A rubidium atomic clock frequency adaptive prediction method and system based on an LSTM network

Abstract

This invention discloses a rubidium atomic clock frequency adaptive prediction method and system based on LSTM network, belonging to the field of time and frequency technology. The invention includes the following steps: Step 1. Data acquisition; Step 2. Filtering and denoising preprocessing; Step 3. Normalization processing; Step 4. LSTM model training and deployment; Step 5. Frequency prediction; Step 6. Online model adaptation. This invention employs a Long Short-Term Memory (LSTM) network, a special type of recurrent neural network, which can effectively learn the complex time dependencies and noise patterns in rubidium atomic clock time and frequency data. The inherent gating mechanism of the LSTM model makes it adept at capturing long-term dependencies in time series, thus enabling high-precision prediction of short-term frequency fluctuations. Through an online fine-tuning mechanism, this invention allows the LSTM model to continuously adapt to the individual drift and aging characteristics of a single rubidium atomic clock, achieving precise optimization for each clock and model, significantly improving the practicality and long-term stability of the method.

Description

Technical Field

[0001] This invention belongs to the field of time and frequency technology, specifically a rubidium atomic clock frequency adaptive prediction method and system based on LSTM network, which is used to improve the short-term frequency stability of rubidium atomic clocks. Background Technology

[0002] As a high-precision frequency source, the core performance of a rubidium atomic clock is mainly measured by key indicators such as frequency stability, frequency accuracy, and frequency drift rate. Frequency stability can be further divided into long-term stability and short-term stability. Short-term frequency stability is mainly used to characterize the random fluctuation characteristics of the frequency within a time scale of seconds or even milliseconds. It is a core parameter for measuring the instantaneous signal purity and phase noise level of the atomic clock, and has a decisive impact on the system's high-precision real-time synchronization and signal processing capabilities.

[0003] Short-term frequency stability plays a crucial role in cutting-edge fields such as high-speed communication, precision radar, satellite navigation and timing, and deep space exploration. When systems have stringent requirements for instantaneous synchronization accuracy, low bit error rate communication, or high-resolution target detection, extremely high demands are placed on the short-term stability of rubidium atomic clocks. The quality of short-term stability directly determines the upper limit of the system's performance in combating phase jitter, suppressing adjacent channel interference, and achieving ultra-high precision measurements. Short-term frequency fluctuations in atomic clocks are the result of the combined effects of multiple internal noise sources, the most significant of which are white frequency modulation noise, flicker frequency modulation noise, and phase noise. These noises mainly originate from the quantum noise of the atomic clock's physical system, the electronic noise of the servo circuit, and the phase noise of the local oscillator itself. Short-term frequency stability is a key indicator for evaluating the rapid response and high-precision instantaneous synchronization capabilities of rubidium atomic clocks. Therefore, optimizing and improving short-term stability is a core aspect of rubidium atomic clock development, especially for modern cutting-edge applications, and it is essential to optimize it to its optimal level.

[0004] To improve short-term stability, existing technologies typically optimize rubidium atomic clocks through the following two approaches: 1. Optimize the performance of voltage-controlled crystal oscillators by selecting crystal resonators with lower phase noise and higher Q values, and improving the design of their oscillation circuits to reduce noise introduced by the local oscillator; 2. Optimize the frequency-locked servo loop by carefully designing the parameters of the loop filter to maximize the servo bandwidth while ensuring loop stability, so as to quickly suppress short-term fluctuations of the crystal oscillator.

[0005] However, the above methods face significant bottlenecks: on the one hand, the development cost of high-precision crystal oscillators with Allan variance less than 1e-12 is high and the technology is complex; on the other hand, the expansion of servo bandwidth is limited by the signal response speed and signal-to-noise ratio of the atomic system itself, and excessively increasing the bandwidth will introduce more atomic shot noise, leading to a deterioration in short-term stability within a specific time interval. Furthermore, considering that the production process of rubidium atomic clocks involves a large number of manual adjustment steps, such as filling the absorber, tuning the microwave cavity, and adjusting the uniformity of the C-field, the discreteness of these processes directly leads to differences in the physical signal strength, signal-to-noise ratio, and servo characteristics of each rubidium atomic clock. Therefore, a new method is urgently needed that can adaptively learn and predict the frequency fluctuation patterns of a single rubidium atomic clock, thereby effectively improving its short-term stability.

[0006] The information disclosed above in this background section is only for enhancing the understanding of the background section of this invention, and therefore may include prior art that is not known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a rubidium atomic clock frequency adaptive prediction method and system based on LSTM network, so as to solve the technical problem that existing technologies are unable to effectively and individually improve the short-term frequency stability of rubidium atomic clocks.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a rubidium atomic clock frequency adaptive prediction method based on LSTM network, comprising the following steps:

[0010] Step 1. Collect frequency difference sequence data of the rubidium atomic clock within the historical time window;

[0011] Step 2. Perform filtering and noise reduction preprocessing on the frequency difference sequence data;

[0012] Step 3. Normalize the preprocessed data to construct a sample set suitable for model training;

[0013] Step 4. Input the constructed sample set into the pre-trained LSTM model, and the LSTM model outputs the frequency prediction values ​​for future time points;

[0014] Step 5. Perform inverse normalization on the predicted values;

[0015] Step 6. Generate a time-frequency signal based on the predicted frequency, which has a better short-term frequency stability index.

[0016] Furthermore, in step 2, the preprocessing of the frequency difference sequence data specifically includes: using a low-pass filter to remove high-frequency noise from the original frequency difference sequence; in step 3, the normalization adopts a relative benchmark normalization method: a normalization method based on the relative rate of change of the first value of the window, X normalized = (X / X first -1, window size is 60.

[0017] Furthermore, the training process of the LSTM model includes: collecting long-period frequency difference data from multiple rubidium atomic clocks as a training dataset; constructing a neural network structure containing an input layer, at least one LSTM hidden layer, and an output layer; and training the network using the mean square error between the predicted frequency value and the actual frequency value as the loss function until the model converges.

[0018] Furthermore, it also includes an online model adaptation step: during actual operation, the frequency difference data of the rubidium atomic clock is continuously collected and used to fine-tune the LSTM model online to adapt to the slow changes in the characteristics of the atomic clock.

[0019] Secondly, this invention provides a rubidium atomic clock frequency adaptive prediction system based on an LSTM network, the system comprising:

[0020] The data acquisition and preprocessing module is used to acquire and preprocess frequency difference sequence data. The data acquisition and preprocessing module includes a frequency difference measurement module, a frequency difference generation module, and a frequency difference acquisition module.

[0021] The LSTM prediction module is used to load the trained LSTM model and predict the frequency value of the rubidium atomic clock output signal at the next moment based on the input frequency difference data.

[0022] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method steps described above.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. This invention employs a special recurrent neural network called Long Short-Term Memory (LSTM) to effectively learn the complex time dependencies and noise patterns in rubidium atomic clock time-frequency data. The inherent gating mechanism of the LSTM model makes it adept at capturing long-term dependencies in time series, thereby enabling high-precision prediction of short-term frequency fluctuations.

[0025] 2. This invention enables the LSTM model to continuously adapt to the individual drift and aging characteristics of a single rubidium atomic clock through an online fine-tuning mechanism, achieving precise optimization of one model per clock and significantly improving the practicality and long-term stability of the method. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0027] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0028] Figure 2 A comparison chart showing the typical Allan variance curve (True Data) and the Allan variance curve (Predicted Data) after LSTM prediction compensation for the rubidium atomic clock 01;

[0029] Figure 3 This is a comparison chart of the typical Allan variance curve (True Data) and the Allan variance curve (Predicted Data) after LSTM prediction compensation for the rubidium atomic clock 02. Detailed Implementation

[0030] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] Please see Figure 1 The present invention provides a rubidium atomic clock frequency adaptive prediction method based on LSTM network, comprising the following steps:

[0032] Step 1: Data Acquisition and Preprocessing

[0033] After the rubidium atomic clock is powered on and stabilized, a high-precision frequency difference measurement module continuously collects the frequency difference data between its output frequency and the reference frequency, forming a time series {x1, x2, ..., x...}. t}

[0034] First, Kalman filtering is applied to the original sequence to initially suppress high-frequency measurement noise. Then, normalization is performed using a relative benchmark normalization method: a normalization method based on the relative rate of change of the first value of the window, X. normalized = (X / X firstThe window size is set to 60, and the normalized sequence is reconstructed into a sample format [X, y], where X is the historical time window data of length n, and y is the data of one or more subsequent future time points.

[0035] Step 2: LSTM Model Training and Deployment

[0036] The output signal of a rubidium atomic clock mainly contains the following noise components: white frequency modulation noise, flicker frequency modulation noise, and phase noise. These noises correspond to the slope characteristics of different time constants τ in the Allan variance curve. Based on these noise characteristics, the LSTM model of this invention pays special attention to modeling short-term fluctuations, i.e., τ = 1 to 100 seconds. Specifically, by analyzing the noise power spectral density and time correlation characteristics, the historical window length, hidden layer dimension, and training strategy of the LSTM model are determined, enabling it to effectively capture non-stationary noise components and long-term drift trends in the signal, thereby achieving high-precision frequency prediction and improved stability.

[0037] A large amount of long-period frequency difference data from multiple rubidium atomic clocks was collected as the training set. The frequency difference data from each rubidium atomic clock exceeded 100,000 data points, with a time resolution of 1 second. 60% of the data was used as the training set, and 40% as the test set. An LSTM network was constructed, with the following structure:

[0038] An input layer receives n historical points;

[0039] Two stacked LSTM hidden layers, using the tanh activation function;

[0040] A Dropout layer to prevent overfitting;

[0041] A fully connected output layer that outputs the predicted values.

[0042] The LSTM model consists of two neural networks, each with 128 neurons, trained for 100 epochs with a learning rate of 0.001. It is trained using the Adam optimizer with mean squared error (MSE) as the loss function. The trained LSTM model parameters are then loaded into the processor's memory.

[0043] Step 3: Frequency Prediction

[0044] When the system is running in real time, the latest time window data X after preprocessing in step 1 will be used. new Input to the model. The LSTM model for X. new Perform forward computation and output the frequency prediction value y for future time periods. pred This prediction essentially incorporates the current drift and noise trends of atomic clocks.

[0045] Step 4: Post-prediction processing

[0046] For y pred The signal is then denormalized to convert it into an actual frequency correction value. A time-frequency signal is generated based on the predicted frequency, which exhibits better short-term frequency stability.

[0047] Step 5: Online Model Adaptation

[0048] To address the long-term, slow changes in the characteristics of atomic clocks, the system periodically stores the latest frequency difference data in a buffer and uses it as a new sample to fine-tune the deployed LSTM model online, ensuring that its predictive ability always matches the current state of the atomic clock.

[0049] Example

[0050] Two commercial rubidium atomic clocks, 01 and 02, were used as the rubidium atomic clocks to be optimized, and one hydrogen clock was used as the reference clock. A time interval analyzer was used to collect the frequency difference between the two clocks and the reference clock. The data acquisition frequency was 1 Hz, and 100,000 data points were continuously collected. The first 60,000 data points were used for training, and the last 40,000 data points were used for testing.

[0051] The implementation steps are as follows:

[0052] Step 1: Data Collection

[0053] The frequency difference between the output of the rubidium atomic clock and the reference clock is collected using a time interval analyzer. A continuous data segment is extracted from the training set to form a sequence {x1, x2, ..., x...}. T}, used for subsequent sample construction, T=60000.

[0054] Step 2: Preprocessing

[0055] High-frequency noise is removed using a first-order low-pass filter; a relative reference normalization method is employed: a normalization method based on the relative rate of change of the first value of the window, X. normalized = (X / X first )-1, window size is 60; construct supervised learning samples: historical window length n=60 seconds, predict the frequency value of the next second.

[0056] Step 3: LSTM Model Training

[0057] The structure of an LSTM network includes:

[0058] Input layer: 60 time steps

[0059] LSTM layer 1: 128 neurons, activation function tanh

[0060] Dropout layer: Drop rate 0.2

[0061] LSTM layer 2: 64 neurons

[0062] Fully connected output layer: 1 neuron

[0063] The training run consisted of 100 epochs with a batch size of 64. The Adam optimizer was used with a learning rate of 0.001 and mean squared error (MSE) as the loss function.

[0064] Step 4: Real-time prediction

[0065] Input the 60-second window of data collected in real time into the trained model, and output the frequency prediction value for the next second.

[0066] Step 5: Online Fine-tuning Mechanism

[0067] Every 12 hours of operation, the system automatically saves the data from the most recent 12 hours as a fine-tuning dataset. The fine-tuning learning rate is set to 0.0001, and the model is incrementally trained to continuously optimize its adaptability.

[0068] Please see Figure 2 The Allan variance comparison of rubidium atomic clock 01 is shown in the table below:

[0069] τ (seconds) True Data Predicted Data Improvement range 1 1.23e-11 3.51e-12 71% 10 4.90e-12 1.73e-12 65% 100 1.87e-12 1.65e-12 12% 1000 6.32e-12 6.39e-12 -1%

[0070] Please see Figure 3 The Allan variance comparison of the rubidium atomic clock 02 is shown in the table below:

[0071] τ (seconds) True Data Predicted Data Improvement range 1 8.17e-12 2.57e-12 69% 10 2.83e-12 1.11e-12 61% 100 9.49e-13 8.39e-13 12% 1000 6.31e-13 6.63e-13 -5%

[0072] As shown above, within the short-term intervals of τ=1 second and 10 seconds, the Allan variance values ​​of both rubidium atomic clocks show a significant decrease, indicating that the LSTM model has high prediction accuracy and can effectively suppress short-term frequency fluctuations. At τ=100 seconds, there is still some optimization effect, but the improvement is smaller. By applying the frequency values ​​predicted by the method of this invention to real-time compensation or signal generation, short-term fluctuations in the output frequency of the rubidium atomic clock can be effectively suppressed without changing the original hardware circuitry, significantly improving the second-level frequency stability, which is particularly suitable for applications requiring high instantaneous synchronization accuracy.

[0073] Furthermore, the rubidium atomic clock frequency adaptive prediction method based on LSTM networks proposed in this invention has universality in principle and architecture. It is not only applicable to rubidium atomic clocks, but also to other high-stability frequency sources such as quartz crystals and atomic clocks. The crystal output signal also contains phase noise and frequency drift. By adapting the corresponding noise characteristics and sampling parameters, this method can effectively predict and compensate for its frequency fluctuations, thereby improving short-term stability.

[0074] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A rubidium atomic clock frequency adaptive prediction method based on LSTM network, characterized in that, Includes the following steps: Step 1. Collect frequency difference sequence data of the rubidium atomic clock within the historical time window; Step 2. Perform filtering and noise reduction preprocessing on the frequency difference sequence data; Step 3. Normalize the preprocessed data to construct a sample set suitable for model training; Step 4. Input the constructed sample set into the pre-trained LSTM model, and the LSTM model outputs the frequency prediction values ​​for future time points; Step 5. Perform inverse normalization on the predicted values; Step 6. Generate a time-frequency signal based on the predicted frequency, which has a better short-term frequency stability index.

2. The adaptive prediction method for rubidium atomic clock frequency based on LSTM network according to claim 1, characterized in that: Step 2, the preprocessing of the frequency difference sequence data specifically includes: using a low-pass filter to remove high-frequency noise from the original frequency difference sequence; and step 3, the normalization adopts a relative benchmark normalization method: a normalization method based on the relative rate of change of the first value of the window. The window size is 60.

3. The adaptive prediction method for rubidium atomic clock frequency based on LSTM network according to claim 1 or 2, characterized in that: The training process of the LSTM model includes: collecting long-period frequency difference data from multiple rubidium atomic clocks as a training dataset; constructing a neural network structure containing an input layer, at least one LSTM hidden layer, and an output layer; and training the network with the mean square error between the predicted frequency value and the actual frequency value as the loss function until the model converges.

4. The adaptive prediction method for rubidium atomic clock frequency based on LSTM network according to claim 3, characterized in that: It also includes an online model adaptation step: during actual operation, the frequency difference data of the rubidium atomic clock is continuously collected and used to fine-tune the LSTM model online to adapt to the slow changes in the characteristics of the atomic clock.

5. A rubidium atomic clock frequency adaptive prediction system based on LSTM network, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire and preprocess frequency difference sequence data. The data acquisition and preprocessing module includes a frequency difference measurement module, a frequency difference generation module, and a frequency difference acquisition module. The LSTM prediction module is used to load the trained LSTM model and predict the frequency value of the rubidium atomic clock output signal at the next moment based on the input frequency difference data.

6. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the method steps as described in any one of claims 1 to 4.

Images