Communication block error rate intelligent prediction and deployment method for cross-domain knowledge migration

Abstract

The invention provides an intelligent prediction and deployment method for a communication block error rate of cross-domain knowledge migration. The method comprises the following steps: extracting upper and lower fields strongly related to the block error rate by using a wireless knowledge graph and dividing field data; analyzing a monotonic relationship between the upper and lower character fields and the block error rate, and constructing a physical inspiration sub-target; weighting and aggregating source domain data, implementing joint representation alignment of context representation and block error rate, and establishing a block error rate knowledge migration sub-target; constructing a generalization upper bound sub-target by using a parameter gradient norm of the block error rate prediction model; performing weighted fusion on the three sub-targets to form a block error rate prediction optimization target; adjusting and optimizing the block error rate prediction model through two stages, and alternately optimizing model parameters and source domain data weight vectors; and continuously monitoring the distribution change of domain data after model deployment, and triggering the model to be re-optimized if significant offset occurs. According to the method, the block error rate physical prior information and the domain self-adaption theory are effectively utilized, and more accurate and robust block error rate prediction support can be provided.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a method for intelligent prediction and deployment of communication error rate through cross-domain knowledge transfer. Background Technology

[0002] The rapid development of mobile communication networks, represented by 5G / 6G, has spurred numerous emerging network application scenarios and services. Considering the high reliability requirements, inherent complexity, and high dynamism of network operation, the verification and deployment of network strategies are becoming increasingly difficult. Traditional on-premises strategy verification carries significant risks and the cost of repeated trials. To address this challenge, network digital twin technology promises to overcome this difficulty. By utilizing data collected from the physical network to construct a high-fidelity virtual network, a virtual interactive environment is provided for network strategies, thereby enabling low-risk, reliable strategy optimization and pre-deployment performance pre-verification.

[0003] In wireless communication systems, the Block Error Rate (BLER) of a Transport Block (TB) is a critical performance metric because it determines whether a TB is successfully transmitted, directly impacting service quality metrics such as effective throughput and packet loss rate. However, since TB BLER is a probabilistic metric, deterministically determining the transmission status of a TB—that is, confirming its transport block transmission status (N)ACK (Non-Acknowledge, ACK / NACK) feedback value—requires a large amount of accurate runtime state information. Therefore, the TB BLER statistically derived from (N)ACKs collected from a specific communication context can be considered a twin target. One of the key factors influencing BLER is the modulation and coding scheme (MCS) employed. However, in practical wireless networks, the available data in certain communication context domains (such as MCS) is often limited (mainly constrained by the communication environment, service behavior, network policy behavior, etc.), thus hindering the accuracy of TB BLER prediction. In machine learning, this is a classic data distribution or label offset problem. Therefore, building a high-fidelity TB BLER twin model requires the ability to perform highly reliable pre-validation outside the data distribution. However, this issue has received little attention. Summary of the Invention

[0004] Purpose of the invention: To address the shortcomings of existing technologies, this invention discloses an intelligent prediction and deployment method for communication block error rate through cross-domain knowledge transfer. It effectively utilizes the physical relationship between communication context and block error rate and domain adaptation theory to achieve cross-domain transfer of block error rate knowledge, which helps improve the accuracy of block error rate prediction and provides robust block error rate prediction technology support for application scenarios such as network digital twins.

[0005] From the perspective of wireless communication network data warehouse storage, this invention mainly utilizes a wireless knowledge graph and a service data warehouse. The wireless knowledge graph details the status and relationships of various wireless communication indicators within the wireless network, while the service data warehouse records data related to the block error rate within the wireless network. By comprehensively utilizing the wireless knowledge graph and the service data warehouse, a domain dataset strongly correlated with the block error rate can be constructed, thereby achieving efficient and accurate prediction of the block error rate.

[0006] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution.

[0007] A method for intelligent prediction and deployment of communication error rate in cross-domain knowledge transfer, characterized by the following steps:

[0008] Step 100: Extract context fields and corresponding datasets that are strongly related to the error rate using wireless knowledge graphs and business data warehouses, and divide the domain data according to their significant impact.

[0009] Step 200: Combine standard encoding and decoding behavior analysis with the monotonicity relationship between context fields and block error rate to construct a physically inspired sub-target;

[0010] Step 300: Aggregate source domain data using source domain data weight vectors, implement joint representation alignment of context representation and error block rate, and establish the knowledge transfer sub-objective of error block rate;

[0011] Step 400: Construct a generalized upper bound sub-objective using the parameter gradient norm of the error block rate prediction model;

[0012] Step 500: Weighted fusion of the three sub-objectives—physical inspiration, block error rate knowledge transfer, and generalization upper bound—to form the block error rate prediction model optimization objective;

[0013] Step 600: Optimize and tune the error block rate prediction model in two stages: optimize the model parameters in the inner layer and optimize the source domain data weight vector in the outer layer.

[0014] Step 700: After the model is deployed, continuously monitor the distribution changes of the domain data. If a significant shift occurs, trigger model retuning.

[0015] Furthermore, step 100 specifically includes the following steps:

[0016] Step 110: Construct a wireless knowledge graph and business data warehouse;

[0017] Step 120: Utilize the wireless knowledge graph to perform correlation analysis, extract context fields that are strongly correlated with the block error rate index, covering field types such as communication resource scheduling and channel quality; then, extract historical data containing the aforementioned context fields and block error rates from the business data warehouse, and use the context unique identifier method to statistically analyze the block error rate under a specific context, thereby generating a block error rate dataset.

[0018] Furthermore, the specific steps of step 120 are as follows: Step 121: For the contextual analysis requirements of the block error rate metric, input the block error rate metric. The wireless knowledge graph calculates and sorts the correlation weights between the associated fields and the required fields for each association category, selecting the top-ranked fields. The contextual fields extracted by the wireless knowledge graph are denoted as... ,in, Indicates the number of resource blocks. Indicates resource block channel quality, Represents the resource block to be scheduled. Indicates the transport block size, Indicates the number of transport blocks, Represents the number of streams, Indicates the modulation order. Indicates bitrate, Indicates the dimension size as The real number field;

[0019] Step 122: Extract data from the business data warehouse based on the extracted context fields; denote the extracted dataset as... ,in, Indicates the number of samples. , and Let the context, (N)ACK feedback value, and coding and modulation scheme order of the i-th sample be represented respectively; the channel quality equivalent operator is scheduled using resource blocks. deal with and , This represents the channel quality of the aggregated and sorted resource blocks, where Represent the Hartmann product; then, Viewed as an empirical distribution, in a specific context The block error rate statistics are as follows: Therefore, we can obtain a size of Block error rate dataset ,in, The operator represents the expectation of a distribution. Domain ;

[0020] Step 123: Based on the fact that the context field values ​​are clearly divided according to preset rules and the dominant domain dataset distribution differences are significant, a certain context field is established as the domain segmentation field; with the MCS order as the domain segmentation field, the joint distribution of "context-block error rate" in each domain shows significant differences;

[0021] Step 124: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The datasets are divided into S domains based on their MCS order. Simultaneously, the target domain dataset is represented as having a sample size of... of Further Classified as block error rate Unknown and block error rate Knowable .

[0022] Furthermore, step 200 details the steps as follows:

[0023] Step 210: Based on the encoding and decoding behavior of the new air interface standard, after the transport block undergoes cyclic redundancy check, the block error rate is... Regarding the context The partial derivatives generally satisfy ,in and These represent context slices where the partial derivatives are less than 0 and greater than 0, respectively.

[0024] Step 220: Let the block error rate prediction model be... In each iteration round, the corresponding communication context is calculated. The partial derivatives violate the direction of the model; to enhance the physical generalization ability of the model, a random sample generation method is adopted to calculate the partial derivatives. ,

[0025] Among them, synthetic samples , and These represent the contexts from the source domain and the target domain, respectively. The coefficients are randomly synthesized. It is an element-wise non-negative operator that operates on vectors. Represent the 1-norm of a vector; by minimizing , It will tend to satisfy step 210. about The direction of the partial derivative, thereby achieving Its generalization ability in terms of monotonicity, while its more complex behavioral characterization will be achieved through the knowledge transfer objective of error rate based on domain adaptation theory;

[0026] Step 230: Construct the physically inspired sub-objective ,in The regularization coefficient is used to penalize violations of the physical monotonicity of the block error rate. .

[0027] Furthermore, step 300 specifically includes the following steps:

[0028] Step 310: Define and initialize the source domain data weight vector with equal weights; the block error rate prediction model specifically includes: a feature extraction neural network, a block error rate prediction neural network, and a pseudo-block error rate prediction neural network;

[0029] Step 320: For target domain data with unknown block error rate, the feature extraction neural network is used to represent the source domain and target domain data; the source domain data representation is weighted and aggregated using the source domain data weight vector to obtain the source domain aggregated representation; the source domain aggregated representation is input into the block error rate prediction neural network to obtain the source domain block error rate prediction error; the block error rate of the target domain is approximated by the pseudo block error rate output by the pseudo block error rate prediction neural network; the Wasserstein-1 distance is used to characterize the difference between the two joint representations "source domain aggregated representation - source domain block error rate" and "target domain data representation - pseudo block error rate" to obtain the unsupervised joint representation distance; the unsupervised joint representation distance and the source domain block error rate prediction error are added together as the unsupervised block error rate knowledge transfer sub-objective;

[0030] Step 330: For target domain error rate data with known error rates, firstly, the same processing procedure as for target domain data with unknown error rates is adopted to obtain the first sub-objective of supervised error rate knowledge transfer; simultaneously, the target domain data with known error rates is input into the feature extraction and error rate prediction neural network to obtain the target domain prediction error, which is used as the second sub-objective of supervised error rate knowledge transfer, and the first sub-objective of supervised error rate knowledge transfer is added according to weights, and the result is used as the supervised error rate knowledge transfer sub-objective.

[0031] Step 340: Combine the unsupervised and supervised block rate knowledge transfer sub-objectives according to weights; using the block rate prediction neural network and the pseudo-block rate prediction neural network, approximate the Wasserstein-1 distance as the maximum difference between the block rate prediction errors of the source domain and the target domain, and introduce a gradient reversal layer into the pseudo-block rate prediction neural network. Then, add the two sub-objectives of unsupervised and supervised block rate knowledge transfer according to weights to finally obtain the block rate knowledge transfer sub-objective.

[0032] Furthermore, in step 310, a source domain data weight vector is defined. ,satisfy Where S is the total number of source domains, and the weight of the i-th source domain data is initialized to... The block error rate prediction model is further specified as a feature extraction neural network, a block error rate prediction neural network, and a pseudo-block error rate prediction neural network.

[0033] Step 320: The specific steps are as follows:

[0034] Step 321: For target domain data with unknown block error rate First, the feature extraction neural network is used to... and Context The latent contexts after representation are defined as follows: and , Represents the contextual representation space;

[0035] Step 322: Aggregate data representations from various source domains The source domain aggregation characterization is obtained. ;

[0036] Step 323: Aggregate and characterize the source domain Input the block error rate prediction neural network, and use its output block error rate Source domain error rate label Calculate the source domain block error prediction error ,in The mean squared error index function, and These represent the model parameters of the block error rate prediction model and the feature extraction neural network, respectively.

[0037] Step 324: Use The representation is defined in the metric space. Unsupervised joint representation distance, i.e., "source domain aggregation representation" -Source domain block error rate "and target domain data representation" - Pseudo-block error rate "The Wasserstein-1 distance between two joint representations, where, express and The joint space above, express The metric L represents the performance of the block error prediction neural network in the representation space. The Lipschitz coefficients on the function, M represents the function The Lipschitz coefficient, express Ideal pseudo-target distribution on, express Ideal distribution of the source domain;

[0038] Step 325: Calculate the distance of the unsupervised joint representation. and the source domain block error prediction error Addition as a sub-objective of knowledge transfer for unsupervised block error rate ,Right now ;

[0039] Step 330: The specific steps are as follows:

[0040] Step 331: For the known block error rate First, the same processing procedure as for target domain data with unknown block error rates is adopted to obtain the first sub-objective of supervised block error rate knowledge transfer. ,Right now , in, express In space The ideal distribution on, The representation is defined in the metric space. The supervised joint representation of distance, and , This indicates the target error rate for prediction;

[0041] Step 332: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] In Input the feature extraction and block error rate prediction neural network, and use its output block error rate Source domain error rate label Obtain the target domain prediction error The first sub-objective of supervised block error rate knowledge transfer is combined with the aforementioned knowledge transfer. The supervised error rate knowledge transfer sub-objective is obtained. , in These are the weighting coefficients; Step 340: The specific steps are as follows: Step 341: Aggregate the unsupervised and supervised block error rate knowledge transfer sub-objectives to obtain the following cross-MCS domain block error rate knowledge transfer objective: , in For weighting coefficients; when hour, It is a supervised objective. The time is an unsupervised target. Indicates a semi-supervised objective; According to the source domain With the target domain The proportion of the dataset is selected accordingly; Step 342: Given a dataset and ,definition ; in, Represents distribution The empirical block error prediction error is in the form of: , in, This represents the model parameters of the pseudo-block error rate prediction model; combined with the target... Minimization requirement, using and approximate Wasserstein-1 distance in and This leads to two sub-optimization problems: min-max and min-max. ; By introducing a gradient inversion layer into the pseudo-block rate prediction neural network Defined as ; in, This refers to the pseudo-block error rate prediction model. Indicates the gradient reversal coefficient; By reversing the gradient direction of a specific portion, the optimization of the maximization objective in the min-max problem can be directly achieved; thus, the objective... In and They are approximated as ; in These are the weighting coefficients; ultimately, minimize... That is, it simultaneously achieves the transfer of knowledge about the block error rate and the approximation of the Wasserstein-1 distance.

[0042] Furthermore, step 400 specifically includes the following steps: Based on data-driven learning generalization and relative entropy theory, the gradient norm of the model parameters of the feature extraction neural network and the block error rate prediction neural network, as well as the number of samples in the domain dataset, are used to establish the upper bound sub-objective of generalization, namely... , in, and Represents the weight coefficients related to the dataset. and K represents the cumulative model parameter gradient norm, and K represents the number of iterations of the stochastic gradient optimization algorithm. This represents the variance of the noise injected during the k-th update of the model parameters. and These represent the update step sizes of the feature extraction neural network and the block error rate prediction neural network parameters, respectively. and They represent the optimization objectives respectively. Regarding model parameters and The gradient.

[0043] Furthermore, the optimization objective of the block error rate prediction model in step 500 is: .

[0044] Furthermore, the specific steps of step 600 are as follows: Step 610: Inner layer optimization stage, given the optimized source domain data weight vector Based on steps 200, 300, 400, and 500, the error block rate prediction model is constructed to optimize the objective. The parameters of the feature extraction neural network, the block error rate prediction neural network, and the pseudo-block error rate prediction neural network are updated using the stochastic gradient Langevin dynamics optimization algorithm. ; Step 620: In the outer layer optimization stage, the source domain data weight vector is... Using the source domain block error prediction error and the model parameter gradient norm obtained in the inner optimization stage as optimization variables, and constraining the sum of the source domain data weight vectors to be one, the following convex optimization problem is constructed: , in, , and These represent the weight coefficients related to the source domain prediction error and the gradient norm of the model parameters, respectively; the convex optimization problem is solved, and the source domain data weight vector is updated using a moving average method. .

[0045] Furthermore, step 700 specifically includes the following steps: After the block error prediction model is deployed, the block error prediction error of the domain dataset is evaluated at fixed time intervals. If the ratio of the block error prediction error in the current evaluation period to the prediction error in the previous period exceeds a set threshold, it is determined that the distribution of the domain data has shifted significantly. At this time, the block error prediction model should be re-tuned according to step 600.

[0046] Beneficial effects: Compared with existing technologies, this invention filters context fields strongly correlated with the block error rate through a wireless knowledge graph, reducing data dependence and processing complexity while ensuring prediction accuracy; it is the first to explicitly incorporate the monotonic relationship between context fields and the block error rate into the model, enhancing interpretability and generalization ability; it proposes a unified unsupervised and supervised knowledge transfer objective, improving the model's adaptability in out-of-distribution scenarios through joint representation alignment; and it constructs a generalization upper bound for block error rate prediction based on data-driven learning theory, constraining model optimization to improve its robustness against distribution changes. This invention can provide more accurate and robust block error rate prediction support for applications such as network digital twins. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 This is a flowchart of a cross-domain knowledge transfer communication block rate intelligent prediction and deployment method provided in a specific embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram illustrating the joint distribution of "context-block error rate" when MCS is used as the domain segmentation field, provided by a specific embodiment of the present invention.

[0050] Figure 3 This is a flowchart illustrating the process of using source domain data weight vectors to aggregate source domain data, aligning context representation with the joint representation of error rate, and establishing the error rate knowledge transfer sub-objective, as provided in a specific embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0052] The block error rate prediction model proposed in this invention includes: a feature extraction neural network, a block error rate prediction neural network, and a pseudo-block error rate prediction neural network.

[0053] like Figure 1As shown, this embodiment of the invention provides a method for intelligent prediction and deployment of communication error rate in cross-domain knowledge transfer, the method comprising: Step 100: Utilize the wireless knowledge graph and business data warehouse to extract contextual fields and corresponding datasets strongly correlated with the block error rate, and segment the data into domains based on their significant impact. Specific steps are as follows:

[0054] Step 110: First, define entities, relationships, and triples based on the endogenous factors of the wireless communication network protocol. Then, calculate sparse representation vectors using node random state weights. Measure and replace edge connections in the graph using cosine similarity, and then calculate node association vectors and feature vectors to construct a wireless knowledge graph. Based on the constructed wireless knowledge graph, generate an initial data classification model using endogenous association reasoning. Then, based on the reasoning model, split and lightly aggregate the data. Next, output a preferred association model through weighted reasoning and sorting of association fields. Finally, combine this model with the aggregated data table to generate a structured data warehouse.

[0055] Step 120: Utilize the wireless knowledge graph to perform correlation analysis, extracting context fields strongly correlated with the block error rate (BER) metric, primarily covering fields related to communication resource scheduling and channel quality. Then, extract historical data containing the aforementioned context fields and BER from the business data warehouse, and use a unique context identifier method to statistically analyze the BER under specific contexts, thereby generating the BER dataset. The specific steps are as follows:

[0056] Step 121: For the contextual analysis requirements of the block error rate metric, input the block error rate metric. The wireless knowledge graph calculates and sorts the correlation weights between the associated fields and the required fields for each association category, selecting the top-ranked fields. The contextual fields extracted by the wireless knowledge graph are denoted as... ,in, Indicates the number of resource blocks. Indicates resource block channel quality, Represents the resource block to be scheduled. Indicates the transport block size, Indicates the number of transport blocks, Represents the number of streams, Indicates the modulation order. Indicates bitrate, Indicates the dimension size as The real number field;

[0057] Step 122: Extract data from the business data warehouse based on the extracted context fields; denote the extracted dataset as... ,in, Indicates the number of samples. , and Let the context, (N)ACK feedback value, and coding and modulation scheme order of the i-th sample be represented respectively; the channel quality equivalent operator is scheduled using resource blocks. deal with and , This represents the channel quality of the aggregated and sorted resource blocks, where Represent the Hartmann product; then, Viewed as an empirical distribution, in a specific context The block error rate statistics are as follows: Therefore, we can obtain a size of Block error rate dataset ,in, The operator represents the expectation of a distribution. Domain ;

[0058] Step 123: Based on the clear division of the context field values ​​according to preset rules and the significant differences in the dominant domain dataset distribution, establish a certain context field as the domain segmentation field; using the MCS order as the domain segmentation field, the joint distribution of "context-block error rate" in each domain shows significant differences, such as... Figure 2 As shown;

[0059] Step 124: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The datasets are divided into S domains based on their MCS order. Simultaneously, the target domain dataset is represented as having a sample size of... of Further Classified as block error rate Unknown and block error rate Knowable .

[0060] Step 200: Analyze the monotonicity relationship between the context field and the block error rate using standard encoding / decoding behavior analysis to construct a physically-inspired sub-objective. The specific steps are as follows:

[0061] Step 210: Based on the encoding and decoding behavior of the new air interface standard, after the transport block undergoes cyclic redundancy check, the block error rate is... Regarding the context The partial derivatives generally satisfy ,in and These represent context slices where the partial derivatives are less than 0 and greater than 0, respectively.

[0062] Step 220: Let the block error rate prediction model be... In each iteration round, the corresponding communication context is calculated. The partial derivatives violate the direction of the model; to enhance the physical generalization ability of the model, a random sample generation method is adopted to calculate the partial derivatives. , Among them, synthetic samples , and These represent the contexts from the source domain and the target domain, respectively. The coefficients are randomly synthesized. It is an element-wise non-negative operator that operates on vectors. Represent the 1-norm of a vector; by minimizing , It will tend to satisfy step 210. about The direction of the partial derivative, thereby achieving Its generalization ability in terms of monotonicity, while its more complex behavioral characterization will be achieved through the knowledge transfer objective of error rate based on domain adaptation theory;

[0063] Step 230: Construct the physically inspired sub-objective ,in The regularization coefficient is used to penalize violations of the physical monotonicity of the block error rate. .

[0064] Step 300: Aggregate source domain data using source domain data weight vectors, align the joint representation of contextual representation and error rate, and establish the error rate knowledge transfer sub-objective. For example... Figure 3 As shown, the specific steps are as follows:

[0065] Step 310: Define the source domain data weight vector ,satisfy Where S is the total number of source domains, and the weight of the i-th source domain data is initialized to... .

[0066] Step 320: For target domain data with unknown block error rate, the feature extraction neural network is used to represent the source domain and target domain data; the source domain data representation is weighted and aggregated using the source domain data weight vector to obtain the source domain aggregated representation; the source domain aggregated representation is input into the block error rate prediction neural network to obtain the source domain block error rate prediction error; the block error rate of the target domain is approximated by the pseudo-block error rate output by the pseudo-block error rate prediction neural network; the Wasserstein-1 distance is used to characterize the difference between the two joint representations, "source domain aggregated representation - source domain block error rate" and "target domain data representation - pseudo-block error rate", to obtain the unsupervised joint representation distance; the unsupervised joint representation distance and the source domain block error rate prediction error are added together as the unsupervised block error rate knowledge transfer sub-objective. The specific steps are as follows:

[0067] Step 321: For target domain data with unknown block error rate First, the feature extraction neural network is used to... and Context The latent contexts after representation are defined as follows: and , Represents the contextual representation space;

[0068] Step 322: Aggregate data representations from various source domains The source domain aggregation characterization is obtained. ; Step 323: Aggregate and characterize the source domain Input the block error rate prediction neural network, and use its output block error rate Source domain error rate label Calculate the source domain block error prediction error ,in The mean squared error index function, and These represent the model parameters of the block error rate prediction model and the feature extraction neural network, respectively. Step 324: Use The representation is defined in the metric space. Unsupervised joint representation distance, i.e., "source domain aggregation representation" -Source domain block error rate "and target domain data representation" - Pseudo-block error rate "The Wasserstein-1 distance between two joint representations, where, express and The joint space above, express The metric L represents the performance of the block error prediction neural network in the representation space. The Lipschitz coefficients on the function, M represents the function The Lipschitz coefficient, express Ideal pseudo-target distribution on, express Ideal distribution of the source domain; Step 325: Calculate the distance of the unsupervised joint representation. and the source domain block error prediction error Addition as a sub-objective of knowledge transfer for unsupervised block error rate ,Right now

[0069] ; Step 330: For target domain error rate data with a known error rate, firstly, the same processing procedure as for target domain data with an unknown error rate is adopted to obtain the first sub-objective of supervised error rate knowledge transfer; simultaneously, the target domain error rate data with a known error rate is input into the feature extraction and error rate prediction neural network to obtain the target domain prediction error, which is used as the second sub-objective of supervised error rate knowledge transfer, and the first sub-objective of supervised error rate knowledge transfer is added to it according to weights, and the result is used as the second sub-objective of supervised error rate knowledge transfer. The specific steps are as follows:

[0070] Step 331: For the known block error rate First, the same processing procedure as for target domain data with unknown block error rates is adopted to obtain the first sub-objective of supervised block error rate knowledge transfer. ,Right now ,

[0071] in, express In space The ideal distribution on, The representation is defined in the metric space. The supervised joint representation of distance, and , This indicates the target error rate for prediction;

[0072] Step 332: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] In Input the feature extraction and block error rate prediction neural network, and use its output block error rate Source domain error rate label Obtain the target domain prediction error The first sub-objective of supervised block error rate knowledge transfer is combined with the aforementioned knowledge transfer. The supervised error rate knowledge transfer sub-objective is obtained. , in These are the weighting coefficients;

[0073] Step 340: Combine the unsupervised and supervised block rate knowledge transfer sub-objectives according to weights; using the block rate prediction neural network and the pseudo-block rate prediction neural network, approximate the Wasserstein-1 distance as the maximum difference between the block rate prediction errors of the source domain and the target domain, and introduce a gradient reversal layer into the pseudo-block rate prediction neural network. Then, add the two sub-objectives of unsupervised and supervised block rate knowledge transfer according to weights to finally obtain the block rate knowledge transfer sub-objective. The specific steps are as follows:

[0074] Step 341: Aggregate the unsupervised and supervised block error rate knowledge transfer sub-objectives to obtain the following cross-MCS domain block error rate knowledge transfer objective: ,

[0075] in For weighting coefficients; when hour, It is a supervised objective. The time is an unsupervised target. Indicates a semi-supervised objective; According to the source domain With the target domain The proportion of the dataset is selected accordingly; Step 342: Given a dataset and ,definition ; in, Represents distribution The empirical block error prediction error is in the form of: ; This represents the model parameters of the pseudo-block error rate prediction model. Combined with the target... Minimization requirement, using and approximate Wasserstein-1 distance in and This leads to two sub-optimization problems: min-max and min-max. ; By introducing a gradient inversion layer into the pseudo-block rate prediction neural network Defined as ; in, This refers to the pseudo-block error rate prediction model. Indicates the gradient reversal coefficient; By reversing the gradient direction of a specific portion, the optimization of the maximization objective in the min-max problem can be directly achieved; thus, the objective... In and They are approximated as ; in These are the weighting coefficients. Finally, minimize... That is, it simultaneously achieves the transfer of knowledge about the block error rate and the approximation of the Wasserstein-1 distance.

[0076] Step 400: Construct a generalized upper bound sub-objective using the parameter gradient norm of the error rate prediction model.

[0077] Based on data-driven learning generalization and relative entropy theory, the gradient norm of the model parameters of the feature extraction neural network and the block error rate prediction neural network, as well as the number of samples in the domain dataset, are used to establish the upper bound sub-objective of generalization, namely... , in, and Represents the weight coefficients related to the dataset. and K represents the cumulative model parameter gradient norm, and K represents the number of iterations of the stochastic gradient optimization algorithm. This represents the variance of the noise injected during the k-th update of the model parameters. and These represent the update step sizes of the feature extraction neural network and the block error rate prediction neural network parameters, respectively. and They represent the optimization objectives respectively. Regarding model parameters and The gradient.

[0078] Step 500: Weighted fusion of the three sub-objectives—physical inspiration, block error rate knowledge transfer, and generalization upper bound—forms the block error rate prediction model optimization objective: .

[0079] Step 600 involves a two-stage optimization and tuning of the block error rate prediction model: the inner layer optimizes the model parameters, and the outer layer optimizes the source domain data weight vector. The specific steps are as follows: Step 610: Inner layer optimization stage, given the optimized source domain data weight vector Based on steps 200, 300, 400, and 500, the error block rate prediction model is constructed to optimize the objective. The parameters of the feature extraction neural network, the block error rate prediction neural network, and the pseudo-block error rate prediction neural network are updated using the stochastic gradient Langevin dynamics optimization algorithm. ; Step 620: In the outer layer optimization stage, the source domain data weight vector is... Using the source domain block error prediction error and the model parameter gradient norm obtained in the inner optimization stage as optimization variables, and constraining the sum of the source domain data weight vectors to be one, the following convex optimization problem is constructed: , in, , and These represent the weight coefficients related to the source domain prediction error and the gradient norm of the model parameters, respectively. The convex optimization problem is solved, and the source domain data weight vector is updated using a moving average method. .

[0080] Step 700: After the model is deployed, continuously monitor the distribution changes of the domain data. If a significant shift occurs, trigger model retuning.

[0081] After the block error prediction model is deployed, the block error prediction error of the domain dataset is evaluated at fixed time intervals. If the ratio of the block error prediction error in the current evaluation period to the prediction error in the previous period exceeds a set threshold, it is determined that the distribution of the domain data has shifted significantly. At this time, the block error prediction model should be re-tuned according to step 600.

Claims

1. A method for intelligent prediction and deployment of communication error rate in cross-domain knowledge transfer, characterized in that, Includes the following steps: Step 100: Extract context fields and corresponding datasets that are strongly related to the error rate using wireless knowledge graphs and business data warehouses, and divide the domain data according to their significant impact. Step 200: Combine standard encoding and decoding behavior analysis with the monotonicity relationship between context fields and block error rate to construct a physically inspired sub-target; Step 300: Aggregate source domain data using source domain data weight vectors, implement joint representation alignment of context representation and error block rate, and establish the knowledge transfer sub-objective of error block rate; Step 400: Construct a generalized upper bound sub-objective using the parameter gradient norm of the error block rate prediction model; Step 500: Weighted fusion of the three sub-objectives—physical inspiration, block error rate knowledge transfer, and generalization upper bound—to form the block error rate prediction model optimization objective; Step 600: Optimize and tune the error block rate prediction model in two stages: optimize the model parameters in the inner layer and optimize the source domain data weight vector in the outer layer. Step 700: After the model is deployed, continuously monitor the distribution changes of the domain data. If a significant shift occurs, trigger model retuning.

2. The intelligent prediction and deployment method for cross-domain knowledge transfer communication block rate according to claim 1, characterized in that, Step 100 specifically includes the following steps: Step 110: Construct a wireless knowledge graph and business data warehouse; Step 120: Utilize the wireless knowledge graph to perform correlation analysis, extract context fields that are strongly correlated with the block error rate index, covering field types such as communication resource scheduling and channel quality; then, extract historical data containing the aforementioned context fields and block error rates from the business data warehouse, and use the context unique identifier method to statistically analyze the block error rate under a specific context, thereby generating a block error rate dataset.

3. The intelligent prediction and deployment method for cross-domain knowledge transfer communication block rate according to claim 2, characterized in that, The specific steps for step 120 are as follows: Step 121: For the contextual analysis requirements of the block error rate metric, input the block error rate metric. The wireless knowledge graph calculates and sorts the correlation weights between the associated fields and the required fields for each association category, selecting the top-ranked fields. The contextual fields extracted by the wireless knowledge graph are denoted as... ,in, Indicates the number of resource blocks. Indicates resource block channel quality, Represents the resource block to be scheduled. Indicates the transport block size, Indicates the number of transport blocks, Represents the number of streams, Indicates the modulation order. Indicates bitrate, Indicates the dimension size as The real number field; Step 122: Extract data from the business data warehouse based on the extracted context fields; denote the extracted dataset as... ,in, Indicates the number of samples. , and Let the context, (N)ACK feedback value, and coding and modulation scheme order of the i-th sample be represented respectively; the channel quality equivalent operator is scheduled using resource blocks. deal with and , This represents the channel quality of the aggregated and sorted resource blocks, where Represent the Hartmann product; then, Viewed as an empirical distribution, in a specific context The block error rate statistics are as follows: Therefore, we can obtain a size of Block error rate dataset ,in, This represents the operator for calculating the expectation of a distribution. Domain ; Step 123: Based on the fact that the values ​​of the context fields are clearly divided according to preset rules and the dominant domain datasets show significant differences in distribution, a certain context field is established as the domain segmentation field; with the MCS order as the domain segmentation field, the joint distribution of "context-block error rate" in each domain shows significant differences; Step 124: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The datasets are divided into S domains based on their MCS order. Simultaneously, the target domain dataset is represented as having a sample size of... of Further Classified as block error rate Unknown and block error rate Knowable .

4. The intelligent prediction and deployment method for communication block rate in cross-domain knowledge transfer according to claim 3, characterized in that, Step 200: The specific steps are as follows: Step 210: Based on the encoding and decoding behavior of the new air interface standard, after the transport block undergoes cyclic redundancy check, the block error rate is... Regarding the context The partial derivatives generally satisfy ,in and These represent context slices where the partial derivatives are less than 0 and greater than 0, respectively. Step 220: Let the block error rate prediction model be... In each iteration round, the corresponding communication context is calculated. The partial derivatives violate the direction of the model; to enhance the physical generalization ability of the model, a random sample generation method is adopted to calculate the partial derivatives. ， Among them, synthetic samples , and These represent the contexts from the source domain and the target domain, respectively. The coefficients are randomly synthesized. It is an element-wise non-negative operator that operates on vectors. Represent the 1-norm of a vector; by minimizing , It will tend to satisfy step 210. about The direction of the partial derivative, thereby achieving Its generalization ability in terms of monotonicity, while its more complex behavioral characterization will be achieved through the knowledge transfer objective of error rate based on domain adaptation theory; Step 230: Construct the physically inspired sub-objective ,in The regularization coefficient is used to penalize violations of the physical monotonicity of the block error rate. .

5. The intelligent prediction and deployment method for communication block rate in cross-domain knowledge transfer according to claim 4, characterized in that, Step 300 specifically includes the following steps: Step 310: Define and initialize the source domain data weight vector with equal weights; the block error rate prediction model specifically includes: a feature extraction neural network, a block error rate prediction neural network, and a pseudo-block error rate prediction neural network; Step 320: For target domain data with unknown block error rate, the feature extraction neural network is used to represent the source domain and target domain data; the source domain data representation is weighted and aggregated using the source domain data weight vector to obtain the source domain aggregated representation; the source domain aggregated representation is input into the block error rate prediction neural network to obtain the source domain block error rate prediction error; the block error rate of the target domain is approximated by the pseudo block error rate output by the pseudo block error rate prediction neural network; the Wasserstein-1 distance is used to characterize the difference between the two joint representations "source domain aggregated representation - source domain block error rate" and "target domain data representation - pseudo block error rate" to obtain the unsupervised joint representation distance; the unsupervised joint representation distance and the source domain block error rate prediction error are added together as the unsupervised block error rate knowledge transfer sub-objective; Step 330: For target domain error rate data with known error rates, firstly, the same processing procedure as for target domain data with unknown error rates is adopted to obtain the first sub-objective of supervised error rate knowledge transfer; simultaneously, the target domain data with known error rates is input into the feature extraction and error rate prediction neural network to obtain the target domain prediction error, which is used as the second sub-objective of supervised error rate knowledge transfer, and the first sub-objective of supervised error rate knowledge transfer is added according to weights, and the result is used as the supervised error rate knowledge transfer sub-objective. Step 340: Combine the unsupervised and supervised block rate knowledge transfer sub-objectives according to weights; using the block rate prediction neural network and the pseudo-block rate prediction neural network, approximate the Wasserstein-1 distance as the maximum difference between the block rate prediction errors of the source domain and the target domain, and introduce a gradient reversal layer into the pseudo-block rate prediction neural network. Then, add the two sub-objectives of unsupervised and supervised block rate knowledge transfer according to weights to finally obtain the block rate knowledge transfer sub-objective.

6. The intelligent prediction and deployment method for communication block rate in cross-domain knowledge transfer according to claim 5, characterized in that, In step 310, the source domain data weight vector is defined. ,satisfy Where S is the total number of source domains, and the weight of the i-th source domain data is initialized to... The block error rate prediction model is further specified as a feature extraction neural network, a block error rate prediction neural network, and a pseudo-block error rate prediction neural network. Step 320: The specific steps are as follows: Step 321: For target domain data with unknown block error rate First, the feature extraction neural network is used to... and Context The latent contexts after representation are defined as follows: and , Represents the contextual representation space; Step 322: Aggregate data representations from various source domains The source domain aggregation characterization is obtained. ; Step 323: Aggregate and characterize the source domain Input the block error rate prediction neural network, and use its output block error rate Source domain error rate label Calculate the source domain block error prediction error ,in The mean squared error index function, and These represent the model parameters of the block error rate prediction model and the feature extraction neural network, respectively. Step 324: Use The representation is defined in the metric space. Unsupervised joint representation distance, i.e., "source domain aggregation representation" -Source domain block error rate "and target domain data representation" - Pseudo-block error rate "The Wasserstein-1 distance between the two joint representations, where, express and The joint space above, express The metric L represents the performance of the block error prediction neural network in the representation space. The Lipschitz coefficients on the function, M represents the function The Lipschitz coefficient, express Ideal pseudo-target distribution on, express Ideal distribution of the source domain; Step 325: Calculate the distance of the unsupervised joint representation. and the source domain block error prediction error Addition as a sub-objective of knowledge transfer for unsupervised block error rate ,Right now ； Step 330: The specific steps are as follows: Step 331: For the known block error rate First, the same processing procedure as for target domain data with unknown block error rates is adopted to obtain the first sub-objective of supervised block error rate knowledge transfer. ,Right now ， in, express In space The ideal distribution on, The representation is defined in the metric space. The supervised joint representation of distance, and , This indicates the target error rate for prediction; Step 332: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] In Input the feature extraction and block error rate prediction neural network, and use its output block error rate Source domain error rate label Obtain the target domain prediction error The first sub-objective of knowledge transfer based on the supervised block error rate is combined. The supervised block error rate knowledge transfer sub-objective is obtained. ， in These are the weighting coefficients; Step 340: The specific steps are as follows: Step 341: Aggregate the unsupervised and supervised block error rate knowledge transfer sub-objectives to obtain the following cross-MCS domain block error rate knowledge transfer objective: ， in For weighting coefficients; when hour, It is a supervised objective. The time is an unsupervised target. Indicates a semi-supervised objective; According to the source domain With the target domain The proportion of the dataset is selected accordingly; Step 342: Given a dataset and ,definition ； in, Represents distribution The empirical block error prediction error is in the form of: ， in This represents the model parameters of the pseudo-block error rate prediction model; combined with the target... Minimization requirement, using and approximate Wasserstein-1 distance in and This leads to two sub-optimization problems: min-max and min-max. ； By introducing a gradient inversion layer into the pseudo-block rate prediction neural network Defined as ； in, This represents the pseudo-block error rate prediction model. Indicates the gradient reversal coefficient; By reversing the gradient direction of a specific portion, the optimization of the maximization objective in the min-max problem can be directly achieved; thus, the objective... In and They are approximated as ； in These are the weighting coefficients; ultimately, minimize... That is, it simultaneously achieves the transfer of knowledge about the block error rate and the approximation of the Wasserstein-1 distance.

7. The intelligent prediction and deployment method for communication block rate in cross-domain knowledge transfer according to claim 6, characterized in that, Step 400 specifically includes the following steps: Based on data-driven learning generalization and relative entropy theory, the gradient norm of the model parameters of the feature extraction neural network and the block error rate prediction neural network, as well as the number of samples in the domain dataset, are used to establish the upper bound sub-objective of generalization, namely... ， in, and Represents the weight coefficients related to the dataset. and K represents the cumulative model parameter gradient norm, and K represents the number of iterations of the stochastic gradient optimization algorithm. This represents the variance of the noise injected during the k-th update of the model parameters. and These represent the update step sizes of the feature extraction neural network and the block error rate prediction neural network parameters, respectively. and They represent the optimization objectives respectively. Regarding model parameters and The gradient.

8. The intelligent prediction and deployment method for communication block rate in cross-domain knowledge transfer according to claim 7, characterized in that, The optimization objective of the block error rate prediction model in step 500 is: 。 9. The intelligent prediction and deployment method for communication block rate in cross-domain knowledge transfer according to claim 8, characterized in that, The specific steps for step 600 are as follows: Step 610: Inner layer optimization stage, given the optimized source domain data weight vector Based on steps 200, 300, 400, and 500, the error block rate prediction model is constructed to optimize the objective. The parameters of the feature extraction neural network, the block error rate prediction neural network, and the pseudo-block error rate prediction neural network are updated using the stochastic gradient Langevin dynamics optimization algorithm. ; Step 620: In the outer layer optimization stage, the source domain data weight vector is... Using the source domain block error prediction error and the model parameter gradient norm obtained in the inner optimization stage as optimization variables, and constraining the sum of the source domain data weight vectors to be one, the following convex optimization problem is constructed: ， in, , and These represent the weight coefficients related to the source domain prediction error and the gradient norm of the model parameters, respectively; the convex optimization problem is solved, and the source domain data weight vector is updated using a moving average method. .

10. The intelligent prediction and deployment method for communication block rate in cross-domain knowledge transfer according to claim 9, characterized in that, Step 700 specifically includes the following steps: After the block error prediction model is deployed, the block error prediction error of the domain dataset is evaluated at fixed time intervals. If the ratio of the block error prediction error in the current evaluation period to the prediction error in the previous period exceeds a set threshold, it is determined that the distribution of the domain data has shifted significantly. At this time, the block error prediction model should be re-tuned according to step 600.

Images