Level attitude real-time calibration method based on multi-modal data fusion

Abstract

The present application relates to the technical field of attitude measurement, and discloses a level attitude real-time calibration method based on multi-modal data fusion, comprising: collecting multi-modal original data and completing unified preprocessing to obtain multi-modal data; reconstructing liquid surface morphology at a physical domain neural operator layer, and outputting physical domain attitude and residual error; inferring noise and drift at a sensing domain neural operator layer, and outputting sensing domain attitude estimation value, bias parameter, scale parameter and residual error; establishing a deviation memory bank and updating it with residual error and historical results, generating long-term drift compensation and superimposing the sensing domain results; inputting the physical domain and the compensated sensing domain results into a dynamic constraint reversible transformation model to output fused attitude and uncertainty; and performing slow variable refinement compensation on the updated parameters to output the final real-time calibration attitude and quality information. The present application realizes real-time, high-precision and long-term stable calibration of level attitude by introducing multi-modal data fusion and double-layer reversible neural operator modeling.

Description

Technical Field

[0001] This invention relates to the field of attitude measurement technology, and in particular to a method for real-time attitude calibration of a level based on multimodal data fusion. Background Technology

[0002] Currently, level gauge attitude measurement technology is widely used in engineering surveying, precision manufacturing, building construction, and equipment installation. Existing levels mostly rely on a single sensor as their primary information source, such as using an inertial measurement unit (IMU) to collect acceleration and angular velocity, or depending on a magnetometer for heading calibration, supplemented by simple filtering methods to achieve attitude calculation. While existing methods can meet basic measurement needs under static conditions, they have significant shortcomings in complex environments or long-term use. On the one hand, IMUs are susceptible to zero-bias drift and random noise, causing attitude calculation results to accumulate errors over time, gradually reducing accuracy. On the other hand, magnetometers are easily affected by external magnetic field disturbances, resulting in unstable heading calculations. Although bubble optical detection methods can provide a physical reference, they typically fail to deeply integrate with other sensor information; their output is often only used as an auxiliary signal, making it difficult to guarantee the continuity and accuracy of attitude measurement under dynamic conditions.

[0003] To address the aforementioned shortcomings, existing technologies often employ Kalman filtering or particle filtering for multi-source data fusion, relying heavily on linear or statistical assumptions. This makes them ineffective in handling complex nonlinear noise environments and lacks adaptive compensation mechanisms for long-term drift problems. The fusion results are often "black box" outputs, lacking close integration with physical constraints, making it prone to situations where the calculated results are physically unrealizable. For example, the geometric constraints and surface tension characteristics of the bubble morphology are not intrinsically modeled, potentially leading to output attitude angles that do not match actual liquid surface characteristics, thus affecting reliability.

[0004] Therefore, how to provide a real-time calibration method for level instrument attitude based on multimodal data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a real-time level attitude calibration method based on multimodal data fusion. This invention fully utilizes a two-layer reversible fusion mechanism of physical domain neural operators and sensor domain neural operators, combined with a self-consistent deviation memory unit to achieve long-term drift compensation, and ensures that the output results conform to physical constraints under a dynamic constraint reversible transformation model. The steps and procedures for achieving accurate level attitude calibration in complex environments are described in detail. This invention possesses advantages such as strong real-time performance, good long-term stability, interpretable results, and physical plausibility, effectively overcoming the shortcomings of existing technologies such as reliance on a single sensor, limited filtering accuracy, drift accumulation, and unstable output.

[0006] A real-time attitude calibration method for a level based on multimodal data fusion according to an embodiment of the present invention includes:

[0007] Collect raw multimodal data, preprocess the raw multimodal data, and obtain preprocessed multimodal data;

[0008] In the physical domain neural operator layer, liquid surface morphology reconstruction and operator inversion are performed based on multimodal data. The normal distribution and curvature distribution of the bubble interface are extracted, liquid surface observation features are generated, and physical domain attitude estimation results and physical residuals are output.

[0009] In the neural operator layer of the sensing domain, operator inference and attitude calculation are performed based on multimodal data to deal with noise and drift, and the sensing domain attitude estimation results, bias parameters, scale parameters and sensing residuals are output.

[0010] A bias memory bank is established, and the bias memory bank is updated using physical residuals, sensing residuals and historical fusion results to generate long-term drift compensation. The long-term drift compensation is then superimposed on the sensor domain attitude estimation result to obtain the compensated sensor domain attitude estimation result.

[0011] The physical domain attitude estimation results, the compensated sensor domain attitude estimation results, and the bias and scale parameters are input into the dynamic constraint reversible transformation model. Reversible coupling transformation is performed under the constraints of angle interval, bubble curvature range, and magnetic disturbance angle range to complete the fusion of two-layer reversible neural operators. The fused attitude estimation results, uncertainty parameters, and updated bias and scale parameters are output. The fused attitude estimation results are then subjected to dual-domain consistency optimization.

[0012] Based on temperature and strain data from multimodal data, slow variable refinement compensation is performed on the updated bias and scale parameters, outputting the final real-time calibration attitude result along with corresponding uncertainty information, modal contribution rate, and anomaly label.

[0013] Optionally, the multimodal raw data includes acceleration and angular velocity data collected by the inertial measurement unit, magnetic field data collected by the magnetometer, bubble image data acquired by the camera and ring light source, and temperature and strain data collected by the temperature sensor and strain sensor.

[0014] Optionally, the preprocessing of the multimodal raw data includes filtering, timestamp synchronization, and data normalization.

[0015] Optionally, the output physical domain attitude estimation result and physical residual include:

[0016] Under phase-coded ring illumination, based on bubble image data, radiometric response calibration, geometric distortion correction, glare suppression and temporal denoising are performed to obtain a preprocessed bubble image sequence.

[0017] Based on the bubble image sequence and cavity geometry prior, liquid surface morphology reconstruction is performed. The technique of combining photometric stereo reconstruction and contact line segmentation is used to extract the normal distribution and curvature distribution of the bubble interface, and generate a liquid surface observation feature set including normal map, curvature map and boundary mask.

[0018] A physical domain neural operator layer is constructed, which consists of a boundary condition encoder, a spectral domain kernel mapping backbone, a cross-scale geometric consistency aggregation unit, a physical constraint gating unit, and an inversion solver head, wherein:

[0019] The boundary condition encoder encodes cavity geometry, contact angle, surface tension, filling height, and temperature parameters;

[0020] The spectral domain kernel mapping backbone uses a combination of Fourier neural operators and low-rank kernel approximation to realize the function mapping from liquid surface observation features to gravity direction related quantities.

[0021] The cross-scale geometric consistency aggregation unit uses feature pyramids and cross-layer connections to align features across multiple scales.

[0022] The physical constraint gating unit adjusts the mapping process based on the relationship between liquid pressure and curvature, as well as the contact angle boundary conditions;

[0023] The inversion solver outputs the physical domain attitude estimation results and physical residuals.

[0024] The operator inversion process is performed, and the liquid surface observation feature set and boundary condition encoding are input into the physical domain neural operator layer. First, the initial physical domain attitude estimation result and residual are generated. Then, the estimation result is updated through a finite number of iterations under the adjustment of the physical constraint gating unit until the residual meets the threshold or reaches the upper limit of the number of iterations. Finally, the physical domain attitude estimation result and physical residual vector are output.

[0025] The physical domain attitude estimation results and physical residual vectors are subjected to quality measurement and region screening. Modal confidence coefficients are generated based on the consistency between residual distribution and reprojection, and abnormal regions are marked.

[0026] Optionally, the output sensing domain attitude estimation result, bias parameter, scale parameter, and sensing residual include:

[0027] Based on multimodal data, acceleration data, angular velocity data, magnetic field data, temperature data, and strain data are organized within a fixed-length time window to form a time-aligned observation sequence;

[0028] A sensor domain neural operator layer is constructed, which consists of three parts: a time-frequency kernel mapping backbone, a cross-modal consistency coupling unit, and a calibration and compensation solver head.

[0029] The time-frequency kernel mapping backbone adopts a joint structure of time-domain convolution stacking and frequency-domain spectrum to extract zero-bias drift, scale drift and random walk patterns from the observation sequence;

[0030] The cross-modal consistency coupling unit uses gravity-acceleration consistency, angular velocity-attitude change consistency and geomagnetic-heading consistency as gating signals to dynamically allocate the weights of each channel;

[0031] The calibration and compensation solver outputs the sensor domain attitude estimation results, bias parameters and scale parameters, and generates sensor residuals.

[0032] Operator inference for noise and drift is performed, and the observation sequence is processed in the time-frequency kernel mapping backbone to identify and characterize noise and drift modes. In the cross-modal uniform coupling unit, the disturbed channel is suppressed and the stable channel is enhanced based on the uniformity deviation. In the calibration and compensation solver head, bias parameters and scale parameters are generated based on temperature and strain data to form a compensated observation sequence.

[0033] The attitude calculation is performed, and the observation sequence is decoded into the attitude estimation result of the sensor domain in the sensor domain neural operator layer. The attitude estimation result of the sensor domain includes pitch angle, roll angle and heading angle, and the sensor residual is output, which consists of acceleration consistency deviation, angular velocity consistency deviation and magnetic field consistency deviation.

[0034] Optionally, obtaining the compensated sensor domain pose estimation result includes:

[0035] A data structure for constructing a deviation memory bank is provided. The data structure consists of a working condition fingerprint key, a deviation state value, a compensation calculation parameter value, and a quality measurement value. The working condition fingerprint key is formed by combining temperature data, strain data, physical residual features, and sensor residual features within a fixed time window.

[0036] Within a preset time window, physical residuals and sensor residuals are collected and combined with temperature data and strain data. The deviation memory is searched according to the working condition fingerprint key. When a match is found, the deviation state value, compensation calculation parameter value and quality measurement value of the corresponding entry are read. When no match is found, a new entry is created and initialized.

[0037] An update method combining exponential forgetting and time-series accumulation is used to update the deviation status value of the hit item. At the same time, the quality metric value is updated according to the residual amplitude distribution, duration and stability. The compensation calculation parameter value is then stratified and corrected according to the updated quality metric value.

[0038] Using the updated deviation state value as input, the physical residual and sensor residual are first feature-encoded, and then the compensation calculation parameter values ​​are gated and weighted according to the working condition fingerprint key. The long-term drift compensation amount is calculated according to the adjusted parameters, and the compensation amount in adjacent time windows is checked for consistency through forward and backward time consistency checks.

[0039] The long-term drift compensation is superimposed on the sensor domain attitude estimation result to obtain the compensated sensor domain attitude estimation result. The compensated result, along with the corresponding working condition fingerprint key, the updated deviation state value, the compensation calculation parameter value, and the quality metric value, is written back to the deviation memory for updating in the next time window.

[0040] Optionally, the output includes the fused attitude estimation result, uncertainty parameter, and updated bias parameter and scale parameter, and performs bi-domain consistency optimization on the fused attitude estimation result, including:

[0041] The input to the dynamic constraint reversible transformation model includes the physical domain attitude estimation results, the compensated sensor domain attitude estimation results, and the bias parameters and scale parameters.

[0042] A dynamic constraint reversible transformation model is constructed, which consists of a reversible coupling stack, a constraint manager, and a dual-domain consistency optimizer. The reversible coupling stack supports forward and backward mapping and contains margin variables to maintain reversibility. The constraint manager is used to apply angular interval constraints, bubble curvature range constraints, and magnetic disturbance angle range constraints within the mapping. The dual-domain consistency optimizer is used to measure the differences between the fused attitude and the physical domain attitude, and between the fused attitude and the compensated sensor domain attitude, and serves as the convergence criterion.

[0043] Forward inference is performed in the reversible coupling layer stack to generate the initial fused pose, the initial updated bias parameters and scale parameters, and the margin variables. During the inference process, the constraint manager performs endogenous constraints on intermediate results that do not conform to the angle, curvature, and magnetic disturbance range.

[0044] The dual-domain consistency optimizer is invoked to perform a finite-step iterative update while maintaining reversibility. The decrease in difference metric is used as the criterion, and the violation of angle, curvature and magnetic disturbance range is included as a penalty term in the objective until the residual threshold or the upper limit of the number of iterations is reached, so as to obtain the converged fusion pose and the updated bias parameters and scale parameters.

[0045] An uncertainty parameter is generated based on a combination of physical residuals, sensing residuals, and consistency differences, so that the uncertainty increases monotonically as the residuals and differences increase.

[0046] The system outputs the fused pose estimation results, updated bias parameters, updated scale parameters, and uncertainty parameters, thus completing the fusion of two-layer reversible neural operators.

[0047] Optionally, the output of the final real-time calibration attitude result and the corresponding uncertainty information, modal contribution rate, and anomaly label includes:

[0048] Based on temperature and strain data, timestamp alignment is performed to form temperature and strain sequences within a fixed-length time window, and the mean, range, and rate of change of the current interval are calculated.

[0049] Based on historical operation records and updated bias and scale parameters, the system is divided into segments according to the range of temperature and strain values. The parameter statistics within each segment are summarized into baseline values ​​and incremental coefficients to form a segmented compensation coefficient table, and monotonicity and upper and lower boundary thresholds are set.

[0050] The current temperature and strain are searched in the segmented compensation coefficient table. The instantaneous compensation coefficient is obtained by linear interpolation of adjacent segments. Boundary coefficients are used for out-of-bounds cases, and median filtering and amplitude limiting smoothing are used for abrupt changes to obtain a stable set of compensation coefficients.

[0051] The compensation coefficient set is applied to the updated bias parameters and scale parameters respectively to obtain the refined bias parameters and scale parameters. The fused pose is then recalculated to obtain the refined fused pose.

[0052] The uncertainty parameters are updated based on the physical residual, sensing residual, and the difference in consistency between the two domains. The modal contribution rates of the physical domain and the sensor domain of the bubble are calculated. Anomaly labels are generated when the temperature or strain exceeds the limit, the residual suddenly increases, or the hard constraint is triggered.

[0053] The beneficial effects of this invention are:

[0054] This invention introduces a multimodal data fusion mechanism into the attitude measurement of a level, fully utilizing inertial measurement units, magnetometers, bubble images, and temperature and strain information to achieve complementarity and enhancement of multi-source information. Unlike existing technologies that rely on a single sensor or use simple filtering methods, this invention establishes a two-layer structure of physical domain neural operators and sensor domain neural operators, simultaneously modeling bubble morphology and sensor noise characteristics. This ensures that the attitude estimation results conform to physical laws and possess adaptive anti-interference capabilities, improving the accuracy of real-time measurements.

[0055] During long-term operation, this invention achieves continuous updating and utilization of historical residual patterns through the introduction of a deviation memory library, automatically generating long-term drift compensation amounts and correcting sensor calculation results. This mechanism compensates for the inability of existing filtering methods to effectively handle accumulated drift, enabling the system to possess empirical self-correction capabilities. Even in complex environments or under conditions of gradual sensor performance degradation, this invention can still maintain the stability and consistency of attitude output, ensuring the reliable completion of long-term measurement tasks.

[0056] This invention introduces a dynamic constraint reversible transformation model into the fusion process, intrinsically embedding angular range constraints, bubble curvature range constraints, and magnetic disturbance angle range constraints during attitude calculation. This fundamentally avoids situations where the output results are physically unrealizable. Compared to the existing approach of calculating first and then trimming, the results of this invention have higher physical rationality and interpretability. This invention not only improves real-time performance and stability but also ensures the physical consistency and traceability of measurement results, expanding the application range of levels in high-precision measurements and complex working conditions. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart of the real-time calibration method for level instrument attitude based on multimodal data fusion proposed in this invention;

[0059] Figure 2 This is a schematic diagram of the dynamic constraint reversible transformation model and dual-domain consistency optimization process of the real-time calibration method for level instrument attitude based on multimodal data fusion proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figure 1 and Figure 2 A real-time attitude calibration method for a level instrument based on multimodal data fusion includes:

[0062] Collect raw multimodal data, preprocess the raw multimodal data, and obtain preprocessed multimodal data;

[0063] In the physical domain neural operator layer, liquid surface morphology reconstruction and operator inversion are performed based on multimodal data. The normal distribution and curvature distribution of the bubble interface are extracted, liquid surface observation features are generated, and physical domain attitude estimation results and physical residuals are output.

[0064] In the neural operator layer of the sensing domain, operator inference and attitude calculation are performed based on multimodal data to deal with noise and drift, and the sensing domain attitude estimation results, bias parameters, scale parameters and sensing residuals are output.

[0065] A bias memory bank is established, and the bias memory bank is updated using physical residuals, sensing residuals and historical fusion results to generate long-term drift compensation. The long-term drift compensation is then superimposed on the sensor domain attitude estimation result to obtain the compensated sensor domain attitude estimation result.

[0066] The physical domain attitude estimation results, the compensated sensor domain attitude estimation results, and the bias and scale parameters are input into the dynamic constraint reversible transformation model. Reversible coupling transformation is performed under the constraints of angle interval, bubble curvature range, and magnetic disturbance angle range to complete the fusion of two-layer reversible neural operators. The fused attitude estimation results, uncertainty parameters, and updated bias and scale parameters are output. The fused attitude estimation results are then subjected to dual-domain consistency optimization.

[0067] Based on temperature and strain data from multimodal data, slow variable refinement compensation is performed on the updated bias and scale parameters, outputting the final real-time calibration attitude result along with corresponding uncertainty information, modal contribution rate, and anomaly label.

[0068] In this embodiment, the multimodal raw data includes acceleration and angular velocity data collected by the inertial measurement unit, magnetic field data collected by the magnetometer, bubble image data acquired by the camera and ring light source, and temperature and strain data collected by the temperature sensor and strain sensor.

[0069] In this embodiment, the preprocessing of the multimodal raw data includes filtering, timestamp synchronization, and data normalization.

[0070] In this embodiment, the output physical domain attitude estimation result and physical residual include:

[0071] Under phase-coded ring illumination, based on bubble image data, radiometric response calibration, geometric distortion correction, glare suppression, and temporal denoising are performed to obtain a preprocessed bubble image sequence, wherein:

[0072] Performing radiation response calibration refers to collecting and fitting the brightness response curves of the bubble image sensor under different light intensity conditions, eliminating the nonlinear deviation between the sensor output and the actual illumination, so that the brightness value of the bubble area can accurately reflect the true illumination intensity distribution.

[0073] Geometric distortion correction refers to correcting barrel or pincushion distortion caused by the lens in a bubble image based on the camera's intrinsic parameters and distortion parameters, so that the bubble boundary and cavity structure maintain their true geometric proportions and shapes in the image.

[0074] Glare suppression and temporal denoising refer to eliminating artifacts caused by high-brightness reflections on the surface of liquid bubbles through spatial filtering and polarization constraint, while suppressing random noise using median filtering of multi-frame time series.

[0075] Liquid surface morphology reconstruction is performed based on bubble image sequences and cavity geometry priors. A combination of photometric stereo reconstruction and contact line segmentation is employed to extract the normal and curvature distributions of the bubble interface, generating a liquid surface observation feature set containing normal maps, curvature maps, and boundary masks. Specifically, the combination of photometric stereo reconstruction and contact line segmentation involves:

[0076] Photometric stereo reconstruction is based on the response relationship of pixel brightness on the surface of a bubble under ring multi-directional illumination as a function of incident direction, and the normal distribution of the bubble surface is derived.

[0077] Contact line segmentation is based on cavity geometry priors, segmenting the contact positions between the bubble and the cavity boundary from the bubble image and generating boundary line information;

[0078] The combination of the two provides surface normal distribution and boundary geometric constraints for vacuole interface reconstruction, forming an observation feature set that includes normal map, curvature map and boundary mask;

[0079] A physical domain neural operator layer is constructed, which consists of a boundary condition encoder, a spectral domain kernel mapping backbone, a cross-scale geometric consistency aggregation unit, a physical constraint gating unit, and an inversion solver head, wherein:

[0080] The boundary condition encoder encodes cavity geometry, contact angle, surface tension, filling height, and temperature parameters, specifically:

[0081] The cavity geometry, contact angle parameters, and surface tension coefficient are represented by feature vectors of fixed dimensions, and an initial encoding for characterizing boundary conditions is generated through an embedding matrix.

[0082] The filling height and temperature parameters are normalized as continuous variables and mapped to the same feature space, and then concatenated with the initial encoding to form a complete boundary condition embedding vector.

[0083] The spectral domain kernel mapping backbone uses a combination of Fourier neural operators and low-rank kernel approximation to achieve a functional mapping from observed features of the liquid surface to gravity-related quantities, specifically:

[0084] The bubble normal distribution, curvature distribution, and boundary mask are transformed to the frequency domain, and the Fourier basis function expansion is used to calculate different frequency components to obtain the multi-scale feature representation of the liquid surface morphology in the frequency domain.

[0085] The frequency domain features are approximated by a low-rank kernel decomposition to extract the main components and compress redundant information, generating a compact response quantity related to the direction of gravity.

[0086] The cross-scale geometric consistency aggregation unit uses feature pyramids and cross-layer connections for multi-scale feature alignment, specifically:

[0087] Hierarchical feature representations of bubbly normal distribution and curvature distribution are constructed at different scales to form a top-down feature pyramid structure;

[0088] By registering and weighting features at adjacent scales through cross-layer connections, an aggregated representation that maintains geometric consistency across the entire scale is obtained.

[0089] The physical constraint gating unit adjusts the mapping process based on the relationship between liquid pressure and curvature, as well as the contact angle boundary conditions, specifically:

[0090] Based on the mathematical relationship between liquid static pressure and liquid surface curvature, pressure consistency adjustment is applied to the normal distribution and curvature distribution generated during the mapping process;

[0091] Based on the contact angle boundary conditions, the characteristics of the contact area between the bubble and the cavity are modified by boundary constraints so that the mapping process conforms to the geometric relationship between the liquid and solid interface within the boundary range.

[0092] The inversion solver outputs the physical domain attitude estimation results and physical residuals, specifically:

[0093] The output features of the spectral domain kernel mapping backbone and the cross-scale geometric consistency aggregation unit are input into the linear transformation and normalization sequence to generate physical domain attitude estimation results composed of pitch angle, roll angle and heading angle.

[0094] The difference between the physical domain attitude estimation results and the constraints provided by the boundary condition encoder is calculated to obtain the physical residual vector composed of the bubble curvature deviation and the normal deviation.

[0095] The operator inversion process is performed, and the liquid surface observation feature set and boundary condition encoding are input into the physical domain neural operator layer. First, the initial physical domain attitude estimation result and residual are generated. Then, the estimation result is updated through a finite number of iterations under the adjustment of the physical constraint gating unit until the residual meets the threshold or reaches the upper limit of the number of iterations. Finally, the physical domain attitude estimation result and physical residual vector are output.

[0096] The physical domain attitude estimation results and physical residual vectors are subjected to quality measurement and region screening. Modal confidence coefficients are generated based on the consistency between residual distribution and reprojection, and abnormal regions are marked.

[0097] This invention introduces bubble image reconstruction under phase-encoded illumination, morphological estimation combining photometric stereo and contact line segmentation, and a physical domain neural operator layer composed of a boundary condition encoder, spectral kernel mapping, cross-scale aggregation, physical constraint gating, and an inversion solver head during the physical domain modeling process. This achieves a deep physical-data fusion unprecedented in traditional level technology. It not only accurately acquires the normal and curvature distribution of the bubble at the image level but also intrinsically introduces liquid pressure relationships and contact angle constraints into the operator layer, ensuring that the output results are physically interpretable and have consistent boundaries. It overcomes the drift accumulation and instability caused by existing technologies relying on empirical filtering and linear assumptions, improves the stability and accuracy of attitude estimation, and has the ability to automatically label abnormal regions, thereby enabling long-term reliable calibration of the level attitude in complex environments.

[0098] In this embodiment, the output sensing domain attitude estimation result, bias parameter, scale parameter, and sensing residual include:

[0099] Based on multimodal data, acceleration data, angular velocity data, magnetic field data, temperature data, and strain data are organized within a fixed-length time window to form a time-aligned observation sequence;

[0100] A sensor domain neural operator layer is constructed, which consists of three parts: a time-frequency kernel mapping backbone, a cross-modal consistency coupling unit, and a calibration and compensation solver head.

[0101] The time-frequency kernel mapping backbone adopts a joint structure of time-domain convolution stacking and frequency-domain spectrum mixing to extract zero-bias drift, scale drift, and random walk patterns from the observation sequence, specifically:

[0102] The acceleration, angular velocity and magnetic field observation sequences are input into a time-domain convolution stack structure. Multi-layer convolution and normalization operations are performed in consecutive time steps to extract the temporal cumulative features of zero bias drift and proportional drift, forming a multi-scale time-domain feature representation.

[0103] The multi-scale time-domain feature representation is further transformed to the frequency domain. The frequency domain spectrum mixing operation is used to decompose and combine the low-frequency slowly changing components and the high-frequency random walk components, and then jointly embeds them with the time-domain feature representation to generate a comprehensive feature representation that includes zero bias drift, proportional drift and random walk patterns.

[0104] The cross-modal consistency coupling unit uses gravity-acceleration consistency, angular velocity-attitude change consistency, and geomagnetic-heading consistency as gating signals to dynamically allocate the weights of each channel, specifically as follows:

[0105] The differences between gravity direction estimation and acceleration observation, angular velocity integral attitude and current attitude estimation, and geomagnetic observation direction and heading angle estimation are calculated as consistency metric signals and used as gating inputs.

[0106] Based on the amplitude and stability of the consistency metric signal, the weight coefficients of the acceleration channel, angular velocity channel and geomagnetic channel are dynamically allocated and normalized to generate a unified feature representation after cross-modal fusion.

[0107] The calibration and compensation solver outputs the sensor domain attitude estimation results, bias parameters, and scale parameters, and generates sensor residuals, specifically:

[0108] The output features of the time-frequency kernel mapping backbone and the cross-modal consistency coupling unit are input into the linear mapping and normalization sequence to solve for the attitude estimation results of the sensing domain consisting of pitch angle, roll angle and heading angle, and the corresponding bias parameters and scale parameters are output at the same time.

[0109] The difference between the attitude estimation results in the sensing domain and the gated consistency signal provided by the cross-modal consistency coupling unit is calculated to obtain the acceleration residual, angular velocity residual and geomagnetic residual, and then combined to generate the sensing residual vector.

[0110] Operator inference for noise and drift is performed, and the observation sequence is processed in the time-frequency kernel mapping backbone to identify and characterize noise and drift modes. In the cross-modal uniform coupling unit, the disturbed channel is suppressed and the stable channel is enhanced based on the uniformity deviation. In the calibration and compensation solver head, bias parameters and scale parameters are generated based on temperature and strain data to form a compensated observation sequence.

[0111] The attitude calculation is performed, and the observation sequence is decoded into the attitude estimation result of the sensor domain in the sensor domain neural operator layer. The attitude estimation result of the sensor domain includes pitch angle, roll angle and heading angle, and the sensor residual is output, which consists of acceleration consistency deviation, angular velocity consistency deviation and magnetic field consistency deviation.

[0112] This invention introduces a joint structure in the sensor domain modeling, combining a time-frequency kernel mapping backbone, a cross-modal consistency coupling unit, and a calibration and compensation solver head. This overcomes the limitations of existing models that rely on single filtering or statistical models, enabling the simultaneous capture of zero-bias drift, proportional drift, and random walk patterns. It also dynamically adjusts the fusion weights of multi-channel data, effectively separating and characterizing long-term drift and transient noise in the observation sequence. Furthermore, it achieves optimal combination of acceleration, angular velocity, and geomagnetic data under consistency gating. By incorporating temperature and strain information to compensate for bias and scale parameters, the impact of environmental disturbances on solution stability is effectively addressed. This improves the accuracy and robustness of attitude calculation results, avoids long-term accumulated errors, and ensures that the output sensor domain attitude estimation and sensing residuals maintain physical rationality and long-term stability even under complex operating conditions.

[0113] In this embodiment, obtaining the compensated sensor domain pose estimation result includes:

[0114] A data structure for constructing a deviation memory bank is provided. The data structure consists of a working condition fingerprint key, a deviation state value, a compensation calculation parameter value, and a quality measurement value. The working condition fingerprint key is formed by combining temperature data, strain data, physical residual features, and sensor residual features within a fixed time window.

[0115] Within a preset time window, physical residuals and sensor residuals are collected and combined with temperature data and strain data. The deviation memory is searched according to the working condition fingerprint key. When a match is found, the deviation state value, compensation calculation parameter value and quality measurement value of the corresponding entry are read. When no match is found, a new entry is created and initialized.

[0116] An update method combining exponential forgetting and time-sequential accumulation is used to update the deviation state value of the hit items. Simultaneously, the quality metric is updated based on the residual amplitude distribution, duration, and stability. The compensation calculation parameter values ​​are then stratified and corrected according to the updated quality metric values. Specifically, updating the deviation state value of the hit items involves:

[0117] An exponential forgetting factor is used to weight and attenuate historical biases, and combined with the current observation residuals, a bias estimate with time memory characteristics is formed.

[0118] The residual values ​​at multiple time points are subjected to sliding weighting based on the time sequence of the residual series to highlight the trend characteristics of the deviation.

[0119] The result of the exponential forgetting update is combined with the result of the time accumulation to obtain the comprehensive corrected deviation state value;

[0120] Using the updated deviation state value as input, the physical residual and sensing residual are first feature-encoded. Then, the compensation calculation parameter values ​​are gated and weighted according to the operating condition fingerprint key. The long-term drift compensation amount is calculated according to the adjusted parameters. The consistency of the compensation amount within adjacent time windows is checked through forward and backward time consistency checks. Specifically, the gated selection and weight adjustment of the compensation calculation parameter values ​​based on the operating condition fingerprint key, and the calculation of the long-term drift compensation amount according to the adjusted parameters, are as follows:

[0121] Input the working condition fingerprint key into the gate control network, and trigger the corresponding subset of compensation parameters according to the working condition characteristics to achieve the filtering and shielding of redundant parameters;

[0122] Within the selected subset of compensation parameters, the parameter weights are dynamically adjusted based on the similarity and stability indices of the working condition fingerprint keys, so that the parameter allocation matches the working condition status.

[0123] Using the parameter set after gating and weighting adjustment, compensation calculations are performed on the updated deviation state values ​​to obtain the corresponding long-term drift compensation amount;

[0124] The long-term drift compensation is superimposed on the sensor domain attitude estimation result to obtain the compensated sensor domain attitude estimation result. The compensated result, along with the corresponding working condition fingerprint key, the updated deviation state value, the compensation calculation parameter value, and the quality metric value, is written back to the deviation memory for updating in the next time window.

[0125] This invention introduces a deviation memory structure with a working condition fingerprint key in the deviation compensation stage, breaking through the traditional method of relying on short-term filtering and static parameter compensation, and realizing a long-term drift compensation mechanism based on environmental state and residual characteristics. By combining temperature, strain, physical residuals, and sensor residuals to form a working condition fingerprint key, the system can classify, store, and retrieve deviation patterns under different working conditions, and dynamically update the deviation state value within a time window. Combining exponential forgetting and time-sequential accumulation update methods, the deviation memory has time memory characteristics, which can effectively capture long-term trend changes. Through gating selection and weight adjustment, redundant compensation parameters are screened and matched with the working condition state, ensuring that the calculated drift compensation amount is both targeted and stable, improving the adaptive ability of attitude estimation in complex environments and long-term operation, overcoming the cumulative drift problem, making the compensation results more accurate and reliable, and realizing the long-term stability and physical rationality of level attitude calculation.

[0126] In this embodiment, the output includes the fused attitude estimation result, uncertainty parameter, updated bias parameter and scale parameter, and the two-domain consistency optimization is performed on the fused attitude estimation result, including:

[0127] The input to the dynamic constraint reversible transformation model includes the physical domain attitude estimation results, the compensated sensor domain attitude estimation results, and the bias parameters and scale parameters.

[0128] A dynamic constraint reversible transformation model is constructed, which consists of a reversible coupling stack, a constraint manager, and a dual-domain consistency optimizer. The reversible coupling stack supports forward and backward mapping and contains margin variables to maintain reversibility. The constraint manager is used to apply angular interval constraints, bubble curvature range constraints, and magnetic disturbance angle range constraints within the mapping. The dual-domain consistency optimizer is used to measure the differences between the fused attitude and the physical domain attitude, and between the fused attitude and the compensated sensor domain attitude, and serves as the convergence criterion.

[0129] Forward inference is performed in the reversible coupled layer stack to generate the initial fused pose, the initially updated bias and scale parameters, and the margin variable. During inference, the constraint manager applies intrinsic constraints to intermediate results that do not conform to the angle, curvature, and magnetic disturbance ranges. Specifically, the forward inference in the reversible coupled layer stack involves:

[0130] The physical domain attitude estimation results, the compensated sensor domain attitude estimation results, and the bias and scale parameters are input into a multi-layer reversible coupling stack. In each layer, coupling mapping and reverse recoverable operation are performed to generate the initial fused attitude and the corresponding bias and scale parameter updates layer by layer.

[0131] In the forward propagation process, a margin variable is introduced to explicitly represent the unmodeled components remaining in the attitude calculation process, and the initial fused attitude, the initial updated bias parameters and scale parameters, and the margin variable are used as the output of the reversible coupled stack.

[0132] The constraint manager applies intrinsic constraints to intermediate results that do not conform to the angle, curvature, and magnetic disturbance range during the inference process, specifically as follows:

[0133] After each layer of reversible coupling calculation is completed, the generated intermediate attitude vector is checked for angle intervals, and the pitch angle, roll angle and yaw angle are limited to the preset effective range. The bubble curvature feature components are checked for boundaries, and values ​​that exceed the range are thresholded and reduced.

[0134] The perturbation amplitude of the component containing the magnetic field direction is detected, the difference with the current environmental magnetic field reference value is calculated, and when the difference exceeds the set threshold, the weight and amplitude of the relevant components are automatically adjusted, and the corrected intermediate results are fed back.

[0135] The dual-domain consistency optimizer is invoked to perform a finite-step iterative update while maintaining reversibility. The decrease in difference metric is used as the criterion, and the violation of angle, curvature and magnetic disturbance range is included as a penalty term in the objective until the residual threshold or the upper limit of the number of iterations is reached, so as to obtain the converged fusion pose and the updated bias parameters and scale parameters.

[0136] An uncertainty parameter is generated based on a combination of physical residuals, sensing residuals, and consistency differences, so that the uncertainty increases monotonically as the residuals and differences increase.

[0137] The system outputs the fused pose estimation results, updated bias parameters, updated scale parameters, and uncertainty parameters, thus completing the fusion of two-layer reversible neural operators.

[0138] This invention introduces a dynamically constrained reversible transformation model in the attitude fusion stage, innovatively combining a reversible coupling stack, a constraint manager, and a dual-domain consistency optimizer. This breaks through the traditional approach of post-processing pruning or static constraints, realizing a mechanism that embeds physical constraints and consistency optimization within the inference process. The reversible coupling stack maintains a bidirectional mapping relationship between input and output, ensuring the system's interpretability and reversibility. The constraint manager imposes intrinsic constraints on angle, curvature, and magnetic disturbance range during inference, guaranteeing consistency between intermediate solutions and physical conditions. The dual-domain consistency optimizer uses the difference between the physical and sensing domains as the convergence criterion, introducing constraint violation penalties to achieve stable convergence of the attitude fusion results. This reduces the risk of generating physically unrealizable solutions during fusion, improves the stability and reliability of the attitude output, and quantifies the credibility of the results through uncertainty parameters, enabling the level instrument attitude calibration to possess real-time performance, physical rationality, and long-term stability.

[0139] In this embodiment, the output of the final real-time calibration attitude result and the corresponding uncertainty information, modal contribution rate, and anomaly label includes:

[0140] Based on temperature and strain data, timestamp alignment is performed to form temperature and strain sequences within a fixed-length time window, and the mean, range, and rate of change of the current interval are calculated.

[0141] Based on historical operation records and updated bias and scale parameters, the system is divided into segments according to the range of temperature and strain values. The parameter statistics within each segment are summarized into baseline values ​​and incremental coefficients to form a segmented compensation coefficient table, and monotonicity and upper and lower boundary thresholds are set.

[0142] The current temperature and strain are searched in the segmented compensation coefficient table. The instantaneous compensation coefficient is obtained by linear interpolation of adjacent segments. Boundary coefficients are used for out-of-bounds cases, and median filtering and amplitude limiting smoothing are used for abrupt changes to obtain a stable set of compensation coefficients.

[0143] The compensation coefficient set is applied to the updated bias parameters and scale parameters respectively to obtain the refined bias parameters and scale parameters. The fused pose is then recalculated to obtain the refined fused pose.

[0144] The uncertainty parameters are updated based on physical residuals, sensing residuals, and differences in consistency between the two domains. The modal contribution rates of the bubble physical domain and the sensor domain are calculated. Anomaly labels are generated when temperature or strain exceeds limits, residuals suddenly increase, or hard constraints are triggered. Specifically, the calculation of the modal contribution rates of the bubble physical domain and the sensor domain is as follows:

[0145] The physical domain residual energy value and the sensing domain residual energy value are obtained by squaring each element of the physical residual vector and summing them separately, and then summing the two to obtain the total residual energy.

[0146] Dividing the physical domain residual energy value and the sensor domain residual energy value by the total residual energy respectively yields the physical domain modal contribution rate and the sensor domain modal contribution rate. The values ​​range from zero to one and can be converted into percentages to reflect the relative influence of the two types of modes in the final attitude estimation.

[0147] This invention introduces a piecewise compensation coefficient table and a modal contribution rate calculation mechanism in the slow variable refinement compensation stage, overcoming the limitations of traditional compensation methods that rely solely on single temperature or strain corrections. By performing time windowing processing on temperature and strain data, a piecewise compensation coefficient table with monotonicity and boundary constraints is established, enabling the compensation process to adapt to nonlinear characteristic changes in different operating conditions. Combining linear interpolation of adjacent segments with amplitude-limiting smoothing strategies under abrupt changes ensures the stability and continuity of the compensation coefficients. By calculating the energy ratio of physical residuals and sensing residuals, the modal contribution rates of the bubble physical domain and sensor domain in attitude estimation are quantified, achieving transparency and interpretability of the fusion solution. When temperature or strain exceeds limits, residuals suddenly increase, or hard constraints are triggered, the system can automatically generate anomaly tags, improving operational safety and intelligent monitoring capabilities. Dynamic adaptation of compensation parameters and quantification of result reliability are achieved, enabling attitude calibration to possess long-term stability, environmental adaptability, and anomaly traceability. Example

[0148] To verify the feasibility of this invention in practice, it was applied to the installation project of a large optical inspection platform. The construction unit needed to install a high-precision measuring device on a 50-meter-long guide rail, with a levelness requirement within 0.1 degrees. Traditional digital levels rely primarily on inertial measurement units (IMUs) to provide attitude angle information and process the data using Kalman filtering. While this maintains accuracy for short-term measurements, during continuous installation operations exceeding six hours, the attitude calculation gradually deviates from the actual value due to zero-bias drift and noise accumulation in the inertial devices, eventually exceeding 0.5 degrees. Furthermore, the presence of welding machines and large motors at the construction site caused significant interference to the magnetometer, resulting in noticeable fluctuations in attitude calculation and compromising installation accuracy.

[0149] To address the aforementioned issues, the real-time level instrument attitude calibration method based on multimodal data fusion provided by this invention was applied in this field. The system first collects multimodal raw data, including acceleration and angular velocity, magnetic field data, bubble images, temperature and strain information, and preprocesses it to form a time-aligned dataset. Then, in the physical domain neural operator layer, the surface curvature and normal distribution are reconstructed from the bubble images, and combined with surface tension and contact angle boundary conditions to obtain the physical domain attitude estimate. In the sensor domain neural operator layer, noise and drift modeling is performed on the acceleration, angular velocity, and magnetic field data, outputting the sensor domain attitude estimate, along with bias and scale parameters. The bias memory continuously updates historical residual patterns during continuous operation, forming a long-term drift compensation amount which is superimposed on the sensor domain attitude estimation result, avoiding the accumulation of drift errors. A dynamically constrained reversible transformation model inputs the physical domain and the compensated sensor domain attitude estimation results, completing the coupling mapping under constraints of angle range, bubble curvature range, and magnetic disturbance angle range, and generating the fused attitude through dual-domain consistency optimization. Finally, the system uses temperature and strain data to perform slow variable refinement compensation for bias and scale parameters, and outputs the final real-time calibration attitude, uncertainty and anomaly markers.

[0150] During the eight-hour continuous on-site test, we used a laser interferometer as a reference instrument every hour to record the attitude results of the traditional level and the method of this invention.

[0151] Table 1. Attitude measurement errors of traditional level instruments and the method of this invention in field comparative tests.

[0152]

[0153] As can be seen from the data in Table 1 above, under the same environmental and time conditions, there is a significant difference in the attitude measurement results between the traditional level and the method of this invention. With the extension of the test time, the error of the traditional level gradually accumulates, increasing from 0.03° after one hour to 0.62° after eight hours, exhibiting typical zero-bias drift and noise accumulation characteristics. This error will seriously affect the reliability of the measurement results during long-term continuous use, especially in high-precision installation scenarios where horizontal accuracy requirements are within ±0.1°, where the traditional level clearly cannot meet the requirements.

[0154] In contrast, the method of this invention maintained high stability and accuracy throughout the entire eight-hour test. The error was 0.02° in one hour and only 0.10° in eight hours, with extremely small fluctuations and a much lower error growth trend than traditional levels. Even under environmental changes where the temperature gradually increased from 24°C to 29°C, the method of this invention was still able to maintain stable output results through deviation memory and slow variable refinement compensation. Therefore, this invention not only effectively suppresses attitude measurement drift but also enhances robustness under environmental temperature fluctuations and long-term operating conditions.

[0155] This method, through the dual-layer fusion of neural operators in the physical and sensing domains, long-term drift compensation of the self-consistent bias memory unit, and the physical constraint mechanism of the dynamic constraint reversible transformation model, enables the level to output high-precision, interpretable, and physically reasonable attitude results even under long-term working conditions, which is significantly better than the performance of traditional levels.

[0156] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A real-time attitude calibration method for a level instrument based on multimodal data fusion, characterized in that, include: Collect raw multimodal data, preprocess the raw multimodal data, and obtain preprocessed multimodal data; In the physical domain neural operator layer, liquid surface morphology reconstruction and operator inversion are performed based on multimodal data. The normal distribution and curvature distribution of the bubble interface are extracted, liquid surface observation features are generated, and physical domain attitude estimation results and physical residuals are output. In the neural operator layer of the sensing domain, operator inference and attitude calculation are performed based on multimodal data to deal with noise and drift, and the sensing domain attitude estimation results, bias parameters, scale parameters and sensing residuals are output. A bias memory bank is established, and the bias memory bank is updated using physical residuals, sensing residuals and historical fusion results to generate long-term drift compensation. The long-term drift compensation is then superimposed on the sensor domain attitude estimation result to obtain the compensated sensor domain attitude estimation result. The physical domain attitude estimation results, the compensated sensor domain attitude estimation results, and the bias and scale parameters are input into the dynamic constraint reversible transformation model. Reversible coupling transformation is performed under the constraints of angle interval, bubble curvature range, and magnetic disturbance angle range to complete the fusion of two-layer reversible neural operators. The fused attitude estimation results, uncertainty parameters, and updated bias and scale parameters are output. The fused attitude estimation results are then subjected to dual-domain consistency optimization. Based on temperature and strain data from multimodal data, the updated bias and scale parameters are refined and compensated for slowly, and the final real-time calibration attitude results, along with the corresponding uncertainty information, modal contribution rate, and anomaly labels, are output. The multimodal raw data includes acceleration and angular velocity data collected by the inertial measurement unit, magnetic field data collected by the magnetometer, bubble image data acquired by the camera and ring light source, and temperature and strain data collected by the temperature sensor and strain sensor.

2. The real-time attitude calibration method for a level instrument based on multimodal data fusion according to claim 1, characterized in that, The preprocessing of the multimodal raw data includes filtering, timestamp synchronization, and data normalization.

3. The real-time attitude calibration method for a level instrument based on multimodal data fusion according to claim 1, characterized in that, The output physical domain attitude estimation results and physical residuals include: Under phase-coded ring illumination, based on bubble image data, radiometric response calibration, geometric distortion correction, glare suppression and temporal denoising are performed to obtain a preprocessed bubble image sequence. Based on the bubble image sequence and cavity geometry prior, liquid surface morphology reconstruction is performed. The technique of combining photometric stereo reconstruction and contact line segmentation is used to extract the normal distribution and curvature distribution of the bubble interface, and generate a liquid surface observation feature set including normal map, curvature map and boundary mask. A physical domain neural operator layer is constructed, which consists of a boundary condition encoder, a spectral domain kernel mapping backbone, a cross-scale geometric consistency aggregation unit, a physical constraint gating unit, and an inversion solver head, wherein: The boundary condition encoder encodes cavity geometry, contact angle, surface tension, filling height, and temperature parameters; The spectral domain kernel mapping backbone uses a combination of Fourier neural operators and low-rank kernel approximation to realize the function mapping from liquid surface observation features to gravity direction related quantities. The cross-scale geometric consistency aggregation unit uses feature pyramids and cross-layer connections to align features across multiple scales. The physical constraint gating unit adjusts the mapping process based on the relationship between liquid pressure and curvature, as well as the contact angle boundary conditions; The inversion solver outputs the physical domain attitude estimation results and physical residuals. The operator inversion process is performed, and the liquid surface observation feature set and boundary condition encoding are input into the physical domain neural operator layer. First, the initial physical domain attitude estimation result and residual are generated. Then, the estimation result is updated through a finite number of iterations under the adjustment of the physical constraint gating unit until the residual meets the threshold or reaches the upper limit of the number of iterations. Finally, the physical domain attitude estimation result and physical residual vector are output. The physical domain attitude estimation results and physical residual vectors are subjected to quality measurement and region screening. Modal confidence coefficients are generated based on the consistency between residual distribution and reprojection, and abnormal regions are marked.

4. The real-time attitude calibration method for a level instrument based on multimodal data fusion according to claim 1, characterized in that, The output sensing domain attitude estimation results, bias parameters, scale parameters, and sensing residuals include: Based on multimodal data, acceleration data, angular velocity data, magnetic field data, temperature data, and strain data are organized within a fixed-length time window to form a time-aligned observation sequence; A sensor domain neural operator layer is constructed, which consists of three parts: a time-frequency kernel mapping backbone, a cross-modal consistency coupling unit, and a calibration and compensation solver head. The time-frequency kernel mapping backbone adopts a joint structure of time-domain convolution stacking and frequency-domain spectrum to extract zero-bias drift, scale drift and random walk patterns from the observation sequence; The cross-modal consistency coupling unit uses gravity-acceleration consistency, angular velocity-attitude change consistency and geomagnetic-heading consistency as gating signals to dynamically allocate the weights of each channel; The calibration and compensation solver outputs the sensor domain attitude estimation results, bias parameters and scale parameters, and generates sensor residuals. Operator inference for noise and drift is performed, and the observation sequence is processed in the time-frequency kernel mapping backbone to identify and characterize noise and drift modes. In the cross-modal uniform coupling unit, the disturbed channel is suppressed and the stable channel is enhanced based on the uniformity deviation. In the calibration and compensation solver head, bias parameters and scale parameters are generated based on temperature and strain data to form a compensated observation sequence. The attitude calculation is performed, and the observation sequence is decoded into the attitude estimation result of the sensor domain in the sensor domain neural operator layer. The attitude estimation result of the sensor domain includes pitch angle, roll angle and heading angle, and the sensor residual is output, which consists of acceleration consistency deviation, angular velocity consistency deviation and magnetic field consistency deviation.

5. The real-time attitude calibration method for a level instrument based on multimodal data fusion according to claim 1, characterized in that, The obtained compensated sensor domain pose estimation result includes: A data structure for constructing a deviation memory bank is provided. The data structure consists of a working condition fingerprint key, a deviation state value, a compensation calculation parameter value, and a quality measurement value. The working condition fingerprint key is formed by combining temperature data, strain data, physical residual features, and sensor residual features within a fixed time window. Within a preset time window, physical residuals and sensor residuals are collected and combined with temperature data and strain data. The deviation memory is searched according to the working condition fingerprint key. When a match is found, the deviation state value, compensation calculation parameter value and quality measurement value of the corresponding entry are read. When no match is found, a new entry is created and initialized. An update method combining exponential forgetting and time-series accumulation is used to update the deviation status value of the hit item. At the same time, the quality metric value is updated according to the residual amplitude distribution, duration and stability. The compensation calculation parameter value is then stratified and corrected according to the updated quality metric value. Using the updated deviation state value as input, the physical residual and sensor residual are first feature-encoded, and then the compensation calculation parameter values ​​are gated and weighted according to the working condition fingerprint key. The long-term drift compensation amount is calculated according to the adjusted parameters, and the compensation amount in adjacent time windows is checked for consistency through forward and backward time consistency checks. The long-term drift compensation is superimposed on the sensor domain attitude estimation result to obtain the compensated sensor domain attitude estimation result. The compensated result, along with the corresponding working condition fingerprint key, the updated deviation state value, the compensation calculation parameter value, and the quality metric value, is written back to the deviation memory for updating in the next time window.

6. The real-time attitude calibration method for a level instrument based on multimodal data fusion according to claim 1, characterized in that, The output includes the fused attitude estimation result, uncertainty parameter, updated bias parameter, and scale parameter. A two-domain consistency optimization is then performed on the fused attitude estimation result, including: The input to the dynamic constraint reversible transformation model includes the physical domain attitude estimation results, the compensated sensor domain attitude estimation results, and the bias parameters and scale parameters. A dynamic constraint reversible transformation model is constructed, which consists of a reversible coupling stack, a constraint manager, and a dual-domain consistency optimizer. The reversible coupling stack supports forward and backward mapping and contains margin variables to maintain reversibility. The constraint manager is used to apply angular interval constraints, bubble curvature range constraints, and magnetic disturbance angle range constraints within the mapping. The dual-domain consistency optimizer is used to measure the differences between the fused attitude and the physical domain attitude, and between the fused attitude and the compensated sensor domain attitude, and serves as the convergence criterion. Forward inference is performed in the reversible coupling layer stack to generate the initial fused pose, the initial updated bias parameters and scale parameters, and the margin variables. During the inference process, the constraint manager performs endogenous constraints on intermediate results that do not conform to the angle, curvature, and magnetic disturbance range. The dual-domain consistency optimizer is invoked to perform a finite-step iterative update while maintaining reversibility. The decrease in difference metric is used as the criterion, and the violation of angle, curvature and magnetic disturbance range is included as a penalty term in the objective until the residual threshold or the upper limit of the number of iterations is reached, so as to obtain the converged fusion pose and the updated bias parameters and scale parameters. An uncertainty parameter is generated based on a combination of physical residuals, sensing residuals, and consistency differences, so that the uncertainty increases monotonically as the residuals and differences increase. The system outputs the fused pose estimation results, updated bias parameters, updated scale parameters, and uncertainty parameters, thus completing the fusion of two-layer reversible neural operators.

7. The real-time attitude calibration method for a level instrument based on multimodal data fusion according to claim 1, characterized in that, The output includes the final real-time calibration attitude result and the corresponding uncertainty information, modal contribution rate, and anomaly label, including: Based on temperature and strain data, timestamp alignment is performed to form temperature and strain sequences within a fixed-length time window, and the mean, range, and rate of change of the current interval are calculated. Based on historical operation records and updated bias and scale parameters, the system is divided into segments according to the range of temperature and strain values. The parameter statistics within each segment are summarized into baseline values ​​and incremental coefficients to form a segmented compensation coefficient table, and monotonicity and upper and lower boundary thresholds are set. The current temperature and strain are searched in the segmented compensation coefficient table. The instantaneous compensation coefficient is obtained by linear interpolation of adjacent segments. Boundary coefficients are used for out-of-bounds cases, and median filtering and amplitude limiting smoothing are used for abrupt changes to obtain a stable set of compensation coefficients. The compensation coefficient set is applied to the updated bias parameters and scale parameters respectively to obtain the refined bias parameters and scale parameters. The fused pose is then recalculated to obtain the refined fused pose. The uncertainty parameters are updated based on the physical residual, sensing residual, and the difference in consistency between the two domains. The modal contribution rates of the physical domain and the sensor domain of the bubble are calculated. Anomaly labels are generated when the temperature or strain exceeds the limit, the residual suddenly increases, or the hard constraint is triggered.

Images