and the wireless signal category prediction result is output through a fully connected layer, a global pooling layer, or a classification head. The network architecture design agent can also modify the number of network layers, number of channels, convolutional kernel size, number of attention heads, hidden layer dimensions, feature fusion method, and classification head structure based on the candidate network's recognition accuracy, number of parameters, floating-point operation volume, inference latency, or other evaluation results, thereby generating new candidate network architectures.

[0125] Based on the aforementioned pipeline configuration space, the wireless signal task status is further extracted from the RadioML 2016.04c training samples and the target deployment environment. This embodiment at least statistically analyzes the sample ratio, sample sequence length, sampling rate, normalized bandwidth, frequency offset statistics, phase change statistics, fourth-order cumulant, and time-frequency concentration for each signal-to-noise ratio interval, and incorporates the maximum number of parameters, maximum floating-point operations, and maximum inference latency allowed by the target computing device into the deployment constraints.

[0126]

[0127] This embodiment uses all the symbol definitions from step two of the invention. Among them, the dataset... Specifically, it is RadioML2016.04c; It represents the proportion distribution of samples corresponding to each signal-to-noise ratio interval in the dataset, and is used to describe the composition of samples with different signal-to-noise ratios; Statistics representing carrier frequency offset in a dataset, such as average frequency offset or frequency offset estimate, are used to describe the overall frequency offset characteristics of the dataset. The variance represents the phase change of the signal and is used to characterize phase stability and phase noise intensity. Indicates the bandwidth occupied by the signal; Indicates the sampling rate; This indicates the number of sampling points contained in each training sample; The cyclic stationarity index is used to describe the cyclic spectrum characteristics or periodic statistical properties of a signal. It represents the fourth-order cumulant of a signal, used to characterize higher-order statistical features and modulation characteristics; Indicates time-frequency concentration, used to describe the degree of concentration of signal energy in the time-frequency plane; It consists of the parameters, maximum floating-point operations, and latency limit given by the target device.

[0128] Each search node stores the complete configuration c, as well as the intermediate states of the preprocessing, characterization, enhancement, and network architecture stages. For example, the intermediate state represented by the short-time Fourier transform records at least two-dimensional time-frequency domain, number of frequency points, number of time frames, and number of channels; the network architecture agent generates the network input layer with input channels and two-dimensional feature dimensions matching accordingly. IQ represents in-phase-quadrature complex baseband signal, where I represents the in-phase component and Q represents the quadrature component; SNR represents signal-to-noise ratio.

[0129] Step 3: Define a multi-objective evaluation function for hierarchical robustness of signal-to-noise ratio.

[0130] The validation set is divided into B intervals according to the signal-to-noise ratio label in the dataset, and the accuracy of each interval is calculated. In one implementation, the low signal-to-noise ratio (SNR) interval is assigned a weight no less than that of the high SNR interval, and a performance dispersion penalty is used to prevent search results from only improving high SNR samples.

[0131]

[0132]

[0133]

[0134] This represents the accuracy of the hierarchical weighted average of the signal-to-noise ratio. Indicates configuration The corresponding maximum number of floating-point operations; Indicates configuration The corresponding number of model parameters; Indicates configuration The corresponding inference delay; , , These represent the corresponding reference values. All candidate configurations use the same statistical caliber, inference device, warm-up times, and official timing times.

[0135] In this embodiment, the total number of signal-to-noise ratio intervals Determined based on the actual number of SNR tags in RadioML 2016.04c; This represents the total number of signal-to-noise ratio intervals. Indicates a range index; Indicates candidate configuration In the The recognition accuracy of each signal-to-noise ratio interval; Indicates the weight of the corresponding interval; This represents the weighted average accuracy. This represents the penalty coefficient for the dispersion of interval performance; This represents the hierarchical robustness accuracy of the signal-to-noise ratio for candidate configurations. It can be achieved using... Equal weighting can also increase the weight of low SNR intervals. to Each parameter is controlled separately for robustness accuracy, maximum floating-point operation volume, parameter quantity, and latency evaluation items. This represents the value of the comprehensive evaluation function.

[0136] Step 4: Generate single-module atomic modifications and enforce wireless signal compatibility constraints.

[0137] Planning agent receives task status The current node's accuracy, resource consumption, and historical edge rewards for each signal-to-noise ratio (SNR) interval are considered. If the accuracy in a low SNR interval is lower than that in other SNR intervals or does not contribute sufficiently to improving the robustness of SNR hierarchical performance, a robustness improvement target is prioritized for the preprocessing, characterization, or enhancement agent. If a resource constraint violation occurs, a structure compression target is prioritized for the network architecture agent. The selected specialized agent outputs an atomic action. The configuration of the other three stages remains unchanged, making .in Indicates from parent configuration To sub-configuration The number of pipeline stages whose configurations change between each stage; a value of 1 indicates that only one stage is changed at a time.

[0138]

[0139] Read the intermediate state of the parent node And atomic action a. If a random frequency domain mask is used, its target must be a spectrum or time-frequency representation; if the input representation is changed from a dual-channel IQ signal to a short-time Fourier time-frequency graph, the network input layer must simultaneously meet the two-dimensional time-frequency tensor size requirement; if the enhancement operation is frequency offset injection, phase perturbation, sampling rate deviation, or IQ imbalance, its parameter range is determined based on the corresponding statistics in the task state, and the label preservation conditions before and after enhancement are verified. If the compatibility judgment fails, model training is not started, but the failure type is recorded for the corresponding agent to avoid in the future.

[0140]

[0141] In the formula, each sub-decision term is either 0 or 1, representing whether the operation domain, tensor shape, processing order, label preservation, and deployment resource requirements are met, respectively; when any sub-decision term is 0, .

[0142]

[0143] In this embodiment, Indicates the task status Lower parent configuration Execute action The compatibility judgment value; 1 indicates that the extension is allowed, and 0 indicates that the extension is rejected; expand(c,a) means that the extension is configured. and actions Establish and evaluate the expansion operations for the corresponding tree edges; overall compatibility determination. It is obtained by multiplying five sub-determinants: operation domain, tensor shape, processing order, label preservation, and device constraints. If any sub-determinant is 0, expand(c,a) will not be executed.

[0144] Step 5: Perform a single-factor counterfactual paired assessment.

[0145] For sub-configurations that pass the compatibility check Model building and training are performed. The parent and child configurations use the same training set, validation set, batch size, optimizer, learning rate, number of training epochs, and random seed. The accuracy, FLOPs, parameter count, and latency of the parent and child configurations are calculated for each signal-to-noise ratio interval, yielding the following results:

[0146]

[0147] In the formula, Indicates action Relative to parent configuration The resulting pairing revenue; This indicates the sub-configuration formed after the action is executed; two items. and These represent the combined evaluation values ​​of the child configuration and the parent configuration, respectively.

[0148] Simultaneously save the interval-specific return vectors This is used to distinguish between average performance improvement and low signal-to-noise ratio robustness improvement.

[0149] In the formula, ΔA(c,a) represents the action. The resulting vector of accuracy change across the signal-to-noise ratio interval; to They represent the first to the second, respectively. The accuracy across signal-to-noise ratio intervals. The parent and child configurations change only one stage and use the same data partitioning, random seed, and training budget.

[0150] Step 6: Build a task-conditional experience base and update the experience reliability.

[0151] The reflexive agent encodes task states, atomic actions, compatibility results, interval-based rewards, and resource changes as experience entries. It uses the distance between normalized task states to determine context clusters q; and only counts the number of successful actions within the same or nearby context clusters. Update the minimum effective return threshold separately. and And calculate the Beta posterior at quantiles The lower confidence boundary .

[0152]

[0153]

[0154]

[0155] In this embodiment, Represents the task context cluster; summation index This represents the historical evaluation samples within the cluster; and Let these represent the success prior and failure prior of the Beta distribution, respectively. and Each represents an action In context cluster The posterior success parameter and posterior failure parameter are in the formula; I[·] represents the indicator function, which takes the value 1 when the condition is true and 0 otherwise; This represents the minimum effective return threshold; This represents the quantile probability used to calculate the lower confidence boundary; The quantile function represents the posterior distribution of Beta; This represents the lower confidence bound of the obtained empirical reliability. It can be set. , And determine based on the fluctuation of repeated experiments in paired evaluation Beta posterior mean It only reflects the success rate, while It also reflects the number of historical samples.

[0156] Step 7: Perform tree edge selection based on state constraints and conditional experience.

[0157] for Candidate tree edges, maintain average pairing payoff Parent node access count Number of visits to the tree and estimated assessment costs And calculate:

[0158]

[0159] In this embodiment, Score(c,a|z) represents the task state. Select parent configuration below and atomic actions Tree-side rating; This represents the historical average pairing return of the tree edge; Indicates the number of times the parent node has been visited; The number of times the corresponding tree edge is visited is represented; ln represents the natural logarithm. Indicates the exploration coefficient; Indicates the empirical reliability weight; This indicates the estimated training and evaluation costs; This represents the penalty coefficient for evaluating costs. To explore the coefficient, This is the cost penalty coefficient. When satisfying... The tree edge with the highest score is selected for expansion based on the resource constraints of the deployment. Thus, the selection target of the Monte Carlo tree search changes from a general parent node to a combination of atomic actions constrained by the wireless signal state.

[0160] Step 8: Return the pairing rewards and update the multi-agent cooperative strategy.

[0161] After completing the matching evaluation in step five, Return along the current access path and update the corresponding tree edges. , , and in step six , The reflective agent generates structured feedback based on changes in revenue, reasons for compatibility failures, and resource changes across each SNR interval; the planning agent adjusts the next priority modification phase accordingly, and each dedicated agent only receives valid experience and failure constraints relevant to this phase.

[0162] Step 9: Iteratively search and output the target pipeline.

[0163] Repeat steps four through eight until the preset number of search iterations is reached, or the comprehensive evaluation function stabilizes over several consecutive rounds. In this embodiment, the search budget can be set to 50 valid candidate evaluations; actions that fail compatibility checks and have not yet started training are not counted as valid evaluations. Finally, nodes that meet the deployment constraints are selected. The highest configuration is used, and the accuracy, average accuracy, lowest interval accuracy, number of parameters, FLOPs, and inference latency are reported for each signal-to-noise ratio interval on the test set.

[0164] In this embodiment, This represents the value of the comprehensive evaluation function. Indicates the signal-to-noise ratio hierarchical robustness accuracy. This indicates the dataset used only for the final test. FLOPs, Params, and Latency represent the final configuration's floating-point operations, parameter count, and inference latency, respectively.

[0165] To verify the effects of each distinguishing technical feature, random search, Monte Carlo tree search using standard upper confidence bounds (UCB), tree search with compatibility criteria, tree search with single-factor counterfactual evaluation, tree search with task-conditionalized experience, and the complete method of this invention were set up as controls. Under the same effective evaluation budget, the number of training iterations required to reach the target evaluation value, the proportion of unexecutable configurations, the accuracy in low signal-to-noise ratio intervals, the performance variance in different signal-to-noise ratio intervals, and the final deployment cost were statistically analyzed.

[0166] Furthermore, if the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present 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 the present 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.

[0167] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0168] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), fiber optic devices, and portable optical disc read-only memory (CDROM). Additionally, computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0169] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of carrying out the invention as currently considered, or those features that are not relevant to implementing the invention) may be omitted.

[0170] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those skilled in the art who benefit from this disclosure, the development effort will be a routine work of design, manufacturing, and production without requiring much experimentation.

[0171] The above embodiments are only used to illustrate the technical solutions of the present invention; those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions described in the foregoing embodiments without causing the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent cooperative configuration method for wireless signal recognition, characterized in that, include: Extract the wireless signal task state and construct a hierarchical configuration search tree. The search nodes in the hierarchical configuration search tree store the complete configuration as well as the intermediate states of the preprocessing, characterization, enhancement and network architecture stages. The planning agent receives wireless signal task status, accuracy of each signal-to-noise ratio interval of the current node, resource consumption, and historical edge gains. The selected dedicated agent outputs atomic actions, which change the configuration of one stage. Read the intermediate state and atomic actions of the parent node, and perform a total compatibility determination based on the operation domain, tensor shape, processing order, label retention and device constraints. If the total compatibility determination passes, perform configuration extension to generate sub-configurations. By performing single-factor counterfactual pairing evaluation on the sub-configuration based on compatibility determination, the parent and sub-configurations use the same training set, validation set, batch size, optimizer, learning rate, number of training rounds, and random seed to obtain pairing benefits, interval benefit vectors, and resource changes. The task status, atomic actions, compatibility results, interval benefits, and resource changes are encoded as experience items. A task-conditional experience base is constructed and the experience reliability is updated. The target pipeline is obtained by iterative search based on the updated experience reliability.

2. The method according to claim 1, characterized in that, The process by which the planning agent determines the phase to be optimized and the selected specialized agent outputs atomic actions includes: The plan includes the task status of the intelligent agent receiving wireless signals, the accuracy of the current node in each signal-to-noise ratio range, resource consumption, and historical edge rewards. When the accuracy in the low signal-to-noise ratio (SNR) interval is lower than that in other SNR intervals, or when the contribution to the robust performance improvement of SNR stratification meets a preset threshold, the planning agent prioritizes sending robust improvement targets to the preprocessing agent, the characterization agent, or the enhancement agent. When resource constraints are violated, the planning agent prioritizes sending the structure compression target to the network architecture agent; The selected specialized intelligent agent outputs atomic actions based on the robustness improvement objective or the structural compression objective.

3. The method according to claim 2, characterized in that, The process of an atomic action changing a stage configuration includes: Atomic actions include stage number, operation type, and operation parameters; Determine one of the stages—preprocessing, characterization, enhancement, and network architecture—based on the stage number, and change the configuration of the corresponding stage according to the operation type and operation parameters. The phase configuration corresponding to the atomic action changes, while the other three phase configurations remain unchanged, so that there is only one phase configuration difference between the parent and child configurations, i.e., the number of phase differences is 1.

4. The method according to claim 3, characterized in that, The process of performing a comprehensive compatibility assessment includes: Read the intermediate state and atomic actions of the parent node, and obtain the judgment results of five sub-decisions: operation domain, tensor shape, processing order, label preservation, and device constraints. The results of the five sub-decisions are either 0 or 1. The results of the five sub-decisions are multiplied together to obtain the overall compatibility determination result. If the result of any sub-decision is 0, the result of the overall compatibility determination is 0, and the configuration extension is not executed. When the results of the five sub-decisions are all 1, the result of the overall compatibility determination is 1, and the configuration extension is executed.

5. The method according to claim 4, characterized in that, The process for determining the operational domain, tensor shape, and label preservation includes: When using a random frequency domain mask, the random frequency domain mask is applied to the spectrum or time-frequency representation; When the input representation is changed from IQ dual-channel to short-time Fourier time-frequency diagram, the network input layer synchronously meets the two-dimensional time-frequency tensor size requirement; When the enhancement operation involves frequency offset injection, phase perturbation, sampling rate deviation, or IQ imbalance, the parameter range of the enhancement operation is determined based on the corresponding statistics in the wireless signal mission state, and the tag retention conditions before and after enhancement are verified. If the compatibility check fails, model training will not be started, and the failure type will be recorded.

6. The method according to claim 5, characterized in that, The process of performing a one-factor counterfactual paired evaluation on sub-configurations that pass the compatibility determination includes: Model building and training are performed on the sub-configurations that pass the compatibility test; The parent and child configurations use the same training set, validation set, batch size, optimizer, learning rate, number of training epochs, and random seed. Calculate the accuracy, FLOPs, number of parameters, and latency of the parent and child configurations in each signal-to-noise ratio range. The paired revenue is obtained based on the comprehensive evaluation function value of the parent and child configurations, the interval revenue vector is saved, and the resource changes are obtained based on FLOPs, parameter quantity, and latency.

7. The method according to claim 6, characterized in that, The process of obtaining the paired payoff and interval payoff vectors includes: Subtract the comprehensive evaluation function value of the parent configuration from the comprehensive evaluation function value of the child configuration to obtain the paired benefit caused by the atomic action; Subtract the accuracy of the parent configuration in the first SNR interval from the accuracy of the sub-configuration in the first SNR interval, and so on, until the accuracy of the sub-configuration in the B SNR interval is subtracted from the accuracy of the parent configuration in the B SNR interval. Then, form a sub-interval profit vector in the order of the first to the B SNR intervals.

8. The method according to claim 7, characterized in that, The process of building a task-conditional experience base includes: Encode task status, atomic actions, compatibility results, interval benefits, and resource changes into experience items; Context clusters are determined using the distance between standardized task states; The number of successful actions within the same context cluster is counted, and the empirical reliability is updated based on historical samples within the same context cluster.

9. The method according to claim 8, characterized in that, The process of updating empirical reliability based on historical samples within the same context cluster includes: Using the minimum effective return threshold as a comparison benchmark, an indicator function is used to determine whether the pairing return of each historical sample is greater than the minimum effective return threshold. The results of the indicator function for pairing returns greater than the minimum effective return threshold are accumulated into the prior parameters of success. The indicator function is used to determine whether the pairing payoff of each historical sample is less than or equal to the minimum effective payoff threshold. The results of the indicator function when the pairing payoff is less than or equal to the minimum effective payoff threshold are added to the prior parameters of the failure. The Beta posterior is determined based on the successful and unsuccessful posterior parameters, and the empirical lower confidence bounds of the Beta posterior at quantiles are calculated.

10. The method according to claim 1, characterized in that, The method for obtaining the target pipeline through iterative search based on the updated empirical reliability is as follows: the compatible tree edges are scored and selected according to the updated empirical reliability, historical average pairing revenue and expected evaluation cost, and the pairing revenue is back-propagated along the search path to update the search state, and the target pipeline is obtained through iterative search.

Images