Communication protocol self-evolution system based on pre-trained large model

Abstract

This invention discloses a communication protocol self-evolution system based on a pre-trained large-scale model. The system, belonging to the field of communication technology, includes a model building module and a risk assessment module. The model building module extracts historical iterative drift and non-iterative drift protocol evolution cases, constructing a protocol evolution influence chain graph and a pre-trained large-scale model respectively. The risk assessment module inputs the current first communication requirement into the pre-trained large-scale model, generates the protocol change content to be evaluated, substitutes it into the protocol evolution influence chain graph to match similar starting nodes, calculates risk scores, and filters out qualified protocols to be deployed. This invention achieves risk prediction and accurate optimization of protocol self-evolution by combining model generation and graph evaluation, improving protocol stability and adaptability, and is applicable to various communication scenarios.

Description

Technical Field

[0001] This invention relates to the field of communication technology, specifically to a communication protocol self-evolution system based on a pre-trained large model. Background Technology

[0002] As the core principle of data interaction between devices, the self-evolution capability of communication protocols directly affects the operating efficiency of communication systems. With the diversification of communication scenarios, the self-evolution technology of traditional communication protocols has gradually revealed the following shortcomings: First, the self-evolution of existing communication protocols largely relies on a pre-set rule base, lacking a systematic reuse of historical drift-free and drift cases, which makes it easy for the generated initial protocol changes to deviate from actual communication needs. Secondly, existing technologies have not established an effective risk prediction mechanism, and cannot identify the long-term cumulative chain reaction caused by the first protocol change in advance. This causes the problem to continue to amplify with iteration, eventually leading to iteration drift phenomena such as increased protocol redundancy. Third, the existing system lacks dynamic optimization capabilities and cannot continuously improve its self-evolution logic by combining new evolution cases. Its adaptability and flexibility are insufficient, making it difficult to meet the needs of the rapid development of communication technology. Therefore, there is an urgent need for a communication protocol self-evolution system based on a pre-trained large model. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a communication protocol self-evolution system based on a pre-trained large model, which solves the problems of easy iterative drift and insufficient risk prediction in existing communication protocol self-evolution.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a communication protocol self-evolution system based on a pre-trained large model, comprising: The model building module extracts all historical protocol evolution cases with and without iterative drift, and constructs a protocol evolution influence chain graph based on the iterative drift protocol evolution cases, and constructs a pre-trained large model based on the protocol evolution cases without iterative drift. The risk assessment module inputs the current first communication request into the pre-trained large model to obtain the corresponding protocol change content to be evaluated. It then substitutes this content into the protocol evolution impact chain graph, matches it with similar starting nodes, and calculates the risk score of the protocol change content to be evaluated based on the similar initial nodes. If the risk score exceeds the preset threshold, the protocol is directly discarded; otherwise, it is determined to be a qualified protocol to be deployed.

[0005] As a further aspect of the present invention, the specific operation of constructing a protocol evolution influence chain graph based on iterative drift protocol evolution cases is as follows: The complete link information of historical iteration drift cases is statistically analyzed, including the protocol change content corresponding to the first communication request, the direct derivative problems that occur in the short term after the change, the chain problems caused by long-term accumulation, and the results of iteration drift. Extract the protocol change content corresponding to the first communication requirement as the starting node and mark the specific parameters; extract the direct derivative problems that appear in the short term after the change as the first-level child nodes and mark the occurrence time; extract the chain problems caused by long-term accumulation as the second-level child nodes and mark the iterative correlation; extract the iterative drift result as the termination node; finally form a single-case node chain of starting node → first-level child node → second-level child node → termination node; Collect all the disassembled single-case node chains, count the co-occurrence frequency between different nodes, and if the co-occurrence frequency is ≥2 times, establish an association edge between the two nodes and mark the co-occurrence frequency on the association edge to form a protocol evolution influence chain graph that includes nodes, association edges, and co-occurrence frequencies.

[0006] As a further aspect of the present invention, the direct derivative problems that occur in the short term after the change refer to problems affecting device functionality that occur within 7 calendar days after the deployment of the protocol change content corresponding to the first communication requirement and without subsequent self-evolution iterations; the chain problems that occur after long-term accumulation refer to problems affecting device functionality that occur more than 30 calendar days after the deployment of the protocol change content corresponding to the first communication requirement and after at least 2 subsequent self-evolution iterations.

[0007] As a further aspect of the present invention, when a new iterative drift case appears in the system, it is decomposed into a single-case node chain; if there are unrecorded nodes in the node chain, they are added to the node library, and the co-occurrence frequency of the new node and the existing nodes is counted; if there are recorded nodes in the node chain, the co-occurrence frequency of its associated edge is incremented by 1; invalid associated edges with a co-occurrence frequency ≤ 1 and no new case matching for more than 6 months are removed.

[0008] As a further aspect of the present invention, the specific operation of constructing a pre-trained large model based on protocol evolution cases without iterative drift is as follows: For each protocol evolution case without iterative drift, the first communication requirement and the corresponding protocol change are collected, and cases with missing parameters are removed to form a structured fine-tuning dataset. The structured fine-tuning dataset is transformed into an instruction-output training sample format, where the instruction is the first communication requirement and the output is the corresponding protocol change content. A general text generation model was chosen as the base model. The LoRA lightweight fine-tuning technique was used, with a low-rank matrix inserted only in the attention layer. The training hyperparameters for the LoRA lightweight fine-tuning task were: batch size of 8, learning rate of 1e-4, and training epochs of 5. Input a 10% untrained structured fine-tuning dataset. If the output protocol change content is consistent with the protocol change content corresponding to the instruction-output, the validation is successful. If the percentage of successful validations exceeds 90%, the basic model parameters and LoRA low-rank matrix parameters are packaged and saved to obtain the pre-trained large model.

[0009] As a further aspect of the present invention, the specific method for determining whether the output protocol change content is consistent with the protocol change content corresponding to the instruction-output is as follows: By breaking down the protocol changes, we can obtain the core operations and objectives of the protocol changes, resulting in a set of operations and a set of objectives. Find the intersection and union of the operation sets of the output protocol change content and the corresponding operation set of the instruction-output protocol change content, and use the ratio of the intersection and union as the operation similarity. Similarly, obtain the target similarity, and calculate the overall similarity = (operation similarity + target similarity) / 2. If the overall similarity is ≥80%, the protocol change content output by the model will be deployed and run in the test environment for 30 consecutive days to monitor for any systemic problems. If no systemic problems occur, it can be determined that the output protocol change content is consistent with the protocol change content corresponding to the instruction-output.

[0010] As a further aspect of the present invention, the specific operation for matching similar starting nodes is as follows: The actions and objects of the changes are extracted from the changes to be evaluated to obtain a set of real-time semantic features; the specific quantitative indicators involved in the changes are extracted from the changes to be evaluated to obtain a set of real-time parameter features. For each starting node, it is decomposed into a semantic feature set and a parameter feature set; a single feature in each set is used as the index key, and the index value is the ID of all starting nodes containing that feature; each index key is marked with a feature frequency level, that is, the frequency level is divided according to the frequency of the feature in all starting nodes, forming a complete feature inverted index table; Based on the generated real-time semantic feature set and real-time parameter feature set, the constructed feature inverted index table is retrieved, and all starting nodes containing any feature are extracted to obtain candidate starting nodes. The number of features that each candidate starting node has in common with the real-time semantic feature set and the real-time parameter feature set is counted and used as the number of co-occurring features. In co-occurrence features, the ratio of the number of high-frequency features to the number of co-occurrence features is calculated to obtain the proportion of high-frequency features; If the number of co-occurring features is greater than or equal to 60% of the total number of baseline features and the proportion of high-frequency features is greater than or equal to 50%, then the initial node is taken as the similarity starting node.

[0011] As a further aspect of the present invention, the frequency level is divided according to the frequency of occurrence of the feature in all starting nodes: if the frequency of occurrence is ≥ 30% of the total number of starting nodes, it is marked as a high-frequency feature; if the frequency of occurrence is between 10% and 30% of the total number of starting nodes, it is marked as a medium-frequency feature; if the frequency of occurrence is < 10% of the total number of starting nodes, it is marked as a low-frequency feature.

[0012] As a further aspect of the present invention, the specific operation for calculating the risk score of the changes to the agreement to be evaluated is as follows: Starting from a similar starting node, along the associated edges in the protocol evolution influence chain graph, extract all first-level child nodes and second-level child nodes directly connected to it, and obtain the co-occurrence frequency of each associated edge. The extracted content is then structured and organized to form a list of associated links from similar starting nodes to first-level child nodes to second-level nodes. Collect the first-level child nodes corresponding to similar initial nodes to construct a first-level problem set. Calculate the intersection of all first-level problem sets corresponding to similar initial nodes, denoted as intersection one. Collect the second-level child nodes corresponding to similar initial nodes to construct a second-level problem set. Calculate the intersection of all second-level problem sets corresponding to similar initial nodes, denoted as intersection two. Calculate the intersection of intersection one and intersection two, denoted as intersection three. Calculate the arithmetic mean of the risk scores of all similar starting nodes and use it as the risk score of the protocol change content to be evaluated.

[0013] As a further aspect of the present invention, the specific steps for calculating the risk score of similar starting nodes are as follows: Calculate the coverage of intersection 1 = number of problems contained in intersection 1 / total number of problems contained in the first-level child nodes corresponding to all similar initial nodes; calculate the coverage of intersection 2 = number of problems contained in intersection 2 / total number of problems contained in the second-level child nodes corresponding to all similar initial nodes; calculate the coverage of intersection 3 = number of problems contained in intersection 3 / sum of the total number of problems contained in the first-level child nodes corresponding to all similar initial nodes and the total number of problems contained in the second-level child nodes. The levels are classified according to the transmissibility of the problem's impact: Intersection 3 is level 3, with a level coefficient of 30; Intersection 2 is level 2, with a level coefficient of 20; Intersection 1 is level 1, with a level coefficient of 10. Calculate the risk score of similar initial nodes = (intersection 1 coverage × grade coefficient 3) + (intersection 2 coverage × grade coefficient 2) + (intersection 3 coverage × grade coefficient 1).

[0014] This invention provides a communication protocol self-evolution system based on a pre-trained large model, which has the following advantages compared with the prior art: (1) This invention constructs a pre-trained large model based on successful cases of no iterative drift, so that the generated first protocol change content fits the core logic of communication requirements, reduces the triggering conditions of iterative drift from the source, and ensures the initial reliability of protocol self-evolution; (2) The impact chain graph constructed by the present invention through iterative drift cases can identify in advance the short-term derivative problems and long-term cumulative chain reactions that may be caused by the change of the protocol to be evaluated. Combined with similar node matching and risk quantification calculation, risk prevention can be achieved to avoid the problem from amplifying with iteration and causing the protocol to fail. Attached Figure Description

[0015] Figure 1 This is the system principle block diagram of the present invention; Figure 2 A flowchart illustrating the steps involved in constructing a protocol evolution impact chain graph for this invention; Figure 3 This is a flowchart illustrating the steps involved in calculating the risk score of changes to the agreement to be evaluated in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0017] like Figure 1 This invention provides a communication protocol self-evolution system based on a pre-trained large model, comprising: The model building module extracts all historical protocol evolution cases with and without iterative drift, and constructs a protocol evolution influence chain graph based on the iterative drift protocol evolution cases, and constructs a pre-trained large model based on the protocol evolution cases without iterative drift. The protocol evolution case of iterative drift refers to an evolution case in which, after multiple consecutive evolutions, although each optimization achieves short-term goals such as reducing latency, long-term accumulation leads to systemic problems in the protocol, such as increased redundancy, decreased device compatibility, and deterioration of transmission stability, ultimately causing the protocol to fail and requiring manual reset. The protocol evolution case without iterative drift specifically refers to an evolution case in which the protocol has undergone multiple consecutive evolutions, and after achieving short-term goals in a single optimization, no systemic problems have occurred in long-term operation, the performance of the entire link remains stable, and no manual intervention or reset is required. The specific steps for building a pre-trained large model based on protocol evolution cases without iterative drift are as follows: For each protocol evolution case without iterative drift, the first communication requirement and the corresponding protocol change are collected, and cases with missing parameters are removed to form a structured fine-tuning dataset. For example, the first communication requirement is: reduce the latency of vital signs data to ≤80ms, and the corresponding protocol change is: adjust the transmission time slot of the ECG monitor from 10ms / frame to 8ms / frame. The generation logic of large models is to imitate the paradigm of training data. If incomplete parameters or negative cases are mixed in, it may lead to increased hesitation in model generation. Directly learning the optimal protocol change path is more in line with the need to generate the protocol corresponding to the first communication requirement. The structured fine-tuning dataset is transformed into a command-output training sample format, where the command is the first communication requirement and the output is the corresponding protocol change content. A general text generation model was selected as the base model. LoRA lightweight fine-tuning technique was used, with a low-rank matrix inserted only in the attention layer. The training hyperparameters for the LoRA lightweight fine-tuning task were: batch size of 8, learning rate of 1e-4, and training epochs of 5. In each epoch, the output performance was evaluated using 20% ​​of the structured fine-tuning dataset. The evaluation criteria were: whether the changed content contained specific parameters and whether it conformed to the successful case paradigm. Input a 10% untrained structured fine-tuning dataset. When the output protocol change content is consistent with the protocol change content corresponding to the instruction-output, the validation is successful. If the percentage of successful validations exceeds 90%, the basic model parameters and LoRA low-rank matrix parameters are packaged and saved to obtain the pre-trained large model. The specific method for determining whether the protocol change content output matches the protocol change content corresponding to the instruction output is as follows: By breaking down the two into their respective core operations for protocol changes, such as adding interfaces, adjusting frequencies, and retaining verification logic, and the targets of these operations, such as old infusion pumps and transmission time slots, we obtain the set of operations and the set of targets. Find the intersection and union of the operation sets of the two entities, and use the ratio of the intersection to the union as the operation similarity; find the intersection and union of the target sets of the two entities, and use the ratio of the intersection to the union as the target similarity; thus, the overall similarity is calculated as (operation similarity + target similarity) / 2. If the overall similarity is ≥80%, it means that the semantic similarity meets the standard. The protocol change content output by the model is then deployed and run in the test environment for 30 consecutive days to monitor for any systemic problems. If no systemic problems occur, it can be determined that the output protocol change content is consistent with the protocol change content corresponding to the instruction-output.

[0018] The risk assessment module inputs the current first communication request into the pre-trained large model to obtain the corresponding protocol change content to be assessed, and substitutes it into the protocol evolution impact chain graph to match similar starting nodes. The specific operation is as follows: The actions and objects of the changes are extracted from the changes to be evaluated to obtain a set of real-time semantic features; the specific quantitative indicators involved in the changes are extracted from the changes to be evaluated to obtain a set of real-time parameter features. The starting nodes of the protocol evolution impact chain graph are preprocessed: each starting node is decomposed into a semantic feature set and a parameter feature set; a single feature in each set is used as the index key, and the index value is the ID of all starting nodes containing that feature; each index key is marked with a feature frequency level, that is, the frequency level is divided according to the frequency of the feature in all starting nodes, forming a complete feature inverted index table. The frequency level is determined based on the frequency of occurrence of this feature in all starting nodes: if the frequency of occurrence is ≥ 30% of the total number of starting nodes, it is marked as a high-frequency feature; if the frequency of occurrence is between 10% and 30% of the total number of starting nodes, it is marked as a medium-frequency feature; if the frequency of occurrence is < 10% of the total number of starting nodes, it is marked as a low-frequency feature. For example, the starting node library contains 3 starting nodes. Starting node 1 is characterized by {new device-specific interface, IEC60601-2-27, 10ms / frame}, starting node 2 is characterized by {new device-specific interface, IEEE 802.11ax, 12ms / frame}, and starting node 3 is characterized by {adjust transmission time slot, 8ms / frame, JPEG2000}. After constructing the inverted index, the new device-specific interface corresponds to starting nodes 1 and 2, 8ms / frame corresponds to starting nodes 1 and 3, and IEC 60601-2-27 corresponds to starting node 1. Based on the generated real-time semantic feature set and real-time parameter feature set, the constructed feature inverted index table is retrieved, and all starting nodes containing any feature are extracted to obtain candidate starting nodes. The number of features that each candidate starting node has in common with the real-time semantic feature set and the real-time parameter feature set is counted and used as the number of co-occurring features. In co-occurrence features, the ratio of the number of high-frequency features to the number of co-occurrence features is calculated to obtain the proportion of high-frequency features; If the number of co-occurring features is greater than or equal to 60% of the total number of baseline features and the proportion of high-frequency features is greater than or equal to 50%, then the initial node is taken as the similarity starting node. Calculate the risk score of the changes to the protocol to be evaluated. If the risk score exceeds the preset threshold, it means that using the changes to the protocol to be evaluated will lead to long-term drift risks. The protocol should be discarded directly. At the same time, a new protocol to be evaluated needs to be generated again through the pre-trained large model. Conversely, if the risk score does not exceed the preset threshold, it means that the protocol is relatively accurate in judging based on historical drift data and is determined to be a qualified protocol to be deployed. The selected qualified protocols are deployed to the communication system, and their effects are monitored. If a new iteration drift problem occurs after the protocol is run, the single-case node chain of the new case is extracted, in the form of: starting node → first-level child node → second-level child node → ending node, and added to the protocol evolution impact chain graph to complete the dynamic update of the graph. Example 2

[0019] This embodiment further discloses a method for constructing a protocol evolution influence chain graph based on embodiment 1, such as... Figure 2 As shown, the specific content includes: The complete link information of historical iteration drift cases is statistically analyzed, including the protocol changes corresponding to the first communication request, such as adjustments to transmission parameters, modifications to frame structure, and additions to adaptation rules; direct derivative problems that occur in the short term after the changes, such as incompatibility of certain types of devices and increased local transmission latency; chain problems caused by long-term accumulation, such as increased redundancy and failure of multiple device access; and the results of iteration drift, such as complete protocol failure and the need for manual reset. The protocol change corresponding to the first communication requirement forms the basis of the entire self-evolution process. All subsequent iterations must be based on its core logic. If there are problems, it will lead to the accumulation of subsequent problems and iteration drift. The direct derivative problems that occur in the short term after the change refer to the problems that affect the device function that occur within 7 calendar days after the deployment of the protocol change content corresponding to the first communication requirement, and without subsequent self-evolution iteration. The chain reaction of problems caused by long-term accumulation refers to the problems that affect the device function after more than 30 calendar days and no less than 2 subsequent self-evolution iterations since the deployment of the protocol change content corresponding to the first communication demand. Because the short-term problem was not resolved, subsequent self-evolution added patchwork changes to compensate for the problem, which in turn triggered new problems, forming a chain reaction of initial problem → patchwork change → new problem → further patchwork. For example, after the deployment of the protocol change corresponding to the first communication requirement, occasional verification failures occurred in older blood pressure monitors within 3 days, which was classified as a direct derivative problem that appeared in the short term after the change. After 45 days, after two iterations: to compensate for the verification failures, iteration 1 added a retransmission mechanism, which reduced the power consumption of portable monitors from 8 hours to 5 hours (new problem 1); to reduce power consumption, iteration 2 adjusted the transmission rate, which caused the data packet loss rate of new CT imaging equipment to increase from 0.01% to 0.1% (new problem 2). At this point, both the increased power consumption and the image packet loss were long-term cumulative chain problems. Extract the protocol change content corresponding to the first communication requirement as the starting node and mark the specific parameters; extract the direct derivative problems that appear in the short term after the change as the first-level child nodes and mark the occurrence time; extract the chain problems caused by long-term accumulation as the second-level child nodes and mark the iterative relationship, such as the addition of a repeated transmission mechanism in iteration 1 → the power consumption of portable monitors is reduced to 5 hours; extract the iteration drift results as the termination node; finally form a single-case node chain of starting node → first-level child node → second-level child node → termination node. Collect all the disassembled single-case node chains, count the co-occurrence frequency between different nodes, and if the co-occurrence frequency is ≥2 times, establish an association edge between the two nodes and mark the co-occurrence frequency on the association edge to form a protocol evolution influence chain graph containing nodes, association edges, and co-occurrence frequencies. When a new iterative drift case appears in the system, it is broken down into a single-case node chain. If there are unrecorded nodes in the node chain, they are added to the node library, and the co-occurrence frequency of the new node and the existing nodes is counted. If there are recorded nodes in the node chain, the co-occurrence frequency of its associated edges is incremented by 1. Invalid associated edges with a co-occurrence frequency of ≤1 and no new case matching for more than 6 months are removed to ensure that the graph is concise and effective. Example 3

[0020] This embodiment, based on Embodiments 1 and 2, further discloses a method for calculating the risk score of the changed content of the agreement to be evaluated, such as... Figure 3 The specific process is as follows: Starting from a similar starting node, along the associated edges in the protocol evolution influence chain graph, extract all first-level child nodes and second-level child nodes directly connected to it, and obtain the co-occurrence frequency of each associated edge. The extracted content is then structured and organized to form a list of associated links from similar starting node to first-level child node (association strength) to second-level node (association strength). For example, the matched similar initial node is historical node 1, which is {new device dedicated interface, IEC 60601-2-27, 10ms / frame}. The associated content is extracted as follows: the first-level sub-node is {old blood pressure monitor verification failure (association strength 3 times), new device delay slightly increased (association strength 2 times)}; the second-level sub-node is {protocol redundancy increased (association strength 3 times), multiple devices disconnected (association strength 2 times)}; and the associated link list is organized as follows: historical node 1 → old blood pressure monitor verification failure (3), new device delay slightly increased (2) → protocol redundancy increased (3), multiple devices disconnected (2); Collect the first-level child nodes corresponding to similar initial nodes to construct a first-level problem set. Calculate the intersection of all first-level problem sets corresponding to similar initial nodes, denoted as Intersection 1: the direct problems caused by all similar initial nodes, representing the problem with the highest recurrence probability. Similarly, collect the second-level child nodes corresponding to similar initial nodes to construct a second-level problem set. Calculate the intersection of all second-level problem sets corresponding to similar initial nodes, denoted as Intersection 2: the long-term problems caused by all similar nodes. Calculate the intersection of Intersection 1 and Intersection 2, denoted as Intersection 3: the common direct problems and common long-term problems of all similar initial nodes, representing the most critical risk point. Calculate the arithmetic mean of the risk scores of all similar starting nodes and use it as the risk score of the protocol change content to be evaluated; The specific steps for calculating the risk score of similar starting nodes are as follows: Calculate the coverage of intersection 1 = number of problems contained in intersection 1 / total number of problems contained in the first-level child nodes corresponding to all similar initial nodes; calculate the coverage of intersection 2 = number of problems contained in intersection 2 / total number of problems contained in the second-level child nodes corresponding to all similar initial nodes; calculate the coverage of intersection 3 = number of problems contained in intersection 3 / sum of the total number of problems contained in the first-level child nodes corresponding to all similar initial nodes and the total number of problems contained in the second-level child nodes. The levels are classified according to the transmissibility of the problem's impact: Intersection 3 (which includes problems shared by first-level and second-level child nodes) is level 3, with a level coefficient of 30; Intersection 2 (which includes problems with second-level child nodes) is level 2, with a level coefficient of 20; Intersection 1 (which includes problems with first-level child nodes) is level 1, with a level coefficient of 10. Calculate the risk score of similar initial nodes = (intersection 1 coverage × grade coefficient 3) + (intersection 2 coverage × grade coefficient 2) + (intersection 3 coverage × grade coefficient 1).

[0021] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0022] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A communication protocol self-evolution system based on a pre-trained large model, characterized in that, include: The model building module extracts all historical protocol evolution cases with and without iterative drift, and constructs a protocol evolution influence chain graph based on the iterative drift protocol evolution cases, and constructs a pre-trained large model based on the protocol evolution cases without iterative drift. The risk assessment module inputs the current first communication request into the pre-trained large model to obtain the corresponding protocol change content to be evaluated. It then substitutes this content into the protocol evolution impact chain graph, matches it with similar starting nodes, and calculates the risk score of the protocol change content to be evaluated based on the similar initial nodes. If the risk score exceeds the preset threshold, the protocol is directly discarded; otherwise, it is determined to be a qualified protocol to be deployed.

2. The communication protocol self-evolution system based on a pre-trained large model according to claim 1, characterized in that, The specific steps for constructing a protocol evolution impact chain graph based on iterative drift protocol evolution case studies are as follows: The complete link information of historical iteration drift cases is statistically analyzed, including the protocol change content corresponding to the first communication request, the direct derivative problems that occur in the short term after the change, the chain problems caused by long-term accumulation, and the results of iteration drift. Extract the protocol change content corresponding to the first communication request as the starting node and annotate the specific parameters; Extract directly derived problems that appear in the short term after the change as first-level child nodes and mark their occurrence time; extract chain problems that have accumulated over a long period of time as second-level child nodes and mark their iterative relationships. Extract the iterative drift result as the termination node; finally form a single-case node chain of starting node → first-level child node → second-level child node → termination node; Collect all the disassembled single-case node chains, count the co-occurrence frequency between different nodes, and if the co-occurrence frequency is ≥2 times, establish an association edge between the two nodes and mark the co-occurrence frequency on the association edge to form a protocol evolution influence chain graph that includes nodes, association edges, and co-occurrence frequencies.

3. The communication protocol self-evolution system based on a pre-trained large model according to claim 2, characterized in that, The direct derivative problems that occur in the short term after the change refer to problems affecting device functionality that occur within 7 calendar days after the deployment of the protocol change content corresponding to the first communication requirement and without subsequent self-evolution iterations; the chain problems that occur after long-term accumulation refer to problems affecting device functionality that occur more than 30 calendar days after the deployment of the protocol change content corresponding to the first communication requirement and after at least 2 subsequent self-evolution iterations.

4. The communication protocol self-evolution system based on a pre-trained large model according to claim 2, characterized in that, When a new iterative drift case appears in the system, it is broken down into a single-case node chain. If there are unrecorded nodes in the node chain, they are added to the node library, and the co-occurrence frequency of the new node and the existing nodes is counted. If there are recorded nodes in the node chain, the co-occurrence frequency of its associated edges is incremented by 1. Invalid associated edges with a co-occurrence frequency of ≤1 and no new case matching for more than 6 months are removed.

5. The communication protocol self-evolution system based on a pre-trained large model according to claim 1, characterized in that, The specific steps for building a pre-trained large model based on protocol evolution cases without iterative drift are as follows: For each protocol evolution case without iterative drift, the first communication requirement and the corresponding protocol change are collected, and cases with missing parameters are removed to form a structured fine-tuning dataset. The structured fine-tuning dataset is transformed into an instruction-output training sample format, where the instruction is the first communication requirement and the output is the corresponding protocol change content. A general text generation model was chosen as the base model. The LoRA lightweight fine-tuning technique was used, with a low-rank matrix inserted only in the attention layer. The training hyperparameters for the LoRA lightweight fine-tuning task were: batch size of 8, learning rate of 1e-4, and training epochs of 5. Input a 10% untrained structured fine-tuning dataset. If the output protocol change content is consistent with the protocol change content corresponding to the instruction-output, the validation is successful. If the percentage of successful validations exceeds 90%, the basic model parameters and LoRA low-rank matrix parameters are packaged and saved to obtain the pre-trained large model.

6. The communication protocol self-evolution system based on a pre-trained large model according to claim 5, characterized in that, The specific method for determining whether the protocol change content output matches the protocol change content corresponding to the instruction output is as follows: By breaking down the protocol changes, we can obtain the core operations and objectives of the protocol changes, resulting in a set of operations and a set of objectives. Find the intersection and union of the operation sets of the output protocol change content and the corresponding operation set of the instruction-output protocol change content, and use the ratio of the intersection and union as the operation similarity. Similarly, obtain the target similarity, and calculate the overall similarity = (operation similarity + target similarity) / 2. If the overall similarity is ≥80%, the protocol change content output by the model will be deployed and run in the test environment for 30 consecutive days to monitor for any systemic problems. If no systemic problems occur, it can be determined that the output protocol change content is consistent with the protocol change content corresponding to the instruction-output.

7. The communication protocol self-evolution system based on a pre-trained large model according to claim 1, characterized in that, The specific operation for matching similar starting nodes is as follows: The actions and objects of the changes are extracted from the changes to be evaluated to obtain a set of real-time semantic features; the specific quantitative indicators involved in the changes are extracted from the changes to be evaluated to obtain a set of real-time parameter features. For each starting node, it is decomposed into a semantic feature set and a parameter feature set; a single feature in each set is used as the index key, and the index value is the ID of all starting nodes containing that feature; each index key is marked with a feature frequency level, that is, the frequency level is divided according to the frequency of the feature in all starting nodes, forming a complete feature inverted index table; Based on the generated real-time semantic feature set and real-time parameter feature set, the constructed feature inverted index table is retrieved, and all starting nodes containing any feature are extracted to obtain candidate starting nodes. The number of features that each candidate starting node has in common with the real-time semantic feature set and the real-time parameter feature set is counted and used as the number of co-occurring features. In co-occurrence features, the ratio of the number of high-frequency features to the number of co-occurrence features is calculated to obtain the proportion of high-frequency features; If the number of co-occurring features is greater than or equal to 60% of the total number of baseline features and the proportion of high-frequency features is greater than or equal to 50%, then the initial node is taken as the similarity starting node.

8. The communication protocol self-evolution system based on a pre-trained large model according to claim 7, characterized in that, The frequency level is determined based on the frequency of occurrence of this feature in all starting nodes: if the frequency of occurrence is greater than or equal to 30% of the total number of starting nodes, it is marked as a high-frequency feature; if the frequency of occurrence is between 10% and 30% of the total number of starting nodes, it is marked as a medium-frequency feature; and if the frequency of occurrence is less than 10% of the total number of starting nodes, it is marked as a low-frequency feature.

9. The communication protocol self-evolution system based on a pre-trained large model according to claim 1, characterized in that, The specific steps for calculating the risk score of the changes to the agreement to be evaluated are as follows: Starting from a similar starting node, along the associated edges in the protocol evolution influence chain graph, extract all first-level child nodes and second-level child nodes directly connected to it, and obtain the co-occurrence frequency of each associated edge. The extracted content is then structured and organized to form a list of associated links from similar starting nodes to first-level child nodes to second-level nodes. Collect the first-level child nodes corresponding to similar initial nodes to construct a first-level problem set. Calculate the intersection of all first-level problem sets corresponding to similar initial nodes, denoted as intersection one. Collect the second-level child nodes corresponding to similar initial nodes to construct a second-level problem set. Calculate the intersection of all second-level problem sets corresponding to similar initial nodes, denoted as intersection two. Calculate the intersection of intersection one and intersection two, denoted as intersection three. Calculate the arithmetic mean of the risk scores of all similar starting nodes and use it as the risk score of the protocol change content to be evaluated.

10. The communication protocol self-evolution system based on a pre-trained large model according to claim 9, characterized in that, The specific steps for calculating the risk score of similar starting nodes are as follows: Calculate the coverage of intersection 1 = number of problems contained in intersection 1 / total number of problems contained in the first-level child nodes corresponding to all similar initial nodes; calculate the coverage of intersection 2 = number of problems contained in intersection 2 / total number of problems contained in the second-level child nodes corresponding to all similar initial nodes; calculate the coverage of intersection 3 = number of problems contained in intersection 3 / sum of the total number of problems contained in the first-level child nodes corresponding to all similar initial nodes and the total number of problems contained in the second-level child nodes. The levels are classified according to the transmissibility of the problem's impact: Intersection 3 is level 3, with a level coefficient of 30; Intersection 2 is level 2, with a level coefficient of 20; Intersection 1 is level 1, with a level coefficient of 10. Calculate the risk score of similar initial nodes = (intersection 1 coverage × grade coefficient 3) + (intersection 2 coverage × grade coefficient 2) + (intersection 3 coverage × grade coefficient 1).

Images