A method and system for predicting clock errors in traceability and comparison timekeeping systems

Abstract

This application discloses a clock difference prediction method and system for a timekeeping system for source tracing and comparison. The method includes: determining whether the acquisition of a superior common-view file is successful, wherein the superior common-view file is used in the timekeeping system for source tracing and comparison; if the acquisition of the superior common-view file fails, predicting clock difference data to obtain predicted clock difference data, wherein the clock difference data is predicted using a machine learning model trained on real clock difference data, and the trained machine learning model is used to output predicted clock difference data based on real clock difference data; and submitting the predicted clock difference data to the timekeeping system for source tracing and comparison. This application solves the problem that, during common-view tracing, the failure to acquire the common-view file of the primary timekeeping agency leads to the inability to trace the local time base in a timely manner, thereby improving the accuracy of clock difference prediction in practical applications.

Description

Technical Field

[0001] This application relates to the field of time and frequency metrology, and more specifically, to a clock error prediction method and system for tracing and comparing timekeeping systems. Background Technology

[0002] With the continuous advancement of social technology, high-precision time references are playing an increasingly important role in navigation and positioning, power transmission, urban transportation, and information communication. To possess a stable and accurate local high-precision time reference, a reliable clock system is a solid foundation and crucial prerequisite. The basic components of a timekeeping system include: an atomic clock subsystem, a traceability and comparison subsystem, a clock difference measurement subsystem, and a control subsystem. The atomic clock subsystem, composed of hydrogen and cesium clocks, is responsible for generating the original time and frequency signals; the clock difference measurement subsystem, composed of multi-channel counters, is responsible for measuring the clock difference between the reference clock and other atomic clocks within the clock group; the traceability and comparison subsystem, composed of one or more common-view receivers, is responsible for acquiring the clock difference between the local time and GNSS time; and the control subsystem, composed of phase micrometers, is responsible for adjusting the frequency and phase of the local time physical signal according to an adjustment strategy based on the acquired clock difference signal, thereby synchronizing the local time with GNSS time.

[0003] For the timekeeping subsystem, this subsystem is based on common-view timekeeping comparison technology. Receiver A, deployed at the primary timekeeping agency, and receiver B, deployed at the user's timekeeping laboratory, simultaneously monitor the time difference with GNSS. Then, based on the data transmission link between stations A and B, the common-view files from the receivers are aggregated at the computer in the timekeeping laboratory. Finally, a weighted average method is used to obtain the time difference value between GNSS_A and GNSS_B. However, malfunctions of receiver A and network disconnections between receivers A and B can prevent the user's timekeeping laboratory from calculating the clock difference data, resulting in the inability to trace the local time reference and ultimately causing the local time signal to deviate. Summary of the Invention

[0004] This application provides a clock difference prediction method and system for tracing and comparing timekeeping systems, so as to at least solve the problem that the local time base cannot be traced in a timely manner due to the failure to obtain the common view file of the first-level timekeeping agency during common view tracing.

[0005] According to one aspect of this application, a clock difference prediction method for a source tracing and comparison timekeeping system is provided, comprising: determining whether the acquisition of a superior common-view file is successful, wherein the superior common-view file is used in the source tracing and comparison timekeeping system; if the acquisition of the superior common-view file fails, predicting clock difference data to obtain predicted clock difference data, wherein the clock difference data is predicted using a machine learning model, the machine learning model is trained using real clock difference data, and the trained machine learning model is used to output predicted clock difference data based on real clock difference data; and submitting the predicted clock difference data to the source tracing and comparison timekeeping system.

[0006] Furthermore, it also includes: if the upper-level shared file is successfully acquired, collecting the clock difference data of the current complete time window, and using the collected clock difference data of the current complete time window to update the weights in the machine learning model.

[0007] Furthermore, the machine learning model has been trained and validated using clock difference data from the previous complete time window.

[0008] Furthermore, before determining whether the acquisition of the superior shared view file is successful, the method further includes: dividing the real clock difference data in the previous complete time window into a training set, a validation set, and a test set; performing multiple rounds of training using the training data, and after each round of training, validating and adjusting the obtained machine learning model using the validation set; and after the completion of the multiple rounds of training, testing and evaluating the trained machine learning model using the test set.

[0009] Furthermore, the real clock difference data is divided according to a 6:2:2 time series length to construct the training set, validation set, and test set.

[0010] Furthermore, the machine learning model is trained through the following steps: initializing and constructing the encoder and decoder structure of the clock sequence prediction model, which consists of an dilated causal convolutional network, a regularization network, a fully convolutional network, and a residual network, and allocating initial parameters such as the sliding window size; and training the constructed initial model using the training set to obtain the machine learning model.

[0011] According to another aspect of this application, a clock difference prediction system for a source tracing and comparison timekeeping system is also provided, comprising: a judgment module for judging whether the acquisition of a superior common-view file is successful, wherein the superior common-view file is used in the source tracing and comparison timekeeping system; a prediction module for predicting clock difference data when the acquisition of the superior common-view file fails, wherein the clock difference data is predicted using a machine learning model, the machine learning model is trained using real clock difference data, and the trained machine learning model is used to output the predicted clock difference data based on the real clock difference data; and a submission module for submitting the predicted clock difference data to the source tracing and comparison timekeeping system.

[0012] Furthermore, it also includes: a training module, used to collect clock difference data of the current complete time window when the upper-level common-view file is successfully acquired, and to update the weights in the machine learning model using the collected clock difference data of the current complete time window.

[0013] Furthermore, the machine learning model has been trained and validated using clock difference data from the previous complete time window.

[0014] Furthermore, the training module is used to: divide the real clock difference data of the previous complete time window into a training set, a validation set, and a test set; perform multiple rounds of training using the training data, and after each round of training, validate and adjust the obtained machine learning model using the validation set; and after the multiple rounds of training are completed, test and evaluate the trained machine learning model using the test set.

[0015] Furthermore, the real clock difference data is divided according to a 6:2:2 time series length to construct the training set, validation set, and test set.

[0016] Furthermore, the machine learning model is trained through the following steps: initializing and constructing the encoder and decoder structure of the clock sequence prediction model, which consists of an dilated causal convolutional network, a regularization network, a fully convolutional network, and a residual network, and allocating initial parameters such as the sliding window size; and training the constructed initial model using the training set to obtain the machine learning model.

[0017] In this embodiment, a method is employed to determine whether the acquisition of the superior common-view file was successful. This superior common-view file is used in the timekeeping system for tracing and comparison. If the acquisition of the superior common-view file fails, clock difference data is predicted to obtain predicted clock difference data. This predicted clock difference data is obtained using a machine learning model trained on real clock difference data. The trained machine learning model then outputs the predicted clock difference data based on the real clock difference data. The predicted clock difference data is then submitted to the timekeeping system for tracing and comparison. This application solves the problem that, during common-view tracing, the failure to acquire the common-view file from the primary timekeeping agency leads to the inability to trace the local time base in a timely manner, thereby improving the accuracy of clock difference prediction in practical applications. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a flowchart of clock sequence prediction using a temporal convolutional network model according to an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of a neural network model architecture according to an embodiment of this application.

[0021] Figure 3 This is a flowchart of a clock error prediction method for a timekeeping system for tracing and comparison, according to an embodiment of this application. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0024] This embodiment provides a clock error prediction method for tracing and comparing timekeeping systems. Figure 3 This is a flowchart of a clock error prediction method for a timekeeping system based on an embodiment of this application. The following describes... Figure 3 The steps involved will be explained.

[0025] Step S302: Determine whether the acquisition of the superior shared view file was successful, wherein the superior shared view file is used for the source tracing and comparison timekeeping system;

[0026] Step S304: If it fails to obtain the upper-level common view file, predict the clock difference data to obtain the predicted clock difference data. The clock difference data is predicted using a machine learning model. The machine learning model is trained using real clock difference data. The trained machine learning model is used to output the predicted clock difference data based on the real clock difference data.

[0027] Step S306: Submit the predicted clock difference data to the source tracing and comparison timekeeping system.

[0028] The above steps solve the problem that the local time base cannot be traced in a timely manner when the common view file of the primary timekeeping agency fails to be obtained during common view tracing, thereby improving the accuracy of clock difference prediction in practical applications.

[0029] In the above steps, if the upper-level shared file is successfully acquired, the clock difference data of the current complete time window is collected, and the weights in the machine learning model are updated using the collected clock difference data of the current complete time window.

[0030] At this point, the machine learning model has been trained and validated using clock difference data from the previous complete time window.

[0031] When training a machine learning model, the real clock difference data in the previous complete time window can be divided into a training set, a validation set, and a test set (for example, the real clock difference data can be divided according to a 6:2:2 time series length to construct the training set, validation set, and test set); multiple rounds of training are performed using the training data, and after each round of training, the obtained machine learning model is validated and adjusted using the validation set; after the multiple rounds of training are completed, the trained machine learning model is tested and evaluated using the test set.

[0032] The following description uses an optional embodiment to illustrate a clock error prediction method for a source tracing and comparison timekeeping system. This method employs a combined offline and online temporal convolutional deep learning model. Based on the complete data of the previous time window divided by the phase micro-jump clock error adjustment strategy, it predicts the clock error of the source tracing and comparison timekeeping system at the next moment online. By determining the acquisition status of the common-view file data of the source tracing and comparison timekeeping system at the next moment, it selectively supplements the missing clock error data of the system or performs online learning on the current online deep model through hedging backpropagation, fine-tuning the weight parameters. Simultaneously, it collects the clock error data of the current complete time window and updates the model weights in the offline state, preparing a pre-trained model for clock error prediction in the next time window.

[0033] In this embodiment, the feature selection and data preprocessing module first obtains clock difference data based on the time window of the current time to be predicted, then removes outliers using the 3Sigma rule, and finally normalizes the data to complete feature selection and data preprocessing. The training, validation, and test dataset construction module divides the preprocessed clock difference data into training, validation, and test sets based on a 6:2:2 time series length ratio.

[0034] The network construction and parameter allocation module first initializes and builds a clock error sequence prediction model consisting of a dilated causal convolutional network, a regularized network, a fully convolutional network, and a residual network, and then performs initial parameter allocation. In this embodiment, the clock error sequence prediction process of the temporal convolutional network model mainly includes the following modules: feature selection and data preprocessing module, training, validation, and test dataset construction module, network construction and parameter allocation module, model training, validation, network adjustment, and evaluation module, and online model import for prediction module. The model training, validation, network adjustment, and evaluation module trains the initial model built by the network construction and parameter allocation module using the training set built by the training, validation, and test dataset construction module. After 5 rounds of training, a validation is performed and the network weight results of the current training are saved once, for a total of 60 rounds. After the training rounds are completed, the prediction accuracy of the network model is calculated using the validation set to find the optimal solution for clock error prediction. It is then determined whether to retrain, fine-tune the network and parameters in the network construction and parameter allocation module, or evaluate the test set. If the evaluation result does not meet the predetermined indicators of the offline model, the feature selection and data preprocessing module and the training, validation, and test dataset construction module need to be re-performed, and clock error data from other time windows need to be added or deleted for preprocessing and dataset construction.

[0035] The online model import and prediction module uses the model training and validation and network adjustment and evaluation modules to evaluate the model and weight parameters that meet the predetermined offline learning indicators. These models then process the real-time clock difference stream in the current source tracing and comparison timekeeping system, outputting predicted clock difference values ​​for the next time step. This process also supplements missing clock difference data in the source tracing and comparison timekeeping system or uses hedging backpropagation to learn the current online deep model online, fine-tuning the weight parameters. Simultaneously, clock difference data for the current complete time window is collected, and the model weights are updated offline to prepare a pre-trained model for clock difference prediction in the next time window.

[0036] Its online deep learning model is designed for real-time clock difference data streams and does not always learn online. When the source tracing and comparison timekeeping system fails to obtain the upper-level common-view file, its output clock difference prediction value will be submitted to the timekeeping system as supplementary data. When the source tracing and comparison timekeeping system successfully obtains the upper-level common-view file, it uses the real clock difference value obtained by the system as the data truth value for online learning, trains the online model through hedging backpropagation, and updates the weight parameters to achieve online fine-tuning of the clock difference data pattern within the same time window.

[0037] The use of clock difference data from a complete time window above the predicted time for training, validation, and testing when pre-training or updating the weights of the offline deep learning model is only one of the embodiments proposed in this solution. It includes, but is not limited to, using clock difference data from other time windows with high similarity in feature selection to adjust or expand the data of the offline model, or pruning some data with low similarity, or using historical prediction data as available data for future rounds of offline model weight updates.

[0038] The clock difference adjustment strategy based on the phase micro-jump meter does not have a fixed time window size for dividing the real-time clock difference data stream of the traceability and timekeeping system. A feasible implementation is the product of the clock difference adjustment period t and the clock difference adjustment level h, so as to fully cover all clock difference periods generated by all adjustment schemes, or set other feasible window sizes.

[0039] The following is combined with Figure 1 and Figure 2The following explanation is provided. The clock sequence prediction process of the temporal convolutional network model in this embodiment mainly includes five modules: feature selection and data preprocessing module, training, validation and test dataset construction module, network construction and parameter allocation module, model training, validation and network adjustment and evaluation module, and online model import for prediction module. The feature selection and data preprocessing module uses clock difference sequence data from the previous 60 days before the current prediction time, combined with a normal distribution, to filter the deviation values ​​of clock difference points, thus completing feature selection and data preprocessing. The training, validation, and test dataset construction module divides the clock difference sequence data from the feature selection and data preprocessing module into training, validation, and test sets based on a 6:2:2 time series length ratio. The network construction and parameter allocation module first initializes the encoder and decoder structure of the clock difference sequence prediction model, consisting of a dilated causal convolutional network, a regularized network, a fully convolutional network, and a residual network, and allocates initial parameters such as the sliding window size. The model training, validation, network adjustment, and evaluation module trains the initial model built by the network construction and parameter allocation module using the training set constructed by the training, validation, and test dataset construction module. Validation is performed every 5 training epochs, and the network weights of the current training are saved. As a result, a total of 60 training rounds were conducted. After each training round, the prediction accuracy of the offline model was calculated using the validation set to find the optimal solution for clock difference prediction. This determined whether to retrain, fine-tune the network and parameters in the network construction and parameter allocation module, or conduct a test set evaluation. If the evaluation results did not meet the predetermined offline learning metrics, the feature selection and data preprocessing module and the training, validation, and test dataset construction module needed to be re-executed. Clock difference data from other time windows were added or removed for preprocessing and dataset construction. The online model import and prediction module used the model training, validation, network adjustment, and evaluation modules to evaluate the model and weight parameters that met the predetermined offline learning metrics after test set evaluation. This processed the real-time clock difference stream in the current source tracing and comparison timekeeping system, outputting the predicted value of the clock difference data for the next time moment. Missing clock difference data in the source tracing and comparison timekeeping system was supplemented, or the current online deep model was trained online through hedging backpropagation, fine-tuning the weight parameters. Simultaneously, clock difference data for the current complete time window was collected, and the model weights were updated in the offline state to prepare a pre-trained model for clock difference prediction in the next time window.

[0040] The feature selection and data preprocessing module comprises three steps: data preparation, feature selection, and data preprocessing. This module is the first step in the clock difference sequence prediction process. The training, validation, and test dataset construction module comprises one step: building the training set, validation set, and test set. This module is the second step in the clock difference sequence prediction process. The network construction and parameter allocation module comprises two steps: building the temporal convolutional network and initializing network parameters. This module is the third step in the clock difference sequence prediction process. The model training, validation, network tuning, and evaluation module comprises two steps: model training and validation, network tuning, and test set evaluation. This module is the fourth step in the clock difference sequence prediction process. The online model import and prediction module comprises importing the model online to predict the clock difference value at time T+1, supplementing the missing clock difference at time T+1 in the timekeeping system through source comparison, fine-tuning the weights at time T+2 through online learning, and based on T... n The offline weight update of the time window data consists of four steps, and the workflow is the fifth step in the clock difference sequence prediction process.

[0041] Figure 1 This is a flowchart of clock error sequence prediction for a temporal convolutional network model according to an embodiment of this application. Figure 1 In the prediction process, after the current time window is detected, an offline dataset for that window is constructed: feature selection and data preprocessing are performed, data is prepared according to appropriate feature conditions, and offline training, validation, and test sets are constructed; simultaneously, offline learning is performed: network parameters are initialized, and model training and validation are performed. Based on whether it is the optimal fit, offline test set evaluation or retraining is conducted. After meeting the preset offline metrics, offline model weights are updated (only when this time window is the initial window, offline model pre-training is completed); subsequently, the offline model weights are imported online, and online prediction and learning begin: predicted values ​​are output, it is determined whether there is missing data in the source tracing and comparison timekeeping system, missing values ​​are supplemented, or online learning is performed, online weights are fine-tuned, and online updates are performed. Finally, it is determined whether to end; otherwise, the clock difference real-time stream is reconnected for online prediction; otherwise, the process ends, and the next time window begins.

[0042] Figure 2 This is a schematic diagram of a neural network model architecture according to an embodiment of this application, such as... Figure 2 As shown, after the input data passes through the input layer, it simultaneously enters a network consisting of a dilated causal convolution and a regularization layer composed of a batch normalization layer and an activation layer. After passing through a one-dimensional fully convolutional network, the feature vectors output from the two branches are combined and enter the fully connected layer to generate the output.

[0043] In another optional implementation, a clock error prediction method for a source tracing and comparison timekeeping system is provided. This method uses a temporal convolutional deep model that combines offline and online learning to specifically supplement the potential missing clock errors at any time in the source tracing and comparison timekeeping system. By offline pre-training and weight updates, it avoids the problems of poor reliability, inaccuracy, or even invalidity of clock error information features obtained by a single batch deep learning model within different clock error adjustment periods of the phase micro-jump meter, thus ensuring the robustness of the clock error prediction line. By online learning, it accurately simulates the changing patterns of clock error data within the same time window, thereby improving the accuracy of clock error prediction in practical applications.

[0044] Its online deep learning model is designed for real-time clock difference data streams. When the source tracing and comparison timekeeping system fails to obtain the superior common-view file, its output clock difference prediction value will be submitted to the timekeeping system as supplementary data. When the source tracing and comparison timekeeping system successfully obtains the superior common-view file, online learning and weight fine-tuning will be performed. Its offline deep learning model uses one or more clock difference data in appropriate time windows after feature selection, or historical prediction data, for pre-training and weight updates. There are multiple feasible implementations for its data volume and type composition.

[0045] In this embodiment, a time window is set for tracing and comparing the real-time clock difference data stream of the timekeeping system based on the clock difference compensation strategy of the phase micro-jump meter. The offline model is updated through different time windows, and there are several feasible embodiments for the specific division of the window size.

[0046] The network composition in this embodiment can take many forms. For example, a temporal convolutional network composed of dilated causal convolutional networks, regularized networks, fully convolutional networks, and residual networks has suitable prediction efficiency and lightweight nature, and is suitable for building clock error prediction models that combine offline and online learning.

[0047] Based on the network in this embodiment, the following steps can be taken in this embodiment: (1) Let the current time be T and the time window in which the current time is located be Tn. Obtain the complete clock difference data of the previous time window Tn-1 of the current time to be predicted T+1, and perform outlier removal and normalization on the clock difference data by the 3Sigma rule to complete the data preprocessing;

[0048] (2) Use the preprocessed target time series to construct the training set, validation set and test set of the offline deep learning model;

[0049] (3) Construct a clock error sequence prediction model consisting of a dilated causal convolutional network, a regularized network, a fully convolutional network, and a residual network, and allocate the initial parameters of the model.

[0050] (4) Use the training set constructed in step (2) to train the clock sequence prediction model built in step (3). After training, calculate the prediction accuracy of the network model through the validation set and determine whether the prediction value meets the predetermined offline learning index. If it does not meet the index, fine-tune the network and parameters of the model in step (3) until the prediction result meets the predetermined index. Then, use this model as the pre-training model for online prediction and online learning.

[0051] (5) Online import step (4) The pre-trained model obtained through offline learning outputs the predicted value of the clock difference data at time T+1, supplements the missing clock difference at time T+1 of the timekeeping system through source comparison, or performs weight fine-tuning at time T+2 through online learning, and performs offline weight update based on the data of time window Tn.

[0052] In this embodiment, an electronic device is provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the methods described in the above embodiments.

[0053] The aforementioned program can run on a processor or be stored in memory (or computer-readable medium). Computer-readable medium includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable medium does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0054] These computer programs may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented using different modules, and different steps can be implemented using different modules.

[0055] This embodiment provides such an apparatus or system. The system, referred to as a clock difference prediction system for a source tracing and comparison timekeeping system, includes: a judgment module for judging whether the acquisition of a superior common-view file is successful, wherein the superior common-view file is used in the source tracing and comparison timekeeping system; a prediction module for predicting clock difference data when the acquisition of the superior common-view file fails, wherein the clock difference data is predicted using a machine learning model trained on real clock difference data, and the trained machine learning model outputs the predicted clock difference data based on the real clock difference data; and a submission module for submitting the predicted clock difference data to the source tracing and comparison timekeeping system.

[0056] The system or apparatus is used to implement the functions of the methods in the above embodiments. Each module in the system or apparatus corresponds to each step in the method, as has been described in the method and will not be repeated here.

[0057] For example, it also includes a training module, used to collect clock difference data for the current complete time window when the parent common-view file is successfully acquired, and to update the weights in the machine learning model using the collected clock difference data for the current complete time window. Optionally, the machine learning model has at least been trained and validated using clock difference data from the previous complete time window.

[0058] For example, the training module is used to: divide the real clock difference data of the previous complete time window into a training set, a validation set, and a test set; perform multiple rounds of training using the training data, and after each round of training, validate and adjust the obtained machine learning model using the validation set; after the multiple rounds of training are completed, test and evaluate the trained machine learning model using the test set. Optionally, the real clock difference data can be divided according to a 6:2:2 time series length to construct the training set, validation set, and test set.

[0059] The above embodiments solve the problem that the local time base cannot be traced in a timely manner when the common view file of the primary timekeeping agency fails to be obtained during common view tracing, thereby improving the accuracy of clock difference prediction in practical applications.

[0060] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A clock error prediction method for a traceability and comparison timekeeping system, characterized in that, include: Determine whether the acquisition of the superior shared view file was successful, wherein the superior shared view file is used for tracing and comparing the timekeeping system; In the event that the acquisition of the superior common view file fails, the clock difference data is predicted to obtain the predicted clock difference data. The clock difference data is predicted using a machine learning model, which is trained using real clock difference data. The trained machine learning model is used to output the predicted clock difference data based on the real clock difference data. The predicted clock difference data is submitted to the source tracing and comparison timekeeping system; If the superior shared file is successfully acquired, the clock difference data of the current complete time window is collected, and the weights in the machine learning model are updated using the collected clock difference data of the current complete time window. The machine learning model includes a network consisting of a regularized layer cascaded with dilated causal convolution, a batch normalization layer, and an activation layer, as well as a one-dimensional fully convolutional network and a fully connected layer. After the input data passes through the input layer, it simultaneously enters the network consisting of a regularized layer cascaded with dilated causal convolution, a batch normalization layer, and an activation layer, and then the one-dimensional fully convolutional network. The feature vectors output from the two branches are combined and then enter the fully connected layer to generate the output.

2. The method according to claim 1, characterized in that, The machine learning model has been trained and validated using clock difference data from at least the previous complete time window.

3. The method according to claim 1 or 2, characterized in that, Before determining whether the acquisition of the shared parent file was successful, the method further includes: The actual clock difference data from the previous complete time window are divided into training set, validation set, and test set; Multiple rounds of training are performed using training data. After each round of training, the resulting machine learning model is validated and adjusted using a validation set. After the multiple rounds of training are completed, the trained machine learning model is tested and evaluated using the test set.

4. The method according to claim 3, characterized in that, The real clock difference data is divided into training, validation and test sets based on a time series length of 6:2:

2.

5. The method according to claim 1 or 2, wherein the machine learning model is trained by the following steps: initializing and constructing an encoder and decoder structure for a clock sequence prediction model composed of a dilated causal convolutional network, a regularization network, a fully convolutional network, and a residual network, and allocating initial parameters, including the sliding window size; and training the constructed initial model using a training set to obtain the machine learning model.

6. A clock error prediction system for a traceability and comparison timekeeping system, characterized in that, include: The judgment module is used to determine whether the acquisition of the superior shared view file was successful, wherein the superior shared view file is used for tracing and comparing the timekeeping system; The prediction module is used to predict clock difference data when the acquisition of the superior common view file fails, and to obtain the predicted clock difference data. The clock difference data is predicted using a machine learning model, which is trained using real clock difference data. The trained machine learning model is used to output the predicted clock difference data based on the real clock difference data. The submission module is used to submit the predicted clock difference data to the source tracing and comparison timekeeping system; The training module is used to collect clock difference data of the current complete time window when the upper-level common-view file is successfully acquired, and to update the weights in the machine learning model using the collected clock difference data of the current complete time window. The machine learning model includes a network consisting of a regularized layer cascaded with dilated causal convolution, a batch normalization layer, and an activation layer, as well as a one-dimensional fully convolutional network and a fully connected layer. After the input data passes through the input layer, it simultaneously enters the network consisting of a regularized layer cascaded with dilated causal convolution, a batch normalization layer, and an activation layer, and then the one-dimensional fully convolutional network. The feature vectors output from the two branches are combined and then enter the fully connected layer to generate the output.

7. The system according to claim 6, characterized in that, The machine learning model has been trained and validated using clock difference data from at least the previous complete time window.

8. The system according to claim 6 or 7, characterized in that, The training module is used for: The actual clock difference data from the previous complete time window are divided into training set, validation set, and test set; Multiple rounds of training are performed using training data. After each round of training, the resulting machine learning model is validated and adjusted using a validation set. After the multiple rounds of training are completed, the trained machine learning model is tested and evaluated using the test set.

Images