A method and apparatus for clock error prediction using a dual-channel neural network

Abstract

This application relates to a method and apparatus for clock error prediction using a dual-channel neural network. The method includes: performing differential processing on satellite clock error data; calculating the generation time of gross errors based on median robust estimation and the first differential result; locating the gross errors using their generation times and removing them from their corresponding positions; supplementing missing data using cubic spline interpolation; sampling the supplemented clock error data using a pre-set sampling interval and sequence length; segmenting the sequence using the sampled data; normalizing the segmented feature sequences; constructing a dual-channel neural network model and training it using a pre-set training set; inputting the normalized feature sequences into the trained dual-channel neural network model and outputting predicted values; inverse normalizing the predicted values; and recovering the clock error based on the inverse normalized predicted values ​​to obtain the predicted clock error value. This method can reduce error accumulation and improve prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of satellite navigation technology, and in particular to a clock bias prediction method and apparatus using a dual-channel neural network. Background Technology

[0002] Atomic clocks are a crucial payload in satellite navigation systems, used to maintain the system's time. To achieve high-precision positioning and timing, accurate clock error prediction models need to be provided to users to correct satellite clock errors and obtain the system's time reference.

[0003] Currently, by extracting features and modeling noise in satellite clock bias, researchers have established various efficient clock bias prediction models, commonly including linear models (LP), quadratic polynomial models (QP), and grey models (GM), which can achieve a certain level of prediction accuracy. However, these models all model the linear trend of the clock bias sequence, while spaceborne atomic clocks exhibit nonlinear fluctuations due to environmental influences, making it impossible to accurately characterize this feature. To further improve prediction accuracy, a series of deep learning-based clock bias prediction models have been proposed, including extreme learning machines, RBF neural networks, BP neural networks, and LSTM neural networks, which have achieved performance improvements to some extent. However, error accumulation in long-term clock bias prediction leads to performance degradation. To further improve network prediction performance while reducing error accumulation in multi-step predictions, a new clock bias prediction network architecture and method are needed. Summary of the Invention

[0004] Therefore, it is necessary to provide a clock error prediction method and apparatus using a dual-channel neural network that can reduce error accumulation and improve prediction accuracy, in order to address the aforementioned technical problems.

[0005] A clock bias prediction method using a dual-channel neural network, the method comprising:

[0006] Acquire satellite clock bias data; perform differential processing on the satellite clock bias data to obtain a first differential result;

[0007] The time of occurrence of gross errors is calculated based on the median robust estimation method and the first difference result; after locating the gross errors using the time of occurrence, the gross errors at the corresponding positions are removed, and the missing data is completed using the cubic spline interpolation method to obtain the completed clock error data.

[0008] The completed clock difference data is sampled using a pre-set sampling interval and sequence length, and the sampled data is used to segment the sequence; the segmented feature sequence is then normalized to obtain the normalized feature sequence.

[0009] A dual-channel neural network model is constructed and trained using a pre-set training set. The dual-channel neural network model includes two channels: one channel is composed of LSTM units to extract the temporal features of the sequence, and the other channel is composed of a one-dimensional convolutional layer to extract the spatial features of the sequence. The outputs of the two channels are added together and fed into a three-layer fully connected layer to fuse the temporal and spatial features.

[0010] The normalized feature sequence is input into the trained dual-channel neural network model, which outputs the predicted value; the predicted value is denormalized to obtain the denormalized predicted value; the clock error is recovered based on the denormalized predicted value to obtain the clock error predicted value.

[0011] A clock bias prediction device based on a dual-channel neural network, characterized in that the device comprises:

[0012] The clock error completion module is used to acquire satellite clock error data; perform differential processing on the satellite clock error data to obtain a first differential result; calculate the time of occurrence of gross errors based on the median robust estimation method and the first differential result; locate the gross errors using the time of occurrence of the gross errors, remove the gross errors at the corresponding positions, and complete the missing data using cubic spline interpolation to obtain the completed clock error data.

[0013] The sampling and sequence segmentation module is used to sample the completed clock difference data using a pre-set sampling interval and sequence length, and to sample the sequence using the sampled data; the segmented feature sequence is normalized to obtain the normalized feature sequence.

[0014] A module for constructing a dual-channel neural network model is used to build a dual-channel neural network model and train it using a pre-set training set. The dual-channel neural network model includes two channels: one channel is composed of LSTM units to extract the temporal features of the sequence, and the other channel is composed of a one-dimensional convolutional layer to extract the spatial features of the sequence. The outputs of the two channels are added together and fed into a three-layer fully connected layer to fuse the temporal and spatial features.

[0015] The clock bias prediction module is used to input the normalized feature sequence into the trained dual-channel neural network model and output the predicted value; to perform inverse normalization on the predicted value to obtain the inverse normalized predicted value; and to recover the clock bias based on the inverse normalized predicted value to obtain the clock bias prediction value.

[0016] The aforementioned dual-channel neural network clock error prediction method and apparatus first acquires and differentially processes clock error data to remove trend components. It then uses median robust estimation to detect and remove outliers, and employs cubic spline interpolation to complete missing data, ensuring data integrity and continuity. Next, the data is sampled and segmented into smaller sequences using a preset sampling interval and sequence length, followed by normalization to eliminate non-stationarity and dimensional differences. Subsequently, a dual-channel neural network is constructed, where the LSTM channel extracts temporal features and the convolutional channel extracts spatial features. These two types of information are then fused through a fully connected layer, ultimately achieving high-precision clock error prediction. This effectively reduces noise accumulation and fully exploits the nonlinear patterns in the clock error sequence, significantly improving prediction accuracy. This application eliminates extreme outliers in the data through gross error removal, reducing interference with subsequent modeling processes. Interpolation completion ensures data continuity and smoothness, avoiding the negative impact of missing data on model training and improving data quality. This leads to more stable subsequent model training, reducing error accumulation and improving clock error prediction accuracy. Furthermore, data sampling, segmentation, and normalization improve training efficiency and model generalization ability. Through local training, the model can better adapt to data from different satellites or different time periods, thereby improving overall prediction accuracy. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a clock error prediction method using a dual-channel neural network in one embodiment.

[0018] Figure 2 This is a schematic diagram of the network structure of a dual-channel neural network model in one embodiment;

[0019] Figure 3 This is a basic schematic diagram of the data sampling process in one embodiment;

[0020] Figure 4 This is a schematic diagram illustrating the clock error prediction residuals of this application on satellite G02 in another embodiment, and comparing them with the residuals of models such as linear polynomial (LP), quadratic polynomial (QP), grey model (GM), autoregressive moving average (ARMA), random forest (RF), and long short-term memory network (LSTM).

[0021] Figure 5 This is a comparison chart of the clock error prediction residuals of this application on satellite G07 in one embodiment, and the residuals of other models such as linear polynomial (LP), quadratic polynomial (QP), grey model (GM), autoregressive moving average (ARMA), random forest (RF), and long short-term memory network (LSTM).

[0022] Figure 6This is a schematic diagram of a clock error prediction device using a dual-channel neural network in one embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] In one embodiment, such as Figure 1 As shown, a clock bias prediction method using a two-channel neural network is provided, comprising the following steps:

[0025] Step 102: Obtain satellite clock bias data; perform differential processing on the satellite clock bias data to obtain the first differential result; calculate the time of occurrence of gross errors based on the median robust estimation method and the first differential result; locate the gross errors using the time of occurrence of the gross errors, remove the gross errors at the corresponding positions, and use cubic spline interpolation to complete the missing data to obtain the completed clock bias data.

[0026] Satellite clock bias data is collected to provide foundational data for subsequent modeling. The satellite clock bias data undergoes differencing; primary differencing weakens linear components and exposes nonlinear characteristics, which is beneficial for subsequent gross error removal. Specifically, the primary differencing result is as follows:

[0027]

[0028] in, The result of a first difference at time t. Let t be the clock difference at time t.

[0029] The timing of outlier occurrences is calculated using the median robust estimation method and first-order differencing. Removing outliers through median robust estimation ensures that the dataset contains no outliers that could affect the overall trend. This is crucial for data stationarity and consistency, especially when processing time series data, where outliers often cause prediction errors and model instability. By locating and removing outliers, data quality is significantly improved. Missing data or outliers, if not removed, can lead to error accumulation during model training, affecting prediction accuracy. Using cubic spline interpolation to complete missing data smoothly restores the data, avoiding large fluctuations or abrupt changes that may occur with simple interpolation methods. This process not only maintains data continuity but also avoids errors or unevenness that may arise during interpolation. Data processed in this way better reflects actual changes, improving model prediction accuracy. Furthermore, in multi-step prediction of time series data, any small prediction error can gradually amplify with increasing steps, leading to performance degradation (i.e., error accumulation). Gross error removal and interpolation completion methods effectively ensure data quality at each time step, thereby reducing the propagation and accumulation of initial errors. By ensuring that the input data at each step undergoes rigorous preprocessing (such as gross error removal, differencing, and interpolation completion), the error during model training is controlled at a low level, preventing the prediction error from increasing over time due to inherent data defects. This is particularly important for long-term predictions, as error accumulation can significantly affect the accuracy of prediction results.

[0030] Step 104: Sample the completed clock difference data using a pre-set sampling interval and sequence length, and use the sampled data to sample the sequence; normalize the segmented feature sequence to obtain the normalized feature sequence.

[0031] By pre-setting sampling intervals and sequence lengths, clock difference data is sampled systematically to ensure that each sampling point contains sufficient information within a given time period. This ensures that each sampling point represents the key features of the clock difference sequence as accurately as possible, without being affected by noise or extreme fluctuations in the data. The sampling interval and sequence length effectively control the data resolution. An appropriate sampling interval avoids redundant information caused by oversampling, while ensuring that key information is preserved and reducing the impact of noise. Furthermore, when the data sequence is too long, the model may encounter training difficulties or inaccurate predictions due to long-term dependencies or error accumulation. By dividing the data into multiple fixed-length subsequences, shorter time series are more likely to capture local patterns and reduce the gradual accumulation of errors.

[0032] Meanwhile, data sampling and segmentation effectively reduce the size of the input data, enabling the model to be trained more efficiently. Compared to processing excessively long original sequences, the segmented subsequences are easier to process in batches, reducing the computational overhead during training.

[0033] Atomic clocks exhibit frequency drift, and data after single-difference often displays a degree of non-stationarity. To improve the model's generalization ability and its ability to learn sequence features, normalizing the feature sequence helps reduce the impact of these non-stationary features on the model, allowing it to process all features equally. When dealing with tasks containing multiple features, normalization ensures that each feature contributes equally to the prediction result, thus improving the model's generalization ability, especially when processing data across satellites or time periods.

[0034] Step 106: Construct a dual-channel neural network model and train it using a pre-set training set. The dual-channel neural network model includes two channels: one channel is composed of LSTM units to extract the temporal features of the sequence, and the other channel is composed of a one-dimensional convolutional layer to extract the spatial features of the sequence. The outputs of the two channels are added together and fed into a three-layer fully connected layer to fuse the temporal and spatial features.

[0035] like Figure 2 The diagram shows the network structure of a dual-channel neural network model. One channel consists of LSTM units to extract the temporal features of the sequence, while the other channel consists of a one-dimensional convolutional layer to extract the spatial features. The outputs of the two channels are summed and fed into a three-layer fully connected layer to fuse the temporal and spatial features, thereby improving the accuracy of sequence prediction. The network is trained using segmented training data to obtain its optimal parameters. Modeling solely based on temporal or spatial features may not fully capture the complexity of clock difference sequences. The dual-channel network combines the advantages of time series modeling and local feature extraction, enabling it to more fully extract the nonlinear patterns in clock difference data and improve prediction accuracy.

[0036] Step 108: Input the normalized feature sequence into the trained dual-channel neural network model and output the predicted value; perform inverse normalization on the predicted value to obtain the inverse normalized predicted value; recover the clock error based on the inverse normalized predicted value to obtain the clock error predicted value.

[0037] The aforementioned dual-channel neural network-based clock error prediction method first acquires and differentially processes clock error data to remove trend components. It then uses median robust estimation to detect and remove outliers, and employs cubic spline interpolation to complete missing data, ensuring data integrity and continuity. Next, the data is sampled and segmented into smaller sequences using a preset sampling interval and sequence length, followed by normalization to eliminate non-stationarity and dimensional differences. Subsequently, a dual-channel neural network is constructed, where the LSTM channel extracts temporal features and the convolutional channel extracts spatial features. These two information are then fused through a fully connected layer, ultimately achieving high-precision clock error prediction. This method effectively reduces noise accumulation and fully exploits the nonlinear patterns in the clock error sequence, significantly improving prediction accuracy. This application eliminates extreme outliers in the data through gross error removal, reducing interference with subsequent modeling processes. Interpolation completion ensures data continuity and smoothness, avoiding the negative impact of missing data on model training and improving data quality. This leads to more stable subsequent model training, reducing error accumulation and improving clock error prediction accuracy. Furthermore, data sampling, segmentation, and normalization improve training efficiency and model generalization ability. Through local training, the model can better adapt to data from different satellites or different time periods, thereby improving overall prediction accuracy.

[0038] In one embodiment, calculating the time of occurrence of gross errors based on the robust median estimation method and the first difference result includes:

[0039] The timing of gross errors is calculated based on the median robust estimation method and the first difference result.

[0040]

[0041] in, for The middle number, This is the result of a first-order difference at time t.

[0042] In one embodiment, the gross error is located using the moment it occurs, including:

[0043] The clock error discrimination threshold is designed based on the time of occurrence of gross errors. ,when When the current difference result is determined to be a gross error.

[0044] In one embodiment, the sampled data is

[0045]

[0046] in, The matrix is Matrix, front As a feature, the first Columns serve as data labels. Indicates the sequence length. Indicates the sampling interval. Indicates the data length. This represents the difference data.

[0047] In one embodiment, the sequence is segmented using the sampled data, including:

[0048] No. i The feature sequences are represented as follows:

[0049]

[0050] in, Indicates the length of the feature sequence. Indicates the sequence length.

[0051] In one embodiment, normalizing the segmented feature sequence to obtain a normalized feature sequence further includes:

[0052] Normalize the segmented feature sequence to obtain the normalized feature sequence.

[0053]

[0054] in, Indicates the first i A feature sequence, This represents the minimum value in the sequence. This represents the maximum value in the sequence.

[0055] In one embodiment, the normalized feature sequence is input into a trained dual-channel neural network model, and the output predicted value includes:

[0056] When the length of the input normalized feature sequence Less than the sampling interval When, then cut off the previous part. Each point is output as a predicted value; when the length of the sequence to be predicted is... Greater than the sampling interval At that time, calculate / The result is rounded up to the nearest integer, which represents the number of network iterations for prediction.

[0057] In one embodiment, the predicted value is denormalized to obtain the denormalized predicted value, including:

[0058] The predicted values ​​are denormalized to obtain the denormalized predicted values.

[0059]

[0060] in, It is the first i One inverse normalized prediction value, It is the first network i One non-normalized predicted value, This represents the minimum value in the sequence. This represents the minimum value in the sequence.

[0061] In one embodiment, clock bias is recovered from the inverse normalized prediction value to obtain the clock bias prediction value, including:

[0062] Clock bias is recovered from the inverse normalized prediction values, resulting in the predicted clock bias value.

[0063]

[0064] in, To begin predicting the clock difference value of the previous moment, It is the first i One inverse normalized prediction value.

[0065] In a specific embodiment, such as Figure 3 As shown, this is the process of constructing a dataset based on sampling, which can significantly improve the accuracy of long-term neural network predictions and reduce the accumulation of errors caused by a large number of iterative predictions.

[0066] like Figure 4 , Figure 5 The figure shows the 24-hour prediction residual curves generated from a prediction experiment using the designed network on GPS satellites G02 and G07. In the experiment, p was set to 36 and L to 32, meaning sampling occurred every 36 points, and the feature sequence length was 32. Compared with classic algorithms such as LP, QP, GM, ARMA, RF, and LSTM, the dual-channel neural network exhibited the smallest prediction residual, demonstrating the effectiveness of this application.

[0067] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0068] In one embodiment, such as Figure 6 As shown, a clock error prediction device based on a dual-channel neural network is provided, including: a clock error completion module 602, a sampling and sequence segmentation module 604, a dual-channel neural network model construction module 606, and a clock error prediction module 608;

[0069] The clock error completion module 602 is used to acquire satellite clock error data; perform differential processing on the satellite clock error data to obtain a first differential result; calculate the time of occurrence of gross errors based on the median robust estimation method and the first differential result; locate the gross errors using the time of occurrence of the gross errors, remove the gross errors at the corresponding positions, and complete the missing data using cubic spline interpolation to obtain the completed clock error data.

[0070] The sampling and sequence segmentation module 604 is used to sample the completed clock difference data using a pre-set sampling interval and sequence length, and to sample the sequence using the sampled data; and to normalize the segmented feature sequence to obtain the normalized feature sequence.

[0071] A dual-channel neural network model module 606 is used to construct a dual-channel neural network model and train it using a pre-set training set. The dual-channel neural network model includes two channels: one channel is composed of LSTM units to extract the temporal features of the sequence, and the other channel is composed of a one-dimensional convolutional layer to extract the spatial features of the sequence. The outputs of the two channels are added together and fed into a three-layer fully connected layer to fuse the temporal and spatial features.

[0072] The clock error prediction module 608 is used to input the normalized feature sequence into the trained dual-channel neural network model and output the predicted value; to perform inverse normalization on the predicted value to obtain the inverse normalized predicted value; and to recover the clock error based on the inverse normalized predicted value to obtain the clock error prediction value.

[0073] Specific limitations regarding the clock error prediction device using a dual-channel neural network can be found in the limitations of the clock error prediction method using a dual-channel neural network described above, and will not be repeated here. Each module in the aforementioned clock error prediction device using a dual-channel neural network can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0074] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0075] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0076] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A clock bias prediction method using a dual-channel neural network, characterized in that, The method includes: Acquire satellite clock bias data; perform differential processing on the satellite clock bias data to obtain a first differential result; The time of occurrence of gross errors is calculated based on the median robust estimation method and the first difference result; after locating the gross errors using the time of occurrence, the gross errors at the corresponding positions are removed, and the missing data is completed using cubic spline interpolation to obtain the completed clock error data; The completed clock difference data is sampled using a pre-set sampling interval and sequence length, and the sampled data is used to segment the sequence; the segmented feature sequence is normalized to obtain the normalized feature sequence. A dual-channel neural network model is constructed and trained using a pre-set training set. The dual-channel neural network model includes two channels: one channel is composed of LSTM units to extract the temporal features of the sequence, and the other channel is composed of a one-dimensional convolutional layer to extract the spatial features of the sequence. The outputs of the two channels are added together and fed into a three-layer fully connected layer to fuse the temporal and spatial features. The normalized feature sequence is input into the trained dual-channel neural network model to output the predicted value; the predicted value is denormalized to obtain the denormalized predicted value; the clock error is recovered based on the denormalized predicted value to obtain the clock error predicted value. The timing of gross errors is calculated based on the robust median estimation method and the first difference result, including: The timing of gross errors is calculated based on the median robust estimation method and the first difference result. MAD = Median{|y(t)-m| / 0.6745} Where m = Median{y(t)} is the median of y(t), and y(t) is the result of the first difference at time t; The predicted values ​​are denormalized to obtain denormalized predicted values, including: The predicted value is denormalized to obtain the denormalized predicted value. y reverse (i)=y pre (i)×(max(F_seq(i))-min(F_seq(i)))+min(F_seq(i)) Among them, y reverse (i) is the i-th inverse normalized prediction value, y pre (i) is the i-th nonnormalized prediction value of the network, min(F_seq(i)) represents the minimum value in the sequence, and max(F_seq(i)) represents the maximum value in the sequence.

2. The method according to claim 1, characterized in that, Locating gross errors by utilizing the moment of their occurrence includes: The clock error discrimination threshold is designed based on the time of occurrence of gross errors, which is (m+l·MAD). When |y(t)|>(m+l·MAD), the current difference result is determined to be a gross error.

3. The method according to claim 1, characterized in that, The sampled data is The data matrix is ​​an (N-pL)×(L+1) matrix, where the first L columns are used as features, the L+1th column is used as the data label, L represents the sequence length, p represents the sampling interval, N represents the data length, and y represents the first difference data.

4. The method according to claim 1, characterized in that, The sequence is segmented using the sampled data, including: The i-th feature sequence is represented as: F_seq(i)=[y i ,y i+M ,y i+2M ,...y i+(L-1)M ] Where M represents the length of the feature sequence and L represents the sequence length.

5. The method according to claim 1, characterized in that, The segmented feature sequences are normalized to obtain normalized feature sequences, which also include: Normalize the segmented feature sequence to obtain the normalized feature sequence. Where F_seq(i) represents the i-th feature sequence, min(F_seq(i)) represents the minimum value in the sequence, and max(F_seq(i)) represents the maximum value in the sequence.

6. The method according to claim 1, characterized in that, The normalized feature sequence is input into the trained dual-channel neural network model, and the predicted value is output, including: when the length M of the input normalized feature sequence is less than the sampling interval p, the first M points are truncated as the predicted value output; when the length M of the sequence to be predicted is greater than the sampling interval p, M / p is calculated and rounded up, which is the number of network iterations for prediction.

7. The method according to claim 1, characterized in that, Clock bias is recovered from the inverse normalized prediction values, resulting in predicted clock bias values, including: Clock bias is recovered from the inverse normalized prediction values, resulting in the predicted clock bias value. Where x0 is the clock difference value at the moment before the start of prediction, and y reverse (i) is the i-th inverse normalized prediction.

8. A clock bias prediction device based on a dual-channel neural network, characterized in that, The device includes: A clock error completion module is used to acquire satellite clock error data; perform differential processing on the satellite clock error data to obtain a first-order differential result; calculate the occurrence time of gross errors based on the median robust estimation method and the first-order differential result; locate the gross errors using their occurrence times, remove the corresponding gross errors, and complete the missing data using cubic spline interpolation to obtain the completed clock error data; the calculation of the occurrence time of gross errors based on the median robust estimation method and the first-order differential result includes: The timing of gross errors is calculated based on the median robust estimation method and the first difference result. MAD = Median{|y(t)-m| / 0.6745} Where m = Median{y(t)} is the median of y(t), and y(t) is the result of the first difference at time t; The sampling and sequence segmentation module is used to sample the completed clock difference data using a pre-set sampling interval and sequence length, and to sample the sequence using the sampled data; the segmented feature sequence is normalized to obtain the normalized feature sequence. A dual-channel neural network model module is used to construct a dual-channel neural network model and train it using a pre-set training set. The dual-channel neural network model includes two channels: one channel is composed of LSTM units to extract the temporal features of the sequence, and the other channel is composed of a one-dimensional convolutional layer to extract the spatial features of the sequence. The outputs of the two channels are added together and fed into a three-layer fully connected layer to fuse the temporal and spatial features. The clock bias prediction module is used to input the normalized feature sequence into a trained dual-channel neural network model and output a predicted value; to denormalize the predicted value to obtain a denormalized predicted value; to recover the clock bias based on the denormalized predicted value to obtain a clock bias prediction value; and to denormalize the predicted value to obtain a denormalized predicted value, including: The predicted value is denormalized to obtain the denormalized predicted value. y reverse (i)=y pre (i)×(max(F_seq(i))-min(F_seq(i)))+min(F_seq(i)) Among them, y reverse (i) is the i-th inverse normalized prediction value, y pre (i) is the i-th nonnormalized prediction value of the network, min(F_seq(i)) represents the minimum value in the sequence, and max(F_seq(i)) represents the maximum value in the sequence.

Images