A tactical communication intent translation method based on physically driven bimodal bayesian inference

Abstract

The application discloses a tactical communication intention translation method based on physical driving bimodal Bayesian inference, and belongs to the technical field of information engineering and artificial intelligence fusion. The application constructs a joint feature of tactical communication intention and battlefield environment, respectively establishes semantic guidance physical prior, historical case data likelihood and communication physical hard constraint feasible region; through precision weighted Bayesian fusion, physical constraint truncated projection inhibits large model illusion and guarantees KPI physical compliance, and then constructs normal work and link interruption bimodal mixed distribution, depicts KPI bimodal characteristics and long tail risk; finally, the KPI probability result containing expected value, confidence interval, default probability and comprehensive risk score is output. The application converts the natural language operation intention of commanders into a physical feasible probabilistic KPI index, quantifies cognitive and environmental random uncertainty, and can meet the high reliability requirement of strong confrontation, high dynamic tactical communication network command decision and communication configuration.

Description

Technical Field

[0001] This invention belongs to the field of information engineering and artificial intelligence integration application, specifically involving a tactical communication intent translation method based on physical-driven bimodal Bayesian inference. Background Technology

[0002] In modern battlefield confrontation, the speed of operational command and decision-making directly determines the efficiency of the OODA (Observe-Orient-Decide-Action) cycle suppression. Intent-Based Networking (IBN), as a core enabling technology for next-generation autonomous networks, aims to automatically translate high-level user intentions into network configuration strategies, and extensive research has been conducted both domestically and internationally. In tactical communication scenarios, intent translation is a crucial intermediate layer connecting commanders' natural language operational intentions with the execution configuration of tactical communication networks. Its core task is to transform ambiguous operational intentions into quantifiable key performance indicators (KPIs), which is the core link in achieving accurate mapping from operational intentions to communication capabilities.

[0003] Existing operational intent translation and network KPI inference technologies mainly include: intent element extraction and compilation methods based on named entity recognition, intent decomposition and mapping methods driven by large language models, multi-agent intent recognition frameworks, service orchestration methods driven by reinforcement learning, and knowledge fusion methods based on retrieval-enhanced generation (RAG). However, most of these technologies are geared towards stable civilian network scenarios and cannot adapt to complex scenarios such as high link interruption rates in tactical communications, limited coverage of historical case data, stringent physical constraints on communications, and dynamic and ever-changing battlefield electromagnetic interference. They suffer from three major technical shortcomings:

[0004] Without the ability to quantify risk, only outputting deterministic point estimates: Existing methods all map natural language intent to fixed configuration parameters or single KPI point estimates, which cannot quantify the cognitive uncertainty of model inference, nor can they characterize the random uncertainty brought about by battlefield electromagnetic interference and terrain fading; they cannot answer the core decision-making questions of "whether the combat intent can be achieved and what the probability of achieving it is", and tactical decisions lack risk quantification basis.

[0005] Lacking physical constraints, the risk of large-scale model illusion is prominent: the intent translation method based on deep learning and large language models is a black-box inference mode, which does not embed the constraints of communication physical laws. It is easy to output KPI prediction values ​​that exceed the Shannon-Hartley theorem and the theoretical upper limit of channel capacity. When facing tasks related to physical laws, the prediction error can reach several times, and the generated results have the fundamental risk of physical infeasibility, which seriously affects the reliability of tactical communication.

[0006] The fusion of multi-source knowledge is superficial and lacks a unified probabilistic inference framework: Intent translation requires the integration of three types of heterogeneous knowledge: semantic understanding, historical combat cases, and communication physical laws. Existing research only focuses on single directions such as text retrieval enhancement, model output fusion, and knowledge graph modeling, without incorporating physical channel priors, historical case likelihoods, and physical constraints into the same Bayesian probabilistic framework for fusion. In new combat missions and extreme battlefield environments, the system is prone to failure and lacks generalization ability.

[0007] In summary, existing methods have significant shortcomings in risk quantification, physical compliance, and multi-source knowledge fusion, and cannot meet the reliable decision-making requirements of highly dynamic and highly adversarial tactical communication scenarios. Summary of the Invention

[0008] This invention addresses the problems existing in the prior art by providing a tactical communication intent translation method based on physical-driven bimodal Bayesian inference. The method of this invention can transform the commander's ambiguous natural language instructions into a KPI probability distribution that is physically feasible and includes expectation and confidence intervals. With the achievable bandwidth of the link and the end-to-end latency as the core prediction targets, it can characterize bimodal and long-tail risks, thereby providing a reliable risk quantification basis for tactical decision-making.

[0009] To address the above technical problems, this invention provides the following technical solution: a tactical communication intent translation method based on physics-driven bimodal Bayesian inference, comprising the following steps:

[0010] S1. Construct a basic model for a tactical communication intent translation system, including intent parsing, environmental perception, and joint intent-environment feature construction. While receiving the natural language combat mission description input by the commander, it collects real-time battlefield environment features, including communication link distance, electromagnetic interference level, terrain type, and core situational parameters. Through structured parsing and feature encoding, it completes the joint feature representation of combat intent semantic information and battlefield environment situational information.

[0011] S2. Based on the aforementioned intention-environment joint features, complete independent modeling of three types of heterogeneous knowledge: semantic guidance physical prior, historical case data likelihood, and hard constraints of communication physical laws.

[0012] S3. A physical-driven bimodal Bayesian inference algorithm is adopted to achieve unified probabilistic fusion of multi-source knowledge through three progressive inference layers: The first layer performs precision-weighted Bayesian fusion to generate a Gaussian posterior distribution; the second layer performs truncated projection on the Gaussian posterior distribution based on the feasible region constrained by the physical laws of communication to obtain a truncated normal distribution; the third layer estimates the link interruption probability by combining battlefield environment characteristics to construct a bimodal hybrid distribution of normal working state and link interruption state.

[0013] S4. Based on the dual-modal hybrid distribution, calculate the comprehensive expected value, confidence interval, business default probability and comprehensive risk score of the network's key performance indicators (KPIs), and output the tactical communication intent translation results containing probability distribution and risk quantification, providing a basis for tactical communication network configuration and command decision-making.

[0014] Furthermore, the semantic-guided physical prior distribution modeling in step S2 specifically includes:

[0015] A large language model is used to classify the constraint types of combat intentions and output the classification confidence scores; combined with real-time battlefield environment parameters, the expected values ​​of KPIs are calculated based on the communication channel propagation model and channel capacity theory; and a semantic-guided physical prior Gaussian distribution is constructed by adaptively scaling the prior standard deviation based on the classification confidence scores.

[0016] When there is a contradiction between the classification results of the large language model and the physical characteristics of communication distance, the prior standard deviation is magnified by a factor; when the classification confidence is lower, the prior standard deviation is nonlinearly amplified to quantify the cognitive uncertainty of the model.

[0017] Furthermore, the likelihood distribution modeling of historical case data in step S2 mentioned above specifically includes:

[0018] Construct a historical tactical communication case library containing operational intent text, environmental features, actual KPI values, and link connectivity status; employ the FAISS vector retrieval framework, integrating text semantic similarity and environmental feature similarity, to retrieve Top-K historical cases matching the current scenario; statistically analyze the KPI mean and discrete variance of the retrieved cases to construct a Gaussian distribution of data likelihood.

[0019] When the highest semantic similarity of the retrieved case is lower than the preset threshold, it is judged as an extreme unknown scenario with zero samples, and the data likelihood precision is set to zero. The posterior distribution completely degenerates into a semantically guided physical prior distribution.

[0020] Furthermore, the feasible region modeling of the hard constraints of communication physical laws in the aforementioned step S2 specifically includes:

[0021] A mapping table between tactical communication constraint types and KPI physical feasible domains is pre-established, wherein the KPIs include link reachable bandwidth and end-to-end delay. Based on the Shannon-Hartley theorem, 3GPP channel propagation theory, and typical tactical communication service coding delay specifications, the upper and lower bounds of the bandwidth feasible domain, the physical feasible domain of delay, and the service decision threshold corresponding to each constraint type are determined as hard boundaries for subsequent distributed compliance verification.

[0022] Furthermore, the aforementioned step S3, the first-layer precision-weighted Bayesian fusion, includes the following steps:

[0023] 3.1 Precision Differentiation Configuration: Set a prior precision enhancement factor higher than the bandwidth KPI for the latency KPI, and calibrate the physical model's measured precision for the prior standard deviation of latency;

[0024] 3.2 Prior-Data Conflict Adaptive Decay: When the deviation between the prior mean and the data mean exceeds twice the data standard deviation, the prior accuracy is adaptively decayed according to the standardized conflict metric to suppress the prior bias caused by misclassification of large models.

[0025] 3.3 Power-prior dynamic weight adjustment: The prior discount factor is calculated based on the historical case retrieval quality. When the data case coverage is high, the data likelihood dominates the fusion. When the data is sparse, the semantic guidance and physical prior are the last resort.

[0026] 3.4 Environmental Interference Perception Scaling: After fusing and generating a Gaussian posterior distribution, the posterior standard deviation is scaled according to the difference in electromagnetic interference level to match the random uncertainty change law of the battlefield environment;

[0027] Furthermore, the aforementioned step S3, the second-layer physical constraint truncation projection, specifically includes:

[0028] Using the feasible region constrained by the physical laws of communication as the boundary, the mean and variance of the Gaussian posterior distribution are solved by a truncated normal closed-form solution. If the posterior mean exceeds the boundary of the physical feasible region, the mean is automatically pulled back into the feasible region, and the variance is adjusted synchronously to represent additional uncertainty. Adaptive observation noise linked to electromagnetic interference level and KPI type is superimposed, and a probability distribution calibration is performed using a KPI-interference level joint differential scaling coefficient to ensure that the predicted standard deviation is accurately aligned with the root mean square of the actual prediction error, thus suppressing the physical infeasibility prediction caused by the illusion of a large model.

[0029] Furthermore, the aforementioned step S3, modeling the third-layer bimodal hybrid distribution, specifically includes:

[0030] The truncated normal distribution after projection is used as the probability distribution of the normal working state of the link; the Dirac function is used to approximate the probability distribution of the link outage failure state, with the bandwidth outage value set to 0 Mbps and the delay outage value set to the preset maximum degradation delay;

[0031] The probability of link interruption is physically modeled by three-dimensional factors including electromagnetic interference level, communication distance, and terrain type. Based on the interruption probability, the normal operation state and the interruption failure state are probabilistically weighted and mixed to construct a dual-modal mixed distribution, which characterizes the bimodal distribution characteristics of KPI and the long-tail failure risk on the battlefield.

[0032] Furthermore, the calculation method for the business default probability in step S4 above is as follows:

[0033] For bandwidth KPIs where larger is better, the probability of default is the cumulative probability that the actual value of the KPI is lower than the minimum threshold of the business; for latency KPIs where smaller is better, the probability of default is the cumulative probability that the actual value of the KPI is higher than the maximum threshold of the business.

[0034] The comprehensive risk score is constructed by weighting the probability of default and the probability of link interruption. The weighting coefficients can be adaptively configured according to the sensitivity of the combat mission to business default and link interruption.

[0035] The present invention also provides a tactical communication intent translation system based on physical-driven bimodal Bayesian inference, which is applied to the aforementioned tactical communication intent translation method based on physical-driven bimodal Bayesian inference, including an input layer, a perception layer, a decision layer, and an output layer;

[0036] The input layer is used to receive natural language combat intent commands and real-time battlefield environment feature parameters;

[0037] The perception layer is used to complete the encoding of combat intent features, independent modeling of multi-source heterogeneous knowledge, and to generate semantically guided physical priors, historical case data likelihood and communication physical hard constraint feasible domains.

[0038] The decision layer incorporates a precision-weighted Bayesian fusion module, a physical constraint truncation projection module, and a dual-modal hybrid distribution modeling module to achieve three-layer progressive Bayesian probabilistic inference.

[0039] The output layer is used to encapsulate KPI probability distribution parameters, confidence intervals, default probabilities, and comprehensive risk scores, and output the quantified results of tactical communication intent translation.

[0040] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the aforementioned tactical communication intent translation method based on physical-driven bimodal Bayesian inference.

[0041] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0042] This invention addresses the core need for tactical communication intent translation in complex battlefield environments. Based on the commander's natural language combat intent and real-time battlefield situation information, it employs a physical-driven bimodal Bayesian inference algorithm (PDBBI) to form a complete strategy for probabilistic mapping and risk quantification of combat intent to network KPIs.

[0043] Simulation results demonstrate that the algorithm proposed in this invention effectively addresses the core issues of existing intent translation methods, such as the lack of risk quantification capabilities, significant large-model illusion risks, and insufficient depth of multi-source knowledge fusion. In complex tactical scenarios with varying levels of electromagnetic interference, terrain environments, and combat mission types, its performance significantly outperforms existing deterministic intent translation algorithms, exhibiting more accurate risk characterization, higher physical compliance, and stronger cross-scenario generalization capabilities. Unlike existing black-box deep learning or large language model methods, the inference process in this invention is constructed based on a physical channel model and a Bayesian statistical framework. The inputs, outputs, and calculation rules for each step all have clear mathematical expressions, possessing complete physical interpretability. Attached Figure Description

[0044] Figure 1 This is the overall architecture diagram of the PDBBI tactical communication intent translation system;

[0045] Figure 2 This is an overall flowchart of a tactical communication intent translation method based on physics-driven bimodal Bayesian inference;

[0046] Figure 3 This is a flowchart of the three-layer inference process of PDBBI;

[0047] Figure 4 This is a schematic diagram of precision-weighted Bayesian fusion;

[0048] Figure 5 This is a schematic diagram of the physical constraint truncation projection mechanism;

[0049] Figure 6 This is a diagram of the theoretical model of the bimodal mixture probability distribution;

[0050] Figure 7 This is a radar comparison chart of five comparison methods for multidimensional KPI forecasting indicators;

[0051] Figure 8 This is a comparison chart of the MAE of the proposed method and the KNN baseline under different data sparsity conditions; Detailed Implementation

[0052] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0053] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0054] Before implementing this invention, system initialization preparations are first completed. A large language model (Qwen3-4B in this embodiment, which excels at Chinese tactical intent understanding and constraint type classification) is deployed on an edge server using the Ollam framework for subsequent intent constraint type classification and confidence assessment. Six types of tactical communication link cases based on real battlefield data are collected, with each record containing combat intent text (natural language) and environmental features. ={distance_km, interference_level, terrain_type}, actual KPI values ​​{bandwidth_mbps, latency_ms}, link connectivity status is_connected (0=interrupted, 1=connected), and constraint type identifier. The above six link constraint types are designed cross-dimensionally based on two dimensions: communication distance (short-range / long-range / hybrid) and service type (command / command / data feedback / coordinated operations). They cover typical tactical communication scenarios ranging from platoon-level short-range reconnaissance to battalion-level long-range command, from low-bandwidth command issuance to high-bandwidth video feedback, and from routine tasks to time-sensitive strike coordination. Their KPI requirements differ significantly (bandwidth thresholds range from 0.05 to 2.0 Mbps, latency thresholds range from 20 to 1000 ms), making them highly representative. The six link constraint types and their corresponding parameters are shown in Table 1. The intent text of the case was encoded using the BGE-Large-zh-v1.5 embedding model (1024 dimensions), and a FAISS vector index was constructed, supporting sub-second semantic similarity retrieval. Initialize the three-layer Bayesian fusion parameters: prior decay exponent =1.0 (preferred value, range [0.5, 2.0]), environmental feature retrieval weight =0.3 (preferred value, range [0.1, 0.5]), data retrieval Top-K=5 (preferred value, range [3, 10]).

[0055] Table 1 shows the parameter table for communication system models facing complex time-varying situations: the bandwidth feasible region column corresponds to the physical feasible region of bandwidth for each constraint type. In addition to time-sensitive strike coordination types, bandwidth service thresholds Take the physical upper bound for each constraint type Time-sensitive strike coordination prioritizes small strike information message volume and ultra-low latency, and the bandwidth service threshold is taken as the upper limit of the operation. =1.0 Mbps. The value is determined based on measured statistical data from typical tactical communication scenarios. The lower bound of the bandwidth for video backhaul links (recon_video_short) is 0.5 Mbps, set according to the lowest available bitrate for H.264 video encoding. The physical feasible region for latency is uniformly set to [1ms, 1000ms] for all constraint types.

[0056] Table 1

[0057] recon_cmd_short Short-range reconnaissance order issued 0.02~0.05 100 recon_video_short Short-range video image return 0.5~2.0 1000 bn_cmd_long Battalion-level long-range command orders 0.02~2.0 300 bn_info_long Long-range battlefield information reporting 0.02~2.0 1000 strike_timesensitive Time-sensitive strike coordination 0.02~2.0 20 combat_coord_mixed Squad-level combat coordination 0.02~2.0 300

[0058] The overall system architecture is as follows Figure 1 As shown, it consists of four parts: an input layer, a perception layer, a decision layer, and an output layer. The input layer receives natural language operational intentions and battlefield environment characteristics; the perception layer is responsible for independent modeling of multi-source heterogeneous knowledge, generating semantically guided physical priors and likelihoods from historical case data; the decision layer uses a three-layer progressive inference process—precision-weighted Bayesian fusion, physically constrained truncated projection, and bimodal hybrid distribution modeling—to sequentially complete multi-source knowledge fusion, physical compliance verification, and long-tail risk quantification; the output layer encapsulates KPI probability distributions and comprehensive risk quantification results, providing a basis for tactical decision-making.

[0059] like Figure 2 The overall flowchart of the present invention shown below illustrates the complete deduction process using a specific tactical mission as an example: The steps are as follows:

[0060] Step 1: Receive the commander's natural language operational intent. The system receives the natural language operational mission description input by the commander as the raw input for intent translation.

[0061] Step 2: Collect real-time battlefield environment characteristics. The system synchronously collects real-time battlefield environment characteristics, including core situational parameters such as communication link distance, electromagnetic interference level, and terrain type, providing environmental input for subsequent historical case retrieval, physical feasible domain determination, and link interruption probability estimation.

[0062] Step 3: LLM constraint type classification and confidence assessment. The system enhances constraint generation by combining retrieval from the tactical communication knowledge base. The large language model classifies the constraint types of operational intentions and outputs the model's self-assessed confidence level for the classification results.

[0063] Step 4, Semantic-guided physical prior distribution modeling. Based on the constraint type classification results of Step 3, using the classified constraint type as an index, and combining battlefield environment parameters (distance, interference level, terrain type), the link SINR is derived based on the 3GPP TR 38.901 logarithmic distance path loss model: the bandwidth prior mean is calculated via SINR→CQI→spectral efficiency chain; the delay prior mean is calculated via SINR→BLER→HARQ expected retransmission rounds (3GPP TS 38.213)→air interface delay, and then combined with M / D / 1 queuing delay, protocol processing delay, and propagation delay. The prior standard deviation is calculated based on the static weak prior corresponding to the constraint type. Based on this, initial scaling is first performed according to the physical consistency between the classification results and the communication distance (when the classification and distance match). Halving, contradiction (Doubled), and then further amplified using a confidence-adaptive scaling formula, to achieve a quantitative representation of the cognitive uncertainty in model reasoning and construct a prior probability distribution. The higher the confidence level, The closer (The stronger the prior knowledge); the confidence level approaches zero. Approaching 4 (Prior information tends to no information).

[0064] Step 5: Historical Case RAG Hybrid Retrieval and Likelihood Estimation. From the tactical communication historical case database, the top-K historical cases that best match the current operational intent are retrieved using a hybrid textual semantic similarity and environmental feature similarity method. The mean and variance of the KPIs of the retrieved cases are statistically analyzed to construct a data likelihood distribution. When the highest similarity score of a retrieved case is lower than a preset threshold, it is determined to be a zero-sample scenario. The system automatically triggers a degradation strategy, reducing the data likelihood precision to zero, and the posterior completely degenerates to a semantically guided physical prior.

[0065] Step 6: Determining the Feasible Domain of Communication Physical Laws. Based on the tactical communication service specifications, the physical feasible domains of each KPI are pre-determined using constraint types as indexes (see Table 1). This step retrieves the corresponding bandwidth feasible domains based on the constraint types obtained in Step 3. With business decision threshold .

[0066] Step 7, First-layer precision-weighted Bayesian fusion. This layer takes the semantic-guided physical prior from Step 4 and the data likelihood from Step 5 as input, and completes precision-weighted fusion through the following five mechanisms to generate a Gaussian posterior distribution:

[0067] (1) Improved KPI differentiation accuracy: Set higher precision weights for delay priors, since the physical channel model has better prediction accuracy for delay than text retrieval; at the same time, calibrate the standard deviation of delay priors to a fixed value that matches the actual prediction accuracy of the physical model, replacing the conservative setting in the original weak priors. Value. (2) Prior and data conflict adaptive decay: When the deviation between the prior mean and the data mean exceeds twice the data standard deviation, the accuracy improvement factor is smoothly decayed according to formula (11) to suppress the prior bias caused by LLM misclassification. (3) Power Prior discount: Based on retrieval quality The prior discount factor is calculated based on the retrieval quality. When the retrieval quality is high, the data dominates, and when the retrieval quality is low, the prior dominates. (4) Physical consistency accuracy penalty: When the classification and distance contradict each other, the prior accuracy enhancement factor is reduced to 1 / 4 of the original value, and the data layer takes over the correction. (5) Interference level perception Scaling: After fusion, the posterior standard deviation is scaled differentially according to the interference level to match the variation of channel physical uncertainty with interference intensity. Fusion demonstration Figure 4For a detailed deduction process, see Figure 3 .

[0068] Step 8, Second-layer physical constraint truncated projection. This layer takes the Gaussian posterior from step 7 as input and performs a truncated projection on the posterior distribution based on the physical feasible region determined in step 6, ensuring that the output fully conforms to the laws of communication physics. (See diagram of the truncation mechanism.) Figure 5 For a detailed deduction process, see Figure 3 .

[0069] Step 9, Modeling the Third-Layer Bimodal Hybrid Distribution. This layer is based on the truncated normal distribution from Step 8, combined with the link interruption probability estimated from the battlefield environment characteristics in Step 2, to construct a bimodal hybrid distribution of the normal operation state and the link interruption state. (See diagram of the bimodal distribution.) Figure 6 For a detailed deduction process, see Figure 3 .

[0070] Step 10: Output the KPI probability distribution, default probability, and overall risk score. The system calculates the default probability and overall risk score for each KPI based on a bimodal mixed distribution, encapsulating the probabilistic information of multi-dimensional KPIs into a unified KPI probability state vector, which serves as the risk quantification basis for tactical decision-making.

[0071] like Figure 3 The diagram shown is a flowchart of the three-layer inference process of the PDBBI of this invention. Figure 2 The four pieces of knowledge obtained in steps 3-6 serve as input, and the detailed inference process is as follows:

[0072] Step 1: Input semantic-guided physical prior parameters. Receive semantic-guided physical prior distribution parameters driven by intent classification. , and LLM self-assessment confidence level For each KPI Semantic-guided physical prior modeling is based on a Gaussian distribution:

[0073] (1)

[0074] in, The SINR is derived from the 3GPP TR 38.901 logarithmic distance path loss model based on the intent classification results. Then, the KPI prior estimates are dynamically calculated using the CQI-spectral efficiency chain (bandwidth) and the BLER-HARQ-multi-component delay chain (delay), respectively. This represents the prior uncertainty of the KPI corresponding to the constraint type, and its value will be adaptively scaled according to the LLM classification confidence level in subsequent steps. For the first The actual observed value of each KPI It is a joint input feature vector that includes semantic information of combat intent and characteristics of the battlefield environment.

[0075] Step 2, Input Data Likelihood Parameters. Receive the data likelihood distribution parameters obtained from historical case retrieval. , and search quality The data likelihood is modeled as a Gaussian distribution:

[0076] (2)

[0077] in, To retrieve the average KPI of the cases, This represents the variance of the case KPIs, reflecting the aleatoric uncertainty of the battlefield environment. It's important to note that when the highest semantic similarity scores of the Top-K search cases are all below a preset threshold... (Preferred value) When the value ranges from [0.3, 0.5], it is determined that the current situation is an extremely unknown scenario with zero samples (Zero-Shot). In this case, let... The posterior completely degenerates into a semantically guided physical prior distribution.

[0078] Step 3, input the physical feasible region parameters. Receive the KPI physical feasible region determined by the physical laws of communication. With business decision threshold :

[0079] (3)

[0080] in, The physical lower bound of KPI. The physical upper bound of KPI. This is the threshold for business decisions.

[0081] Step 4: Input environmental characteristic parameters. Receive battlefield environmental characteristics (electromagnetic interference level, communication distance, terrain type) for subsequent link interruption probability. Physical modeling.

[0082] Step 5: Confidence-Driven Prior Adjustment. Existing methods typically use fixed prior standard deviations or linear confidence weighting, which cannot effectively suppress the overconfidence of LLMs in low-confidence scenarios. This invention proposes using LLM self-assessment confidence. The exponential scaling mechanism for the independent variable performs nonlinear dynamic adjustment of the prior standard deviation:

[0083] (4)

[0084] in, The adjusted prior standard deviation is obtained after scaling the LLM self-assessment confidence level (conf) index. This is the static basic prior standard deviation corresponding to the constraint type.

[0085] when At that time, the scaling factor is 4 0 =1.0, a narrow prior distribution indicates that the model is highly confident in the classification result; when At this point, the scaling factor is 4¹=4.0, and the prior standard deviation is magnified to 4 times, significantly increasing cognitive uncertainty. Compared with linear scaling, exponential decay provides a stronger penalty in the low confidence interval, effectively suppressing the undesirable bias of overconfident priors on the posterior during LLM misclassification.

[0086] Step 6, calculate the prior precision. Convert the adjusted prior standard deviation to precision (the reciprocal of the variance):

[0087] (5)

[0088] in, For the first The prior precision of each KPI is equal to the adjusted prior variance. The reciprocal of the precision value indicates that the prior knowledge is more certain, and thus the greater the fusion weight obtained in precision-weighted Bayesian fusion.

[0089] Step 7: Calculate data precision and calibrate data layer precision. Existing RAG retrieval methods directly use the variance of historical case KPIs as the likelihood distribution parameter, which leads to severe overconfidence problems due to retrieval quality degradation in strong interference scenarios. This invention proposes a data layer standard deviation lower limit constraint mechanism that combines KPI and interference level differentiation, which differs from existing fixed lower limits or globally uniform lower limits:

[0090] (6)

[0091] (7)

[0092] in, The standard deviation of the calibration data after applying the lower limit constraint of the interference level difference. For the corresponding data precision, The standard deviation of the raw data is directly obtained from the statistical variance of the KPIs of the Top-K historical cases. This is the lower limit of the standard deviation of the data layer, differentiated according to interference level. In this invention, the bandwidth does not have a lower limit across all interference levels. The lower limit of latency increases with the level of interference: ms (corresponding to Lv0 / Lv1 / Lv2 / Lv3). The basis for this differentiated design is as follows: In the interference-free environment of Lv0, KNN retrieval accuracy is high (data layer prediction is reliable), and no lower limit is set to avoid artificial inflation. This leads to an increase in NLL (Network Limiting). As the interference level increases, the KNN retrieval quality decreases due to channel state fluctuations, and the latency is more significantly affected by random factors such as instantaneous channel state (HARQ retransmission count, scheduling delay, queue depth). Therefore, an incremental lower limit needs to be set to prevent the data layer from becoming overconfident. The Lv3 lower limit of 15.0 ms is determined by referencing the B2 (pure KNN) latency MAE of 31.61 ms in the experiment, and approximately MAE / 2 is taken as the minimum uncertainty estimate under strong interference scenarios.

[0093] Step 8: Prior Accuracy Calibration and KPI Differentiation Accuracy Enhancement. This step includes two innovative mechanisms: prior standard deviation calibration and KPI differentiation prior accuracy enhancement factor. .

[0094] (1) Prior standard deviation calibration. This invention implements an adaptive calibration strategy for the prior accuracy of KPIs: using the measured accuracy of the physical model as a benchmark, when the scene is static prior... When the model is on the same order of magnitude as the physical model MAE, confidence scaling is used. Fine-tuning is performed; when the scene has a static prior... When the deviation from the physical model accuracy is significant (more than 2 times), the physical calibration value is directly used to replace it, avoiding invalid scaling on an incorrect baseline. Static priors in SCENE_PRIORS The settings were conservatively chosen to address LLM misclassification. However, the physical channel model (3GPP TR 38.901 logarithmic distance path loss → SINR → CQI / BLER → HARQ retransmission → M / D / 1 queuing + 3GPP TS 38.213 delay model) showed significantly higher actual delay prediction accuracy than the static model. The implied level. In the experiment, the B3 (pure LLM) delay MAE = 22.65 ms, corresponding to the actual prediction error standard deviation. ms is much smaller than static (50~200 ms). Therefore, the prior standard deviation is calibrated and replaced within the fusion layer:

[0095] (8)

[0096] in, The prior standard deviation after internal calibration of the fusion layer; This refers to the physical calibration replacement value set for the k-th KPI. This is a fixed replacement value for the prior standard deviation calibrated based on the measured accuracy of the physical channel model, and is only enabled for KPIs where the prediction accuracy of the physical model is significantly better than that of the static weak prior. In the example, =25.0 ms (preferred value, range [20.0, 35.0]), bandwidth without calibration. This value is calculated based on B3 (pure LLM) delay MAE = 22.65ms: According to the semi-normal distribution theory, when the prediction error follows a zero-mean normal distribution... hour, , reverse reasoning ms; rounded and with a certain conservative margin, it is set to 25.0 ms (slightly less than the theoretical value, so that the prior receives a slightly higher weight in the fusion than the theoretical accuracy, automatically preventing overconfidence through the conflict decay mechanism). This calibration matches the prior accuracy with its actual predictive ability, avoiding excessively conservative... This leads to the underestimation of high-quality priors during fusion.

[0097] (2) Improved Prior Precision for Differentiated KPIs. Based on the differences in the physical modelability of different KPIs, a prior precision improvement factor for differentiated KPIs is introduced. :

[0098] (9)

[0099] in, This is an enhanced prior accuracy scaled by the KPI differentiation precision enhancement factor, used to replace the original prior accuracy. Participate in subsequent integration; For the first Differentiated prior accuracy enhancement factor for each KPI, subscript Corresponding bandwidth KPI, subscript Corresponding latency KPI; =1.0 (Bandwidth: Data retrieval is more reliable, prior accuracy remains at the standard level without improvement). =15.0 (Latency: The prediction accuracy of the physical channel model is significantly better than that of data retrieval, thus increasing the weight of prior accuracy). The standard deviation of the latency KPI is the data layer standard deviation, typically ranging from 15 to 30 ms, and is derived from the latency variance statistics of Top-K retrieval cases. The value is determined by the ratio of prior knowledge to data accuracy calibration: in =25 ms, typical Under the condition of 15~30 ms, =15 makes the prior effective precision =15 / (25²)=0.024 and data precision =1 / (20²)=0.0025 are on the same order of magnitude, ensuring that both paths of knowledge make substantial contributions to the fusion. The physical basis for the >1 is that latency is mainly determined by deterministic physical processes (air interface propagation delay + protocol processing delay + HARQ retransmission), and the physical channel model can accurately model its main causes (in the experiment, the B3 pure LLM latency MAE = 22.65 ms < the B2 pure KNN latency MAE = 31.61 ms, verifying that physical prior is superior to data retrieval), while bandwidth is greatly affected by random factors (fast fading, multi-user scheduling contention), and the physical model can only give a rough theoretical upper limit (B3 bandwidth MAE = 0.612 > B2 bandwidth MAE = 0.340, verifying that data retrieval is superior to physical prior), therefore, =1.0 No improvement.

[0100] (3) Adaptive attenuation of prior-data conflict. Existing Bayesian fusion methods still forcibly fuse priors and likelihoods with fixed weights when there is a severe conflict, leading to uncontrollable bias of the prior on the posterior in LLM misclassification. This invention proposes an adaptive attenuation mechanism for prior accuracy based on standardized conflict metrics to automatically suppress the influence of priors in misclassification scenarios:

[0101] (10)

[0102] (11)

[0103] in, This is a standardized conflict measure between the prior and the data, representing the ratio of the deviation of the prior mean from the data mean to the data standard deviation. The larger the value, the more significant the conflict. For the first The semantically guided physical prior mean of each KPI (i.e., in step 1) ); For the first Historical case data likelihood mean of KPIs (i.e., in step 2) ); This is the effective accuracy improvement factor after conflict adaptive decay; the greater the conflict between prior knowledge and data, the better. The smaller; The original value of the KPI differential prior accuracy enhancement factor defined in step 8(2).

[0104] The selection of the attenuation threshold z=2 is based on the statistical principle of 2. Criterion: Under the Gaussian distribution assumption, prior... With data The deviation is in 2 Within 2 is considered normal statistical fluctuation (95.4% confidence interval); exceeding 2 There are statistically significant reasons to suspect that the prior and the data come from different distributions (i.e., LLM may misclassify). Specifically, when When ≤2, Unchanged, that is = The prior accuracy improvement is fully effective; when =5 (moderate conflict) Down to 1 / 4; when =50 (extreme conflict, a typical misclassification case). The impact of prior knowledge was reduced to 1 / 49 of its original value, and the influence of prior knowledge was almost completely suppressed. This mechanism fully utilizes high-quality prior knowledge while automatically resisting the risk of misclassification caused by LLM illusion.

[0105] Step 9: Calculate the prior discount weights. Existing knowledge fusion methods typically allocate the contributions of prior knowledge and data likelihood with manually set fixed weights, which cannot adapt to the dynamic changes in historical case coverage. This invention introduces the power-prior theory into the RAG retrieval quality adaptive scenario to improve retrieval quality. As a quantitative proxy for data coverage, it enables fully automatic dynamic adjustment of prior effective weights without manual intervention:

[0106] (12)

[0107] in, For the first The prior discount weight of each KPI, with a value range of [0,1]. The prior decay exponent (preferred value) =1.0, with a value range of [0.5, 2.0]. The larger the prior (the faster the decay). When hour (Data-driven), when hour (Priority-driven).

[0108] like Figure 4 The diagram illustrates a precision-weighted Bayesian fusion model. Subgraphs (a) and (b) show the fusion process of prior and likelihood in two typical scenarios: sufficient data and sparse data. In (a), historical case data is abundant, data dominates, and the posterior distribution almost coincides with the data likelihood, with semantic-guided physical prior providing only slight corrections. In (b), historical data is scarce, and retrieval quality... Low, prior discount weight Automatically increasing, the posterior shifts significantly towards semantically guided physical priors, with the physical priors playing a fallback role. The two subgraphs visually illustrate the core adaptive fusion mechanism of the Power Priors, which automatically adjusts prior weights based on retrieval quality, achieving a smooth transition from data-driven to prior-backup.

[0109] Step 10, First-layer precision-weighted Bayesian fusion. Based on the conjugate prior theory of Gaussian distributions, when both the prior and likelihood distributions are Gaussian, the posterior distribution is also Gaussian, and the posterior precision equals the sum of the prior precision and the likelihood precision—the higher the precision of the distribution, the greater its contribution to the posterior. This property naturally realizes an adaptive fusion mechanism where the more certain distribution dominates. Combining the prior precision enhancement in Step 8 and the Power Prior discount in Step 9, the fusion formula is:

[0110] (13)

[0111] (14)

[0112] (15)

[0113] in, For the first The posterior precision of a KPI is equal to the sum of the data precision and the weighted prior precision. Higher precision indicates a more concentrated posterior distribution. For the first The posterior mean of each KPI is obtained by weighting the data likelihood mean and the prior mean according to their respective precision, and is the expected value of the fused KPI. For the first The posterior standard deviation of each KPI is equal to the square root of the reciprocal of the posterior precision, reflecting the magnitude of the overall uncertainty after fusion.

[0114] Step 11: Obtain the Gaussian posterior distribution. The fusion result is a Gaussian posterior distribution. It integrates two knowledge streams: semantic guidance physical prior driven by intent classification and likelihood from historical case data.

[0115] Step 12, truncate the projection of the second-layer physical constraint. (e.g.) Figure 5 The diagram illustrates the physical constraint truncation projection mechanism. Subgraphs (a) and (b) show two typical scenarios: slight truncation of the posterior mean within the physical feasible region and illusion suppression of the posterior mean exceeding the boundary. These visually demonstrate the core mechanism by which the truncation operation pulls the out-of-bounds mean back into the feasible region and simultaneously adjusts the variance. The posterior distribution... Projected onto the physically feasible region The truncated normal distribution is obtained. The mean and variance of a truncated normal distribution have a closed-form solution:

[0116] make , ,

[0117] (16)

[0118] (17)

[0119] in, The standard normal cumulative distribution function is... It is the standard normal probability density function. is a normalization constant, equal to the probability mass within the cutoff interval; For the first The truncated normal mean of each KPI after physical constraint truncation projection is automatically pulled back into the feasible region when the posterior mean goes out of bounds. The corresponding truncated normal variance is truncated. The truncation operation shrinks the variance compared to the posterior variance and reflects additional boundary uncertainties.

[0120] When the posterior mean is within the physically feasible region and far from the boundary, the truncation transformation has a small impact on the distribution; when the posterior mean is close to or exceeds the boundary of the physically feasible region (i.e., large model illusion occurs), the truncation transformation automatically pulls the mean back into the physically feasible region, while adjusting the variance to reflect additional uncertainty, thus suppressing the physical infeasibility risk caused by large model illusion at its source. Simultaneously applying... Lower bound constraint ×0.01 to avoid overconfidence.

[0121] Step 13, Superimpose Adaptive Observation Noise. The observation noise variance is superimposed after the truncated projection to obtain the final prediction variance. The observation noise uses an adaptive scaling factor, proportional to the truncated mean, and the noise figure varies according to the interference level:

[0122] (18)

[0123] (19)

[0124] in, For the first The adaptive observation noise standard deviation of each KPI is proportional to the truncated mean and scaled differently according to the interference level. Truncate the normal mean (as input); To truncate the normal standard deviation (as input); The final prediction standard deviation after superimposing observation noise comprehensively reflects the uncertainty of truncated projection and channel random fluctuations; and These represent the upper and lower limits of the noise amplitude.

[0125] This refers to the relative noise figure differentiated according to interference levels. The bandwidth noise figure in this invention... =[0.15, 0.20, 0.25, 0.30] (corresponding to Lv0 / Lv1 / Lv2 / Lv3), delay noise figure =[0.15, 0.25, 0.35, 0.50] (corresponding to Lv0 / Lv1 / Lv2 / Lv3). The physical basis for the increasing interference levels is as follows: Fast fading in a wireless channel is approximately constant in the logarithmic domain, and the power fluctuation of typical Rayleigh fading is ±3~7 dB (3GPP TR 38.901), corresponding to the power standard deviation in the linear domain. The higher the interference level, the more severe the channel fluctuations: at Lv0 with no interference, the fast fading fluctuations are approximately ±1~2 dB (c≈0.15), while at Lv3 with strong suppression, the fluctuations can reach ±5~7 dB (c≈0.30~0.50) after the superimposed effects of interference backoff and frequent HARQ retransmissions. This differentiated design ensures that at Lv0... To avoid artificial inflation (and thus prevent damage to the existing optimal NLL), at Lv3 The bandwidth is sufficiently widened to cover long-tail fluctuations caused by strong interference. The noise figure of the delay is higher than that of the bandwidth because the noise sources of the delay, in addition to channel fading, also include queuing jitter (random queue depth), HARQ retransmission (each retransmission adds about 8 ms), and interference backoff (random backoff window), resulting in a higher relative fluctuation amplitude than the bandwidth. In this invention, the bandwidth noise range is [0.005, 0.35] Mbps: the lower limit of 0.005 Mbps corresponds to the minimum physical jitter (the minimum resolvable throughput change corresponding to the equipment thermal noise floor), and the upper limit of 0.35 Mbps is approximately 17.5% of the maximum nominal bandwidth of the tactical radio, 2.0 Mbps, limiting the noise level to a reasonable level in large signal scenarios. The latency noise range is [5.0, 50.0] ms: the lower limit of 5.0 ms corresponds to the minimum processing latency of the protocol stack (1 TTI = 1 ms scheduling cycle plus the inherent processing overhead of the protocol stack), and the upper limit of 50.0 ms corresponds to the upper limit of the superposition of multiple HARQ retransmissions under high interference (maximum 6 times × 8 ms / time ≈ 48 ms) and queuing jitter. The design consideration for superimposing noise after truncation is as follows: if superimposed before truncation, the standard deviation of the noise-inflated distribution will be used to calculate the mean of the severely distorted small KPI through the closed moment of the truncated normal distribution (for example, in a scenario with a bandwidth of 0.038 Mbps, 0.35 Mbps of fixed noise causes the truncated mean to shift from 0.038 to 0.33).

[0126] Step 14, Interference Level Perception Scaling. Existing uncertainty estimation methods use uniform global calibration coefficients, which cannot adapt to the fundamental differences in KPI distribution patterns under different interference levels (the GBR capacity saturation quasi-degradation distribution in interference-free scenarios versus the long-tailed burst distribution in strong interference scenarios). The requirements are drastically different. This invention proposes a diagnostic calibration scaling mechanism based on a dual-dimensional joint index of KPI and interference level, deriving the scaling coefficient based on NLL optimal solution theory, and achieving... Precise alignment with RMSE at each level:

[0127] (20)

[0128] in, The final prediction standard deviation after interference level perception scaling is used for subsequent bimodal mixed distribution modeling and probability calculation; The predicted standard deviation after adding observation noise (as input); A table is looked up for the scaling factor based on the combined differentiation of KPI and interference level. The bandwidth scaling factor in this invention... =[1.70, 0.88, 0.86, 1.56] (corresponding to Lv0 / Lv1 / Lv2 / Lv3), delay scaling factor =[1.42, 0.85, 0.67, 1.44] (corresponding to Lv0 / Lv1 / Lv2 / Lv3). The determination method is based on the diagnostic calibration principle: first, calculate the per-Lv RMSE (root mean square of the true prediction error) for each interference level on the test set. (The mean of the predicted standard deviation), then with =RMSE / One-time precise calibration, ensuring that the calibrated Aligned with RMSE at each level of interference. The theoretical basis for this method is: under the assumption of a Gaussian prediction distribution, The optimal solution is exactly =RMSE (i.e.) When / RMSE=1.0, NLL takes the global minimum value); where, This represents the actual observed value of the KPI; This is the truncated predicted mean of the corresponding KPI. The scaling factor exhibits a U-shaped characteristic, high at both ends and low in the middle (the coefficients for Lv0 and Lv3 are greater than 1, while the coefficients for Lv1 and Lv2 are less than 1). The physical interpretation is as follows: In an interference-free environment at Lv0, according to the QoS GBR model specified in 3GPP TS 23.501, the link reaches capacity saturation under high SNR conditions. The true KPI values ​​are largely concentrated near the upper constraint bound, exhibiting a quasi-degradable distribution, resulting in a high actual RMSE, requiring scaling. To match; under moderate disturbances at Lv1 / Lv2, the KPI distribution is relatively regular, and the preceding mechanism has provided appropriate matching. It is estimated that only minor adjustments are needed; however, under strong suppression at Lv3, long-tail events erupt and the truth variance increases sharply (the RMSE of latency expands from 22.5 ms at Lv0 to 66.0 ms at Lv3), requiring significant amplification. To cover the long-tail distribution. This step, along with the four accuracy calibration methods (a priori) in the preceding steps, is related to this. Calibration, data layer The lower bound, KPI differentiation accuracy enhancement, and conflict adaptive decay together constitute the probability distribution accuracy adaptive system of this method: the first four methods deal with the knowledge source level differentiation in the prior-data fusion stage, while this step deals with the environmental intensity level differentiation after fusion output. The two work together from different dimensions to ensure accurate calibration of the probability distribution in various scenarios. Experimental results after calibration show that all eight Lv×KPI combinations... The RMSE values ​​were all accurate to 1.00, verifying the effectiveness of the diagnostic calibration method.

[0129] Step 15: The third layer constructs a bimodal hybrid distribution. For example... Figure 6 The figure shown is a theoretical model diagram of a dual-modal mixture probability distribution. The continuous main peak on the left side of the figure represents the normal working state (truncated normal distribution, weight 1-). The right-hand spike represents a link interruption failure state (approximately). Function, weight ), business decision threshold The main peak is divided into a normal operation zone and a performance degradation default zone, fully characterizing the bimodal probability structure and long-tail risk of KPIs under both operating states. The link outage probability is estimated by combining the environmental characteristics from step 4. Construct a hybrid distribution of normal operating state and link interruption state:

[0130] (twenty one)

[0131] The first item is , is the truncated normal distribution component of the normal working state, where, This represents the probability of link interruption. This represents the probability mass weight of the component. For the first The values ​​of each KPI; To truncate the predicted mean; To truncate the variance parameter of the normal distribution; The first The lower and upper bounds of the physical feasible region for each KPI. The second term is... It is the Dirac state of the link interruption. Functional components; where, The Dirac function represents the concentration of all interruption probability mass at a single point; For the first The interruption degradation value for each KPI, with bandwidth taken as... Delay .

[0132] In practice, due to The function is a theoretically infinitely narrow pulse, which cannot be directly calculated numerically. This invention uses the following three equivalent calculation methods to handle this discrete term: When calculating the probability density, it uses an extremely narrow Gaussian distribution. approximate function, where In calculating the cumulative distribution function, a step function is used instead, i.e. ;

[0133] in, This is the KPI value lookup point; To truncate the predicted mean; Scaling the forecast standard deviation ; This represents the probability of link interruption. For the characteristic function, when The value is 1 if true and 0 otherwise. In Monte Carlo sampling, the probability is... Directly assign the sampled value as the interruption degradation value .

[0134] Link interruption probability Physical modeling is performed based on battlefield environment characteristics, comprehensively considering three factors: electromagnetic interference, communication distance, and terrain.

[0135] (twenty two)

[0136] in, The interference outage probability is mapped to {0.02, 0.08, 0.15, 0.30} based on the interference levels {0, 1, 2, 3}. This mapping references the HARQ maximum retransmission count and signal-to-interference-plus-noise ratio (SINR) degradation model in the 5G NR standard: interference level 0 (no interference) only has extremely low probability events such as equipment failure (~2%); interference level 1 (slight interference, SNR penalty of about 5 dB) corresponds to occasional retransmission timeouts (~8%); interference level 2 (moderate interference, SNR penalty of about 12 dB) corresponds to frequent retransmission failures (~15%); interference level 3 (strong interference, SNR penalty of about 25 dB) causes the signal quality to fall below the demodulation threshold, and the outage probability increases significantly (~30%). The probability of distance interruption increases linearly when the communication distance exceeds 5 km (after exceeding the typical coverage radius of a tactical radio, path loss causes a rapid decrease in SNR). The probability of terrain interruption increases by 0.03 in indoor and mountainous scenes (due to additional shadow decay caused by non-line-of-sight propagation).

[0137] Under the bimodal distribution, the overall expected value for:

[0138] (twenty three)

[0139] Comprehensive variance for:

[0140] (twenty four)

[0141] Step 16: Calculate the probability of default, given a business decision threshold. Considering the physical meaning of different KPIs in tactical communication scenarios, the calculation of default probability should differentiate between indicator types:

[0142] For metrics where larger is always better (such as bandwidth): default is defined as the actual value falling below a minimum threshold. ,Right now Default probability The formula is:

[0143] (25)

[0144] For metrics where smaller is better (such as latency): a default is defined as an actual value exceeding the maximum threshold. ,Right now The formula for the probability of default is:

[0145] (26)

[0146] Step 17, Calculate the comprehensive risk score. To characterize long-tail disaster risk, define the comprehensive risk score:

[0147] (27)

[0148] in, Indicates the first The comprehensive risk score of each KPI reflects the weighted superposition of the risk of constraint violation and link interruption faced by the KPI prediction result, and is used to rank the risks of multiple KPI prediction results and provide decision support. and For the weighting coefficients, satisfying + =1, which determines the relative severity of penalties for default and interruption events. In this invention, a preferred value is selected. =0.3、 =0.7 (This can be adjusted by those skilled in the art based on the relative sensitivity of the combat mission to breaches and interruptions).

[0149] Step 18: Output the KPI probability distribution and risk quantification results. Encapsulate the probabilistic information of multi-dimensional KPIs into a unified KPI probability state vector:

[0150] (28)

[0151] Where K represents the total number of KPIs. The system output concludes that the link is in a high-risk state, with a serious risk of bandwidth default. =97.1%) and interruption risk ( (15%), it is recommended to take immediate link switching or redundancy protection measures.

[0152] In summary, this invention, using a typical tactical mission of short-range video backhaul as an example, fully demonstrates the entire inference process of the PDBBI algorithm, from the commander's natural language operational intent input, through semantic-guided physical prior modeling, historical case RAG retrieval, three-layer progressive Bayesian inference, to the output of KPI probability distribution and risk quantification results. Taking bandwidth KPI as an example, the final output expected value is 1.134 Mbps, prediction standard deviation is 0.364 Mbps, default probability is 97.1%, and comprehensive risk score is 0.396; comprehensively judging that the link is currently in a high-risk state, it is recommended to immediately take link switching or redundancy protection measures. Latency KPI is calculated and output independently according to the same inference process, and will not be described further.

[0153] Taking bandwidth KPI as an example, the complete inference value chain is as follows. Input: Operational intent "The battalion command vehicle notifies the forward reconnaissance UAV cluster to immediately activate video transmission and transmit the dynamic real-time data of targets within 500 meters in front of the position back to the command post," environmental characteristics .

[0154] (1) LLM classification: constraint_type="recon_video_short", conf=0.82. (2) Semantic-guided physical priors: Mbps; static = 0.6 Mbps, after physical consistency scaling (short range, classification consistency, ×0.5) to get 0.3 Mbps, then after confidence scaling ( )have to Mbps, a priori accuracy (3) Data likelihood: Mbps, Mbps, Data accuracy Power Priority Discount Physically feasible region [0.5, 2.0] Mbps, service threshold Mbps. (4) Precision-weighted Bayesian fusion (first layer): , Mbps, Mbps. (5) Physical constraint truncation projection (second layer): Mbps, Mbps. (6) Observation noise and Scaling (lv=2): Mbps, Mbps, interference perception scaling Mbps. (7) Dual-modal hybrid distribution (third layer): Overall expectations Mbps. (8) Risk quantification: probability of default The overall risk score is Risk_Score = 0.3 × 0.971 + 0.7 × 0.15 = 0.396 (high risk).

[0155] Furthermore, to further verify the performance advantages of this invention in different battlefield scenarios, a performance evaluation of the method was conducted in a specific test embodiment. On 1309 independent test set samples, the performance of five methods—B1 (rule lookup table method, directly querying nominal values ​​based on constraint type without probability modeling), B2 (pure KNN mean method, using only data layer retrieval mean prediction without LLM and physical constraints), B3 (pure LLM prior method, using only semantically guided physical priors calculated by the physical channel model after LLM intent classification without using historical case data), and B4 (LLM+RAG concatenation method, concatenating retrieved historical cases and their KPI data into prompt words, directly generating KPI prediction values ​​by a large language model, representing a typical paradigm of existing LLM+RAG shallow fusion)—was compared with the complete three-layer fusion method presented in this paper. Evaluation metrics included MAE (mean absolute error, measuring point estimation accuracy), NLL (negative log-likelihood, measuring probability distribution prediction quality), CRPS (continuous ranking probability score), and 95% confidence interval coverage. The calculation of NLL is based on the log-likelihood of the bimodal distribution:

[0156] (29)

[0157] in, =1 means Under normal observation conditions, =0 indicates an interrupt event. This is to truncate the normal density function.

[0158] like Figure 7As shown, in the multi-dimensional indicator radar charts for the two KPIs of link reachable bandwidth and end-to-end delay, the orange area of ​​our proposed method is the largest among the five methods, indicating that our proposed method has the best overall performance across the six dimensions of MAE, RMSE, NLL, CRPS, 95% CI coverage, and calibration error. B2 pure KNN shows a significant shrinkage in the NLL dimension, reflecting its worst probability distribution calibration capability; B1 rule lookup and B3 pure LLM lag significantly behind in the CRPS dimension; B4 LLM+RAG is in the middle overall, but does not surpass our proposed method in any dimension. (Two sub-charts are shown.) Figure 1 The results show that the proposed method does not have any obvious shortcomings and achieves synergistic improvement in three complementary dimensions: point estimation accuracy, probability distribution quality, and uncertainty calibration.

[0159] like Figure 8 As shown, using historical case library coverage (5%~100%) as the independent variable, the MAE of this invention and the B2 pure KNN baseline are compared in terms of bandwidth and latency KPIs. When the coverage is only 5%, the prior weights... =0.74, with physical prior inference as the primary driver, the latency MAE of this invention is reduced by 30.5% compared to B2; as coverage increases to 100%, The latency automatically decreased to 0.21, and the data gradually became dominant. The difference in latency MAE between the two methods narrowed from 10.03 ms to 1.77 ms, all without manual intervention. These results verify the robust fallback capability of the PowerPrior mechanism in data-sparse scenarios and its adaptive transition characteristics as data accumulates.

[0160] In summary, this invention addresses the challenges of accurately mapping operational intentions to key network performance indicators (KPIs) in complex battlefield environments, the lack of risk quantification capabilities in existing methods, the frequent occurrence of physically infeasible predictions due to large-scale model illusions, and insufficient depth of multi-source knowledge fusion. It proposes a tactical communication intention translation method (PDBBI) based on physics-driven bimodal Bayesian inference. This method integrates three types of heterogeneous knowledge: semantically guided physical priors driven by intention classification, likelihood from historical case data, and hard constraints from communication physics laws. Through a three-layer progressive inference process—precision-weighted Bayesian fusion, physics-constrained truncated projection, and bimodal hybrid distribution modeling—it achieves a complete mapping of operational intentions to KPI probability distributions and risk quantification results. Simulation results demonstrate that this invention comprehensively outperforms existing deterministic intention translation methods in three core indicators: KPI point estimation accuracy, probability distribution calibration quality, and physical constraint compliance. Furthermore, it maintains stable performance advantages in complex tactical scenarios with varying levels of electromagnetic interference, terrain types, and operational mission types, exhibiting strong cross-scenario generalization capabilities.

[0161] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A tactical communication intent translation method based on physics-driven bimodal Bayesian inference, characterized in that, Includes the following steps: S1. Construct a basic model for a tactical communication intent translation system, including intent parsing, environmental perception, and joint intent-environment feature construction. While receiving the natural language combat mission description input by the commander, it collects real-time battlefield environment features, including communication link distance, electromagnetic interference level, terrain type, and core situational parameters. Through structured parsing and feature encoding, it completes the joint feature representation of combat intent semantic information and battlefield environment situational information. S2. Based on the aforementioned intention-environment joint features, complete independent modeling of three types of heterogeneous knowledge: semantic guidance physical prior, historical case data likelihood, and hard constraints of communication physical laws. S3. A physical-driven bimodal Bayesian inference algorithm is adopted to achieve unified probabilistic fusion of multi-source knowledge through three progressive inference layers: The first layer performs precision-weighted Bayesian fusion to generate a Gaussian posterior distribution; the second layer performs truncated projection on the Gaussian posterior distribution based on the feasible region constrained by the physical laws of communication to obtain a truncated normal distribution; the third layer estimates the link interruption probability by combining battlefield environment characteristics to construct a bimodal hybrid distribution of normal working state and link interruption state. S4. Based on the dual-modal hybrid distribution, calculate the comprehensive expected value, confidence interval, business default probability and comprehensive risk score of the network's key performance indicators (KPIs), and output the tactical communication intent translation results containing probability distribution and risk quantification, providing a basis for tactical communication network configuration and command decision-making.

2. The tactical communication intent translation method based on physics-driven bimodal Bayesian inference according to claim 1, characterized in that, Step S2, semantic-guided physical prior distribution modeling, specifically includes: A large language model is used to classify the constraint types of combat intentions and output the classification confidence scores; combined with real-time battlefield environment parameters, the expected values ​​of KPIs are calculated based on the communication channel propagation model and channel capacity theory; and a semantic-guided physical prior Gaussian distribution is constructed by adaptively scaling the prior standard deviation based on the classification confidence scores. When there is a contradiction between the classification results of the large language model and the physical characteristics of communication distance, the prior standard deviation is magnified by a factor; when the classification confidence is lower, the prior standard deviation is nonlinearly amplified to quantify the cognitive uncertainty of the model.

3. The tactical communication intent translation method based on physics-driven bimodal Bayesian inference according to claim 1, characterized in that, Step S2, the likelihood distribution modeling of historical case data, specifically includes: Construct a historical tactical communication case library containing operational intent text, environmental features, actual KPI values, and link connectivity status; employ the FAISS vector retrieval framework, integrating text semantic similarity and environmental feature similarity, to retrieve Top-K historical cases matching the current scenario; statistically analyze the KPI mean and discrete variance of the retrieved cases to construct a Gaussian distribution of data likelihood. When the highest semantic similarity of the retrieved case is lower than the preset threshold, it is judged as an extreme unknown scenario with zero samples, and the data likelihood precision is set to zero. The posterior distribution completely degenerates into a semantically guided physical prior distribution.

4. The tactical communication intent translation method based on physics-driven bimodal Bayesian inference according to claim 2, characterized in that, Step S2, the modeling of the feasible region under hard constraints of communication physical laws, specifically includes: A mapping table between tactical communication constraint types and KPI physical feasible domains is pre-established, wherein the KPIs include link reachable bandwidth and end-to-end delay. Based on the Shannon-Hartley theorem, 3GPP channel propagation theory, and typical tactical communication service coding delay specifications, the upper and lower bounds of the bandwidth feasible domain, the physical feasible domain of delay, and the service decision threshold corresponding to each constraint type are determined as hard boundaries for subsequent distributed compliance verification.

5. The tactical communication intent translation method based on physics-driven bimodal Bayesian inference according to claim 1, characterized in that, In step S3, the first-layer precision-weighted Bayesian fusion includes the following steps: S3.1 Precision Differentiation Configuration: Set a prior precision enhancement factor higher than the bandwidth KPI for the latency KPI, and perform physical model measured precision calibration on the prior standard deviation of latency; S3.2, Prior-Data Conflict Adaptive Decay: When the deviation between the prior mean and the data mean exceeds twice the data standard deviation, the prior accuracy is adaptively decayed according to the standardized conflict metric to suppress the prior bias caused by misclassification of large models. S3.3, Power-Prior Dynamic Weight Adjustment: Based on the historical case retrieval quality, the prior discount factor is calculated. When the data case coverage is high, the data likelihood dominates the fusion. When the data is sparse, the semantic guidance and physical prior are the last resort. S3.4 Environmental Interference Perception Scaling: After fusing and generating a Gaussian posterior distribution, the posterior standard deviation is scaled according to the electromagnetic interference level to match the random uncertainty change pattern of the battlefield environment.

6. The tactical communication intent translation method based on physics-driven bimodal Bayesian inference according to claim 1, characterized in that, Step S3, the second-layer physical constraint truncation projection, specifically includes: using the feasible region of hard constraints based on communication physical laws as the boundary, performing a truncated normal closed-form solution on the mean and variance of the Gaussian posterior distribution; if the posterior mean exceeds the boundary of the physical feasible region, automatically pulling the mean back into the feasible region, and simultaneously adjusting the variance to represent additional uncertainty; superimposing adaptive observation noise linked to electromagnetic interference level and KPI type, and using a KPI-interference level joint differential scaling coefficient for probability distribution calibration, so that the predicted standard deviation is precisely aligned with the root mean square of the actual prediction error, suppressing the physical infeasibility prediction caused by the illusion of a large model.

7. The tactical communication intent translation method based on physics-driven bimodal Bayesian inference according to claim 1, characterized in that, Step S3, the third-layer bimodal mixture distribution modeling, specifically includes: The truncated normal distribution after projection is used as the probability distribution of the normal working state of the link; the Dirac function is used to approximate the probability distribution of the link outage failure state, with the bandwidth outage value set to 0 Mbps and the delay outage value set to the preset maximum degradation delay; The probability of link interruption is physically modeled by three-dimensional factors including electromagnetic interference level, communication distance, and terrain type. Based on the interruption probability, the normal operation state and the interruption failure state are probabilistically weighted and mixed to construct a dual-modal mixed distribution, which characterizes the bimodal distribution characteristics of KPI and the long-tail failure risk on the battlefield.

8. The tactical communication intent translation method based on physics-driven bimodal Bayesian inference according to claim 1, characterized in that, In step S4, the probability of service default is calculated as follows: for bandwidth KPIs that are larger and better, the probability of default is the cumulative probability that the actual value of the KPI is lower than the minimum threshold of the service; for latency KPIs that are smaller and better, the probability of default is the cumulative probability that the actual value of the KPI is higher than the maximum threshold of the service; the comprehensive risk score is constructed by weighting the probability of default and the probability of link interruption, and the weight coefficient can be adaptively configured according to the sensitivity of the combat mission to service default and link interruption.

9. A tactical communication intent translation system based on physics-driven bimodal Bayesian inference, used to implement the tactical communication intent translation method according to any one of claims 1-8, characterized in that, It comprises an input layer, a perception layer, a decision layer, and an output layer. The input layer receives natural language operational intent commands and real-time battlefield environment feature parameters. The perception layer performs operational intent feature encoding, independent modeling of multi-source heterogeneous knowledge, and generates semantically guided physical priors, historical case data likelihoods, and feasible regions with communication physical hard constraints. The decision layer incorporates a precision-weighted Bayesian fusion module, a physical constraint truncated projection module, and a dual-modal hybrid distribution modeling module to achieve three-layer progressive Bayesian probabilistic inference. The output layer encapsulates KPI probability distribution parameters, confidence intervals, default probabilities, and comprehensive risk scores, outputting the quantified results of tactical communication intent translation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the tactical communication intent translation method based on physical-driven bimodal Bayesian inference as described in any one of claims 1-8.

Images