A channel quality evaluation method and system based on multi-modal and quantum state tomography

Abstract

The application relates to the technical field of communication, and specifically discloses a channel quality evaluation method and system based on multi-modal and quantum state tomography, which comprises the following steps: dynamically optimizing a sampling interval based on deep learning and reinforcement learning, combining quantum frequency synthesis and compressed sensing to realize high-precision adaptive sampling, and guaranteeing data credibility through a memristor and a hash chain; constructing collected data into a three-dimensional tensor, performing Tucker parallel decomposition, normalization flow, VAE and GAN joint optimization, generating an average spectrum curve and an adaptive noise threshold; fusing time-frequency features through a double attention mechanism and extracting fine structures in combination with quantum state tomography; evaluating channel quality based on multi-modal deep neural network fusion of time-frequency features and spatial domain data; and dynamically optimizing a resource allocation strategy by using reinforcement learning. The application realizes a complete closed loop from adaptive sampling, intelligent processing, fine feature extraction to evaluation and decision-making, and effectively improves the precision and intelligent level of channel monitoring in a complex electromagnetic environment.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a channel quality assessment method and system based on multimodal and quantum state tomography. Background Technology

[0002] In existing technologies, channel quality monitoring and evaluation of wireless communication systems typically rely on fixed sampling frequencies or uniform sampling strategies, combined with traditional Fourier analysis, moving averages, energy detection, and other methods for spectrum measurement. However, these methods cannot dynamically optimize sampling points based on real-time electromagnetic environment changes, limiting both sampling efficiency and accuracy. Furthermore, existing frequency synthesis techniques mostly employ PLLs (Phase-Locked Loops) or DDSs (Direct Digital Synthesizers), which lack sufficient accuracy and stability, and have not yet incorporated quantum frequency synthesis to improve sampling quality. While compressed sensing can be used for sparse signal sampling, its data lacks reliable and tamper-proof security mechanisms during storage and transmission.

[0003] Furthermore, current spectral data processing methods primarily rely on linear filtering, windowing, and simple statistics, which are insufficient for handling high-dimensional and complex spectra. They also lack stability under noise perturbations and fail to incorporate deep learning models such as tensor computation, normalized flow, VAEs (Variational Autoencoders), or GANs (Generative Adversarial Networks) for high-precision spectral modeling. Existing background noise thresholds are mostly fixed or empirically set, lacking adaptability and leading to high misclassification rates. Traditional channel feature extraction typically focuses only on the time or frequency domains, failing to achieve multi-domain fusion and lacking the integration of high-precision methods such as quantum state tomography, resulting in incomplete features.

[0004] In terms of channel quality assessment, existing technologies mostly use single-modal deep learning models such as CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory), which are difficult to handle multi-source data such as time domain, frequency domain, and spatial domain, resulting in limited assessment accuracy. Resource allocation strategies mostly rely on static rules or simple heuristic algorithms, which cannot achieve intelligent and real-time optimization.

[0005] In summary, existing technologies have significant shortcomings in terms of sampling planning, sampling accuracy, data security, intelligent spectrum processing, adaptive noise threshold, multimodal feature extraction, accuracy of channel quality assessment, and intelligent resource allocation, making it difficult to meet the needs for high-precision, reliable, and intelligent channel quality monitoring in complex electromagnetic environments. Summary of the Invention

[0006] This invention provides a channel quality assessment method and system based on multimodal and quantum state tomography. The technical problem it solves is how to achieve high-precision sampling, stable spectrum modeling, fine feature extraction, and intelligent assessment decision-making in complex electromagnetic environments.

[0007] To address the above technical problems, this invention provides a channel quality assessment method based on multimodal and quantum state tomography, comprising the following steps:

[0008] S1. Generate a sampling frequency based on deep learning, and collect and store channel spectrum data within a preset time based on the generated sampling frequency.

[0009] S2. Process the stored spectrum data to generate an average spectrum curve and an adaptive background noise threshold curve;

[0010] S3. Based on the average spectral curve, time-frequency features are fused by dual attention and refined time-frequency feature vectors are extracted using quantum state tomography.

[0011] S4. Utilize the extracted time-frequency fine feature vectors to evaluate channel quality through a multimodal deep neural network, and combine reinforcement learning to dynamically optimize resource allocation strategies.

[0012] S5, output channel quality assessment results, adaptive background noise threshold curve and resource allocation strategy.

[0013] Furthermore, step S3 specifically includes the following steps:

[0014] S31. Extract the time-frequency matrix from the average spectrum data;

[0015] S32. Apply the self-attention mechanism to the time dimension and frequency dimension of the time-frequency matrix respectively to obtain the weighted time-domain features and frequency-domain features, and then fuse them to obtain the fused features.

[0016] S33. Encode the fused features into quantum states, and reconstruct the density matrix using quantum measurement and tomography algorithms to extract quantum features;

[0017] S34. Combine the fused features with the quantum features to obtain the final time-frequency fine feature vector.

[0018] Furthermore, step S33 specifically includes:

[0019] Fusion features Normalized to unit length, resulting in a normalized vector. ;

[0020] Amplitude coding is used for normalized vectors Constructing quantum pure states and its density matrix :

[0021] Select a set of observables For quantum pure state Perform projection measurements to obtain the desired value. ;

[0022] The density matrix was reconstructed from multiple measurement results using a quantum state tomography algorithm. And extract its features or eigenvalues ​​to form a quantum eigenvector. .

[0023] Furthermore, step S4 specifically includes the following steps:

[0024] S41. Construct a multimodal deep neural network model, input the time-frequency fine feature vector and spatial domain data into the multimodal deep neural network model, and output the channel quality estimate after fusion.

[0025] S42. Construct a state space from the channel quality estimate and the system resource state, define the action space and reward function, and learn the optimal resource allocation strategy through reinforcement learning.

[0026] S43. Adjust system resources based on the actions made by reinforcement learning decisions and provide feedback on the new state.

[0027] Furthermore, step S41 specifically includes:

[0028] Time-frequency fine feature vector Decomposed into time domain part and frequency domain part ;

[0029] Multimodal deep neural network models for the time domain Frequency domain part and spatial domain data Extract high-level features separately;

[0030] The high-level features extracted from the three modalities are concatenated to obtain multimodal fusion features;

[0031] The multimodal fusion features are processed through a fully connected layer and an activation function to output a channel quality estimate.

[0032] Furthermore, step S1 specifically includes:

[0033] S11. When the system starts, it uses historical channel data and deep learning models such as long short-term memory networks or gated recurrent units to predict the initial set of sampling parameters, including the starting frequency, preset frequency interval, sampling frequency and preset time.

[0034] S12. Construct the adjustment strategy of the preset frequency interval into a Markov decision process: Define the state space and action space, the agent selects an action according to the current state, and evaluates the effect of the action through the reward function to obtain the optimal preset frequency interval.

[0035] S13. Based on the obtained optimal frequency interval and the predicted starting frequency, determine each sampling frequency point through quantum frequency synthesis;

[0036] S14. At each sampling frequency point, compressed sensing technology is used to compress and collect data at the predicted sampling frequency and preset time to obtain compressed data.

[0037] S15. The collected compressed data is physically stored using a memristor array.

[0038] Furthermore, in step S12, the reward function Designed for , For the agent in state Execute action The reward value obtained immediately afterwards Indicates the current state Spectrum utilization rate Indicates the current state The spectrum sampling error is as follows. They are respectively and Weighting coefficients;

[0039] In step S13, when the optimal preset frequency interval is determined... Then, the corresponding physical sampling point frequency , For the first Each sampling frequency point , The number of sampling frequency points, The starting frequency;

[0040] In step S14, the compressed data is represented as , The original signal, For the measurement matrix, It is a sparse base.

[0041] Furthermore, step S2 specifically includes the following steps:

[0042] S21. Organize the collected spectrum data into a three-dimensional tensor including three dimensions: time dimension length, number of frequency points, and number of measurements. Use Tucker decomposition to perform parallel computation on the tensor to obtain the average spectrum data.

[0043] S22. Input the average spectrum data into the normalized flow model, and map it to the standard normal distribution through an invertible transformation to obtain the normalized latent variables.

[0044] S23. Input the normalized latent variables into the variational autoencoder for dimensionality reduction and feature extraction. Input the extracted key latent space features into the generative adversarial network. The generative adversarial network generates the average spectral curve.

[0045] S24. Construct a deep reinforcement learning model, using the real-time channel noise parameters in the average spectrum curve as the state, the adjustment amount of the background noise threshold as the action, and design a reward function with the goal of reducing the false positive rate. The agent learns the optimal strategy by interacting with the environment and finally outputs a dynamic adaptive noise threshold curve.

[0046] Furthermore, in step S23, the encoder of the variational autoencoder normalizes the latent variables. Mapped to a lower-dimensional latent space Decoder reconstruction The loss function is designed as a weighted sum of the reconstruction loss and the KL divergence.

[0047] This invention also provides a channel quality assessment system based on multimodal and quantum state tomography, the key feature of which is that it includes:

[0048] The adaptive sampling and trusted storage module, the parallel computing and intelligent curve generation module, the dual attention fusion and quantum state tomography module, the channel quality assessment and reinforcement learning resource allocation module, and the result output module are respectively used to implement steps S1 to S5 in the channel quality assessment method based on multimodal and quantum state tomography.

[0049] The present invention provides a channel quality assessment method and system based on multimodal and quantum state tomography, the advantages of which are as follows:

[0050] By using deep learning prediction and reinforcement learning to dynamically optimize the sampling interval, combined with quantum frequency synthesis, the adaptive adjustment of the sampling strategy and high-precision frequency generation are achieved, solving the problem of difficulty in balancing sampling efficiency and accuracy in complex electromagnetic environments.

[0051] Sparse data acquisition is achieved by using RIP-based compressed sensing technology, which significantly reduces the amount of data and hardware pressure. Combined with memristor physical storage and hash chain verification, a dual security defense of "physical + logic" is built to ensure data trustworthiness.

[0052] By constructing the spectral data into a three-dimensional tensor and performing Tucker decomposition and distributed parallel computing, the speed and robustness of high-dimensional spectrum processing are significantly improved. Then, through normalized flow, VAE and GAN joint optimization, an average spectral curve with both stability and high fidelity is generated, avoiding the over-smoothing or noise residue of traditional methods.

[0053] By using deep reinforcement learning to dynamically optimize the background noise threshold, the probability of false positives and false negatives is effectively reduced, and the detection accuracy is improved.

[0054] By deeply fusing the time and frequency dimension features of the time-frequency matrix through a dual attention mechanism, and introducing quantum state tomography to extract fine structures that are difficult to capture by classical methods, the problems of incomplete feature extraction and insufficient resolution are solved.

[0055] Constructing a multimodal deep neural network to fuse fine time-frequency features and spatial domain data significantly improves the accuracy and robustness of channel quality assessment;

[0056] By combining the evaluation results with the resource status, and using reinforcement learning to dynamically optimize the allocation strategies for power, time slots, and frequency bands, a closed-loop adaptive optimization of resource allocation is achieved.

[0057] The final output includes channel quality, noise threshold, and resource decisions, guiding the communication system to perform adaptive modulation and coding, spectrum switching, and other operations, forming a complete closed loop from acquisition to decision-making. This fully realizes high-precision, intelligent, and reliable channel monitoring and decision-making in complex electromagnetic environments. Attached Figure Description

[0058] Figure 1 This is a flowchart of the channel quality monitoring and evaluation method provided in the embodiments of the present invention;

[0059] Figure 2 This is a flowchart of the sampling point frequency setting based on reinforcement learning provided in an embodiment of the present invention;

[0060] Figure 3 This is a flowchart of generating an average spectrum curve provided in an embodiment of the present invention;

[0061] Figure 4 This is a flowchart of step S3 provided in an embodiment of the present invention. Detailed Implementation

[0062] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. The embodiments are given for illustrative purposes only and should not be construed as limiting the present invention. The accompanying drawings are for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0063] This invention provides a channel quality assessment method based on multimodal and quantum state tomography, such as... Figure 1 The flowchart shown includes the following steps:

[0064] S1: Adaptive sampling and reliable storage of channel spectrum data: Based on deep learning, a sampling frequency is generated, and channel spectrum data is collected and stored within a preset time based on the generated sampling frequency.

[0065] S2: Parallel computation, normalization and intelligent curve generation of channel spectrum data: Process the spectrum data stored in step S1 to generate a high-quality average spectrum curve and an adaptive background noise threshold curve.

[0066] S3: Dual attention fusion and quantum state tomography analysis of time-frequency features: Based on the average spectral curve, time-frequency features are fused by dual attention and fine time-frequency feature vectors are extracted using quantum state tomography;

[0067] S4: Channel quality assessment and reinforcement learning-driven resource allocation decision: Utilizing the extracted time-frequency fine feature vectors, channel quality is assessed through a multimodal deep neural network, and reinforcement learning is combined to dynamically optimize resource allocation strategies;

[0068] S5: Output of detection results: Outputs channel quality assessment results, adaptive background noise threshold curves, and resource allocation strategies to guide the adaptive modulation and coding and spectrum switching operations of the communication system.

[0069] Through steps S1 to S5 above, this invention achieves a complete closed loop from spectrum data acquisition, processing, feature extraction to quality assessment and resource allocation, effectively adapting to complex electromagnetic environments and providing high-precision, intelligent channel monitoring and decision-making capabilities. The following provides a more detailed explanation of each step.

[0070] (1) Step S1: Sampling and storage of channel spectrum data

[0071] This step aims to address the problem that traditional fixed sampling strategies cannot adapt to dynamic electromagnetic environments. It predicts initial sampling parameters through deep learning, dynamically optimizes the sampling interval using reinforcement learning, achieves high-precision frequency generation through quantum frequency synthesis, and finally realizes efficient data acquisition and reliable storage through compressed sensing and memristor blockchain technology. This step specifically includes the following sub-steps:

[0072] S11. When the system starts, it uses historical channel data and deep learning models such as long short-term memory networks or gated recurrent units to predict the initial set of sampling parameters, including the starting frequency, preset frequency interval, sampling frequency and preset time.

[0073] S12. Construct the preset frequency interval adjustment strategy as a Markov decision process: Define the state space (containing real-time channel parameters such as current signal strength and noise power spectral density) and the action space (i.e., the set of frequency interval adjustment amounts). The agent selects an action based on the current state and evaluates the effect of the action through the reward function to obtain the optimal preset frequency interval.

[0074] S13. Based on the obtained optimal frequency interval and the predicted starting frequency, determine each sampling frequency point through quantum frequency synthesis;

[0075] S14. At each sampling frequency point, compressed sensing technology is used to compress and collect data at the predicted sampling frequency and preset time to obtain compressed data.

[0076] S15. The collected compressed data is physically stored through a memristor array, utilizing its non-volatile and high-density characteristics to achieve fast read and write.

[0077] Through the intelligent decision-making and high-precision execution of the above steps, step S1 achieves a complete closed loop from perceiving the environment to making decisions, physical execution, and security recording: the initial prediction parameters provide a benchmark for reinforcement learning, accelerate the dynamic optimization process, and ensure that the sampling framework is consistent with historical patterns; the dynamic optimization of reinforcement learning enables the sampling interval to adapt to environmental changes and improve sampling efficiency; quantum frequency synthesis ensures high precision of the sampling point frequency; compressed sensing reduces the amount of data and alleviates hardware pressure; memristors and hash chains ensure the reliability and immutability of the data.

[0078] In the prediction parameters of step S11, the initial frequency This refers to the lowest frequency point on the frequency axis where the spectrum is sampled, i.e., the left boundary of the frequency band to be monitored. In actual spectrum monitoring, it is usually necessary to focus on a continuous frequency band (e.g., 2.4 GHz to 2.5 GHz), with the starting frequency... This refers to the starting point of the frequency band. Since the channel environment (such as the location of interference sources and available frequency bands) may change over time, the system needs to predict the starting point of the most worthwhile frequency band to monitor based on historical data, avoiding wasting sampling resources in irrelevant frequency bands. Starting frequency Interval with preset frequency Together they determine the location of all subsequent sampling frequency points:

[0079] (1)

[0080] in, For the first Each sampling frequency point , The number of sampling frequency points (determined by the monitoring range or system bandwidth).

[0081] Preset frequency interval The frequency step size is the sampling interval between two adjacent sampling frequencies, controlling the density of frequency domain sampling. A smaller interval results in higher frequency resolution but increases the number of sampling points, data volume, and computational burden; a larger interval reduces the number of sampling points, improving efficiency but potentially losing spectral details. The optimal frequency interval varies depending on the channel conditions. For example, a larger interval can be used in interference-sparse frequency bands to improve efficiency, while a smaller interval is needed in interference-dense or rapidly changing signal bands to capture details. Deep learning models predict an initial reasonable interval based on historical data, which the reinforcement learning agent can then dynamically adjust for adaptive optimization. Preset frequency interval. and starting frequency They jointly define the sampling frequency grid; at the same time, the frequency interval The choice of this parameter will affect the design of subsequent compressed sensing measurement matrices (e.g., the construction of sparse bases) and spectral efficiency. The calculation.

[0082] Sampling frequency The sampling frequency refers to the rate at which a signal is converted from analog to digital in the time domain, i.e., the number of samples collected per second (in Hz). It determines the time resolution and directly affects the signal's bandwidth and anti-aliasing capability. According to the Nyquist sampling theorem, to recover a signal without distortion, the sampling frequency... It should be at least twice the highest frequency of the signal. However, in this embodiment, due to the introduction of compressed sensing technology, the actual sampling frequency is... It can be lower than the Nyquist rate. While compressed sensing allows undersampling, the sampling frequency... Certain conditions still need to be met (such as the RIP property of the measurement matrix). Predicting a suitable sampling frequency is crucial. This ensures reliable spectrum reconstruction while reducing data volume. Furthermore, the sampling frequency... This also affects the number of sample points within the time window, thus impacting the frequency resolution of the spectral analysis. Sampling frequency With preset time Together they determine the number of time samples at each sampling frequency point. These samples are used for subsequent spectral calculations (such as FFT) or tensor construction.

[0083] preset time This refers to the duration of each spectral sample, i.e., the window length (in seconds) for time-domain sampling of each frequency point, which determines the frequency resolution. (If Fourier transform is used). Longer... It can improve frequency domain resolution, but it will reduce time response speed; shorter... Conversely, in dynamic channel environments, a trade-off between time-frequency resolution and frequency resolution is necessary. Depending on the rate of channel change (e.g., the duration of interference signals), the system needs to predict an appropriate time window length to ensure the accuracy and real-time performance of spectrum measurements. In tensor construction, a preset time... The size of the corresponding time dimension affects subsequent tensor decomposition and feature extraction.

[0084] During the initialization phase, the system uses a deep learning model to predict the initial values ​​of these four types of parameters, adapting them to historical channel characteristics. During operation, the reinforcement learning agent primarily adjusts the frequency spacing... The system dynamically adjusts to adapt to real-time environmental changes, and other parameters may also be fine-tuned based on feedback. This "prediction + adaptation" mechanism enables the system to achieve high-precision and high-efficiency spectrum monitoring in complex electromagnetic environments. This set of initial parameters provides a starting point for subsequent reinforcement learning (such as the initial interval of the action space). ) and fixed parameters (such as , , This reduces the exploration space, accelerates system convergence, and ensures that the sampling framework matches historical patterns.

[0085] In step S12, as Figure 2 As shown, the sampling point frequency setting process based on reinforcement learning includes: defining the state space, action space, and reward function; the agent learns the optimal policy through interaction with the environment. Specifically, the state space... (Current moment) The state (including channel status) contains real-time parameters such as signal strength, noise power spectral density, and the location and intensity of interference sources. Action space (Current moment) The agent's actions are set at preset frequency intervals. ( A set of adjustment strategies (for a preset number of frequency intervals). The RL agent adjusts its strategies based on the state. With reward function Learning the optimal preset frequency interval Choose a strategy.

[0086] reward function To maximize spectrum utilization As the target, while considering sampling error The expression is:

[0087] (2)

[0088] in, For the agent in state Execute action The reward value obtained immediately after the action reflects the immediate effect of the current action. Indicates the current state Spectrum utilization rate Indicates the current state The spectrum sampling error is as follows. They are respectively and The weighting coefficients.

[0089] The agent iteratively learns the optimal preset frequency interval through the following Bellman equation. Configuration strategy:

[0090] (3)

[0091] in, Indicates the current moment The agent is in a state Time, Execution of Action The cumulative expected reward that can be obtained This is a discount factor (its value is usually between [0,1], used to measure the importance of future rewards relative to current rewards). Indicates the next state All possible actions that a medium-sized intelligent agent can choose. The maximum Q value that can be obtained (representing the best cumulative reward that can be obtained from the next moment by following the optimal strategy).

[0092] In step S13, quantum frequency synthesis technology, compared to traditional phase-locked loops or direct digital frequency synthesizers, has higher frequency accuracy, providing accurate sampling positions for subsequent compressed sensing reconstruction and avoiding errors introduced by frequency drift. Once the optimal preset frequency interval is determined... The corresponding physical sampling point frequencies are shown in equation (1). Frequency error Satisfying Relationship:

[0093] (4)

[0094] in, For the coherence time of a quantum bit, These are constants related to specific physical implementations. The longer, The smaller. (Through) This can reduce errors and ensure that the actual sampling frequency is highly consistent with the target frequency.

[0095] In step S14, the compressed data can be represented as:

[0096] (5)

[0097] in, The original signal, For the measurement matrix, It is a sparse base. For sparse representation. To ensure signal reconstructability, the measurement matrix must satisfy the RIP (Restricted Isometry) condition. Through careful design It can acquire signals at a sampling rate far lower than the Nyquist sampling rate, significantly reducing the amount of data.

[0098] preset time and sampling frequency Together they determine the number of original samples at each frequency point. The measurement matrix of compressed sensing number of rows Less than The specific value is determined by the sparsity of the signal and the RIP condition, but the preset time... and sampling frequency This indirectly affected the reconstruction accuracy.

[0099] Compressed sensing reduces the sampling rate and data volume, alleviating the pressure on hardware storage and transmission. At the same time, it utilizes the sparsity of the signal to ensure reconstruction quality, enabling the system to achieve efficient monitoring over a wide frequency range.

[0100] In step S15, the spectrum data is finally physically stored using a memristor, and immutability verification is achieved using a hash chain:

[0101] (6)

[0102] This ensures data security. For the first One data block, Its hash value, For XOR operation, For hash operation, This is the hash value of the previous data block. By verifying the integrity of the hash chain, any data tampering can be detected.

[0103] The reason for designing step S1 in this way is as follows:

[0104] Through reinforcement learning, the density of sampling points can be automatically adjusted according to the real-time electromagnetic environment (increased sampling in areas with more interference and sparse sampling in idle frequency bands), which solves the problem of low efficiency of traditional fixed-interval sampling.

[0105] Compared to traditional phase-locked loops or direct digital frequency synthesizers, quantum frequency synthesis technology significantly reduces frequency offset, ensures the accuracy of sampling point positions, and provides a high-quality data source for subsequent data analysis.

[0106] By using compressed sensing, the sampling rate can be much lower than that required by the traditional Nyquist sampling theorem, which greatly reduces the sampling pressure of the analog-to-digital converter and the data transmission bandwidth requirements.

[0107] The physical storage characteristics of memristors, combined with the hash chain verification of blockchain, construct a dual security defense of "physical + logical". This not only prevents the physical loss of data at the storage medium level, but also prevents malicious tampering during network transmission or storage, thus meeting the credibility requirements when spectrum data is used as evidence or key analysis objects.

[0108] (2) Step S2: Parallel computation, normalization and intelligent curve generation of spectrum data

[0109] This step follows the high-precision spectrum data collected and stored in step S1. It aims to perform in-depth processing of the raw spectrum data using advanced mathematical and artificial intelligence methods to generate high-quality average spectrum curves and adaptive background noise threshold curves, providing stable and accurate input for the subsequent time-frequency feature extraction and channel quality assessment in step S3. Figure 3 As shown in the flowchart, this step specifically includes the following steps:

[0110] S21. Organize the spectrum data collected in step S1 into a three-dimensional tensor including three dimensions: time dimension length, number of frequency points, and number of measurements. Use Tucker decomposition to perform parallel calculations on the tensor to obtain the average spectrum data.

[0111] S22. Input the average spectrum data into the normalizing flow model, and map it to the standard normal distribution through an invertible transformation to obtain the normalized latent variables;

[0112] S23. Input the normalized latent variables into the variational autoencoder (VAE) for dimensionality reduction and feature extraction. Input the extracted key latent space features into the generative adversarial network (GAN). The GAN generates the average spectral curve.

[0113] S24. Construct a deep reinforcement learning model, with real-time channel noise parameters (such as mean, variance, historical threshold, etc.) in the average spectrum curve as the state, the adjustment amount of the background noise threshold as the action, and the reward function designed with the goal of reducing the misjudgment rate. The agent learns the optimal policy by interacting with the environment (i.e., processing spectrum data according to the current threshold and evaluating the misjudgment situation), and finally outputs a dynamic adaptive noise threshold curve.

[0114] Through a series of processes from S21 to S24, step S2 transforms the raw spectrum data collected in step S1 into a high-precision average spectrum curve and an adaptive background noise threshold curve, laying a solid data foundation for time-frequency feature extraction and channel quality assessment in step S3. The high-precision average spectrum curve is generated through tensor parallel computation, normalized flow, and joint optimization of VAE and GAN, resulting in a smooth curve that retains key latent space features. The adaptive background noise threshold curve is dynamically generated through deep reinforcement learning, allowing it to adapt to environmental changes in real time.

[0115] In step S21, the three-dimensional tensor can be represented as , Represents the length of time. Represents the number of frequency points. The number of measurements. Tucker decomposition is expressed by the formula:

[0116] (7)

[0117] in, For the core tensor, The factor matrix (represented by low-dimensional representations of time, frequency, and measurement dimensions, respectively), with subscripts... Corresponding to the first, second, and third dimensions. Core tensor It reflects the complex coupling relationship between the three dimensions, while the factor matrix The main patterns of change were captured in the dimensions of time, frequency, and number of measurements.

[0118] Parallel solution on a distributed computing cluster The average spectral data is then reconstructed from the factor matrix through inverse transformation.

[0119] Step S21 significantly improves the computational efficiency of the average spectrum curve through tensor decomposition and parallel computing, meeting the needs of real-time monitoring; and enhances the robustness of the spectrum data by using multi-dimensional information compression, suppressing the influence of random noise.

[0120] In step S22, normalized flow techniques are used to perform a series of reversible transformations. (m items) will input data Mapping to the standard normal distribution This step eliminates the problem of unstable distribution of the original spectral data, provides standardized input for stable training of VAE and GAN, and maintains data reversibility, which facilitates subsequent analysis.

[0121] In step S23, the VAE encoder will Mapped to a lower-dimensional latent space Decoder reconstruction The loss function of a VAE consists of two parts: reconstruction loss and KL divergence (weighted sum). Reconstruction loss (typically using mean squared error or cross-entropy) measures the encoder-decoder's ability to reconstruct the input data. This part forces the decoder to recover the original spectral data from the latent space variables as accurately as possible, ensuring that the latent space encoding contains the key information of the input. KL divergence measures the difference between the latent variable distribution of the encoder output and the prior distribution (standard normal distribution). By minimizing KL divergence, the latent space is regularized to a continuous, smooth distribution, allowing latent variables randomly sampled from the prior to decode reasonable spectral data, enhancing the generation capability. The joint optimization of both is essentially maximizing the lower bound of the data likelihood (ELBO), ensuring the accuracy of feature extraction while providing a structured latent space input for subsequent GAN generation.

[0122] The loss function of GANs is based on the adversarial training concept and consists of two mutually interacting parts (weighted sum): a generator and a discriminator. The discriminator's loss aims to accurately distinguish between real spectral data and fake data generated by the generator. The discriminator outputs a high probability for real samples and a low probability for fake samples, and its loss function typically uses cross-entropy to maximize the probability of correct classification. This forces the discriminator to continuously improve its discrimination ability. The generator's loss aims to deceive the discriminator, causing it to classify the generated fake samples as real data. The generator attempts to minimize the discriminator's probability of correctly recognizing its output, thereby learning the distribution characteristics of real data. Through this zero-sum game, the generator gradually approximates the distribution of real spectral data. The generated curves are not only smooth and continuous but also capture the subtle structure of the real spectrum, avoiding the oversmoothing or noise residue problems common in traditional statistical averaging methods.

[0123] The VAE first performs dimensionality reduction and feature extraction on the normalized streaming spectral data, providing the GAN with a compact and semantically rich latent space input. The GAN then further optimizes the curve details, ensuring that the final generated spectral curve combines the global structure preservation capability of the VAE with the local detail realism of the GAN. This joint loss function design ensures that the generated spectral curve satisfies both statistical stability and high fidelity, providing a reliable foundation for subsequent channel quality assessment.

[0124] In step S24, the reward function is designed as follows:

[0125] (8)

[0126] in, The false positive rate (the probability of misclassifying noise as a signal). The false negative rate (the probability of misclassifying a signal as noise). These are the weighting coefficients.

[0127] Step S24 outputs an adaptive noise threshold curve that varies with the environment. This threshold, along with the average spectrum curve, is used in the time-frequency fusion analysis of S3 as a benchmark for signal detection, improving the accuracy of channel quality assessment. Compared to a fixed threshold, the adaptive threshold dynamically responds to noise changes, significantly reducing the false positive and false negative rates. The deep reinforcement learning model continuously learns and optimizes the threshold, enabling the system to have long-term adaptability. An accurate noise threshold enhances the reliability of subsequent feature extraction and channel assessment.

[0128] (3) Step S3: Dual attention fusion of time-frequency features and quantum state tomography analysis

[0129] This step follows the average spectrum curve and adaptive background noise threshold curve generated in step S2. It aims to deeply fuse time-domain and frequency-domain information through a dual-attention mechanism and combine this with quantum state tomography to extract fine-grained channel structure features, providing a more comprehensive and high-resolution feature representation for subsequent channel quality assessment. Figure 4 The flowchart shown includes the following sub-steps in step S3:

[0130] S31. Extract the time-frequency matrix from the average spectrum data obtained in step S21. , The duration (based on a preset time) Decide), The number of frequency points (determined by the number of sampling frequency points), each row of the matrix This represents the spectral distribution at a given moment, with each column... Represents a time series at a specific frequency point;

[0131] S32. Apply the self-attention mechanism to the time and frequency dimensions of the time-frequency matrix respectively to obtain weighted time-domain features and frequency-domain features, and then fuse them (using a concatenation method) to obtain the fused features. ;

[0132] S33. Encode the fused features into quantum states, and reconstruct the density matrix using quantum measurement and tomography algorithms to extract quantum features. ;

[0133] S34, Integrating Features With quantum characteristics By merging, the final time-frequency refined feature vector is obtained. This serves as the input for channel quality assessment in step S4.

[0134] Through the above sub-steps, step S3 achieves the extraction of high-resolution time-frequency fine features from average spectrum data, providing a more comprehensive and accurate feature foundation for subsequent intelligent evaluation.

[0135] The time-frequency matrix in step S31 comprehensively reflects the distribution characteristics of the signal in the time-frequency domain, providing basic data for the subsequent attention mechanism.

[0136] In step S32, the time-frequency matrix is... The query matrix is ​​obtained through linear transformation. Key matrix Sum matrix ,in For learnable weight matrix, For the dimensions of the query and key vector, Let be the dimension of the value vector. Then the temporal attention output is:

[0137] (9)

[0138] in, This refers to the softmax function.

[0139] By transposing the time-frequency matrix, frequency-domain attention is obtained using the same process as calculating time-domain attention.

[0140] Time-frequency fusion is a time-frequency fusion network based on a dual attention mechanism. The time-domain attention mechanism calculates the time-domain signal... (Right now The weights are assigned based on the correlation between different time points. Let the time-domain signal... The length is T, and the temporal attention weights are... Through formula Calculation, where Frequency domain attention mechanisms weight different frequency components based on their energy distribution and importance. Let the frequency domain signal... (i.e., frequency domain attention output) The frequency range is F, and the frequency domain attention weights are... Through formula Calculation, where Then, the weighted time-domain and frequency-domain features are fused to obtain the fused features. .

[0141] The rationale for step S32 is that time-domain attention can capture the long-term dependencies of a signal in the time dimension (such as the duration of interference pulses), and frequency-domain attention can enhance key frequency components (such as the main signal frequency). The combination of the two can comprehensively highlight the important features in the time-frequency structure, providing a purer input for subsequent quantum state tomography.

[0142] In step S33, the fusion features are first... Normalized to unit length, resulting in a normalized vector. (D represents the total dimension of the fused features). Amplitude encoding is used to construct quantum pure states:

[0143] (10)

[0144] Its density matrix is Select a set of observables (For example, the tensor product of Pauli operators), by projecting measurements onto the quantum state, we obtain the expected value:

[0145] (11)

[0146] in, Indicates a trace.

[0147] Reconstructing the density matrix from multiple measurements using quantum state tomography algorithms (such as maximum likelihood estimation) And extract its features, such as diagonal elements (probability distribution) or eigenvalues, to construct a quantum feature vector. .

[0148] The rationale for step S33 is that quantum state tomography can reveal quantum coherence and entanglement information that are difficult to capture by classical methods. When applied to classical signals, it can reflect subtle features such as weak modulation and phase noise, thereby improving the resolution and accuracy of channel quality assessment.

[0149] In step S34, the final time-frequency fine feature vector is obtained. The time-frequency information, which includes dual attention enhancement, and the fine structure information extracted by quantum tomography, will serve as an important input to the channel quality assessment model in step S4.

[0150] (4) Step S4: Channel quality assessment and reinforcement learning-driven resource allocation decision

[0151] This step, based on the refined features extracted in step S3, utilizes a multimodal deep neural network (MDNN) to comprehensively evaluate channel quality and combines reinforcement learning to dynamically optimize network resource allocation, achieving adaptive decision-making. Specifically, it includes the following sub-steps:

[0152] S41. Construct a multimodal deep neural network (MDNN) model, and use the time-frequency fine feature vector obtained in step S34. The spatial domain data is input into the MDNN model, and the channel quality estimate is output after fusion.

[0153] S42. Construct a state space from the channel quality estimate and the system resource state, define the action space and reward function, and learn the optimal resource allocation strategy through reinforcement learning.

[0154] S43. Adjust system resources based on the actions made by reinforcement learning decisions and provide feedback on the new state.

[0155] In step S41, a multimodal deep neural network (MDNN) model is constructed to evaluate channel quality. The input includes the fused features obtained from step S34. It can be decomposed into a time domain part. and frequency domain part In addition, it also includes spatial domain data. (For example, antenna array phase, interference source location, signal angle of arrival, etc., where S is the spatial feature dimension, set to zero if none exists). Each feature extraction sub-network (e.g., fully connected layer or convolutional layer) extracts high-level features:

[0156] (12)

[0157] in, This represents the mapping function of each feature extraction sub-network (e.g., a two-layer fully connected network), with output feature dimensions of respectively. .

[0158] The outputs of the three sub-networks are concatenated to obtain the multimodal fusion feature:

[0159] (13)

[0160] Finally, the multimodal fusion features are processed through a fully connected layer and an activation function to output a channel quality estimate:

[0161] (14)

[0162] in, This is the weight matrix. For bias, C is the number of output categories (e.g., 1 for binary classification, and the number of categories for multi-class classification). Y is the activation function (such as the ReLU function). Y can be the channel quality level (excellent / good / poor) or a continuous value (such as the signal-to-noise ratio estimate).

[0163] The rationale for designing step S41 is that a single modal feature is difficult to fully reflect the channel state. By processing time-domain, frequency-domain, and spatial-domain features separately, MDNN can fuse multi-source information and improve the accuracy and robustness of channel quality assessment.

[0164] In step S42, dynamic optimization of resource allocation is achieved based on reinforcement learning. The state space of the reinforcement learning agent is defined. This includes channel quality assessment results and system resource status (such as remaining power). Number of idle time slots Available frequency bands etc.), motion space For different sets of resource allocation strategies, such as power allocation strategies Time slot allocation strategy Frequency band allocation strategy Reward function The goal is to maximize system performance (such as spectral efficiency (SE) and reduce bit error rate). The agent learns the optimal resource allocation strategy by interacting with the network environment and applying the Bellman equation.

[0165] The rationale for designing step S42 is that channel quality changes over time, and static resource allocation strategies cannot adapt to dynamic environments; reinforcement learning can learn the optimal resource allocation strategy through continuous interaction with the environment, thereby maximizing spectrum efficiency while ensuring communication quality.

[0166] In step S43, the transmission power, time slot allocation, or operating frequency band of the communication system is adjusted according to the action decided by reinforcement learning. Specifically, if the action is a transmission power update, the transmission power is updated; if the action is a time slot adjustment, the time slot allocation table is adjusted; if the action is a frequency band switching command, the system switches to the specified frequency band. The adjusted resource state will be fed back to the agent as input for the next time step, forming a closed-loop adaptive loop.

[0167] (5) Step S5: Output of detection results

[0168] This step outputs the final results of channel quality monitoring and assessment, which guide the communication system's adaptive modulation and coding, spectrum switching, or interference avoidance operations. Specifically, it includes the following sub-steps:

[0169] S51. Output channel quality assessment results: including the channel quality level or signal-to-noise ratio estimate Y obtained in step S41.

[0170] S52. Output adaptive background noise threshold curve: that is, the dynamic adaptive noise threshold curve generated in step S24, which is used to indicate the current background noise level.

[0171] S53, Output resource allocation decision: that is, the optimal resource allocation action output in step S42, including power adjustment amount, time slot allocation table, frequency band switching command, etc.

[0172] The above results can be transmitted to the communication control module in real time through the system interface to achieve adaptive modulation and coding (adjusting the modulation and coding method according to the channel quality), dynamic spectrum access (selecting idle frequency bands based on noise thresholds and frequency band decisions), power control, etc., thereby improving the reliability and spectrum efficiency of the communication system.

[0173] It should be noted that the various processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved. This embodiment does not impose any limitations on these steps.

[0174] Based on the above steps, this embodiment of the invention also provides a channel quality assessment system based on multimodal and quantum state tomography, comprising:

[0175] The adaptive sampling and trusted storage module is used to implement step S1 in the above method, namely, predicting the initial sampling parameters based on deep learning, dynamically optimizing the preset frequency interval by combining reinforcement learning, determining the sampling frequency point by quantum frequency synthesis, compressing the data by compressed sensing technology, physically storing the collected data through a memristor array, and constructing a hash chain to achieve tamper-proof verification.

[0176] The parallel computing and intelligent curve generation module is used to implement step S2 in the above method, that is, to construct the collected spectrum data into a three-dimensional tensor, generate average spectrum data through Tucker decomposition and distributed parallel computing, and then process it through a normalized flow model, variational autoencoder and generative adversarial network in sequence to generate a high-precision average spectrum curve, and dynamically optimize the background noise threshold through a deep reinforcement learning model to output an adaptive noise threshold curve.

[0177] The dual attention fusion and quantum state tomography module is used to implement step S3 in the above method, that is, to extract the time-frequency matrix from the average spectrum data, apply the self-attention mechanism to the time dimension and frequency dimension respectively for weighted fusion, encode the fusion features into quantum states, reconstruct the density matrix and extract quantum features through quantum measurement and tomography algorithms, and finally merge the fusion features and quantum features into a time-frequency fine feature vector.

[0178] The channel quality assessment and reinforcement learning resource allocation module is used to implement step S4 in the above method, that is, to input the time-frequency fine feature vector and spatial domain data into the multimodal deep neural network, output the channel quality estimate, and then use the estimate and the system resource state to form a state space, learn the optimal resource allocation strategy through reinforcement learning, and adjust the system resources accordingly.

[0179] The result output module is used to implement step S5 in the above method, that is, to output the channel quality assessment result, the adaptive background noise threshold curve and the resource allocation decision, so as to guide the communication system to perform adaptive modulation and coding, spectrum switching or interference avoidance operations.

[0180] The embodiments described in this invention can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with the implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0181] Computer programs for implementing the methods and systems of the present invention may be written in any combination of one or more programming languages ​​and stored in a computer-readable storage medium. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0182] Computer-readable storage media can be tangible media that may contain or store computer programs for use by or in conjunction with an instruction execution system, apparatus, or device. Computer-readable storage media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, computer-readable storage media can be machine-readable signal media. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, optical disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0183] In summary, the channel quality assessment method and system based on multimodal and quantum state tomography provided by the embodiments of the present invention have the following beneficial effects:

[0184] 1. To address the problems of low efficiency and insufficient frequency synthesis accuracy of fixed sampling strategies in existing technologies, this invention (step S1) uses deep learning to predict initial parameters and combines reinforcement learning to dynamically optimize the sampling interval, thereby achieving adaptive adjustment of the sampling strategy and solving the problem of not being able to balance sampling efficiency and accuracy in complex electromagnetic environments. At the same time, it introduces quantum frequency synthesis technology, which significantly improves the frequency generation accuracy compared to traditional PLL / DDS, avoids errors introduced by frequency drift, and lays the foundation for high-precision channel monitoring.

[0185] 2. To address the problems of large sample data volume, high hardware pressure, and lack of reliable security mechanisms in existing technologies, this invention (step S1) adopts RIP-based compressed sensing technology to achieve sparse spectrum acquisition at a rate far lower than the Nyquist rate, significantly reducing the sample data volume and hardware storage and transmission pressure. At the same time, by combining memristor physical storage with hash chain verification, a dual security defense of "physical + logic" is constructed, solving the problem of the lack of an immutable security mechanism for spectrum data during storage and transmission.

[0186] 3. To address the problem that existing technologies struggle to handle high-dimensional complex spectra and lack stability under noise disturbances, this invention (step S2) constructs the spectral data as a three-dimensional tensor and employs Tucker decomposition and distributed parallel computing, significantly improving the generation speed and robustness of the average spectral curve and solving the computational load problem of high-dimensional spectral data. Simultaneously, by combining normalized flow, VAE, and GAN joint optimization, the generated spectral curve possesses both statistical stability and high fidelity, avoiding the over-smoothing or noise residue problems caused by traditional linear filtering, windowing, and other methods.

[0187] 4. In view of the problem that background noise thresholds in existing technologies are mostly fixed or empirically set and lack adaptability, the present invention (step S2) constructs a deep reinforcement learning model, uses real-time channel noise parameters as the state and aims to reduce the false positive rate to dynamically adjust the noise threshold, thereby achieving adaptive optimization of the background noise threshold, greatly reducing the probability of misjudging noise as a signal or vice versa, and improving the accuracy of signal detection.

[0188] 5. To address the problem that feature extraction in existing technologies is usually based only on a single time domain or frequency domain and fails to achieve multi-domain fusion, this invention (step S3) applies self-attention to the time and frequency dimensions of the time-frequency matrix respectively through a dual attention mechanism, realizing the deep fusion of long-term time domain dependencies and key frequency domain components, and comprehensively highlighting the important features in the time-frequency structure; at the same time, quantum state tomography is introduced to encode the fused features into quantum states and reconstruct the density matrix, extracting fine features such as weak modulation and phase noise that are difficult to capture by classical methods, thus solving the problems of incomplete feature extraction and insufficient resolution.

[0189] 6. To address the problem that existing technologies often use single-modal deep learning models and have difficulty processing multi-source data, this invention (step S4) constructs a multi-modal deep neural network to fuse time-frequency fine features with spatial domain data (such as antenna array phase, interference source location, etc.) to evaluate channel quality, thus solving the problem of limited evaluation accuracy of single-modal models; by fusing multi-source information in the time domain, frequency domain, and spatial domain, the accuracy and robustness of channel quality evaluation are significantly improved.

[0190] 7. In view of the problem that resource allocation in the prior art mostly relies on static rules or simple heuristic algorithms and cannot achieve real-time optimization, the present invention (step S4) constructs a state space with the channel quality assessment results and the system resource status, and dynamically optimizes the resource allocation strategies such as power, time slots and frequency bands through reinforcement learning, thereby realizing closed-loop adaptive optimization of resource allocation; compared with static allocation strategies, it can respond to channel changes in real time and maximize spectrum efficiency while ensuring communication quality.

[0191] 8. The present invention (step S5) outputs the channel quality assessment results, adaptive background noise threshold curve and resource allocation decision in real time to guide the adaptive modulation and coding, spectrum switching or interference avoidance of the communication system. It realizes a complete closed loop from spectrum data acquisition, processing, feature extraction to quality assessment and resource allocation. It can effectively adapt to complex electromagnetic environments and provide high-precision, intelligent and reliable channel monitoring and decision-making capabilities, and comprehensively solves the defects of the prior art mentioned in the background art.

[0192] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A channel quality assessment method based on multimodal and quantum state tomography, characterized in that, Including the following steps: S1. Generate a sampling frequency based on deep learning, and collect and store channel spectrum data within a preset time based on the generated sampling frequency. S2. Process the stored spectrum data to generate an average spectrum curve and an adaptive background noise threshold curve; S3. Based on the average spectral curve, time-frequency features are fused using dual attention and refined time-frequency feature vectors are extracted using quantum state tomography; Step S3 specifically includes the following steps: S31. Extract the time-frequency matrix from the average spectrum data; S32. Apply the self-attention mechanism to the time dimension and frequency dimension of the time-frequency matrix respectively to obtain the weighted time-domain features and frequency-domain features, and then fuse them to obtain the fused features. S33. Encode the fused features into quantum states, and reconstruct the density matrix using quantum measurement and tomography algorithms to extract quantum features; Step S33 specifically involves: Fusion features Normalized to unit length, resulting in a normalized vector. ; Amplitude coding is used for normalized vectors Constructing quantum pure states , and its density matrix : Select a set of observables For quantum pure state Perform projection measurements to obtain the desired value. ; The density matrix was reconstructed from multiple measurement results using a quantum state tomography algorithm. And extract its features or eigenvalues ​​to form a quantum eigenvector. ; S34. Combine the fused features with the quantum features to obtain the final time-frequency fine feature vector; S4. Utilize the extracted time-frequency fine feature vectors to evaluate channel quality through a multimodal deep neural network, and combine reinforcement learning to dynamically optimize resource allocation strategies. S5, output channel quality assessment results, adaptive background noise threshold curve and resource allocation strategy.

2. The channel quality assessment method based on multimodal and quantum state tomography according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Construct a multimodal deep neural network model, input the time-frequency fine feature vector and spatial domain data into the multimodal deep neural network model, and output the channel quality estimate after fusion. S42. Construct a state space from the channel quality estimate and the system resource state, define the action space and reward function, and learn the optimal resource allocation strategy through reinforcement learning. S43. Adjust system resources based on the actions made by reinforcement learning decisions and provide feedback on the new state.

3. The channel quality assessment method based on multimodal and quantum state tomography according to claim 2, characterized in that, Step S41 is as follows: Time-frequency fine feature vector Decomposed into time domain part and frequency domain part ; Multimodal deep neural network models for the time domain Frequency domain part and spatial domain data Extract high-level features separately; The high-level features extracted from the three modalities are concatenated to obtain multimodal fusion features; The multimodal fusion features are processed through a fully connected layer and an activation function to output a channel quality estimate.

4. The channel quality assessment method based on multimodal and quantum state tomography according to claim 1, characterized in that, Step S1 specifically includes: S11. When the system starts, it uses historical channel data and a long short-term memory network or a gated recurrent unit deep learning model to predict the initial set of sampling parameters, including the starting frequency, preset frequency interval, sampling frequency and preset time. S12. Construct the adjustment strategy of the preset frequency interval into a Markov decision process: Define the state space and action space, the agent selects an action according to the current state, and evaluates the effect of the action through the reward function to obtain the optimal preset frequency interval. S13. Based on the obtained optimal frequency interval and the predicted starting frequency, determine each sampling frequency point through quantum frequency synthesis; S14. At each sampling frequency point, compressed sensing technology is used to compress and collect data at the predicted sampling frequency and preset time to obtain compressed data. S15. The collected compressed data is physically stored using a memristor array.

5. The channel quality assessment method based on multimodal and quantum state tomography according to claim 4, characterized in that: In step S12, the reward function Designed for , For the agent in state Execute action The reward value obtained immediately afterwards Indicates the current state Spectrum utilization rate Indicates the current state The spectrum sampling error is as follows. They are respectively and Weighting coefficients; In step S13, when the optimal preset frequency interval is determined... Then, the corresponding physical sampling point frequency , For the first Each sampling frequency point , The number of sampling frequency points, The starting frequency; In step S14, the compressed data is represented as , The original signal, For the measurement matrix, It is a sparse base.

6. The channel quality assessment method based on multimodal and quantum state tomography according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Organize the collected spectrum data into a three-dimensional tensor including three dimensions: time dimension length, number of frequency points, and number of measurements. Use Tucker decomposition to perform parallel computation on the tensor to obtain the average spectrum data. S22. Input the average spectrum data into the normalized flow model, and map it to the standard normal distribution through an invertible transformation to obtain the normalized latent variables. S23. Input the normalized latent variables into the variational autoencoder for dimensionality reduction and feature extraction. Input the extracted key latent space features into the generative adversarial network. The generative adversarial network generates the average spectral curve. S24. Construct a deep reinforcement learning model, using the real-time channel noise parameters in the average spectrum curve as the state, the adjustment amount of the background noise threshold as the action, and design a reward function with the goal of reducing the false judgment rate. The agent learns the optimal strategy by interacting with the environment and finally outputs a dynamic adaptive noise threshold curve.

7. The channel quality assessment method based on multimodal and quantum state tomography according to claim 6, characterized in that: In step S23, the encoder of the variational autoencoder normalizes the latent variables. Mapped to a lower-dimensional latent space Decoder reconstruction The loss function is designed as a weighted sum of the reconstruction loss and the KL divergence.

8. A channel quality assessment system based on multimodal and quantum state tomography, characterized in that, include: The adaptive sampling and trusted storage module, the parallel computing and intelligent curve generation module, the dual attention fusion and quantum state tomography module, the channel quality assessment and reinforcement learning resource allocation module, and the result output module are respectively used to implement steps S1 to S5 in the channel quality assessment method based on multimodal and quantum state tomography as described in any one of claims 1 to 7.

Images