Communication network performance prediction method and system based on large model

Abstract

The invention discloses a communication network performance prediction method and system based on a large model, and relates to a communication network and artificial intelligence. The method comprises the following steps: acquiring performance time series data from a historical database of a network management system, preprocessing the data in an offline training stage, and generating data description according to a prompt structure; the preprocessed performance time sequence data are respectively input into a time sequence large model TimeGPT for model fine tuning and a patch reprogramming module to generate a group of new vector representations; converting the data description into a time sequence prompt prefix by using a pre-trained large language model, and splicing the time sequence prompt prefix with a vector output by a patch reprogramming module to form a new time sequence to represent the understanding ability of the pre-trained large language model training model to the time sequence data; in the real-time stage, the TimeGPT carries out anomaly and performance prediction according to input performance time sequence data, detection and prediction results are input into a pre-trained large language model, and efficient and accurate detection results, anomaly description, prediction results and potential fault description are output.

Description

Technical Field

[0001] The present invention relates to the fields of communication network technology, artificial intelligence and big data analysis, and in particular to a communication network performance prediction method and system based on a large model. Background Art

[0002] With the rapid development of communication networks, the scale of networks is becoming larger and the structure is becoming more complex. Network anomaly detection and performance prediction have become key operation and maintenance technologies to ensure network stability and improve user experience. Traditional methods mostly rely on manual rules or simple statistical models, which are difficult to cope with large-scale and highly dynamic network data. In recent years, with the rise of deep learning technology, especially the expanded application of various large models in the fields of text reasoning and data analysis, new solutions have been provided for communication network anomaly detection and performance prediction.

[0003] The technical problems existing in the performance prediction of existing communication networks mainly include model bias caused by data imbalance, the complexity of high-dimensional data processing, the difficulty of modeling complex systems, and the problem of capturing long-term dependencies. In addition, existing communication network anomaly detection technologies face multiple challenges. On the one hand, traditional methods rely on fixed rules and feature libraries, which are difficult to adapt to the continuous evolution and complexity of the network, resulting in limited detection accuracy and easy omissions or false alarms. On the other hand, in the face of the explosive growth of network data volume, traditional technology has low processing efficiency and is difficult to achieve real-time or near-real-time detection, making it impossible to respond to network anomalies in a timely manner. In addition, most systems lack the ability to predict future network performance and potential threats, and cannot provide network administrators with forward-looking decision support. These problems and challenges require models to have stronger data processing capabilities, feature extraction capabilities, and model reasoning capabilities to cope with complex and changing practical application scenarios. Summary of the invention

[0004] The purpose of the present invention is to utilize the advantages of artificial intelligence technology to collaboratively solve the anomaly detection and performance prediction technical problems existing in communication networks through the text reasoning capabilities of the Large Language Model (LLM) and the detection and prediction capabilities of the large time series model TimeGPT. To this end, one purpose of the present invention is to provide a communication network performance prediction method based on a large model, which can timely and accurately detect various types of network anomalies, predict potential network performance degradation in the future, and provide corresponding description reasoning to assist in fault location, which can improve network operation and maintenance efficiency while reducing operation and maintenance difficulty, thereby providing strong guarantees for the operation of communication networks.

[0005] Another purpose of the present invention is to use the ability of large models to process massive data and powerful learning and reasoning capabilities to provide a communication network performance prediction system based on large models to provide technical support for network intelligent operation and maintenance. By refining the characteristic data of communication network equipment, an intelligent prediction and analysis algorithm based on large language model (LLM) and time series large model TimeGPT is adopted to automatically generate communication network performance data prediction and abnormal detection warning reports, realize real-time tracking of key network equipment performance indicators (Key Performance Indicator, KPI), and help users quickly grasp the overall situation of network operations.

[0006] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions.

[0007] The present invention provides a communication network performance prediction method based on a large model, comprising the following steps:

[0008] 1) Obtain performance time series data from the network management system historical database, perform data preprocessing and other normalization operations on the data during the offline training phase, and the operation and maintenance personnel generate data descriptions based on the designed prompt structure;

[0009] 2) The preprocessed performance time series data are input into the time series large model TimeGPT for model fine-tuning and patch reprogramming module to generate a set of new vector representations;

[0010] 3) Use the pre-trained large language model to convert the data description provided by the operation and maintenance personnel into a time series prompt prefix, and splice it with the vector output by the patch reprogramming module into a new time series representation, and then input it into the pre-trained large language model again to train the model's ability to understand time series data;

[0011] 4) In the real-time stage, the time series big model TimeGPT performs anomaly detection and performance prediction based on the input performance time series data, and inputs the detection and prediction results into the pre-trained big language model, which outputs the detection results and anomaly descriptions, prediction results, and potential fault descriptions;

[0012] 5) Regularly collect feedback from operation and maintenance personnel on the model prediction results to update the patch reprogramming module and large model, and improve the accuracy and reliability of the model through continuous learning and adjustment.

[0013] In step 1), the performance time series data includes various monitoring data of the network management system, including but not limited to connection rate, throughput, packet loss rate, bandwidth utilization, etc.; the data preprocessing and other normalization operations include: firstly, missing value processing is performed, including but not limited to interpolation method, missing value prediction method based on machine learning, and missing value filling method based on rules; then, the time series data is normalized into a sequence with zero mean and unit standard deviation through reversible instance normalization (Revin) to reduce the time series distribution drift.

[0014] In step 1), the operation and maintenance personnel generate data descriptions according to the designed prompt structure, and construct effective data descriptions as prefixes for subsequent time series for communication network operation and maintenance scenarios, including three key components: domain knowledge, task descriptions, and statistical information; domain knowledge provides LLM with basic background information about input time series, which usually exhibit different characteristics in different domains; task descriptions play a crucial guiding role in the patch embedding conversion process of specific tasks; statistical information (such as data trends) is used to enrich input time series to facilitate pattern recognition and reasoning.

[0015] In step 2), the preprocessed performance time series data are respectively input into the time series large model TimeGPT for model fine-tuning. Specifically, TimeGPT is a Transformer-based time series model with a self-attention mechanism. It is trained on the largest public time series collection and used for anomaly detection and time series prediction tasks. It can identify unknown anomalies and has strong generalization capabilities. Therefore, the preprocessed performance time series data can be used to fine-tune the TimeGPT model to improve the accuracy and performance of the model in anomaly detection and time series prediction tasks in network operation and maintenance scenarios.

[0016] It should be noted that performance prediction can be achieved through time series prediction. First, it is necessary to determine the key performance indicators (KPIs) related to network performance, such as network latency, throughput, packet loss rate, bandwidth utilization, etc. KPIs can reflect the current performance status of the network; therefore, TimeGPT can be used to perform KPI predictions to predict future trends in network performance, helping operation and maintenance personnel to promptly discover and respond to potential problems such as network degradation, and ensure network stability and reliability.

[0017] In step 2), the preprocessed performance time series data is input into the patch reprogramming module to generate a set of new vector representations, including: defining the first The input time series data is , , patch the preprocessed data, that is, cut it into pieces of length For consecutive overlapping patches of length The length of each patch is , the sliding window step size is , That is, there will be overlaps between patches, and the total number of patches is , better preserve local semantic information by aggregating local information into each patch;

[0018] The patch reprogramming module blocks and embeds the time series data, and aligns it with the semantic prototype of the natural language text through a cross-attention mechanism to generate a set of new vector representations; these vector representations contain both the information of the original time series and the semantic information related to the natural language text; therefore, these vectors can be regarded as an "encoded" or "transformed" form of the time series data, rather than the original time series data itself;

[0019] Specifically, the patch reprogramming module uses pre-trained word embeddings Reprogramming ,in is the vocabulary size; through linear probing To maintain a small set of text prototypes, denoted by ,in ; The present invention adopts multi-head cross attention layer to perform patch reprogramming operation, for each attention head , define the query matrix , the bond matrix , value matrix ,in is the patched sequence representation, , ; D is the hidden dimension of LLM, ; The operation of reprogramming the time series patches in each attention head is defined as:

[0020]

[0021] By aggregating each get , which is then linearly projected to align the hidden dimensions with the backbone LLM model, resulting in a new vector representation :

[0022]

[0023] Among them, Concat represents the concatenation operation. is the output weight matrix.

[0024] In step 3), the data description is converted into a time series prompt prefix by using a pre-trained large language model. Specifically, the data description generated in step 1) is input into the pre-trained LLM to generate a prompt prefix token. It should be noted that the pre-trained LLM is used as the backbone model and its parameters are kept frozen. The advantage of this is that the language knowledge and reasoning ability that the LLM has learned can be utilized without fine-tuning the LLM. The time series prompt prefix is ​​spliced ​​with the vector output by the patch reprogramming module into a new time series representation and input into the frozen LLM. The LLM performs reasoning learning based on the input time series representation and the prompt prefix, so that the LLM has the ability to analyze time series data.

[0025] Optionally, the pre-trained LLM mentioned above can use an existing open source large language model, such as Falcon, LLaMA, and Guanaco.

[0026] In step 4), the real-time stage TimeGPT performs anomaly detection and performance prediction based on the input performance time series data, and inputs the detection and prediction results into the pre-trained large language model, which outputs the detection results and anomaly descriptions, prediction results and potential fault descriptions; specifically, after TimeGPT completes anomaly detection and performance prediction, it passes these results (including anomaly points, abnormal time periods, predicted values, etc.) as input to the pre-trained LLM; after receiving the output of TimeGPT, the pre-trained LLM will use the analysis capabilities of time series data trained in step 3) and its natural language generation capabilities to generate easy-to-understand detection results and prediction descriptions; for the anomaly detection part, LLM will output a specific description of the anomaly, possible causes and recommended countermeasures; for the performance prediction part, LLM will output prediction results, potential risk points and suggestions or warnings based on the prediction results; these outputs not only contain technical details, but are also presented in a human-readable form, so that non-technical personnel can quickly understand and take corresponding actions.

[0027] In step 5), the feedback phase regularly updates the patch reprogramming module and the large model according to the operation and maintenance feedback results; specifically, the system regularly updates the patch reprogramming module and the large model according to the operation and maintenance feedback results to ensure that the model can continue to adapt to new environments, data distributions, and operation and maintenance requirements;

[0028] It should be noted that the system regularly collects feedback from the interactive modules on model performance and prediction accuracy, including but not limited to:

[0029] Anomaly detection accuracy: evaluates whether the anomalies detected by the model actually exist, and whether there are any missed or false detections;

[0030] Validity of prediction results: Verify whether the model accurately predicts future performance and whether the prediction helps to detect and solve problems in advance;

[0031] Clarity of output descriptions: Assess the clarity of the model output descriptions of test results and predictions to ensure that the information is easy to understand and act on.

[0032] Operation and maintenance feedback should include specific cases, error types (such as false positives, false negatives), prediction deviations, and other information, which is crucial for subsequent analysis and model updates;

[0033] After collecting operation and maintenance feedback, process the feedback data, including but not limited to data statistics, case extraction, etc.; update the big model based on the feedback and data analysis results;

[0034] It should be noted that the large model includes the large language model and the TimeGPT model;

[0035] The model update cycle is determined according to actual needs. The update strategy includes but is not limited to:

[0036] Time series data optimization: Based on feedback data, update the data preprocessing strategy to optimize the quality of time series data;

[0037] Expand training data: Extract operation-related text data to expand the text description of time series and optimize LLM input to improve its understanding of time series; feedback data (anomaly detection and prediction result feedback) is used as training data to fine-tune the TimeGPT model to improve detection and prediction performance;

[0038] Optimize output templates: Based on feedback from operations and maintenance personnel, optimize the structure of LLM's output templates and descriptive text to improve the clarity and usefulness of the output results.

[0039] The present invention provides a communication network performance prediction system based on a large model, which is provided with a data module, a large model module and an interaction module in sequence;

[0040] The data module includes a data acquisition module and a data processing module:

[0041] The data acquisition module is used to obtain the communication terminal performance time series data, and includes an interface with the data storage system (such as a database, a distributed file system, etc.) to achieve persistent storage of data; the data processing module is used to pre-process the performance data before the data is input into the large model module, including but not limited to data cleaning, data integration, data standardization, etc.;

[0042] The large model module includes a model training and reasoning module and a model optimization and update module:

[0043] The model training and reasoning module includes patch reprogramming of data, generation of prompt prefixes by the large language model, reasoning learning of the large language model based on the input patches and prompt prefixes, and fine-tuning of the TimeGPT model; the model optimization and update module regularly updates the patch reprogramming module and the large model based on operation and maintenance feedback data;

[0044] The interactive module includes a visualization module and a feedback module:

[0045] The visualization module is responsible for visually displaying the status of the communication network, performance data, and detection and prediction results of the large model to the operation and maintenance personnel in the form of graphics, charts, dashboards, etc. The feedback module is the feedback channel between the system and the operation and maintenance personnel. It is responsible for receiving feedback information submitted by the operation and maintenance personnel through the interactive interface, including but not limited to the evaluation of anomaly detection and performance prediction results, anomaly handling operations, and missed anomaly supplementation. The feedback module also includes some automated processing processes, including but not limited to calculating key indicators such as false alarm rate and missed alarm rate to quantify the performance of the model on specific tasks, and realize closed-loop optimization of operation and maintenance.

[0046] By integrating the reasoning ability of LLM and the advantages of TimeGPT in time series prediction, the present invention can achieve higher-precision anomaly detection and performance prediction compared to the prior art, and at the same time has powerful data reasoning capabilities, which is helpful for timely fault alarm and location. The present invention can achieve efficient and accurate anomaly detection and performance prediction, and can identify unknown anomalies and performance degradation. The present invention makes full use of data information, improves the comprehensibility of operation and maintenance, reduces the difficulty of operation and maintenance, helps to improve the efficiency of network operation and maintenance, and provides a new solution for the intelligent management of communication networks. The present invention can not only realize real-time analysis of massive data in communication networks, accurately identify network anomalies, and predict the changing trend of network performance in advance, but also provide easy-to-understand operation and maintenance information, thereby providing network administrators with timely and accurate operation and maintenance decision support, reducing the difficulty and cost of operation and maintenance, and providing strong support for building efficient and intelligent communication networks.

[0047] The present invention uses the TimeGPT model to process time series data, which can accurately identify various types of network anomalies and predict network performance. By combining the analysis and reasoning of the TimeGPT output results with a large language model, the anomaly description and performance prediction report are easier to understand and use, thereby improving network operation and maintenance efficiency. The innovation of the present invention lies in the deep integration of advanced large model technology with the actual needs of communication network operation and maintenance, providing strong support for the intelligent management of communication networks. Telecom operators can use the method and system of the present invention to conduct real-time monitoring and intelligent operation and maintenance of their large and complex communication networks, improve network stability and service quality, and reduce operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is an overall flow chart of a communication network performance prediction method based on a large model according to an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of the communication network performance prediction principle according to an embodiment of the present invention;

[0050] Figure 3 This is an example diagram of patch reprogramming in communication network performance prediction based on a large model according to an embodiment of the present invention;

[0051] Figure 4 The present invention is a schematic diagram of a communication network performance prediction system module based on a large model. DETAILED DESCRIPTION

[0052] The invention content of the present invention is described in detail below, and examples of the invention content are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0053] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific embodiments of the specification.

[0054] The embodiment of the present invention includes the following steps:

[0055] 1) Obtain performance time series data from the network management system historical database, perform data preprocessing and other normalization operations on the data during the offline training phase, and the operation and maintenance personnel generate data descriptions based on the designed prompt structure;

[0056] 2) The preprocessed performance time series data are input into the time series large model TimeGPT for model fine-tuning and patch reprogramming module to generate a set of new vector representations;

[0057] 3) Use the pre-trained large language model to convert the data description into a time series prompt prefix, and splice it with the vector output by the patch reprogramming module into a new time series representation, and then input it into the pre-trained large language model again to train the model's ability to understand time series data;

[0058] 4) In the real-time stage, TimeGPT performs anomaly detection and performance prediction based on the input performance time series data, and inputs the detection and prediction results into the pre-trained large language model. The pre-trained large language model outputs the detection results and anomaly descriptions, prediction results and potential fault descriptions;

[0059] 5) During the feedback phase, patch reprogramming modules and large models are regularly updated based on operation and maintenance feedback results.

[0060] Figure 1 The overall flow chart of a communication network performance prediction method based on a large model according to one invention content of the present invention specifically includes the following steps:

[0061] S101, during the offline training phase, communication network performance time series data is input for data preprocessing and other normalization operations, and the operation and maintenance personnel generate data descriptions according to the designed prompt structure.

[0062] Specifically, the input communication network performance time series data is first preprocessed to ensure the quality and consistency of the data, which is convenient for subsequent analysis and modeling; data preprocessing includes but is not limited to missing value processing, outlier detection and correction, data standardization or normalization, etc.; the operation and maintenance personnel generate data descriptions based on the designed prompt structure. This data description is to explain the characteristics and background of the time series data. For the communication network operation and maintenance scenario, an effective data description is constructed as a prefix for the subsequent time series, including three key components: domain knowledge, task description, and statistical information. Domain knowledge provides LLM with basic background information about the input time series, such as data source, business characteristics, industry standards, etc.; these input time series usually show different characteristics in different domains. The task description describes the relevant information of a specific task, such as task objectives, task nature, specific indicators required for the task, etc., which plays a vital guiding role in the patch embedding conversion process of a specific task, and can help the model clarify the goal, so as to select the appropriate algorithm and parameters for training. Statistical information describes some important statistical characteristics of time series data, such as data trends, periodicity, volatility, etc. Statistical information is used to enrich the input time series and help the model better understand the laws and patterns of the input time series to promote pattern recognition and reasoning. For example, if the data shows obvious periodic characteristics, the model can perform period prediction and anomaly detection based on this.

[0063] Furthermore, domain knowledge includes, but is not limited to, indicators related to access issues, such as RRC connection establishment success rate, QoS Flow establishment success rate, wireless connection rate, etc.

[0064] The task description includes but is not limited to: identifying and describing future trends of KPIs and related fault information, detecting network anomalies through KPIs, etc.

[0065] Statistical information includes, but is not limited to: statistical information of each KPI (for example, the minimum value of wireless connection rate is <minimum value>, the maximum value is <maximum value>, and it is generally within <interval>), KPI information related to specific problems (for example, low handover problem and <kpi1> 、 <kpi2>and <kpi3>related), KPI trend information (for example, the overall trend is <slow or fast>, <rising or falling>, <stable or sudden>, etc.).

[0066] S102, input the preprocessed performance time series data into the time series large model respectively, fine-tune the TimeGPT model according to the specific task, and generate a set of new vector representations by the patch reprogramming module.

[0067] Definition The input time series data is , , patch the preprocessed data, that is, cut it into pieces of length For consecutive overlapping patches of length The length of each patch is , the sliding window step size is , That is, there will be overlaps between patches, and the total number of patches is , better preserve local semantic information by aggregating local information into each patch.

[0068] Specifically, the patch reprogramming module uses pre-trained word embeddings Reprogramming ,in is the vocabulary size. Through linear probing To maintain a small set of text prototypes, denoted by ,in ; This embodiment uses a multi-head cross attention layer to perform patch reprogramming operations. For each attention head , define the query matrix , the bond matrix , value matrix , to capture the association information between patches; is the patched sequence representation, , . D is the hidden dimension of LLM, The operation of reprogramming the time series patches in each attention head is defined as:

[0069]

[0070] By aggregating each get , which is then linearly projected to align the hidden dimensions with the backbone LLM model, resulting in a new vector representation :

[0071]

[0072] Among them, Concat represents the concatenation operation. is the output weight matrix.

[0073] S103, based on the information provided by the operation and maintenance personnel, a data description containing domain knowledge, task instructions and statistical information is generated, and the data description is converted into a time series prompt prefix using a pre-trained large language model (LLM); the new vector representation output by the patch reprogramming module is concatenated with the time series prompt prefix after the data description is converted; the concatenated new time series representation is used as input and input into the pre-trained large language model again to train the model's ability to understand time series data. Through training, the model can better understand the communication network performance time series data and generate accurate detection results and prediction descriptions.

[0074] S104, in the real-time stage, TimeGPT performs anomaly detection and performance prediction based on the input performance time series data, and inputs the detection and prediction results into the pre-trained large language model LLM. The pre-trained large language model outputs the detection results and anomaly descriptions, prediction results and potential fault descriptions for reference by operation and maintenance personnel.

[0075] Specifically, after TimeGPT completes anomaly detection and performance prediction, it passes these results (including anomaly points, abnormal time periods, predicted values, etc.) as input to the pre-trained LLM. After receiving the output of TimeGPT, the pre-trained LLM uses its ability to analyze time series data and its natural language generation capabilities to generate easy-to-understand detection results and prediction descriptions. For the anomaly detection part, LLM will output a specific description of the anomaly, possible causes, and recommended countermeasures. For the performance prediction part, LLM will output the prediction results, potential risk points, and suggestions or warnings based on the prediction results. These outputs not only contain technical details, but are also presented in a human-readable form, allowing non-technical personnel to quickly understand and take appropriate actions.

[0076] S105, use the validation set to evaluate the trained model, including the accuracy of anomaly detection and the accuracy of performance prediction. Based on the evaluation results, determine whether the model needs to be further optimized, repeat steps S102 to S104, and conduct a new round of model training and evaluation until satisfactory performance is achieved. This ensures that the model has better performance and accuracy in practical applications.

[0077] S106, during the feedback phase, the large model is regularly updated based on the operation and maintenance feedback results.

[0078] It should be noted that the system regularly collects feedback from the interactive modules on model performance and prediction accuracy, including but not limited to:

[0079] Anomaly detection accuracy: Evaluate whether the anomalies detected by the model actually exist and whether there are any missed or false detections.

[0080] Validity of prediction results: Verify whether the model accurately predicts future performance and whether the predictions help detect and resolve issues in advance.

[0081] Clarity of output descriptions: Assess the clarity of the model’s output descriptions of detection results and predictions to ensure that the information is easy to understand and act on.

[0082] Operation and maintenance feedback should include specific cases, error types (such as false positives, missed negatives), prediction deviations and other information, which is crucial for subsequent analysis and model updates.

[0083] After collecting operation and maintenance feedback, process the feedback data, including but not limited to data statistics, case extraction, etc. Update the big model based on operation and maintenance feedback and data analysis results.

[0084] It should be noted that the large model includes the large language model and the TimeGPT model.

[0085] The model update cycle is determined according to actual needs. The update strategy includes but is not limited to:

[0086] Time series data optimization: Based on the feedback data, update the data preprocessing strategy to optimize the quality of time series data.

[0087] Expand training data: Extract operation-related text data to expand the text description of time series and optimize LLM input to improve its understanding of time series; feedback data (anomaly detection and prediction result feedback) is used as training data to fine-tune the TimeGPT model to improve detection and prediction performance.

[0088] Optimize output templates: Based on operation and maintenance feedback, optimize the structure of LLM's output templates and descriptive text to improve the clarity and usefulness of the output results.

[0089] like Figure 2 As shown, the schematic diagram of network performance prediction according to the content of the present invention specifically includes: in the offline training stage, normalization operations such as data preprocessing are performed on the time series data, and the operation and maintenance personnel generate data descriptions according to the designed prompt structure. Then, the preprocessed performance time series data are respectively input into the time series large model TimeGPT for model fine-tuning and the patch reprogramming module to generate a set of new vector representations. The data description is converted into a time series prompt prefix using the pre-trained LLM, and spliced ​​with the vector output by the patch reprogramming module into a new time series representation, which is input into the pre-trained LLM training model again to understand the time series data. In the real-time stage, TimeGPT performs anomaly detection and performance prediction based on the input performance time series data, and inputs the detection and prediction results into the pre-trained LLM. The pre-trained LLM outputs the detection results and anomaly descriptions, prediction results and potential fault descriptions. In the feedback stage, the patch reprogramming module and the large model are regularly updated according to the operation and maintenance feedback results.

[0090] It should be noted that the pre-trained LLM and TimeGPT in the figure are both open source large models. They do not need to be retrained and can be fine-tuned using a small amount of data. This reduces operation and maintenance costs while using the large model's massive data processing capabilities to improve anomaly detection and performance prediction efficiency.

[0091] The following is an example of data description design in the communication network performance prediction method based on a large model. Taking the 5G access network problem as an example, the 5G access network problem includes five major categories of KPI indicators. The data description includes three key components: domain knowledge, task description, and statistical information.

[0092] The following is the 5GKPI input data information:

[0093] <Begin>

[0094] [Data field]: Indicators related to access issues include RRC connection establishment success rate, QoS Flow establishment success rate,

[0095] Wireless connection rate...

[0096] [Task Description]: Identify and describe future trends of KPIs and related fault information...

[0097] [Statistics 1]: The minimum value of wireless connection rate is <minimum value>, the maximum value is <maximum value>, and it is generally within <range>...

[0098] [Statistics 2]: Low switching issues and <kpi1> 、 <kpi2>and <kpi3>related...

[0099] [Statistical information n]: The overall trend is <slow or fast>, <rising or falling>, <stable or sudden>...

[0100] <END>

[0101] Specifically, the data description includes start and end markers, as well as:

[0102] Domain knowledge includes but is not limited to: indicators related to access issues, such as RRC connection establishment success rate, QoSFlow establishment success rate, wireless connection rate, etc.

[0103] The task description includes but is not limited to: identifying and describing future trends of KPIs and related fault information, detecting network anomalies through KPIs, etc.

[0104] Statistical information includes, but is not limited to: statistical information of each KPI (for example, the minimum value of wireless connection rate is <minimum value>, the maximum value is <maximum value>, and it is generally within <interval>), KPI information related to specific problems (for example, low handover problem and <kpi1> 、 <kpi2>and <kpi3>related), KPI trend information (for example, the overall trend is <slow or fast>, <rising or falling>, <stable or sudden>, etc.).

[0105] like Figure 3 Shown is an example diagram of patch reprogramming in communication network performance prediction based on a large model according to the present invention.

[0106] As shown in the figure, language cues such as "downward", "steady" and "short-term" "upward" are learned by exploring a small text prototype, which are combined into patch 1 to represent the local patch information "short-term upward and then stable downward". This approach allows adaptive selection of relevant source information.

[0107] Specifically, a simple linear layer is used to extract text prototypes. This linear layer projects the original vocabulary onto a smaller prototype space, thereby reducing the number of "words" that need to be aligned. These text prototypes can be seen as a bridge between time series data and text data. The embedding representation of the temporal input features is used as a query and passed through a multi-head cross-attention layer with the embedding of the extracted text prototypes. In this way, the temporal input features are reprogrammed into a representation that is more compatible with natural language text.

[0108] In order to achieve the above invention, the embodiment of the present invention proposes a communication network performance prediction system based on a large model, such as Figure 4 As shown, the system includes: a data module, a large model module and an interaction module;

[0109] The data module includes a data acquisition module and a data processing module:

[0110] The data acquisition module is used to obtain the performance time series data of the communication terminal, and collects data from different nodes and terminals in the communication network through technical means such as API calls and log file reading. The data acquisition module is equipped with an interface with the data storage system (such as database, distributed file system, etc.) to realize the persistent storage of data; the data acquisition module sets the frequency and scope of data collection according to actual needs to ensure the comprehensiveness and timeliness of the data; in the process of data collection, strictly abide by relevant laws and regulations to ensure data privacy and security. The data processing module is used to pre-process the performance data before the data is input into the large model module to improve the data quality and model training effect; including but not limited to data cleaning module, data integration module, data standardization module, etc.; the data cleaning module is used to process missing values, outliers, etc., the data integration module is used to merge data from different sources; the data standardization module is used to convert data into a unified format and range, etc.

[0111] Optionally, the data processing module may be provided with a quality assessment module for checking the integrity, consistency and accuracy of the data after preprocessing to ensure data quality.

[0112] The large model module includes a model training and reasoning module and a model optimization and update module:

[0113] The model training and reasoning module includes a patch reprogramming unit, a prompt prefix generation unit, an inference learning unit, and a TimeGPT model fine-tuning unit; the patch reprogramming unit is used to patch the performance data and convert the time series data into a format recognizable by the model through a specific algorithm. The prompt prefix generation unit is used to use a large language model to generate prompt prefixes related to communication network performance to help the model better understand the data. The reasoning learning unit is used for the large language model to perform inference learning based on the input patches and prompt prefixes to predict the communication network performance. The TimeGPT model fine-tuning unit is used to fine-tune the TimeGPT model for the communication network performance prediction task to improve the prediction accuracy.

[0114] The model optimization and update module is used to regularly update the patch reprogramming module and large model based on operation and maintenance feedback data.

[0115] The interactive module includes a visualization module and a feedback module:

[0116] The visualization module is responsible for visually displaying the status of the communication network, performance data, and detection and prediction results of the large model to the operation and maintenance personnel in the form of graphics, charts, dashboards, etc. The feedback module is the feedback channel between the system and the operation and maintenance personnel, which is used to receive feedback information submitted by the operation and maintenance personnel through the interactive interface, including but not limited to the evaluation of performance prediction results, abnormal handling operations, abnormal missed detection supplements, etc., and automatically processes this information to improve the performance of the model, including but not limited to calculating key indicators such as false alarm rate and missed alarm rate to quantify the performance of the model on specific tasks. Achieve closed-loop optimization of operation and maintenance. Operation and maintenance personnel can submit feedback information through web forms, mobile applications, etc.

[0117] Optionally, the interaction module may be provided with a privacy and security module to ensure the privacy and security of feedback information and prevent abuse and false alarms.

[0118] In summary, a communication network performance prediction method and system based on a large model provided in accordance with an embodiment of the present invention includes the following steps: first, performance time series data is obtained from a historical database of a network management system, and data preprocessing and other normalization operations are performed on the data in an offline training phase, and the operation and maintenance personnel generate data descriptions according to the designed prompt structure. Then, the preprocessed performance time series data are respectively input into the time series large model TimeGPT for model fine-tuning and the patch reprogramming module to generate a set of new vector representations. The data description is converted into a time series prompt prefix using a pre-trained large language model, and the vector output by the patch reprogramming module is spliced ​​into a new time series representation, which is then input into the pre-trained large language model again to train the model's ability to understand time series data. In the real-time phase, TimeGPT performs anomaly detection and performance prediction based on the input performance time series data, and inputs the detection and prediction results into the pre-trained large language model, which outputs the detection results and anomaly descriptions, prediction results, and potential fault descriptions. In the feedback phase, the patch reprogramming module and the large model are regularly updated according to the operation and maintenance feedback results. By integrating the reasoning ability of LLM with the advantages of TimeGPT in time series prediction, the present invention can achieve higher-precision anomaly detection and performance prediction compared to the prior art, and at the same time has powerful data reasoning capabilities, which is helpful for timely fault alarm and location. The present invention makes full use of data information, improves the comprehensibility of operation and maintenance, reduces the difficulty of operation and maintenance, helps to improve the operation and maintenance efficiency of communication networks, and provides a new solution for the intelligent management of communication networks. The present invention uses the TimeGPT model to process time series data, which can accurately identify various types of network anomalies and predict network performance. By combining the analysis and reasoning of the TimeGPT output results with a large language model, the anomaly description and performance prediction report are easier to understand and use, thereby improving the efficiency of network operation and maintenance.

[0119] The above embodiments are only preferred embodiments of the present invention and cannot be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention. < / kpi1> < / kpi1> < / kpi1>

Claims

1. A communication network performance prediction method based on a large model, characterized in that The following steps are involved: 1) Obtain performance time series data from the network management system historical database, preprocess the data during the offline training phase, and the operation and maintenance personnel generate data descriptions based on the designed prompt structure; 2) The preprocessed performance time series data are input into the time series large model TimeGPT for model fine-tuning and patch reprogramming module to generate a set of new vector representations; 3) Use the pre-trained large language model to convert the data description provided by the operation and maintenance personnel into a time series prompt prefix, and splice it with the vector output by the patch reprogramming module into a new time series representation, and then input it into the pre-trained large language model again to train the model's ability to understand time series data; 4) In the real-time stage, the time series big model TimeGPT performs anomaly detection and performance prediction based on the input performance time series data, and inputs the detection and prediction results into the pre-trained big language model, which outputs the detection results and anomaly descriptions, prediction results, and potential fault descriptions; 5) Regularly collect feedback from operation and maintenance personnel on the model prediction results, update patch reprogramming modules and large models, and improve the accuracy and reliability of the model through continuous learning and adjustment.

2. A communication network performance prediction method based on a large model as claimed in claim 1, characterized in that In step 1), the performance time series data includes various monitoring data of the network management system, including but not limited to connection rate, throughput, packet loss rate, and bandwidth utilization; the data preprocessing includes: firstly performing missing value processing, including but not limited to interpolation method, missing value prediction method based on machine learning, and missing value filling method based on rules; then normalizing the time series data into a sequence with zero mean and unit standard deviation through reversible instance normalization to reduce time series distribution drift.

3. A communication network performance prediction method based on a large model as claimed in claim 1, characterized in that In step 1), the operation and maintenance personnel generate data descriptions according to the designed prompt structure, and construct effective data descriptions as prefixes for subsequent time series for the communication network operation and maintenance scenario, including three key components: domain knowledge, task description, and statistical information; Domain knowledge provides LLM with basic background information about the input time series, which usually exhibit different characteristics in different domains; The task description plays a crucial guiding role in the task-specific patch embedding conversion process; the statistical information is used to enrich the input time series to facilitate pattern recognition and reasoning.

4. A communication network performance prediction method based on a large model as claimed in claim 1, characterized in that In step 2), the preprocessed performance time series data are input into the large time series model TimeGPT for model fine-tuning. TimeGPT is a Transformer-based time series model with a self-attention mechanism. It is trained on the largest public time series collection and used for anomaly detection and time series prediction tasks. It can identify unknown anomalies and has strong generalization capabilities. Therefore, the preprocessed performance time series data is used to fine-tune the TimeGPT model to improve the accuracy and performance of the model in anomaly detection and time series prediction tasks in network operation and maintenance scenarios.

5. A communication network performance prediction method based on a large model as claimed in claim 1, characterized in that In step 2), the preprocessed performance time series data is input into the patch reprogramming module to generate a set of new vector representations, including: defining the first The input time series data is , , patch the preprocessed data, that is, cut it into pieces of length For consecutive overlapping patches of length The length of each patch is , the sliding window step size is , That is, there will be overlaps between patches, and the total number of patches is , better preserve local semantic information by aggregating local information into each patch; The patch reprogramming module blocks and embeds the time series data, and aligns it with the semantic prototype of the natural language text through a cross-attention mechanism to generate a set of new vector representations; these vector representations contain both the information of the original time series and the semantic information related to the natural language text; therefore, these vectors are regarded as an "encoded" or "transformed" form of the time series data, rather than the original time series data itself; Specifically, the patch reprogramming module uses pre-trained word embeddings Reprogramming ,in is the vocabulary size; through linear probing To maintain a small set of text prototypes, denoted by ,in ; Use multi-head cross attention layer for patch reprogramming operation, for each attention head , define the query matrix , the bond matrix , value matrix ,in is the patched sequence representation, , ; D is the hidden dimension of LLM, ; The operation of reprogramming the time series patches in each attention head is defined as: ； By aggregating each get , which is then linearly projected to align the hidden dimensions with the backbone LLM model, resulting in a new vector representation : ； Among them, Concat represents the concatenation operation. is the output weight matrix.

6. A communication network performance prediction method based on a large model as claimed in claim 1, characterized in that In step 3), the data description is converted into a time series prompt prefix using a pre-trained large language model. Specifically, the data description generated in step 1) is input into the pre-trained LLM to generate a prompt prefix token; the pre-trained LLM is used as the backbone model, its parameters are kept frozen, and the language knowledge and reasoning ability that the LLM has learned are utilized without fine-tuning the LLM; the time series prompt prefix is ​​concatenated with the vector output by the patch reprogramming module into a new time series representation and input into the frozen LLM. The LLM performs reasoning learning based on the input time series representation and the prompt prefix, so that the LLM has the ability to analyze time series data.

7. A communication network performance prediction method based on a large model as claimed in claim 1, characterized in that In step 4), in the real-time stage, TimeGPT performs anomaly detection and performance prediction based on the input performance time series data, and inputs the detection and prediction results into the pre-trained large language model, which outputs the detection results and anomaly descriptions, prediction results and potential fault descriptions. Specifically, after TimeGPT completes anomaly detection and performance prediction, it passes these results as input to the pre-trained LLM. After receiving the output of TimeGPT, the pre-trained LLM uses the time series data analysis capabilities trained in step 3) and its natural language generation capabilities to generate easy-to-understand detection results and prediction descriptions. For the anomaly detection part, LLM will output a specific description of the anomaly, possible causes, and recommended countermeasures. For the performance prediction part, LLM will output prediction results, potential risk points, and suggestions or warnings based on the prediction results. These outputs not only contain technical details, but are also presented in a human-readable form, so that non-technical personnel can quickly understand and take appropriate actions.

8. A communication network performance prediction method based on a large model as claimed in claim 1, characterized in that In step 5), feedback from operation and maintenance personnel on the model prediction results is collected regularly, including but not limited to: Anomaly detection accuracy: evaluates whether the anomalies detected by the model actually exist, and whether there are any missed or false detections; Validity of prediction results: Verify whether the model accurately predicts future performance and whether the prediction helps to detect and solve problems in advance; Clarity of output descriptions: Assess the clarity of the model output descriptions of test results and predictions to ensure that the information is easy to understand and act on. Operation and maintenance feedback should include specific cases, error types, and prediction deviations. This information is crucial for subsequent analysis and model updates. After collecting operation and maintenance feedback, process the feedback data, including but not limited to data statistics and case extraction; Update the big model based on feedback and data analysis results.

9. A communication network performance prediction method based on a large model as claimed in claim 1, characterized in that In step 5), the large model includes a large language model and a TimeGPT model; The model update cycle is determined according to actual needs. The update strategy includes but is not limited to: Time series data optimization: Based on feedback data, update the data preprocessing strategy to optimize the quality of time series data; Expand training data: Extract operation-related text data to expand the text description of time series and optimize LLM input to improve its understanding of time series; use feedback data as training data to fine-tune the TimeGPT model to improve detection and prediction performance; Optimize output templates: Based on feedback from operations and maintenance personnel, optimize the structure of LLM's output templates and descriptive text to improve the clarity and usefulness of the output results.

10. A communication network performance prediction system based on a large model, characterized in that There are data module, large model module and interaction module in sequence; The data module includes a data acquisition module and a data processing module; Among them, the data acquisition module is used to obtain the performance time series data of the communication terminal, and the data acquisition module includes an interface with the data storage system to realize the persistent storage of data; the data processing module is used to pre-process the performance data before the data is input into the large model module to improve the data quality and model training effect; it includes a data cleaning module, a data integration module, and a data standardization module; the data cleaning module is used to process missing values ​​and abnormal values, and the data integration module is used to merge data from different sources; the data standardization module is used to convert the data into a unified format and range; The large model module includes a model training and reasoning module and a model optimization and update module: Among them, the model training and reasoning module includes a patch reprogramming unit, a prompt prefix generation unit, an inference learning unit, and a TimeGPT model fine-tuning unit; the patch reprogramming unit is used to patch the performance data and convert the time series data into a format recognizable by the model through a specific algorithm; the prompt prefix generation unit is used to use a large language model to generate prompt prefixes related to communication network performance to help the model better understand the data; the inference learning unit is used for the large language model to perform inference learning based on the input patches and prompt prefixes to predict the communication network performance; the TimeGPT model fine-tuning unit is used to fine-tune the TimeGPT model for the communication network performance prediction task to improve the prediction accuracy; the model optimization and update module is used to regularly update the patch reprogramming module and the large model according to the operation and maintenance feedback data; The interactive module includes a visualization module and a feedback module: Among them, the visualization module is responsible for intuitively displaying the status of the communication network, performance data, and detection and prediction results of the large model to the operation and maintenance personnel in the form of graphics, charts, dashboards, etc.; the feedback module is the feedback channel between the system and the operation and maintenance personnel, which is used to receive feedback information submitted by the operation and maintenance personnel through the interactive interface, including but not limited to the evaluation of performance prediction results, abnormal handling operations, and abnormal missed detection supplements, and automatically processes this information to improve model performance, including but not limited to calculating false alarm rate and missed alarm rate, so as to quantify the performance of the model on specific tasks and realize operation and maintenance closed-loop optimization.

Images