Micro-thrust measurement ultra-low frequency drift compensation system and method based on trend prediction
Patent Information
- Application Number
- CN202511048121.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2025-07-28
- Filing Date
- 2025-07-29
- Publication Date
- 2026-09-22
- Estimated Expiration
- 2045-07-29
AI Technical Summary
[0005]本发明正是针对现有技术中微推力测量方法的不稳定性和误差较大的问题，提供一种基于趋势预测的微推力测量超低频漂移补偿系统和方法，使用微推力测量数据采集模块持续在施加待测微推力状态、无外力施加状态和施加标准力的三种状态中循环，并采集获取连续的微推力测量信号数据集；利用TVD滤波提取前后段数据骨架线消除噪声干扰，利用改进加权最小二乘拟合补偿中间段信号的漂移；通过信号分解来提取残余的超低频分量作为漂移预测数据集；建立长序列时间预测模型并通过模型预测漂移，通过误差修正实现对于预测误差的补偿，最终实现微推力测量超低频漂移补偿，显著提高微推力测量的可靠性及准确性
[0049]与现有技术相比，本发明具有的技术优势及技术效果是：本发明提供的一种基于趋势预测的微推力测量超低频漂移补偿系统及方法，综合考虑了预测方法和信号分解等方面，采用深度学习的方法，构建基于Mamba的长序列时间预测模型，充分提取超低频漂移特征，通过通道位置编码和Mamba编码解码，提高了漂移预测的效率和准确性，降低了超低频漂移对微推力测量的干扰。
Smart Images
Figure CN120949560B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of micro-thrust measurement, and mainly relates to ultra-low frequency drift compensation and drift compensation of a high-stability low-frequency micro-signal measurement system for micro-thrust measurement. Specifically, it relates to an ultra-low frequency drift compensation system and method for micro-thrust measurement based on trend prediction. Background Technology
[0002] In space gravitational wave detection, the satellite formation platform needs to be in an ultra-stable and ultra-quiet state, requiring drag-free control technology to overcome non-conservative force interference. As the micro-thrust compensation actuator for drag-free control technology, the micro-propulsion system should be capable of providing high-precision, low-noise micro-Newton-level thrust at the 0.1 mHz and 0.1 μN level. Therefore, the development and application of micro-thrust measurement methods and related measurement devices are of paramount importance.
[0003] Researchers both domestically and internationally have developed various micro-thrust measurement systems with different structural compositions. These systems employ direct or indirect measurement schemes for research. The research on micro-thrust measurement systems aligns with the development trend of the output thrust range of micro-thrusters, and can be divided into the following stages: the earliest micro-thrust measurement systems used balance and inverted pendulum structures, enabling thrust measurements at the mN-N level; single and double pendulum structures enabled thrust measurements at the μN-mN level; and torsion wire suspension and two-point support torsion pendulum structures enabled micro-thrust measurements at the mHz and μN levels.
[0004] However, environmental vibrations, temperature drift, structural creep, and the effects of personnel activity and airflow can all cause unpredictable drift in the measurement signal. In particular, ultra-low frequency drift accumulates continuously in thrust signal measurements, far exceeding the 0.1 μN level. This type of drift is difficult to address through material selection or structural optimization of the force measurement system, making it challenging to achieve ultra-low frequency sub-micro Newton level micro-thrust measurements at 0.1 mHz and 0.1 μN. Furthermore, conventional filtering algorithms cannot effectively remove ultra-low frequency drift in micro-thrust signals, potentially leading to step-edge distortion of the true thrust signal. Therefore, to mitigate measurement signal drift and ensure measurement accuracy, it is necessary to develop high-precision, high-stability, and high environmental reliability ultra-low frequency drift compensation methods for micro-thrust measurement systems from a drift compensation perspective. Summary of the Invention
[0005] This invention addresses the instability and large errors of existing micro-thrust measurement methods by providing a trend prediction-based ultra-low frequency drift compensation system and method for micro-thrust measurement. The system uses a micro-thrust measurement data acquisition module to continuously cycle through three states: applied micro-thrust, no external force applied, and applied standard force, acquiring a continuous dataset of micro-thrust measurement signals. TVD filtering is used to extract the skeleton lines of the preceding and following data segments to eliminate noise interference, and improved weighted least squares fitting is used to compensate for the drift in the middle segment of the signal. Residual ultra-low frequency components are extracted through signal decomposition as a drift prediction dataset. A long-sequence time prediction model is established and drift is predicted using the model. Error correction is used to compensate for the prediction error, ultimately achieving ultra-low frequency drift compensation for micro-thrust measurement and significantly improving the reliability and accuracy of micro-thrust measurement.
[0006] To achieve the above objectives, the technical solution adopted by this invention is: a micro-thrust measurement ultra-low frequency drift compensation system based on trend prediction, comprising at least a micro-thrust measurement data acquisition module, a nonlinear drift removal module, an ultra-low frequency component extraction module, a long-time series prediction module, a prediction error correction module, and a drift identification module, wherein:
[0007] The micro-thrust measurement data acquisition module is used to acquire micro-thrust measurement data under different conditions; the different conditions include at least three conditions: the state of applying the micro-thrust to be measured, the state of no external force applied, and the state of applying a standard force.
[0008] The nonlinear drift removal module: based on the TVD filter skeleton line of the two segments of data under different states, a fitting algorithm is used to fit and compensate for the low-frequency drift of the intermediate signal.
[0009] The ultra-low frequency component extraction module: performs signal decomposition on the residual signal after compensation of two data segments under different states using a signal decomposition method to extract the ultra-low frequency drift component;
[0010] The long-time series prediction module: predicts drift based on a long-time series prediction model to achieve ultra-low frequency drift prediction on a small scale;
[0011] The prediction error correction module is used to correct the errors in the output of the long-term series prediction module. The errors include, but are not limited to, system drift error and cumulative error from multiple prediction iterations.
[0012] The drift identification module: by performing time-frequency analysis on the compensated intermediate applied force data, it determines whether there is thrust drift in the thrust signal, thereby decoupling the background signal drift and the thrust signal drift.
[0013] As an improvement to the present invention, the fitting algorithm used in the nonlinear drift removal module is an improved weighted least squares fitting algorithm, the objective function of which is:
[0014]
[0015] in, It is a measurement signal. It is a design matrix. It is the estimated parameter vector. Indicates weight, The total number of observations;
[0016] Define one-dimensional micro-thrust data as The time interval it occupies is ; middle The value at time t is The intermediate data segment to be compensated Defined as:
[0017]
[0018] in Indicates the middle data segment The value is empty. express The time period in which it occurs;
[0019] Different weighting strategies are used for the two data segments, specifically:
[0020]
[0021]
[0022] in This indicates the weight of the preceding data points. This indicates the weight of the data points in the latter part. This represents the error observation data.
[0023] As an improvement of the present invention, the signal decomposition method in the ultra-low frequency component extraction module includes, but is not limited to, the signal decomposition method based on wavelet transform. The two segments of data after fitting compensation are decomposed respectively. Based on the principle of minimizing reconstruction error, Symlets are selected as wavelet basis functions. At this time, the ultra-low frequency drift component is represented as an approximate component, and the mid-to-high frequency noise component is represented as a detail component. The ultra-low frequency drift component is retained to extract the residual drift feature component, and the noise component is discarded.
[0024] As another improvement of the present invention, the long-series time prediction model in the long-series prediction module includes at least a Mamba encoder, a channel fusion location embedding, and a Mamba decoder, and includes the following steps:
[0025] S1, determine the time series data to be predicted, including multiple feature value inputs and feature value outputs;
[0026] S2, divide the dataset into training set, validation set, and test set;
[0027] S3, data standardization;
[0028] S4, Define the length of the history window and future prediction window length ;
[0029] S5, divide the input and output sequences into blocks respectively; for an input sequence... Divide into small blocks in order to obtain subsequences , Indicates the length of the subsequence. This represents the number of blocks in the input sequence; similarly, for the number of blocks in the output sequence, Indicates the length of the subsequence. This indicates the number of blocks in the output sequence.
[0030]
[0031] After dividing the input sequence into blocks, each subsequence Mapped through hidden layer to The linear transformation layer is defined as follows:
[0032]
[0033] in For a learnable parameter matrix, A learnable parameter vector;
[0034] S6, Linear Layer Output The input is fed into the Mamba encoder to extract sequence features; to mitigate information loss in the sequence, time series encoding and channel-independent encoding between variables are integrated to further enhance feature extraction.
[0035]
[0036] in , , This indicates the number of channels; the encoded output and positionally encoded output are fused using the concatenate operation, and then output to the Mamba decoder for decoding through a linear layer;
[0037] To obtain the predicted output sequence, the decoder output passes through a linear layer, and then is reshaped to restore it to the output dimension. ;
[0038] S7 adds RevIN before the network input and after the output for standardization, so as to achieve a consistent input and output distribution and solve the problem of data distribution drift.
[0039] As another improvement of the present invention, the system drift error in the prediction error correction module is removed by a first-order detrending algorithm. Linear regression analysis is performed on the data y(t) to fit a linear model y(t)=at+b, where a is the slope and b is the intercept. The fitted value y^(t)=at+b at each time point t is calculated. The fitted linear trend is subtracted from the original data to obtain the detrended data.
[0040] As another improvement of the present invention, in the drift identification module, a standard force is applied to the low-noise signal after fitting and compensation. By performing time-frequency analysis on the data of the intermediate applied force after prediction and compensation, the TVD algorithm is used to extract the stepped standard force after compensation. The force extracted by TVD is compared with the standard force, and the thrust is judged to contain drift based on the maximum amplitude error.
[0041] To achieve the above objectives, the present invention also adopts the following technical solution: a micro-thrust measurement ultra-low frequency drift compensation method based on trend prediction, comprising at least the following steps:
[0042] S1: Collect micro-thrust measurement data under different states, including at least the state of no actual external force applied, the state of applying actual external force and then removing the external force under the state of no actual external force applied, and the state of applying standard force and then removing the standard force under the state of no standard force applied. Record the start time and time interval of the latter two states.
[0043] S2: Based on the data collected in step S1, the latter two states include force-free data of equal length at the beginning and end and force-containing data in the middle. The TVD filtering algorithm is used to extract the filter skeleton line of the force-free data of equal length at the beginning and end, and the force-containing data in the middle is replaced with NAN. The TVD filter skeleton line of the beginning and end segments is used to input the data to remove nonlinear drift based on improved least squares fitting, and the whole segment of data after fitting compensation is obtained.
[0044] S3: The wavelet decomposition algorithm is used to decompose the data before and after fitting compensation. According to the principle of minimizing reconstruction error, Symlets is selected as the wavelet basis function and the order and decomposition level are determined. The ultra-low frequency drift component is retained to extract the residual drift feature component, thereby obtaining the prediction dataset. If the index requirements are met, the extraction of ultra-low frequency drift component is completed; otherwise, the order and decomposition level are re-determined, and the signal is decomposed again.
[0045] S4: A long-term series prediction neural network model based on Mamba, which is trained to output the prediction results of intermediate segment drift;
[0046] S5: Based on the drift prediction output of step S4, remove the system drift error by using a first-order detrending algorithm, and perform scaling error correction by manually inferring the accumulated error trajectory to remove the accumulated error from multiple iterations.
[0047] S6: Repeat steps S1-S5, output the compensated thrust signal, apply a standard force to the fitted compensated low-noise signal, and perform time-frequency analysis on the data of the intermediate applied force after prediction compensation to determine whether the thrust contains drift, thereby achieving decoupling between background signal drift and thrust signal drift.
[0048] As another improvement of the present invention, the parameter settings of the long-term sequence prediction neural network model based on Mamba in step S4 are as follows: Epochs is 50, Batch_size is 32, learning rate is 0.0001, convolutional kernel dimension is 3, early stopping is 10, subsequence length is 20, hidden layer state dimension D_model is 32, number of network layers is 64, state expansion factor is 2, and L1 paradigm loss function is used; the ratio of the training set, validation set and test set of the model is 6:2:2.
[0049] Compared with the prior art, the technical advantages and effects of this invention are as follows: This invention provides a trend prediction-based micro-thrust measurement ultra-low frequency drift compensation system and method, which comprehensively considers prediction methods and signal decomposition, and adopts deep learning methods to construct a long-sequence time prediction model based on Mamba, fully extracting ultra-low frequency drift features. Through channel position encoding and Mamba encoding and decoding, the efficiency and accuracy of drift prediction are improved, and the interference of ultra-low frequency drift on micro-thrust measurement is reduced. Attached Figure Description
[0050] Figure 1 This is a flowchart of the steps of the ultra-low frequency drift compensation method for micro-thrust measurement based on trend prediction of the present invention;
[0051] Figure 2 This is a flowchart of the wavelet decomposition algorithm used in step S3 of the method of the present invention.
[0052] Figure 3 This is a schematic diagram of the Mamba model structure used in step S4 of the method of the present invention;
[0053] Figure 4 This is a comparison diagram before and after compensation in Embodiment 1 of the present invention;
[0054] Figure 5 This is a comparison chart of standard thrust drift identification in Embodiment 1 of the present invention. Detailed Implementation
[0055] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0056] Example 1
[0057] A trend-prediction-based ultra-low frequency drift compensation system for micro-thrust measurement can compensate for ultra-low frequency drift in micro-thrust. It includes at least a micro-thrust measurement data acquisition module, a nonlinear drift correction module, an ultra-low frequency component extraction module, a long-time series prediction module, a prediction error correction module, and a drift identification module. Wherein:
[0058] The micro-thrust measurement data acquisition module is designed to generate data for micro-thrust measurement. The acquisition states can include three states: micro-thrust under test, no external force applied, and standard force applied. It acquires a continuous micro-thrust measurement signal dataset of more than 10,000 seconds. During the acquisition process, the background signal and the start time of applying micro-thrust / standard force need to be marked.
[0059] The nonlinear drift removal module uses TVD filter skeleton lines of equal length from the first and last data segments (excluding force) to fit and compensate for the low-frequency drift of the middle 10000s signal (including / excluding external force).
[0060] When adjusting the parameters of the TVD algorithm to extract the skeleton line, it indicates that the skeleton lines of the two segments only have ultra-low frequency drift and no low frequency noise. The fitting algorithm used is an improved weighted least squares fitting algorithm, which obtains the best drift compensation effect by minimizing the weighted residuals. The objective function expression obtained by the weighted least squares algorithm is:
[0061]
[0062] in, It is a measurement signal. It is a design matrix. It is the estimated parameter vector. Representing weights, the formula is equivalent to:
[0063] in These are error observation data.
[0064] Transform the formula into a matrix to obtain
[0065]
[0066] in for A diagonal matrix.
[0067] To minimize the objective function, we take the derivative and set it to zero to obtain the estimated parameter vector. To use the front and rear skeleton lines extracted by TVD filtering to fit and compensate for the intermediate data segments, the one-dimensional micro-thrust data is defined as follows: The time interval it occupies is ; middle The value at time t is The intermediate data segment to be compensated. Defined as
[0068]
[0069] Therefore, different weighting strategies should be used for the first and second parts. The weighting strategies are as follows:
[0070]
[0071]
[0072] in This indicates the weight of the preceding data points. This indicates the weight of the data points in the later part.
[0073] By initializing the data, the fitting error at each time step is calculated, and for the fitting error... Weighting, exist Arranged in ascending order and assigned corresponding... Weight, exist Arranged in descending order and assigned corresponding... Weights are applied to large errors and small errors to reduce fitting error, improve fitting accuracy, and ultimately achieve optimal compensation for the entire data segment.
[0074] The ultra-low frequency component extraction module uses a signal decomposition method to decompose the residual signals after compensation in the two segments and extract the ultra-low frequency components as the residual drift components.
[0075] The signal decomposition method in this embodiment is wavelet decomposition. The residual signal is decomposed, and the ultra-low frequency drift component is represented as an approximate component, while the mid-to-high frequency noise component is represented as a detail component. The ultra-low frequency drift component is retained to extract the residual drift feature component, while the noise component is discarded.
[0076] The long-term series prediction module uses a long-term time prediction model to predict drift and achieve ultra-low frequency drift prediction on a small scale.
[0077] The long-term series prediction module, within a Mamba-based long-term series prediction neural network model, includes a Mamba encoder, channel fusion location embedding, and a Mamba decoder, and comprises the following steps:
[0078] S1, determine the time series data to be predicted, including multiple feature value inputs and feature value outputs;
[0079] S2, divide the dataset into training set, validation set, and test set;
[0080] S3, data standardization;
[0081] S4, Define the length of the history window and future prediction window length ;
[0082] S5, divide the input and output sequences into blocks respectively; for an input sequence... Divide into small blocks in order to obtain subsequences , Indicates the length of the subsequence. This represents the number of blocks in the input sequence; similarly, for the number of blocks in the output sequence, Indicates the length of the subsequence. Indicates the number of blocks in the output sequence:
[0083]
[0084] After dividing the input sequence into blocks, each subsequence Mapped through hidden layer to The linear transformation layer is defined as follows:
[0085]
[0086] in For a learnable parameter matrix, A learnable parameter vector;
[0087] S6, Linear Layer Output The input is fed into the Mamba encoder to extract sequence features; to mitigate information loss in the sequence, time series encoding and channel-independent encoding between variables are integrated to further enhance feature extraction.
[0088]
[0089] in , , This indicates the number of channels; then, the encoded output and the positionally encoded output are fused using the concatenate operation, and then the output is passed through a linear layer to the Mamba decoder for decoding.
[0090] To obtain the predicted output sequence, the decoder output passes through a linear layer, and then is reshaped to restore it to the output dimension. ;
[0091] S7 adds RevIN before the network input and after the output for standardization, so as to achieve a consistent input and output distribution and solve the problem of data distribution drift.
[0092] Based on the prediction error correction module: the prediction error of the long sequence prediction model has error accumulation and system drift; the system drift is removed by the detrend algorithm, the error dataset is analyzed, and the scale error is corrected by manually inferring the error accumulation trajectory to obtain the corrected drift prediction result.
[0093] Drift identification module: It identifies the presence or absence of drift in the background signal data and identifies whether there is thrust drift in the background signals at both ends and the thrust applied signal in the middle, thereby decoupling the background signal drift and the thrust signal drift.
[0094] The ultra-low frequency drift compensation method for micro-thrust measurement based on trend prediction, using the system described above, can be applied to micro-thrust measurement systems of various structures, including balance structures, torsion pendulum structures, simple pendulum structures, and deformation structures. This application does not impose any restrictions on the specific type of measurement system. Figure 1 As shown, the specific steps include the following:
[0095] Step S1: Using the micro-thrust measurement data acquisition module, without considering the influence of temperature, continuously collect data for more than 8 hours under the following three conditions at a sampling frequency of 10Hz:
[0096] (1) No actual external force is applied;
[0097] (2) The state of no actual external force applied, followed by the state of applying actual external force, and then the state of removing actual external force;
[0098] (3) The state with no standard force applied, followed by the state with the standard force applied, and finally the state with the standard force removed;
[0099] In the actual data acquisition process, for states (2) and (3), it is necessary to mark the background signal and the starting time of the applied external force / standard force to distinguish the time intervals of applied force and unapplied force.
[0100] Step S2: Based on the collected measured data, select data from the three states collected above for a time of more than 10,000 seconds. In this embodiment, a data segment of 20,000 seconds is used as an example. States (2) and (3) include 5,000 seconds of data (excluding force) with equal length before and after, and 10,000 seconds of data (including force) in the middle.
[0101] The TVD filtering algorithm is used to extract the filter skeleton line of the data with equal lengths before and after, and the middle 10000s data is replaced with NAN. Then, the TVD filter skeleton line of the front and back segments is used to input the data into the improved least squares fitting to remove nonlinear drift, and the fitted and compensated 20000s data is obtained.
[0102] Step S3: The residual signals after compensation are decomposed using a signal decomposition method, and the ultra-low frequency component is extracted as the residual drift component. In this embodiment, the signal decomposition method used is a wavelet transform-based method. Figure 2 Here is a flowchart of the steps in the wavelet transform-based signal decomposition method:
[0103] First, analyze the noise spectral density function image of the 20,000 s data after fitting and compensation near 0.1 mHz to determine if it meets the requirements. If not, continue with the following steps: Use wavelet decomposition algorithm to decompose the 5,000 s data before and after fitting and compensation. Based on the principle of minimizing reconstruction error, select Symlets as the wavelet basis function and determine the order as 8 and the decomposition level as 4. At this time, the ultra-low frequency drift component appears as an approximate component, and the mid-high frequency noise component appears as a detail component. Retain the ultra-low frequency drift component to extract the residual drift feature component, and discard the noise component. Thus, the main component component below 0.1 Hz with the approximate component is obtained as the prediction dataset, and the remaining components and residual components are all in the mid-high frequency range (greater than 0.1 Hz). Otherwise, redetermine the order and the number of decomposition levels.
[0104] Step S4: Establish a Mamba-based neural network model. The network model structure is as follows: Figure 3 As shown. For the obtained front and back datasets, the back dataset needs to be flipped to train prediction of the middle data; then, the training set, validation set, and test set are divided according to a 6:2:2 ratio; the data is standardized by 0-1, and the dataset is constructed with the input format (Batch_size, L, Dimension) and the output format (Batch_size, H, Dimension), and an appropriate historical data length is selected. and the length of the predicted data ,here It is 400. The model's parameters are as follows: Epochs = 50, Batch size = 32, learning rate = 0.0001, kernel dimension = 3, early stopping = 10, subsequence length = 20, hidden state dimension D_model = 32, number of network layers = 64, state expansion factor = 2, and L1 paradigm loss function is used. The training set is then input into a long-term series prediction model based on the Mamba model for training, evaluated on the validation set, and the best network model is selected and saved. Finally, the test set is input into the trained network model for testing, outputting evaluation metrics, and iteratively predicting 10,000 seconds of data.
[0105] Step S5: Long-term series prediction has systematic errors and cumulative errors. Systematic errors usually exhibit a non-linear drift trend, so the detrend algorithm is used to remove systematic drift errors. Cumulative errors usually exhibit diverse changes, affected by factors such as the number of iterations and data length. Based on the pattern that the cumulative error continuously increases and then stabilizes, and shows a continuous decay and then stabilizes in the signal amplitude, an appropriate scaling interval is selected. For example, in predicting the later data using the data before and after, the first 10,000 points of the prediction result are kept unchanged, and the remaining points of the prediction result are amplified by 100 times, while ensuring the consistency of the data phase before and after compensation.
[0106] Step S6: Execute steps S1 to S5 and output the compensated result. For example... Figure 4 As shown, the measured thrust noise floor signal is selected. Then, after executing step S6, a power spectral density comparison analysis is performed. The predicted data of the compensated intermediate applied force is lower than the original 3820µN / Hz at 0.1Hz. 1 / 2 It dropped to 0.08µN / Hz 1 / 2 It outperforms the first-order detrending algorithm at 0.91µN / Hz. 1 / 2 ;like Figure 5 As shown, a standard force is applied to the fitted and compensated noise floor signal, and the TVD algorithm is used to extract the compensated step-shaped standard force. The force extracted by TVD is basically consistent with the standard force, and the maximum amplitude error is less than 0.05μN. Further analysis shows that the force extracted by TVD does not contain obvious drift.
[0107] In summary, the ultra-low frequency drift compensation method for micro-thrust measurement based on trend prediction will significantly reduce the impact of drift on the measurement accuracy under external interference, especially reducing the impact of drift caused by temperature changes, thus improving the stability and accuracy of the measurement.
[0108] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.
Claims
1. A micro-thrust measurement ultra-low frequency drift compensation system based on trend prediction, characterized in that: It includes at least a micro-thrust measurement data acquisition module, a nonlinear drift removal module, an ultra-low frequency component extraction module, a long-time series prediction module, a prediction error correction module, and a drift identification module, wherein: The micro-thrust measurement data acquisition module is used to acquire micro-thrust measurement data under different conditions; the different conditions include at least three conditions: the state of applying the micro-thrust to be measured, the state of no external force applied, and the state of applying a standard force. The nonlinear drift removal module: based on the TVD filter skeleton line of two data segments under different states, a fitting algorithm is used to fit and compensate for the low-frequency drift of the intermediate signal. The ultra-low frequency component extraction module: performs signal decomposition on the residual signal after compensation of two data segments under different states using a signal decomposition method to extract the ultra-low frequency drift component; The long-time series prediction module: predicts drift based on a long-time series prediction model to achieve ultra-low frequency drift prediction on a small scale; The prediction error correction module is used to correct the errors in the output of the long-term series prediction module, including system drift error and cumulative error from multiple prediction iterations. The drift identification module: by performing time-frequency analysis on the compensated intermediate applied force data, it determines whether there is thrust drift in the thrust signal, thereby decoupling the background signal drift and the thrust signal drift.
2. The ultra-low frequency drift compensation system for micro-thrust measurement based on trend prediction as described in claim 1, characterized in that: The fitting algorithm used in the nonlinear drift removal module is an improved weighted least squares fitting algorithm, and its objective function is: ； in, It is a measurement signal. It is a design matrix. It is the estimated parameter vector. Indicates weight, The total number of observations; Define one-dimensional micro-thrust data as The time interval it occupies is ; middle The value at time t is The intermediate data segment to be compensated Defined as: ； in Indicates the middle data segment The value is empty. express The time period in which it occurs; Different weighting strategies are used for the two data segments, specifically: ； ； in This indicates the weight of the preceding data points. This indicates the weight of the data points in the latter part. This represents the error observation data.
3. The ultra-low frequency drift compensation system for micro-thrust measurement based on trend prediction as described in claim 1, characterized in that: The signal decomposition method in the ultra-low frequency component extraction module includes a wavelet transform-based signal decomposition method. The two segments of data after fitting and compensation are decomposed separately. Based on the principle of minimizing reconstruction error, Symlets are selected as wavelet basis functions. At this time, the ultra-low frequency drift component is represented as an approximate component, and the mid-to-high frequency noise component is represented as a detail component. The ultra-low frequency drift component is retained to extract the residual drift feature component, and the noise component is discarded.
4. The ultra-low frequency drift compensation system for micro-thrust measurement based on trend prediction as described in claim 1, characterized in that: The long-series time prediction module includes at least a Mamba encoder, a channel fusion location embedding, and a Mamba decoder. The encoded output and the location encoded output are fused using a concatenate operation, and then the output is passed through a linear layer to the Mamba decoder for decoding. The decoder output is passed through a linear layer and reshaped to restore the output dimension. RevIN is added before the network input and after the output for standardization to ensure consistent input and output distribution.
5. The ultra-low frequency drift compensation system for micro-thrust measurement based on trend prediction as described in claim 1, characterized in that: The system drift error in the prediction error correction module is removed by a first-order detrending algorithm. Linear regression analysis is performed on the data y(t) to fit a linear model y(t)=at+b, where a is the slope and b is the intercept. The fitted value y^(t)=at+b at each time point t is calculated. The fitted linear trend is subtracted from the original data to obtain the detrended data.
6. The ultra-low frequency drift compensation system for micro-thrust measurement based on trend prediction as described in claim 1, characterized in that: In the drift identification module, a standard force is applied to the low-noise signal after fitting and compensation. By performing time-frequency analysis on the data of the intermediate applied force after prediction and compensation, the TVD algorithm is used to extract the stepped standard force after compensation. The force extracted by TVD is compared with the standard force, and the maximum amplitude error is used to determine whether the thrust contains drift.
7. The ultra-low frequency drift compensation method for micro-thrust measurement based on trend prediction using the system described in claim 1, characterized in that, It should include at least the following steps: S1: Collect micro-thrust measurement data under different states, including at least the state of no actual external force applied, the state of applying actual external force and then removing the external force under the state of no actual external force applied, and the state of applying standard force and then removing the standard force under the state of no standard force applied. Record the start time and time interval of the latter two states. S2: Based on the data collected in step S1, the latter two states include force-free data of equal length at the beginning and end and force-containing data in the middle. The TVD filtering algorithm is used to extract the filter skeleton line of the force-free data of equal length at the beginning and end, and the force-containing data in the middle is replaced with NAN. The TVD filter skeleton line of the beginning and end segments is used to input the data to remove nonlinear drift based on improved least squares fitting, and the whole segment of data after fitting compensation is obtained. S3: The wavelet decomposition algorithm is used to decompose the data before and after fitting compensation. According to the principle of minimizing reconstruction error, Symlets is selected as the wavelet basis function and the order and decomposition level are determined. The ultra-low frequency drift component is retained to extract the residual drift feature component, thereby obtaining the prediction dataset. If the index requirements are met, the extraction of ultra-low frequency drift component is completed; otherwise, the order and decomposition level are re-determined, and the signal is decomposed again. S4: A long-term series prediction neural network model based on Mamba, which is trained to output the prediction results of intermediate segment drift; S5: Based on the drift prediction output of step S4, remove the system drift error by using a first-order detrending algorithm, and perform scaling error correction by manually inferring the accumulated error trajectory to remove the accumulated error from multiple iterations. S6: Repeat steps S1-S5, output the compensated thrust signal, apply a standard force to the fitted compensated low-noise signal, and perform time-frequency analysis on the data of the intermediate applied force after prediction compensation to determine whether the thrust contains drift, thereby achieving decoupling between background signal drift and thrust signal drift.
8. The ultra-low frequency drift compensation method for micro-thrust measurement based on trend prediction as described in claim 7, characterized in that: The parameter settings for the Mamba-based long-term sequence prediction neural network model in step S4 are as follows: Epochs = 50, Batch_size = 32, learning rate = 0.0001, kernel dimension = 3, early stopping = 10, subsequence length = 20, hidden layer state dimension D_model = 32, number of network layers = 64, state expansion factor = 2, and L1 paradigm loss function is used; the ratio of the training set, validation set and test set of the model is 6:2:2.