Beidou short message adaptive power control method and system based on AI

Abstract

The application relates to the technical field of wireless communication, in particular to a Beidou short message adaptive power control method and system based on AI. The method generates a channel feature vector by collecting channel parameters, carries out competitive mapping in a self-organizing topology space to obtain an activated node identifier and a node activation degree, iteratively generates a basic power value through an adversarial generation network based on a seed power value, constructs a directed acyclic causal graph, separates causal effect quantities to carry out counterfactual reasoning compensation to obtain a target power value, adjusts transmission power according to the target power value and sends a message, calculates a power loss difference when transmission fails, and updates a Bayesian posterior probability through kernel density estimation in combination with the channel feature vector, selects a maximum posterior probability node to reconstruct the target power value and resend the message. The application realizes adaptive optimization and adjustment of power, improves transmission reliability and reduces energy consumption.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to an AI-based adaptive power control method and system for BeiDou short messages. Background Technology

[0002] In existing technologies, power control in BeiDou short message terminals mostly employs fixed thresholds or linear adjustment methods based on signal strength indications. The conventional approach is to pre-set a set of power levels, and the terminal switches between these levels based on the received signal strength or signal-to-noise ratio estimate. Some solutions introduce lookup table methods, obtaining the mapping relationship between channel state and power values ​​through offline training, and the terminal directly determines the transmit power based on real-time measured channel parameters. These methods rely on empirical formulas or statistical models, and their adaptability to dynamically changing environments is limited.

[0003] Conventional power adjustment methods have long feedback cycles, typically requiring multiple send-receive acknowledgments before adjusting the power value. This can easily lead to lag when the channel changes rapidly, resulting in wasted power or transmission failures. For example, in scenarios with sudden interference or rapid terminal movement, linear adjustment cannot respond promptly to drastic fluctuations in channel parameters, causing packet loss or redundant transmit power. The lack of a reverse verification mechanism for power adjustment results means that in the event of a transmission failure, simply increasing the power level for retrying often fails to effectively separate the combined effects of channel fading and external interference on transmission quality, leading to inefficient retry strategies and even power oscillations. This adjustment method, based on experience or linear assumptions, struggles to balance power efficiency and transmission reliability in complex electromagnetic environments. Summary of the Invention

[0004] This invention provides an AI-based adaptive power control method and system for BeiDou short messages, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an AI-based adaptive power control method for BeiDou short messages, comprising:

[0006] The channel parameters of the Beidou short message terminal are collected to generate a channel feature vector. The channel feature vector is then competitively mapped in the self-organizing topology space, and the active node identifier and node activation degree are output.

[0007] The seed power value is extracted based on the activation node identifier. A generative adversarial network is constructed to iteratively generate candidate power values ​​based on the seed power value and node activation degree until the transmission reliability score output by the discriminator reaches the preset reliability threshold. The optimized candidate power value is then output as the base power value.

[0008] A directed acyclic causal graph is constructed from the channel feature vector. The causal effect quantities of the channel fading factor and the interference factor are separated by intervention operation. The target power value is obtained by counterfactual reasoning compensation based on the causal effect quantities of the base power value.

[0009] Adjust the transmit power according to the target power value and send the message to obtain the transmission status identifier and the actual received power;

[0010] When the transmission status is marked as failed, the power loss difference between the target power value and the actual received power is calculated. The power loss difference is combined with the channel feature vector as observation evidence. The Bayesian posterior probability of the neighboring nodes of the active node is updated by kernel density estimation. The node with the largest posterior probability is selected as the reconstruction node identifier. The target power value is regenerated based on the reconstruction node identifier and the message is sent again.

[0011] The channel parameters of the BeiDou short message terminal are collected to generate a channel feature vector. This channel feature vector is then competitively mapped in a self-organizing topology space, outputting the active node identifier and node activation degree, including:

[0012] The received signal strength, signal-to-noise ratio, and bit error rate of the BeiDou short message terminal within a continuous time window are collected as channel parameters. The channel parameters are sampled by a time-series sliding window. The sampling results are used to construct a two-dimensional matrix according to the time dimension and the parameter dimension and then normalized to generate a channel feature vector.

[0013] Calculate the Euclidean distance between the channel feature vector and the weight vectors of each node in the self-organized topology space, identify the node with the smallest Euclidean distance, and extract the node's spatial location index as the identifier of the active node.

[0014] Based on the activation node identifier, locate the corresponding node weight vector, and calculate the cosine similarity between the channel feature vector and the located node weight vector as the initial activation degree.

[0015] Extract the received signal strength component from the channel feature vector, calculate the channel quality attenuation factor by normalizing the ratio of the received signal strength component to the preset signal strength reference value, and then weight and correct the initial activation degree according to the channel quality attenuation factor to obtain the node activation degree. Output the active node identifier and node activation degree.

[0016] The seed power value is extracted based on the activated node identifier. A generative adversarial network is constructed to iteratively generate candidate power values ​​based on the seed power value and node activation degree until the transmission reliability score output by the discriminator reaches a preset reliability threshold. The optimized candidate power values ​​are then output as the base power values, including:

[0017] Based on the activation node identifier, the corresponding initial power value is retrieved from the preset power mapping table as the seed power value;

[0018] Construct an adversarial generative network, which includes a generator and a discriminator. The generator maps node activations to power adjustment vectors and incrementally adjusts the seed power value based on the power adjustment vectors to generate candidate power values.

[0019] The discriminator receives candidate power values ​​and constructs a virtual transmission scenario. In the virtual transmission scenario, it simulates the signal propagation process of the candidate power values ​​and calculates the received power of the signal reaching the Beidou satellite. Based on the received power and the preset Beidou satellite receiving sensitivity threshold, it calculates the power margin and generates a transmission reliability score.

[0020] When the transmission reliability score does not reach the preset reliability threshold, the power compensation amount is calculated based on the deviation between the received power and the Beidou satellite receiving sensitivity threshold.

[0021] The power adjustment vector is adjusted according to the power compensation amount, and the seed power value is incrementally adjusted to generate a new candidate power value. This process is repeated until the transmission reliability score reaches the preset reliability threshold. The candidate power value that reaches the preset reliability threshold is then output as the base power value.

[0022] The discriminator receives candidate power values ​​and constructs a virtual transmission scenario. Within this scenario, it simulates the signal propagation process of the candidate power values ​​and calculates the received power of the signal reaching the BeiDou satellite. Based on the received power and a preset BeiDou satellite receiver sensitivity threshold, it calculates the power margin and generates a transmission reliability score, including:

[0023] The discriminator receives candidate power values, extracts the received signal strength component and signal-to-noise ratio component from the channel feature vector, and calculates the multipath propagation delay spread parameter and Doppler frequency shift parameter as multipath fading parameters based on the fading amplitude of the received signal strength component and the noise intensity of the signal-to-noise ratio component.

[0024] The time-domain fading distribution in the virtual transmission scenario is determined based on the multipath propagation delay spread parameter, and the frequency-domain fading distribution in the virtual transmission scenario is determined based on the Doppler frequency shift parameter. A virtual transmission scenario with time-frequency fading characteristics is constructed, and the candidate power value is used as the transmit power to start the signal propagation process simulation in the virtual transmission scenario.

[0025] Based on the spatial distance between the BeiDou short message terminal and the BeiDou satellite, free space path attenuation is applied to the candidate power value. Based on the time domain fading distribution and frequency domain fading distribution, multipath fading attenuation is applied to the power after free space path attenuation to obtain the received power of the signal reaching the BeiDou satellite.

[0026] Obtain the preset BeiDou satellite receiving sensitivity threshold, calculate the power margin between the received power and the preset BeiDou satellite receiving sensitivity threshold, and generate a transmission reliability score by weighting the power margin and node activation degree according to the preset weight coefficient.

[0027] A directed acyclic causal graph is constructed from the channel feature vectors. The causal effects of the channel fading factor and interference factor are separated through intervention operations. Based on these causal effects, the target power value is obtained by counterfactual inference compensation of the baseline power value, including:

[0028] The received signal strength component, signal-to-noise ratio component, Doppler frequency shift component, and time delay spread component are extracted from the channel feature vector. The received signal strength component and signal-to-noise ratio component are used as observation variable nodes, and the Doppler frequency shift component and time delay spread component are used as factor variable nodes. A directed edge connection is established based on the influence relationship between the observation variable nodes and the factor variable nodes to construct a directed acyclic causal graph.

[0029] The value of the Doppler frequency shift component is fixed and the change amplitude of the conditional probability distribution of the observed variable node is calculated as the causal effect of the channel fading factor. The value of the delay spread component is fixed and the change amplitude of the conditional probability distribution of the observed variable node is calculated as the causal effect of the interference factor.

[0030] The causal effect of the channel fading factor is set to the maximum fading level, and the fading counterfactual value of the received signal strength component is derived in reverse. The causal effect of the interference factor is set to the maximum interference level, and the interference counterfactual value of the signal-to-noise ratio is derived in reverse.

[0031] The fading compensation amount is obtained by calculating the difference between the fading counterfactual value and the actual value of the received signal strength component. The interference compensation amount is obtained by calculating the difference between the interference counterfactual value and the actual value of the signal-to-noise ratio component. The fading compensation amount and the interference compensation amount are added to the base power value to obtain the target power value.

[0032] The causal effect of fixing the value of the Doppler frequency shift component and calculating the change in the conditional probability distribution of the observed variable nodes as the causal effect of the channel fading factor, and fixing the value of the delay spread component and calculating the change in the conditional probability distribution of the observed variable nodes as the causal effect of the interference factor, include:

[0033] Construct a joint probability distribution of Doppler frequency shift components and time delay spread components in historical transmission records, and sample from the joint probability distribution to generate an intervention sample set containing multiple sets of numerical combinations;

[0034] The Doppler frequency shift component values ​​in the intervention sample set are sequentially fixed to the factor variable nodes corresponding to the Doppler frequency shift components in the directed acyclic causal graph, and the input edges of the factor variable nodes are cut off. The mean of the conditional probability distribution of the observation variable nodes corresponding to the received signal strength components after each fixation is calculated, and the standard deviation of the mean of all conditional probability distributions is used as the causal effect of the channel fading factor.

[0035] The time delay spread component values ​​in the intervention sample set are sequentially fixed to the factor variable nodes corresponding to the time delay spread components in the directed acyclic causal graph, and the input edges of the factor variable nodes are cut off. The mean of the conditional probability distribution of the observation variable nodes corresponding to the signal-to-noise ratio component after each fixation is calculated, and the standard deviation of the mean of all conditional probability distributions is used as the causal effect of the interference factor.

[0036] The power loss difference and channel feature vector are combined as observational evidence. The Bayesian posterior probability of the active node's neighboring nodes is updated using kernel density estimation. The node with the highest posterior probability is selected as the reconstructed node identifier. The target power value is regenerated based on the reconstructed node identifier, and the message is retransmitted, including:

[0037] The power loss difference is combined with the channel feature vector to form an observation evidence vector that contains both power dimension information and channel dimension information.

[0038] In a directed acyclic causal graph, all adjacent connected nodes of the node corresponding to the activated node are extracted as a neighborhood node set. For each node in the neighborhood node set, the observed evidence vector is used as the sample input to construct a kernel density estimation function. The probability density value of the kernel density estimation function at the node position is calculated as the likelihood probability. The likelihood probability is multiplied by the prior probability of the node according to the Bayes formula and then normalized to obtain the Bayes posterior probability of the node.

[0039] Traverse the Bayesian posterior probabilities of all nodes in the neighborhood node set and select the node with the largest value as the reconstruction node identifier. Extract the causal effect of the channel fading factor and the causal effect of the interference factor corresponding to the reconstruction node identifier. Based on the causal effect of the channel fading factor and the causal effect of the interference factor, perform counterfactual reasoning compensation on the base power value to regenerate the target power value. Use the regenerated target power value to send the BeiDou short message again.

[0040] A second aspect of this invention provides an AI-based adaptive power control system for BeiDou short messages, comprising:

[0041] The channel feature unit is used to collect the channel parameters of the Beidou short message terminal, generate the channel feature vector, perform competitive mapping of the channel feature vector in the self-organizing topology space, and output the active node identifier and node activation degree.

[0042] The power optimization unit is used to extract seed power values ​​based on the active node identifier, and to generate candidate power values ​​iteratively based on the seed power values ​​and node activation by constructing an adversarial generative network until the transmission reliability score output by the discriminator reaches a preset reliability threshold. The optimized candidate power values ​​are then output as the base power values.

[0043] The causal effect unit is used to construct a directed acyclic causal graph from the channel feature vector. It separates the causal effect quantities of the channel fading factor and the interference factor through intervention operations, and obtains the target power value by counterfactual reasoning compensation based on the causal effect quantities of the base power value.

[0044] The power adjustment unit is used to adjust the transmission power according to the target power value and send messages, and to obtain the transmission status identifier and the actual received power.

[0045] The failure retransmission unit is used to calculate the power loss difference between the target power value and the actual received power when the transmission status is identified as failure. The power loss difference is combined with the channel feature vector as observation evidence. The Bayesian posterior probability of the neighboring nodes of the active node is updated by kernel density estimation. The node with the largest posterior probability is selected as the reconstruction node identifier. The target power value is regenerated based on the reconstruction node identifier and the message is sent again.

[0046] A third aspect of the present invention provides an electronic device, comprising:

[0047] processor;

[0048] Memory used to store processor-executable instructions;

[0049] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0050] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0051] This invention, through competitive mapping of self-organizing topology space, can accurately extract seed power values ​​corresponding to channel features. Combined with iterative optimization of generative adversarial networks, it can quickly obtain basic power values ​​that meet transmission reliability requirements, significantly reducing manual parameter tuning steps and improving the automation level of power configuration. Based on counterfactual reasoning compensation using directed acyclic causal graphs, the causal effects of channel fading and interference are separated, allowing the target power value to actively offset environmental interference and maintain stable communication quality even under complex channel conditions. The rapid recovery mechanism after transmission failure effectively shortens retransmission response time. By calculating the power loss difference and performing kernel density estimation in conjunction with channel feature vectors, the Bayesian posterior probabilities of neighboring nodes are updated, accurately identifying key nodes causing failure and reconstructing the target power value. This avoids power waste caused by blind retransmission, improving retransmission success rate and resource utilization. Deep integration of power control and causal reasoning enables adaptive adjustment in dynamic channel environments, significantly reducing message loss rate. Simultaneously, the dual optimization based on topology competition and generative adversarial networks reduces the number of power adjustment iterations, lowering terminal energy consumption. This is particularly suitable for resource-constrained BeiDou short message terminals, improving the overall robustness and energy efficiency of the system. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the AI-based adaptive power control method for BeiDou short messages according to an embodiment of the present invention.

[0053] Figure 2 This is a flowchart illustrating the operation of the channel power compensation algorithm based on causal inference in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0056] Figure 1 This is a flowchart illustrating the AI-based adaptive power control method for BeiDou short messages according to an embodiment of the present invention.

[0057] AI-based adaptive power control methods for BeiDou short messages include:

[0058] The channel parameters of the Beidou short message terminal are collected to generate a channel feature vector. The channel feature vector is then competitively mapped in the self-organizing topology space, and the active node identifier and node activation degree are output.

[0059] The seed power value is extracted based on the activation node identifier. A generative adversarial network is constructed to iteratively generate candidate power values ​​based on the seed power value and node activation degree until the transmission reliability score output by the discriminator reaches the preset reliability threshold. The optimized candidate power value is then output as the base power value.

[0060] A directed acyclic causal graph is constructed from the channel feature vector. The causal effect quantities of the channel fading factor and the interference factor are separated by intervention operation. The target power value is obtained by counterfactual reasoning compensation based on the causal effect quantities of the base power value.

[0061] Adjust the transmit power according to the target power value and send the message to obtain the transmission status identifier and the actual received power;

[0062] When the transmission status is marked as failed, the power loss difference between the target power value and the actual received power is calculated. The power loss difference is combined with the channel feature vector as observation evidence. The Bayesian posterior probability of the neighboring nodes of the active node is updated by kernel density estimation. The node with the largest posterior probability is selected as the reconstruction node identifier. The target power value is regenerated based on the reconstruction node identifier and the message is sent again.

[0063] The channel parameters of the BeiDou short message terminal are collected to generate a channel feature vector. This channel feature vector is then competitively mapped in a self-organizing topology space, outputting the active node identifier and node activation degree, including:

[0064] The received signal strength, signal-to-noise ratio, and bit error rate of the BeiDou short message terminal within a continuous time window are collected as channel parameters. The channel parameters are sampled by a time-series sliding window. The sampling results are used to construct a two-dimensional matrix according to the time dimension and the parameter dimension and then normalized to generate a channel feature vector.

[0065] Calculate the Euclidean distance between the channel feature vector and the weight vectors of each node in the self-organized topology space, identify the node with the smallest Euclidean distance, and extract the node's spatial location index as the identifier of the active node.

[0066] Based on the activation node identifier, locate the corresponding node weight vector, and calculate the cosine similarity between the channel feature vector and the located node weight vector as the initial activation degree.

[0067] Extract the received signal strength component from the channel feature vector, calculate the channel quality attenuation factor by normalizing the ratio of the received signal strength component to the preset signal strength reference value, and then weight and correct the initial activation degree according to the channel quality attenuation factor to obtain the node activation degree. Output the active node identifier and node activation degree.

[0068] In the BeiDou short message communication scenario, accurate channel state perception is a prerequisite for adaptive power control. Given the dynamic wireless channel environment in which the BeiDou short message terminal operates, it is necessary to collect physical layer parameters reflecting the current channel quality from multiple dimensions. Specifically, using continuous time windows as units, the terminal's Received Signal Strength (RSSI), Signal-to-Noise Ratio (SNR), and Bit Error Rate (BER) are periodically collected. Received Signal Strength reflects the combined effect of path loss and shadowing fading; SNR characterizes the current link's anti-interference capability; and BER directly measures the reliability of data transmission. These three parameters characterize the channel state from different perspectives, forming a complementary multi-dimensional observation.

[0069] When performing time-series sliding window sampling on the acquired channel parameters, the window length is set to... Step size is Within each sampling window, the mean, variance, and range of the received signal strength, signal-to-noise ratio, and bit error rate sequences are calculated, thereby compressing the continuous time-series signal into statistical features of finite dimensions. The sampling results are arranged along the time dimension (row direction) and the parameter dimension (column direction) to construct a two-dimensional matrix. Each row corresponds to a time sampling point, and each column corresponds to the statistical characteristics of a type of channel parameter.

[0070] To eliminate the influence of differences in the dimensions of different parameters on the subsequent competitive mapping process, the matrix is... Perform column-by-column max-min normalization. For the matrix... any element of the column Normalized value ,in To prevent the introduction of extremely small positive numbers with a denominator of zero, For row index, Column index. Normalized matrix. Expanding this into a one-dimensional vector yields the channel feature vector. , This vector fully describes the channel dynamics within the current time window.

[0071] The self-organizing topology space adopts a self-organizing map (SOM) network structure, consisting of several nodes arranged in a two-dimensional grid. Each node... Maintain a weight vector with the same dimension as the channel feature vector. The network has been trained offline using a large amount of historical channel data, and the weight vectors of each node represent prototypes of different typical channel states. During the online inference phase, the current channel feature vectors are... Given a pre-trained self-organizing topology space, compute... Euclidean distance to all node weight vectors: ,in Channel feature vector The One portion, For nodes The weight vector of the first One portion, Channel feature vectors and nodes The Euclidean distance between weight vectors. After traversing all nodes, the identification... The node with the minimum value is the winning node that best matches the current channel state. Its spatial location index (a combination of row and column numbers) in the two-dimensional grid is extracted as the active node identifier. This is used for subsequent extraction of seed power values ​​and Bayesian posterior update operations.

[0072] Determine the activation node identifier Then, locate the corresponding node weight vector. Calculate the channel feature vector and The cosine similarity between them serves as the initial activation level, reflecting the degree of matching between the current channel state and the node prototype. The range of values ​​for cosine similarity is: ,when When the value is close to 1, it indicates that the current channel characteristics highly match the typical channel state represented by the active node, and the mapping confidence is high; when... When the value is too low, it indicates that the current channel state is in a transitional region between multiple typical states, and further refinement of the activation degree representation is required through subsequent weighted correction.

[0073] The initial activation only reflects the similarity between the overall feature vector and the node prototype, and does not fully utilize the received signal strength component, which is the most direct indicator of the current link quality. Therefore, the received signal strength component is extracted from the channel feature vector. (The statistical mean of the received signal strength column in the corresponding normalized matrix) is compared with the preset signal strength reference value. The channel quality attenuation factor is obtained by performing normalized ratio calculation. , ,in The received signal strength reference value, empirically set under standard channel conditions, is typically taken as the normalized value corresponding to the rated received power in the system design link budget. Higher than hour, This indicates that the current channel quality is better than the baseline, and the activation level can be appropriately enhanced; when Below hour, This indicates that there is significant fading in the channel, and the activation level needs to be compressed accordingly to avoid overestimating the node matching confidence level and causing power allocation deviation.

[0074] Based on channel quality attenuation factor initial activation After weighted adjustment, the final node activation degree is obtained. : ,in This indicates that the result will be truncated to... The interval ensures that the activation value has probabilistic interpretability. Corrected node activation value. It comprehensively reflects the global similarity between the channel feature vector and the prototype of the active node, as well as the local quality state of the current received signal strength, and can more accurately describe the applicability of the power strategy corresponding to the active node under the current channel conditions.

[0075] Finally, the node identifier will be activated. With node activation The common output is passed to the subsequent Generative Adversarial Network (GAN) module. Activation node identifier. Used to retrieve seed power values ​​and node activation from a pre-stored node-power mapping table. This serves as the modulation weight in the process of iteratively generating candidate power values ​​in the adversarial generative network, enabling the generated candidate power values ​​to be adaptively scaled according to the channel matching confidence. This tends to generate a more conservative power compensation strategy when the channel quality is poor, and avoids unnecessary power waste when the channel quality is good, thus achieving accurate matching between transmit power and channel state.

[0076] The seed power value is extracted based on the activated node identifier. A generative adversarial network is constructed to iteratively generate candidate power values ​​based on the seed power value and node activation degree until the transmission reliability score output by the discriminator reaches a preset reliability threshold. The optimized candidate power values ​​are then output as the base power values, including:

[0077] Based on the activation node identifier, the corresponding initial power value is retrieved from the preset power mapping table as the seed power value;

[0078] Construct an adversarial generative network, which includes a generator and a discriminator. The generator maps node activations to power adjustment vectors and incrementally adjusts the seed power value based on the power adjustment vectors to generate candidate power values.

[0079] The discriminator receives candidate power values ​​and constructs a virtual transmission scenario. In the virtual transmission scenario, it simulates the signal propagation process of the candidate power values ​​and calculates the received power of the signal reaching the Beidou satellite. Based on the received power and the preset Beidou satellite receiving sensitivity threshold, it calculates the power margin and generates a transmission reliability score.

[0080] When the transmission reliability score does not reach the preset reliability threshold, the power compensation amount is calculated based on the deviation between the received power and the Beidou satellite receiving sensitivity threshold.

[0081] The power adjustment vector is adjusted according to the power compensation amount, and the seed power value is incrementally adjusted to generate a new candidate power value. This process is repeated until the transmission reliability score reaches the preset reliability threshold. The candidate power value that reaches the preset reliability threshold is then output as the base power value.

[0082] After obtaining the active node identifier, the seed power value corresponding to that node needs to be extracted from the preset power mapping table. The power mapping table is constructed offline during system initialization. Each node identifier corresponds to an empirical initial power value, derived from statistical analysis of historical transmission data, reflecting the typical transmit power level required for the channel state category represented by that node. The initial power value is retrieved directly from the power mapping table based on the active node identifier and recorded as the seed power value. The unit is dBm. The seed power value serves as the starting point for subsequent adversarial generative network iterative optimization, and its rationality directly affects the iteration convergence speed. Therefore, the quality of the power mapping table has a significant impact on the overall efficiency of the method.

[0083] In constructing a generative adversarial network, the generator and discriminator each play different functional roles. The generator's input is the node activation level. It is mapped to a power adjustment vector through a multi-layer fully connected network. Power adjustment vector It is a scalar sequence representing the incremental adjustment based on the seed power value. Its sign and magnitude reflect the required power compensation direction and intensity under the current channel conditions. Candidate power value It is obtained by superimposing the seed power value and the power adjustment vector, i.e. The generator's network structure employs a fully connected structure with a batch normalization layer, uses ReLU as the activation function, and employs linear activation in the output layer to ensure that the power adjustment vector takes values ​​in the continuous real number domain, thereby supporting fine-grained power increment control.

[0084] The discriminator receives candidate power values Subsequently, a virtual transmission scenario was constructed internally to simulate the signal propagation process from the BeiDou short message terminal to the BeiDou satellite. The virtual transmission scenario was modeled based on a free-space path loss model combined with fading parameters in the current channel eigenvector. It comprehensively considered factors such as atmospheric refraction loss, multipath fading margin, and terminal antenna gain to calculate the received power when the signal reaches the BeiDou satellite. Specifically, path loss Based on the communication distance, carrier frequency, and fading coefficient in the channel eigenvector, the received power is estimated to meet the following conditions: ,in These are antenna gain correction values, all in dBm or dB.

[0085] In the calculation Then, it is compared with the preset BeiDou satellite receiver sensitivity threshold. Compare and calculate power margin ,satisfy .when When the signal is greater than zero, it indicates that the signal energy is sufficient to be reliably received by the satellite; when... A value less than zero indicates that the current candidate power value is insufficient to guarantee successful transmission. (Based on power margin) The discriminator maps it to a transmission reliability score using the Sigmoid function. ,satisfy ,in This is the sensitivity adjustment coefficient, used to control the slope of the score's response to changes in power margin. Transmission reliability score. The value range is (0, 1), and the higher the score, the more likely the current candidate power value is to achieve reliable transmission in the virtual transmission scenario.

[0086] Preset reliability threshold The system uses a pre-set judgment benchmark based on the quality requirements of the BeiDou short message service, typically calculated based on the target block error rate or link budget margin requirements. When the transmission reliability score... Not achieved At that time, the discriminator will and The deviation is used as the power compensation amount Feedback is given to the generator to satisfy Power compensation amount It reflects the energy gap between the current candidate power value and the target transmission requirement. Its dimension is dBm. A positive value indicates that the transmission power needs to be increased, while a negative value indicates that the power can be appropriately reduced to avoid interference.

[0087] The generator is based on the power compensation amount Power adjustment vector The correction will be made using a gradient-guided update strategy. As an additional input concatenated to the generator's input layer, along with node activation values... The co-driving network re-infers the adjusted power adjustment vector. The updated candidate power value is The data is then input into the discriminator again for virtual scene evaluation, and the above process is repeated. In each iteration, the generator's network weights are updated using the backpropagation algorithm based on the difference between the discriminator's output transmission reliability score and the preset reliability threshold. The discriminator's network weights are updated based on the physical constraints in the virtual transmission scene and the distribution of real transmission samples. The two work alternately to optimize, forming an adversarial training mechanism.

[0088] The iteration termination condition for generative adversarial networks is: transmission reliability score. Reaching or exceeding a preset reliability threshold Or the number of iterations reaches the preset maximum number of iteration rounds. During the iteration process, to prevent the candidate power values ​​from exceeding the hardware transmission power range of the BeiDou short message terminal, [measures were taken]. Apply upper and lower bound constraints, upper bound The lower bound is determined by the terminal hardware specifications. Determined by the minimum communication link budget requirement. If in a certain iteration... If the value exceeds the constraint range, it will be trimmed to the boundary value before being sent to the discriminator to ensure that the candidate power value is always within the physically realizable range.

[0089] When the iteration process meets the termination condition, the current transmission reliability score is output. The candidate power value is used as the base power value If the iteration terminates due to reaching the maximum number of iterations, the candidate power value corresponding to the highest transmission reliability score in each iteration is selected as the base power value to ensure that the output result is optimal within the iteration capability range. Base Power Value This will serve as the input for the subsequent causal reasoning compensation stage, and will be further refined by combining the causal effect quantities of the channel fading factor and the interference factor, ultimately yielding the target power value for controlling the actual transmit power.

[0090] The training phase of the Generative Adversarial Network (GAN) was completed offline before system deployment. Training data was derived from historical BeiDou short message transmission records, including node activation levels, seed power values, and transmission result annotations under different channel conditions. During training, the discriminator used real successful transmission samples as positive samples and virtual transmission results corresponding to candidate power values ​​generated by the generator as negative samples. A binary cross-entropy loss function was used to drive the discriminator to improve its ability to judge the reliability of candidate power values. The generator, on the other hand, aimed to maximize the transmission reliability score output by the discriminator, using adversarial loss to guide it to continuously generate more reasonable power adjustment vectors. In the online inference phase, only the pre-trained generator and discriminator were used to perform forward inference and a finite number of iterations, eliminating the need for retraining and ensuring real-time performance.

[0091] The discriminator receives candidate power values ​​and constructs a virtual transmission scenario. Within this scenario, it simulates the signal propagation process of the candidate power values ​​and calculates the received power of the signal reaching the BeiDou satellite. Based on the received power and a preset BeiDou satellite receiver sensitivity threshold, it calculates the power margin and generates a transmission reliability score, including:

[0092] The discriminator receives candidate power values, extracts the received signal strength component and signal-to-noise ratio component from the channel feature vector, and calculates the multipath propagation delay spread parameter and Doppler frequency shift parameter as multipath fading parameters based on the fading amplitude of the received signal strength component and the noise intensity of the signal-to-noise ratio component.

[0093] The time-domain fading distribution in the virtual transmission scenario is determined based on the multipath propagation delay spread parameter, and the frequency-domain fading distribution in the virtual transmission scenario is determined based on the Doppler frequency shift parameter. A virtual transmission scenario with time-frequency fading characteristics is constructed, and the candidate power value is used as the transmit power to start the signal propagation process simulation in the virtual transmission scenario.

[0094] Based on the spatial distance between the BeiDou short message terminal and the BeiDou satellite, free space path attenuation is applied to the candidate power value. Based on the time domain fading distribution and frequency domain fading distribution, multipath fading attenuation is applied to the power after free space path attenuation to obtain the received power of the signal reaching the BeiDou satellite.

[0095] Obtain the preset BeiDou satellite receiving sensitivity threshold, calculate the power margin between the received power and the preset BeiDou satellite receiving sensitivity threshold, and generate a transmission reliability score by weighting the power margin and node activation degree according to the preset weight coefficient.

[0096] The discriminator receives the candidate power values ​​output by the generator. Next, the received signal strength component and signal-to-noise ratio (SNR) component are extracted from the channel feature vector. The received signal strength component reflects the overall fading degree of the current channel, while the SNR component characterizes the combined intensity of noise and interference in the channel. By statistically analyzing the fluctuation amplitude of the received signal strength component in the time domain, the amplitude fluctuations caused by multipath effects can be quantified; by analyzing the spectral distribution of noise intensity in the SNR component, the frequency shift characteristics caused by the Doppler effect can be extracted.

[0097] Multipath propagation delay spread parameters This parameter describes the dispersion of multipath signals in the time domain and is calculated from the variance of the fading amplitude of the received signal strength component. A larger variance in fading amplitude indicates a more significant difference in the time delay of signals arriving from different paths, and a more severe time delay spread. (Doppler frequency shift parameter) It is estimated from the spectral peak offset of the noise intensity in the signal-to-noise ratio component, reflecting the degree of influence of the relative motion speed between the terminal and the satellite on the signal frequency. and Together, they constitute a multipath fading parameter combination, providing a physical basis for the subsequent construction of virtual transmission scenarios.

[0098] During the construction phase of the virtual transmission scenario, based on the multipath propagation delay spread parameter... Determine the temporal fading distribution. Specifically, When the value is large, the temporal fading distribution exhibits a wide exponential decay profile, indicating that the energy of each multipath component is dispersed over a long time range. When the value is small, the time-domain fading distribution is concentrated within a shorter time delay range, and the time dispersion effect of the channel is relatively limited. This is based on the Doppler frequency shift parameter. Determine the frequency domain fading distribution. When the value is large, the bandwidth of the frequency domain fading distribution expands, the signal spectrum broadens significantly, and the coherence bandwidth decreases. When the frequency is low, the frequency domain fading distribution is relatively concentrated, and the channel's frequency selectivity is weak. Combining the time domain fading distribution with the frequency domain fading distribution creates a virtual transmission scenario with joint time-frequency fading characteristics. This scenario can simultaneously characterize the non-stationary fading behavior of the channel in both the time and frequency dimensions, making the power assessment more closely resemble the physical process of BeiDou short messages in a real propagation environment.

[0099] In virtual transmission scenarios, candidate power values The initial transmit power from the transmitter initiates the simulation of the signal propagation process. The signal originates from the BeiDou short message terminal and first undergoes free-space path attenuation. The free-space path attenuation amount... The spatial distance between the BeiDou short message terminal and the BeiDou satellite and signal carrier frequency The decision is made jointly, and the calculation method is as follows: ,in, The speed of light is constant. After attenuation along a free-space path, the signal power... for: All the above calculations are in dBm. In the BeiDou satellite communication scenario, the spatial distance between the terminal and the satellite is usually on the order of tens of thousands of kilometers. Free space path loss is one of the main factors affecting received power, so accurate modeling of this loss is crucial for power assessment.

[0100] Building upon free-space path attenuation, multipath fading attenuation is further applied. Multipath fading attenuation amount. The fading effect is determined jointly by the time-domain and frequency-domain fading distributions. The time-domain fading distribution determines the relative power weights of each multipath component, while the frequency-domain fading distribution determines the fading depth experienced by each subcarrier within the current signal bandwidth. The combined multipath fading attenuation is obtained by superimposing the fading effects of both. The signal reception power to the BeiDou satellite. for: ,in, This represents the antenna gain correction, indicating the compensation correction of the BeiDou short message terminal antenna directivity gain for the received power. Through the power calculations in the above three stages, the candidate power values ​​in the virtual transmission scenario are finally obtained. Corresponding satellite receiving power .

[0101] Obtain the preset BeiDou satellite receiver sensitivity threshold Then, calculate the received power. and Power margin between , When the value is positive, it indicates that the candidate power value can ensure that the received power at the satellite end exceeds the sensitivity threshold, thus meeting the basic conditions for reliable transmission. A negative value indicates that the current candidate power value is insufficient to support reliable transmission. The value of directly reflects the transmission margin of the candidate power value under the current virtual channel conditions.

[0102] In generating transmission reliability scores At that time, the power margin With node activation Node activation degree is calculated by weighting the results according to preset weighting coefficients. The output from the self-organizing topology mapping stage reflects the degree of matching between the current channel feature vector and the weight vector of the activated nodes. A higher value indicates that the current channel state is closer to the typical channel mode represented by that node, and the corresponding seed power value has higher reference value. Node activation degree is introduced. As a weighting factor, it can further integrate channel matching quality information based on power margin assessment, enabling the transmission reliability score to more comprehensively reflect the overall applicability of candidate power values. Let the preset weighting coefficient of power margin be... The preset weighting coefficient for node activation is: Then the transmission reliability score The calculation method is as follows: ,in, For power margin The dimensionless values ​​after normalization ensure that the power margin and node activation are comparable in magnitude. and satisfy The specific values ​​of both are pre-configured based on the emphasis placed on transmission reliability and channel matching in the BeiDou short message application scenario. When Reaching the preset reliability threshold When the discriminator outputs a pass signal, the generator stops iterating, and the current candidate power value is used as the base power value. Output to subsequent causal reasoning and compensation stages; when Not achieved At this point, the discriminator feeds back the evaluation results to the generator, driving the generator to adjust the power adjustment vector in the next iteration. New candidate power values ​​are generated and continue to participate in the evaluation until the convergence condition is met or the preset maximum number of iterations is reached. .

[0103] Through the construction of the virtual transmission scenario and the multi-stage power attenuation modeling, the discriminator can perform a high-fidelity reliability assessment of the candidate power values ​​output by the generator without relying on feedback from real satellite links. This effectively guides the power optimization iteration process of the adversarial generative network and ultimately outputs a basic power value that meets the transmission quality requirements of BeiDou short messages.

[0104] like Figure 2 As shown, Figure 2 This is a flowchart illustrating the operation of the channel power compensation algorithm based on causal inference in an embodiment of the present invention.

[0105] A directed acyclic causal graph is constructed from the channel feature vectors. The causal effects of the channel fading factor and interference factor are separated through intervention operations. Based on these causal effects, the target power value is obtained by counterfactual inference compensation of the baseline power value, including:

[0106] The received signal strength component, signal-to-noise ratio component, Doppler frequency shift component, and time delay spread component are extracted from the channel feature vector. The received signal strength component and signal-to-noise ratio component are used as observation variable nodes, and the Doppler frequency shift component and time delay spread component are used as factor variable nodes. A directed edge connection is established based on the influence relationship between the observation variable nodes and the factor variable nodes to construct a directed acyclic causal graph.

[0107] The value of the Doppler frequency shift component is fixed and the change amplitude of the conditional probability distribution of the observed variable node is calculated as the causal effect of the channel fading factor. The value of the delay spread component is fixed and the change amplitude of the conditional probability distribution of the observed variable node is calculated as the causal effect of the interference factor.

[0108] The causal effect of the channel fading factor is set to the maximum fading level, and the fading counterfactual value of the received signal strength component is derived in reverse. The causal effect of the interference factor is set to the maximum interference level, and the interference counterfactual value of the signal-to-noise ratio is derived in reverse.

[0109] The fading compensation amount is obtained by calculating the difference between the fading counterfactual value and the actual value of the received signal strength component. The interference compensation amount is obtained by calculating the difference between the interference counterfactual value and the actual value of the signal-to-noise ratio component. The fading compensation amount and the interference compensation amount are added to the base power value to obtain the target power value.

[0110] After generating the base power value, it is necessary to further quantitatively separate the fading effect and interference effect in the channel through a causal reasoning mechanism, and then compensate the base power value accordingly to obtain the target power value that can cope with the worst channel conditions.

[0111] Four key components are extracted from the channel feature vector: Received Signal Strength Indicator (RSSI), Signal-to-Noise Ratio (SNR), Doppler Shift, and Delay Spread. Among these, RSSI and SNR directly reflect the transmission quality of the current link and are observable outcome variables; therefore, they are designated as observation variable nodes in the directed acyclic causal graph. The Doppler Shift reflects the frequency shift effect caused by the relative motion between the terminal and the BeiDou satellite, while the Delay Spread reflects the time-domain spread effect caused by multipath propagation. Both are causal variables affecting transmission quality and are therefore designated as factor variable nodes.

[0112] When constructing a directed acyclic causal graph, the directed edge connections between nodes are determined based on the physical channel propagation mechanism: the Doppler frequency shift component has an impact path on both the received signal strength component and the signal-to-noise ratio (SNR) component; that is, the Doppler effect causes changes in the signal coherence bandwidth, leading to fluctuations in received power and a decrease in SNR. Similarly, the delay spread component has an impact path on both the received signal strength component and the SNR component; that is, multipath delay spread causes inter-symbol interference, directly degrading the SNR and causing distributed received power. Therefore, the directed acyclic causal graph contains four directed edges: from the Doppler frequency shift node to the received signal strength node, from the Doppler frequency shift node to the SNR node, from the delay spread node to the received signal strength node, and from the delay spread node to the SNR node, forming a complete causal topology. This graph satisfies the definition constraints of a directed acyclic graph, contains no directed loops, and guarantees the feasibility of causal reasoning.

[0113] When separating the causal effects of the channel fading factor and the interference factor, the do-calculus operation in the Pearl causal framework is used. When estimating the causal effects of the channel fading factor, the value of the Doppler shift component is fixed to a specific value in the observed data distribution. All incoming edges between the Doppler shift node and its upstream node are severed, so that the value of that node is no longer affected by other variables and is determined solely by external intervention. Under this intervention condition, the change in the conditional probability distribution of the observed variable nodes (received signal strength component and signal-to-noise ratio component) relative to the uninterrupted state is calculated. This change is defined as the causal effect of the channel fading factor, denoted as […]. . Specifically, This characterizes the degree of shift in the mean distribution of the observed variable nodes when the Doppler frequency shift takes a specific value, i.e.: ,in, This indicates the value of the observed variable node (received signal strength component or signal-to-noise ratio component). This indicates the value at which the Doppler frequency shift component is fixed during intervention. This represents the expected value of the observed variable node after intervention. This represents the expected value of the observed variable node under no-intervention conditions.

[0114] When estimating the causal effect size of the interference factor, the same intervention operation framework is used, and the value of the time delay spread component is fixed to a specific value. Cut off all incoming edges of the delay-spreading node, calculate the magnitude of change in the conditional probability distribution of the observed variable node, and define this magnitude as the causal effect of the interference factor, denoted as . ,Right now: ,in, Indicates the delay spread component, This indicates the value of the time delay spread component that is fixed during intervention. and The independent contributions of the Doppler frequency shift effect and the multipath delay spread effect to the transmission quality were quantified, and the causal decoupling of the two types of physical effects was achieved.

[0115] After obtaining the causal effect quantity, counterfactual reasoning is performed to calculate the power compensation quantity. For the channel fading factor, its causal effect quantity is... Set as the maximum fading level, that is, assuming that the Doppler frequency shift effect acts on the channel to the worst extent, derive the counterfactual fading value that the received signal strength component should exhibit under this extreme condition, denoted as . . Specifically, The effect size corresponding to the maximum fading level is obtained by superimposing the actual observed values ​​of the currently received signal strength components. ,in, This represents the actual observed value of the received signal strength component. This represents the causal effect of the channel fading factor at the maximum fading level. This value is usually taken as the extreme quantile of the causal effect in historical observation data (such as the 95th percentile).

[0116] For the interference factor, its causal effect size Set to the maximum interference level, the counterfactual values ​​of the signal-to-noise ratio components under extreme interference conditions are derived in reverse, denoted as . , ,in, This represents the actual observed value of the signal-to-noise ratio component. This represents the causal effect size of the interference factor at the maximum interference level.

[0117] Calculate fading compensation The interference compensation amount is calculated based on the difference between the counterfactual fading value and the actual value of the received signal strength component. To interfere with the difference between the counterfactual numerical value and the actual value of the signal-to-noise ratio component, The additional power margin required to counteract the maximum fading effect was quantified. The additional power margin required to counteract the maximum interference effect is quantified. Both compensation values ​​are expressed in the linear domain of signal strength or signal-to-noise ratio, and need to be converted into power domain compensation values ​​through link budget relationships before being added to the base power value.

[0118] fading compensation With interference compensation amount After being converted into the corresponding power compensation increment, it is added to the base power value. Obtain the target power value , ,in, The conversion coefficient from fading compensation to the power domain is given. The conversion coefficients from interference compensation to the power domain are determined based on the mapping relationship between received signal strength and transmitted power, and the mapping relationship between signal-to-noise ratio and transmitted power in the link budget model. This is the target power value after causal reasoning compensation. This value, based on the base power value, fully considers the fading and interference effects under worst-case channel conditions, ensuring the reliability of BeiDou short message transmission even under extreme channel conditions. In practical applications, It also needs to undergo power upper and lower bound constraint trimming to ensure that it does not exceed the maximum transmit power range allowed by the terminal hardware, thereby achieving a balance between power efficiency and transmission reliability.

[0119] The causal effect of fixing the value of the Doppler frequency shift component and calculating the change in the conditional probability distribution of the observed variable nodes as the causal effect of the channel fading factor, and fixing the value of the delay spread component and calculating the change in the conditional probability distribution of the observed variable nodes as the causal effect of the interference factor, include:

[0120] Construct a joint probability distribution of Doppler frequency shift components and time delay spread components in historical transmission records, and sample from the joint probability distribution to generate an intervention sample set containing multiple sets of numerical combinations;

[0121] The Doppler frequency shift component values ​​in the intervention sample set are sequentially fixed to the factor variable nodes corresponding to the Doppler frequency shift components in the directed acyclic causal graph, and the input edges of the factor variable nodes are cut off. The mean of the conditional probability distribution of the observation variable nodes corresponding to the received signal strength components after each fixation is calculated, and the standard deviation of the mean of all conditional probability distributions is used as the causal effect of the channel fading factor.

[0122] The time delay spread component values ​​in the intervention sample set are sequentially fixed to the factor variable nodes corresponding to the time delay spread components in the directed acyclic causal graph, and the input edges of the factor variable nodes are cut off. The mean of the conditional probability distribution of the observation variable nodes corresponding to the signal-to-noise ratio component after each fixation is calculated, and the standard deviation of the mean of all conditional probability distributions is used as the causal effect of the interference factor.

[0123] After constructing the directed acyclic causal graph, it is necessary to accurately quantify the causal effects of the channel fading factor and the interference factor. Directly using statistical correlations from observational data is insufficient to reflect the true causal relationship because there may be interference from common confounding variables between the Doppler shift component and the delay spread component. Therefore, a causal inference framework based on intervention computation is adopted. By constructing a joint probability distribution and generating an intervention sample set, the independent causal effects of each factor are systematically isolated.

[0124] For the Doppler frequency shift component and time delay spread component data accumulated in historical transmission records, a joint probability distribution model for both is first constructed. Historical transmission records contain snapshots of channel parameters collected during each message transmission. The Doppler frequency shift component reflects the frequency drift caused by the relative motion between the terminal and the BeiDou satellite, while the time delay spread component reflects the degree of signal arrival time dispersion caused by multipath propagation. The Doppler frequency shift component is denoted as... The delay spread component is denoted as Based on historical records For each sample, the kernel density estimation method is used to fit the two-dimensional joint probability density function. The kernel function uses a Gaussian kernel, and the bandwidth is adaptively determined using the Silverman criterion. After establishing the joint probability distribution, samples are extracted from it using Monte Carlo sampling. The numerical combinations constitute the intervention sample set. ,in For sample index, The value of needs to be balanced between statistical significance and computational cost, and is usually set to no less than 200 to ensure the stability of the standard deviation estimate.

[0125] In the calculation of the causal effect of the channel fading factor, for each set of numerical combinations in the intervention sample set, the Doppler frequency shift component values ​​are... The factors are sequentially fixed to the corresponding factor variable nodes in the directed acyclic causal graph, and a cut-off operation is performed, i.e., all input edges of the node are deleted, so that the value of the node is no longer affected by its parent node and is determined only by the external intervention value. This operation corresponds to the do operator in causal inference theory, denoted as […]. After cutting the input edge, the structure of the directed acyclic causal graph undergoes a local change; the Doppler shift component node becomes an exogenous variable, while the conditional dependencies of other nodes remain unchanged. Under this graph structure, message passing is performed along the causal path to calculate the conditional probability distribution of the observed variable nodes corresponding to the received signal strength component. The first... The mean of the conditional probability distribution of the observed variable nodes for the received signal strength component under each intervention is denoted as . For all Mean sequence obtained from the intervention Calculate the standard deviation as the causal effect of the channel fading factor. The calculation method is as follows: ,in It is the arithmetic mean of the means of all conditional probability distributions, i.e. . The physical meaning is that when the Doppler frequency shift component changes within its natural distribution range, the greater the fluctuation of the expected value of the received signal strength component, the more significant the causal impact of channel fading on the received power, and a larger compensation weight should be allocated to the fading factor in subsequent power compensation.

[0126] The calculation logic for the causal effect size of the interference factor is symmetrical to the above process, but the objects of action are different. The time-delayed spread component of each numerical combination in the intervention sample set is calculated. Sequentially fix the factor variable nodes corresponding to the time delay spread components in the directed acyclic causal graph, and similarly cut off all input edges of those nodes, then execute... Intervention operation. Under the modified graph structure, calculate the conditional probability distribution of the observed variable nodes corresponding to the signal-to-noise ratio components. (The remaining text appears to be incomplete and requires further context.) The mean of the conditional probability distribution of the signal-to-noise ratio component observation variable nodes under each intervention is denoted as . For all Mean sequence obtained from the intervention Calculate the standard deviation as the causal effect size of the confounding factor. The calculation method is as follows: ,in It is the arithmetic mean of the conditional probability distributions of all signal-to-noise ratio components, i.e. An increase in the delay spread component usually indicates increased multipath interference, resulting in selective fading of the signal in the frequency domain and a decrease in the signal-to-noise ratio; through Quantifying the strength of this causal link can provide an objective basis for subsequent interference compensation.

[0127] In the calculation of conditional probability distributions, the conditional dependencies between nodes in the directed acyclic causal graph are parametrically modeled using Gaussian process regression. The conditional probability distribution of each child node is determined by the value of its parent node through a linear mixture model, with residual noise terms added. After the severance operation is performed, the set of parent nodes of the intervened node becomes empty, and its conditional probability distribution degenerates into a Gaussian distribution with the intervention value as the mean and the historical residual variance as the variance. Downstream nodes recursively calculate their own conditional probability distributions based on the updated graph structure until the observed variable node is reached. The entire message passing process is executed sequentially under the topological sorting of the directed acyclic graph, avoiding circular dependencies.

[0128] Regarding the size of the intervention sample set The selection of [specific parameter] can be dynamically adjusted based on the terminal's computing resources in actual engineering deployment. When the terminal has strong floating-point computing capabilities, the [specific parameter] can be appropriately increased. To improve the accuracy of causal effect estimation; in resource-constrained scenarios, it can... The confidence interval was reduced to between 50 and 100, and a bootstrap method was used to correct the estimation results to compensate for the bias caused by insufficient sample size. Furthermore, the timeliness of historical transmission records has a significant impact on the accuracy of the joint probability distribution; therefore, it is recommended to use only the most recent records. The system records an exponentially decaying weight for earlier records to accommodate the slow drift of channel statistics over time.

[0129] After completion and After calculation, the two will serve as the core basis for counterfactual reasoning compensation, and will be used to determine the adjustment ratio of the fading compensation amount and the interference compensation amount relative to the basic power value, thereby generating the final target power value and ensuring the transmission reliability of BeiDou short messages under complex channel conditions.

[0130] The power loss difference and channel feature vector are combined as observational evidence. The Bayesian posterior probability of the active node's neighboring nodes is updated using kernel density estimation. The node with the highest posterior probability is selected as the reconstructed node identifier. The target power value is regenerated based on the reconstructed node identifier, and the message is retransmitted, including:

[0131] The power loss difference is combined with the channel feature vector to form an observation evidence vector that contains both power dimension information and channel dimension information.

[0132] In a directed acyclic causal graph, all adjacent connected nodes of the node corresponding to the activated node are extracted as a neighborhood node set. For each node in the neighborhood node set, the observed evidence vector is used as the sample input to construct a kernel density estimation function. The probability density value of the kernel density estimation function at the node position is calculated as the likelihood probability. The likelihood probability is multiplied by the prior probability of the node according to the Bayes formula and then normalized to obtain the Bayes posterior probability of the node.

[0133] Traverse the Bayesian posterior probabilities of all nodes in the neighborhood node set and select the node with the largest value as the reconstruction node identifier. Extract the causal effect of the channel fading factor and the causal effect of the interference factor corresponding to the reconstruction node identifier. Based on the causal effect of the channel fading factor and the causal effect of the interference factor, perform counterfactual reasoning compensation on the base power value to regenerate the target power value. Use the regenerated target power value to send the BeiDou short message again.

[0134] When the transmission status is marked as failed, it means that the current target power value fails to meet the reliable transmission requirements of BeiDou short messages. In this case, the power loss difference is combined with the channel feature vector to form an observation evidence vector, which contains both power and channel dimension information. Specifically, the power loss difference reflects the deviation between the target power value and the actual received power in this transmission, revealing the actual level of channel path loss; the channel feature vector carries multi-dimensional channel state information such as received signal strength, signal-to-noise ratio, Doppler shift, and delay spread. The two are concatenated to form the observation evidence vector, denoted as... Its dimensions are ,in is the dimension of the channel feature vector, with the additional dimension corresponding to the power loss difference.

[0135] In a directed acyclic causal graph, the node corresponding to the activated node has already been determined in the current transmission round. Extract all adjacent connected nodes of this node in the causal graph topology, including parent and child nodes, to form a neighborhood node set, denoted as . The range of the neighborhood node set is limited to the node level directly connected to the active node, and does not extend to more distant hops, in order to ensure the locality and computational efficiency of the posterior update.

[0136] For the set of neighboring nodes Each node in ( ), and the observation evidence vector As input samples, construct a node The historical activation records serve as the reference distribution's kernel density estimation function. The construction of the kernel density estimation function depends on the nodes. The set of observation evidence vector samples corresponding to the activation in the historical transmission record, denoted as The total number of samples is The corresponding historical activation count. Kernel density estimation uses a Gaussian kernel function, and the bandwidth parameter is adaptively determined using the Silverman rule to balance the smoothness of the estimation with detail resolution. The kernel density estimation function at the node... Location of observation evidence vector The probability density value is denoted as This value serves as a node. The likelihood probability reflects the probability at the node. The probability of generating the current observation evidence vector under the described channel conditions.

[0137] node Prior probability Node activation frequency statistics derived from self-organizing topology space, i.e., nodes The proportion of times a node is selected as the active node out of the total number of transmissions in all historical transmission records. After multiplying the likelihood probability and prior probability using Bayes' theorem, the product is normalized over all nodes in the neighborhood node set to obtain the node... Bayesian posterior probability The calculation process is as follows: In this case, the denominator represents the set of neighboring nodes. The product of the likelihood probability and the prior probability of all nodes is summed to perform normalization, ensuring that the sum of the Bayesian posterior probabilities of all nodes is 1. This normalization process eliminates the influence of the absolute probability density magnitude, making the posterior probabilities of different nodes comparable.

[0138] Traverse the set of neighboring nodes The Bayesian posterior probabilities of all nodes are used to determine the reconstructed node identifier, denoted as . , Reconstruct node identifiers This represents the node in the neighborhood node set that best matches the actual channel state under the current observational evidence conditions. The channel feature distribution described by its weight vector has the highest posterior consistency with the channel environment under the current failed transmission scenario.

[0139] Extract Reconstruction Node Identifiers The corresponding causal effects of the channel fading factor and interference factor are represented in the directed acyclic causal graph. By performing intervention operations on the causal graph, the independent contributions of the channel fading path and interference path to the observed variable nodes are separated, and the reconstructed nodes are obtained respectively. Corresponding fading causal effect size With interference causal effect size Unlike the initial generation of the target power value which uses the causal effect of the active node, this method uses the causal effect of the reconstructed node, thus incorporating the channel state deviation information revealed by failed transmissions into the power compensation calculation process.

[0140] according to and For the base power value Counterfactual inference compensation is performed to regenerate the target power value. The compensation logic is consistent with the counterfactual inference compensation mechanism: the fading compensation amount is derived from the fading causal effect amount through a conversion coefficient. Mapped to the power domain, the interference compensation amount is derived from the interference causality effect amount through the conversion coefficient. Mapped to the power domain, the two are superimposed to the base power value. The regenerated target power value is obtained. : The regenerated target power value Before application, a power range constraint test must be performed to ensure that it does not exceed the maximum allowable transmission power of the BeiDou short message terminal, nor is it lower than the minimum transmission power required to maintain a basic communication link. If it exceeds the range, it will be truncated to the corresponding boundary value.

[0141] Using the regenerated target power value After adjusting the transmission power, the BeiDou short message is sent again. If the transmission still fails, a new observation evidence vector is constructed using the power loss difference from this transmission and the channel feature vector. As a new activation node identifier, the above-mentioned neighborhood node Bayesian posterior probability update and reconstruction node selection process is repeated to achieve multiple rounds of adaptive reconstruction until the transmission is successful or the preset maximum number of retries is reached.

[0142] The core advantage of the above mechanism lies in the fact that kernel density estimation can model the probability distribution of node activation history samples in a non-parametric manner, without assuming a specific parameterized distribution form, and has good adaptability to the multimodal distribution of channel states. Bayesian posterior update fuses the direct evidence of transmission failure, namely power loss difference, with the multidimensional information of channel feature vectors. This makes the selection of reconstruction nodes not only dependent on the geometric similarity of channel features, but also makes full use of the physical constraint information of power propagation loss, thereby achieving more accurate channel state localization and power reconstruction in self-organizing topology space.

[0143] A second aspect of this invention provides an AI-based adaptive power control system for BeiDou short messages, comprising:

[0144] The channel feature unit is used to collect the channel parameters of the Beidou short message terminal, generate the channel feature vector, perform competitive mapping of the channel feature vector in the self-organizing topology space, and output the active node identifier and node activation degree.

[0145] The power optimization unit is used to extract seed power values ​​based on the active node identifier, and to generate candidate power values ​​iteratively based on the seed power values ​​and node activation by constructing an adversarial generative network until the transmission reliability score output by the discriminator reaches a preset reliability threshold. The optimized candidate power values ​​are then output as the base power values.

[0146] The causal effect unit is used to construct a directed acyclic causal graph from the channel feature vector. It separates the causal effect quantities of the channel fading factor and the interference factor through intervention operations, and obtains the target power value by counterfactual reasoning compensation based on the causal effect quantities of the base power value.

[0147] The power adjustment unit is used to adjust the transmission power according to the target power value and send messages, and to obtain the transmission status identifier and the actual received power.

[0148] The failure retransmission unit is used to calculate the power loss difference between the target power value and the actual received power when the transmission status is identified as failure. The power loss difference is combined with the channel feature vector as observation evidence. The Bayesian posterior probability of the neighboring nodes of the active node is updated by kernel density estimation. The node with the largest posterior probability is selected as the reconstruction node identifier. The target power value is regenerated based on the reconstruction node identifier and the message is sent again.

[0149] A third aspect of the present invention provides an electronic device, comprising:

[0150] processor;

[0151] Memory used to store processor-executable instructions;

[0152] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0153] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0154] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based adaptive power control method for BeiDou short messages, characterized in that, include: The channel parameters of the Beidou short message terminal are collected to generate a channel feature vector. The channel feature vector is then competitively mapped in the self-organizing topology space, and the active node identifier and node activation degree are output. The seed power value is extracted based on the activation node identifier. A generative adversarial network is constructed to iteratively generate candidate power values ​​based on the seed power value and node activation degree until the transmission reliability score output by the discriminator reaches the preset reliability threshold. The optimized candidate power value is then output as the base power value. A directed acyclic causal graph is constructed from the channel feature vector. The causal effect quantities of the channel fading factor and the interference factor are separated by intervention operation. The target power value is obtained by counterfactual reasoning compensation based on the causal effect quantities of the base power value. Adjust the transmit power according to the target power value and send the message to obtain the transmission status identifier and the actual received power; When the transmission status is marked as failed, the power loss difference between the target power value and the actual received power is calculated. The power loss difference is combined with the channel feature vector as observation evidence. The Bayesian posterior probability of the neighboring nodes of the active node is updated by kernel density estimation. The node with the largest posterior probability is selected as the reconstruction node identifier. The target power value is regenerated based on the reconstruction node identifier and the message is sent again.

2. The method according to claim 1, characterized in that, The channel parameters of the BeiDou short message terminal are collected to generate a channel feature vector. This channel feature vector is then competitively mapped in a self-organizing topology space, and the active node identifier and node activation degree are output, including: The received signal strength, signal-to-noise ratio, and bit error rate of the BeiDou short message terminal within a continuous time window are collected as channel parameters. The channel parameters are sampled by a time-series sliding window. The sampling results are used to construct a two-dimensional matrix according to the time dimension and the parameter dimension and then normalized to generate a channel feature vector. Calculate the Euclidean distance between the channel feature vector and the weight vectors of each node in the self-organized topology space, identify the node with the smallest Euclidean distance, and extract the node's spatial location index as the identifier of the active node. Based on the activation node identifier, locate the corresponding node weight vector, and calculate the cosine similarity between the channel feature vector and the located node weight vector as the initial activation degree. Extract the received signal strength component from the channel feature vector, calculate the channel quality attenuation factor by normalizing the ratio of the received signal strength component to the preset signal strength reference value, and then weight and correct the initial activation degree according to the channel quality attenuation factor to obtain the node activation degree. Output the active node identifier and node activation degree.

3. The method according to claim 1, characterized in that, The seed power value is extracted based on the activated node identifier. A generative adversarial network is constructed to iteratively generate candidate power values ​​based on the seed power value and node activation degree until the transmission reliability score output by the discriminator reaches a preset reliability threshold. The optimized candidate power values ​​are then output as the base power values, including: Based on the activation node identifier, the corresponding initial power value is retrieved from the preset power mapping table as the seed power value; Construct an adversarial generative network, which includes a generator and a discriminator. The generator maps node activations to power adjustment vectors and incrementally adjusts the seed power value based on the power adjustment vectors to generate candidate power values. The discriminator receives candidate power values ​​and constructs a virtual transmission scenario. In the virtual transmission scenario, it simulates the signal propagation process of the candidate power values ​​and calculates the received power of the signal reaching the Beidou satellite. Based on the received power and the preset Beidou satellite receiving sensitivity threshold, it calculates the power margin and generates a transmission reliability score. When the transmission reliability score does not reach the preset reliability threshold, the power compensation amount is calculated based on the deviation between the received power and the BeiDou satellite receiving sensitivity threshold. The power adjustment vector is adjusted according to the power compensation amount, and the seed power value is incrementally adjusted to generate a new candidate power value. This process is repeated until the transmission reliability score reaches the preset reliability threshold. The candidate power value that reaches the preset reliability threshold is then output as the base power value.

4. The method according to claim 3, characterized in that, The discriminator receives candidate power values ​​and constructs a virtual transmission scenario. Within this scenario, it simulates the signal propagation process of the candidate power values ​​and calculates the received power of the signal reaching the BeiDou satellite. Based on the received power and a preset BeiDou satellite receiver sensitivity threshold, it calculates the power margin and generates a transmission reliability score, including: The discriminator receives candidate power values, extracts the received signal strength component and signal-to-noise ratio component from the channel feature vector, and calculates the multipath propagation delay spread parameter and Doppler frequency shift parameter as multipath fading parameters based on the fading amplitude of the received signal strength component and the noise intensity of the signal-to-noise ratio component. The time-domain fading distribution in the virtual transmission scenario is determined based on the multipath propagation delay spread parameter, and the frequency-domain fading distribution in the virtual transmission scenario is determined based on the Doppler frequency shift parameter. A virtual transmission scenario with time-frequency fading characteristics is constructed, and the candidate power value is used as the transmit power to start the signal propagation process simulation in the virtual transmission scenario. Based on the spatial distance between the BeiDou short message terminal and the BeiDou satellite, free space path attenuation is applied to the candidate power value. Based on the time domain fading distribution and frequency domain fading distribution, multipath fading attenuation is applied to the power after free space path attenuation to obtain the received power of the signal reaching the BeiDou satellite. Obtain the preset BeiDou satellite receiving sensitivity threshold, calculate the power margin between the received power and the preset BeiDou satellite receiving sensitivity threshold, and generate a transmission reliability score by weighting the power margin and node activation degree according to the preset weight coefficient.

5. The method according to claim 1, characterized in that, A directed acyclic causal graph is constructed from the channel feature vectors. The causal effects of the channel fading factor and interference factor are separated through intervention operations. Based on these causal effects, the target power value is obtained by counterfactual inference compensation of the baseline power value, including: The received signal strength component, signal-to-noise ratio component, Doppler frequency shift component, and time delay spread component are extracted from the channel feature vector. The received signal strength component and signal-to-noise ratio component are used as observation variable nodes, and the Doppler frequency shift component and time delay spread component are used as factor variable nodes. A directed edge connection is established based on the influence relationship between the observation variable nodes and the factor variable nodes to construct a directed acyclic causal graph. The value of the Doppler frequency shift component is fixed and the change amplitude of the conditional probability distribution of the observed variable node is calculated as the causal effect of the channel fading factor. The value of the delay spread component is fixed and the change amplitude of the conditional probability distribution of the observed variable node is calculated as the causal effect of the interference factor. The causal effect of the channel fading factor is set to the maximum fading level, and the fading counterfactual value of the received signal strength component is derived in reverse. The causal effect of the interference factor is set to the maximum interference level, and the interference counterfactual value of the signal-to-noise ratio is derived in reverse. The fading compensation amount is obtained by calculating the difference between the fading counterfactual value and the actual value of the received signal strength component. The interference compensation amount is obtained by calculating the difference between the interference counterfactual value and the actual value of the signal-to-noise ratio component. The fading compensation amount and the interference compensation amount are added to the base power value to obtain the target power value.

6. The method according to claim 5, characterized in that, The causal effect of fixing the value of the Doppler frequency shift component and calculating the change in the conditional probability distribution of the observed variable nodes as the causal effect of the channel fading factor, and fixing the value of the delay spread component and calculating the change in the conditional probability distribution of the observed variable nodes as the causal effect of the interference factor, include: Construct a joint probability distribution of Doppler frequency shift components and time delay spread components in historical transmission records, and sample from the joint probability distribution to generate an intervention sample set containing multiple sets of numerical combinations; The Doppler frequency shift component values ​​in the intervention sample set are sequentially fixed to the factor variable nodes corresponding to the Doppler frequency shift components in the directed acyclic causal graph, and the input edges of the factor variable nodes are cut off. The mean of the conditional probability distribution of the observation variable nodes corresponding to the received signal strength components after each fixation is calculated, and the standard deviation of the mean of all conditional probability distributions is used as the causal effect of the channel fading factor. The time delay spread component values ​​in the intervention sample set are sequentially fixed to the factor variable nodes corresponding to the time delay spread components in the directed acyclic causal graph, and the input edges of the factor variable nodes are cut off. The mean of the conditional probability distribution of the observation variable nodes corresponding to the signal-to-noise ratio component after each fixation is calculated, and the standard deviation of the mean of all conditional probability distributions is used as the causal effect of the interference factor.

7. The method according to claim 1, characterized in that, The power loss difference and channel feature vector are combined as observational evidence. The Bayesian posterior probability of the active node's neighboring nodes is updated using kernel density estimation. The node with the highest posterior probability is selected as the reconstructed node identifier. The target power value is regenerated based on the reconstructed node identifier, and the message is retransmitted, including: The power loss difference is combined with the channel feature vector to form an observation evidence vector that contains both power dimension information and channel dimension information. In a directed acyclic causal graph, all adjacent connected nodes of the node corresponding to the activated node are extracted as a neighborhood node set. For each node in the neighborhood node set, the observed evidence vector is used as the sample input to construct a kernel density estimation function. The probability density value of the kernel density estimation function at the node position is calculated as the likelihood probability. The likelihood probability is multiplied by the prior probability of the node according to the Bayes formula and then normalized to obtain the Bayes posterior probability of the node. Traverse the Bayesian posterior probabilities of all nodes in the neighborhood node set and select the node with the largest value as the reconstruction node identifier. Extract the causal effect of the channel fading factor and the causal effect of the interference factor corresponding to the reconstruction node identifier. Based on the causal effect of the channel fading factor and the causal effect of the interference factor, perform counterfactual reasoning compensation on the base power value to regenerate the target power value. Use the regenerated target power value to send the BeiDou short message again.

8. An AI-based BeiDou short message adaptive power control system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The channel feature unit is used to collect the channel parameters of the Beidou short message terminal, generate the channel feature vector, perform competitive mapping of the channel feature vector in the self-organizing topology space, and output the active node identifier and node activation degree. The power optimization unit is used to extract seed power values ​​based on the active node identifier, and to generate candidate power values ​​iteratively based on the seed power values ​​and node activation by constructing an adversarial generative network until the transmission reliability score output by the discriminator reaches a preset reliability threshold. The optimized candidate power values ​​are then output as the base power values. The causal effect unit is used to construct a directed acyclic causal graph from the channel feature vector. It separates the causal effect quantities of the channel fading factor and the interference factor through intervention operations, and obtains the target power value by counterfactual reasoning compensation based on the causal effect quantities of the base power value. The power adjustment unit is used to adjust the transmission power according to the target power value and send messages, and to obtain the transmission status identifier and the actual received power. The failure retransmission unit is used to calculate the power loss difference between the target power value and the actual received power when the transmission status is identified as failure. The power loss difference is combined with the channel feature vector as observation evidence. The Bayesian posterior probability of the neighboring nodes of the active node is updated by kernel density estimation. The node with the largest posterior probability is selected as the reconstruction node identifier. The target power value is regenerated based on the reconstruction node identifier and the message is sent again.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images