Priori fusion risk calibration spectrum prediction method, device, equipment and storage medium

Abstract

The application relates to a priori fusion risk calibration spectrum prediction method, device, equipment and storage medium. The method comprises the following steps: acquiring historical spectrum observation data, regulatory-physical knowledge graph and environmental context information; estimating the time-frequency activation probability of a candidate emission source and generating a physical priori spectrum field; inputting the data into a spectrum prediction network to obtain a future three-dimensional spectrum state tensor prediction result; constructing a non-consistency score by using the prediction residual of a calibration sample set to generate a multi-dimensional conformal prediction tube covering the future three-dimensional spectrum state tensor; screening a safe action set according to the conservative upper limit provided by the prediction tube and regulatory and interference constraints, adjusting the conservativeness by using a risk calibration parameter, and selecting a dynamic spectrum access action with the maximum communication utility. The method can fuse regulatory and physical priori, generate a multi-dimensional prediction tube, realize risk-controllable dynamic spectrum access, and improve the safety and efficiency of spectrum utilization in a non-stationary environment.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a priori fusion risk calibration spectrum prediction method, apparatus, device and storage medium. Background Technology

[0002] Dynamic spectrum access technology is a key means to alleviate the contradiction between spectrum scarcity and low spectrum utilization under the static licensing allocation model. Its core idea is to allow unlicensed users (secondary users) to opportunistically access idle spectrum resources without causing harmful interference to licensed users (primary users). To achieve secure and efficient dynamic spectrum access, accurate prediction of future spectrum occupancy and quantification of the uncertainty of the prediction results are prerequisites for supporting decision-making and mitigating interference risks.

[0003] Existing spectrum prediction methods can be mainly divided into three categories: methods based on statistical models or purely data-driven deep learning, which use historical spectrum observations as input to predict future spectrum states; methods based on physical propagation models, which use emission source parameters and geographical environment to calculate the spatial spectrum field distribution; and methods based on uncertainty quantification, which provide prediction intervals through quantile regression or Bayesian neural networks. These methods have already been applied to some extent in practical systems.

[0004] However, the aforementioned existing technologies still have the following shortcomings. On the one hand, existing methods struggle to integrate heterogeneous knowledge from multiple sources, such as regulatory approvals, frequency band usage, prior knowledge of the transmitter, and geographical propagation. Purely data-driven models lack generalization ability in cross-regional and non-stationary scenarios, while physical propagation models rely on precise transmitter parameters, which are difficult to obtain accurately in practice. On the other hand, most existing uncertainty estimates are point-by-point intervals for a single frequency point or a single spatial location, failing to characterize the joint error structure of the time-frequency-space tensor composed of multiple time steps, multiple frequencies, and multiple spatial regions, making it difficult to support spectrum access decisions involving multiple resource blocks. Furthermore, concept drift in non-stationary spectrum environments can cause pre-calibrated prediction intervals to become invalid, and existing methods lack adaptive adjustment mechanisms. Finally, existing technologies lack a complete technical solution to effectively transform prediction uncertainties into dynamic spectrum access action-level risk constraints (such as primary user collision probability, interference threshold, and regulatory compliance) and guide security decisions. Summary of the Invention

[0005] Based on this, it is necessary to provide a priori fusion risk calibration spectrum prediction method, device, equipment, and storage medium that can integrate regulatory and physical priors, generate multi-dimensional predictive pipes, and realize risk-controllable dynamic spectrum access to address the above-mentioned technical problems.

[0006] A priori fusion risk calibration spectrum prediction method, the method comprising: Acquire historical spectrum observation data, characterize spectrum regulatory rules, regulatory-physical knowledge graphs of physical space and emission source association structures, and environmental context information describing the spatiotemporal and operational environment in which the prediction task is located; Based on the regulatory-physical knowledge graph and the environmental context information, the time-frequency activation probability of the candidate emission source is estimated, and a physical prior spectrum field is generated based on the activation probability and the propagation model. The historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field are input into the spectrum prediction network to obtain the prediction result of the future three-dimensional spectrum state tensor. The future three-dimensional spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels and multiple spatial locations in the future. A non-consistent score is constructed using the prediction residuals on the calibration sample set. Conformal quantiles are calculated on the non-consistent score based on a preset target coverage rate, and a multidimensional conformal prediction tube covering the future three-dimensional spectral state tensor is generated. A set of safe actions is selected based on the conservative upper bound of the spectrum state, spectrum regulatory constraints, and interference protection thresholds provided by the multidimensional conformal prediction tube. The conservatism of the set of safe actions is adjusted by risk calibration parameters, and dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility are selected from the set of safe actions after risk calibration.

[0007] In one embodiment, the regulatory-physical knowledge graph includes candidate emission source nodes, frequency nodes, spatial region nodes, service type nodes, regulatory rule nodes, and time event nodes; The edge relationships in the regulatory-physical knowledge graph include: the authorized use relationship between candidate transmitter nodes and frequency nodes, the location affiliation and coverage relationship between candidate transmitter nodes and spatial region nodes, the adjacent or interference relationship between different frequency nodes, the protection constraint relationship between regulatory rule nodes and frequency nodes or spatial region nodes, and the time scheduling relationship between time event nodes and the activation state of candidate transmitter nodes.

[0008] In one embodiment, estimating the time-frequency activation probability of candidate emission sources based on the regulatory-physical knowledge graph and the environmental context information includes: The target output dimension of the prediction task includes multiple future prediction times and multiple prediction frequencies; Based on the node attributes in the regulatory-physical knowledge graph, generate the initial embedding of each node; The initial embeddings of the candidate emission source nodes are extracted from the regulatory-physical knowledge graph as attribute embeddings, the initial embeddings of the frequency nodes are extracted as frequency embeddings, the initial embeddings of the service type nodes are extracted as service embeddings, and the regulatory authorization embeddings are extracted. Extract the time feature code and event context feature corresponding to the predicted time from the environmental context information; The extracted embeddings and features are concatenated into vectors, and the activation score is obtained after linear transformation. The activation score is input into the sigmoid activation function to obtain the activation probability of each candidate emission source at each prediction time and each prediction frequency.

[0009] In one embodiment, generating a physical prior spectrum field based on the activation probability and propagation model includes: Based on the propagation model and incorporating a learnable correction term, calculate the path loss from the candidate emission source to the spatial location at the predicted frequency; Based on the transmit power, antenna gain, path loss, and activation probability of the candidate transmitter at the prediction time and prediction frequency, calculate the linear received power of the candidate transmitter at the spatial location. The linear received power of all candidate transmitters at the same prediction time, the same prediction frequency, and the same spatial location is linearly superimposed with the noise power to obtain the bus linear received power. The bus-type received power is converted to the dB domain to obtain the value of the physical prior spectrum field at the prediction time, prediction frequency, and spatial location.

[0010] In one embodiment, the spectrum prediction network employs an attention-based neural network, wherein the attention logit of the attention mechanism incorporates physical propagation bias, regulatory relationship bias, and geographic proximity bias. The spectrum prediction network takes the historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field as input, and uses physical prior plus residual learning to output the point prediction results of the future three-dimensional spectrum state tensor.

[0011] In one embodiment, constructing a non-consistent score using prediction residuals on a calibration sample set includes: The prediction residual of each sample in the calibration sample set is vectorized to obtain a high-dimensional residual vector; The high-dimensional residual vector is mapped to a low-dimensional space by an encoder to obtain a low-dimensional residual representation; Based on the context group to which the calibration sample belongs, the mean and covariance matrix of the low-dimensional residual representation within the group are calculated, wherein the context group is obtained by dividing the calibration sample according to the region type, frequency band type, service type and / or time period; For each calibration sample, a joint inconsistency score is constructed based on the Mahalanobis distance between the corresponding low-dimensional residual representation and the mean of its group, wherein the joint inconsistency score is used to measure the overall prediction error of the future three-dimensional spectral state tensor.

[0012] In one embodiment, adjusting the conservatism of the set of safety actions using risk calibration parameters includes: Obtain a risk calibration sample set, which includes historical prediction results, historical context information, and the actual spectrum state observed after historical decisions; For each of the preset multiple candidate risk calibration parameters, the comprehensive risk generated by performing dynamic spectrum access actions when using a certain candidate risk calibration parameter is evaluated using the risk calibration sample set to obtain the corresponding empirical risk; Calculate the upper confidence bound of the empirical risk corresponding to each candidate parameter; The smallest candidate parameter that satisfies the upper confidence bound not exceeding the preset target risk level is selected as the risk calibration parameter.

[0013] This application also proposes a priori fusion risk calibration spectrum prediction device, the device comprising: The data acquisition module is used to acquire historical spectrum observation data, characterize spectrum regulatory rules, regulatory-physical knowledge graphs of physical space and emission source association structures, and environmental context information describing the spatiotemporal and operational environment in which the prediction task is located; The prior generation module is used to estimate the time-frequency activation probability of candidate emission sources based on the regulatory-physical knowledge graph and the environmental context information, and generate a physical prior spectrum field based on the activation probability and the propagation model. The spectrum prediction module is used to input the historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field into the spectrum prediction network to obtain the prediction result of the future three-dimensional spectrum state tensor. The future three-dimensional spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels and multiple spatial locations in the future. The conformal calibration module is used to construct non-consistency scores using the prediction residuals on the calibration sample set, calculate conformal quantiles on the non-consistency scores according to the preset target coverage rate, and generate a multidimensional conformal prediction tube covering the future three-dimensional spectral state tensor. The risk decision module is used to select a set of safe actions based on the upper conservative bound of the spectrum state, spectrum regulatory constraints and interference protection thresholds provided by the multi-dimensional conformal prediction tube, adjust the conservatism of the set of safe actions through risk calibration parameters, and select dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility from the set of safe actions after risk calibration.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the specific steps in the above-described prior fusion risk calibration spectrum prediction method.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the specific steps in the aforementioned prior fusion risk calibration spectrum prediction method.

[0016] The aforementioned prior fusion risk calibration spectrum prediction method, apparatus, equipment, and storage medium acquire historical spectrum observation data, a regulatory-physical knowledge graph characterizing spectrum regulatory rules, physical space, and the associated structure of emission sources, and environmental context information describing the spatiotemporal and operational environment of the prediction task. Based on the regulatory-physical knowledge graph and environmental context information, it estimates the time-frequency activation probability of candidate emission sources and generates a physical prior spectrum field based on the activation probability and propagation model. The historical spectrum observation data, regulatory-physical knowledge graph, and physical prior spectrum field are then input into a spectrum prediction network to obtain the prediction result of the future three-dimensional spectrum state tensor. The spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels, and multiple spatial locations in the future. Non-consistency scores are constructed using the prediction residuals on the calibration sample set. Conformal quantiles are calculated on the non-consistency scores based on the preset target coverage rate, generating a multi-dimensional conformal prediction tube that covers the future three-dimensional spectrum state tensor. A set of safe actions is selected based on the conservative upper bound of the spectrum state, spectrum regulatory constraints, and interference protection thresholds provided by the multi-dimensional conformal prediction tube. The conservatism of the set of safe actions is adjusted by risk calibration parameters, and dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility are selected from the risk-calibrated set of safe actions.

[0017] Beneficial effects: This method integrates regulatory-physical knowledge graphs with physical prior spectrum fields, enabling spectrum prediction to combine the interpretability of regulatory rules with the physical consistency of propagation. This reduces the dependence of pure data-driven models on large-scale labeled samples and the distortion risk of cross-domain migration. By generating a multidimensional conformal prediction tube covering the future three-dimensional spectrum state tensor using inconsistency scores constructed based on low-dimensional residual representation and Mahalanobis distance, it can quantify the time-frequency-space joint prediction error and provide a statistically guaranteed uncertainty boundary, overcoming the shortcomings of traditional point-by-point intervals that cannot characterize the overall error structure. Furthermore, by adaptively adjusting the conservatism of the set of safety actions through risk calibration parameters, and selecting the access action with the highest communication utility under the constraints of primary user protection, interference threshold, and regulatory compliance, this method achieves reliable prediction and risk-controllable dynamic spectrum access in non-stationary spectrum environments, effectively reducing the probability of harmful interference and improving spectrum utilization efficiency. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a priori fusion risk calibration spectrum prediction method in one embodiment; Figure 2 This is a schematic diagram illustrating the specific framework of the method proposed in this paper in one embodiment; Figure 3 This is a structural block diagram of a priori fusion risk calibration spectrum prediction device in one embodiment; Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] To address the problems of existing technologies, such as the difficulty in integrating regulatory approval with physical propagation priors, the inability to characterize the joint time-frequency-space prediction error structure, the lack of adaptive mechanisms to environmental drift, and the difficulty in transforming prediction uncertainty into controllable risk constraints, this application addresses these issues. Figure 1 As shown, a priori fusion risk calibration spectrum prediction method is provided, which specifically includes the following steps: Step S100: Obtain historical spectrum observation data, a regulatory-physical knowledge graph characterizing spectrum regulatory rules, physical space and emission source association structures, and environmental context information describing the spatiotemporal and operational environment in which the prediction task is located.

[0021] Step S110: Based on the regulatory-physical knowledge graph and environmental context information, estimate the time-frequency activation probability of the candidate emission source, and generate a physical prior spectrum field based on the activation probability and propagation model.

[0022] Step S120: Input historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field into spectrum prediction network to obtain prediction results of future three-dimensional spectrum state tensor. The future three-dimensional spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels and multiple spatial locations in the future.

[0023] Step S130: Construct non-consistency scores using the prediction residuals on the calibration sample set, calculate conformal quantiles on the non-consistency scores based on the preset target coverage rate, and generate a multidimensional conformal prediction tube covering the future three-dimensional spectral state tensor.

[0024] Step S140: Based on the conservative upper bound of spectrum state, spectrum regulatory constraints and interference protection threshold provided by the multidimensional conformal prediction tube, a set of safe actions is selected. The conservatism of the set of safe actions is adjusted by risk calibration parameters. Then, dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility are selected from the set of safe actions after risk calibration.

[0025] In step S100, at the current time Can obtain the past The spectral observation tensor at each time step is the historical spectral observation data. The spectral observation matrix at each time step express A frequency channel or frequency point in The received power on a spatial grid, monitoring node, or geographical area. This method requires predicting future... Spectral tensor at each time step .

[0026] Specifically, historical spectrum observation data can be represented as: (1) The future spectrum observation data to be predicted can be represented as: (2) The overall input-output relationship of this method can be expressed as follows: (3) In formula (3), For future spectrum point prediction results, The target coverage rate is The prediction set, For the final dynamic spectrum access action, To allow for risky water.

[0027] For ease of subsequent description, the main symbols involved in this application are defined as follows: Indicates the current discrete-time index; Indicates the length of the historical observation window; Indicates the predicted time span; Indicates the number of frequency channels, frequency points, or sub-bands; Indicates the number of spatial grids, monitoring nodes, or geographical areas; Indicates the number of candidate launch sources; Represents the current regulatory-physical knowledge graph; This represents geographical, business, temporal, and environmental contextual characteristics, i.e., environmental contextual information. Indicates the source of the emission At any moment ,frequency The potential activation probability on; Indicates the target coverage rate The prediction set; This indicates the final dynamic spectrum access action selected.

[0028] In this embodiment, regulatory data, candidate transmitters, frequency bands, regions, sensors, services, and time events are uniformly encoded into a heterogeneous knowledge graph. Graph nodes include, but are not limited to: transmitter nodes, frequency nodes, spatial region nodes, monitoring nodes, service type nodes, regulatory rule nodes, and time event nodes. Graph edges include, but are not limited to: authorized usage relationships between candidate transmitter nodes and frequency nodes; location affiliation and coverage relationships between candidate transmitter nodes and spatial region nodes; adjacency or interference relationships between different frequency nodes; protection and constraint relationships between regulatory rule nodes on frequency nodes or spatial region nodes; and time scheduling relationships between time event nodes and the activation status of candidate transmitter nodes.

[0029] In this embodiment, the knowledge graph is represented as: (4) In formula (4), For a set of nodes, Let be the set of edges. For a set of relation types, and These are node features and edge features, respectively.

[0030] Specifically, for any node in the knowledge graph Its initial node representation can be obtained by concatenating node type embedding and attribute encoding: (5) In formula (5), Represents the original attributes of a node. This represents an attribute encoder. For candidate transmitter nodes, attributes may include latitude and longitude, antenna height, rated power, antenna pattern, affiliated service, licensed frequency band, and license validity period; for regional nodes, attributes may include terrain height, building density, vegetation cover, road density, and whether they are located in a protected area.

[0031] Furthermore, in the graph coding stage, targeting relation types... Relationship-aware attention can be used to aggregate neighboring nodes: (6) Among them, the neighbor messages after relational attention aggregation are denoted as The node is updated as follows: (7) In this way, licenses, protected areas, frequency band usage, transmission source locations, and service time patterns are converted into structured embeddings that can be used by subsequent spectrum prediction networks.

[0032] In step S110, the time-frequency activation probability of candidate emission sources is estimated based on the regulatory-physical knowledge graph and the environmental context information. This includes: the target output dimension of the prediction task includes multiple future prediction times and multiple prediction frequencies. Initial embeddings for each node are generated based on the node attributes in the regulatory-physical knowledge graph. The initial embeddings of candidate emission source nodes are extracted from the regulatory-physical knowledge graph as attribute embeddings, the initial embeddings of frequency nodes are extracted as frequency embeddings, the initial embeddings of the service type nodes are extracted as service embeddings, and regulatory authorization embeddings are extracted. The time feature encoding and event context features corresponding to the prediction times are extracted from the environmental context information. The extracted embeddings and features are concatenated into vectors, and after linear transformation, an activation score is obtained. The activation score is input into the sigmoid activation function to obtain the activation probability of each candidate emission source at each prediction time and each prediction frequency.

[0033] Specifically, for candidate launch sources ,time and frequency First, calculate the activation score: (8) Next, the activation probability is calculated: (9) In formulas (8) and (9), For the sigmoid function, This indicates vector concatenation.

[0034] In this embodiment, regulatory authorization is not directly considered as a true spectrum occupancy label, but rather as a noisy prior. For candidate emission sources... ,time and frequency Set binary latent variables and in accordance with the above As an estimate of its activation probability. The activation score and activation probability are given by formulas (8) and (9), respectively. If some real emission logs are obtained (e.g., through sensors or historical records), the above activation probability estimation process can be trained using supervised cross-entropy loss, expressed as: (10) If the actual emission log is unavailable, the activation probability can be constrained using sparse regularization and smooth regularization to avoid overfitting or drastic jumps, as shown below: (11) In formula (11), sparsity regularization prevents the system from treating all candidate emission sources as active, while smoothing regularization is used to suppress unreasonable and drastic jumps in a short period of time. For intermittent services such as radar pulses and emergency communications, the weight of smoothing regularization can be adjusted or service-related pulse priors can be introduced.

[0035] It should be noted that the above activation probability estimation can be achieved through a neural network model, which is trained using the cross-entropy loss of formula (10) or the regularization loss of formula (11). In this way, the activation probability estimation can flexibly adapt to different regulatory information and data availability scenarios, providing a reasonable emission source activation estimate for the subsequent generation of the physical prior spectrum field.

[0036] In this embodiment, generating a physical prior spectrum field based on activation probability and propagation model includes: calculating the path loss from the candidate transmitter to the spatial location at the predicted frequency according to the propagation model and in combination with learnable correction terms; calculating the linear received power of the candidate transmitter at the spatial location based on the transmit power, antenna gain, path loss, and activation probability of the candidate transmitter at the predicted time and predicted frequency; linearly superimposing the linear received power of all candidate transmitters at the same predicted time, the same predicted frequency, and the same spatial location with the noise power to obtain the bus-level received power; and converting the bus-level received power to the dB domain to obtain the value of the physical prior spectrum field at the predicted time, predicted frequency, and spatial location.

[0037] Specifically, for candidate launch sources To spatial location In frequency The path loss can be expressed as the sum of the output of the empirical propagation model (or the ray tracing approximation model, the learnable propagation model) and the learnable correction term: (12) In formula (12), This is a learnable correction term used to locally correct the empirical model to adapt it to actual geographical environments and propagation conditions. To maintain the interpretability of the correction term and avoid overfitting, a propagation correction regularization term can be introduced: (13) The linear received power contributed by a single candidate transmitter is determined by the transmit power, antenna gain, path loss, and activation probability, and is expressed as: (14) In formula (14), The activation probability is the one estimated above. As the source of launch In frequency The transmit power (dBm) on the device. For antenna gain, The path loss is (dB). It should be noted that the activation probability is explicitly multiplied in formula (14), which reflects the prior uncertainty that "the source may not emit".

[0038] For all candidate emission sources, at the same prediction time same frequency and the same spatial location The linear received power is linearly superimposed with the noise power to obtain the bus linear received power, expressed as: (15) In formula (15), The total number of candidate launch sources. The noise is linear power. If the channel bandwidth is... The receiver noise figure is Then the linear power of the noise can be approximated as: (16) In formula (16), dBm / Hz is an approximation of the power spectral density of room temperature thermal noise. The physical prior spectrum field does not replace the actual measurement, but rather provides a baseline for the prediction model that satisfies the propagation law.

[0039] Finally, the bus-mode received power is converted to the dB domain to obtain the values ​​of the physical prior spectrum field at the corresponding prediction time, frequency, and spatial location, where the physical prior spectrum field is expressed as: (17) In formula (17), It is a small positive number used to avoid taking the logarithm of zero.

[0040] It should be noted that the aforementioned physical prior spectrum field does not replace actual measurements, but rather provides a baseline for the prediction model that satisfies the laws of propagation and regulatory priors. The true future spectrum can be modeled as the sum of physical priors, system residuals, and random noise, expressed as: (18) In formula (18), For the future The physical prior spectrum field of the step, The system residuals are learned by the subsequent spectrum prediction network. This represents the random noise term. Through the above method, the physical prior spectral field, while maintaining physical consistency, provides a good starting point for data-driven residual learning.

[0041] In step S120, the spectrum prediction network adopts a neural network based on the attention mechanism. The attention mechanism introduces physical propagation bias, regulatory relationship bias and geographical proximity bias in the attention logit. The spectrum prediction network takes historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field as input, and adopts physical prior plus residual learning to output the point prediction result of the future three-dimensional spectrum state tensor.

[0042] Specifically, the overall prediction process of the spectrum prediction network can be represented as follows: (19) In formula (19), For the future The physical prior spectrum field of the step, This is the residual correction term learned by the neural network. This architecture encourages the network to learn only the deviation between the physical prior and the true spectrum, rather than fitting the entire spectrum state from scratch, thereby improving data efficiency and generalization ability.

[0043] In this embodiment, the spectrum prediction network adopts a Transformer structure to jointly encode the historical spectrum tensor, physical prior field, and knowledge graph embedding. First, the historical spectrum tensor is divided into time-frequency-spatial local patches. The token for each patch is represented as: (20) In formula (20), For the first Spectral observations within a patch It is a learnable linear mapping matrix. , , These are location codes for time, frequency, and spatial location, respectively.

[0044] As for the predicted future location Construct a future query token, represented as: (twenty one) In formula (21), For the graph context vector related to future location, To predict the embedding of time steps, For frequency embedding, For spatial embedding, This is the encoding function for physical prior values.

[0045] In this embodiment, the network's attention mechanism incorporates physical propagation bias, supervisory relationship bias, geographic proximity bias, and missing observation mask bias. The attention logit is represented as: (twenty two) The attention output is represented as: (twenty three) In formulas (22) and (23), For physical propagation bias (such as setting based on path loss or distance between the transmitter and receiver). For regulatory relationship bias (such as attention weight between authorized users or within protected areas). Geographic proximity bias (gaining higher attention between spatially adjacent grids). Used to mask missing observation locations; This is a learnable bias strength coefficient.

[0046] Furthermore, through the encoder and decoder with the aforementioned bias attention, the network output residual correction term is expressed as: (twenty four) In formula (24), This is the feature vector at the corresponding position output by the decoder. The final point prediction result is the sum of the physical prior and the residual correction term, expressed as: (25) In addition to point prediction, the spectrum prediction network also outputs upper and lower quantile estimates and local scale parameters for subsequent uncertainty quantification. (26) (27) In formulas (26) and (27), You can take the lower quantile and the upper quantile, for example and . Ensure the output is positive. This represents a very small constant to prevent it from being zero.

[0047] In this embodiment, multiple loss functions are jointly optimized when training the spectrum prediction network. The point prediction loss uses Huber loss, and the observation mask is used... (Mark whether there are valid observations at this location) Calculate below: (28) Quantile loss, using pinball loss, is expressed as: (29) In addition, other regularization terms can be introduced, such as occupancy loss. Propagation modification regularity and residual sparsity regularization Then the total loss function during the training phase is expressed as: (30) This total loss function encourages the spectral prediction network to learn only the necessary residual corrections based on physical priors, rather than relying entirely on free fitting, thereby improving the model's physical consistency, interpretability, and cross-scenario generalization ability.

[0048] In step S130, inconsistency scores are constructed using the prediction residuals on the calibration sample set, and conformal quantiles are calculated based on a preset target coverage rate, ultimately generating a multidimensional conformal prediction tube capable of covering the future three-dimensional spectral state tensor. Conformal prediction is a statistical inference framework that does not rely on assumptions about data distribution. Its core idea is to quantify the uncertainty of the prediction model on new samples using inconsistency scores on the calibration set, and thereby construct a prediction set with marginal coverage guarantees.

[0049] Specifically, since the future spectrum state is a tensor containing three dimensions—time, frequency, and space—traditional point-by-point prediction intervals cannot characterize the joint error structure between different dimensions. Therefore, this method designs a non-consistent score that reflects the correlation of high-dimensional residuals, and generates a prediction tube covering the entire future three-dimensional spectrum state tensor based on this score, thus providing a statistically reliable uncertainty boundary for subsequent dynamic spectrum access decisions.

[0050] In this embodiment, constructing a non-consistency score using the prediction residuals on the calibration sample set includes: vectorizing the prediction residuals of each sample in the calibration sample set to obtain a high-dimensional residual vector; mapping the high-dimensional residual vector to a low-dimensional space using an encoder to obtain a low-dimensional residual representation; calculating the mean and covariance matrix of the low-dimensional residual representations within the group according to the context group to which the calibration sample belongs; wherein the context group is obtained by dividing the calibration samples according to region type, frequency band type, service type and / or time period; and for each calibration sample, constructing a joint non-consistency score based on the Mahalanobis distance between the corresponding low-dimensional residual representation and the mean of the group to which it belongs; wherein the joint non-consistency score is used to measure the overall prediction error of the future three-dimensional spectral state tensor.

[0051] Specifically, after the trained spectrum prediction network is fixed, an independent calibration set is used to calculate the inconsistency score and conformal quantile. For the calibration samples... First, calculate its predicted residual vector and quantize it: (31) In formula (31), For the true spectrum tensor, This is the result for the predicted points. Due to the dimensionality of the original residual vector... The values ​​are typically very high, and direct use would lead to computational and statistical difficulties. Therefore, this method uses an encoder... Mapping the high-dimensional residual vector to a low-dimensional space yields the low-dimensional residual representation: (32) Specifically, the encoder can be implemented using an autoencoder, principal component analysis, or a trainable neural network, with the aim of reducing dimensionality while preserving the main correlation structure of the residuals.

[0052] To accommodate the differences in residual distribution across various scenarios, a context grouping mechanism is introduced. The context grouping function is defined as follows: (33) The aforementioned context grouping function divides calibration samples into different groups based on the geographic region type, frequency band type, service type, and / or time period of the samples. In the middle. For each group Assuming it contains For each calibration sample, calculate the mean and covariance matrix of the low-dimensional residual representation within that group: (34) (35) In formula (35), This is a regularization term that ensures the covariance matrix is ​​invertible.

[0053] Based on this, a joint non-consistent score is constructed, defined as the Mahalanobis distance between the low-dimensional residual representation and the mean of its group, expressed as: (36) The fraction represented by formula (36) can reflect the magnitude of the joint prediction error of the entire three-dimensional spectral tensor: when the residual direction of each dimension is consistent with the principal direction within the group or the residual amplitude is large, the Mahalanobis distance will increase significantly.

[0054] In this embodiment, relying solely on joint inconsistency scores may fail to capture extreme prediction errors occurring at a local location (e.g., a specific time, frequency, or spatial point). Therefore, this method further introduces local inconsistency scores. The lower quantile estimate from the output of the spectral prediction network is then used. Upper quantile estimation and local scale parameters ,definition: (37) The fraction represented by formula (37) measures the maximum relative extent to which the true value exceeds the predicted quantile interval at any position in the three-dimensional tensor. If all positions fall within the quantile interval, then... Otherwise, take the largest excess amount.

[0055] To avoid the limitations of joint fractions and local fractions, they are combined into a composite nonconformist fraction, represented as: (38) In formula (38), The score representing the local inconsistency is obtained from quantile estimation and local scale parameters. The balancing coefficient is used to adjust the influence of local extreme errors on the overall score. By taking the maximum value, the composite score can simultaneously reflect both the overall joint error and local extreme errors.

[0056] Furthermore, for each context group The composite non-consistency score of all calibration samples within the group. Sort in ascending order and based on the preset target coverage rate. Calculate the conformal quantiles. In the basic implementation, sorting statistics are used: (39) Take the sorted number The quantile values ​​are used as the quantile threshold for this group. For a new sample to be predicted, its grouping is first determined based on its context information. And calculate its composite non-consistency score. (in (assuming the true future spectrum tensor), then the multidimensional conformal prediction tube is defined as: (40) The prediction tube covers all future three-dimensional spectral state tensors that ensure the non-consistency score does not exceed the grouping quantile. According to conformal prediction theory, the probability that this predictor covers the true future spectral tensor is at least... (In a marginal sense), and independent of the specific form of data distribution.

[0057] To address temporal shifts in the spectrum environment (such as day-night cycles and sudden events) and maintain coverage assurance for the predictive tubes, this method may also include an online adaptive calibration mechanism. Specifically, for each group... Maintain an online calibration buffer And an exponentially decaying weight is used to assign higher importance to recent samples: (41) Weighted quantiles are defined as the smallest quantiles that satisfy the following formula : (42) At each time step, after the true spectrum is observed, the coverage loss flag is calculated: (43) And based on this, the target coverage rate of the group is adaptively updated: (44) In addition, a drift detection mechanism can be set up to actively shorten the time decay constant when a significant difference is detected between the distribution of recent inconsistent scores and the historical reference distribution. And reduce the target coverage rate This temporarily expands the prediction tube to maintain safe coverage. For sparsely grouped samples, hierarchical quantile shrinkage can be used to smooth the quantiles towards the quantiles of the parent group.

[0058] Specifically, the drift detection distance is defined as follows: (45) when This shortens the time decay constant and reduces the target coverage rate: (46) This temporarily expands the prediction tube to maintain safe coverage. For sparsely grouped samples, hierarchical quantile shrinkage can be used to smooth the quantiles towards the quantiles of the parent group. (47) (48) In another embodiment, to handle the non-Gaussian, multimodal, skewed, or abrupt characteristics that the residual distribution may exhibit, this method employs a Conditional Normalizing Flow to map the high-dimensional residuals to a latent variable space. Specifically, an invertible transformation is defined: (49) In formula (49), For the residual vector, For conditional context (such as knowledge graph embeddings and temporal features), this transformation maps the original residual distribution to the latent variable space of a standard normal distribution (or some simple distribution). In this case, the latent variable manifold inconsistency scores can be represented using the L2 norm: (50) Alternatively, negative log-likelihood can be used: (51) The corresponding conformal prediction set is represented as: (52) It is important to emphasize that the normalized flow model is only used to improve the geometric representation of inconsistent fractions; its coverage guarantee is still provided by the conformal quantiles on the calibration set, and therefore it also has statistical reliability. Those skilled in the art can choose an appropriate fraction construction method (joint Mahalanobis distance, local fractions, manifold fractions, or combinations thereof) based on the actual data characteristics.

[0059] Through the above step S130, the multidimensional conformal prediction tube generated by this method can cover the future time-frequency-space three-dimensional spectrum state tensor. It not only reflects the overall joint prediction error, but also penalizes local extreme deviations and adapts to environmental drift online, thus providing a solid uncertainty quantification basis for subsequent risk-controllable dynamic spectrum access decisions.

[0060] In step S140, based on the conservative upper bound of the spectrum state provided by the multidimensional conformal prediction tube generated in step S130, and combined with spectrum regulatory constraints and interference protection thresholds, a set of safe actions is selected. The conservatism of this set is then adjusted using risk calibration parameters. Finally, dynamic spectrum access actions that satisfy risk constraints and maximize communication utility are selected from the risk-calibrated set of safe actions. The following details this step in terms of action representation, regulatory filtering, interference assessment, safe set construction, risk calibration parameter selection, and utility maximization.

[0061] In this embodiment, dynamic spectrum access action is defined as a combination of frequency band set, transmit power, access duration, beam configuration, and secondary user location, expressed as: (53) In formula (53), The set of frequency bands to be accessed (which may contain one or more frequency channels). For transmission power, For the expected access duration, Configure beam parameters (such as direction and beamwidth). The expected spatial location for secondary users.

[0062] However, not all actions are permitted under regulation. This method is first based on a regulatory-physical knowledge graph. The feasibility of candidate actions is assessed to form a set of regulatoryly feasible actions, represented as follows: (54) In formula (54), For regulatory feasibility functions, when the action Returns 1 if all applicable regulatory rules (such as licensed frequency band usage rights, transmit power limits, no-transmission regulations for specific areas, time window constraints, etc.) are met; otherwise, returns 0.

[0063] To protect the main user from harmful interference under the premise of prediction uncertainty, this method extracts a conservative upper bound of the spectral state from the multidimensional conformal prediction tube.

[0064] Specifically, regarding future moments ,frequency Spatial location Its conservative upper bound is defined as the sum of the quantile upper bound and the conformal quantile correction term: (55) In formula (55), The upper quantile of the spectrum prediction network output. For local scale parameters, For the corresponding context grouping obtained in step S130 The conformal quantile threshold. This conservative upper bound is based on... The confidence level covers the true spectral power, meaning the probability that the true value exceeds this upper bound is no higher than [a certain percentage]. .

[0065] Furthermore, based on a conservative upper bound, a conservative occupancy determination is constructed: (56) In formula (56), For action The future time-frequency-space region covered For frequency ,Location The tolerance threshold (or protection threshold) for the primary user. To calibrate parameters with risk Changing safety margin ( The larger, The smaller the size, the tighter the safety set.

[0066] For the protected primary user receiver Secondary user actions The generated interference power can be calculated as follows: (57) In formula (57), For action In frequency Linear transmit power on For the secondary user to primary user receiver in frequency Linear antenna gain on This represents linear path loss.

[0067] Specifically, the primary user collision loss is defined as: if the action A collision occurs if the actual spectral power at any covered time-frequency-spatial location exceeds the protection threshold. (58) Specifically, harmful interference loss takes into account the degree of interference exceeding the threshold at the time of the collision: (59) In formula (59), For the action The set of primary user receivers affected. This represents the maximum tolerable interference power. The regulatory violation loss is directly given by the regulatory feasibility function: (60) The overall risk loss is the weighted sum of losses from user collisions, harmful interference, and regulatory violations: (61) In formula (61), the weights The settings can be configured according to different emphases on main user protection, interference tolerance, and regulatory compliance based on actual business needs.

[0068] After obtaining the conservative upper bound, a set of safe actions is constructed in this method. An action is considered "safe" if it simultaneously meets two conditions: (1) at any covered time-frequency-space location, the conservative upper bound does not exceed the protection threshold minus the safety margin; (2) for all affected primary user receivers, the interference power generated by secondary users does not exceed their maximum tolerable interference minus the corresponding safety margin. This can be expressed as: (62) Furthermore, formula (62) also satisfies the interference safety condition: (63) In formulas (62) and (63), As a risk calibration parameter, it adjusts the safety margin. and To control the conservatism of the set of safety actions: The larger the value, the smaller the safety margin and the narrower the safety set (higher risk but potentially higher efficiency). The smaller the value, the larger the safety margin and the wider the safety set (more conservative, but fewer available actions).

[0069] To achieve the optimal balance between security and spectrum utilization, a risk calibration parameter is also introduced in this method. And adaptively select the most suitable risk calibration sample set. value.

[0070] In this embodiment, adjusting the conservatism of the set of security actions through risk calibration parameters includes: obtaining a risk calibration sample set, which contains historical prediction results, historical context information, and the actual spectrum state observed after historical decisions; for each of the preset multiple candidate risk calibration parameters, using the risk calibration sample set to evaluate the comprehensive risk generated by performing dynamic spectrum access actions when using a certain candidate risk calibration parameter, obtaining the corresponding empirical risk; calculating the upper confidence bound of the empirical risk corresponding to each candidate parameter; and selecting the smallest candidate parameter that satisfies the upper confidence bound not exceeding the preset target risk level as the risk calibration parameter.

[0071] Specifically, for each of the preset candidate risk calibration parameters, the comprehensive risk arising from performing dynamic spectrum access actions when using that candidate parameter is evaluated using the risk calibration sample set, and the corresponding empirical risk is obtained: (64) In formula (64), In the first On each risk calibration sample, parameters are used. The action chosen is based on the subsequent utility maximization strategy.

[0072] Specifically, calculate the upper confidence bound of the empirical risk corresponding to each candidate parameter (e.g., based on the Hoeffding inequality): (65) In formula (65), The confidence level parameter is selected. The upper confidence bound is chosen to satisfy the condition that the risk level does not exceed a preset target risk level. The smallest candidate parameter, used as the risk calibration parameter, is expressed as: (66) Selected by the above method It can guarantee with a high probability that the actual risk does not exceed It is neither overly conservative nor excessively conservative, thereby maximizing available spectrum opportunities.

[0073] Furthermore, after determining the risk calibration parameters Afterwards, a set of safety actions. This is also determined. The system selects the action that maximizes communication efficiency from this set: (67) Communication utility functions can comprehensively consider throughput, power consumption, and perception overhead. An example is as follows: (68) In formula (68), For the action in the predicted time step The actual duration of use For access bandwidth, For the linear gain from the secondary user to its receiver, For the predicted interference power, For noise power, The power cost factor, For the perceived overhead coefficient, To perform the action The required sensing costs (such as the number of channel detections, sensor power consumption, etc.).

[0074] In extreme cases, if the set of safety actions If the value is empty, meaning no action can simultaneously satisfy the conservative upper bound constraint, the interference protection threshold, and the regulatory constraint, the system will execute a fallback strategy. Fallback strategies may include: reducing transmit power, shortening access duration, switching to a more conservative licensed frequency band, increasing the number of sensing rounds to obtain a more accurate interference estimate, or temporarily abandoning access and waiting for the next decision cycle.

[0075] Through the above step S140, this method quantifies the uncertainty information provided by the multidimensional conformal prediction tube into an engineering-operable safety boundary, and uses risk calibration parameters to achieve an adaptive optimal balance between protecting the primary user and maximizing spectrum utilization, ultimately outputting a safe and efficient dynamic spectrum access action.

[0076] In this embodiment, the above method is divided into three stages: offline training, offline calibration, and online deployment. This reasonable division into three stages ensures both sufficient optimization of the model and parameters, and enables rapid response to new samples and environmental adaptability.

[0077] Specifically, the first phase involves offline training. This includes constructing a regulatory-physical knowledge graph, estimating the activation probabilities of candidate emission sources, generating a physical prior field based on a path loss model, and training a spectrum prediction network guided by the physical knowledge graph. The training, calibration, and test sets should be divided chronologically to avoid future information leakage.

[0078] Specifically, the second stage is offline calibration. A fixed prediction network generates prediction results on the calibration set, calculates joint inconsistency scores, local inconsistency scores, or latent variable manifold scores, and estimates conformal quantiles for different context groups. Simultaneously, the access risk under different risk calibration parameters is assessed using the risk calibration set, and parameters that meet the target risk constraints are selected.

[0079] Specifically, the third stage involves online deployment. At each moment, the system receives the latest spectrum observations, updates the historical window, knowledge graph, and contextual features; generates a physical prior field and outputs future predictions; reads or updates conformal quantiles and constructs a prediction set; based on the prediction set, it constructs a set of security actions and selects access actions. Once the actual future spectrum is observed, the system calculates the actual inconsistency score and decision loss, and writes them to the online calibration buffer, thus forming a continuous self-calibration closed loop.

[0080] Through the three stages described above, this method achieves a complete closed loop from offline learning to online deployment, ensuring both global optimization of the model and parameters and real-time adaptive capability to environmental changes. It can continuously provide reliable and secure decision support in dynamic spectrum access scenarios.

[0081] like Figure 2 As shown, a schematic diagram of the specific framework of this method is also provided. From Figure 2 As can be seen, the entire methodology is divided into six levels from top to bottom: "historical data monitoring," "regulatory-physical prior modeling," "spectrum transformer guided by physical knowledge graph," "adaptive grouping calibration," "conformal risk control dynamic spectrum access," and "output and decision-making." Specifically, historical spectrum observation data, along with the knowledge graph and contextual information, are input into the regulatory-physical prior modeling module. Through entity relationship extraction, heterogeneous knowledge graph construction, and differentiable propagation prior generation, the physical prior field is obtained. Subsequently, the physical prior field, historical spectrum data, and knowledge graph are input into the spectrum Transformer, and the output point prediction is generated using a physical / regulatory bias attention mechanism. quantiles Local scale and residual representation During the adaptive group calibration phase, the system is grouped by region, frequency band, service, and time period, and drift sensing updates are performed to calculate the group-specific calibration status. , , , Furthermore, a multidimensional prediction pipeline is constructed through joint / local scores. and conservative power upper limit Finally, the conformal risk control module performs regulatory filtering, primary user collision risk control, and harmful interference constraints on candidate actions, and outputs the optimal secure access action. It also provides spectral point predictions and multidimensional prediction sets. Figure 2 It fully demonstrates the end-to-end process from data acquisition to final decision output, corresponding one-to-one with the aforementioned steps S100 to S140.

[0082] The aforementioned prior fusion risk calibration spectrum prediction method reduces the reliance of pure data-driven models on large-scale labeled samples by fusing regulatory-physical knowledge graphs to license, service, frequency band, geographic, and propagation rules. It constructs a physical prior field using the superposition of potential transmitter activation probabilities and linear power domains, ensuring prediction results conform to basic propagation laws and reducing distortion during cross-regional and cross-frequency band migrations. The method introduces physical bias and regulatory bias attention mechanisms into the spectrum prediction network, enabling it to simultaneously utilize statistical correlation, propagation correlation, and regulatory structure correlation. By using a multidimensional conformal prediction tube instead of point-by-point confidence intervals, it can cover the future time-frequency-space joint spectrum tensor, making it more suitable for safety-critical dynamic spectrum access decisions. Through grouped adaptive calibration and drift detection mechanisms, it can dynamically adjust the prediction set size under non-stationary scenarios such as weekdays / weekends, day / night cycles, sudden events, and mobile interference sources. Finally, through risk calibration parameters, the prediction uncertainty is mapped to a set of safety actions, maximizing communication utility while meeting primary user protection and regulatory constraints.

[0083] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0084] In one embodiment, such as Figure 3 As shown, a priori fusion risk calibration spectrum prediction device is provided, comprising: a data acquisition module 200, a priori generation module 210, a spectrum prediction module 220, a conformal calibration module 230, and a risk decision module 240, wherein: The data acquisition module 200 is used to acquire historical spectrum observation data, characterize spectrum regulatory rules, regulatory-physical knowledge graphs of physical space and emission source association structures, and environmental context information describing the spatiotemporal and operational environment in which the prediction task is located; The prior generation module 210 is used to estimate the time-frequency activation probability of candidate emission sources based on the regulatory-physical knowledge graph and the environmental context information, and generate a physical prior spectrum field based on the activation probability and the propagation model. The spectrum prediction module 220 is used to input the historical spectrum observation data, the regulatory-physical knowledge graph and the physical prior spectrum field into the spectrum prediction network to obtain the prediction result of the future three-dimensional spectrum state tensor, wherein the future three-dimensional spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels and multiple spatial locations in the future. The conformal calibration module 230 is used to construct non-consistency scores using the prediction residuals on the calibration sample set, calculate conformal quantiles on the non-consistency scores according to the preset target coverage rate, and generate a multidimensional conformal prediction tube covering the future three-dimensional spectral state tensor. The risk decision module 240 is used to select a set of safe actions based on the upper conservative bound of the spectrum state, spectrum regulatory constraints and interference protection threshold provided by the multidimensional conformal prediction tube, adjust the conservatism of the set of safe actions through risk calibration parameters, and select dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility from the set of safe actions after risk calibration.

[0085] Specific limitations regarding the prior fusion risk calibration spectrum prediction device can be found in the limitations of the prior fusion risk calibration spectrum prediction method described above, and will not be repeated here. Each module in the aforementioned prior fusion risk calibration spectrum prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0086] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a priori fusion risk calibration spectrum prediction method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0087] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0088] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: Acquire historical spectrum observation data, characterize spectrum regulatory rules, regulatory-physical knowledge graphs of physical space and emission source association structures, and environmental context information describing the spatiotemporal and operational environment in which the prediction task is located; Based on the regulatory-physical knowledge graph and the environmental context information, the time-frequency activation probability of the candidate emission source is estimated, and a physical prior spectrum field is generated based on the activation probability and the propagation model. The historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field are input into the spectrum prediction network to obtain the prediction result of the future three-dimensional spectrum state tensor. The future three-dimensional spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels and multiple spatial locations in the future. A non-consistency score is constructed using the prediction residuals on the calibration sample set. Conformal quantiles are calculated on the non-consistency score based on a preset target coverage rate, and a multidimensional conformal prediction tube covering the future three-dimensional spectral state tensor is generated. A set of safe actions is selected based on the conservative upper bound of the spectrum state, spectrum regulatory constraints, and interference protection thresholds provided by the multidimensional conformal prediction tube. The conservatism of the set of safe actions is adjusted by risk calibration parameters, and dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility are selected from the set of safe actions after risk calibration.

[0089] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire historical spectrum observation data, characterize spectrum regulatory rules, regulatory-physical knowledge graphs of physical space and emission source association structures, and environmental context information describing the spatiotemporal and operational environment in which the prediction task is located; Based on the regulatory-physical knowledge graph and the environmental context information, the time-frequency activation probability of the candidate emission source is estimated, and a physical prior spectrum field is generated based on the activation probability and the propagation model. The historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field are input into the spectrum prediction network to obtain the prediction result of the future three-dimensional spectrum state tensor. The future three-dimensional spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels and multiple spatial locations in the future. A non-consistency score is constructed using the prediction residuals on the calibration sample set. Conformal quantiles are calculated on the non-consistency score based on a preset target coverage rate, and a multidimensional conformal prediction tube covering the future three-dimensional spectral state tensor is generated. A set of safe actions is selected based on the conservative upper bound of the spectrum state, spectrum regulatory constraints, and interference protection thresholds provided by the multidimensional conformal prediction tube. The conservatism of the set of safe actions is adjusted by risk calibration parameters, and dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility are selected from the set of safe actions after risk calibration.

[0090] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0091] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A priori fusion risk calibration spectrum prediction method, characterized in that, The method includes: Acquire historical spectrum observation data, characterize spectrum regulatory rules, regulatory-physical knowledge graphs of physical space and emission source association structures, and environmental context information describing the spatiotemporal and operational environment in which the prediction task is located; Based on the regulatory-physical knowledge graph and the environmental context information, the time-frequency activation probability of the candidate emission source is estimated, and a physical prior spectrum field is generated based on the activation probability and the propagation model. The historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field are input into the spectrum prediction network to obtain the prediction result of the future three-dimensional spectrum state tensor. The future three-dimensional spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels and multiple spatial locations in the future. A non-consistency score is constructed using the prediction residuals on the calibration sample set. Conformal quantiles are calculated on the non-consistency score based on a preset target coverage rate, and a multidimensional conformal prediction tube covering the future three-dimensional spectral state tensor is generated. A set of safe actions is selected based on the conservative upper bound of the spectrum state, spectrum regulatory constraints, and interference protection thresholds provided by the multidimensional conformal prediction tube. The conservatism of the set of safe actions is adjusted by risk calibration parameters, and dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility are selected from the set of safe actions after risk calibration.

2. The prior fusion risk calibration spectrum prediction method according to claim 1, characterized in that, The regulatory-physical knowledge graph includes candidate emission source nodes, frequency nodes, spatial region nodes, service type nodes, regulatory rule nodes, and time event nodes; The edge relationships in the regulatory-physical knowledge graph include: the authorized use relationship between candidate transmitter nodes and frequency nodes, the location affiliation and coverage relationship between candidate transmitter nodes and spatial region nodes, the adjacent or interference relationship between different frequency nodes, the protection constraint relationship between regulatory rule nodes and frequency nodes or spatial region nodes, and the time scheduling relationship between time event nodes and the activation state of candidate transmitter nodes.

3. The prior fusion risk calibration spectrum prediction method according to claim 2, characterized in that, Based on the regulatory-physical knowledge graph and the environmental context information, the time-frequency activation probability of candidate emission sources is estimated, including: The target output dimension of the prediction task includes multiple future prediction times and multiple prediction frequencies; Based on the node attributes in the regulatory-physical knowledge graph, generate the initial embedding of each node; The initial embeddings of the candidate emission source nodes are extracted from the regulatory-physical knowledge graph as attribute embeddings, the initial embeddings of the frequency nodes are extracted as frequency embeddings, the initial embeddings of the service type nodes are extracted as service embeddings, and the regulatory authorization embeddings are extracted. Extract the time feature code and event context feature corresponding to the predicted time from the environmental context information; The extracted embeddings and features are concatenated into vectors, and the activation score is obtained after linear transformation. The activation score is input into the sigmoid activation function to obtain the activation probability of each candidate emission source at each prediction time and each prediction frequency.

4. The prior fusion risk calibration spectrum prediction method according to claim 3, characterized in that, Generate a physical prior spectrum field based on the activation probability and propagation model, including: Based on the propagation model and incorporating a learnable correction term, calculate the path loss from the candidate emission source to the spatial location at the predicted frequency; The linear received power of the candidate transmitter at the spatial location is calculated based on the transmit power of the candidate transmitter, the antenna gain, the path loss, and the activation probability of the candidate transmitter at the prediction time and prediction frequency. The linear received power of all candidate transmitters at the same prediction time, the same prediction frequency, and the same spatial location is linearly superimposed with the noise power to obtain the bus linear received power. The bus-type received power is converted to the dB domain to obtain the value of the physical prior spectrum field at the prediction time, prediction frequency, and spatial location.

5. The prior fusion risk calibration spectrum prediction method according to claim 1, characterized in that, The spectrum prediction network employs an attention-based neural network, in which physical propagation bias, regulatory relationship bias, and geographic proximity bias are introduced into the attention logit of the attention mechanism. The spectrum prediction network takes the historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field as input, and uses physical prior plus residual learning to output the point prediction results of the future three-dimensional spectrum state tensor.

6. The prior fusion risk calibration spectrum prediction method according to claim 1, characterized in that, Constructing inconsistent scores using prediction residuals on a calibration sample set includes: The prediction residual of each sample in the calibration sample set is vectorized to obtain a high-dimensional residual vector; The high-dimensional residual vector is mapped to a low-dimensional space by an encoder to obtain a low-dimensional residual representation; Based on the context group to which the calibration sample belongs, the mean and covariance matrix of the low-dimensional residual representation within the group are calculated, wherein the context group is obtained by dividing the calibration sample according to the region type, frequency band type, service type and / or time period; For each calibration sample, a joint inconsistency score is constructed based on the Mahalanobis distance between the corresponding low-dimensional residual representation and the mean of its group, wherein the joint inconsistency score is used to measure the overall prediction error of the future three-dimensional spectral state tensor.

7. The prior fusion risk calibration spectrum prediction method according to claim 1, characterized in that, The adjustment of the conservatism of the set of safety actions through risk calibration parameters includes: Obtain a risk calibration sample set, which includes historical prediction results, historical context information, and the actual spectrum state observed after historical decisions; For each of the preset multiple candidate risk calibration parameters, the comprehensive risk generated by performing dynamic spectrum access actions when using a certain candidate risk calibration parameter is evaluated using the risk calibration sample set to obtain the corresponding empirical risk; Calculate the upper confidence bound of the empirical risk corresponding to each candidate parameter; The smallest candidate parameter that satisfies the upper confidence bound not exceeding the preset target risk level is selected as the risk calibration parameter.

8. A priori fusion risk calibration spectrum prediction device, characterized in that, The device includes: The data acquisition module is used to acquire historical spectrum observation data, characterize spectrum regulatory rules, regulatory-physical knowledge graphs of physical space and emission source association structures, and environmental context information describing the spatiotemporal and operational environment in which the prediction task is located; The prior generation module is used to estimate the time-frequency activation probability of candidate emission sources based on the regulatory-physical knowledge graph and the environmental context information, and generate a physical prior spectrum field based on the activation probability and the propagation model. The spectrum prediction module is used to input the historical spectrum observation data, regulatory-physical knowledge graph and physical prior spectrum field into the spectrum prediction network to obtain the prediction result of the future three-dimensional spectrum state tensor. The future three-dimensional spectrum state tensor contains spectrum power state information at multiple time steps, multiple frequency channels and multiple spatial locations in the future. The conformal calibration module is used to construct non-consistency scores using the prediction residuals on the calibration sample set, calculate conformal quantiles on the non-consistency scores according to the preset target coverage rate, and generate a multidimensional conformal prediction tube covering the future three-dimensional spectral state tensor. The risk decision module is used to select a set of safe actions based on the upper conservative bound of the spectrum state, spectrum regulatory constraints and interference protection thresholds provided by the multi-dimensional conformal prediction tube, adjust the conservatism of the set of safe actions through risk calibration parameters, and select dynamic spectrum access actions that meet the risk constraints and have the greatest communication utility from the set of safe actions after risk calibration.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images