Dynamic stabilization and automatic calibration method for aiming point fused with data of inertial measurement unit

Abstract

The invention relates to an aiming point dynamic stabilization and automatic calibration method fusing inertial measurement unit data, and belongs to the technical field of aiming equipment dynamic control. The method comprises the following steps: acquiring attitude data of an inertial measurement unit, and constructing a multi-source data matrix in combination with real-time position data of an aiming point and environmental interference data; after time sequence synchronization and drift suppression are carried out on multi-source data, associated feature vectors are extracted through multi-modal feature fusion; inputting a self-adaptive extended state observer to estimate total disturbance and generate an anti-interference compensation amount, constructing a kinetic model in combination with a nonlinear model prediction controller, solving a multi-objective optimization problem, and generating an attitude compensation control sequence to realize dynamic stability of an aiming point; and finally, aiming deviation is calculated in real time to trigger double-closed-loop calibration, an inner ring corrects the drift error of the inertial measurement unit, and an outer ring optimizes parameters of the observer and the controller. According to the invention, environmental interference and inertial drift are effectively suppressed, the dynamic aiming stability is improved, and the long-term aiming reliability is guaranteed through closed-loop calibration.

Description

Technical Field

[0001] This invention belongs to the field of dynamic control technology for aiming equipment, specifically relating to a method for dynamic stabilization and automatic calibration of aiming points that integrates inertial measurement unit data. Background Technology

[0002] Traditional aiming stabilization schemes often rely on mechanical anti-shake structures combined with simple inertial compensation algorithms. For example, a standard Kalman filter is used to process the attitude data output by the inertial measurement unit (IMU) to offset aiming offset caused by changes in device attitude. However, such schemes have significant drawbacks: First, the mechanical structure has a limited response speed and cannot cope with complex time-varying disturbances such as sudden vibrations and temperature drift, leading to cumulative drift in the IMU attitude data and consequently, aiming point offset. Second, traditional filtering algorithms are mostly based on linear assumptions to build models, while the actual motion of the aiming device exhibits significant nonlinearity. Linear models cannot accurately describe these nonlinear behaviors, resulting in increased dynamic offset estimation errors and making it difficult to meet the stabilization accuracy requirements of high-precision scenarios.

[0003] Furthermore, existing solutions often rely solely on IMU attitude data for stabilization control, failing to fully integrate real-time aiming point position data with environmental interference data. For example, when environmental interference causes fluctuations in the aiming point position, IMU attitude compensation alone cannot completely offset these fluctuations, nor can it establish a correlation between "attitude-position-interference," resulting in insufficient ability to predict and compensate for dynamic interference, further reducing the dynamic stabilization effect.

[0004] The long-term accuracy of the aiming system relies on periodic calibration to correct for issues such as IMU device drift and control model parameter deviations. Existing calibration schemes suffer from two typical drawbacks: First, manual calibration accounts for a high proportion, requiring operators to manually adjust the calibration target position and manually record IMU reference data, which is not only inefficient but also prone to introducing human error. Second, the automatic calibration mechanism is imperfect; most schemes only achieve single-stage calibration and do not consider the impact of environmental interference on calibration accuracy. For example, when calibrating the IMU in a vibration environment, if the interference of vibration noise on the calibration attitude data is not filtered out, the zero-bias compensation parameters obtained from the calibration will be biased. At the same time, existing calibrations do not form a closed loop of "calibration-feedback-optimization." After calibration, the calibration parameters cannot be dynamically adjusted according to the actual aiming deviation. When environmental conditions change, the calibrated parameters quickly become invalid, requiring frequent recalibration, which is difficult to adapt to dynamically changing application scenarios.

[0005] Existing technologies have significant shortcomings in the data processing stage: on the one hand, the timing synchronization accuracy of multi-source data is low, for example, the timestamp deviation between IMU data and aiming point position data is large, which leads to errors in subsequent data fusion and feature extraction; on the other hand, the data fusion mechanism is simple, mostly using direct splicing or weighted summation, without constructing a unified feature space to extract deep correlation features of "attitude-position-disturbance", resulting in insufficient feature quality for subsequent disturbance estimation and control model input.

[0006] Regarding the optimization of control and calibration models, existing solutions lack a dynamic update mechanism. For example, the constraints and optimization weight matrices of the control model are mostly fixed values ​​and are not adaptively adjusted based on model errors discovered during calibration. The IMU's calibration coefficients are only set at the factory and are not dynamically corrected in conjunction with the accumulated drift errors during long-term use. This leads to a gradual disconnect between model parameters and actual operating conditions, further exacerbating the problems of aiming deviation and insufficient calibration accuracy.

[0007] In summary, current aiming point dynamic stabilization and automatic calibration technologies suffer from significant technical deficiencies in areas such as anti-interference capability, nonlinear adaptability, automatic calibration robustness, and multi-source data processing, making it difficult to meet the application requirements of high precision, high dynamics, and long-term stability. Therefore, a technical solution is needed that can integrate IMU data and multi-source information to achieve high-precision dynamic stabilization control and closed-loop automatic calibration, thereby addressing the pain points of existing technologies. Summary of the Invention

[0008] To address the aforementioned problems in the existing technology, this invention provides a method for dynamic stabilization and automatic calibration of the aiming point by integrating inertial measurement unit data. The objective of this invention can be achieved through the following technical solution: A method for dynamic stabilization and automatic calibration of the aiming point that integrates inertial measurement unit data includes: S1: Acquire the attitude data output by the inertial measurement unit and construct a multi-source data matrix with the real-time position data of the aiming point and environmental interference data; S2: First, the attitude data, real-time position data of the aiming point, and environmental interference data are processed for time-series synchronization. Drift suppression is performed on the synchronized attitude data. Through a multimodal feature fusion mechanism, the data is mapped to a unified feature space, and the associated feature vectors of attitude-position-interference are extracted. S3: Estimate the total disturbance in real time based on the extended state, dynamically adjust the observation gain, and generate the disturbance rejection compensation amount; construct a nonlinear attitude dynamics model based on the associated feature vector and the disturbance rejection compensation amount, set control constraints, solve the multi-objective optimization problem within the sampling period, and generate the attitude compensation control sequence of the sampling period. S4: Real-time calculation of the deviation between the actual position of the aiming point and the theoretical position of the target after dynamic stabilization, triggering a dual closed-loop calibration process; acquiring calibration target reference data, inertial measurement unit calibration attitude data, and calibration position data; the inner loop processes the inertial measurement unit calibration attitude data, calculates the difference with the standard reference attitude, generates cumulative drift error, and updates the zero bias compensation parameters and calibration coefficients; the outer loop, based on the difference between the calibration target reference data and the calibration position, combines environmental interference characteristics, fits the deviation and interference mapping relationship, generates model error correction coefficients, and updates the observation gain, optimized weight matrix, and control constraint threshold.

[0009] Specifically, when constructing the multi-source data matrix, feature decomposition and correlation mapping are performed on various types of data: for the attitude data, three core features are extracted: attitude change rate, attitude angle deviation, and attitude stability index; for the real-time position data of the aiming point, it is decomposed into three features: position coordinate deviation, position change trend, and position fluctuation frequency; for the environmental interference data, it is classified into slow-varying interference features and sudden interference features; the features of the three types of data are aligned based on the time dimension, and a time series dimension is constructed with the same time interval as the unit. The attitude data features, aiming point position data features, and environmental interference data features in each time unit together constitute a data vector, which is arranged in chronological order to generate the multi-source data matrix including the time-data type-feature dimension.

[0010] Specifically, during the timing synchronization process, the internal clock of the inertial measurement unit is used as a reference, and the time stamps of the real-time position data of the aiming point and the environmental interference data are adjusted by a timestamp comparison algorithm. When the attitude data after synchronization is drift suppressed, a dynamic filtering algorithm based on the environmental interference data is used to adjust the filtering window and filtering coefficient according to the temperature change amplitude and vibration intensity in the environment.

[0011] Specifically, when the multimodal feature fusion mechanism maps data to a unified feature space, it first performs feature standardization on the attitude data, real-time aiming point position data, and environmental interference data respectively: the attitude data is normalized to a preset feature interval based on the angular velocity-linear acceleration dimension, the aiming point position data is normalized based on the target coordinate range, and the environmental interference data is quantized based on the interference intensity level; then, through feature correlation analysis, fusion weights are assigned, and finally, through feature splicing and dimensional compression, the dimensionally unified attitude-position-interference associated feature vector is generated.

[0012] Specifically, when processing the associated feature vector through the adaptive extended state observer, the motion state feature components of the aiming device are extracted as input states. When estimating the total disturbance in real time based on the extended state, the extended state is defined as a comprehensive disturbance term of the unmodeled dynamics of the aiming device and the total environmental interference, and the estimated value of the total disturbance is corrected. When dynamically adjusting the observation gain, based on the deviation between the estimated value of the total disturbance and the actual observed interference data, proportional-integral adjustment logic is used to generate the anti-interference compensation amount. The estimated total disturbance is transformed into a reverse compensation component based on the direction and intensity of the influence.

[0013] Specifically, when constructing the nonlinear attitude dynamics model, based on the attitude motion law of the aiming device, a coupled motion equation of pitch angle and azimuth angle is established, incorporating the inertial delay characteristics of attitude change and the response delay characteristics of the actuator; the attitude data in the associated feature vector is used as the state input, the disturbance compensation amount is used as the disturbance compensation input, and the parameters are verified through historical attitude motion data.

[0014] Specifically, when setting the control constraints, the constraints include the attitude adjustment rate of the actuator, control energy consumption, and a preset threshold for the aiming point offset; during the transformation of the constraints, the constraints are transformed into inequality constraints of the optimization problem, and critical constraint situations are handled through a constraint boundary softening mechanism.

[0015] Specifically, when solving the multi-objective optimization problem, weights are assigned based on three major optimization objectives: interference suppression effect, control energy consumption, and attitude adjustment. A time domain length is preset, and the future attitude change trend is predicted based on the nonlinear attitude dynamics model. Combined with control constraints, the attitude compensation control sequence that minimizes the weighted sum of the multi-objectives is solved through an iterative optimization algorithm.

[0016] Specifically, the judgment process for triggering the dual closed-loop calibration process includes: calculating the deviation between the actual position of the aiming point and the theoretical position of the target in real time, setting deviation judgment rules; monitoring calibration trigger scenarios, triggering initialization calibration when the inertial measurement unit is powered on and initializing, and triggering timed calibration when the preset time period is reached.

[0017] Specifically, the inner loop uses the Kalman smoothing algorithm to process the calibration attitude data, sets the smoothing window length, and filters out short-term random errors in the calibration attitude data; when calculating the difference with the standard reference attitude, the attitude angle difference is calculated point by point based on the sampling order, and the cumulative drift error of the inertial measurement unit within the calibration period is generated through integration; when updating the zero bias compensation parameters, the angular velocity zero bias and linear acceleration zero bias are adjusted using the incremental correction method according to the amplitude and trend of the cumulative drift error; when updating the scale coefficients, the inter-axis coupling coefficient and temperature compensation coefficient are corrected by combining the correlation between drift error and attitude change range.

[0018] Specifically, the process of fitting the outer loop deviation and the mapping relationship between interference includes: using the difference between the calibration target reference data and the calibration position data as the deviation sample, and using the environmental interference characteristics at the corresponding time as the interference sample to construct a sample dataset; using the least squares method to fit the sample dataset to establish a linear or nonlinear mapping equation between the deviation value and the interference characteristics.

[0019] Specifically, the dual-loop calibration process stores the deviation data, drift error data, and model correction coefficient data generated during the current calibration process according to timestamps and calibration scenarios, generating a historical calibration database. When the calibration process is triggered again, historical calibration data of the same or similar scenarios are first retrieved from the historical database, and the matching degree between the current scenario and the historical scenario is calculated through a similarity matching algorithm.

[0020] This method addresses the technical pain points of insufficient dynamic stability accuracy of aiming equipment, manual calibration reliance, and poor robustness. Through technological innovations such as multi-source data collaborative processing, disturbance rejection control, and dual closed-loop calibration, it achieves multi-dimensional performance improvements, with the following specific benefits: This method employs a fusion architecture of "adaptive extended state observer + nonlinear model predictive control": the adaptive extended state observer can estimate the total disturbance of the aiming device in real time (including slowly varying environmental disturbances, sudden vibrations, and unmodeled dynamics), and eliminate disturbance estimation bias by dynamically adjusting the observation gain; the nonlinear model predictive control combines the attitude coupling characteristics of the aiming device with the constraints of the actuator to solve a multi-objective optimization problem and generate a precise control sequence. Compared with traditional linear control methods, it can effectively suppress aiming point offset caused by attitude coupling and environmental disturbances, ensuring that the aiming point position deviation is always controlled within a preset threshold during dynamic stabilization, making it particularly suitable for aiming requirements in complex disturbance scenarios.

[0021] The designed dual-loop calibration process forms a collaborative mechanism of "IMU drift correction - model parameter optimization": the inner loop uses Kalman smoothing to process IMU calibration attitude data, accurately calculates cumulative drift error, and updates zero-bias compensation parameters and calibration coefficients, avoiding accuracy degradation caused by IMU drift over long-term use; the outer loop combines environmental interference characteristics to fit the deviation mapping relationship, optimizes observation gain and control constraint thresholds, and ensures that the model adapts to environmental changes. Compared with traditional manual calibration, initialization, timing, or deviation exceeding threshold calibration can be triggered without manual intervention, and scene matching through historical calibration database improves the consistency of calibration accuracy under different environments, increasing calibration efficiency while reducing human error.

[0022] By constructing a multi-source data matrix, synchronizing time sequences, and fusing multimodal features, this approach addresses the problems of "time asynchrony, high noise interference, and fragmented features" in traditional data processing. It achieves time alignment of three types of data based on the IMU clock, dynamically adjusts filtering parameters to suppress IMU drift based on environmental interference, and generates a unified "attitude-position-interference" correlated feature vector through feature standardization and correlation-weighted fusion. This process effectively eliminates data noise and redundant information, ensuring that the input data for subsequent observers and controllers is both timely and correlated, avoiding control deviations or calibration errors caused by data quality issues, and guaranteeing the reliability of the overall technical solution.

[0023] The core technology modules of this method possess scenario adaptability: the adaptive extended state observer dynamically adapts to disturbance changes through proportional-integral adjustment logic, eliminating the need to redesign the observer structure for different devices; the nonlinear attitude dynamics model incorporates attitude delay and actuator characteristics, and can verify the motion patterns of different types of aiming devices through historical data; the threshold and parameter optimization mechanism of dual closed-loop calibration can dynamically adjust the calibration strategy according to the usage frequency of the aiming device and the intensity of environmental interference. Compared with traditional methods with fixed parameters, it can adapt to different models of aiming devices without extensive customized debugging, reducing the adaptation cost and usage threshold of the technical solution.

[0024] Through a closed-loop iteration of "dynamic stabilization - calibration feedback - parameter update," this method establishes a long-term accuracy maintenance mechanism: real-time feedback during the dynamic stabilization process optimizes model parameters, the calibration process periodically corrects IMU drift and model bias, and the historical calibration database provides data accumulation for subsequent optimization. As usage time increases, model and IMU parameters continuously approach their optimal state, avoiding the problem of "accuracy decaying with usage time" in traditional methods. This extends the high-precision working cycle of the aiming equipment, reduces equipment maintenance or replacement costs due to insufficient accuracy, and improves the overall efficiency of equipment use. Attached Figure Description

[0025] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart illustrating a method for dynamic stabilization and automatic calibration of aiming points that integrates inertial measurement unit data according to the present invention. Figure 2 This is a structural diagram of the anti-interference controller in this invention; Figure 3 This is a logic diagram of the dual closed-loop calibration process in this invention. Detailed Implementation

[0027] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0028] Please see Figure 1-3 A method for dynamic stabilization and automatic calibration of aiming points that integrates inertial measurement unit data, comprising: S1: Acquire the attitude data output by the inertial measurement unit and construct a multi-source data matrix with the real-time position data of the aiming point and environmental interference data; S2: First, the attitude data, real-time position data of the aiming point, and environmental interference data are processed for time-series synchronization. Drift suppression is performed on the synchronized attitude data. Through a multimodal feature fusion mechanism, the data is mapped to a unified feature space, and the associated feature vectors of attitude-position-interference are extracted. S3: Estimate the total disturbance in real time based on the extended state, dynamically adjust the observation gain, and generate the disturbance rejection compensation amount; construct a nonlinear attitude dynamics model based on the associated feature vector and the disturbance rejection compensation amount, set control constraints, solve the multi-objective optimization problem within the sampling period, and generate the attitude compensation control sequence of the sampling period. S4: Real-time calculation of the deviation between the actual position of the aiming point and the theoretical position of the target after dynamic stabilization, triggering a dual closed-loop calibration process; acquiring calibration target reference data, inertial measurement unit calibration attitude data, and calibration position data; the inner loop processes the inertial measurement unit calibration attitude data, calculates the difference with the standard reference attitude, generates cumulative drift error, and updates the zero bias compensation parameters and calibration coefficients; the outer loop, based on the difference between the calibration target reference data and the calibration position, combines environmental interference characteristics, fits the deviation and interference mapping relationship, generates model error correction coefficients, and updates the observation gain, optimized weight matrix, and control constraint threshold.

[0029] This embodiment is applied to a laser aiming system (hereinafter referred to as "aiming system"), which includes an inertial measurement unit (IMU), a laser aiming point detection module and an environmental interference monitoring module. It needs to achieve stable control of the aiming point under complex environments such as temperature drift and sudden vibration.

[0030] Specifically, when constructing the multi-source data matrix, it is necessary to first decompose the features of various types of data and perform cross-type association mapping: for the attitude data output by the inertial measurement unit, three core features are extracted—attitude change rate, attitude angle deviation, and attitude stability index; for the real-time position data of the aiming point, it is decomposed into three key features—position coordinate deviation, position change trend, and position fluctuation frequency; for environmental interference data, it is classified into slow-varying interference features and sudden interference features according to the change characteristics—slow-varying interference features include temperature drift and continuous low-intensity vibration, and sudden interference features include instantaneous impact vibration and instantaneous electromagnetic interference.

[0031] After feature decomposition, the features of the three types of data are strictly aligned based on the time dimension: using the sampling time interval of the inertial measurement unit as a unified benchmark, the sampling time points of the real-time position data of the aiming point and the environmental interference data are adjusted to completely match the sampling time points of the inertial measurement unit, ensuring that each time unit contains complete attitude data features, aiming point position data features, and environmental interference data features; the three types of data features in each time unit are combined to form a multi-dimensional data vector, and then the multi-dimensional data vectors corresponding to all time units are arranged in chronological order to finally generate the three-dimensional structure of the multi-source data matrix containing the time dimension, data type dimension (attitude, position, interference), and feature dimension.

[0032] Specifically, during the timing synchronization process, the internal clock of the inertial measurement unit (IMU) is used as the time reference—because the sampling frequency and time stability of the IMU are usually higher than those of other data acquisition modules, its internal clock can provide a more accurate time reference. A timestamp comparison algorithm is used to synchronize the time of other data with the IMU data: first, the timestamp corresponding to each set of attitude data output by the IMU is extracted; then, the timestamps of the real-time position data of the aiming point and the environmental interference data are extracted separately. The timestamps of the latter two types of data are compared one by one with the timestamp of the IMU. For data points with mismatched timestamps, linear interpolation (suitable for scenarios with gradual data changes) or adjacent time point matching (suitable for scenarios with drastic data changes) is used to adjust the sampling values ​​of the real-time position data of the aiming point and the environmental interference data to a time node completely consistent with the timestamp of the IMU, ensuring strict synchronization of the three types of data in the time dimension and avoiding data correlation deviations caused by time misalignment.

[0033] When suppressing drift in the synchronized attitude data, a dynamic filtering algorithm based on the environmental interference data is adopted. This algorithm can sense the changing state of environmental interference in real time. When the temperature difference per unit time exceeds the preset sensing threshold, it is determined that the attitude data is susceptible to temperature drift. The length of the filtering window is then increased (by extending the data smoothing period, the suppression effect on slow drift is enhanced). When the vibration amplitude in the environment exceeds the preset vibration threshold, it is determined that the attitude data contains a lot of high-frequency vibration noise. The filtering coefficient is then adjusted (by increasing the attenuation ratio of high-frequency noise). By dynamically adapting the filtering parameters, attitude drift is effectively suppressed while retaining the effective information related to the actual movement of the device in the attitude data, avoiding attitude response delay caused by over-filtering.

[0034] Data Acquisition: After the aiming system is started, the attitude data output by the IMU is acquired in real time. This data includes angular velocity data ω (angular velocity ω1 in the pitch direction and angular velocity ω2 in the azimuth direction of the aiming system) and linear acceleration data a (linear acceleration a1 along the aiming axis and linear acceleration a2 perpendicular to the aiming axis). The laser aiming point detection module acquires real-time position data of the aiming point, which is decomposed into position coordinate deviation d (horizontal deviation d1 and vertical deviation d2 between the aiming point and the target center) and position change trend t (horizontal position change rate t1 and vertical position change rate t2). At the same time, the position fluctuation frequency f (f1 is the horizontal fluctuation frequency and f2 is the vertical fluctuation frequency) is recorded. Environmental interference data is acquired through the environmental interference monitoring module and classified into slow-varying interferences s (interferences caused by temperature drift s1 and humidity changes s2) and sudden interferences b (interferences caused by equipment vibration b1 and external impacts b2).

[0035] Construction of multi-source data matrix: Align the features of the above data based on the time dimension, and construct the time series dimension (e.g., T1, T2, T3…T) using the same time interval T as the unit. n (as a continuous time unit) Within each time unit T, the attitude data features (ω1, ω2, a1, a2), the aiming point position data features (d1, d2, t1, t2, f1, f2), and the environmental interference data features (s1, s2, b1, b2) are combined to form the data vector V_T; V_T1, V_T2...V_T are arranged in chronological order. n Arrange the data to generate a multi-source data matrix M containing time, data type, and feature dimensions. The rows of matrix M correspond to the time unit T, and the columns correspond to various data features.

[0036] Specifically, when the multimodal feature fusion mechanism maps various types of data to a unified feature space, it undergoes three core steps: feature standardization, feature weight allocation, and feature concatenation and dimensionality compression. First, feature standardization is performed—for the attitude data, normalization is performed based on the two core dimensions of angular velocity and linear acceleration. Angular velocity data is mapped to a preset standard interval according to its physical range using a min-max normalization algorithm, ensuring that angular velocity data of different directions and magnitudes have a unified numerical scale. Linear acceleration data is normalized to another preset standard interval based on its range using similar logic. The real-time position data of the aiming point is normalized to the standard range of the corresponding axis based on the spatial coordinate range of the target (i.e., the maximum and minimum values ​​of the X-axis, Y-axis, and Z-axis coordinates that the target may appear in), eliminating the influence of differences in the numerical range of different coordinate dimensions. For environmental interference data, it is quantified based on the interference intensity level: first, the intensity of various interferences (temperature drift, vibration, etc.) is divided into several preset levels (e.g., from 0 to 5 levels from low to high), and then the interference data is converted into a unified level quantification value according to the actual interference detection value, so as to achieve the standardization of interference data.

[0037] After standardization, fusion weights are assigned through feature correlation analysis: the correlation between each feature of each data class and the core objective of "aiming point stability" is calculated. Features with high correlation (such as attitude angle deviation and position coordinate deviation) are assigned higher fusion weights, while features with low correlation (such as minor features of some environmental interference) are assigned lower weights, ensuring that important features play a dominant role in the fusion process. Finally, feature concatenation and dimensionality compression are performed: the standardized features of the three data classes are concatenated in a preset order to form a high-dimensional feature vector, and then the high-dimensional vector is compressed using principal component analysis—preserving principal components that can explain the main variation information of the data (such as principal components whose cumulative variance contribution rate exceeds a preset threshold), and eliminating redundant feature dimensions, ultimately generating the attitude-position-interference correlation feature vector with uniform dimensions and high information density.

[0038] Attitude data drift suppression: A dynamic filtering algorithm based on environmental interference data is used to suppress drift in the synchronized attitude data (ω1, ω2, a1, a2): Extract the temperature change amplitude s1 (a slow-varying disturbance s) and vibration intensity b1 (a sudden disturbance b) from the environmental disturbance data. Set the filter window adjustment rules: if s1 > S_th (S_th is the temperature disturbance threshold) or b1 > B_th (B_th is the vibration disturbance threshold), then adjust the filter window W to W1 (W1 > W, W is the initial window), and at the same time adjust the filter coefficient K to K1 (K1 > K, K is the initial coefficient); if s1 ≤ S_th and b1 ≤ B_th, then keep the filter window W and the filter coefficient K. The attitude data is filtered by adjusting W (or W1) and K (or K1) to remove the drift components ω_drift (ω1 drift caused by temperature ω_drift1) in ω1 and ω2 and the drift components a_drift (a1 drift caused by vibration a_drift1) in a1 and a2, so as to obtain the drift-free attitude data ω' (ω'1, ω'2) and a' (a'1, a'2).

[0039] Multimodal feature fusion: The three types of preprocessed data are mapped to a unified feature space through a multimodal feature fusion mechanism: Feature standardization: The attitude data ω' and a' are normalized to the preset interval [X,Y] based on the angular velocity-linear acceleration dimension (ω'1 is normalized to ω''1∈[X,Y], and a'1 is normalized to a''1∈[X,Y]); the aiming point position data d and t are normalized based on the target coordinate range R (d1 is normalized to d''1∈[0,R], and t1 is normalized to t''1∈[0,R]); the environmental interference data s and b are quantized based on the interference intensity level L (s1 is quantized to s''1∈[1,L], and b1 is quantized to b''1∈[1,L]). Feature correlation analysis: Calculate the correlation coefficients between standardized features, such as the correlation coefficient r1 between ω''1 and d''1, the correlation coefficient r2 between a''1 and t''1, and the correlation coefficient r3 between s''1 and ω''2. Assign fusion weights W1 (e.g., W1_ω''1 > W1_d''1 when r1 > r2) and W2 (e.g., W2_s''1 > W2_ω''2 when r3 > r1) to each feature based on r1, r2, and r3. Feature concatenation and dimensionality compression: The weighted features (W1_ω'', W1_a'', W1_d'', W1_t'', W2_s'', W2_b'') are concatenated into a high-dimensional feature vector. Dimensionality compression is performed through principal component analysis to generate a pose-position-perturbation correlation feature vector V with uniform dimensionality (V contains feature components v1=W1_ω''1, v2=W1_a''1, v3=W1_d''1, v4=W2_s''1, v5=W2_b''1).

[0040] Specifically, when processing the associated feature vector, the adaptive extended state observer first extracts feature components directly related to the motion state of the aiming device from the associated feature vector as the observer's input state—specifically including attitude angles and attitude change rates in the attitude data, and position coordinates and position change trends in the aiming point position data. These feature components can fully reflect the current motion state of the device, providing basic state information for disturbance estimation. When estimating the total disturbance in real time based on the extended state, the extended state is explicitly defined as a comprehensive disturbance term of the unmodeled dynamics of the aiming device and the total environmental disturbance: the unmodeled dynamics include dynamic characteristics that are difficult to describe by precise models, such as the elastic deformation of the device's own structure and the nonlinear dead zone of the actuator; the total environmental disturbance includes the superposition effect of slow-varying disturbances and sudden disturbances. The observer incorporates this comprehensive disturbance term as an additional state variable into the observation system, and, combined with the change law of the input state, calculates the estimated value of the total disturbance in real time through the extended state equation. Compared with traditional observers that can only estimate environmental disturbances, this method can more comprehensively cover various disturbance factors affecting the stability of the device's attitude, improving the completeness and accuracy of disturbance estimation.

[0041] When dynamically adjusting the observation gain, a bias-based proportional-integral (PI) adjustment logic is adopted: First, the deviation between the estimated total disturbance and the actual observed disturbance data is calculated. When the deviation is positive (i.e., the estimated value is less than the actual disturbance), the observation gain is increased through a proportional loop to accelerate the response speed of the disturbance estimation, while the deviation is accumulated through an integral loop to eliminate static errors. When the deviation is negative (i.e., the estimated value is greater than the actual disturbance), the observation gain is decreased through a proportional loop to avoid excessive fluctuations in the estimated value, while the integral loop continues to adjust to gradually converge to the actual disturbance value. Through PI-integral coordinated adjustment, adaptive optimization of the observation gain is achieved. When generating the disturbance rejection compensation, the direction and intensity of the estimated total disturbance's influence on the aiming point stability are reversed: if the total disturbance causes the aiming point to drift in a certain direction (e.g., a larger attitude angle causes the aiming point to drift upwards), a compensation component with the opposite direction and matching intensity (e.g., a compensation amount for adjusting the attitude angle downwards) is generated.

[0042] Specifically, when constructing the nonlinear attitude dynamics model, the actual attitude motion law of the aiming device is taken as the core. First, the coupled motion equations of pitch and azimuth are established: considering that when the device adjusts its attitude, the change in pitch will affect the azimuth through the mechanical coupling relationship of the device structure, and vice versa, a coupling coefficient (used to characterize the degree of mutual influence between the two attitude angles) is introduced into the equation. The specific value of the coupling coefficient is determined by fitting analysis of the historical attitude motion data of the device, so that the equation can accurately describe the coordinated change relationship between the two attitude angles. At the same time, the inertial delay characteristics of attitude change and the response delay characteristics of the actuator are incorporated into the model: the inertial delay characteristics are reflected by adding an inertial term to the equation, describing the phenomenon that the device cannot respond to the control command instantaneously due to its own inertia; the response delay characteristics of the actuator are reflected by adding a delay element, describing the time lag between the issuance of the control command and the actual action of the actuator. The incorporation of these characteristics enables the model to more realistically reflect the dynamic response process of the device and avoids control deviations caused by ignoring delays.

[0043] Regarding the input design of the model, the attitude data (attitude angle, attitude change rate) in the associated feature vector is used as the state input to reflect the current attitude state of the device in real time; the disturbance rejection compensation amount is used as the disturbance compensation input to offset the influence of disturbances not fully covered by the model and improve the model's disturbance rejection capability. After the model is built, the parameters are verified using historical attitude motion data: select the attitude motion data of the device under different past operating conditions (including attitude input, control commands, and output attitude), substitute the input data into the model to calculate the predicted attitude, compare it with the actual output attitude, and calculate the deviation value between the two. If the deviation value exceeds the preset verification threshold, adjust the coupling coefficient, inertia term parameters, and delay element parameters in the model, and repeat the verification process until the deviation value meets the requirements, ensuring the accuracy of the model parameters and the prediction accuracy of the model.

[0044] Specifically, when setting the control constraints, the specific content of the constraints is first clarified, covering the physical performance limitations and aiming accuracy requirements of the actuator: the attitude adjustment rate constraint of the actuator, i.e., the maximum amplitude of the attitude angle that the actuator can adjust per unit time; the control energy consumption constraint, i.e., the maximum energy consumption of the actuator per unit time during attitude adjustment; and the preset threshold constraint for aiming point offset, i.e., the maximum distance that the aiming point is allowed to deviate from the theoretical position of the target during attitude adjustment. When these constraints are transformed into inequalities that can be solved in the optimization problem, mathematical modeling is used: for example, the attitude adjustment rate constraint is transformed into the inequality "attitude angle adjustment amount in each sampling period ≤ maximum allowable adjustment amount in a unit sampling period", the control energy consumption constraint is transformed into the inequality "energy consumption value corresponding to the control command ≤ maximum allowable energy consumption value", and the aiming point offset constraint is transformed into the inequality "aiming point position deviation at any time during adjustment ≤ preset offset threshold". To address critical constraints that may arise in practical applications (such as attitude adjustment rate approaching maximum amplitude or offset approaching threshold), a constraint boundary softening mechanism is employed: a small boundary buffer is introduced into the constraint conditions. When the control quantity approaches the constraint boundary, it is no longer directly forcibly truncated. Instead, the control quantity is fine-tuned through a gradually changing penalty coefficient (such as the penalty coefficient gradually increasing as it approaches the boundary, causing the control quantity to slowly converge to within the boundary). This avoids sudden changes in the control quantity caused by forced constraints, which could lead to drastic fluctuations in the device's attitude.

[0045] Specifically, when solving the multi-objective optimization problem, the priority and weight allocation logic of the optimization objectives are first clarified: the interference suppression effect is the primary optimization objective, the attitude adjustment stability is the secondary objective, and the energy consumption control is the auxiliary objective. Based on this priority, corresponding weight coefficients are assigned to the three objectives respectively (the primary objective has the highest weight, and the auxiliary objective has the lowest weight). The specific values ​​of the weight coefficients are determined through performance testing of the equipment under different operating conditions (e.g., appropriately increasing the weight of the interference suppression effect in scenarios with strong interference, and appropriately increasing the weight of energy consumption control in scenarios with limited energy).

[0046] Next, a preset optimization time domain length is established: the model predicts the attitude change trend over multiple consecutive sampling periods. Based on the nonlinear attitude dynamics model, the current associated feature vector and disturbance rejection compensation amount are used as inputs to predict the changes in key variables such as device attitude angle, aiming point position, and control energy consumption in each future sampling period. Based on the prediction, and combined with the aforementioned transformed control constraints, an iterative optimization algorithm is used to solve the multi-objective optimization problem. Algorithms suitable for multi-objective optimization, such as non-dominated sorting genetic algorithms or sequential quadratic programming algorithms, are employed. Using the "weighted sum minimization of three optimization objectives" as the optimization criterion, the optimal combination of control quantities that satisfies the constraints is iteratively searched within each sampling period. These control quantity combinations are arranged in chronological order, thus generating the attitude compensation control sequence for multiple future sampling periods, ensuring that the control quantities in each sampling period can achieve multi-objective optimization effects while satisfying the constraints.

[0047] Input the associated feature vector V into AESO to achieve total disturbance estimation and disturbance rejection compensation: Feature extraction: Extract the motion state feature components v1 (ω''1) and v2 (a''1) of the aiming system from V, and use them as the input state of AESO; define the extended state E of AESO as the comprehensive disturbance term of the unmodeled dynamic U of the aiming system (such as the friction characteristics of the actuator) and the total environmental disturbance F (F=s''+b'', where s'' is the total slow-varying disturbance after quantization and b'' is the total sudden disturbance after quantization); Perturbation estimation: Based on the input states v1, v2 and the extended state E, the total perturbation estimate F_est (F_est=EU) is calculated in real time using the AESO observation equation. The observation gain adjustment logic is set as follows: if |F_est-F_act|>E_th (E_th is the perturbation estimation bias threshold, and F_act is the total actual observed interference), then the observation gain G is adjusted to G1 (G1>G, G is the initial gain); if |F_est-F_act|≤E_th, then the observation gain is kept at G. Disturbance compensation quantity generation: Based on the influence direction and intensity of the total disturbance estimate F_est, F_est is converted into a reverse compensation component C_comp (if F_est is a positive disturbance, C_comp is a negative compensation; if F_est is a negative disturbance, C_comp is a positive compensation), ensuring that the amplitude of C_comp is equal to that of F_est and the direction is opposite.

[0048] Nonlinear Model Predictive Controller (NMPC) Control Sequence Generation The associated feature vector V and the disturbance rejection compensation amount C_comp are input into NMPC to construct a control model and solve the optimization problem: Nonlinear attitude dynamics model construction: Based on the attitude motion law of the aiming system, a coupled motion equation for pitch angle θ and azimuth angle φ is established (the rate of change of θ θ=k_θ・a''1-f_θ・φ, the rate of change of φ φ=k_φ・ω''1-f_φ・θ, where k_θ and k_φ are attitude coefficients, and f_θ and f_φ are coupling coefficients); the inertial delay characteristics τ of θ and φ are incorporated into the model (the actual response lag of θ is τ_θ, the actual response lag of φ is τ_φ, τ_θ=τ・k_τθ, τ_φ=τ・k_τφ, where k_τθ and k_τφ are delay coefficients). At the same time, C_comp is incorporated into the equation as a disturbance compensation input term. The model parameters (k_θ, k_φ, f_θ, f_φ, τ) are verified by historical attitude motion data (such as the θ and φ change curves in the past T_hist time period) to ensure that the deviation between the model output and the actual motion is ≤M_th (M_th is the model accuracy threshold). Control constraint settings: Set three types of constraints: ① Actuator attitude adjustment rate constraint R_adj (the adjustment rate of θ ≤ R_θ, the adjustment rate of φ ≤ R_φ, and R_θ and R_φ are the upper limits of the rate); ② Control energy consumption constraint P_adj (actuator drive power P≤P_max, P_max is the maximum energy consumption); ③ Aiming point offset constraint Q_adj (actual aiming point offset Δd≤Q_max, where Q_max is the offset threshold); The three types of constraints are transformed into inequality constraints of the NMPC optimization problem (such as θ-R_θ ≤0, P-P_max≤0, Δd-Q_max≤0). The critical constraint case is handled by the constraint boundary softening mechanism (when θ is close to R_θ, a softening coefficient α is introduced to make the constraint θ-R_θ≤α・R_θ, where α is a softening factor). Multi-objective optimization solution: The optimization objectives are set as follows: ① interference suppression effect (suppression rate η ≥ η_min of F_est), ② energy consumption control (P ≤ P_avg, where P_avg is the average energy consumption), ③ attitude adjustment smoothness (the rate of change of θ and φ fluctuates Δθ ≤ Δθ_max, Δφ ≤ Δφ_max). Weights are assigned to each objective: W_obj1 (interference suppression), W_obj2 (energy consumption), and W_obj3 (smoothness) (W_obj1 > W_obj2 when η_min > P_avg). The prediction time domain length is preset to H (H consecutive sampling cycles). Based on the nonlinear attitude dynamics model, the trend of θ and φ changes within the next H T is predicted. Combined with the control constraints, the weighted sum minimization problem of "W_obj1・η+W_obj2・P+W_obj3・(Δθ+Δφ)" is solved by the gradient descent iterative optimization algorithm to generate a continuous attitude compensation control sequence C_ctl for the next H T (C_ctl contains the θ compensation amount Δθ and φ compensation amount Δφ in each T, C_ctl1=[Δθ1,Δφ1] corresponds to T1, C_ctl2=[Δθ2,Δφ2] corresponds to T2). Control execution and feedback: The control sequence C_ctl is sent to the servo actuator of the aiming system in real time, driving the actuator to adjust θ and φ according to C_ctl (e.g., adjust θ to θ1=θ0+Δθ1 according to Δθ1, where θ0 is the current pitch angle; adjust φ to φ1=φ0+Δφ1 according to Δφ1, where φ0 is the current azimuth angle); at the same time, the actual adjustment amount Δθ_act and Δφ_act of the actuator (e.g., the actual adjustment value of θ Δθ_act and the actual adjustment value of φ Δφ_act) and the aiming point position feedback data d_act (e.g., the actual position deviation d_act after adjustment) are sent back to AESO and NMPC to update the observation gain G (or G1) of AESO and the model parameters (e.g., k_θ and k_φ) of NMPC, forming a closed-loop control.

[0049] Specifically, the judgment process for triggering the dual closed-loop calibration process needs to combine deviation monitoring and scenario monitoring logic: In terms of deviation monitoring, the deviation value between the actual position of the aiming point and the theoretical position of the target after dynamic stabilization is calculated in real time, and a clear deviation judgment rule is set—when the deviation value exceeds the preset deviation threshold for multiple consecutive sampling periods, it is determined that the current device attitude or model parameters have deviated, and the calibration process needs to be initiated; the setting of the deviation threshold needs to be determined according to the aiming accuracy requirements of the device. In terms of scenario monitoring, two specific calibration scenarios are monitored: one is the power-on initialization scenario of the inertial measurement unit, at which time the device switches from the power-off state to the working state, and the inertial measurement unit may have initial drift due to long-term shutdown, and calibration is needed to ensure the accuracy of the initial attitude data; the other is the scenario of reaching the preset timed calibration cycle, at which time the device has been working continuously for a period of time, and the drift error of the inertial measurement unit may accumulate with the working time, and the influence of environmental interference on the model parameters may also gradually appear, and timed calibration is needed to eliminate accumulated errors and update model parameters. In actual judgment, deviation monitoring and scenario monitoring complement each other: deviation monitoring is used to deal with sudden, non-periodic deviations, while scenario monitoring is used to deal with periodic, cumulative deviations. Together, they ensure that the calibration process can be triggered in a timely manner when needed, avoiding a decrease in aiming accuracy due to failure to calibrate in time.

[0050] Specifically, when the inner loop processes the inertial measurement unit (IMU) calibration attitude data, it first preprocesses the data using the Kalman smoothing algorithm. This algorithm combines the predicted and observed values ​​of the data to perform posterior smoothing on the calibration attitude data, which can more effectively filter out short-term random errors compared to traditional filtering algorithms. In the application of the algorithm, an appropriate smoothing window length needs to be set—if the window length is too short, it cannot sufficiently smooth random errors; if the window length is too long, it will cause data lag and loss of detailed information about attitude changes. The specific value of the window length is determined by analyzing the noise characteristics of the calibration attitude data. When calculating the difference with the standard reference attitude, the comparison is performed point by point in the order of sampling time: first, the IMU calibration attitude angle (including pitch and azimuth angles) corresponding to each sampling period during the calibration process is obtained; then, the standard reference attitude (the ideal attitude angle pre-calibrated by a high-precision calibration device) is obtained; and the difference between the calibration attitude angle and the standard reference attitude angle (i.e., single-point drift error) in each sampling period is calculated point by point. When generating the cumulative drift error through integration, the single-point drift error of each sampling period is multiplied by the sampling time interval to obtain the drift error increment within that sampling period. Then, the drift error increments of all sampling periods are summed to obtain the cumulative drift error of the inertial measurement unit within the entire calibration period.

[0051] When updating the zero-bias compensation parameters, an incremental correction method is used: the correction increment is determined based on the magnitude of the cumulative drift error, and the angular velocity zero bias (used to compensate for fixed deviations in angular velocity measurement) and linear acceleration zero bias (used to compensate for fixed deviations in linear acceleration measurement) of the inertial measurement unit are adjusted incrementally, taking into account the error variation trend. When updating the scale coefficients, the correlation between the cumulative drift error and the attitude change range of the inertial measurement unit is analyzed, and the inter-axis coupling coefficient and temperature compensation coefficient are corrected based on this relationship.

[0052] Specifically, the process of fitting the outer loop deviation and mapping the interference relationship requires three steps: sample construction, model fitting, and mapping equation generation. In the sample construction stage, the difference between the calibration target reference data and the calibration position data is used as the deviation sample. The calibration target reference data is the ideal position data of the calibration target pre-calibrated by a high-precision measuring device, and the calibration position data is the actual position data measured when the aiming device is aligned with the calibration target. The difference between the two directly reflects the position deviation of the current aiming device. At the same time, the environmental interference features at the corresponding time are used as interference samples. The environmental interference features of each sampling period during the calibration process are extracted from the associated feature vector to ensure that each deviation sample corresponds to the interference sample at the same time. All deviation samples and corresponding interference samples are combined in chronological order to form a complete sample dataset. The number of samples must meet the fitting accuracy requirements.

[0053] In the model fitting stage, the least squares method is used to fit the sample dataset. This algorithm finds the model parameters that best reflect the relationship between deviation and interference by minimizing the sum of squares between the deviation samples and the predicted values ​​of the fitted model. Based on the actual correlation characteristics between deviation and interference, an appropriate fitting model type is selected—if data analysis reveals a linear relationship between deviation and interference, a linear fitting model is used; if a non-linear relationship is found, a non-linear fitting model is used. In the mapping equation generation stage, based on the fitted model parameters, a specific mapping equation between the deviation value and the interference feature is established: for example, linear fitting generates the equation "deviation value = interference feature × linear coefficient + constant term," while non-linear fitting generates the equation "deviation value = higher-order term of interference feature × corresponding coefficient + constant term." This mapping equation can quantitatively describe the variation law of deviation under different interference intensities, providing a basis for the subsequent generation of model error correction coefficients—through this equation, the theoretically existing deviation under the current environmental interference can be calculated, and the difference between the actual deviation and the theoretical deviation is used as the model error, thereby generating the corresponding correction coefficient.

[0054] Specifically, the dual-loop calibration process stores the deviation data, drift error data, and model correction coefficient data generated during the current calibration process according to timestamps and calibration scenarios, generating a historical calibration database. When the calibration process is triggered again, historical calibration data of the same or similar scenarios are first retrieved from the historical database, and the matching degree between the current scenario and the historical scenario is calculated through a similarity matching algorithm.

[0055] Calibration trigger judgment: Calculate in real time the deviation Δp (Δp=|p_act-p_ref|) between the actual position p_act of the aiming point and the theoretical position p_ref of the target after dynamic stabilization, and set the calibration trigger rule: Deviation Trigger: If Δp exceeds the offset threshold Q_max for N consecutive sampling periods T (N≥2), deviation calibration is triggered. Scenario triggering: When the IMU is powered on and initialized, initialization calibration is triggered (at this time, the zero bias parameter of the IMU needs to be initially corrected); when the system running time reaches the timing period T_cal (e.g., T_cal=K_cal・T, where K_cal is the period coefficient), timing calibration is triggered. In this embodiment, when Δp exceeds Q_max for two consecutive cycles T1 and T2, and the system running time reaches T_cal, the dual closed-loop calibration process is triggered.

[0056] Inner loop calibration (IMU drift calibration) Data acquisition: Obtain the reference data B (reference coordinates B_ref of the target center) of the preset high-precision calibration target, control the aiming system to align with the calibration target, and collect the calibration attitude data ω_cal (calibration angular velocity ω_cal1, ω_cal2) and a_cal (calibration linear acceleration a_cal1, a_cal2) of the IMU at this time, as well as the calibration position data d_cal (calibration position deviation d_cal1) of the aiming point detection module. Data processing: The Kalman smoothing algorithm is used to process ω_cal and a_cal. The smoothing window length W_sm is set (W_sm=W・K_sm, where W is the filter window and K_sm is the smoothing coefficient). Short-term random error ω_rand in ω_cal (such as ω_cal1 fluctuation caused by instantaneous noise) and short-term random error a_rand in a_cal (such as a_cal1 fluctuation caused by impact noise) are filtered out to obtain the smoothed calibration attitude data ω_cal_sm and a_cal_sm. Drift error calculation: Obtain the standard reference attitude data of the IMU, ω_std (standard angular velocity calibrated at the factory) and a_std (standard linear acceleration calibrated at the factory). Calculate the difference Δω between ω_cal_sm and ω_std (Δω=ω_cal_sm-ω_std) and the difference Δa between a_cal_sm and a_std (Δa=a_cal_sm-a_std) point by point based on the sampling order. Generate the cumulative drift error E_drift of the IMU (E_drift=∫Δωdt+∫Δadt, integration interval is [0,T_cal]) through integration (integrating Δω over the calibration period T_cal). Parameter Update: Based on the amplitude and trend of E_drift, the IMU's bias compensation parameters are updated using an incremental correction method: the angular velocity bias parameter Z_ω is adjusted to Z_ω_new = Z_ω_old + ΔZ_ω (ΔZ_ω is the increment calculated based on E_drift, ΔZ_ω ∝ the angular velocity drift component in E_drift), and the linear acceleration bias parameter Z_a is adjusted to Z_a_new = Z_a_old + ΔZ_a (ΔZ_a ∝ the linear acceleration drift component in E_drift); combined with E_drift and the attitude change range (such as the change range of θ, φ)... The correlation between [θ_min,θ_max] and [φ_min,φ_max] is determined, and the scale coefficients are corrected: the inter-axis coupling coefficient K_couple is adjusted to K_couple_new=K_couple_old・(1-ΔK_couple) (ΔK_couple is the correction amount calculated based on E_drift), and the temperature compensation coefficient K_temp is adjusted to K_temp_new=K_temp_old+ΔK_temp (ΔK_temp∝ the temperature drift component in E_drift, corresponding to environmental disturbance s).

[0057] Outer loop calibration (model parameter optimization) Mapping relationship fitting: The difference Δd_cal between the calibration target reference data B and the calibration position data d_cal (Δd_cal = |B-d_cal|, belonging to Δp) is used as the deviation sample. The environmental disturbance features s_cal (such as temperature drift s_cal1 during calibration, belonging to s) and b_cal (such as vibration intensity b_cal1 during calibration, belonging to b) at the corresponding time are used as disturbance samples to construct a sample dataset D_sample (D_sample contains multiple sets of [Δd_cal, s_cal, b_cal] data pairs). The least squares method is used to fit D_sample to establish a nonlinear mapping equation M_map (Δd_cal = A・s_cal) between the deviation value and the disturbance features. 2 +B・b_cal+C, where A, B, and C are fitting coefficients); Correction coefficient generation: Based on the fitting error ΔM of the mapping equation M_map (ΔM=|Δd_cal_est-Δd_cal_act|, where Δd_cal_est is the predicted value of M_map and Δd_cal_act is the actual deviation value), model error correction coefficient K_corr is generated (K_corr=1-ΔM / Δd_cal_max, where Δd_cal_max is the maximum deviation value in the sample set). Model parameter update: Update key model parameters in S3 based on K_corr: Adjust the observation gain G (or G1) of AESO to G_new=G_old・K_corr (if K_corr<1, it indicates that the observation bias is large, and G needs to be increased to improve the observation accuracy); adjust the optimized weight matrix W_obj of NMPC (including W_obj1, W_obj2, W_obj3) to W_obj_new=W_obj_old・K_corr (e.g., W_obj1_new=W_obj1_old・K_corr, to enhance the interference suppression weight); adjust the control constraint thresholds (Q_max, R_θ, P_max) of NMPC to Q_max_new=Q_max_old・K_corr, R_θ_new=R_θ_old・K_corr, P_max_new=P_max_old・K_corr to ensure that the constraints match the current model accuracy.

[0058] Historical calibration data management: The deviation data Δp (e.g., Δp1, Δp2), drift error data E_drift (e.g., E_drift1, E_drift2), and model correction coefficient data K_corr (e.g., K_corr1, K_corr2) generated during this calibration process are classified and stored according to timestamp T_stamp (e.g., calibration start time T_stamp1, end time T_stamp2) and calibration scenario S_cal (e.g., deviation calibration scenario S_cal1, timed calibration scenario S_cal2), generating a historical calibration database D_hist; When the calibration process is triggered again, historical calibration data D_hist_sim that is the same as or similar to the current scene S_cal_curr (or the historical data corresponding to S_cal2 if the current calibration is timed) is called from D_hist (e.g., E_drift and K_corr from the past 3 S_cal2 scenes). The matching degree S_sim (S_sim=cosθ, where θ is the angle between the current feature vector and the historical feature vector) is calculated by the cosine similarity algorithm between the current scene features (e.g., current s_curr, b_curr) and the historical scene features (e.g., historical s_hist, b_hist). If S_sim > S_th (S_th is the similarity threshold), then K_corr in D_hist_sim is used as the initial correction coefficient to shorten the iteration optimization time of this calibration.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for dynamic stabilization and automatic calibration of aiming points integrating inertial measurement unit data, characterized in that, include: S1: Acquire the attitude data output by the inertial measurement unit and construct a multi-source data matrix with the real-time position data of the aiming point and environmental interference data; S2: First, the attitude data, real-time position data of the aiming point, and environmental interference data are processed for time-series synchronization. Drift suppression is performed on the synchronized attitude data. Through a multimodal feature fusion mechanism, the data is mapped to a unified feature space, and the associated feature vectors of attitude-position-interference are extracted. S3: Estimate the total disturbance in real time based on the extended state, dynamically adjust the observation gain, and generate the disturbance rejection compensation amount; construct a nonlinear attitude dynamics model based on the associated feature vector and the disturbance rejection compensation amount, set control constraints, solve the multi-objective optimization problem within the sampling period, and generate the attitude compensation control sequence of the sampling period. S4: Real-time calculation of the deviation between the actual position of the aiming point and the theoretical position of the target after dynamic stabilization, triggering a dual closed-loop calibration process; acquiring calibration target reference data, inertial measurement unit calibration attitude data, and calibration position data; the inner loop processes the inertial measurement unit calibration attitude data, calculates the difference with the standard reference attitude, generates cumulative drift error, and updates the zero bias compensation parameters and calibration coefficients; the outer loop, based on the difference between the calibration target reference data and the calibration position, combines environmental interference characteristics, fits the deviation and interference mapping relationship, generates model error correction coefficients, and updates the observation gain, optimized weight matrix, and control constraint threshold.

2. The method according to claim 1, characterized in that, In S1, when constructing the multi-source data matrix, feature decomposition and association mapping are performed on various types of data: for the attitude data, three core features are extracted: attitude change rate, attitude angle deviation, and attitude stability index; for the real-time position data of the aiming point, it is decomposed into three features: position coordinate deviation, position change trend, and position fluctuation frequency; for the environmental interference data, it is classified into slow-varying interference features and sudden interference features; the features of the three types of data are aligned based on the time dimension, and a time series dimension is constructed with the same time interval as the unit. The attitude data features, aiming point position data features, and environmental interference data features in each time unit together constitute a data vector, which is arranged in chronological order to generate the multi-source data matrix including the time-data type-feature dimension.

3. The method according to claim 1, characterized in that, In S2, when performing the timing synchronization process, the internal clock of the inertial measurement unit is used as a reference, and the time stamp of the real-time position data of the aiming point and the environmental interference data are adjusted by a timestamp comparison algorithm. When performing drift suppression on the synchronized attitude data, a dynamic filtering algorithm based on the environmental interference data is adopted, and the filtering window and filtering coefficient are adjusted according to the temperature change amplitude and vibration intensity in the environment.

4. The method according to claim 1, characterized in that, In S2, when the multimodal feature fusion mechanism maps data to a unified feature space, it first performs feature standardization on the attitude data, real-time aiming point position data, and environmental interference data: the attitude data is normalized to a preset feature interval based on the angular velocity-linear acceleration dimension, the aiming point position data is normalized based on the target coordinate range, and the environmental interference data is quantized based on the interference intensity level; then, through feature correlation analysis, fusion weights are assigned, and finally, through feature splicing and dimensional compression, the dimensionally unified attitude-position-interference associated feature vector is generated.

5. The method according to claim 1, characterized in that, In S3, when processing the associated feature vector through the adaptive extended state observer, the motion state feature components of the aiming device are extracted as input states. When estimating the total disturbance in real time based on the extended state, the extended state is defined as a comprehensive disturbance term of the unmodeled dynamics of the aiming device and the total environmental interference, and the estimated value of the total disturbance is corrected. When dynamically adjusting the observation gain, based on the deviation between the estimated value of the total disturbance and the actual observed interference data, proportional-integral adjustment logic is used to generate the anti-interference compensation amount. The estimated total disturbance is transformed into a reverse compensation component based on the direction and intensity of the influence.

6. The method according to claim 1, characterized in that, In S3, when constructing the nonlinear attitude dynamics model, the coupled motion equations of pitch angle and azimuth angle are established based on the attitude motion law of the aiming device, incorporating the inertial delay characteristics of attitude change and the response delay characteristics of the actuator; the attitude data in the associated feature vector is used as the state input, the disturbance compensation amount is used as the disturbance compensation input, and the parameters are verified through historical attitude motion data.

7. The method according to claim 1, characterized in that, In S3, when setting the control constraints, the constraints include the attitude adjustment rate of the actuator, control energy consumption, and a preset threshold for the aiming point offset. During the transformation of the constraints, the constraints are transformed into inequality constraints of the optimization problem, and critical constraint situations are handled through a constraint boundary softening mechanism.

8. The method according to claim 1, characterized in that, In S3, when solving the multi-objective optimization problem, weights are assigned based on the three major optimization objectives: interference suppression effect, control energy consumption, and attitude adjustment. With a preset time domain length, the future attitude change trend is predicted based on the nonlinear attitude dynamics model. Combined with control constraints, the attitude compensation control sequence that minimizes the weighted sum of multiple objectives is solved through an iterative optimization algorithm.

9. The method according to claim 1, characterized in that, In S4, the judgment process for triggering the dual closed-loop calibration process includes: calculating the deviation between the actual position of the aiming point and the theoretical position of the target in real time, setting the deviation judgment rule; monitoring the calibration triggering scenario, triggering initialization calibration when the inertial measurement unit is powered on and initialized, and triggering timed calibration when the preset time period is reached.

10. The method according to claim 1, characterized in that, In S4, the inner loop uses the Kalman smoothing algorithm to process the calibration attitude data, sets the smoothing window length, and filters out short-term random errors in the calibration attitude data. When calculating the difference with the standard reference attitude, the attitude angle difference is calculated point by point based on the sampling order, and the cumulative drift error of the inertial measurement unit within the calibration period is generated through integration. When updating the zero bias compensation parameters, the angular velocity zero bias and linear acceleration zero bias are adjusted using the incremental correction method according to the amplitude and trend of the cumulative drift error. When updating the scale coefficients, the inter-axis coupling coefficient and temperature compensation coefficient are corrected by combining the correlation between drift error and attitude change range.

11. The method according to claim 1, characterized in that, In S4, the process of fitting the outer loop deviation and the interference mapping relationship includes: using the difference between the calibration target reference data and the calibration position data as the deviation sample, and using the environmental interference characteristics at the corresponding time as the interference sample to construct a sample dataset; using the least squares method to fit the sample dataset to establish a linear or nonlinear mapping equation between the deviation value and the interference characteristics.

12. The method according to claim 1, characterized in that, In S4, the dual closed-loop calibration process stores the deviation data, drift error data, and model correction coefficient data generated during the current calibration process according to timestamps and calibration scenarios, generating a historical calibration database. When the calibration process is triggered again, the historical calibration data of the same or similar scenarios are first retrieved from the historical database, and the matching degree between the current scenario and the historical scenario is calculated through a similarity matching algorithm.

Images