Multi-mode communication protocol optimization method based on deep learning

Abstract

The invention discloses a multi-mode communication protocol optimization method based on deep learning, and the method comprises the following steps: S1, collecting operation data of a communication terminal, and carrying out the preprocessing of the operation data; s2, constructing a multi-mode communication protocol structure diagram, and generating an adjacent matrix; s3, constructing a structured LSTM model, and modeling time sequence characteristics of a communication environment and structural dependence information between protocols; s4, performing performance prediction by using the structured LSTM model; s5, constructing a protocol cause subgraph between the time sequence feature vector and the performance prediction vector; and S6, performing multi-target comprehensive evaluation on the communication protocol, and determining an optimal communication protocol configuration scheme. According to the method, structured deep learning and causal modeling are fused, a multimode communication protocol selection process is optimized, and the method has the advantages of high prediction precision, high decision interpretability and good environmental adaptability.

Description

Technical Field

[0001] The present invention relates to the field of multimode communications, and in particular to a multimode communication protocol optimization method based on deep learning. Background Art

[0002] With the rapid development of wireless communication technology and the widespread popularity of multi-mode terminal devices, users are faced with an environment where multiple communication protocols coexist in their daily communications, such as LTE, 5G, Wi-Fi, Bluetooth, etc. Different protocols have their own advantages in terms of coverage, transmission rate, energy consumption, connection stability, etc. Dynamically switching protocols based on network status by communication terminals has become a key strategy to improve communication quality and resource utilization. In existing technologies, some studies have attempted to achieve switching control of multi-mode communication protocols through rule-driven or historical statistics-based methods. However, such methods rely on preset logic, have weak adaptability to the environment, and are difficult to make optimal judgments in dynamic and complex scenarios. In addition, existing studies have a relatively superficial understanding of the correlation between multi-mode communication protocols and often ignore the complex dependencies between communication protocols in switching behavior, mechanism structure, and resource hierarchy.

[0003] In recent years, with the widespread application of deep learning technology, some work has begun to introduce methods such as recurrent neural networks and graph neural networks into communication systems for tasks such as modeling network status and predicting communication indicators. However, these methods often focus on a single feature or static network topology, making it difficult to achieve joint modeling of communication protocol structure and performance in the protocol dimension.

[0004] On the other hand, there are not only switching paths between communication protocols, but also differences at the mechanism level such as frame structure, scheduling method, retransmission mechanism, as well as competitiveness at the resource level such as spectrum resource occupancy and mutual exclusion. Current research lacks the ability to uniformly model these multi-dimensional protocol dependencies, and fails to fully reflect the structural dependencies between protocols at the modeling level. For protocol performance evaluation, most still adopt strategies based on fixed weights or black box predictions, and fail to explicitly model the causal mechanism between changes in the communication environment and changes in protocol performance, resulting in insufficient sensitivity and adaptability of optimization strategies to environmental changes. Summary of the Invention

[0005] One purpose of the present invention is to propose a multi-mode communication protocol optimization method based on deep learning. The present invention integrates structured deep learning and causal modeling methods, constructs a multi-mode communication protocol structure diagram and embeds an LSTM model, combines the communication environment characteristics with the causal relationship between protocol performance, optimizes protocol selection and switching strategies, and has the advantages of high prediction accuracy, transparent decision-making, strong environmental adaptability, and support for multi-objective evaluation, significantly improving the intelligence level and stability of the multi-mode communication system.

[0006] A multi-mode communication protocol optimization method based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] S1. Collecting operation data of the communication terminal in a multi-mode communication environment, and pre-processing the operation data to generate a time series feature vector;

[0008] S2. Construct a multi-mode communication protocol structure diagram, model the structural dependency between multi-mode communication protocols, and generate an adjacency matrix;

[0009] S3. Build a structured LSTM model. Based on the time series feature vector and adjacency matrix, perform information transfer of the structured LSTM model, model the time series characteristics of the communication environment and the structural dependency information between protocols, and generate a structure-enhanced hidden state vector.

[0010] S4, inputting the time series feature vector and the adjacency matrix into the structured LSTM model, and outputting the performance prediction vector of the communication protocol based on the structure-enhanced hidden state vector;

[0011] S5. Generate a causal weight matrix based on the performance prediction vector and the hidden state vector of the structured LSTM model, and construct a protocol causal subgraph between the time series feature vector and the performance prediction vector;

[0012] S6. Based on the causal weight matrix and the protocol causal subgraph, a multi-objective comprehensive evaluation of the communication protocol is performed, and the optimal communication protocol configuration scheme is determined based on the set comprehensive scoring function.

[0013] Optionally, the operating data includes channel quality parameters, network load status, and terminal coordinates. The channel quality parameters include reference signal power, reference signal reception quality, signal-to-interference-and-noise ratio, channel quality indication, and block error rate. The network load status includes the number of user connections, physical resource block utilization, and scheduling queue length.

[0014] Optionally, the preprocessing includes time alignment, feature normalization, outlier removal and missing value filling.

[0015] Optionally, the S2 specifically includes:

[0016] S21. Setting a communication protocol set, where the communication protocol set includes LTE, 5G, Wi-Fi, and Bluetooth;

[0017] S22. Construct a protocol switching matrix based on the historical communication protocol switching records of the communication terminal:

[0018]

[0019] Among them, A i,j Represents the elements in the protocol switching matrix, describing the communication protocol p i With pj Whether there is a switching relationship between them;

[0020] S23. Construct the mechanism characteristics of each communication protocol into a mechanism feature vector, and combine them to form a protocol mechanism similarity matrix. Each element in the protocol mechanism similarity matrix represents the structural similarity of the communication protocol. The mechanism feature vector is composed of structural mechanism feature codes. The mechanism features include frame structure, scheduling mode, retransmission mechanism, control signaling structure, operating frequency band, and access mode. Each mechanism feature is one-hot encoded.

[0021] S24. Identify resource conflict relationships between protocols and build a resource competition matrix:

[0022]

[0023] Among them, R i,j Represents the elements in the resource competition matrix, describing the communication protocol p i With p j The resource hierarchical mutually exclusive relationship between them;

[0024] S25. Integrate the protocol switching matrix, the protocol mechanism similarity matrix, and the resource competition matrix, and define the multimode communication protocol structure graph as G = (V, E), where V represents the node set of the multimode communication protocol structure graph, and E represents the edge set of the multimode communication protocol structure graph:

[0025] E={(p i ,p j )∣A i,j =1∨S i,j >θ s ∨R i,j =1};

[0026] Where E represents the edge set of the multi-mode communication protocol structure graph, describing the establishment of a connection when there is a switching behavior between communication protocols, the structural similarity is greater than the threshold, or the resource competition condition is met, θ s represents the similarity threshold, ∨ represents the “and” operation, p i represents the i-th communication protocol, p j represents the jth communication protocol;

[0027] S26. Represent the multi-mode communication protocol structure diagram as an adjacency matrix, where the adjacency matrix indicates whether there is an edge connection relationship between nodes, and the edge connection relationship includes switching behavior, structural similarity, and resource contention.

[0028] Optionally, the S3 specifically includes:

[0029] S31, setting the input data to a time series feature vector;

[0030] S32. Construct a structured LSTM model, wherein the structured LSTM model includes an input layer, a structured LSTM hidden layer, and an output layer. The structured LSTM hidden layer, based on the traditional LSTM hidden layer, introduces a multi-mode communication protocol structure diagram as an explicit control signal to structurally constrain the information transmission path and strength between each hidden unit;

[0031] S33. At each time step, calculate the input gate, forget gate, and output gate vectors of the structured LSTM model:

[0032] i t =σ(W i x t +U i h t-1 +b i );

[0033] f t =σ(W f x t +U f h t-1 +b f );

[0034] o t =σ(W o x t +U o h t-1 +b o );

[0035] Among them, i t represents the input gate vector at time step t, f t Represents the forget gate vector at time step t, o t represents the output gate vector at time step t, W i ,W f ,W o represents the input weight matrix, U i ,U f ,U o represents the hidden state transfer weight matrix, b i ,b f ,b o represents the bias vector, σ(·) represents the Sigmoid activation function, h t-1 represents the hidden state vector at time step t-1, x t represents the time series feature vector at time step t;

[0036] S34. During the information transmission process, the candidate memory unit state, memory unit state and hidden state vector are calculated:

[0037]

[0038] h t =o t ⊙tanh(c t );

[0039] in, represents the candidate memory cell state vector at time step t, c t represents the memory cell state vector at time step t, h t represents the hidden state vector at time step t, W c Represents the input weight matrix of the candidate state, U c Represents the hidden state weight matrix of the candidate state, b c Represents the bias vector of the candidate state, tanh(·) represents the hyperbolic tangent activation function, and ⊙ represents the Hadamard product, i.e., element-wise multiplication;

[0040] S35. Use the adjacency matrix to introduce a structural dependency control mechanism during the propagation process at each time step, and introduce the aggregated influence from adjacent nodes to the hidden state of each protocol:

[0041]

[0042] in, Indicates the communication protocol p i The structure-enhanced hidden state vector at time step t+1, Indicates the communication protocol p i The structure-enhanced hidden state vector at time step t, G represents the adjacency matrix, and G(i, j) represents the protocol p i With p j The structural connection relationship of M j→i (·) represents the protocol p j For protocol p i The mapping operation is performed using a multi-layer perceptron. Indicates the communication protocol p j The structure-enhanced hidden state vector at time step t, N represents the total number of communication protocols.

[0043] Optionally, the S4 specifically includes:

[0044] S41. Protocol performance prediction based on the structure-enhanced hidden state vector output by the structured LSTM model;

[0045] S42. Construct a corresponding performance prediction header for each communication protocol, process the structure-enhanced hidden state vector using a protocol-specific nonlinear mapping function, and output a performance prediction vector for the communication protocol. The performance prediction vector includes throughput, communication delay, and energy consumption:

[0046] y k=W k2 ReLU(W k1 ·h T +b k1 )+b k2 ;

[0047] Among them, y k Denotes the performance prediction vector of the communication protocol, W k1 ,W k2 Represents the weight matrix of the two-layer fully connected structure, b k1 , b k2 represents the bias vector, ReLU represents the nonlinear activation function, and h T represents the structure-enhanced hidden state vector;

[0048] S43. Execute performance prediction on all protocols in the communication protocol set to construct a protocol performance prediction set:

[0049]

[0050] in, represents the set of protocol performance predictions, y k Represents the performance prediction vector corresponding to the k-th communication protocol.

[0051] Optionally, the S5 specifically includes:

[0052] S51, obtaining a time series feature vector of the communication environment, a hidden state vector of the structured LSTM model, and a performance prediction vector of the communication protocol;

[0053] S52. Expand the performance prediction vectors of all communication protocols into a performance indicator set, where each element in the performance indicator set represents a performance indicator, including throughput, communication delay, and energy consumption. Extract the time series aggregation vector of each time series feature vector by time-decay weighted averaging:

[0054]

[0055] α t =exp(-λ·(Tt));

[0056] Among them, q i represents the time series aggregation vector of the time series feature vector, α t represents the decay weight of time step t, λ represents the time decay factor, represents the i-th time series feature vector in the t-th time step, Z represents the normalization coefficient of the weights of all time steps, T represents the total number of time steps, and exp represents the natural exponential function;

[0057] S53. Construct a performance indicator vector for each performance indicator. Take the hidden state vector of the structured LSTM model as input and obtain it by superimposing a nonlinear activation function on two layers of linear transformation:

[0058] k j =φ(W j2 ·(W j1 ·h t +b j1 )+b j2 );

[0059] Among them, k j represents the performance indicator vector, W j1 Represents the weight matrix of the first layer transformation, W j2 represents the weight matrix of the second layer transformation, b j1 represents the first layer bias vector, b j2 represents the second layer bias vector, φ(·) represents the activation function, using the ReLU function, h t represents the hidden state vector at time step t;

[0060] S54. Use the time series aggregation vector and performance indicator vector to construct a causal weight matrix:

[0061]

[0062] Among them, C represents the causal weight matrix, C i,j Represents the elements of the causal weight matrix, describing the effect of the time series feature vector on the performance indicator y j The causal influence weight, k j Represents the performance index y j The performance indicator vector, k l Represents the performance index y l performance indicator vector, exp(·) represents the exponential function, and B represents the transposition operation;

[0063] S55. Set the causal influence threshold, filter the index pairs whose causal influence weight is greater than the causal influence threshold, and construct the causal edge set:

[0064] ε k ={(x i ,y j )∣C i,j >θ c};

[0065] Among them, ε k represents the set of causal edges, θ c represents the causal influence threshold, y j represents the performance index, x i represents the time series feature vector, (x i ,yj ) represents the causal edge between the time series feature vector and the performance indicator;

[0066] S56. Construct causal graph G k =(V k ,ε k )), where V k Represents the node set of the causal graph, including all time series feature vectors and performance indicators, ε k Represent the set of causal edges and construct the protocol causal subgraph, including;

[0067] For each communication protocol, determine a set of performance indicators, and use the performance indicators of this communication protocol as the target node;

[0068] Traverse all causal edges in the causal graph and filter the edges whose target nodes belong to the performance indicator set. When the causal weight of the causal edge is greater than or equal to the preset threshold, the edge is determined to be a valid influence path. The causal edges that meet the screening conditions and their source nodes and target nodes are combined into the edge set and node set of the initial protocol causal subgraph to construct the initial protocol causal subgraph;

[0069] The maximum backtracking depth is set. Starting from each timing feature vector node in the initial protocol causal subgraph, a breadth-first search algorithm is used to recursively trace back its upstream causal variables, identify indirect causal paths that do not exceed the maximum backtracking depth, and add the intermediate nodes and edges involved in the indirect causal paths to the initial protocol causal subgraph to obtain a complete protocol causal subgraph. The protocol causal subgraph describes the key causal dependency between the performance indicators and timing characteristics of the current communication protocol in the communication environment.

[0070] Optionally, the S6 specifically includes:

[0071] S61. Calculate the weight factors of all causal edges in the causal graph based on the causal graph and the causal weight matrix:

[0072]

[0073] Among them, ξ represents the weight factor of the causal graph, ε k represents the set of causal edges, (x i ,y j ) represents the causal edge between the time series feature vector and the performance indicator;

[0074] S62. For each communication protocol, obtain the protocol causal subgraph of the performance indicator, extract the relevant causal edge set from the protocol causal subgraph, and calculate the indicator causal weight of the communication protocol:

[0075]

[0076] Among them, wT , w D , w E They represent the causal weights of the communication protocol indicators in the three dimensions of throughput, communication delay and energy consumption, ε k represents the set of causal edges, T k ,D k ,E k represent throughput, communication delay and energy consumption respectively, Respectively represent the performance index T k , D k , E k The causal influence weight, (x i ,T k ) represents the causal edge between the timing feature vector and throughput, (x i ,D k ) represents the causal edge between the timing feature vector and the communication delay, (x i ,E k ) represents the causal edge between the time series feature vector and energy consumption, ε r Represents the set of relevant causal edges in the protocol causal subgraph;

[0077] S63. Adjust the performance indicators based on the causal weights of the indicators to construct weighted performance indicators of the communication protocol:

[0078]

[0079] in, represent the weighted throughput, weighted delay and weighted energy consumption after causal enhancement respectively;

[0080] S64. Calculate statistics based on the weighted performance indicators of all communication protocols and construct a comprehensive scoring function for the communication protocols:

[0081]

[0082] Among them, S k represents the comprehensive scoring function of the communication protocol, α, β, γ represent the weighted coefficients of the throughput, delay and energy consumption scoring dimensions respectively, ∈ represents a positive number to prevent the introduction of zero division errors, μ T ,σ D ,σ E They represent the mean of weighted throughput, the standard deviation of weighted delay, and the standard deviation of weighted energy consumption respectively;

[0083] S65. Select the communication protocol index with the highest score:

[0084]

[0085] Among them, k *represents the index of the communication protocol with the highest score in the comprehensive scoring function, S k represents the comprehensive score of the kth communication protocol, and argmax represents the variable value that maximizes the function value;

[0086] S66. Generate optimal communication protocol configuration solution in represents the optimal communication protocol, Indicates the kth * The performance prediction vector corresponding to the communication protocol, represents k * The protocol causal subgraph of a communication protocol.

[0087] Optionally, the statistics include the mean of weighted throughput after modulation of all communication protocols, the standard deviation of weighted delay, and the standard deviation of weighted energy consumption.

[0088] The beneficial effects of the present invention are:

[0089] First, by introducing a structured deep learning model and a causal reasoning mechanism, this invention addresses the key issues in the optimization process of multimodal communication protocols and proposes a new method with environmental adaptability, structural perception, and decision interpretability, thereby improving the protocol selection capability and communication quality of the communication system in complex dynamic environments.

[0090] Secondly, unlike the existing protocol switching strategies based on empirical rules or black box models, the present invention first constructs a multi-mode communication protocol structure diagram that integrates communication protocol switching behavior, mechanism structure similarity and resource competition relationship, fully expresses the structural dependency relationship between different communication protocols, and embeds it into the structured LSTM model as control information in the form of an adjacency matrix. The introduction of this structure diagram enables the model to dynamically capture and utilize the intrinsic correlation between protocols while processing timing features, achieving more accurate protocol-level modeling and context-aware hidden state calculation, thereby improving the accuracy of performance prediction.

[0091] In addition, the present invention further combines the structure-enhanced hidden state vector with the performance prediction results to construct a protocol causal subgraph. By introducing the time-decay aggregation and indicator embedding mechanism, it clearly reveals the causal impact path of key factors in the communication environment on communication performance. This causal modeling mechanism not only enhances the interpretability of the model, but also makes the protocol selection process traceable and quantifiable. It can dynamically identify the main environmental variables and their action mechanisms that affect throughput, latency and energy consumption, thereby supporting more targeted protocol evaluation and optimization.

[0092] Finally, during the policy decision-making phase, the present invention constructs a multi-objective comprehensive scoring function based on a combination of weighted performance metrics and causal weights. This function considers the multi-dimensional objectives of maximizing throughput, minimizing latency, and controlling energy consumption. A statistical robustness indicator is also introduced to mitigate the influence of short-term fluctuations on the decision-making results. This integrated scoring mechanism ensures that the selected communication protocol is more consistent with the actual state of the current network environment and business needs, achieving comprehensive optimization of communication protocol switching decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0094] Figure 1 This is a flowchart of a multi-mode communication protocol optimization method based on deep learning proposed by the present invention;

[0095] Figure 2 This is a schematic diagram of the structure of a structured LSTM model for a multi-mode communication protocol optimization method based on deep learning proposed in the present invention;

[0096] Figure 3 This is a schematic diagram of the causal weight matrix construction and protocol causal subgraph generation of a deep learning-based multi-mode communication protocol optimization method proposed in the present invention. DETAILED DESCRIPTION

[0097] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0098] refer to Figure 1-3 , a multi-mode communication protocol optimization method based on deep learning, comprising the following steps:

[0099] S1. Collecting operation data of the communication terminal in a multi-mode communication environment, and pre-processing the operation data to generate a time series feature vector;

[0100] S2. Construct a multi-mode communication protocol structure diagram, model the structural dependency between multi-mode communication protocols, and generate an adjacency matrix;

[0101] S3. Build a structured LSTM model. Based on the time series feature vector and adjacency matrix, perform information transfer of the structured LSTM model, model the time series characteristics of the communication environment and the structural dependency information between protocols, and generate a structure-enhanced hidden state vector.

[0102] S4, inputting the time series feature vector and the adjacency matrix into the structured LSTM model, and outputting the performance prediction vector of the communication protocol based on the structure-enhanced hidden state vector;

[0103] S5. Generate a causal weight matrix based on the performance prediction vector and the hidden state vector of the structured LSTM model, and construct a protocol causal subgraph between the time series feature vector and the performance prediction vector;

[0104] S6. Based on the causal weight matrix and the protocol causal subgraph, a multi-objective comprehensive evaluation of the communication protocol is performed, and the optimal communication protocol configuration scheme is determined based on the set comprehensive scoring function.

[0105] The present invention proposes a multi-mode communication protocol optimization method based on deep learning, which can integrate the environmental operation data of the communication terminal with the protocol structure information for modeling, support accurate prediction of the performance of multiple communication protocols and intelligent optimization selection of protocol configuration schemes. By introducing a structured LSTM model and a causal graph modeling mechanism, it realizes the joint modeling and interpretation of the dynamic characteristics of the communication environment, the protocol structure dependency and the causal effect between performance indicators, effectively improving the intelligence, rationality and environmental adaptability of the protocol switching strategy, and enhancing the stability and resource scheduling efficiency of the communication system in a multi-protocol coexistence environment.

[0106] In this embodiment, the operating data includes channel quality parameters, network load status, and terminal coordinates. The channel quality parameters include reference signal power, reference signal reception quality, signal-to-interference-and-noise ratio, channel quality indicator, and block error rate. The network load status includes the number of user connections, physical resource block utilization, and scheduling queue length.

[0107] The present invention clearly proposes the composition of operating data, including channel quality parameters, network load status and terminal location coordinates, and refines the specific composition of channel quality parameters and network load indicators, so that the input data has rich contextual environment semantics in the timing feature modeling stage. Compared with the traditional optimization method that relies on a single physical layer parameter, the data collected by the present invention comprehensively covers the communication channel status, network congestion and terminal spatial distribution, which can improve the model's ability to characterize the actual communication environment, thereby enhancing the accuracy and robustness of protocol performance prediction and selection results.

[0108] In this embodiment, the preprocessing includes time alignment, feature normalization, outlier removal and missing value filling.

[0109] The present invention ensures the integrity, stability and standardized expression of the input time series feature vector by performing preprocessing operations such as time alignment, feature normalization, outlier removal and missing value filling after running data collection, providing a high-quality, structurally consistent data foundation for the deep learning model. This preprocessing process can effectively eliminate sampling offsets, scale inconsistencies and noise effects between multi-source data, improve the accuracy and generalization ability of communication environment modeling, and ensure that the downstream structured LSTM model can efficiently learn protocol performance patterns under different network conditions.

[0110] In this embodiment, S2 specifically includes:

[0111] S21. Setting a communication protocol set, where the communication protocol set includes LTE, 5G, Wi-Fi, and Bluetooth;

[0112] S22. Construct a protocol switching matrix based on the historical communication protocol switching records of the communication terminal:

[0113]

[0114] Among them, A i,j Represents the elements in the protocol switching matrix, describing the communication protocol p i With p j Whether there is a switching relationship between them;

[0115] S23. Construct the mechanism characteristics of each communication protocol into a mechanism feature vector, and combine them to form a protocol mechanism similarity matrix. Each element in the protocol mechanism similarity matrix represents the structural similarity of the communication protocol. The mechanism feature vector is composed of structural mechanism feature codes. The mechanism features include frame structure, scheduling mode, retransmission mechanism, control signaling structure, operating frequency band, and access mode. Each mechanism feature is one-hot encoded.

[0116] S24. Identify resource conflict relationships between protocols and build a resource competition matrix:

[0117]

[0118] Among them, R i,j Represents the elements in the resource competition matrix, describing the communication protocol p i With p j The resource hierarchical mutually exclusive relationship between them;

[0119] S25. Integrate the protocol switching matrix, the protocol mechanism similarity matrix, and the resource competition matrix, and define the multimode communication protocol structure graph as G = (V, E), where V represents the node set of the multimode communication protocol structure graph, and E represents the edge set of the multimode communication protocol structure graph:

[0120] E={(p i,p j )∣A i,j =1∨S i,j >θ s ∨R i,j =1};

[0121] Where E represents the edge set of the multi-mode communication protocol structure graph, describing the establishment of a connection when there is a switching behavior between communication protocols, the structural similarity is greater than the threshold, or the resource competition condition is met, θ s represents the similarity threshold, ∨ represents the “and” operation, p i represents the i-th communication protocol, p j represents the jth communication protocol;

[0122] S26. Represent the multi-mode communication protocol structure diagram as an adjacency matrix, where the adjacency matrix indicates whether there is an edge connection relationship between nodes, and the edge connection relationship includes switching behavior, structural similarity, and resource contention.

[0123] The present invention constructs a multi-mode communication protocol structure diagram by comprehensively considering the switching behavior, mechanism structure similarity and resource competition relationship of the communication protocol, and converts it into an adjacency matrix and embeds it into a neural network, thereby improving the deep model's structural understanding of the dependency relationship between protocols. This method can explicitly model the switching feasibility, mechanism compatibility and resource mutual exclusivity between communication protocols, thereby achieving accurate modeling of the protocol collaboration relationship, providing structural control information for the structured LSTM, and effectively improving the contextual modeling capability of protocol performance prediction and the overall system scheduling optimization level.

[0124] In this embodiment, S3 specifically includes:

[0125] S31, setting the input data to a time series feature vector;

[0126] S32. Construct a structured LSTM model, wherein the structured LSTM model includes an input layer, a structured LSTM hidden layer, and an output layer. The structured LSTM hidden layer, based on the traditional LSTM hidden layer, introduces a multi-mode communication protocol structure diagram as an explicit control signal to structurally constrain the information transmission path and strength between each hidden unit;

[0127] S33. At each time step, calculate the input gate, forget gate, and output gate vectors of the structured LSTM model:

[0128] i t =σ(W i x t +U i h t-1 +b i );

[0129] ft =σ(W f x t +U f h t-1 +b f );

[0130] o t =σ(W o x t +U o h t-1 +b o );

[0131] Among them, i t represents the input gate vector at time step t, f t Represents the forget gate vector at time step t, o t represents the output gate vector at time step t, W i ,W f ,W o represents the input weight matrix, U i ,U f ,U o represents the hidden state transfer weight matrix, b i ,b f ,b o represents the bias vector, σ(·) represents the Sigmoid activation function, h t-1 represents the hidden state vector at time step t-1, x t represents the time series feature vector at time step t;

[0132] S34. During the information transmission process, the candidate memory unit state, memory unit state and hidden state vector are calculated:

[0133]

[0134] h t =o t ⊙tanh(c t );

[0135] in, represents the candidate memory cell state vector at time step t, c t represents the memory cell state vector at time step t, h t represents the hidden state vector at time step t, W c Represents the input weight matrix of the candidate state, U c Represents the hidden state weight matrix of the candidate state, b c Represents the bias vector of the candidate state, tanh(·) represents the hyperbolic tangent activation function, and ⊙ represents the Hadamard product, i.e., element-wise multiplication;

[0136] S35. Use the adjacency matrix to introduce a structural dependency control mechanism during the propagation process at each time step, and introduce the aggregated influence from adjacent nodes to the hidden state of each protocol:

[0137]

[0138] in, Indicates the communication protocol p i The structure-enhanced hidden state vector at time step t+1, Indicates the communication protocol p i The structure-enhanced hidden state vector at time step t, G represents the adjacency matrix, and G(i, j) represents the protocol p i With p j The structural connection relationship of M j→i (·) represents the protocol p j For protocol p i The mapping operation is performed using a multi-layer perceptron. Indicates the communication protocol p j The structure-enhanced hidden state vector at time step t, N represents the total number of communication protocols.

[0139] Based on the LSTM model structure, the present invention introduces a multimodal communication protocol structure diagram as an explicit structural prior. Through structured LSTM units, the path and strength of the hidden state transmission process are structurally constrained, so that the model not only learns the dynamic changes in the time series, but also can simultaneously perceive the structural associations between communication protocols. This method improves the ability of neural networks in modeling the evolution of communication environments and the dependence of protocol states, avoids the problem of traditional LSTM ignoring protocol collaboration information, and enhances the expressive power of hidden states and the accuracy of the model in predicting communication performance.

[0140] In this embodiment, the S4 specifically includes:

[0141] S41. Protocol performance prediction based on the structure-enhanced hidden state vector output by the structured LSTM model;

[0142] S42. Construct a corresponding performance prediction header for each communication protocol, process the structure-enhanced hidden state vector using a protocol-specific nonlinear mapping function, and output a performance prediction vector for the communication protocol. The performance prediction vector includes throughput, communication delay, and energy consumption:

[0143] y k =W k2 ReLU(W k1 ·h T +b k1 )+b k2 ;

[0144] Among them, yk Denotes the performance prediction vector of the communication protocol, W k1 ,W k2 Represents the weight matrix of the two-layer fully connected structure, b k1 , b k2 represents the bias vector, ReLU represents the nonlinear activation function, and h T represents the structure-enhanced hidden state vector;

[0145] S43. Execute performance prediction on all protocols in the communication protocol set to construct a protocol performance prediction set:

[0146]

[0147] in, represents the set of protocol performance predictions, y k Represents the performance prediction vector corresponding to the k-th communication protocol.

[0148] The present invention designs an independent performance prediction head for each communication protocol and maps the structure-enhanced hidden state vector into three key performance indicators: throughput, latency and energy consumption, thereby achieving fine-grained performance prediction at the communication protocol level. This design improves the model's ability to model protocol behavior differences, avoids the representation ambiguity caused by parameter sharing, processes protocol-specific states through dedicated mapping functions, effectively supports parallel prediction and comparison of various protocols in a multi-mode communication environment, improves the accuracy and reliability of the overall performance evaluation, and provides accurate input for protocol optimization.

[0149] In this embodiment, the S5 specifically includes:

[0150] S51, obtaining a time series feature vector of the communication environment, a hidden state vector of the structured LSTM model, and a performance prediction vector of the communication protocol;

[0151] S52. Expand the performance prediction vectors of all communication protocols into a performance indicator set, where each element in the performance indicator set represents a performance indicator, including throughput, communication delay, and energy consumption. Extract the time series aggregation vector of each time series feature vector by time-decay weighted averaging:

[0152]

[0153] α t =exp(-λ·(Tt));

[0154] Among them, q i represents the time series aggregation vector of the time series feature vector, α t represents the decay weight of time step t, λ represents the time decay factor, represents the i-th time series feature vector in the t-th time step, Z represents the normalization coefficient of the weights of all time steps, T represents the total number of time steps, and exp represents the natural exponential function;

[0155] S53. Construct a performance indicator vector for each performance indicator. Take the hidden state vector of the structured LSTM model as input and obtain it by superimposing a nonlinear activation function on two layers of linear transformation:

[0156] k j =φ(W j2 ·(W j1 ·h t +b j1 )+b j2 );

[0157] Among them, k j represents the performance indicator vector, W j1 Represents the weight matrix of the first layer transformation, W j2 represents the weight matrix of the second layer transformation, b j1 represents the first layer bias vector, b j2 represents the second layer bias vector, φ(·) represents the activation function, using the ReLU function, h t represents the hidden state vector at time step t;

[0158] S54. Use the time series aggregation vector and performance indicator vector to construct a causal weight matrix:

[0159]

[0160] Among them, C represents the causal weight matrix, C i,j Represents the elements of the causal weight matrix, describing the effect of the time series feature vector on the performance indicator y j The causal influence weight, k j Represents the performance index y j The performance indicator vector, k l Represents the performance index y l performance indicator vector, exp(·) represents the exponential function, and B represents the transposition operation;

[0161] S55. Set the causal influence threshold, filter the index pairs whose causal influence weight is greater than the causal influence threshold, and construct the causal edge set:

[0162] ε k ={(x i ,y j )∣C i,j >θ c};

[0163] Among them, ε k represents the set of causal edges, θc represents the causal influence threshold, y j represents the performance index, x i represents the time series feature vector, (x i ,y j ) represents the causal edge between the time series feature vector and the performance indicator;

[0164] S56. Construct causal graph G k =(V k ,ε k )), where V k Represents the node set of the causal graph, including all time series feature vectors and performance indicators, ε k Represent the set of causal edges and construct the protocol causal subgraph, including;

[0165] For each communication protocol, determine a set of performance indicators, and use the performance indicators of this communication protocol as the target node;

[0166] Traverse all causal edges in the causal graph and filter the edges whose target nodes belong to the performance indicator set. When the causal weight of the causal edge is greater than or equal to the preset threshold, the edge is determined to be a valid influence path. The causal edges that meet the screening conditions and their source nodes and target nodes are combined into the edge set and node set of the initial protocol causal subgraph to construct the initial protocol causal subgraph;

[0167] The maximum backtracking depth is set. Starting from each timing feature vector node in the initial protocol causal subgraph, a breadth-first search algorithm is used to recursively trace back its upstream causal variables, identify indirect causal paths that do not exceed the maximum backtracking depth, and add the intermediate nodes and edges involved in the indirect causal paths to the initial protocol causal subgraph to obtain a complete protocol causal subgraph. The protocol causal subgraph describes the key causal dependency between the performance indicators and timing characteristics of the current communication protocol in the communication environment.

[0168] The present invention constructs a causal weight matrix by using the hidden state vector and performance prediction results of the structured LSTM model, and generates a protocol-level causal subgraph based on the attenuated aggregation method and multi-layer perception mechanism. It introduces a causal interpretability mechanism in communication protocol optimization. This mechanism not only clearly points out which characteristic variables in the communication environment have a significant impact on protocol performance, but also supports traceability analysis of protocol selection results. This method improves the model's interpretability, enhances decision-making transparency, and helps to deploy safe and controllable protocol switching strategies in actual communication systems.

[0169] In this embodiment, S6 specifically includes:

[0170] S61. Calculate the weight factors of all causal edges in the causal graph based on the causal graph and the causal weight matrix:

[0171]

[0172] Among them, ξ represents the weight factor of the causal graph, ε k represents the set of causal edges, (x i ,y j ) represents the causal edge between the time series feature vector and the performance indicator;

[0173] S62. For each communication protocol, obtain the protocol causal subgraph of the performance indicator, extract the relevant causal edge set from the protocol causal subgraph, and calculate the indicator causal weight of the communication protocol:

[0174]

[0175] Among them, w T ,w D ,w E They represent the causal weights of the communication protocol indicators in the three dimensions of throughput, communication delay and energy consumption, ε k represents the set of causal edges, T k ,D k ,E k represent throughput, communication delay and energy consumption respectively, Respectively represent the performance index T k , D k , E k The causal influence weight, (x i ,T k ) represents the causal edge between the timing feature vector and throughput, (x i ,D k ) represents the causal edge between the timing feature vector and the communication delay, (x i ,E k ) represents the causal edge between the time series feature vector and energy consumption, ε r Represents the set of relevant causal edges in the protocol causal subgraph;

[0176] S63. Adjust the performance indicators based on the causal weights of the indicators to construct weighted performance indicators of the communication protocol:

[0177]

[0178] in, represent the weighted throughput, weighted delay and weighted energy consumption after causal enhancement respectively;

[0179] S64. Calculate statistics based on the weighted performance indicators of all communication protocols and construct a comprehensive scoring function for the communication protocols:

[0180]

[0181] Among them, S k represents the comprehensive scoring function of the communication protocol, α, β, γ represent the weighted coefficients of the throughput, delay and energy consumption scoring dimensions respectively, ∈ represents a positive number to prevent the introduction of zero division errors, μ T ,σ D ,σ E They represent the mean of weighted throughput, the standard deviation of weighted delay, and the standard deviation of weighted energy consumption respectively;

[0182] S65. Select the communication protocol index with the highest score:

[0183]

[0184] Among them, k * represents the index of the communication protocol with the highest score in the comprehensive scoring function, S k represents the comprehensive score of the kth communication protocol, and argmax represents the variable value that maximizes the function value;

[0185] S66. Generate optimal communication protocol configuration solution in represents the optimal communication protocol, Indicates the kth * The performance prediction vector corresponding to the communication protocol, represents k * The protocol causal subgraph of a communication protocol.

[0186] Based on causal graph analysis, this paper constructs a multi-objective weighted scoring function that integrates the causal weights and statistical characteristics of protocol performance indicators to achieve quantifiable and stable evaluation of protocol comprehensive scoring and selection. This method comprehensively considers the optimization objectives of throughput, latency, and energy consumption, and introduces statistical terms such as standard deviation to regulate prediction volatility, making the final protocol selection more robust, accurate, and adaptable to actual scenario requirements. This strategy significantly improves the quality of protocol optimization results and the reliability of engineering deployment.

[0187] In this embodiment, the statistics include the mean of weighted throughput after modulation of all communication protocols, the standard deviation of weighted delay, and the standard deviation of weighted energy consumption.

[0188] This invention introduces the weighted mean throughput and the standard deviation of weighted latency and energy consumption as statistical indicators during the communication protocol scoring phase. This allows the comprehensive scoring function to take into account both performance expectations and risk fluctuations. This design effectively prevents protocol selection errors caused by individual abnormal predictions or environmental fluctuations, thereby improving the stability and security of the communication protocol switching solution. This mechanism enhances the system's robustness to dynamic network environments and is a key component in ensuring the intelligent and practical nature of protocol selection.

[0189] Example 1:

[0190] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a multi-mode wireless communication system in the core area of ​​a city. The average daily number of active communication terminals in this area exceeds 800,000, covering four mainstream communication protocols: LTE, 5G, Wi-Fi and Bluetooth. The scene types include business districts, high-speed transportation hubs, underground spaces, office-dense areas and residential areas. Due to the large differences in user behavior patterns, strong network load fluctuations, and large spans in device movement speeds, the existing methods for selecting communication protocols based on fixed rules or empirical thresholds frequently experience problems such as unstable performance, protocol switching errors, and excessive energy consumption.

[0191] The present invention was deployed in the communication dispatch system of a mainstream operator in the region, and the communication protocol optimization method was implemented on 100 multimode terminal nodes in the core area with a continuous 72-hour window. First, the terminal device collects real-time data from its own operating environment, including channel quality parameters (RSRP, RSRQ, SINR, CQI), network load status (base station load rate, queue length, number of users), terminal location information (GPS coordinates, access point ID), device speed (calculated by location change), etc., and completes preprocessing processes such as gap filling, normalization and time alignment to generate a continuous time series feature vector.

[0192] Subsequently, the system constructs a multi-mode communication protocol structure diagram based on the communication protocol switching records in the past 7 days, protocol mechanism characteristics (such as frame structure, access method, scheduling method) and the mutually exclusive relationship between spectrum resources, and converts it into an adjacency matrix and embeds it into a structured LSTM model to build a structure-enhanced communication protocol prediction model. At each time step, the structured LSTM model predicts the expected throughput, communication delay and energy consumption values ​​of each protocol in the current environment, and constructs a causal weight matrix between timing characteristics and performance indicators.

[0193] To enhance interpretability and selection stability, the system uses causal weights to screen effective causal paths, constructs a protocol causal subgraph for each communication protocol, and calculates weighted performance indicators of the communication protocol based on the importance of the prediction results and causal paths. Finally, all candidate protocols are evaluated and ranked through a multi-objective comprehensive scoring function. This comprehensive scoring function takes into account throughput maximization, latency minimization, and energy consumption constraints, and introduces a standard deviation control term to reduce switching instability caused by short-term prediction fluctuations.

[0194] During the experiment, the research team set up a control group and compared the traditional rule-based protocol switching algorithm with the method proposed in this invention. The test platform operated in the same network resource scheduling environment, and all terminal device communication behavior records were uploaded to the server through log synchronization for comparison and analysis.

[0195] Table 1 Comparison of core performance of multi-mode communication protocol optimization

[0196]

[0197] In terms of protocol switching frequency, the method of the present invention has an average switching frequency of 4.88 times per hour, which is about 27.3% less than the 6.71 times of the traditional method. This shows that the present invention can effectively reduce unnecessary protocol switching, reduce system switching overhead, and improve communication stability. Further analysis of the protocol switching misjudgment rate shows that it has dropped from 11.4% to 6.7%, a reduction of more than 40%, indicating that the present invention has a higher judgment accuracy in the protocol selection process and can reduce the frequent reselection or switching failure problems caused by misjudgment.

[0198] In terms of communication performance, the average throughput of the present invention is improved to 69.7Mbps, an increase of 19.6% compared with the 58.3Mbps of the traditional method, indicating that the present invention is more inclined to select protocols with stronger data transmission capabilities during the protocol selection process. At the same time, the average communication delay is reduced from 92.4ms to 74.1ms, a reduction of about 19.8%. Combined with the reduction of unit bit energy consumption from 0.0159mJ / bit to 0.0124mJ / bit, it can be seen that the present invention not only improves communication efficiency, but also effectively reduces energy consumption overhead, and is particularly friendly to battery-powered terminals.

[0199] In terms of network stability, the drop rate in congested areas dropped from 3.8% to 2.7%, indicating that the present invention has stronger congestion adaptability and anti-burst interference capabilities, which is conducive to improving the overall availability of the system. In terms of user experience, the subjective satisfaction score increased from 6.8 points to 8.2 points, indicating that end users can clearly perceive the improvement in communication experience. Finally, the protocol switching success rate increased from 87.3% to 94.9%, which means that the present invention is more reliable in the execution stage of protocol switching and reduces the interference of switching failures on communication continuity.

[0200] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A multi-mode communication protocol optimization method based on deep learning, characterized in that: The steps include: S1. Collecting operation data of the communication terminal in a multi-mode communication environment, and pre-processing the operation data to generate a time series feature vector; S2. Construct a multi-mode communication protocol structure diagram, model the structural dependency between multi-mode communication protocols, and generate an adjacency matrix; S3. Build a structured LSTM model. Based on the time series feature vector and adjacency matrix, perform information transfer of the structured LSTM model, model the time series characteristics of the communication environment and the structural dependency information between protocols, and generate a structure-enhanced hidden state vector. S4, inputting the time series feature vector and the adjacency matrix into the structured LSTM model, and outputting the performance prediction vector of the communication protocol based on the structure-enhanced hidden state vector; S5. Generate a causal weight matrix based on the performance prediction vector and the hidden state vector of the structured LSTM model, and construct a protocol causal subgraph between the time series feature vector and the performance prediction vector; S6. Based on the causal weight matrix and the protocol causal subgraph, a multi-objective comprehensive evaluation of the communication protocol is performed, and the optimal communication protocol configuration scheme is determined based on the set comprehensive scoring function.

2. The multi-mode communication protocol optimization method based on deep learning according to claim 1, characterized in that: The operating data includes channel quality parameters, network load status, and terminal coordinates. The channel quality parameters include reference signal power, reference signal reception quality, signal-to-interference-and-noise ratio, channel quality indicator, and block error rate. The network load status includes the number of user connections, physical resource block utilization, and scheduling queue length.

3. The multi-mode communication protocol optimization method based on deep learning according to claim 1, characterized in that: The preprocessing includes time alignment, feature normalization, outlier removal and missing value filling.

4. The multi-mode communication protocol optimization method based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21. Setting a communication protocol set, where the communication protocol set includes LTE, 5G, Wi-Fi, and Bluetooth; S22. Construct a protocol switching matrix based on the historical communication protocol switching records of the communication terminal; S23, constructing the mechanism characteristics of each communication protocol into a mechanism feature vector, and combining them to form a protocol mechanism similarity matrix; S24, identifying resource conflict relationships between protocols and constructing a resource competition matrix; S25. Integrate the protocol switching matrix, the protocol mechanism similarity matrix, and the resource competition matrix to define a multi-mode communication protocol structure diagram; S26. Represent the multi-mode communication protocol structure diagram as an adjacency matrix, where the adjacency matrix indicates whether there is an edge connection relationship between nodes, and the edge connection relationship includes switching behavior, structural similarity, and resource contention.

5. The multi-mode communication protocol optimization method based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, setting the input data to a time series feature vector; S32. Construct a structured LSTM model, wherein the structured LSTM model includes an input layer, a structured LSTM hidden layer, and an output layer. The structured LSTM hidden layer, based on the traditional LSTM hidden layer, introduces a multi-mode communication protocol structure diagram as an explicit control signal to structurally constrain the information transmission path and strength between each hidden unit; S33. At each time step, calculate the input gate, forget gate, and output gate vectors of the structured LSTM model; S34. During the information transmission process, calculating the candidate memory unit state, the memory unit state, and the hidden state vector; S35. Use the adjacency matrix to introduce a structural dependency control mechanism during the propagation process at each time step, and introduce the aggregated influence from adjacent nodes to the hidden state of each protocol: in, Indicates the communication protocol p i The structure-enhanced hidden state vector at time step t+1, Indicates the communication protocol p i The structure-enhanced hidden state vector at time step t, G represents the adjacency matrix, and G(i, j) represents the protocol p i With p j The structural connection relationship of M j→i (·) represents the protocol p j For protocol p i The mapping operation is performed using a multi-layer perceptron. Indicates the communication protocol p j The structure-enhanced hidden state vector at time step t, N represents the total number of communication protocols.

6. The multi-mode communication protocol optimization method based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. Protocol performance prediction based on the structure-enhanced hidden state vector output by the structured LSTM model; S42. Construct a corresponding performance prediction header for each communication protocol, process the structure-enhanced hidden state vector using a protocol-specific nonlinear mapping function, and output a performance prediction vector for the communication protocol, wherein the performance prediction vector includes throughput, communication delay, and energy consumption; S43. Execute performance prediction on all protocols in the communication protocol set to construct a protocol performance prediction set.

7. The multi-mode communication protocol optimization method based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51, obtaining a time series feature vector of the communication environment, a hidden state vector of the structured LSTM model, and a performance prediction vector of the communication protocol; S52, expanding the performance prediction vectors of all communication protocols into a set of performance indicators, and extracting the time series aggregation vector of each time series feature vector by time decay weighted averaging; S53. Construct a performance indicator vector for each performance indicator, using the hidden state vector of the structured LSTM model as input, and obtain it by superimposing a nonlinear activation function on two layers of linear transformation; S54. Use the time series aggregation vector and performance indicator vector to construct a causal weight matrix: Among them, C represents the causal weight matrix, C i,j Represents the elements of the causal weight matrix, describing the effect of the time series feature vector on the performance indicator y j The causal influence weight, k j Represents the performance index y j The performance indicator vector, k l Represents the performance index y l performance indicator vector, exp(·) represents the exponential function, and B represents the transposition operation; S55. Set a causal influence threshold, select index pairs whose causal influence weight is greater than the causal influence threshold, and construct a causal edge set; S56. Construct causal graph G k =(V k ,ε k )), where V k Represents the node set of the causal graph, including all time series feature vectors and performance indicators, ε k Represent the set of causal edges and construct the protocol causal subgraph, including; For each communication protocol, determine a set of performance indicators, and use the performance indicators of this communication protocol as the target node; Traverse all causal edges in the causal graph and filter the edges whose target nodes belong to the performance indicator set. When the causal weight of the causal edge is greater than or equal to the preset threshold, the edge is determined to be a valid influence path. The causal edges that meet the screening conditions and their source nodes and target nodes are combined into the edge set and node set of the initial protocol causal subgraph to construct the initial protocol causal subgraph; The maximum backtracking depth is set. Starting from each timing feature vector node in the initial protocol causal subgraph, a breadth-first search algorithm is used to recursively trace back its upstream causal variables, identify indirect causal paths that do not exceed the maximum backtracking depth, and add the intermediate nodes and edges involved in the indirect causal paths to the initial protocol causal subgraph to obtain a complete protocol causal subgraph. The protocol causal subgraph describes the key causal dependency between the performance indicators and timing characteristics of the current communication protocol in the communication environment.

8. The multi-mode communication protocol optimization method based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. Calculate weight factors of all causal edges in the causal graph according to the causal graph and the causal weight matrix; S62. For each communication protocol, obtain a protocol causal subgraph of the performance indicator, extract a set of relevant causal edges from the protocol causal subgraph, and calculate the indicator causal weight of the communication protocol; S63. Adjust the performance indicators based on the indicator causal weights to construct a weighted performance indicator for the communication protocol; S64. Calculate statistics based on the weighted performance indicators of all communication protocols and construct a comprehensive scoring function for the communication protocols: Among them, S k represents the comprehensive scoring function of the communication protocol, α, β, γ represent the weighted coefficients of the throughput, delay and energy consumption scoring dimensions respectively, ∈ represents a positive number to prevent the introduction of zero division errors, μ T ,σ D ,σ E They represent the mean of weighted throughput, the standard deviation of weighted delay, and the standard deviation of weighted energy consumption respectively; S65. Select the communication protocol index with the highest score; S66. Generate an optimal communication protocol configuration solution.

9. The multi-mode communication protocol optimization method based on deep learning according to claim 8, characterized in that: The statistics include the mean of weighted throughput after modulation of all communication protocols, the standard deviation of weighted delay and the standard deviation of weighted energy consumption.

Images