Atomic clock aging prediction method based on feature fusion and quantum constraint learning

Abstract

The application relates to the technical field of precision timing instrument testing, in particular to an atomic clock aging prediction method based on feature fusion and quantum constraint learning. The atomic clock aging prediction method based on multi-modal feature fusion and quantum constraint learning provided by the application adopts a multi-modal tensor fusion means, simultaneously collects time domain microwave probe signals, space domain atomic cloud density distribution and frequency domain Ramsey fringe signals for feature extraction, and improves the completeness of the features; when a nonlinear model is established, Bloch equation constraint is adopted to ensure that the network model conforms to quantum physical laws. The method can improve the atomic clock aging feature expression effect through an artificial intelligence means, effectively avoids the overfitting defects caused by simply relying on training data, makes the atomic clock aging prediction model more conform to physical laws, and improves the atomic clock aging prediction precision and optimizes the prediction model complexity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of precision timing instrument testing, and particularly relates to an atomic clock aging prediction method based on feature fusion and quantum constraint learning. BACKGROUND

[0002] As a high-precision time and frequency reference, the long-term stability and life of an atomic clock directly affect the reliability of key fields such as navigation, communication and deep space exploration.

[0003] The existing aging state evaluation technology of an atomic clock (such as a rubidium or cesium atomic clock) mainly accumulates frequency drift data of the atomic clock through accelerated aging experiments or natural aging monitoring (6-12 months) to establish an empirical model; a statistical model (such as multiple linear regression or a support vector machine) or a physical model (such as a quantum noise theory model) is usually used to analyze the influence of environmental disturbances such as temperature and magnetic field on aging; and a threshold alarm (such as Allan variance overrun) or spectral analysis (FFT detects abnormal harmonics) is used to identify early frequency drift signs.

[0004] The traditional atomic clock aging state evaluation technology has the following three major pain points: (1) It relies on long-term physical experiments and has poor timeliness. The traditional method needs to collect frequency drift data through continuous aging experiments for 6-12 months, resulting in a long research and development cycle and high cost. In the design of a new type of atomic clock, repeated experiments seriously delay the iteration efficiency, and cannot meet the real-time life evaluation demand. (2) It is difficult to model the nonlinear relationship between the environment and aging. The aging of an atomic clock is affected by multiple parameters such as temperature, magnetic field and power supply noise, and its nonlinear dynamic characteristics are difficult to accurately characterize by a traditional mathematical model (such as linear regression or polynomial fitting). The existing method ignores the dynamic disturbance of the environment, resulting in a large long-term prediction error. (3) Data overfitting is a key challenge, which may cause the model to fail in actual application. An overfitting model performs well on training data, but the prediction error of new data is significantly increased. The aging data of an atomic clock usually has high noise and complex nonlinear characteristics, and overfitting the training data (such as short-term fluctuations or measurement noise) will cause the model to fail to capture the true long-term aging trend (such as frequency drift). Relying on an overfitting model may trigger unnecessary calibration or replacement operations, increasing the operation and maintenance cost. For example, the maintenance cost of a satellite atomic clock is extremely high, and a false prediction will directly affect the mission life. Overfitting may mask data quality problems (such as sensor errors and uneven sampling intervals), causing researchers to ignore the importance of data preprocessing. An overfitting model performs well on training data, but the prediction error of new data is significantly increased.

[0005] Therefore, the person skilled in the art is committed to developing an atomic clock aging prediction method that can model the nonlinear relationship between the environment and aging, and better represent the true physical degradation law (such as the frequency drift of a rubidium clock or the cavity aging of a hydrogen maser). SUMMARY

[0006] The application provides an atomic clock aging prediction method based on feature fusion and quantum constraint learning. Through a multi-modal feature fusion architecture, cross-dimensional correlation features are extracted, a bloch quantum physics constrained neural network is used, and an environment-aging nonlinear model is constructed to solve the technical problem of real-time atomic clock aging prediction.

[0007] The specific embodiments of the application will be introduced as follows.

[0008] The application provides an atomic clock aging prediction method based on feature fusion and quantum constraint learning. The steps include:

[0009] Synchronously collect time-domain microwave probe signals, spatial atomic cloud density distribution, and frequency-domain Ramsey fringe signals; construct a three-dimensional tensor input matrix X e R^(T x S x N), where T is the time step, S is the spatial grid point, and N is the frequency channel number;

[0010] A 3D convolutional neural network is used to extract deep features of the three-dimensional tensor, and a feature map F e R^(k x d) is output, where k is the time segment number and d is the feature dimension;

[0011] A double-channel GRU network is constructed, the main channel inputs the features F, and outputs the predicted frequency offset Δω; the auxiliary channel embeds the Bloch equation constraint, calculates the differential constraint term of the Hamiltonian H and the wave function ψ; and a hybrid loss function is defined:

[0012] wherein, L MSE is the mean square error term, is the Bloch equation constraint term, H is the Hamiltonian of the system, describing the internal energy level structure of the atomic clock and the action of the external field, is the reduced Planck constant; and α and β are weight parameters for balancing the contributions of the two loss terms. If β>α, the model will more strictly follow the physical law, but the fitting ability to data may be reduced, and vice versa, the model will be more dependent on data and may deviate from the physical constraint;

[0013] An environmental stress transfer function is established: dω / dt=K·∫[f(T(t))·g(P(t))·h(E(t))]dt+λ∫ω_hist(t-τ)dτ, wherein K is a stress coupling coefficient, λ is a historical aging memory factor, f(T(t)) is a temperature stress function, g(P(t)) is a pressure stress function, and h(E(t)) is an electromagnetic field stress function;

[0014] A lightweight model is deployed on an atomic clock edge computing unit to output the aging prediction value in real time; a dynamic calibration threshold is set, and when the predicted aging rate exceeds the dynamic calibration threshold, an active compensation signal is triggered to the servo system.

[0015] In a possible implementation, the time attention weight W_t, the space attention weight W_s and the frequency attention weight W_f are applied on the three-dimensional tensor input matrix in parallel; the fusion features are generated by cross-correlation calculation wherein σ is a Sigmoid activation function.

[0016] In a possible implementation, the time-domain microwave probe signal is collected in a specific time window of a microwave pulse using a lock-in amplifier or a high-speed ADC, and the signal is accumulated by multiple repeated pulse sequences to improve the signal-to-noise ratio.

[0017] In a possible implementation, the spatial domain atomic cloud density distribution is obtained by absorption imaging or fluorescence imaging.

[0018] In a possible implementation, the frequency domain Ramsey fringe signal is collected by precisely controlling the frequency of the microwave source to scan around the atomic transition frequency, and the step frequency interval needs to be less than the fringe width; after the microwave field is separated twice and interacts with the atoms, the proportion of atoms in the excited state is measured by a detection laser or an electronic method; the atomic state population probability is statistically counted by repeating the experiment multiple times for each scanning frequency point; the center frequency is extracted by fitting the Ramsey fringe.

[0019] In a possible implementation, the temperature stress function f(T(t)) adopts an Arrhenius model or a polynomial fitting, f(T(t))=exp(-Ea / kBT(t)), where Ea is an activation energy, kB is a Boltzmann constant, or wherein the coefficient a i The experimental data is calibrated.

[0020] In a possible implementation, the pressure stress function g(P(t))=b0+b1P(t)+b2P 2 (t).

[0021] In a possible implementation, the pressure stress function is simplified as a linear response according to the influence of pressure, g(P(t))=b0+b1P(t).

[0022] In a possible implementation, the electromagnetic field stress function h(E(t))=c0||E(t)||+c1||E(t)| 2 , wherein E(t) is an electromagnetic field strength.

[0023] In a possible implementation, the calibration threshold is 3σ historical mean.

[0024] This scheme avoids underfitting by leveraging the rich learning and data processing capabilities of the neural network nonlinear model, while also increasing detailed features by integrating spatiotemporal frequency feature types. Furthermore, it avoids overfitting by adding quantum constraint regularization terms to the neural network learning model, thus enabling the learning model to possess both data and physical interpretation capabilities.

[0025] The beneficial effects of this application are as follows: by extracting cross-domain correlations through multimodal spatiotemporal feature fusion and improving the generalization of the model using quantum physics-constrained neural networks, high-precision aging verification of atomic clocks can be achieved. This scheme solves the bottleneck problem of aging assessment of atomic sensors in complex environments through cross-modal correlation modeling and embedding of physical laws.

[0026] The beneficial effects of this application are also reflected in: multimodal tensor fusion improves feature completeness; Bloch equation constraints ensure that the network model conforms to the laws of quantum physics. Attached Figure Description

[0027] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0028] Figure 1 Flowchart of atomic clock aging prediction in existing technology;

[0029] Figure 2 The flowchart of atomic clock aging prediction based on multimodal feature fusion and quantum constraint learning used in Example 1 of this scheme;

[0030] Figure 3 : Schematic diagram of underfitting, best fitting, and overfitting in machine learning data;

[0031] Figure 4 A diagram illustrating the complexity of machine learning models when data is underfitted, best-fitted, or overfitted.

[0032] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0033] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0034] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the drawings.

[0035] Embodiment 1

[0036] In order to clearly understand the technical solutions of Embodiment 1 of the present application, the solutions of the prior art will be described in detail first.

[0037] The aging of an atomic clock is mainly manifested as a long-term drift of the output frequency, and the reasons include: slight changes in atomic transition frequency (such as energy level shift of cesium or rubidium atomic clock). External environmental disturbances (temperature, magnetic field fluctuations). Performance degradation of hardware components (such as oscillators, lasers).

[0038] Atomic clock aging prediction refers to predicting the performance changes of an atomic clock over time, especially its frequency stability and accuracy, through specific algorithms and techniques. The aging of an atomic clock is mainly due to physical changes of internal components and influences of external environment, resulting in frequency drift and precision decline.

[0039] Atomic clocks use the transition frequency between atomic energy levels for timing, and this frequency is extremely stable, providing an extremely precise standard for time measurement. The core of an atomic clock is a specific type of atom (such as cesium or rubidium), whose electrons will undergo transitions between different energy levels when subjected to electromagnetic radiation. By measuring this transition frequency, an atomic clock can achieve high-precision time measurement. However, over time, the internal components of an atomic clock (such as crystal oscillators, electronic components, etc.) will change, leading to a decrease in frequency stability and accuracy, which is the aging phenomenon of an atomic clock.

[0040] Deep neural networks are an effective prediction method. Deep neural networks can learn and predict complex nonlinear mapping relationships by simulating the way the human brain processes information. In atomic clock clock difference prediction, deep neural networks can capture deep features of atomic clock performance changes by analyzing historical data, thereby predicting future performance changes. In addition, the convolutional layer in deep neural networks reduces the number of model parameters through parameter sharing, improves computational efficiency, and further optimizes prediction accuracy.

[0041] Ramsey fringes are interference patterns generated by the technique of separated oscillatory field (two time-interval microwave pulses interact with atoms), which play key roles including: frequency sensitivity: the central peak of the fringe corresponds to the eigen transition frequency of the atom, frequency drift will directly manifest as the shift of the fringe center. high resolution: the narrow linewidth of the fringe (inversely proportional to the interaction time) can detect extremely small frequency changes (such as 10^-16 order of magnitude). aging indicator: long-term monitoring of the shift of the fringe center position can quantify the aging trend of the atomic clock.

[0042] Practical application effect: calibration reference: through Ramsey fringes, the frequency of the local oscillator is corrected in real time to compensate for aging drift. fault warning: distortion of the fringe shape (such as contrast reduction, linewidth change) may reflect hardware degradation (such as microwave power instability).

[0043] Ramsey fringes are both high-precision frequency scales and "diagnostic maps" of hardware status in atomic clock aging prediction. As AI features, their physical parameters (center frequency, linewidth, etc.) are transformed into quantifiable aging indicators through signal processing and machine learning, combining physical models and data-driven methods to achieve cross-scale prediction from microscopic quantum effects to macroscopic performance degradation.

[0044] In the world of machine learning and deep learning, we always hope that the model can not only fit the training data well, but also have excellent generalization ability for unknown data. However, reality is always full of challenges, and the model may fall into the dilemma of overfitting or underfitting.

[0045] Underfitting: when the model cannot capture the potential trend of the data. Underfitting means that the model performs poorly on the training set, validation set and test set. This usually means that the model does not have enough ability to capture the characteristics of the data, or the feature extraction of the training sample is not sufficient. Causes of underfitting: insufficient model complexity: the model does not have enough ability to fit the data. Insufficient feature extraction: the features of the training sample are extracted too little, resulting in the model being unable to match.

[0046] Overfitting: When a model overfits the training data. Overfitting is when a model performs well on the training set but poorly on the test set. This is usually because the model has overfit the training data, resulting in a decrease in the generalization ability to unknown data. Overfitting performance: Training error and test error start to separate after reaching a certain critical point. The model performs well on the training set, but poorly on the test set. Methods to solve overfitting: Increase training data: Although increasing data is usually helpful in solving underfitting, it has limited effect on overfitting. Reduce model complexity: For example, reduce the number of hidden layers and hidden units of neural network, or use simpler algorithm. Increase regularization constraints: Regularization is to prevent model overfitting, by increasing regularization parameter or introducing regularization term to limit the complexity of the model.

[0047] The existing atomic clock aging prediction technology currently adopts the process as shown in the figure, which is generally based on atomic clock frequency signal data collection, and predicts based on data through machine learning. The disadvantages are that on the one hand, the feature expression of the collected data is limited, and the detailed features that can be modeled and learned are not rich and comprehensive, and on the other hand, in the nonlinear modeling (usually a neural network model), the standard mean square error is often used to establish constraints, which may cause overfitting, resulting in complex model, inaccurate prediction, violation of physical laws, etc. Figure 1

[0048] Therefore, the embodiment 1 of the present application aims to provide an atomic clock aging prediction method based on multi-modal feature fusion and quantum constraint learning, which can avoid underfitting by the rich learning and data processing ability of the neural network nonlinear model, and increase the detailed features by fusing time, space and frequency feature types, and can avoid data overfitting by adding quantum constraint regularization term in the neural network learning model, so that the learning model has both data and physical interpretation ability. The method steps are as shown in the figure, which include: Figure 2

[0049] Step 1: Multi-modal data collection and three-dimensional tensor construction.

[0050] 1.1 Synchronously collect time domain microwave probe signal, space domain atomic cloud density distribution (obtained by absorption imaging) and frequency domain Ramsey fringe signal;

[0051] 1.2 Time alignment and normalization processing is performed on the three types of signals to construct a time-space-frequency three-dimensional tensor input matrix X∈R^(T×S×N), wherein T is the time step, S is the spatial grid point, and N is the frequency channel number.

[0052] Step 2: Cross-dimensional feature extraction.

[0053] 2.1 Design a time-space attention mechanism:

[0054] ​​The time attention weight W_t, the space attention weight W_s and the frequency attention weight W_f are applied in parallel on the tensor input;

[0055] The fusion feature is generated by cross-correlation calculation: where σ is the Sigmoid activation function;

[0056] 2.2 A 3D convolutional neural network is used to extract tensor deep features, and a feature map F ∈ R^(k×d) is output, where k is the number of time segments and d is the feature dimension.

[0057] Step 3: Time sequence modeling with quantum physics constraints.

[0058] 3.1 A dual-channel GRU network (Gated Recurrent Unit, a variant of recurrent neural network RNN) is constructed:

[0059] The main channel input feature F, and output the predicted frequency offset Δω;

[0060] The auxiliary channel embeds the Bloch equation constraint, and calculates the differential constraint term of the Hamiltonian H and the wave function ψ;

[0061] 3.2 Define the hybrid loss function:

[0062] where L MSE is the mean squared error term, which represents the mean squared error (Mean Squared Error) between the model prediction value and the true aging data, which is a standard loss term in supervised learning; is the Bloch equation constraint term, which requires the output of the neural network (wave function ψ or its derivative) to approximately satisfy the dynamic equation of the quantum system, is the gradient of the wave function with respect to time or parameters (depending on the input variable of the model), and H is the Hamiltonian of the system, which describes the internal energy level structure of the atomic clock and the action of the external field, is the reduced Planck constant; the weight parameters α and β are used to balance the contributions of the two loss terms. If β > α, the model will more strictly follow the physical law, but the fitting ability to data may be reduced, and vice versa.

[0063] Step 4: Establish the environmental stress transfer function:

[0064] dω / dt = K·∫[f(T(t))·g(P(t))·h(E(t))]dt + λ∫ω_hist(t-τ)dτ, where K is the stress coupling coefficient and λ is the historical aging memory factor.

[0065] Step 5: Online prediction and calibration.

[0066] 5.1 Deploy a lightweight model on the edge computing unit of the atomic clock to output aging prediction values ​​in real time;

[0067] 5.2 Set dynamic calibration threshold: When the predicted aging rate exceeds the 3σ historical average, trigger an active compensation signal to the servo system.

[0068] Example 1 overcomes the limitations of traditional single-mode monitoring by fusing spatiotemporal frequency features using three-dimensional tensors; it uses the fundamental equations of quantum mechanics as constraints for the neural network to ensure that the prediction results conform to physical laws.

[0069] In one possible improved embodiment of Example 1, the time-domain microwave probe signal is acquired using a lock-in amplifier or a high-speed ADC within a specific time window of the microwave pulse, and the signal-to-noise ratio is improved by accumulating multiple repeated pulse sequences.

[0070] In one possible improved embodiment of Example 1, the spatial atomic cloud density distribution is obtained using absorption imaging or fluorescence imaging.

[0071] In a possible improved embodiment of Example 1, the acquisition steps of the frequency domain Ramsey fringe signal are as follows: by precisely controlling the frequency of the microwave source, scanning is performed near the atomic transition frequency, and the step frequency interval must be less than the fringe width; after the microwave field interacts with the atoms in a secondary separation, the proportion of atoms in the excited state is measured by probe laser or electronic methods; the experiment is repeated multiple times for each scanning frequency point, and the probability of the atomic state population is statistically analyzed; Ramsey fringes are fitted, and the center frequency is extracted.

[0072] In a possible improved embodiment of Example 1, the temperature stress function f(T(t)) is fitted using an Arrhenius model or a polynomial, f(T(t)) = exp(-E a / k B T(t)), where E a The activation energy is given by kB, where kB is the Boltzmann constant, or Where the coefficient a i Calibration was performed using experimental data.

[0073] In a possible improved embodiment of Example 1, the pressure stress function g(P(t)) = b0 + b1P(t) + b2P 2 (t).

[0074] In a possible improved embodiment of Example 1, the pressure stress function is simplified to a linear response according to the influence of pressure, g(P(t))=b0+b1P(t).

[0075] In a possible improved embodiment of Example 1, the electromagnetic field stress function h(E(t)) = c0||E(t)|| + c1||E(t)||2 where E(t) is the electromagnetic field strength.

[0076] By the above embodiment 1 and each improved embodiment thereof, the multi-modal tensor fusion means is used to improve the completeness of the features; the Bloch equation constraint is used to ensure that the network model conforms to the quantum physics law; the method can improve the atomic clock aging feature expression effect through the artificial intelligence means and effectively avoid the overfitting shortcomings (such as Figure 3 indicated) caused by simply relying on the training data, so as to make the atomic clock aging prediction model more in line with the physical law (such as Figure 4 indicated) and thus avoid meaningless model over-complexity.

[0077] In the above embodiment 1 and each improved embodiment thereof, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments. Each technical feature of the above embodiments can be combined arbitrarily, and in order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope disclosed in the specification.

[0078] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The specification and examples given are intended as illustrative only and not limiting of the true scope and spirit of the application. What is claimed is:

[0079] It should be understood that the application is not limited to the precise construction that has been described above and shown in the accompanying drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the appended claims.

Claims

1. An atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning, characterized in that, Including the following steps: Simultaneously acquire time-domain microwave probe signals, spatial atomic cloud density distribution, and frequency-domain Ramsey fringe signals; construct a three-dimensional spatiotemporal frequency tensor input matrix X∈R^(T×S×N), where T is the time step, S is the spatial grid point, and N is the number of frequency channels; A 3D convolutional neural network is used to extract the deep features of the three-dimensional tensor, and the output feature map F∈R^(k×d) is generated, where k is the number of time segments and d is the feature dimension. A dual-channel GRU network is constructed. The main channel takes the feature F as input and outputs the predicted frequency offset Δω. The auxiliary channel is constrained by the Bloch equation, and the differential constraint terms of the Hamiltonian H and wave function ψ are calculated. A hybrid loss function is defined. Among them, L MSE It is the mean square error term. These are constraint terms in the Bloch equation, where H is the Hamiltonian of the system, describing the internal energy level structure of the atomic clock and the influence of external fields. α is the reduced Planck constant; α and β are weight parameters used to balance the contributions of the two losses. If β > α, the model will more strictly follow the physical laws, but reduce its ability to fit the data. Conversely, it will rely more on the data and deviate from the physical constraints. Establish the environmental stress transfer function: dω / dt=K·∫[f(T(t))·g(P(t))·h(E(t))]dt+λ∫ω_hist(t-τ)dτ, where K is the stress coupling coefficient, λ is the historical aging memory factor, f(T(t)) is the temperature stress function, g(P(t)) is the pressure stress function, and h(E(t)) is the electromagnetic field stress function; A lightweight model is deployed on the edge computing unit of the atomic clock to output aging prediction values ​​in real time; a dynamic calibration threshold is set, and when the predicted aging rate exceeds the dynamic calibration threshold, an active compensation signal is triggered to the servo system.

2. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to claim 1, characterized in that, Temporal attention weights W_t, spatial attention weights W_s, and frequency attention weights W_f are applied in parallel to the three-dimensional tensor input matrix; fused features are generated through cross-correlation calculation. Where σ is the Sigmoid activation function.

3. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to claim 1, characterized in that, The time-domain microwave probe signal is acquired using a lock-in amplifier or a high-speed ADC within a specific time window of the microwave pulse, and the signal-to-noise ratio is improved by accumulating multiple repeated pulse sequences.

4. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to claim 1, characterized in that, The spatial atomic cloud density distribution was obtained using absorption imaging or fluorescence imaging.

5. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to any one of claims 1-4, characterized in that, The acquisition steps of the frequency domain Ramsey fringe signal are as follows: by precisely controlling the frequency of the microwave source, scanning is performed near the atomic transition frequency, and the step frequency interval must be less than the fringe width; after the microwave field interacts with the atoms in the secondary separation, the proportion of atoms in the excited state is measured by probe laser or electronic methods; the experiment is repeated multiple times for each scanning frequency point, and the probability of the atomic state population is statistically analyzed; the Ramsey fringe is fitted, and the center frequency is extracted.

6. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to any one of claims 1-4, characterized in that, The temperature stress function f(T(t)) is fitted using an Arrhenius model or a polynomial, f(T(t)) = exp(-E a / k B T(t)), where E a To activate energy, k B Boltzmann constant, or Where the coefficient a i Calibration was performed using experimental data.

7. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to any one of claims 1-4, characterized in that, The pressure stress function g(P(t)) = b0 + b1P(t) + b2P 2 (t).

8. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to claim 7, characterized in that, The pressure stress function is simplified to a linear response based on the influence of pressure, g(P(t))=b0+b1P(t).

9. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to any one of claims 1-4, characterized in that, The electromagnetic field stress function h(E(t)) = c0‖E(t)‖ + c1‖E(t)‖ 2 , where E(t) is the electromagnetic field strength.

10. The atomic clock aging prediction method based on multimodal feature fusion and quantum constraint learning according to claim 1, characterized in that, The calibration threshold is the 3σ historical average.

Images