A Comprehensive Atomic Time Calculation Method and System Based on LSTM Neural Network

Abstract

This application belongs to the technical field of time-frequency signals, and discloses a comprehensive atomic time calculation method and system based on an LSTM neural network. By obtaining the measured clock difference data of each atomic clock, preprocessing the measured clock difference data to obtain the preprocessed clock difference data, and dividing the preprocessed clock difference data into training clock difference data and test clock difference data according to a preset ratio; using the training clock difference data to train the LSTM neural network, adjusting the network parameters to obtain a trained LSTM neural network; inputting the test clock difference data into the trained LSTM neural network to obtain the predicted clock difference sequence of each atomic clock, and then calculating the predicted clock difference data; calculating the weight value of each atomic clock, and based on its corresponding weight value, predicted clock difference data and measured clock difference data, calculating the comprehensive atomic time through the basic equation of the time scale of the weighted average algorithm. The stability performance of generating the comprehensive atomic time is improved.

Description

Technical Field

[0001] The present application relates to the technical field of time-frequency signals, and in particular, to a method and system for calculating integrated atomic time based on an LSTM neural network. Background Art

[0002] With the rapid development of science and technology, in recent years, the space information technology and national defense security have increasingly high requirements for the accuracy, stability, and reliability of high-precision time-frequency signals. High-precision time is becoming increasingly important in fields such as space information technology and national defense security. In practical applications, high-precision time can be obtained by an integrated atomic time algorithm. The integrated atomic time algorithm mainly includes a weighted average algorithm, a Kalman filtering algorithm, a wavelet decomposition algorithm, etc. Among them, the most widely used is the weighted average algorithm, including the ALGOS algorithm and the AT1 algorithm. The ALGOS algorithm in the weighted average algorithm uses the least squares method for clock error prediction, and the AT1 algorithm uses exponential filtering for clock error prediction. However, in the weighted average algorithm, due to different measurement errors and noises at different times being converted into estimation errors and noises in different states at corresponding times after the least squares calculation, and the states at different times are not related to each other and restricted, the clock error prediction results are rough and messy, resulting in poor stability of the final timekeeping system and affecting the performance of the finally generated integrated atomic time. Summary of the Invention

[0003] To this end, the present application provides a method for calculating integrated atomic time based on an LSTM neural network, which improves the stability performance of the obtained integrated atomic time.

[0004] In a first aspect, the present application provides a method for calculating integrated atomic time based on an LSTM neural network.

[0005] The present application is achieved through the following technical solutions:

[0006] A method for calculating integrated atomic time based on an LSTM neural network, the method comprising:

[0007] Obtain the measured clock error data of each atomic clock, preprocess the measured clock error data to obtain the preprocessed clock error data, and divide the preprocessed clock error data into training clock error data and test clock error data according to a preset ratio;

[0008] Use the training clock error data to train an LSTM neural network, adjust the network parameters, and obtain a trained LSTM neural network;

[0009] Input the test clock error data into the trained LSTM neural network to obtain a predicted clock error sequence for each atomic clock, and obtain the predicted clock error data for each atomic clock based on the predicted clock error sequence;

[0010] Calculate the weight value of each of the atomic clocks, and based on the weight value, the predicted clock error data, and the measured clock error data, calculate the combined atomic time through the basic equation of the time scale of the weighted average algorithm.

[0011] In a preferred example of the present application, it can be further set that the step of using the training clock error data to train the LSTM neural network and adjust the network parameters includes: using the training clock error data as the network input and the expected output of the LSTM neural network to obtain the actual output of the LSTM neural network, calculating the difference between the actual output and the expected output, and based on the difference, performing a first-order gradient optimization on the objective function through the Adam optimizer, and optimizing the hyperparameters of the LSTM neural network.

[0012] In a preferred example of the present application, it can be further set that the LSTM neural network includes: an input layer, a hidden layer, and an output layer; the input layer is used to perform normalization processing on the measured clock error data input to the LSTM neural network to obtain standard input data; the hidden layer performs data feature extraction on the standard input data through a deep recurrent neural network of the neurons of the LSTM neural network to obtain feature data of the standard input data; the output layer is used to output a predicted clock error sequence according to the feature data.

[0013] In a preferred example of the present application, it can be further set that the step of preprocessing the measured clock error data includes: performing a first-order difference operation on the measured clock error data to obtain first-order difference data, fitting the trend term of the first-order difference data by the least squares method, and removing the trend term from the first-order difference data to obtain preprocessed clock error data.

[0014] In a preferred example of the present application, it can be further set that the step of calculating the weight value of each of the atomic clocks is as follows:

[0015] Calculate the Allan variance of each of the atomic clocks, calculate the reciprocal of the Allan variance, and use the reciprocal of the Allan variance as the weight value of each of the atomic clocks, where the calculation formula of the Allan variance is:

[0016]

[0017] where represents the Allan variance, τ represents the smoothing time, usually an integer multiple of the measurement time interval, x(i + 2) is the time difference at the (i + 2)-th moment, x(i + 1) is the time difference at the (i + 1)-th moment, x(i) is the time difference at the i-th moment, and N represents the number of time difference data points.

[0018] In a second aspect, the present application provides a combined atomic time calculation system based on an LSTM neural network.

[0019] This application is realized through the following technical solutions:

[0020] An integrated atomic time calculation system based on an LSTM neural network, the system comprising:

[0021] A data acquisition module, configured to acquire the measured clock difference data of each atomic clock, preprocess the measured clock difference data to obtain preprocessed clock difference data, and divide the preprocessed clock difference data into training clock difference data and test clock difference data according to a preset ratio;

[0022] A model training module, configured to train an LSTM neural network using the training clock difference data, adjust network parameters, and obtain a trained LSTM neural network;

[0023] A clock difference prediction module, configured to input the test clock difference data into the trained LSTM neural network to obtain a predicted clock difference sequence for each atomic clock, and obtain predicted clock difference data for each atomic clock based on the predicted clock difference sequence;

[0024] A calculation module, configured to calculate the weight value of each atomic clock, and calculate the integrated atomic time through the time scale basic equation of the weighted average algorithm based on the weight value, the predicted clock difference data, and the measured clock difference data.

[0025] In a preferred example of this application, it can be further set that the system further comprises:

[0026] An evaluation module, configured to calculate the stability of the integrated atomic time and judge the performance of the integrated atomic time.

[0027] In a third aspect, this application provides a computer device.

[0028] This application is realized through the following technical solutions:

[0029] A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of any of the above integrated atomic time calculation methods based on an LSTM neural network are implemented.

[0030] In a fourth aspect, this application provides a computer-readable storage medium.

[0031] This application is realized through the following technical solutions:

[0032] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any one of the above-mentioned comprehensive atomic time calculation methods based on the LSTM neural network are implemented.

[0033] In summary, compared with the prior art, the beneficial effects brought by the technical solution provided by the embodiment of the present application at least include: obtaining measurement clock difference data of an atomic clock group through measurement, forecasting the measurement clock difference data through the LSTM neural network of machine learning to obtain forecast clock difference data; calculating the weight value of each atomic clock based on stability, and substituting the forecast clock difference data, measurement clock difference data and weight value into the basic equation of the time scale to obtain the comprehensive atomic time. The LSTM neural network uses the characteristics of the network structure to find the sequence correlation of the input measurement clock difference data, and uses its excellent performance in non-linear mapping ability to avoid the problem of rough and messy forecast results caused by using the least squares method for clock difference forecasting in the traditional weighted average algorithm, and further improves the clock difference forecasting accuracy. Substituting the clock difference forecast result into the basic equation of the weighted average algorithm time scale improves the stability performance of generating the comprehensive atomic time. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a schematic flowchart of a comprehensive atomic time calculation method based on an LSTM neural network provided by an exemplary embodiment of the present application;

[0035] Figure 2 is a schematic diagram of a comprehensive atomic time calculation system based on an LSTM neural network provided by another exemplary embodiment of the present application. DETAILED DESCRIPTION

[0036] This specific embodiment is only an explanation of the present application, and it does not limit the present application. Those skilled in the art can make modifications without creative contributions to this embodiment according to needs after reading this specification, but as long as they are within the scope of the claims of the present application, they are protected by the patent law.

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without making creative efforts fall within the scope of protection of the present application.

[0038] In addition, the term "and / or" in this application merely describes the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in this application, the character " / " generally represents an "or" relationship between the front and rear associated objects unless otherwise specified.

[0039] In this application, terms such as "first" and "second" are used to distinguish identical or similar items with basically the same functions and effects. It should be understood that there is no logical or temporal dependence relationship between "first", "second", and "nth", nor are the quantity and execution order limited.

[0040] Currently, the main comprehensive atomic time algorithms used at home and abroad are: weighted average algorithms (including the ALGOS algorithm and the AT1 algorithm), the Kalman filtering algorithm, and the wavelet decomposition algorithm. The weighted average algorithm is a classic combined clock time scale algorithm. The Kalman filtering algorithm and the wavelet decomposition algorithm can be used not only for state estimation of a single atomic clock but also for establishing a time scale of a combined clock. Specifically, the weighted average algorithm obtains the comprehensive atomic time by weighted averaging each clock in the clock group. Typical weighted average algorithms include the ALGOS algorithm and the AT1 algorithm. The ALGOS algorithm calculates the weight of each clock based on the frequency stability of each atomic clock; the AT1 algorithm calculates the weight of each clock based on the prediction error of the atomic clock, and the prediction error refers to the absolute deviation between the predicted clock difference of each clock and the master clock and the actual clock difference. However, in the weighted average algorithm, due to different measurement errors and noises at different times being converted into estimation errors and noises in different states at corresponding times after the least squares method calculation, the states at different times are not connected and restricted with each other, resulting in rough and messy clock difference prediction results, poor stability of the finally calculated comprehensive atomic time, and affecting the performance of the finally generated comprehensive atomic time. This application utilizes the excellent clock difference prediction performance of the LSTM neural network to improve the stability performance of the weighted average comprehensive atomic time.

[0041] The following further describes the embodiments of this application in detail with reference to the accompanying drawings of the specification.

[0042] In an embodiment of this application, a comprehensive atomic time calculation method based on an LSTM neural network is provided, as Figure 1 shown, and the main steps are described as follows:

[0043] S10: Obtain the measured clock difference data of each atomic clock, preprocess the measured clock difference data to obtain the preprocessed clock difference data, and divide the preprocessed clock difference data into training clock difference data and test clock difference data according to a preset ratio.

[0044] Specifically, select an atomic clock with good performance from the atomic clock group as the reference clock j. Measure the measured clock difference data of other atomic clocks i in the clock group relative to the reference clock j through a time interval counter. The number of atomic clocks in the atomic clock group is at least 3. The measured clock difference data is expressed as:

[0045] x ij =h i -h j ,

[0046] where x ij represents the measured clock difference data between atomic clock i and reference clock j, h i represents the clock face reading of atomic clock i, and h j represents the clock face reading of reference clock j. After obtaining the measured clock difference data, for the convenience of subsequent prediction of clock difference data, it is necessary to preprocess the measured clock difference data to obtain the preprocessed clock difference data, and divide the preprocessed clock difference data into two parts according to a preset ratio. Among them, 80% is used as training clock difference data to adjust the network structure parameters and reduce errors; the remaining 20% is used as test clock difference data.

[0047] Preferably, the steps for preprocessing the measured clock difference data are as follows:

[0048] Perform a first-order difference operation on the measured clock difference data to obtain first-order difference data, reduce the number of significant digits of the data, further fit the trend term of the first-order difference data by the least squares method, and then remove the trend term from the first-order difference data. The remaining random term data in the first-order difference data is used as the preprocessed clock difference data. Deducting the trend term in the measured clock difference data in advance is beneficial to improving the prediction accuracy of the subsequent model.

[0049] Let X = {x i , i = 1, 2,..., N} be the measured clock difference data, and the subscript i represents the time corresponding to the clock difference data. The first-order difference data of the measured clock difference data is expressed as ΔX = {Δx i , i = 2,..., N}, Δx i =x i -x i-1 . The first-order difference data of the measured clock difference data after deducting the trend term is expressed as ΔX' = {Δx' i , i = 2,..., N}, Δx' i =Δx i -t i , where t i is the trend term at time i. The data input into the neural network is ΔX'.

[0050] S20: Use the training clock difference data to train the LSTM neural network, adjust the network parameters, and obtain the trained LSTM neural network.

[0051] Specifically, the LSTM neural network includes an input layer, a hidden layer, and an output layer. The input layer is used to standardize the measured clock error data input into the LSTM neural network to obtain standard input data, so that the input clock error time series meets the input requirements of the LSTM neurons. The hidden layer extracts data features from the standard input data based on the deep recurrent neural network of the neurons of the LSTM neural network to obtain the feature data of the above standard input data. The output layer is used to output the predicted clock error sequence according to the above feature data.

[0052] It should be noted that by introducing the "gate" mechanism, the LSTM neural network can make the network have stronger storage capabilities and obtain better results in longer sequences. The hidden layer of the LSTM neural network is composed of neurons. The three gates of the neurons are: the forget gate, the input gate, and the output gate. The forget gate determines how much of the previous cell state is forgotten or retained. The input gate determines how much of the current input is input into the current cell state. The output gate determines how much of the current cell state is input into the output value of the current moment.

[0053] Among them, the expression of the input gate is:

[0054] i t = σ(W i · [h t-1 , x t + b i ),

[0055]

[0056] Among them, i t represents the input gate at time t, represents the selection gate at time t, σ represents the sigmoid activation function, W i represents the weight matrix of the input gate, W c represents the weight matrix of the selection gate, h t-1 represents the hidden state at time t-1, x t represents the input at time t, b i represents the bias of the input gate, b c represents the bias of the selection gate.

[0057] The expression of the forget gate is:

[0058] f t = σ(W f · [h t-1 , x t + b f );

[0059] Among them, f t represents the forget gate at time t, Wf The weight matrix representing the forget gate, b f represents the bias of the forget gate.

[0060] The expression for the output gate is:

[0061] o t = σ(W o · [h t-1 , x t + b o ),

[0062] h t = o t * tanh(C t );

[0063] where, o t represents the output gate at time t, h t represents the hidden state at time t, and C t represents the cell state at time t.

[0064] And the expression for the cell state is:

[0065]

[0066] where, C t-1 represents the cell state at time t - 1.

[0067] When performing atomic clock frequency offset prediction based on the LSTM neural network, the process includes two parts: training and prediction of frequency offset data. Preferably, the steps of training the LSTM neural network using the training frequency offset data and adjusting the network parameters are as follows: taking the training frequency offset data as the network input and expected output of the LSTM neural network, obtaining the actual output of the LSTM neural network, calculating the difference between the actual output and the expected output, performing first-order gradient optimization on the objective function based on the difference through the Adam optimizer, and optimizing the hyperparameters of the LSTM neural network. The hyperparameters to be optimized include: learning rate, number of iterations epoch, batch size batch-size, etc. Through training, the LSTM neural network masters the data mapping relationship, makes the LSTM neural network model converge, and obtains the trained LSTM neural network.

[0068] Using the Adam optimization method has a relatively low memory requirement for calculations and can effectively improve the training and prediction efficiency of the LSTM neural network. Specifically, the Adam optimization method is as follows:

[0069] Determine the step size α, the first-order exponential decay rate β1, the second-order exponential decay rate β2, and the stochastic objective function f(θ), where the values of β1 and β2 are in the range [0, 1);

[0070] Initialize the parameter vector θ0, the first moment vector m0, the second moment vector v0, and the time step t;

[0071] When the parameter θ has not converged, update each part iteratively in a loop, increment the time step t by 1 to obtain the current time step, and update the gradient g of the objective function with respect to the parameter θ at the current time step t , update the first moment estimate m of the bias t and the second raw moment estimate v t , then calculate the bias-corrected first moment estimate and the bias-corrected second moment estimate, and then update the parameters θ of the network model with the values calculated above t .

[0072] S30: Input the test clock error data into the trained LSTM neural network to obtain the predicted clock error sequence for each of the atomic clocks, and obtain the predicted clock error data for each of the atomic clocks based on the predicted clock error sequence.

[0073] Specifically, input the test clock error data into the trained LSTM neural network. Through the input layer, and then extract data features through the hidden layer. Different hidden layer neurons correspond to the weights of the neurons in the input layer and their own biases. The excitation of the input layer is transmitted to the excitation of the hidden layer. Finally, the output layer outputs the feature data of the hidden layer as the predicted clock error sequence based on the weights of different hidden layers and its own biases.

[0074] The calculation steps for obtaining the predicted clock error data based on the predicted clock error sequence are as follows:

[0075] Let ΔX' predict ={Δx' i,predict , i = 3,..., N + 1} be the prediction result of the LSTM neural network, that is, the predicted clock error sequence. First, add the trend term deducted in the preprocessing stage of the measured clock error data to the predicted clock error sequence to obtain the first-order difference prediction result of the clock error: ΔX predict ={Δx i,predict , i = 3,..., N + 1}, where Δx i,predict =Δx' i,predict +t i . Then accumulate the first-order difference prediction result of the clock error to obtain the clock error prediction result: X predict ={x i,predict , i = 2,..., N + 1}, where

[0076] S40: Calculate the weight value of each atomic clock, and calculate the combined atomic time through the basic equation of the time scale of the weighted average algorithm based on the weight value, the predicted clock error data, and the measured clock error data. Considering the stability of each atomic clock when calculating the combined atomic time can improve the stability of the overall combined atomic time.

[0077] Preferably, the weight value of each atomic clock is calculated based on the Allan variance. Calculate the Allan variance of each atomic clock, and then further calculate the reciprocal of the Allan variance. The reciprocal of the Allan variance is used as the weight value of each atomic clock. The calculation formula of the Allan variance is as follows:

[0078]

[0079] Among them, represents the Allan variance, τ represents the smoothing time, usually an integer multiple of the measurement time interval, x(i + 2) is the time difference at the (i + 2)-th moment, x(i + 1) is the time difference at the (i + 1)-th moment, x(i) is the time difference at the i-th moment, and N represents the number of time difference data points.

[0080] The calculation formula of the weight value of each atomic clock is as follows:

[0081]

[0082] Among them, ω i is the weight value of atomic clock i, represents the stability of atomic clock i at the smoothing time of τ, represents the stability of atomic clock k at the smoothing time of τ, and M is the number of atomic clocks in the atomic clock group.

[0083] After calculating the weight value of each atomic clock, the measured clock difference data of each atomic clock measured previously and the predicted clock difference data of each clock obtained through the LSTM neural network are used to calculate the combined atomic time through the time scale basic equation of the weighted average algorithm. The specific calculation formula of the time scale basic equation is as follows:

[0084]

[0085] Among them, x j (t) represents the clock difference data of the reference clock j relative to the combined atomic clock, that is, the calculation result of the combined atomic time. ω i represents the weight value of atomic clock i, h′ i (t) represents the predicted clock difference data of atomic clock i, and x i,j (t) represents the measured clock difference data between atomic clock i and reference clock j obtained by measurement.

[0086] This application also provides a combined atomic time calculation system based on the LSTM neural network. As Figure 2 shown, the system includes: a data acquisition module 201, a model training module 202, a clock difference prediction module 203, and a calculation module 204.

[0087] Among them, the data acquisition module 201 is used to acquire the measured clock difference data of each atomic clock, preprocess the measured clock difference data to obtain the preprocessed clock difference data, and divide the preprocessed clock difference data into training clock difference data and test clock difference data according to a preset ratio;

[0088] The model training module 202 is used to train the LSTM neural network with the training clock difference data, adjust the network parameters, and obtain a trained LSTM neural network;

[0089] The clock difference prediction module 203 is used to input the test clock difference data into the trained LSTM neural network to obtain the predicted clock difference data of each atomic clock;

[0090] The calculation module 204 is used to calculate the weight value of each atomic clock, and based on the weight value, the predicted clock difference data and the measured clock difference data, calculate the integrated atomic time through the basic equation of the time scale of the weighted average algorithm.

[0091] Preferably, the system further includes an evaluation module 205 for calculating the stability of the integrated atomic time and judging the performance of the integrated atomic time.

[0092] In one embodiment, a computer device is provided, and the computer device may be a server.

[0093] The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium has an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements any one of the above integrated atomic time calculation methods based on the LSTM neural network.

[0094] In one embodiment, a computer-readable storage medium is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the computer program to implement any one of the above integrated atomic time calculation methods based on the LSTM neural network.

[0095] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink), DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0096] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the above division of each functional unit and module is used as an example. In actual applications, the above functions can be allocated to different functional units and modules as needed, that is, the internal structure of the system described in the present application can be divided into different functional units or modules to complete all or part of the functions described above.

Claims

1. A comprehensive atomic time calculation method based on LSTM neural network, characterized in that The method includes: Obtaining the measured clock offset data of each atomic clock, preprocessing the measured clock offset data to obtain the preprocessed clock offset data, and dividing the preprocessed clock offset data into training clock offset data and test clock offset data according to a preset ratio; Training the LSTM neural network using the training clock offset data, adjusting the network parameters, and obtaining the trained LSTM neural network; Inputting the test clock offset data into the trained LSTM neural network to obtain the predicted clock offset sequence of each atomic clock, and obtaining the predicted clock offset data of each atomic clock based on the predicted clock offset sequence; Calculating the weight value of each atomic clock, and calculating the combined atomic time through the basic equation of the time scale of the weighted average algorithm based on the weight value, the predicted clock offset data, and the measured clock offset data.

2. The integrated atomic time calculation method based on the LSTM neural network according to claim 1, characterized in that The step of training the LSTM neural network using the training clock offset data and adjusting the network parameters includes: Taking the training clock offset data as the network input and the expected output of the LSTM neural network to obtain the actual output of the LSTM neural network, calculating the difference between the actual output and the expected output, and performing first-order gradient optimization on the objective function through the Adam optimizer based on the difference to optimize the hyperparameters of the LSTM neural network.

3. The integrated atomic time calculation method based on the LSTM neural network according to claim 2, wherein The LSTM neural network includes an input layer, a hidden layer, and an output layer; the input layer is used to perform normalization processing on the measured clock offset data input to the LSTM neural network to obtain the standard input data; the hidden layer performs data feature extraction on the standard input data through a deep recurrent neural network of the neurons of the LSTM neural network to obtain the feature data of the standard input data; the output layer is used to output the predicted clock offset sequence according to the feature data.

4. The integrated atomic time calculation method based on the LSTM neural network according to claim 1, characterized in that The step of preprocessing the measured clock offset data includes: Performing a first-order difference operation on the measured clock offset data to obtain the first-order difference data, fitting the trend term of the first-order difference data by the least squares method, and removing the trend term from the first-order difference data to obtain the preprocessed clock offset data.

5. The integrated atomic time calculation method based on the LSTM neural network according to claim 1, wherein The step of calculating the weight value of each atomic clock is: Calculating the Allan variance of each atomic clock, calculating the reciprocal of the Allan variance, and taking the reciprocal of the Allan variance as the weight value of each atomic clock, where the calculation formula of the Allan variance is: Among them, represents the Allan variance, τ represents the smoothing time, which is usually an integer multiple of the measurement time interval, x(i + 2) is the time difference at the (i + 2)-th moment, x(i + 1) is the time difference at the (i + 1)-th moment, x(i) is the time difference at the i-th moment, and N represents the number of time difference data points.

6. An integrated atomic time calculation system based on an LSTM neural network, characterized in that, The system includes: A data acquisition module, which is used to obtain the measured clock offset data of each atomic clock, preprocess the measured clock offset data to obtain the preprocessed clock offset data, and divide the preprocessed clock offset data into training clock offset data and test clock offset data according to a preset ratio; A model training module, which is used to train the LSTM neural network using the training clock offset data, adjust the network parameters, and obtain the trained LSTM neural network; A clock offset prediction module, which is used to input the test clock offset data into the trained LSTM neural network to obtain the predicted clock offset sequence of each atomic clock, and obtain the predicted clock offset data of each atomic clock based on the predicted clock offset sequence; A calculation module for calculating the weight value of each of the atomic clocks, and calculating the combined atomic time through the basic equation of the time scale of the weighted average algorithm based on the weight value, the predicted clock error data, and the measured clock error data.

7. The integrated atomic time calculation system based on the LSTM neural network according to claim 6, wherein The system further includes: An evaluation module for calculating the stability of the combined atomic time and judging the performance of the combined atomic time.

8. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory, and the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images