A dual-source timing intelligent supervision optimization method based on Beidou and 5G internet of things

Abstract

The application discloses a double-source timing intelligent supervision optimization method based on Beidou and 5G Internet of Things, and relates to the cross field of timing technology and artificial intelligence technology. The application collects Beidou timing data, 5G-A Internet of Things timing data and environment auxiliary data, constructs high-fidelity timing sequence features after preprocessing; based on a CNN-LSTM-Transformer hybrid AI model, combined with distributed satellite common view DCV technology, extracts abnormal degree quantitative values and double-source consistency scores, and integrates to obtain a timing supervision judgment parameter set; based on the parameter set, through a hierarchical response strategy, dynamically adjusts the Beidou and 5G-A timing weights; constructs a traceability database to store early warning information, optimization parameters and full-process data; through a supervision optimization model composed of a reinforcement learning agent and a double-optimization branch, outputs optimization instructions of the hybrid AI model and the hierarchical response strategy. The application realizes double-source timing full-process intelligent supervision and dynamic iterative optimization, effectively improves the precision, stability and reliability of wide-area multi-node nanosecond-level timing.

Description

Technical Field

[0001] This invention relates to the intersection of timing technology and artificial intelligence technology, specifically to a dual-source timing intelligent supervision and optimization method based on BeiDou and 5G Internet of Things. Background Technology

[0002] With the rapid iterative development of key fields such as 5G communication, computing networks, smart grids, industrial internet, and autonomous driving, wide-area, multi-node distributed deployment scenarios are becoming increasingly common. These scenarios have raised the requirements for timing accuracy, operational stability, and long-term reliability to the nanosecond level. Even minute deviations in timing accuracy can lead to serious problems such as data transmission errors, equipment coordination failures, and system malfunctions, directly impacting the normal operation of core areas. Currently, existing timing methods mostly rely on a single timing source, making them susceptible to complex factors such as extreme weather, electromagnetic interference, signal blockage, and link loss, frequently resulting in timing interruptions and sudden drops in accuracy, failing to meet the demands for high reliability and high precision. Against this backdrop, multi-source fusion timing technology has become a research hotspot in the field for addressing the shortcomings of single timing sources and improving timing performance.

[0003] Chinese patent (publication number CN119780585A) discloses a method and related equipment for monitoring the status of power equipment based on BeiDou high-precision time synchronization. This method obtains high-precision time synchronization signals through BeiDou to provide a time reference for monitoring the status of power equipment, realizes the time-series monitoring of the operating status of power equipment, focuses on the application extension of a single BeiDou time synchronization source in a specific industry (power), and ensures the time synchronization of power equipment monitoring data. However, this method relies on a single BeiDou time synchronization source, has limited adaptability to scenarios, and does not introduce artificial intelligence algorithms, so it cannot efficiently achieve accurate quantification of time synchronization anomalies and dynamic iterative optimization of parameters. The time synchronization accuracy, stability and reliability cannot meet the requirements of nanosecond-level time synchronization and full-process supervision.

[0004] Chinese patent (publication number CN120639232B) discloses a hybrid timing method based on BeiDou satellite navigation and TSN. This method obtains basic timing signals from BeiDou satellites and combines the time synchronization advantages of TSN technology to improve the latency and jitter of timing signals during transmission, thereby enhancing the stability of timing signal transmission. It is suitable for scenarios with high transmission synchronization requirements and achieves basic hybrid timing signal output and transmission optimization. However, this method focuses on the transmission synchronization optimization of timing signals and does not involve timing anomaly identification, quantification, and accurate response. Its anti-interference capability is limited, and it cannot cope with timing failure problems in scenarios such as BeiDou signal blockage and extreme weather. Furthermore, it does not introduce artificial intelligence models and reinforcement learning-related technologies, so it cannot achieve dynamic iterative optimization of timing parameters, making it difficult to continuously improve timing accuracy and meet nanosecond-level timing requirements.

[0005] In view of the shortcomings of the existing technologies, there is an urgent need for a timing supervision method that can achieve deep fusion of dual sources, intelligent supervision of the whole process, and dynamic iterative optimization, so as to solve the core problems of inaccurate anomaly identification, inaccurate response, and inability to continuously optimize in the existing technologies. Summary of the Invention

[0006] To address the aforementioned technical issues, this application discloses a dual-source timing intelligent supervision and optimization method based on BeiDou and 5G IoT, specifically including: Collect dual-source timing data from BeiDou and 5G-A IoT, combine it with environmental auxiliary data, and construct a high-fidelity time series feature sequence after preprocessing; Based on the preprocessed data, a wide-area multi-node time synchronization calibration is achieved by using a CNN-LSTM-Transformer hybrid AI model combined with DCV technology, and time synchronization supervision and judgment parameters are output. Based on the timing supervision and judgment parameters, through a hierarchical response strategy, optimized parameters are output, dual-source wide-area collaboration is carried out, and early warning information is recorded and pushed. Based on optimized parameters and early warning information, a source tracing database is constructed; Based on the source database, the supervised optimization model outputs optimization instructions for the CNN-LSTM-Transformer hybrid AI model and the hierarchical response strategy. The supervised optimization model is constructed based on the reinforcement learning subject and dual optimization branches.

[0007] Preferably, the high-fidelity time series feature sequence specifically includes: BeiDou timing data including at least a timestamp, 1PPS signal, and phase error; GA IoT timing data including at least an NTP / NITZ timestamp and time offset; and environmental auxiliary data including at least operating temperature and electromagnetic interference intensity. The three types of collected data undergo multi-stage preprocessing to remove invalid interference, unify data format, and enhance features. Then, through feature weighted fusion, the preprocessed three types of data are collaboratively integrated to obtain a high-fidelity time-series feature sequence. .

[0008] Preferably, the timing supervision and judgment parameters specifically involve: performing hierarchical feature extraction on the input high-fidelity time-series feature sequence and outputting anomaly quantification values. ,based on Determine the level of abnormality ; By using distributed satellite common-view DCV technology and combining time synchronization data collected from multiple nodes in a wide area, the time difference between nodes is calculated to establish a unified wide-area synchronization benchmark, and the time synchronization data after preliminary processing by the hybrid AI model is precisely calibrated. Calculate the dual-source consistency score based on the DCV-calibrated BeiDou and 5G-A dual-source timing data. Quantify the matching degree of dual-source time synchronization data; Integrating dual-source consistency scores Abnormal level By combining the anomaly type and the wide-area synchronization deviation value, a set of timing supervision and judgment parameters is obtained.

[0009] Preferably, the CNN-LSTM-Transformer hybrid AI model specifically involves: in hierarchical feature extraction, the CNN captures short-term abnormal fluctuations in the timing data, as shown in the formula: in, The local feature vectors output by the CNN layer. Here, is the Sigmoid activation function, and is the convolutional kernel weight matrix of the CNN layer. For the bias term of the CNN layer, This is a two-dimensional convolution operation; LSTM, based on the local features output by CNN, focuses on the long-term changes in timing data, capturing the long-range temporal dependence of timing data and capturing the temporal trend. The formula is as follows: in, for The temporal feature vector output by the LSTM layer at time step 1. The hyperbolic tangent activation function is used. Here is the weight matrix of the LSTM layer. For the bias term of the LSTM layer, For feature splicing operations; Transformer uses self-attention to apply temporal features to the LSTM output. Further optimization was performed, strengthening the focus on core features, resulting in... The original temporal feature vector output by the LSTM attention layer at time step After self-attention optimization, the optimized features are processed through the fully connected layer of the model to output anomaly feature vectors, as shown in the formula: in, , , These are the weight matrices for the first and second fully connected layers, respectively. , These are the bias terms for the first and second fully connected layers, respectively. This represents the anomalous feature vector output by the fully connected layer. Based on the output anomaly quantization value Calculate the quantification value of the degree of abnormality. The formula is: in, For the first Core anomaly features in The value at time, The standard threshold for the k-th type of core anomaly feature obtained by training with historical normal data. For the first Weights of core anomaly features The number of core abnormal features.

[0010] Preferably, the precise calibration of the timing data after preliminary processing of the hybrid AI model specifically involves: using distributed satellite common-view (DCV) technology, based on multiple geographically dispersed nodes, simultaneously observing the timing signal of the same BeiDou satellite, calculating the signal transmission time difference between each node and the satellite, and indirectly obtaining the time difference between each node. Setting a time difference threshold based on nanosecond-level timing monitoring requirements ,according to Calibration is required as per the specifications.

[0011] Preferably, the step involves selecting a BeiDou time stamp sequence calibrated by DCV. and 5G-A timing timestamp sequence Using a certain number of continuous sample data, the consistency score of dual-source time synchronization is calculated using the following formula: in, For the dual-source consistency score, For the number of data samples, , They are respectively The timestamp for BeiDou and 5G-A timing after DCV calibration; when Less than the preset score threshold When this occurs, it is determined to be an anomaly in the consistency between two sources.

[0012] Preferably, the hierarchical response strategy specifically comprises: inputting a set of timing supervision and judgment parameters, including response level and anomaly level. Correspondingly, the response intensity gradually increases with the level of anomaly. Level 1 minor anomalies are continuously monitored and slightly adjusted; Level 2 general anomalies are given early warnings and parameter adjustments; Level 3 severe anomalies are coordinated and optimized; and Level 4 critical anomalies are switched to emergency mode and optimized comprehensively. Based on the response level, the optimized dynamic weights for BeiDou and 5G-A timing are determined using the following formula: in, , These are the optimized dynamic weights for BeiDou and 5G-A timing synchronization under the L-level response. Initial weights for BeiDou timing. This is the weighting adjustment coefficient.

[0013] Preferably, the traceability database is specifically: based on early warning information data, optimized parameter data, and full-process related data, after data regularization by unifying timestamps, removing invalid and duplicate data, and correcting data deviations, the data is stored in association through three tables: the optimized parameter table, the early warning information table, and the full-process related data table, sharing a timestamp index; Data storage adopts a hybrid storage mode combining time-series databases and relational databases; A query mechanism is built based on a triple index of timestamp, response level, and exception type.

[0014] Preferably, the supervised optimization model specifically adopts an architecture of reinforcement learning subject and dual optimization branches. The reinforcement learning subject performs global optimization decision-making and benefit calculation, and the dual optimization branches are a hybrid AI model parameter optimization branch and a hierarchical response strategy threshold optimization branch, respectively. Data-driven dynamic optimization is realized based on the traceability database. The reinforcement learning module is constructed using the DQN algorithm, aiming to optimize the overall performance of the time synchronization system. It includes an input layer, a hidden layer, and an output layer. The input layer receives the state. The output layer outputs the optimal action. , , These are optimization instructions for dual optimization branches; The reinforcement learning reward function is: in, for Reward value at any moment, state , for The core dataset after normalization in the database is traced back to its source at all times. for The accuracy of the time synchronization system for Quantitative value of the degree of anomaly at any given moment for Time-based dual-source consistency score, for Maximum deviation of wide-area synchronization at any given time , , , These are the weights of the corresponding parameters.

[0015] Preferably, the dual optimization branches specifically refer to: a hybrid AI model parameter optimization branch receiving optimization instructions. Extract the parameter adjustment amounts of each layer in the instruction, and dynamically adjust all parameters of the hybrid AI model; after optimization, apply the hybrid AI model to real-time timing data processing, calculate the optimized anomaly degree quantification value, compare it with the anomaly degree before optimization, record the optimization difference, and record the optimized model parameters, optimized anomaly degree quantification value, and optimization difference into the traceability database; Hierarchical response strategy threshold optimization branch receiving optimization instructions Extract the threshold adjustment amount of each level of response in the instruction, and optimize the triggering conditions of the graded response and the threshold for the classification of early warning levels; after optimization, record the new threshold parameters and optimization basis, and update the full-process related data table of the traceability database synchronously.

[0016] Compared with the prior art, the technical solution of this application has the following technical effects: This invention integrates BeiDou and 5G-A IoT dual-source timing data and introduces environmental auxiliary data. After preprocessing and weighted fusion, a high-fidelity time series feature sequence is constructed, which effectively avoids the problems of single timing source being susceptible to signal blockage and electromagnetic interference. The dual-source complementarity ensures the continuity of timing and is suitable for complex working conditions with multiple nodes in a wide area.

[0017] This invention uses a CNN-LSTM-Transformer hybrid AI model for hierarchical feature extraction and combines DCV technology to achieve wide-area calibration. It can accurately quantify the degree of time synchronization anomalies and calculate the dual-source consistency score, solving the pain points of inaccurate anomaly identification and inability to quantify existing technologies, and providing accurate basis for subsequent response strategies.

[0018] This invention formulates a graded response strategy based on the anomaly level, dynamically adjusts the timing weights of BeiDou and 5G-A, and accurately matches the response intensity with the anomaly level. This enables differentiated responses to minor anomaly monitoring and fine-tuning, and severe anomaly emergency switching, avoiding insufficient or excessive response and improving timing stability.

[0019] This invention constructs a supervised optimization model based on a source database, which is a reinforcement learning subject with two optimization branches. It can dynamically output optimization instructions of hybrid AI model and hierarchical response strategy, realize adaptive parameter adjustment, continuously improve timing accuracy to the nanosecond level, and meet the long-term timing supervision requirements of wide area and multi-node.

[0020] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0023] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows: Figure 1 The flowchart shows the overall process of a dual-source timing intelligent supervision and optimization method based on BeiDou and 5G IoT. Figure 2 This is a diagram illustrating the overall architecture of a dual-source timing intelligent supervision and optimization method based on BeiDou and 5G IoT. Figure 3 This is an architecture diagram of the CNN-LSTM-Transformer hybrid AI model in this application; Figure 4 This is an architecture diagram of the supervised optimization model in this application; Figure 5 This is a graph showing the iterative data of the CNN-LSTM-Transformer hybrid AI model in the embodiments of this application; Figure 6 This is a comparison chart of the timing deviations of the various methods in the application embodiments under four working conditions; Figure 7 This is a comparison chart of the anomaly detection capabilities of various methods in the application embodiments; Figure 8 This is a comparison chart of the overall performance of the various methods in the application embodiments. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0025] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0026] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0027] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0028] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0029] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0030] Example 1 mainly describes a dual-source timing intelligent supervision and optimization method based on BeiDou and 5G IoT, such as... Figure 1 , Figure 2 As shown, it specifically includes: Collect dual-source timing data from BeiDou and 5G-A IoT, combine it with environmental auxiliary data, and construct a high-fidelity time series feature sequence after preprocessing; Based on the preprocessed data, a wide-area multi-node time synchronization calibration is achieved by using a CNN-LSTM-Transformer hybrid AI model combined with DCV technology, and time synchronization supervision and judgment parameters are output. Based on the timing supervision and judgment parameters, through a hierarchical response strategy, optimized parameters are output, dual-source wide-area collaboration is carried out, and early warning information is recorded and pushed. Based on optimized parameters and early warning information, a source tracing database is constructed; Based on the source database, the supervised optimization model outputs optimization instructions for the CNN-LSTM-Transformer hybrid AI model and the hierarchical response strategy. The supervised optimization model is constructed based on the reinforcement learning subject and dual optimization branches.

[0031] Furthermore, the high-fidelity time series feature sequence specifically includes: BeiDou time synchronization data including at least a timestamp, 1PPS signal, and phase error; GA IoT time synchronization data including at least an NTP / NITZ timestamp and time offset; and environmental auxiliary data including at least operating temperature and electromagnetic interference intensity. The three types of collected data undergo multi-stage preprocessing to remove invalid interference and unify the data format. Then, through feature-weighted fusion, the preprocessed three types of data are collaboratively integrated to obtain a high-fidelity time-series feature sequence. The formula is: in, To achieve high-fidelity temporal feature values ​​after fusion, , , They are respectively Normalized values ​​of preprocessed BeiDou timing data, 5G-A timing data, and environmental auxiliary data. These are the feature weights, where the weight of the BeiDou timing data is... Highest weight: Environmental auxiliary data Dynamically adjust according to actual working conditions to achieve collaborative representation of multi-dimensional features.

[0032] Furthermore, outlier removal employs the 3σ criterion to filter out valid data and eliminate anomalies that occur during the data collection process; time alignment uses the BeiDou 1PPS timing signal as a benchmark to synchronize 5G-A timing data and environmental auxiliary data, ensuring the matching of the three types of data at the same time t; normalization processes uniformly normalize the three types of data to the [0,1] interval to eliminate interference caused by differences in data dimensions and units; feature enhancement slightly enhances the normalized data to strengthen the effective features in the data and weaken random interference.

[0033] Furthermore, the timing supervision and judgment parameters are specifically as follows: hierarchical feature extraction is performed on the input high-fidelity time-series feature sequence, and the anomaly degree quantification value is output. ,based on Determine the level of abnormality The formula is: Calculate the dual-source consistency score based on the DCV-calibrated BeiDou and 5G-A dual-source timing data. Quantify the matching degree of dual-source time synchronization data; Integrating dual-source consistency scores Abnormal level By combining the anomaly type and the wide-area synchronization deviation value, a set of timing supervision and judgment parameters is obtained.

[0034] Furthermore, the CNN-LSTM-Transformer hybrid AI model specifically involves: in hierarchical feature extraction, CNN captures short-term abnormal fluctuations in the timing data, as shown in the formula: in, The local feature vectors output by the CNN layer. Here, is the Sigmoid activation function, and is the convolutional kernel weight matrix of the CNN layer. For the bias term of the CNN layer, This is a two-dimensional convolution operation; LSTM, based on the local features output by CNN, focuses on the long-term changes in timing data, capturing the long-range temporal dependence of timing data and capturing the temporal trend. The formula is as follows: in, Let be the temporal feature vector output by the LSTM layer at time t. The hyperbolic tangent activation function is used. Here is the weight matrix of the LSTM layer. For the bias term of the LSTM layer, For feature splicing operations; Transformer uses self-attention to apply temporal features to the LSTM output. Further optimization is performed to enhance the focus on core features, resulting in the original temporal feature vector output by the LSTM attention layer at time t. After self-attention optimization, the optimized features are processed through the fully connected layer of the model to output anomaly feature vectors, as shown in the formula: in, , , These are the weight matrices for the first and second fully connected layers, respectively. , These are the bias terms for the first and second fully connected layers, respectively. This represents the anomalous feature vector output by the fully connected layer. Based on the output anomaly quantization value Calculate the quantification value of the degree of abnormality. The formula is: in, For the first The value of the core anomaly feature at time t. The standard threshold for the k-th type of core anomaly feature obtained by training with historical normal data. For the first Weights of core anomaly features The number of core abnormal features.

[0035] Furthermore, the tiered response strategy specifically involves: inputting a set of timing monitoring and judgment parameters, and defining the response level and anomaly level. Correspondingly, the response intensity gradually increases with the level of anomaly. Level 1 minor anomalies are continuously monitored and slightly adjusted; Level 2 general anomalies are given early warnings and parameter adjustments; Level 3 severe anomalies are coordinated and optimized; and Level 4 critical anomalies are switched to emergency mode and optimized comprehensively. Based on the response level, the optimized dynamic weights for BeiDou and 5G-A timing are determined using the following formula: in, , These are the optimized dynamic weights for BeiDou and 5G-A timing synchronization under the L-level response. Initial weights for BeiDou timing. This is the weighting adjustment coefficient.

[0036] Furthermore, when the anomaly level L=1 (minor anomaly) and the dual-source consistency score S (consist) ≥0.8 (dual-source time synchronization data is normal) and the maximum deviation ΔT of wide-area synchronization max When the time interval is ≤9ns (wide-area multi-node time synchronization is normal), a Level 1 response is triggered; when the anomaly level L=2 (general anomaly) and the dual-source consistency score is 0.7≤S (consist) <0.8 (a slight deviation exists in the dual-source timing data) and the maximum deviation of wide-area synchronization is 9ns <ΔT max A level 2 response is triggered when the time difference is ≤12ns (small synchronization deviation exists in wide-area multi-node time synchronization); a level 2 response is triggered when the anomaly level L=3 (severe anomaly) and the dual-source consistency score is 0.6≤S. (consist) <0.7 (significant deviation exists in dual-source timing data) and the maximum deviation of wide-area synchronization is 12ns <ΔT max A Level 3 response is triggered when the time interval is ≤15ns (significant synchronization deviation exists in wide-area multi-node time synchronization); a Level 3 response is triggered when the anomaly level L=4 (crisis anomaly) and the dual-source consistency score S... (consist) <0.6 (the dual-source timing data has a large deviation) and the maximum deviation of wide-area synchronization ΔT max A level 4 response is triggered when the time exceeds 15ns (significant synchronization deviation exists in wide-area multi-node time synchronization).

[0037] Furthermore, the core operations of the response strategies at each level are as follows: Level 1 Minor Anomaly: Maintain the existing dual-source timing mode, increase the anomaly monitoring frequency from 100Hz to 200Hz, and only fine-tune the attention weights of the Transformer layer; no need to activate deep dual-source collaboration. Level 2 General Anomaly: Push ordinary early warning information, activate mild dual-source collaboration, fine-tune the basic parameters of each layer of the hybrid AI model and the dual-source timing weights, and correct minor wide-area synchronization deviations. Level 3 Severe Anomaly: Push emergency early warning information, activate deep dual-source collaboration, optimize the core parameters of the hybrid AI model, the dynamic weights of dual-source timing, and the DCV calibration frequency, and intercept interference signals. Level 4 Critical Anomaly: Push special early warning information, activate seamless switching between BeiDou and 5G-A timing sources, comprehensively optimize all models and timing parameters, activate the backup timing module, and ensure uninterrupted timing.

[0038] Furthermore, the supervised optimization model specifically adopts an architecture of reinforcement learning agent and dual optimization branches. The reinforcement learning agent performs global optimization decisions and payoff calculations, while the dual optimization branches are a hybrid AI model parameter optimization branch and a hierarchical response strategy threshold optimization branch. Data-driven dynamic optimization is achieved based on the traceability database. The reinforcement learning module is constructed using the DQN algorithm, aiming to optimize the overall performance of the time synchronization system. It includes an input layer, a hidden layer, and an output layer. The input layer receives the state. The output layer outputs the optimal action. , , These are optimization instructions for dual optimization branches; The reinforcement learning reward function is: in, for Reward value at any moment, state , for The core dataset after normalization in the database is traced back to its source at all times. for The accuracy of the time synchronization system for Quantitative value of the degree of anomaly at any given moment for Time-based dual-source consistency score, for Maximum deviation of wide-area synchronization at any given time , , , These are the weights of the corresponding parameters.

[0039] Furthermore, there are two optimization branches, specifically: a hybrid AI model parameter optimization branch receives optimization instructions. Extract the parameter adjustment amounts of each layer in the instruction, and dynamically adjust all parameters of the hybrid AI model; after optimization, apply the hybrid AI model to real-time timing data processing, calculate the optimized anomaly degree quantification value, compare it with the anomaly degree before optimization, record the optimization difference, and record the optimized model parameters, optimized anomaly degree quantification value, and optimization difference into the traceability database; Hierarchical response strategy threshold optimization branch receiving optimization instructions Extract the threshold adjustment amount of each level of response in the instruction, and optimize the triggering conditions of the graded response and the threshold for the classification of early warning levels; after optimization, record the new threshold parameters and optimization basis, and update the full-process related data table of the traceability database synchronously.

[0040] Furthermore, such as Figure 3The diagram shows the architecture of the CNN-LSTM-Transformer hybrid AI model. The structural parameters of the CNN-LSTM-Transformer hybrid AI model are as follows: The CNN module is a local feature extraction module, which adopts a single-layer convolutional architecture without multiple convolutional layers. The overall process is input layer → convolutional operation layer → activation layer → output layer, and the output is directly connected to the LSTM module. Its input layer receives a pre-processed high-fidelity temporal feature sequence. The convolutional operation layer uses a 3×3 convolutional kernel, a stride of 1, and a same padding method to perform two-dimensional convolutional operations. The activation layer uses the Sigmoid function. The trainable parameters include the initial random initialization, the convolutional kernel weight matrix dynamically optimized through reinforcement learning, and the bias terms initialized and optimized synchronously. The output layer outputs local feature vectors. The output dimension is consistent with the input data dimension. Each dimension corresponds to a local anomaly feature in the timing. In the optimization of the correlation parameters, the parameter adjustment coefficient is fixed at 0.1, and the weight adjustment amount and the bias adjustment amount are both between [-0.01, 0.01]. The LSTM module adopts a single-layer LSTM architecture. The overall process is as follows: input layer → LSTM core layer (containing a three-gate structure) → activation layer → output layer. The input is the local feature vector output by the CNN module, and the output is connected to the Transformer module. Its input layer receives the concatenation data of the local feature vector output by the CNN module at the current time step and the hidden state of the LSTM module at the previous time step. The LSTM core layer includes a forget gate, an input gate, and an output gate. The dimension of the hidden layer is consistent with the dimension of the local feature vector output by the CNN module. The activation layer uses the hyperbolic tangent function with a value range of (-1, 1). The trainable parameters include the weight matrix that is initially randomly initialized and subsequently dynamically optimized through reinforcement learning, as well as the bias term that is initialized and optimized synchronously. The output layer outputs the temporal feature vector at the current time step. The output dimension is consistent with the input concatenation dimension and the output dimension of the CNN module. In the optimization of the associated parameters, the parameter adjustment coefficient is fixed at 0.1, and the weight adjustment amount and the bias adjustment amount are both between [-0.01, 0.01]. The Transformer module employs a single-layer self-attention architecture with fully connected layers and anomaly quantization branches. The overall flow is: input layer → self-attention optimization layer → fully connected layer (two layers in series) → anomaly quantization output layer. The input is the temporal feature vector output by the LSTM module. Its input layer receives the temporal feature vector output by the LSTM module. The self-attention optimization layer has 4 self-attention heads, and the hidden layer dimension is consistent with the temporal feature vector output by the LSTM module. The input temporal feature vector is used as the query, key, and value simultaneously to calculate attention weights and map them to the [0,1] interval through a normalization function. The fully connected layer consists of two layers. The structure is cascaded. The first layer contains a weight matrix and a bias term, paired with a Sigmoid activation function (value range (0,1)). The second layer contains a weight matrix and a bias term. The weight matrices and bias terms of both layers are initially randomly initialized and subsequently dynamically optimized through reinforcement learning. The anomaly quantization output layer outputs an anomaly quantification value with a value range of [0,1]. It focuses on three types of time-series features: phase error feature weight 0.4, timing jitter and signal strength feature weights 0.3 each. In the optimized correlation parameters, the parameter adjustment coefficient is fixed at 0.1, and all weights and bias adjustments are between [-0.01, 0.01].

[0041] Furthermore, such as Figure 4 The diagram shows the architecture of the supervised optimization model. The supervised optimization model adopts a single-level integrated architecture with a reinforcement learning main body and two optimization branches. The overall process is as follows: input layer → reinforcement learning main body layer → two optimization branch layer → iterative update layer → output layer. The input is the normalized core dataset in the source database, and the output is the directly executable optimization instructions. The output connects to the corresponding modules of the time synchronization system. Its input layer receives historical and real-time data read in real time from the source database, and the input dimension is adapted to the input dimension of the reinforcement learning main body layer. The reinforcement learning agent is constructed using the DQN algorithm, which includes an input layer, a hidden layer, and an output layer. The learning rate is set to 0.001, the iteration batch is set to 32, and the initial network parameters of the agent are randomly initialized. The core components include three core parameters: state, action, and reward. The initial state value is the historical dataset of the most recent hour in the source database, the initial action adjustment is set to 0, and the initial reward weights are summed to 1. The dual-optimization branch layer includes a hybrid AI model parameter optimization branch and a hierarchical response strategy threshold optimization branch, which work in parallel and in synergy. The parameter adjustment coefficient is fixed at 0.1, and all parameter adjustments are within the range of [-0.01, 0.01]. The iterative update layer uses a temporal difference learning algorithm with a fixed learning rate of 0.001 and a fixed discount factor of 0.9. The convergence criterion is that the reward value fluctuation range is ≤0.05 for 100 consecutive iterations, and the iteration frequency is one incremental iteration every 24 hours. The output layer outputs the optimal optimization instruction set, including the optimal parameters of the hybrid AI model and the optimal threshold of the hierarchical response strategy. The output parameters are synchronously stored in the traceability database. All trainable parameters are initially randomly initialized and then dynamically optimized through model iteration.

[0042] This embodiment details a dual-source timing intelligent supervision and optimization method based on BeiDou and 5G IoT. It collects BeiDou timing data, 5G-A IoT timing data, and environmental auxiliary data, and after preprocessing, constructs a high-fidelity time-series feature sequence. Based on a CNN-LSTM-Transformer hybrid AI model, combined with distributed satellite common-view (DCV) technology, it extracts anomaly quantification values ​​and dual-source consistency scores, integrating them to obtain a timing supervision and judgment parameter set. Based on this parameter set, it dynamically adjusts the BeiDou and 5G-A timing weights through a hierarchical response strategy. It constructs a traceability database to store early warning information, optimization parameters, and full-process data. Through a supervision and optimization model composed of a reinforcement learning subject and dual optimization branches, it outputs optimization instructions from the hybrid AI model and the hierarchical response strategy.

[0043] Example 2, based on Example 1, details the experimental process of using this method for intelligent supervision of dual-source time synchronization in a wide-area multi-node time synchronization system of a power dispatch center, as follows: This wide-area multi-node timing system primarily serves core services such as power equipment status monitoring, power grid dispatch command synchronization, and fault location and tracing. It has stringent requirements for timing accuracy, operational stability, and timely response to anomalies, needing to achieve nanosecond-level timing accuracy. However, in this scenario, BeiDou timing is easily obstructed by substation buildings, leading to a sharp drop in timing accuracy; and the latency fluctuations of 5G networks during peak periods affect timing synchronization; in addition, extreme environments (high temperature, electromagnetic interference) can easily cause timing anomalies, and these anomalies cannot be accurately identified. Therefore, this method is needed to achieve dual-source fusion of BeiDou and 5G IoT, AI intelligent detection, reinforcement learning optimization, and hierarchical response.

[0044] One hundred substation terminals of the system were selected, and hardware equipment covering timing, monitoring, computing, communication and terminals was deployed. Each substation terminal was equipped with one BD-2 dual-mode Beidou timing receiver, one 5G-A IoT timing module, one set of temperature and humidity sensors, and one electromagnetic interference detector to collect auxiliary data on operating temperature, ambient humidity and electromagnetic interference intensity. Based on the hardware environment, a software system adapted to the technical solution of this invention is built, including a CNN-LSTM-Transformer hybrid AI model, a reinforcement learning model, and a database using a hybrid storage mode of InfluxDB2.0 time series database combined with MySQL8.0 relational database.

[0045] The experiment was conducted over a period of 72 hours, with a data acquisition frequency of 1 time per second. It covered four typical operating conditions: normal operation, weak BeiDou signal, 5G network fluctuation, and extreme environment. This comprehensively simulated the actual operation of a power dispatch center, collecting three types of core data. After preprocessing, the effective data volume reached 761,280 records, including: timestamps, 1PPS signal, phase error, number of satellite locks, and signal strength for BeiDou timing data; NTP / NITZ timestamps, time offset, network latency, packet loss rate, and signal-to-noise ratio for 5G-A IoT timing data; and operating temperature, electromagnetic interference intensity, ambient humidity, and equipment power consumption for environmental auxiliary data.

[0046] Based on collected historical data, a hybrid AI model of CNN-LSTM-Transformer was constructed using the TensorFlow 2.10 framework. The CNN layer uses 3×3 convolutional kernels with a stride of 1 and same padding, and has 32 kernels. The LSTM layer is based on the local feature vectors output by the CNN layer, and is constructed as a 2-layer LSTM network with a hidden layer dimension of 64 and a dropout coefficient of 0.2. The Transformer layer adopts a 4-head self-attention mechanism. 83.3% (618,240 data points) of the 72-hour experimental data was used as the training set, and 16.7% (143,040 data points) was used as the test set. The number of iterations was set to 500, the learning rate to 0.001, and the batch size to 64. The cross-entropy loss function was used. Figure 5 The CNN-LSTM-Transformer hybrid AI model iteration data shown shows that after training, the model converged, the loss value stabilized below 0.02, and the test set fit R² ≥ 0.98.

[0047] Two existing technologies from the background of this application were selected for simultaneous comparative experiments, including the single Beidou high-precision time synchronization monitoring method BHTM and the Beidou-TSN hybrid time synchronization method BTSM. Among them, BHTM only deploys BD-2 type Beidou time synchronization receivers at the terminals of each substation, without deploying 5G-A IoT modules or environmental monitoring modules. The server only installs basic data acquisition software and does not need to build AI models, supervised optimization models, or traceability databases. BHTM obtains high-precision time synchronization signals through Beidou time synchronization receivers and extracts timestamps and 1PPS signals as time references for power equipment status monitoring. BTSM deploys BD-2 Beidou timing receivers at the terminals of each substation, along with TSN time-sensitive network switches to replace 5G-A IoT modules. The servers are equipped with basic data acquisition and transmission optimization software, eliminating the need to build AI models, supervised optimization models, and traceability databases. BTSM acquires basic timing signals through Beidou timing receivers and uses TSN time-sensitive network switches to optimize latency and jitter issues during timing signal transmission, thereby improving the synchronization of timing signal transmission.

[0048] During the implementation of this method, the Beidou timing module, 5G-A IoT module, environmental monitoring module and terminal intelligent monitoring equipment are activated. Based on 5G-A IoT, low-latency data transmission from 100 substation terminals is achieved. The collected data is pushed to the server cluster in real time and stored in a temporary cache to avoid data loss. The 3σ criterion is used to filter the data in the temporary buffer, removing the following invalid and interfering data: Data with a signal strength of less than 10dB and fewer than 4 satellites locked to the BeiDou timing data were excluded. 5G-A timing data is filtered to remove data with a packet loss rate >5% and a signal-to-noise ratio <20dB. Environmental auxiliary data is used to filter out data that is outside the sensor's measurement range; Using the BeiDou 1PPS signal as the time reference, the timestamps of 5G-A timing data and environmental auxiliary data are uniformly aligned to eliminate data time deviations between different terminals and modules; the format of all data is unified into JSON format, and the data fields, units and precision are clearly defined, with timing deviation and phase error retained to 6 decimal places (nanosecond level), and network latency and packet loss rate retained to 2 decimal places. Wavelet transform algorithm is used to enhance the features of preprocessed data, strengthen core abnormal features such as timing deviation, phase error and network delay, and weaken random interference signals; By performing feature-weighted fusion of BeiDou timing data, 5G-A timing data, and environmental auxiliary data, a high-fidelity time-series feature sequence is constructed. ,in, (Weight of BeiDou timing data) (5G-A timing data weight) (Environmental auxiliary data weights), the weights can be dynamically adjusted according to operating conditions; Using a trained CNN-LSTM-Transformer hybrid AI model, high-fidelity temporal feature sequences are extracted hierarchically from the preprocessed dataset. The core features of these sequences are then extracted at the CNN layer using a formula. Capture short-term abnormal fluctuations in timing data; use formulas in the LSTM layer. Capture the long-term temporal dependencies and changing trends of timing data; output anomaly feature vectors through a fully connected layer in the Transformer layer, the formula for which is: .

[0049] Based on distributed satellite common-view (DCV) technology, 100 geographically dispersed substation terminals are activated, allowing all terminals to simultaneously observe the timing signal of the same BeiDou satellite (selected as a BeiDou GEO satellite, satellite number BEIDOU-3G01). The signal transmission time difference between each terminal and the satellite is calculated, indirectly obtaining the time difference between each terminal. Set nanosecond-level timing threshold ,according to The requirement is to accurately calibrate the timing data after the initial processing of the hybrid AI model, establish a unified wide-area synchronization benchmark, and eliminate timing deviations between terminals.

[0050] Selecting the BeiDou time stamp sequence after DCV calibration and 5G-A timing timestamp sequence The continuous sample data (N=1000) are used to calculate the dual-source consistency score using the formula described in this invention. Quantify the matching degree of dual-source time synchronization data; set a score threshold. ,when If this occurs, it is determined to be a dual-source consistency anomaly, and a tiered response needs to be initiated.

[0051] Based on the anomaly severity quantification value ε output by the hybrid AI model, the anomaly level L is determined. The anomaly level classification criteria strictly follow the settings of this invention: L=1 (minor anomaly, ε<0.1), L=2 (moderate anomaly, 0.1≤ε<0.3), L=3 (severe anomaly, 0.3≤ε<0.5), L=4 (critical anomaly, ε≥0.5); integrating the anomaly level L and the dual-source consistency score. The system identifies anomaly types (signal anomalies, network anomalies, environmental interference) and wide-area synchronization deviation values, constructs a complete set of timing supervision and judgment parameters, and pushes them to the hierarchical response module in real time to provide accurate basis for the execution of subsequent hierarchical response strategies.

[0052] Based on the time synchronization monitoring and judgment parameter set, four response levels are set, with the response intensity gradually increasing as the anomaly level rises. Level 1 is a minor anomaly (L=1, ...). ≥0.8, ≤9ns) triggers a minor anomaly response without requiring deep collaborative operation; continuously monitors timing data, increases the anomaly monitoring frequency from 100Hz to 200Hz, fine-tunes the attention weight of the Transformer layer, strengthens the core feature recognition capability, and ensures that anomalies do not escalate; Level 2 general abnormality (L=2, 0.7≤ <0.8, 9ns< ≤12ns) triggers a general anomaly response, pushes ordinary early warning information to the dispatch center monitoring platform; initiates mild dual-source collaboration, fine-tunes the basic parameters of each layer of the hybrid AI model and the weight of dual-source timing, compensates for timing deviation, and avoids anomaly deterioration; Grade 3 severe abnormality (L=3, 0.6≤ <0.7, 12ns< ≤15ns) triggers a severe anomaly response, pushes emergency warning information to the dispatch center and relevant maintenance personnel; initiates deep dual-source collaboration, optimizes the core parameters of the hybrid AI model, the dynamic weight of dual-source timing and the DCV calibration frequency (increased from 1 time / minute to 1 time / 10 seconds), and initiates the electromagnetic interference interception module to reduce the impact of environmental interference; Level 4 Crisis Anomaly (L=4, <0.6, >15ns) triggers a crisis anomaly response, pushes a special warning message, and activates an audible and visual alarm; initiates seamless switching between BeiDou and 5G-A timing sources (if BeiDou signal is abnormal, switch to 5G-A as the main source; if 5G network is abnormal, switch to BeiDou as the main source), comprehensively optimizes all models and timing parameters, activates the backup timing module to ensure uninterrupted timing, and notifies maintenance personnel to intervene urgently.

[0053] Based on the response level L, the timing weights of BeiDou and 5G-A are dynamically adjusted to achieve dual-source wide-area collaborative optimization, ensuring stable timing accuracy. For example, in the case of a Level 3 severe anomaly (L=3), =0.65, When =0.7), Appropriately increase the timing weight of 5G-A to compensate for insufficient BeiDou signal.

[0054] Three core related data tables are constructed: an optimization parameter table, an early warning information table, and a full-process related data table. These three tables share a timestamp index to enable linked data queries. At the same time, a triple query index is constructed based on timestamp, response level, and anomaly type to improve data query and traceability efficiency, ensuring that time synchronization data and anomaly handling details can be quickly traced for any time period and any terminal.

[0055] Based on the PyTorch 1.13 framework, a supervised optimization model with a reinforcement learning subject and dual optimization branches is constructed. The reinforcement learning subject is constructed using the DQN (Deep Q-Network) algorithm, aiming at the optimal overall performance of the time synchronization system, and includes a three-layer structure of input layer, hidden layer, and output layer. The dual optimization branches are a hybrid AI model parameter optimization branch and a hierarchical response policy threshold optimization branch, respectively. The two branches work in parallel to achieve collaborative optimization.

[0056] In reinforcement learning, the input layer receives the state, and the reward value at time t is calculated using a reward function. A higher reward value indicates better system performance. The reward function formula is as follows: ;in, The parameters are weighted to ensure balanced optimization of various performance metrics; the output layer outputs the optimal action. .

[0057] Hybrid AI model parameter optimization branch with dual optimization branches The instructions extract the parameter adjustment amounts for each layer (adjustment range [-0.01, 0.01]) and dynamically adjust all parameters of the CNN-LSTM-Transformer model, including convolutional kernel weights, LSTM hidden layer weights, and Transformer self-attention weights. The optimized anomaly quantification value and optimization difference are recorded in the traceability database as the basis for the next iteration.

[0058] Hierarchical response strategy threshold optimization branch reception The instructions extract the threshold adjustment amounts for each level of response, optimize the anomaly level classification threshold, the dual-source consistency score threshold, and the wide-area synchronization deviation threshold, and update the tiered response triggering conditions; synchronize the new threshold parameters and optimization criteria to the traceability database to ensure the adaptability of the tiered response strategy.

[0059] The model was iterated using a temporal difference learning algorithm with a learning rate of 0.001 and a discount factor of 0.9. The convergence criterion was that the reward value fluctuation range was ≤0.05 for 100 consecutive iterations. The iteration frequency was one incremental iteration every 24 hours, and each iteration only used the newly added full-process data of the day to avoid repeated training and improve iteration efficiency. After multiple iterations, the model reached a stable state, and the performance of the timing system was continuously optimized.

[0060] The timing accuracy of the three methods was tested using 12 hours of test data. The results were compared under four typical operating conditions: normal operation, weak BeiDou signal operation, 5G network fluctuation operation, and extreme environment operation. Table 1 below shows the comparison of timing accuracy data. Table 1 Comparison of Time Synchronization Accuracy Data According to Table 1 and Figure 6 As shown in the comparison chart of timing deviations of the various methods under four operating conditions, the average timing deviation and maximum timing deviation of the present invention are significantly better than those of the BHTM and BTSM methods under all four operating conditions. The accuracy improvement compared to BTSM is more than 70%. The advantages are even more obvious under complex operating conditions such as weak BeiDou signals and extreme environments. It can stably achieve nanosecond-level timing accuracy, which fully meets the stringent requirements of power dispatching.

[0061] The most common anomalies under four operating conditions were selected to test the anomaly detection capabilities of the three methods. The normal operating condition was intermittent equipment clock drift (mild), the BeiDou weak signal operating condition was excessive timing deviation (severe), the 5G network fluctuation operating condition was synchronization deviation caused by latency (moderate), and the extreme environment operating condition was multiple factors superimposed anomalies (severe). The anomaly detection capability comparison data is shown in Table 2 below: Table 2 Comparison of Anomaly Detection Capabilities According to Table 2 and Figure 7 As shown in the comparison chart of anomaly detection capabilities of the various methods, among the three methods, the hybrid AI model of this invention has the best anomaly detection accuracy, ≥98.5% under all four operating conditions. Moreover, the timing deviation can be quickly restored to a stable nanosecond range after graded response, achieving a breakthrough in accuracy and efficiency compared to the comparison methods. The BHTM detection accuracy is 78.5%~83.2%. Due to the limitation of a single timing source, the detection stability is slightly insufficient under complex operating conditions, and there is still room for optimization of the timing deviation after processing. The BTSM detection accuracy is 82.3%~88.6%. It has a good effect on the identification of transmission-related anomalies, but the detection accuracy is limited when multiple factors are superimposed. The timing deviation after processing cannot meet the stringent nanosecond requirements of power dispatching.

[0062] The comprehensive performance indicators of each method were obtained by statistically analyzing the data over the entire 72-hour experimental period, as shown in Table 3 below: Table 3 Comprehensive Performance Indicators According to Table 3 and Figure 8 As shown in the comprehensive performance comparison chart, this method outperforms the comparison methods in core indicators such as average timing deviation (±5.5ns) and anomaly detection accuracy (≥98.5%). It also has automatic response (≤60s) and data traceability capabilities, and the lowest daily anomaly occurrence rate (0.83 times / 100 nodes). All data meet the engineering compliance threshold. In contrast, BHTM and BTSM lack dynamic optimization and traceability functions, and their timing accuracy and anomaly handling capabilities are limited, fully demonstrating the technical advantages and practicality of this method.

[0063] This embodiment details the experimental process of using this method for intelligent supervision of dual-source time synchronization in a wide-area multi-node time synchronization system of a power dispatch center. It compares the method with two existing methods, BHTM and BTSM, and verifies that it can maintain stable nanosecond-level time synchronization accuracy under four typical operating conditions, with an anomaly detection accuracy rate of ≥98.5%. It also demonstrates dynamic iterative optimization, wide-area collaboration, automatic hierarchical response, and full-process traceability capabilities.

[0064] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A dual-source timing intelligent supervision and optimization method based on BeiDou and 5G IoT, characterized in that, include: Collect dual-source timing data from BeiDou and 5G-A IoT, combine it with environmental auxiliary data, and construct a high-fidelity time series feature sequence after preprocessing; Based on the preprocessed data, a wide-area multi-node time synchronization calibration is achieved by using a CNN-LSTM-Transformer hybrid AI model combined with DCV technology, and time synchronization supervision and judgment parameters are output. Based on the timing supervision and judgment parameters, through a hierarchical response strategy, optimized parameters are output, dual-source wide-area collaboration is carried out, and early warning information is recorded and pushed. Based on optimized parameters and early warning information, a source tracing database is constructed; Based on the source database, the supervised optimization model outputs optimization instructions for the CNN-LSTM-Transformer hybrid AI model and the hierarchical response strategy. The supervised optimization model is constructed based on the reinforcement learning subject and dual optimization branches.

2. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things as described in claim 1, characterized in that, The high-fidelity timing feature sequence specifically includes: BeiDou timing data including at least a timestamp, 1PPS signal, and phase error; GA IoT timing data including at least an NTP / NITZ timestamp and time offset; and environmental auxiliary data including at least operating temperature and electromagnetic interference intensity. The three types of collected data undergo multi-stage preprocessing to remove invalid interference, unify data format, and enhance features. Then, through feature weighted fusion, the preprocessed three types of data are collaboratively integrated to obtain a high-fidelity time-series feature sequence. .

3. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things as described in claim 1, characterized in that, The timing supervision and judgment parameters are specifically: performing hierarchical feature extraction on the input high-fidelity time-series feature sequence and outputting an anomaly degree quantification value. ,based on Determine the level of abnormality ; By using distributed satellite common-view DCV technology and combining time synchronization data collected from multiple nodes in a wide area, the time difference between nodes is calculated to establish a unified wide-area synchronization benchmark, and the time synchronization data after preliminary processing by the hybrid AI model is precisely calibrated. Calculate the dual-source consistency score based on the DCV-calibrated BeiDou and 5G-A dual-source timing data. Quantify the matching degree of dual-source time synchronization data; Integrating dual-source consistency scores Abnormal level By combining the anomaly type and the wide-area synchronization deviation value, a set of timing supervision and judgment parameters is obtained.

4. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things according to claim 3, characterized in that, The CNN-LSTM-Transformer hybrid AI model specifically involves the following: In hierarchical feature extraction, the CNN captures short-term abnormal fluctuations in the timing data, as shown in the formula: in, This refers to the local feature vector output by the CNN layer. Here, is the Sigmoid activation function, and is the convolutional kernel weight matrix of the CNN layer. For the bias term of the CNN layer, This is a two-dimensional convolution operation; LSTM, based on the local features output by CNN, focuses on the long-term changes in timing data, capturing the long-range temporal dependence of timing data and capturing the temporal trend. The formula is as follows: in, for The temporal feature vector output by the LSTM layer at time step [time]. The hyperbolic tangent activation function is used. Here is the weight matrix of the LSTM layer. For the bias term of the LSTM layer, For feature splicing operations; Transformer uses self-attention to apply temporal features to the LSTM output. Further optimization was performed, strengthening the focus on core features, resulting in... The original temporal feature vector output by the LSTM attention layer at time step After self-attention optimization, the optimized features are processed through the fully connected layer of the model to output anomaly feature vectors, as shown in the formula: in, , , These are the weight matrices for the first and second fully connected layers, respectively. , These are the bias terms for the first and second fully connected layers, respectively. This represents the anomalous feature vector output by the fully connected layer. Based on the output anomaly quantization value Calculate the quantification value of the degree of abnormality. The formula is: in, For the first Core anomaly features in The value at time, The standard threshold for the k-th type of core anomaly feature obtained by training with historical normal data. For the first Weights of core anomaly features The number of core abnormal features.

5. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things according to claim 3, characterized in that, The precise calibration of the timing data after initial processing by the hybrid AI model involves: using distributed satellite common-view (DCV) technology, based on multiple geographically dispersed nodes, simultaneously observing the timing signal of the same BeiDou satellite, calculating the signal transmission time difference between each node and the satellite, and indirectly obtaining the time difference between each node. Setting a time difference threshold based on nanosecond-level timing monitoring requirements ,according to Calibration is required as per the specifications.

6. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things according to claim 3, characterized in that, Specifically, this involves selecting the BeiDou time stamp sequence after DCV calibration. and 5G-A timing timestamp sequence Using a certain number of continuous sample data, the consistency score of dual-source time synchronization is calculated using the following formula: in, For the dual-source consistency score, For the number of data samples, , They are respectively The timestamp for BeiDou and 5G-A timing after DCV calibration; when Less than the preset score threshold When this occurs, it is determined to be an anomaly in the consistency between two sources.

7. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things according to claim 1, characterized in that, The hierarchical response strategy specifically includes: inputting a set of timing monitoring and judgment parameters, response level, and anomaly level. Correspondingly, the response intensity gradually increases with the level of anomaly. Level 1 minor anomalies are continuously monitored and slightly adjusted; Level 2 general anomalies are given early warnings and parameter adjustments; Level 3 severe anomalies are coordinated and optimized; and Level 4 critical anomalies are switched to emergency mode and optimized comprehensively. Based on the response level, the optimized dynamic weights for BeiDou and 5G-A timing are determined using the following formula: in, , These are the optimized dynamic weights for BeiDou and 5G-A timing synchronization under the L-level response. Initial weights for BeiDou timing. This is the weighting adjustment coefficient.

8. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things according to claim 1, characterized in that, The traceability database is specifically composed of: data based on early warning information data, optimized parameter data, and full-process correlation data. After data regularization by unifying timestamps, removing invalid and duplicate data, and correcting data deviations, the data is stored in association through three tables: the optimized parameter table, the early warning information table, and the full-process correlation data table, sharing a timestamp index. Data storage adopts a hybrid storage mode combining time-series databases and relational databases; A query mechanism is built based on a triple index of timestamp, response level, and exception type.

9. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things according to claim 1, characterized in that, The supervised optimization model specifically adopts an architecture of reinforcement learning subject and dual optimization branches. The reinforcement learning subject performs global optimization decisions and payoff calculations, and the dual optimization branches are a hybrid AI model parameter optimization branch and a hierarchical response strategy threshold optimization branch. Data-driven dynamic optimization is achieved based on the traceability database. The reinforcement learning module is constructed using the DQN algorithm, aiming to optimize the overall performance of the time synchronization system. It includes an input layer, a hidden layer, and an output layer. The input layer receives the state. The output layer outputs the optimal action. , , These are optimization instructions for dual optimization branches; The reinforcement learning reward function is: in, for Reward value at any moment, state , for The core dataset after normalization in the database is traced back to its source at all times. for The accuracy of the time synchronization system for Quantitative value of the degree of anomaly at any given moment for Time-based dual-source consistency score, for Maximum deviation of wide-area synchronization at any given time , , , The weights are for the corresponding parameters.

10. The intelligent supervision and optimization method based on dual-source timing of BeiDou and 5G Internet of Things according to claim 9, characterized in that, The dual optimization branches are specifically: the hybrid AI model parameter optimization branch receives optimization instructions. Extract the parameter adjustment amounts of each layer in the instruction, and dynamically adjust all parameters of the hybrid AI model; after optimization, apply the hybrid AI model to real-time timing data processing, calculate the optimized anomaly degree quantification value, compare it with the anomaly degree before optimization, record the optimization difference, and record the optimized model parameters, optimized anomaly degree quantification value, and optimization difference into the traceability database; Hierarchical response strategy threshold optimization branch receiving optimization instructions Extract the threshold adjustment amount of each level of response in the instruction, and optimize the triggering conditions of the graded response and the threshold for the classification of early warning levels; after optimization, record the new threshold parameters and optimization basis, and update the full-process related data table of the traceability database synchronously.

Images