Pumping optical power error modeling and compensating method based on attention mechanism LSTM model
By employing an error modeling and compensation method based on an attention-mechanism LSTM model, the scaling factor drift and zero-bias instability issues caused by pump power fluctuations in the SERF inertial measurement system were resolved, thereby improving the long-term stability and accuracy of the system.
Patent Information
- Application Number
- CN202511094634.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-06
- Publication Date
- 2025-11-21
AI Technical Summary
In SERF inertial measurement systems, scale factor drift and zero bias instability caused by pump power fluctuations are difficult to resolve effectively, affecting the long-term stability and accuracy of the system.
An attention-based LSTM model is adopted, and the pump power is monitored non-contactly by a ring PD. A real-time estimation model is established by combining Gaussian beam theory. An LSTM deep neural network is constructed for error modeling and compensation, and the model prediction results are used for dynamic compensation.
It significantly suppressed the system output drift trend and improved the long-term operational stability and accuracy of the SERF inertial measurement system.
Smart Images
Figure CN120995077A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of SERF inertial measurement systems, and in particular to a pump power error modeling and compensation method based on an attention mechanism LSTM model, in order to solve the scale factor drift and zero bias instability problems caused by pump power fluctuations in SERF inertial measurement systems. Background Technology
[0002] SERF inertial measurement technology, with its extremely high sensitivity to low-frequency angular velocities, low noise, and zero-magnetic-field operation, has shown broad application prospects in recent years in next-generation high-precision atomic gyroscopes and inertial navigation systems. Its core principle is to utilize the dynamic evolution behavior of atomic spins in extremely low magnetic field environments to achieve high-resolution measurement of angular velocity or acceleration. In this type of system, the optical pumping process directly affects the atomic spin polarization efficiency and is a key factor influencing system performance; therefore, it places extremely high demands on the stability of the laser's output power and wavelength.
[0003] Specifically, laser power fluctuations not only lead to changes in pump rate but also affect electronic polarizability, longitudinal optical frequency shift, and spin relaxation behavior, thereby causing systematic errors such as scaling factor instability and zero-bias drift. These errors directly weaken the repeatability and long-term stability of measurements, becoming a significant bottleneck restricting the practical application of SERF inertial measurement systems. To suppress these uncertainties, traditional methods mainly rely on hardware means such as high-precision current sources and temperature control modules to maintain the constant output power of the laser under thermally stable operating conditions. However, these methods often struggle to cope with dynamic disturbances caused by device aging, optical coupling drift, or environmental temperature fluctuations. Therefore, recent research has shifted towards actively suppressing errors caused by laser power perturbations at the software level through system modeling and compensation strategies, thereby enhancing the robustness and long-term accuracy of the system.
[0004] In summary, effectively modeling and compensating for system output errors caused by pump power fluctuations is one of the key issues in improving the long-term stability and accuracy of SERF inertial measurement systems. To address this issue, this invention proposes a pump power error modeling and compensation method based on an attention-based LSTM model. This method predicts the pump power error and compares it with the true value of the system output signal, thereby achieving real-time compensation for system drift. This is of great significance for improving the long-term operational stability of SERF inertial measurement systems. Summary of the Invention
[0005] This invention addresses the shortcomings of existing technologies by proposing a pump power error modeling and compensation method based on an attention-based LSTM model to solve the scaling factor drift and zero-bias instability problems caused by pump power fluctuations in SERF inertial measurement systems. First, a real-time estimation model of the main pump power is established by non-contact monitoring of the pump laser power using a PD (Power Distribution Device) and incorporating Gaussian beam theory. Second, the system output signal and time-series data of the main pump power derived from the real-time estimation model are collected. A deep neural network error model is constructed using an attention-enhanced LSTM to achieve accurate modeling and dynamic prediction of the output signal. Finally, the output error is calculated based on the model prediction results to compensate the system output signal in real time. This method significantly suppresses the system output drift trend and can be used to improve the long-term operational stability of SERF inertial measurement systems.
[0006] The technical solution of the present invention is as follows:
[0007] A method for modeling and compensating pump optical power error based on an attention-based LSTM model, characterized by the following steps:
[0008] Step 1: Use a ring PD to monitor the pump optical power non-contactly;
[0009] Step 2: Establish a main path power estimation model based on Gaussian beam distribution theory;
[0010] Step 3: Synchronously acquire time-series data of optical power and system output signal in the ring area;
[0011] Step 4: Construct an LSTM deep neural network model based on the attention mechanism to model and predict the output signal;
[0012] Step 5: Calculate the output error using the model prediction results and dynamically compensate for the real-time output signal of the system.
[0013] Step 1 includes: In the SERF inertial measurement device, the pump light is used as the main laser to directly enter the atomic gas cell to achieve atomic spin polarization; a ring PD is introduced in the main pump light path, and the central beam is avoided through the hollow ring structure, allowing most of the laser to pass through without loss, while some edge light power is collected in the ring area, thereby realizing indirect monitoring of the pump light power.
[0014] Step 2 includes: determining the main pump power P based on the Gaussian beam intensity distribution law. pump With the optical power P in the ring region r The ratio is always:
[0015]
[0016] Step 3 includes: synchronously acquiring the optical power P in the ring region. r With system output signal V t The time-series data was used to obtain the main pump optical power P from the main power estimation model. pump The data is preprocessed, including cleaning, normalization, and filling in missing values. The missing values are filled using the nearest neighbor interpolation method to supplement the non-gyroscope signal data to be consistent with the gyroscope signal data, ensuring time series alignment.
[0017] Step 4 includes: constructing an LSTM deep neural network model based on an attention mechanism to model and predict the output signal, wherein the model input is the main path pump optical power P obtained based on the real-time estimation model. pump Time series data, output as a prediction system signal The model uses 10 seconds of optical power data to predict the gyroscope drift data for the following second. The LSTM model architecture includes: an input layer that receives preprocessed time-series data; an LSTM layer that captures long-term dependencies in the time series; an attention mechanism layer that dynamically assigns importance weights to features at different time steps; and an output layer that generates the prediction system signal. The model hyperparameters include: a learning rate of 0.01, a batch size of 512, and 200 iterations. It also employs Dropout regularization with a dropout ratio of 0.2.
[0018] Step 5 includes: calculating the output error using the model prediction results to dynamically compensate the real-time output signal of the system; the actual measured output signal of the system is denoted as V. t The estimation error caused by the power disturbance is:
[0019]
[0020] By using this error term ΔV t By introducing a feedback path for compensation, dynamic correction of the system output can be achieved.
[0021] The technical effects of this invention are as follows: This invention provides a pump power error modeling and compensation method based on an attention-mechanism LSTM model to solve the scaling factor drift and zero-bias instability problems caused by pump power fluctuations in SERF inertial measurement systems. First, the pump laser power is monitored non-contactly using a PD (Power Distribution Device), and a real-time estimation model of the main path pump power is established using Gaussian beam theory. Second, time-series data of the system output signal and pump power are collected, and a deep neural network error model is constructed using an attention-mechanism-enhanced LSTM to achieve accurate modeling and dynamic prediction of the output signal. Finally, the output error is calculated based on the model prediction results to compensate the system output signal in real time. This method significantly suppresses the system output drift trend and can be used to improve the long-term operational stability of SERF inertial measurement systems. Attached Figure Description
[0022] Figure 1 This is a schematic diagram illustrating the process of implementing the pump optical power error modeling and compensation method based on the attention mechanism LSTM model of the present invention. Figure 1 The process includes: Step 1, using a ring-shaped PD to non-contactly monitor pump optical power; Step 2, establishing a main path power estimation model based on Gaussian beam distribution theory; Step 3, synchronously collecting time-series data of optical power in the ring area and system output signal; Step 4, constructing an LSTM deep neural network model based on an attention mechanism to model and predict the output signal; Step 5, using the model prediction results to calculate the output error and dynamically compensate the real-time output signal of the system.
[0023] Figure 2 This is a schematic diagram of the pump optical path for implementing the pump optical power error modeling and compensation method based on an attention mechanism LSTM model of the present invention. The pump optical path includes a laser source, a polarizer, a liquid crystal, an analyzer, and a ring PD. The polarizer, liquid crystal, and analyzer constitute a controllable optical power attenuation module, and the ring PD is used to acquire the power signal that changes synchronously with the pump light.
[0024] Figure 3 This figure shows the experimental results of a pump optical power error modeling and compensation method based on an attention mechanism LSTM model according to the present invention. The blue curve represents the original signal output by the system, with the data corresponding to the left ordinate axis; the red curve represents the signal after experimental compensation, with the data corresponding to the right ordinate axis. Detailed Implementation
[0025] The following is in conjunction with the attached diagram ( Figures 1-3 The invention will be described in the following sections and examples.
[0026] Figure 1 This is a schematic diagram illustrating the process of implementing the pump optical power error modeling and compensation method based on the attention mechanism LSTM model of the present invention. Figure 2This is a schematic diagram of the pump optical path for implementing the pump optical power error modeling and compensation method based on the attention mechanism LSTM model of the present invention. Figure 3 The figure shows the experimental results of a pump optical power error modeling and compensation method based on an attention mechanism LSTM model according to the present invention. (Reference) Figures 1-3 As shown, a pump optical power error modeling and compensation method based on an attention mechanism LSTM model is characterized by the following steps: Step 1, using a ring PD for non-contact monitoring of pump optical power; Step 2, establishing a main path power estimation model based on Gaussian beam distribution theory; Step 3, synchronously acquiring time series data of optical power in the ring region and system output signal; Step 4, constructing an attention mechanism-based LSTM deep neural network model to model and predict the output signal; Step 5, using the model prediction results to calculate the output error and dynamically compensate the real-time output signal of the system.
[0027] In the SERF inertial measurement unit, the pump beam, as the main pump beam, directly enters the atomic gas cell to achieve atomic spin polarization. However, since the pump beam must act entirely within the gas cell, it cannot be directly extracted from the main path for real-time power monitoring. To address this issue, this invention introduces a ring-shaped power PD in the main pump beam path, as shown in the system structure below. Figure 2 As shown, the ring-shaped PD is installed in the pump optical main path. It avoids the central beam through a hollow ring structure, allowing most of the laser to pass through without loss. At the same time, it collects some edge optical power in the ring area, thereby realizing indirect monitoring of the pump optical power.
[0028] The pump laser has an approximate Gaussian distribution, and its intensity decays normally along the transverse direction, typically represented as a two-dimensional Gaussian function:
[0029]
[0030] Where I0 is the maximum light intensity at the beam center, (x0, y0) are the coordinates of the beam center, and σ is the standard deviation, which determines the beam width. The standard deviation σ represents the width of the beam distribution in space; a larger σ indicates a wider beam, and a smaller σ indicates a narrower beam.
[0031] In a Gaussian distribution, the standard deviation σ is directly related to the energy distribution of the light beam. The light intensity decays to its maximum value e at a distance σ from the beam center. -1 / 2 (Approximately 60.7%); at a distance of 2σ from the beam center, the light intensity decays to its maximum value e. -2 (Approximately 13.5%), the 2σ radius is often used to describe the effective width of a beam.
[0032] Light intensity represents the light power per unit area, while light power is the integral of light intensity over a specific area. In the case of a Gaussian beam, the total light power can be obtained by integrating the light intensity, as expressed by:
[0033]
[0034] The optical power P in the region of highest intensity (within the radius σ) σ =2πI0σ 2 (1-e -2 The optical power P of the Gaussian beam in the main energy region (within the 2σ radius) 2σ =2πI0σ 2 (1-e -1 / 2 According to the intensity distribution law of Gaussian beams, the main pump power is P. pump =2πI0σ 2 (1-e -2 The optical power P collected by the ring PD at the 2σ radius and the intermediate ring portion at the σ radius is... r =2πI0σ 2 (e -1 / 2 -e -2 Main pump optical power P pump , with the optical power P in the ring region r The ratio is always:
[0035]
[0036] Since the Gaussian distribution shape is fixed, the proportion of the total power in the ring region is constant. Therefore, the optical power signal P acquired by the ring PD is... r It can stably characterize the main path pump optical power P. pump .
[0037] Synchronous acquisition of optical power P in the ring area r , and the system output signal V t The time-series data was used to obtain the main pump optical power P from the main power estimation model. pump The data is preprocessed, including cleaning, normalization, and filling in missing values. The nearest neighbor interpolation method is used to fill in the missing values, supplementing the non-gyroscope signal data to be consistent with the gyroscope signal data to ensure time series alignment.
[0038] The long-term stability error characteristics of the SERF inertial measurement output signal caused by pump power error are complex, making it difficult to correct for errors by establishing an accurate nonlinear mathematical model. Since data-driven deep learning methods do not require prior models and have strong nonlinear fitting capabilities, this invention constructs an attention-based LSTM deep neural network model to model and predict the output signal. The model input is the main path pump power P obtained from a real-time estimation model. pump Time series data, output as a predictive system signal The model uses 10 seconds of optical power data to predict the gyroscope drift data for the following second. The LSTM model architecture includes: an input layer that receives preprocessed time-series data; an LSTM layer that captures long-term dependencies in the time series; an attention mechanism layer that dynamically assigns importance weights to features at different time steps; and an output layer that generates the prediction system signal. The model hyperparameters include: a learning rate of 0.01, a batch size of 512, and 200 iterations. It also employs Dropout regularization with a dropout ratio of 0.2.
[0039] The specific experimental procedure is divided into three stages, as shown below:
[0040] (1) Offline modeling stage: Simultaneously collect the optical power P of the annular region r , and the voltage value V of the output signal t The main pump optical power P is obtained from the main path power estimation model. pump The collected data is modeled to obtain... The model is then saved to the host computer.
[0041] (2) Online inference stage: Based on the collected optical power P in the annular region r The main pump optical power P is obtained from the main path power estimation model. pump Using model g(P) pump Calculate the pump optical power error model prediction value.
[0042] (3) Output compensation stage: The output error is calculated using the model prediction results to dynamically compensate the real-time output signal of the system. The actual measured output signal of the system is denoted as V. t The estimation error caused by the power disturbance is:
[0043]
[0044] By using this error term ΔV t By introducing a feedback path for compensation, dynamic correction of the system output can be achieved.
[0045] The results of the pump optical power error compensation experiment are as follows: Figure 3 As shown in the figure, the blue curve represents the original signal output by the system, with the data corresponding to the left ordinate axis; the red curve represents the signal after experimental compensation, with the data corresponding to the right ordinate axis. By comparison, the original signal shows a clear trend change, while the "drift" trend of the compensated signal is suppressed. Experimental results indicate that pump power error compensation improves the gyroscope drift range by 25.9%.
[0046] This method requires five steps to model and compensate for pump power error based on an attention-based LSTM model.
[0047] Step 1: Utilize a ring-shaped PD for non-contact monitoring of pump optical power;
[0048] In the SERF inertial measurement unit, the pump beam, as the main laser path, directly enters the atomic gas cell to achieve atomic spin polarization. A ring-shaped PD is introduced into the main pump beam path. The hollow ring structure avoids the central beam, allowing most of the laser to pass through without loss. At the same time, some edge light power is collected in the ring region, thereby achieving indirect monitoring of the pump beam power.
[0049] Step 2: Establish a main path power estimation model based on Gaussian beam distribution theory;
[0050] Based on the intensity distribution law of Gaussian beams, the main pump power P pump , with the optical power P in the ring region r The ratio of optical power to optical power is always:
[0051]
[0052] Step 3: Synchronously acquire time-series data of optical power and system output signal in the ring area;
[0053] Synchronous acquisition of optical power P in the ring area r , and the system output signal V d The time-series data was used to obtain the main pump optical power P from the main power estimation model. pump The data is preprocessed, including cleaning, normalization, and missing value imputation. Missing value imputation uses nearest neighbor interpolation to supplement non-gyroscope signal data to match the amount of gyroscope signal data, ensuring time series alignment.
[0054] Step 4: Construct an LSTM deep neural network model based on the attention mechanism to model and predict the output signal;
[0055] An attention-based LSTM deep neural network model is constructed to model and predict the output signal. The model input is the pump power data of the previous time window, and the output is the predicted system signal. The model uses 10 seconds of optical power data to predict the gyroscope drift data for the following second. The LSTM model architecture includes: an input layer that receives preprocessed time-series data; an LSTM layer that captures long-term dependencies in the time series; an attention mechanism layer that dynamically assigns importance weights to features at different time steps; and an output layer that generates the prediction system signal. The model hyperparameters include: a learning rate of 0.01, a batch size of 512, and 200 iterations. It also employs Dropout regularization with a dropout ratio of 0.2.
[0056] Step 5: Calculate the output error using the model prediction results and dynamically compensate for the real-time output signal of the system.
[0057] The actual output signal measured by the system is denoted as V. t The estimation error caused by the power disturbance is:
[0058]
[0059] By using this error term ΔV t By introducing a feedback path for compensation, dynamic correction of the system output can be achieved.
[0060] By modeling and compensating for pump power error based on an attention-based LSTM model, the system output drift trend is significantly suppressed, which can be used to improve the long-term operational stability of the SERF inertial measurement system.
[0061] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, and / or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.
Claims
1. A method for modeling and compensating pump optical power error based on an attention-based Long Short-Term Memory (LSTM) network model, characterized in that, Includes the following steps: Step 1: Use a ring PD to monitor the pump optical power non-contactly; Step 2: Establish a main path power estimation model based on Gaussian beam distribution theory; Step 3: Synchronously acquire time-series data of optical power and system output signal in the ring area; Step 4: Construct an LSTM deep neural network model based on the attention mechanism to model and predict the output signal; Step 5: Calculate the output error using the model prediction results and dynamically compensate for the real-time output signal of the system.
2. Step 1 includes: In a spin-free exchange relaxation (SERF) inertial measurement unit, the pump light is used as the main laser to directly enter the atomic gas cell to achieve atomic spin polarization. A ring-shaped PD is introduced into the pump optical main path. The hollow ring structure avoids the central beam, allowing most of the laser to pass through without loss. At the same time, some edge optical power is collected in the ring area, thereby realizing indirect monitoring of pump optical power.
3. Step 2 includes: Based on the intensity distribution law of Gaussian beams, the main pump power P pump With the optical power P in the ring region r The ratio is always:
4. Step 3 includes: Synchronous acquisition of optical power P in the ring area r With the system output signal V t The main pump optical power P is obtained from the time series data and the main power estimation model. pump The data is preprocessed, including cleaning, normalization, and filling in missing values. The missing values are filled using the nearest neighbor interpolation method to supplement the non-gyroscope signal data to be consistent with the gyroscope signal data, ensuring time series alignment.
5. Step 4 includes: An attention-based LSTM deep neural network model is constructed to model and predict the output signal. The model input is the main pump optical power P obtained from the real-time estimation model. pump Time series data, output as a predictive system signal The model uses 10 seconds of optical power data to predict the gyroscope drift data for the following second. The LSTM model architecture includes: an input layer that receives preprocessed time-series data; an LSTM layer that captures long-term dependencies in the time series; an attention mechanism layer that dynamically assigns importance weights to features at different time steps; and an output layer that generates the prediction system signal. The model hyperparameters include: a learning rate of 0.01, a batch size of 512, and 200 iterations. It also employs Dropout regularization with a dropout ratio of 0.
2.
6. Step 5 includes: The output error is calculated using the model prediction results to dynamically compensate for the real-time output signal of the system. The actual measured output signal of the system is denoted as V. t If and , then the estimation error caused by the power disturbance is: By using this error term ΔV t By introducing a feedback path for compensation, dynamic correction of the system output can be achieved.