Training method and prediction method for clock error prediction model of atomic clock

Abstract

The embodiment of the invention discloses a training method and a prediction method of an atomic clock error prediction model. The method comprises the following steps: acquiring target real-time clock error sequence data; processing the target real-time clock error sequence data to obtain a local feature sequence; processing the target real-time clock error sequence data to obtain a global feature sequence; calculating a dynamic weight vector by using the local feature sequence and the global feature sequence, and performing weighted summation on the local feature sequence and the global feature sequence to obtain a fused feature sequence; processing the fusion feature sequence and the real-time clock error sequence data to obtain clock error prediction sequence data; processing the clock error prediction sequence data and the target real-time clock error sequence data to obtain an error compensation value, and obtaining compensated clock error prediction sequence data according to the clock error prediction sequence data and the error compensation value; and processing the compensated clock error prediction sequence data and a historical target clock error prediction value by using a Kalman filtering algorithm to obtain a predicted target clock error prediction value.

Description

Technical Field

[0001] This invention relates to the field of time and frequency measurement. More specifically, it relates to a training method and a prediction method for an atomic clock bias prediction model. Background Technology

[0002] Currently, existing atomic clock bias prediction techniques mainly rely on traditional algorithms such as Kalman filtering and grey models. Their core shortcomings lie in the linear assumptions and stationarity requirements. For example, Kalman filtering struggles to model the nonlinear drift caused by atomic clock aging, resulting in errors exceeding 0.3 nanoseconds in predictions exceeding 100 hours. Grey models are sensitive to non-stationary disturbances such as sudden temperature changes and have poor cross-scenario generalization capabilities. Furthermore, these methods rely on manually designed features and cannot adaptively extract complex temporal patterns from clock bias data, leading to insufficient long-term prediction accuracy.

[0003] While emerging deep learning models such as LSTM and Transformer can capture temporal features, they still have significant drawbacks: a single LSTM model suffers from the vanishing gradient problem, resulting in a root mean square error (RMSE) as high as 0.15 nanoseconds in a 720-hour prediction; the standard Transformer is sensitive to local noise, with the mean absolute error (MAE) increasing by 27% under the interference of outliers in cesium clock data. More importantly, existing methods do not consider the differences in atomic clock types (cesium clocks, hydrogen clocks, optical clocks) and dynamic environmental factors (such as temperature drift and equipment aging), requiring model retraining for cross-device migrations, thus limiting their engineering practicality. Summary of the Invention

[0004] The purpose of this invention is to provide a training method and a prediction method for an atomic clock bias prediction model, so as to solve at least one of the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a training method for an atomic clock bias prediction model, the method comprising: Obtain the target real-time clock difference sequence data based on the real-time clock difference sequence data; The target real-time clock difference sequence data is processed using a first machine learning model to obtain a local feature sequence; The target real-time clock difference sequence data is processed using a second machine learning model to obtain... Global feature sequence; A dynamic weight vector is calculated using the local feature sequence and the global feature sequence, and a fused feature sequence is obtained by weighted summation of the local feature sequence and the global feature sequence based on the dynamic weight vector. The fused feature sequence and the real-time clock difference sequence data are processed using a third machine learning model to obtain clock difference prediction sequence data; The clock difference prediction sequence data and the target real-time clock difference sequence data are processed using a fourth machine learning model to obtain the error compensation value of the clock difference prediction sequence data, and the compensated clock difference prediction sequence data is obtained based on the clock difference prediction sequence data and the error compensation value. The predicted target clock difference is obtained by processing the compensated clock difference prediction sequence data and the historical target clock difference prediction values ​​using the Kalman filter algorithm.

[0006] Optionally, obtaining the target real-time clock difference sequence data based on the real-time clock difference sequence data includes: The clock difference timing input data is obtained by aligning the real-time clock difference sequence data with timestamps and using a sliding window. Outlier processing is performed on the clock difference timing input data. Missing value imputation and data standardization yield the target real-time clock difference sequence data.

[0007] Optionally, the step of processing the target real-time clock difference sequence data using the first machine learning model to obtain a local feature sequence includes:

[0008] In the formula, For local feature sequences, ; This represents the local temporal characteristics at time step 1. This represents the local temporal characteristics at time step 2. The local temporal features at time t are... This represents the local temporal characteristics at time N; , Let be the sequence of forward hidden states at time t. Let be the sequence of backward hidden states at time t. For splicing processing; , Let be the forward hidden state sequence at time t-1. For the data at time t in the target real-time clock difference sequence, It is a bidirectional long short-term memory network; , This is the sequence of backward hidden states at time t-1; The length of the local feature sequence; The dimension of the hidden layer in a bidirectional long short-term memory network; It is the set of real numbers.

[0009] Optionally, the second machine learning model is used to process the target real-time clock difference sequence data to obtain... The global feature sequence includes:

[0010] In the formula, for Global feature sequence; For deep learning encoders; For the target real-time clock difference sequence data; For position encoding functions, It is the position index of an element in the sequence; It is the dimension index of the positional encoding vector; The position encoding function is as follows:

[0011] In the formula, This represents the dimension of the position encoding vector.

[0012] Optionally, calculating the dynamic weight vector using the local feature sequence and the global feature sequence includes:

[0013] In the formula, It is a dynamic weight vector; For activation functions; This is the weight matrix. ; For bias terms; The step of weighting and summing the local feature sequence and the global feature sequence according to the dynamic weight vector to obtain the fused feature sequence includes:

[0014] In the formula, For fusion of feature sequences; and These are the first and second components in the dynamic weight vector.

[0015] Optionally, the step of processing the fused feature sequence and the target real-time clock difference sequence data using a third machine learning model to obtain clock difference prediction sequence data includes: At time t, the clock difference data at time t-1 of the fused feature sequence and the target real-time clock difference sequence data are processed using the first long short-term memory network to obtain the clock difference prediction sequence data from time t+1 to time t+M.

[0016] Optionally, the step of processing the clock difference prediction sequence data and the target real-time clock difference sequence data using a fourth machine learning model to obtain the error compensation value of the clock difference prediction sequence data includes: The clock difference prediction sequence data from time t+1 to time t+M and the target real-time clock difference sequence data are processed using a second long short-term memory network to obtain the error compensation value from time t+1 to time t+M. The method of obtaining the compensated clock difference prediction sequence data based on the clock difference prediction sequence data and the error compensation value includes:

[0017] In the formula, The compensated clock error prediction sequence data is from time t+1 to time t+M. This is the clock difference prediction sequence data from time t+1 to time t+M; This represents the error compensation value from time t+1 to time t+M; This is the compensation coefficient.

[0018] Optionally, the step of processing the compensated clock difference prediction sequence data and historical target clock difference prediction values ​​using the Kalman filter algorithm to obtain the predicted target clock difference includes:

[0019] In the formula, The target clock difference prediction value from time t+1 to time t+M; The target clock difference prediction value at time k; The Kalman gain at time k; This is the observation matrix for Kalman filtering; This is the compensated clock difference prediction sequence data from time t to time t+M-1.

[0020] Optionally, the method further includes: The target clock difference prediction values ​​from time t+1 to time t+M are de-standardized to obtain the final clock difference prediction values ​​from time t+1 to time t+M.

[0021] A second aspect of the present invention provides a method for predicting atomic clock bias, the method comprising: Obtain the actual target real-time clock difference sequence data based on the actual real-time clock difference sequence data; The actual local feature sequence is obtained by processing the real-time clock difference sequence data of the actual target using the fifth machine learning model; The actual target real-time clock difference sequence data is processed using the sixth machine learning model to obtain the actual... Global feature sequence; The actual dynamic weight vector is calculated using the actual local feature sequence and the actual global feature sequence, and the actual fused feature sequence is obtained by weighted summation of the actual local feature sequence and the actual global feature sequence based on the actual dynamic weight vector. The actual clock error prediction sequence data is obtained by processing the actual fused feature sequence and the final clock error prediction value using the seventh machine learning model. The actual clock difference prediction sequence data and the actual target real-time clock difference sequence data are processed using the eighth machine learning model to obtain the actual error compensation value of the actual clock difference prediction sequence data, and the compensated actual clock difference prediction sequence data is obtained based on the actual clock difference prediction sequence data and the actual error compensation value. The actual target clock difference prediction value is obtained by processing the actual compensated clock difference prediction sequence data and the historical actual target clock difference prediction value using the Kalman filter algorithm.

[0022] The beneficial effects of this invention are as follows: The technical solution described in this invention captures the local temporal correlation of the clock difference sequence through a first machine learning model, extracts local temporal features, models global long-distance dependencies using a second machine learning model, and achieves multi-scale feature fusion by combining an adaptive gating mechanism. This results in high-precision modeling of atomic clock difference sequences and a significant improvement in prediction accuracy on atomic clock measured data compared to traditional methods. It is applicable to atomic clock management in satellite navigation systems (such as GPS and BeiDou), frequency control in national timekeeping laboratories, and clock synchronization optimization in quantum communication networks. Attached Figure Description

[0023] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0024] Figure 1 The flowchart illustrates the training method for the atomic clock error prediction model provided in an embodiment of the present invention. Detailed Implementation

[0025] To more clearly illustrate the present invention, the following description, in conjunction with embodiments and accompanying drawings, further explains the invention. Similar components in the drawings are indicated by the same reference numerals. Those skilled in the art should understand that the specific description below is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0026] While emerging deep learning models such as LSTM and Transformer can capture temporal features, they still have significant drawbacks: a single LSTM model suffers from the vanishing gradient problem, resulting in a root mean square error (RMSE) as high as 0.15 nanoseconds in a 720-hour prediction; the standard Transformer is sensitive to local noise, with the mean absolute error (MAE) increasing by 27% under the interference of outliers in cesium clock data. More importantly, existing methods do not consider the differences in atomic clock types (cesium clocks, hydrogen clocks, optical clocks) and dynamic environmental factors (such as temperature drift and equipment aging), requiring model retraining for cross-device migrations, thus limiting their engineering practicality.

[0027] In view of this, such as Figure 1 As shown, one embodiment of the present invention provides a training method for an atomic clock bias prediction model. The method includes: obtaining target real-time clock bias sequence data based on real-time clock bias sequence data; processing the target real-time clock bias sequence data using a first machine learning model to obtain a local feature sequence; and processing the target real-time clock bias sequence data using a second machine learning model to obtain... A global feature sequence is generated; a dynamic weight vector is calculated using the local feature sequence and the global feature sequence, and a weighted sum of the local feature sequence and the global feature sequence is obtained based on the dynamic weight vector to obtain a fused feature sequence; a third machine learning model is used to process the fused feature sequence and the real-time clock difference sequence data to obtain clock difference prediction sequence data; a fourth machine learning model is used to process the clock difference prediction sequence data and the target real-time clock difference sequence data to obtain an error compensation value for the clock difference prediction sequence data, and a compensated clock difference prediction sequence data is obtained based on the clock difference prediction sequence data and the error compensation value; a Kalman filter algorithm is used to process the compensated clock difference prediction sequence data and the historical target clock difference prediction value to obtain the predicted target clock difference prediction value.

[0028] In a specific example, a third machine learning model is used, with the fused feature sequence as the initial context state of the decoder and the historical clock difference data of the last moment in the target real-time clock difference sequence data as the starting input of the decoder, to perform multi-step iterative prediction and obtain future clock difference prediction sequence data.

[0029] This embodiment captures the local temporal correlation of the clock difference sequence through a first machine learning model, extracts local time series features, models the global long-distance dependency using a second machine learning model, and combines an adaptive gating mechanism to achieve multi-scale feature fusion. This achieves high-precision modeling of the atomic clock difference sequence and significantly improves the prediction accuracy on atomic clock measured data compared to traditional methods. It is applicable to atomic clock management in satellite navigation systems (such as GPS and BeiDou), frequency control in national timekeeping laboratories, and clock synchronization optimization in quantum communication networks.

[0030] In one possible implementation, obtaining the target real-time clock difference sequence data from the real-time clock difference sequence data includes: performing timestamp alignment and sliding window truncation on the real-time clock difference sequence data to obtain clock difference timing input data; and performing outlier processing on the clock difference timing input data. Missing value imputation and data standardization yield the target real-time clock difference sequence data.

[0031] In a specific example, the method includes a first step: real-time clock difference data acquisition and preprocessing. The input is a sequence of real-time clock difference values, in the form of a sequence of historical time differences. This is then truncated using timestamp alignment and a sliding window to construct a structure of length [length missing]. The time-series input data was used. To address noise and dimensional differences in the data, outlier handling and missing value imputation were employed to detect and remove gross errors. Z-score standardization was then used. The original data is normalized.

[0032] In a specific example, the data input includes: real-time clock difference sequence data, in the form of a sequence of clock difference values ​​at historical moments. , This represents the clock difference value at time 1. For the second moment; For time t; truncated by timestamp alignment and sliding window (window length) It can be configured to run from 24 hours to 72 hours, with a step size of 1 hour, to build time-series input data.

[0033] In a specific example, outlier handling includes: calculating the median (Med) and median absolute deviation (MAD) of the data within the sliding window; if the median of the first data point within the sliding window is... Data points If the absolute difference between the data point and the Med value exceeds 3 × MAD, the data point is considered an outlier and is replaced with the Med value.

[0034] In a specific example Missing value filling includes For missing data, use linear interpolation or dynamic filling based on the mean of the preceding and following windows.

[0035] In a specific example, data standardization includes: adopting Sliding window Z-score standardization Dynamically calculate the mean within the window and standard deviation The normalization formula is:

[0036] In the formula, The data is standardized, i.e., the target real-time clock difference sequence data; This is the original data, i.e., the real-time clock difference sequence data.

[0037] Furthermore, the mean The calculation formula is:

[0038] Furthermore, standard deviation The calculation formula is:

[0039] In the formula, The size of the sliding window is limited to historical data. to ; For the first Data points.

[0040] In one possible implementation, processing the target real-time clock difference sequence data using a first machine learning model to obtain a local feature sequence includes:

[0041] In the formula, For local feature sequences, ; This represents the local temporal characteristics at time step 1. This represents the local temporal characteristics at time step 2. The local temporal features at time t are... This represents the local temporal characteristics at time N; , Let be the sequence of forward hidden states at time t. Let be the sequence of backward hidden states at time t. For splicing processing, It is composed of the forward and backward hidden state sequences; , Let be the forward hidden state sequence at time t-1. The data at time t in the target real-time clock difference sequence data, i.e., the input sequence at time step t. Data points, It is a bidirectional long short-term memory network; , This is the sequence of backward hidden states at time t-1; The length of the local feature sequence; The dimension of the hidden layer in a bidirectional long short-term memory network; It is the set of real numbers.

[0042] In a specific example, the method includes a second step: dual-branch feature extraction. A dual-branch network is constructed to extract clock difference sequences separately. Local temporal features and Global dependency feature.

[0043] In a specific example, constructing an LSTM branch A bidirectional LSTM structure is employed to capture the short-term temporal dependencies and periodic variations in clock error data. LSTM units update their hidden states through a gating mechanism, outputting a feature sequence.

[0044] In a specific example, a bidirectional LSTM is a neural network structure capable of processing input sequences simultaneously from front to back and from back to front. This captures contextual information in both directions within the sequence data, effectively capturing local temporal features. The forward hidden state sequence is computed from front to back, encompassing the sequence from its start to the next time step. Information; the backward hidden state sequence is computed from back to front, containing information from the end of the sequence to the time step. The information is then obtained. Subsequently, it was found that bidirectional LSTM can capture local temporal features in the sequence and output local feature sequences.

[0045] In one possible implementation, the target real-time clock difference sequence data is obtained by processing it using a second machine learning model. The global feature sequence includes:

[0046] In the formula, for Global feature sequence, ; For deep learning encoders; For the target real-time clock difference sequence data; For positional encoding functions, each row of the output is generated by the joint calculation of sine and cosine functions based on the parity of the dimension index j, used to represent... The absolute position information of position pos in the global feature sequence; It is the position index of an element in the sequence; It is the dimension index of the positional encoding vector; This means adding the position code to the input sequence element by element, thereby injecting position information into the input representation; This represents a complete Transformer encoder structure, which achieves deep feature extraction by stacking multiple encoding layers.

[0047] The position encoding function is as follows:

[0048] In the formula, This represents the dimension of the position encoding vector.

[0049] In a specific example, constructing a Transformer branch The algorithm employs an encoder structure that incorporates a multi-head self-attention mechanism and a feedforward network to capture long-range dependencies in clock bias. After adding positional encoding to the input sequence, global weights are calculated through self-attention to output a feature sequence.

[0050] In a specific example, self-attention and multi-head attention are used to directly capture the relationship between any two time points, which is suitable for handling globally dependent features.

[0051] In a specific example, the formula for modeling global dependencies using multi-head self-attention is:

[0052] In the formula, For the final attention output; query matrix Represents the current element or position. , It is the length of the target sequence vector; the key matrix. Used to compare with the query vector to determine the strength of the correlation between different elements. , It is the length of the key vector and the value vector; the value matrix It contains the actual input content used to generate the final attention output. ;use The activation function converts the attention score into a probability distribution, determines the weight of each element, and generates a weighted output. This represents the dimension of the key vector.

[0053] In a specific example, positional encoding is generated based on alternating sine and cosine functions according to the parity of the dimension, providing positional information for the input sequence.

[0054] In a specific example, the dimension of the location encoding vector is usually the same as the dimension of the model's input.

[0055] In a specific example, each layer contains the following core operations: based on The formula employs a multi-head self-attention mechanism, a feedforward neural network, and is supplemented with residual connections and layer normalization standard components to stabilize training and improve model performance.

[0056] In one possible implementation, calculating the dynamic weight vector using the local feature sequence and the global feature sequence includes:

[0057] In the formula, It is a dynamic weight vector; The activation function can be... function or Functions, etc., are used to map the input to a specific range in order to determine the weights of each feature branch; This is the weight matrix, specifically the weight matrix of the fully connected layer, used to map the concatenated features to the weight vector space. ; The offset used to adjust the weight calculation for the bias term; The step of weighting and summing the local feature sequence and the global feature sequence according to the dynamic weight vector to obtain the fused feature sequence includes:

[0058] In the formula, For fusion of feature sequences; and The first and second components in the dynamic weight vector represent the weight coefficients of the LSTM branch and Transformer branch features during the fusion process. Dimensions and and They are the same, both are .

[0059] In a specific example, the method includes a third step: adaptive feature fusion.

[0060] In a specific example, an adaptive weight allocation module is designed to dynamically fuse features from two branches.

[0061] In a specific example, the weighted summation of the bi-branch features yields the fused features, preserving both local and global information through a multi-head self-attention mechanism and a feedforward network, capturing long-range dependencies in clock bias. After adding positional encoding to the input sequence, global weight complementarity is calculated via self-attention.

[0062] In one possible implementation, the step of using a third machine learning model to process the fused feature sequence and the target real-time clock difference sequence data to obtain clock difference prediction sequence data includes: at time t, using a first long short-term memory network to process the clock difference data of the fused feature sequence and the target real-time clock difference sequence data at time t-1 to obtain clock difference prediction sequence data from time t+1 to time t+M.

[0063] In a specific example, the method also includes a fourth step: multi-step prediction of the output. This involves inputting the fused features. Seq2Seq structure prediction module The decoder uses LSTM stacked layers, through Teacher Forcing Strategy training, outputting the future A multi-step clock difference prediction sequence. During decoding, the prediction result of each step serves as the input for the next step, achieving multi-step prediction from sequence to sequence.

[0064] In a specific example, the Seq2Seq decoder structure includes: 2-layer LSTM, 128-dimensional hidden layers, and an initial state composed of fused features. The mean vector is initialized.

[0065] Furthermore, a multi-step prediction mechanism: In each prediction step, the decoder takes the fused features output by the encoder and the predicted value of the previous step (the true value in Teacher Forcing mode) as input, and outputs the clock difference prediction sequence for the next M steps.

[0066] Furthermore, the inputs include: the fused features of the encoder output, and the predicted or true values ​​from the previous step (using the Teacher Forcing strategy during the training phase to accelerate training and improve model stability).

[0067] Furthermore, the output includes: future Clock difference prediction sequence of steps In the formula It can be configured to a forecast range of 24 hours to 168 hours, corresponding to a forecast range of 1 day to 7 days.

[0068] Furthermore, the loss function includes: using the root mean square error (RMSE) to evaluate the difference between the predicted and actual values, calculated as follows:

[0069] In the formula, This represents the predicted value, i.e., the clock difference prediction sequence data at time i. This represents the true value, i.e., the target real-time clock difference sequence data at time i. To predict the step size.

[0070] In one possible implementation, the step of processing the clock difference prediction sequence data and the target real-time clock difference sequence data using a fourth machine learning model to obtain the error compensation value of the clock difference prediction sequence data includes: processing the clock difference prediction sequence data from time t+1 to time t+M and the target real-time clock difference sequence data using a second long short-term memory network to obtain the error compensation value from time t+1 to time t+M; the step of obtaining the compensated clock difference prediction sequence data based on the clock difference prediction sequence data and the error compensation value includes:

[0071] In the formula, The compensated clock error prediction sequence data is from time t+1 to time t+M. This is the clock difference prediction sequence data from time t+1 to time t+M; This represents the error compensation value from time t+1 to time t+M; The compensation coefficient is the adjustable compensation coefficient used to control the strength of error compensation.

[0072] In a specific example, the method includes a fifth step: error compensation. To address the accumulated error from multi-step predictions, a residual learning module is designed to calculate the residual sequence between historical predicted values ​​and actual values. , This represents the residual value at time 1. This represents the residual value at time 2. Let t be the residual value at time t; input it into the LSTM network to learn the residual change pattern, and output the future value. The error compensation value is calculated for each step. This compensation value is then added to the initial prediction value to correct the error-compensated prediction.

[0073] In one possible implementation, the step of processing the compensated clock error prediction sequence data and historical target clock error prediction values ​​using the Kalman filter algorithm to obtain the predicted target clock error value includes:

[0074] In the formula, The target clock difference prediction value from time t+1 to time t+M; The target clock difference prediction value at time k; The Kalman gain at time k; This is the observation matrix for Kalman filtering; This is the compensated clock difference prediction sequence data from time t to time t+M-1.

[0075] In a specific example, the method includes a sixth step: Kalman filter fusion.

[0076] Furthermore, the state equation is:

[0077] In the formula, The state vector (clock difference value and its first difference) is the estimated clock difference of the atomic clock at time k. Let $\mathbf{k-1}$ be the clock difference state vector; and $\mathbf{k-1}$ be the state transition matrix. ; Sampling interval; process noise covariance Process noise It follows a pattern with a mean of 0 and a covariance matrix of . The normal distribution, i.e. .

[0078] Furthermore, the observation equation:

[0079] In the formula, Observations for clock error measurement; observation matrix ]; Observation noise , follows a mean of 0 and a variance of The normal distribution, i.e. ~ Observation noise covariance .

[0080] In a specific example, the initial state is obtained by calculating the mean and variance of the predicted values ​​from the first 10 clock difference sequences to obtain the state vector of the atomic clock difference estimate at the initial moment. This provides an initial state estimate for the Kalman filter.

[0081] In a specific example, the filtering process includes a prediction phase and an update phase.

[0082] Furthermore, the prediction phase includes: estimating the optimal state using the clock difference at time k-1. With state transition matrix Through equation Calculate the prior state estimate of the clock difference at time k. That is, the predicted target clock error value at time k; simultaneously, the clock error estimation error covariance matrix at time k-1 is used. Update clock bias prior estimate error covariance .

[0083] Furthermore, the update phase includes: firstly, calculating the Kalman gain of the clock error estimate. This gain determines the level of confidence in the model predictions and observations during clock difference data fusion; subsequently, the prior estimates and observations are fused to obtain the optimal posterior state estimate for the current time; finally, the posterior covariance is updated. Prepare for the next time step clock difference state estimation.

[0084] In one possible implementation, the method further includes: performing inverse standardization on the target clock difference prediction values ​​from time t+1 to time t+M to obtain the final clock difference prediction values ​​from time t+1 to time t+M.

[0085] In a specific example, the method also includes a seventh step: outputting the final clock bias prediction. The Kalman-filtered predictions obtained after the complete prediction and fusion process are denormalized to restore their actual dimensions, thus yielding a future M-step clock bias prediction sequence with a time range from t+1 to t+M. , This is the predicted clock difference value at time t+1. This is the predicted clock difference value at time t+2. The predicted clock difference value at time t+M is used as the final output.

[0086] Another embodiment of the present invention provides a method for predicting atomic clock bias, the method comprising: obtaining actual target real-time clock bias sequence data based on actual real-time clock bias sequence data; processing the actual target real-time clock bias sequence data using a fifth machine learning model to obtain an actual local feature sequence; and processing the actual target real-time clock bias sequence data using a sixth machine learning model to obtain an actual local feature sequence. The process involves: generating a global feature sequence; calculating an actual dynamic weight vector using the actual local feature sequence and the actual global feature sequence; weighting and summing the actual local feature sequence and the actual global feature sequence according to the actual dynamic weight vector to obtain an actual fused feature sequence; processing the actual fused feature sequence and the final clock error prediction value using a seventh machine learning model to obtain actual clock error prediction sequence data; processing the actual clock error prediction sequence data and the actual target real-time clock error sequence data using an eighth machine learning model to obtain the actual error compensation value of the actual clock error prediction sequence data; and obtaining the compensated actual clock error prediction sequence data based on the actual clock error prediction sequence data and the actual error compensation value; and processing the actual compensated clock error prediction sequence data and the historical actual target clock error prediction value using a Kalman filter algorithm to obtain the predicted actual target clock error prediction value.

[0087] This embodiment captures the local temporal correlation of the clock difference sequence through a first machine learning model, extracts local time series features, models the global long-distance dependency using a second machine learning model, and combines an adaptive gating mechanism to achieve multi-scale feature fusion. This achieves high-precision modeling of the atomic clock difference sequence and significantly improves the prediction accuracy on atomic clock measured data compared to traditional methods. It is applicable to atomic clock management in satellite navigation systems (such as GPS and BeiDou), frequency control in national timekeeping laboratories, and clock synchronization optimization in quantum communication networks.

[0088] In the description of this invention, it should be noted that the terms "upper," "lower," etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two elements. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0089] It should also be noted that in the description of this invention, 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, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0090] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is impossible to exhaustively list all the implementation methods here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A training method for an atomic clock error prediction model, characterized in that, The method includes: Obtain the target real-time clock difference sequence data based on the real-time clock difference sequence data; The target real-time clock difference sequence data is processed using a first machine learning model to obtain a local feature sequence; A global feature sequence is obtained by processing the target real-time clock difference sequence data using a second machine learning model; A dynamic weight vector is calculated using the local feature sequence and the global feature sequence, and a fused feature sequence is obtained by weighted summation of the local feature sequence and the global feature sequence based on the dynamic weight vector. The fused feature sequence and the real-time clock difference sequence data are processed using a third machine learning model to obtain clock difference prediction sequence data; The clock difference prediction sequence data and the target real-time clock difference sequence data are processed using a fourth machine learning model to obtain the error compensation value of the clock difference prediction sequence data, and the compensated clock difference prediction sequence data is obtained based on the clock difference prediction sequence data and the error compensation value. The predicted target clock difference is obtained by processing the compensated clock difference prediction sequence data and the historical target clock difference prediction values ​​using the Kalman filter algorithm.

2. The training method for the atomic clock error prediction model according to claim 1, characterized in that, The step of obtaining the target real-time clock difference sequence data based on the real-time clock difference sequence data includes: The clock difference timing input data is obtained by aligning the real-time clock difference sequence data with timestamps and using a sliding window. The clock difference timing input data is processed for outliers, missing values ​​are filled, and the data is standardized to obtain the target real-time clock difference sequence data.

3. The training method for the atomic clock error prediction model according to claim 2, characterized in that, The process of using the first machine learning model to process the target real-time clock difference sequence data to obtain the local feature sequence includes: In the formula, For local feature sequences, ; This represents the local temporal characteristics at time step 1. This represents the local temporal characteristics at time step 2. The local temporal features at time t are... This represents the local temporal characteristics at time N; , Let be the sequence of forward hidden states at time t. Let be the sequence of backward hidden states at time t. For splicing processing; , Let be the forward hidden state sequence at time t-1. For the data at time t in the target real-time clock difference sequence, It is a bidirectional long short-term memory network; , This is the sequence of backward hidden states at time t-1; The length of the local feature sequence; The dimension of the hidden layer in a bidirectional long short-term memory network; It is the set of real numbers.

4. The training method for the atomic clock error prediction model according to claim 3, characterized in that, The target real-time clock difference sequence data is obtained by processing it using a second machine learning model. The global feature sequence includes: In the formula, for Global feature sequence; For deep learning encoders; For the target real-time clock difference sequence data; For position encoding functions, It is the position index of an element in the sequence; It is the dimension index of the positional encoding vector; The position encoding function is as follows: In the formula, This represents the dimension of the position encoding vector.

5. The training method for the atomic clock error prediction model according to claim 4, characterized in that, The calculation of the dynamic weight vector using the local feature sequence and the global feature sequence includes: In the formula, It is a dynamic weight vector; For activation functions; This is the weight matrix. ; For bias terms; The step of weighting and summing the local feature sequence and the global feature sequence according to the dynamic weight vector to obtain the fused feature sequence includes: In the formula, For fusion feature sequences; and These are the first and second components in the dynamic weight vector.

6. The training method for the atomic clock error prediction model according to claim 5, characterized in that, The process of using a third machine learning model to process the fused feature sequence and the target real-time clock difference sequence data to obtain the clock difference prediction sequence data includes: At time t, the clock difference data at time t-1 of the fused feature sequence and the target real-time clock difference sequence data are processed using the first long short-term memory network to obtain the clock difference prediction sequence data from time t+1 to time t+M.

7. The training method for the atomic clock error prediction model according to claim 6, characterized in that, The step of processing the clock difference prediction sequence data and the target real-time clock difference sequence data using the fourth machine learning model to obtain the error compensation value of the clock difference prediction sequence data includes: The clock difference prediction sequence data from time t+1 to time t+M and the target real-time clock difference sequence data are processed using a second long short-term memory network to obtain the error compensation value from time t+1 to time t+M. The method of obtaining the compensated clock difference prediction sequence data based on the clock difference prediction sequence data and the error compensation value includes: In the formula, The compensated clock error prediction sequence data is from time t+1 to time t+M. This is the clock difference prediction sequence data from time t+1 to time t+M; This represents the error compensation value from time t+1 to time t+M; This is the compensation coefficient.

8. The training method for the atomic clock error prediction model according to claim 7, characterized in that, The process of using the Kalman filter algorithm to process the compensated clock error prediction sequence data and historical target clock error prediction values ​​to obtain the predicted target clock error value includes: In the formula, The target clock difference prediction value from time t+1 to time t+M; The target clock difference prediction value at time k; The Kalman gain at time k; This is the observation matrix for Kalman filtering; This is the compensated clock difference prediction sequence data from time t to time t+M-1.

9. The training method for the atomic clock error prediction model according to claim 8, characterized in that, The method also includes: The target clock difference prediction values ​​from time t+1 to time t+M are de-standardized to obtain the final clock difference prediction values ​​from time t+1 to time t+M.

10. A method for predicting atomic clock bias, characterized in that, The method includes: Obtain the actual target real-time clock difference sequence data based on the actual real-time clock difference sequence data; The actual local feature sequence is obtained by processing the real-time clock difference sequence data of the actual target using the fifth machine learning model; The actual target real-time clock difference sequence data is processed using the sixth machine learning model to obtain the actual... Global feature sequence; The actual dynamic weight vector is calculated using the actual local feature sequence and the actual global feature sequence, and the actual fused feature sequence is obtained by weighted summation of the actual local feature sequence and the actual global feature sequence based on the actual dynamic weight vector. The actual clock error prediction sequence data is obtained by processing the actual fused feature sequence and the final clock error prediction value using the seventh machine learning model. The actual clock difference prediction sequence data and the actual target real-time clock difference sequence data are processed using the eighth machine learning model to obtain the actual error compensation value of the actual clock difference prediction sequence data, and the compensated actual clock difference prediction sequence data is obtained based on the actual clock difference prediction sequence data and the actual error compensation value. The actual target clock difference prediction value is obtained by processing the actual compensated clock difference prediction sequence data and the historical actual target clock difference prediction value using the Kalman filter algorithm.

Images