A method and system for compensating for a mapping auxiliary assembly

Abstract

The present application relates to the technical field of intelligent sensor, solve the technical problem that the precision of the prior art cannot be self-adaptive to the change of the field environment, resulting in the decline under non-stationary conditions, easy to produce cumulative drift, resulting in inaccurate processing results, especially relates to a kind of surveying and mapping auxiliary assembly compensation correction method and system, the steps of the method are: after the acquisition of multi-source data is cleaned and normalized, attitude inference and covariance adaptive estimation are carried out to obtain attitude error;Deformation is fused and feedback correction is obtained to obtain deformation prediction;The final compensation amount is calculated by redundancy weighting and abnormality detection, and finally the sensor parameters are updated by measurement correction and exponential smoothing, the present application realizes the self-adaptive adjustment of filter parameters and the hierarchical fault tolerance of abnormality by the double action of wavelet neural network, and the accurate prediction of thermal deformation and the self-calibration of parameters are realized by thermodynamic model and random forest cooperation, and the fault source is automatically isolated and the fusion safety is improved by inverse variance weighting and mahalanobis distance detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, and in particular to a compensation and correction method and system for a surveying and mapping auxiliary component. Background Technology

[0002] Intelligent sensors are digital measurement units that integrate sensing, processing, and communication functions and can autonomously perform signal conditioning, self-calibration, and data fusion. Compensation and correction of mapping auxiliary components are achieved through intelligent sensors. Generally, the attitude is estimated by filtering with fixed parameters first, then the theoretical deformation is calculated by substituting it into a thermodynamic model with fixed coefficients, and finally the original measurement coordinates are directly corrected. Existing technologies cannot adapt to changes in the field environment, leading to a decrease in accuracy under non-stationary conditions. The lack of anomaly detection mechanisms allows sensor pulse interference to directly disrupt state estimation. Gyroscope bias estimation lacks online correction and is prone to cumulative drift during long-term operation, resulting in inaccurate processing results. For example, the linearization assumption of existing complementary filters fails when tilting at large angles exceeding 30 degrees, resulting in a nonlinear error of about 1.5 degrees in roll angle estimation. In special cases where the gyroscope bias temperature drift rate reaches 0.02 degrees per minute under high-temperature conditions, the fixed process noise of the standard extended Kalman filter cannot track the rapidly changing time-varying bias, and the cumulative attitude error exceeds 0.5 degrees within ten minutes, causing the centering rod measurement point to continuously shift. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a compensation and correction method and system for surveying and mapping auxiliary components. This solves the technical problems of existing technologies being unable to adapt to changes in the field environment, leading to decreased accuracy under non-stationary conditions, lacking online correction for gyroscope zero-bias estimation, and being prone to cumulative drift, resulting in inaccurate processing results.

[0004] To solve the above technical problems, the present invention provides the following technical solution: a compensation and correction method for a surveying and mapping auxiliary component, the method comprising the following steps: collecting the original multi-source sensor data of the target component, and generating a preprocessed sensor dataset by cleaning, normalizing, aligning and encapsulating the original multi-source sensor data; Attitude integral inference estimation is performed based on the preprocessed sensor dataset to generate innovation prediction values ​​and prediction confidence. Covariance adaptive estimation is then performed on the innovation prediction values ​​and prediction confidence to obtain the attitude error estimate. The attitude error estimate is subjected to deformation fusion processing to obtain the preliminary deformation, residual prediction and theoretical deformation. Based on the preliminary deformation, residual prediction and theoretical deformation, the deformation prediction is obtained through feedback correction and encapsulation processing. The confidence weight is obtained by weighting the predicted deformation value with the inverse variance of the redundancy error. The final compensation amount is obtained by weighted detection, transformation and encapsulation based on the confidence weight. The sensor parameters are updated by performing measurement correction exponential smoothing based on the final compensation amount. Among them, the covariance adaptive estimation of the innovation prediction value and the prediction confidence is encapsulated, including: obtaining the adjusted process noise covariance by adaptive covariance processing based on the innovation sequence and the innovation prediction value; Anomaly secondary feedback correction is performed based on prediction confidence, innovation sequence, and innovation prediction value to obtain the corrected attitude estimate; Error estimation is encapsulated by combining the corrected attitude estimate and the preprocessed sensor dataset to obtain the attitude error estimate.

[0005] Preferably, the process of cleaning, normalizing, aligning, and packaging the original multi-source sensor data includes: cleaning the original multi-source sensor data using timestamps to generate synchronized cleaned data; Calibration compensation data is obtained by compensating calibration parameters based on synchronous cleaning data; The calibration compensation data is subjected to multimodal normalization, alignment, and encapsulation to generate a preprocessed sensor dataset.

[0006] Preferably, attitude integral inference estimation based on preprocessed sensor dataset includes: generating initial attitude quaternions by initializing the attitude integral based on the preprocessed sensor dataset; Quaternion recursion is performed based on the initial attitude quaternion and the preprocessed sensor dataset to obtain the attitude estimate and the innovation sequence. Wavelet inference estimation is performed on the innovation sequence to generate innovation prediction values ​​and prediction confidence.

[0007] Preferably, the attitude error estimate is subjected to deformation fusion processing, including: obtaining the theoretical deformation amount by thermodynamic deformation processing based on the attitude error estimate; Random forest residual prediction is performed based on the attitude error estimate to obtain the residual prediction value; The theoretical deformation and the predicted residual value are initially fused to obtain the preliminary deformation.

[0008] Preferably, the process involves feedback correction and encapsulation based on the preliminary deformation, residual prediction, and theoretical deformation, including: adaptive correction based on the residual prediction and theoretical deformation to obtain the corrected linear expansion coefficient, and adjusting the attitude error estimate based on the corrected linear expansion coefficient. Uncertainty assessment is performed based on the initial deformation amount to generate deformation uncertainty; The deformation uncertainty and the initial deformation amount are encapsulated to obtain the deformation prediction value.

[0009] Preferably, the deformation prediction value is weighted by the inverse variance of the redundancy error, which includes: obtaining a multi-source attitude error set by processing the redundancy error based on the deformation prediction value; The compensation calculation is performed based on the multi-source attitude error set to obtain the single-source compensation set; The confidence weights are obtained by applying inverse variance weighting to the single-source compensation set.

[0010] Preferably, the weighted detection transformation and encapsulation based on confidence weight includes: performing weighted detection based on the single-source compensation set and confidence weight to obtain the fused compensation amount and anomaly label; Based on the fusion compensation amount and anomaly markers, back projection consistency verification is performed to obtain the verification compensation amount; The compensation amount after verification is converted and encapsulated to obtain the final compensation amount.

[0011] Preferably, the measurement correction index smoothing process based on the final compensation amount includes: obtaining the corrected measurement result by performing measurement correction processing based on the final compensation amount; Based on the corrected measurement results, residual statistical processing is performed to generate a residual statistical report; The residual statistical report is exponentially smoothed to obtain the updated sensor parameters.

[0012] Preferably, the raw multi-source sensor data includes any one or more of the following: triaxial acceleration data, triaxial angular velocity data, biaxial tilt data, temperature data, and air pressure data.

[0013] This technical solution also provides a compensation and correction system for surveying and mapping auxiliary components, the system comprising: The preprocessing module collects raw multi-source sensor data of the target component, and generates a preprocessed sensor dataset by cleaning, normalizing, aligning and encapsulating the raw multi-source sensor data. The attitude error module performs attitude integral inference estimation based on the preprocessed sensor dataset, generates innovation prediction values ​​and prediction confidence, and encapsulates the innovation prediction values ​​and prediction confidence by covariance adaptive estimation to obtain the attitude error estimate. The deformation prediction module performs deformation fusion processing on the attitude error estimate to obtain the preliminary deformation, residual prediction and theoretical deformation. Based on the preliminary deformation, residual prediction and theoretical deformation, the deformation prediction value is obtained through feedback correction and encapsulation processing. The compensation module calculates the confidence weight by weighting the predicted deformation value using the inverse variance of the redundancy error. Based on the confidence weight, it performs weighted detection, transformation, and encapsulation to obtain the final compensation amount. The parameter update module performs measurement correction exponential smoothing based on the final compensation amount to obtain the updated sensor parameters.

[0014] By employing the above technical solution, the present invention provides a compensation and correction method and system for surveying auxiliary components, which has at least the following beneficial effects: 1. This invention solves the problems of fixed noise covariance of extended Kalman filter and insufficient fault tolerance of sensor anomalies by leveraging the dual role of the innovation prediction value of wavelet neural network. On the one hand, the deviation between the predicted innovation and the measured innovation adaptively corrects the noise covariance of the process, enabling the filter parameters to be adjusted in real time with changes in the environment. On the other hand, the normalized innovation deviation is judged by a dual-threshold classification. In the case of mild anomalies, the predicted innovation directly replaces the contaminated innovation to avoid filter divergence. In the case of severe anomalies, incremental learning is triggered, and the network is repaired online and then corrected again, so that the attitude estimation of the sensor does not jump when subjected to instantaneous impact.

[0015] 2. This invention solves the problem of accurately modeling the thermal deformation of carbon fiber center rods by combining a thermodynamic model with random forest and feedback correction of the linear expansion coefficient. It not only obtains the main trend of deformation by processing the linear expansion coefficient through a thermoelastic beam model, but also supplements nonlinear components such as anisotropy and hygroscopic swelling based on the prediction residuals of random forest, and quantifies the uncertainty with the prediction variance. At the same time, the residual back-calculation of the expansion coefficient correction amount realizes exponential smooth update, so that the physical model continuously approaches the true properties of the material over time.

[0016] 3. This invention solves the problems of inconsistent accuracy of sources and automatic isolation of fault sources in redundant sensor systems by using inverse variance weighting and Mahalanobis distance anomaly detection. On the one hand, the total uncertainty is combined by filtering covariance and deformation uncertainty, and low noise sources automatically obtain high weights. On the other hand, Mahalanobis distance normalizes the deviation of each source with covariance. After being correctly eliminated, the fusion deviation is greatly reduced. The back projection verification further calculates the fusion amount back to the angle space, providing a spatial domain self-consistency secondary verification for the fusion result, and improving the safety of the fusion compensation amount in hidden fault scenarios.

[0017] 4. This invention solves the problems of unquantifiable residual bias and long-term parameter drift by using control point residual statistics and exponential smoothing closed-loop updates. It not only calculates the mean and standard deviation of residuals in real time by comparing the calibrated coordinates with known control points, thus quantifying system bias and random dispersion, but also corrects zero bias based on the gain of the mean residual, and drives the synchronous adjustment of filter noise parameters and expansion coefficients based on the standard deviation. The three parameters are based on the same residual to ensure consistent correction direction. Feedback on updated parameters allows the latest calibration values ​​to be used directly in the next operation, improving the system's long-term accuracy maintenance capability and power-on usability. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a compensation and correction method for a surveying auxiliary component according to the present invention; Figure 2 This is a structural block diagram of a compensation and correction system for a surveying auxiliary component according to the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0020] Example 1: Due to the inability of existing technologies to adapt to changes in the field environment, leading to decreased accuracy under non-stationary conditions, and the lack of online correction for gyroscope bias estimation, which easily results in cumulative drift and inaccurate processing results, please refer to [the relevant documentation / reference]. Figure 1 This embodiment provides a compensation and correction method for a surveying auxiliary component, which enables the system to maintain accuracy under non-stationary conditions, correct online, and reduce the adverse effects of accumulated drift. The method includes the following steps: S1. Collect the original multi-source sensor data of the target component. Based on the original multi-source sensor data, clean, normalize, align and encapsulate to generate a preprocessed sensor dataset. Existing technologies cannot adapt to the local statistical characteristics of signals, resulting in false rejection or missed rejection. The sensor factory calibration parameters are not fully utilized, causing individual zero bias and installation errors to introduce systematic attitude deviations. Direct splicing of multi-channel data does not consider the differences in dimensions and inconsistent sampling rates, resulting in ill-conditioned filtering values ​​and timing misalignment. To solve the above problems, the specific implementation steps are as follows: S11. Based on the original multi-source sensor data, timestamp-cleaned data is generated to produce synchronized cleaned data. In this step, the mean value of each channel sequence in the original multi-source sensor data is obtained using the mean formula within the sliding window, and the standard deviation of the window is obtained using the standard deviation formula. Then, if the absolute value of the difference between the current sample value and the window mean value is greater than 3 times the window standard deviation, the current value is replaced with the result of (previous sample value + next sample value) / 2; otherwise, the original value is retained. Finally, all channels are aligned according to the GPS timestamp to form synchronized cleaned data. The sliding window width is set to 10 to 30 sampling points to adaptively track the local statistical characteristics of the signal. The 3 times standard deviation criterion is based on the normal distribution assumption to separate sensor random noise from transient noise caused by impact or electromagnetic interference. Outliers are effectively separated; linear interpolation replaces the outlier with the average of two normal samples before and after the outlier, eliminating spike noise while preserving the local trend of the signal. The original multi-source sensor data mainly includes: First, triaxial acceleration data, collected by a microelectromechanical system accelerometer, reflecting the linear acceleration of the component in the body coordinate system; Second, triaxial angular velocity data, collected by a microelectromechanical system gyroscope, reflecting the rotational angular rate of the component around each axis; Third, biaxial tilt data, collected by a tilt sensor, directly measuring the static tilt angle of the component relative to the horizontal plane; Fourth, temperature data, collected by a digital temperature sensor, used for thermal deformation compensation calculation; Fifth, barometric pressure data, collected by a barometric altimeter, assisting in elevation direction reference correction.

[0021] S12. Based on the synchronous cleaning data, perform calibration parameter compensation to obtain calibration compensation data; the first step in this step is to read the factory-calibrated zero bias value and scale factor value from the sensor storage chip; the second step is to calculate the calibration compensation value as follows: (original sampled value in synchronous cleaning data - zero bias value) / scale factor; the third step is to calculate the tilt sensor angle compensation value as follows: original angle value - angle offset value. All channel calibration values ​​constitute the calibration compensation data. Due to differences in manufacturing processes, accelerometers and gyroscopes in microelectromechanical systems (MEMS) have fixed zero bias and scaling factor errors. The zero bias is typically on the order of ±0.05 to ±0.2 m / s² or ±0.1 to ±0.5 degrees / s, and the scaling factor is between 0.98 and 1.02. The factory calibration parameters are written into the sensor's internal storage chip and are automatically read and compensated upon power-up, eliminating the systematic deviation of individual sensor differences on subsequent attitude calculations. The angular offset value of the tilt sensor is caused by the mechanical alignment error between the mounting reference surface and the sensor's sensitive axis, and is typically ±0.1 to ±0.5 degrees, obtained through static level calibration.

[0022] S13. Perform multimodal normalization, alignment, and encapsulation on the calibration compensation data to generate a preprocessed sensor dataset. The first step in this process is to obtain the channel mean for each channel in the calibration compensation data (acceleration, angular velocity, temperature, tilt angle, and air pressure) using the mean formula and the channel standard deviation using the standard deviation formula. The standardized value is calculated as (original value - channel mean) / channel standard deviation. The second step involves dividing the standardized channels into data frames according to a unified time window. The feature vector for each frame is calculated as follows: standardized X-axis acceleration, standardized Y-axis acceleration, standardized Z-axis acceleration, standardized X-axis angular velocity, standardized Y-axis angular velocity, standardized Z-axis angular velocity, standardized X-axis tilt angle, standardized Y-axis tilt angle, standardized temperature, and standardized air pressure. The third step is to use the collection of all frames to form the preprocessed sensor dataset. The order of normalization followed by encapsulation ensures that each channel is independently standardized before being stitched together, avoiding the aliasing of values ​​with different dimensions that would cause the global statistics to lose their physical meaning. Acceleration is on the order of approximately 10 meters per second squared, angular velocity is on the order of approximately 3 radians per second, air pressure is on the order of approximately 1000 hPa, tilt angle is on the order of approximately 0.5 degrees, and temperature is on the order of -10 to 50 degrees Celsius. Direct stitching would lead to an excessively large condition number in the observation covariance matrix of the extended Kalman filter in S2, causing numerical ill-conditioning problems. The large differences in the input scale of the wavelet neural network would cause uneven gradient updates. After Z-score standardization, each channel is mapped to a distribution with a mean of 0 and a standard deviation of 1. The uniform time window length is set to 0.1 to 0.5 seconds, consistent with the filter recursion period. The step size is half the window length to achieve inter-frame overlap. The standardization parameters are pre-calculated and stored during the first static acquisition for online operation.

[0023] This invention achieves anomaly removal, individual error calibration, dimensional unification, and temporal alignment of multi-source sensor data through a progressive design that includes sliding window Raida criterion anomaly removal, sensor factory calibration zero-bias scaling factor compensation, multi-channel independent Z-score standardization, and unified time window overlapping alignment encapsulation. This enables extended Kalman filtering and wavelet neural networks to operate on numerically stable and dimensionally consistent inputs, avoiding ill-conditioned matrices and gradient imbalances. It also ensures the timing accuracy of thermodynamic calculations by receiving strictly synchronized temperature and attitude data from the thermal deformation model, thus guaranteeing the accuracy and robustness of the entire mapping compensation and correction system from the data preprocessing source.

[0024] S2. Based on the preprocessed sensor dataset, attitude integral inference estimation is performed to generate innovation prediction values ​​and prediction confidence. The innovation prediction values ​​and prediction confidence are encapsulated by covariance adaptive estimation to obtain attitude error estimates. Existing technologies cannot adapt to changes in the field environment, which leads to a decrease in accuracy under non-stationary conditions. The lack of an innovation anomaly detection mechanism allows sensor pulse interference to directly destroy state estimation. The gyroscope zero bias estimation lacks online correction and is prone to cumulative drift during long-term operation, resulting in inaccurate processing results. To solve the above problems, the specific implementation steps are as follows: S21. Initialize the attitude quaternion by attitude integration based on the preprocessed sensor dataset; the first step in this step is to extract triaxial acceleration data from the initial still frame of the preprocessed sensor dataset; the second step is to set the initial pitch angle as the arctangent function (Y-axis acceleration / Z-axis acceleration), the initial roll angle as the arctangent function (-X-axis acceleration / (Y-axis acceleration^2+Z-axis acceleration^2)^2), and the initial yaw angle as 0; the third step is to set the initial attitude quaternion as the Euler angle to quaternion conversion function (initial pitch angle, initial roll angle, initial yaw angle). The initial attitude quaternion is initialized within 2 to 3 seconds after the system is powered on and remains stationary. The stationary condition is determined by whether the three-axis module of the accelerometer is close to 1 times the gravitational acceleration, with a threshold of 0.9 to 1.1 times the gravitational acceleration. The arctangent calculation of pitch and roll angles uses the components of the gravity vector in each axis of the body coordinate system to infer the attitude relative to the ground. The yaw angle is initially set to 0, assuming that the measurement start direction is north of the reference coordinate system. The conversion of Euler angles to quaternions adopts the standard aviation sequence, such as the order of yaw, pitch, and roll, mapping the three Euler angles to unit quaternions to avoid gimbal lock problems.

[0025] S22. Based on the initial attitude quaternion and the preprocessed sensor dataset, perform quaternion recursion to obtain the attitude estimate and innovation sequence. In this step, the first step is that the state vector consists of seven dimensions: four components of the quaternion and the zero bias of the three-axis gyroscope. The second step is time update: prediction quaternion = previous time step quaternion + 0.5 × quaternion right multiplication matrix × (current angular velocity - previous time step gyroscope zero bias estimate) × time interval, prediction covariance = state transition matrix × previous time step covariance × state transition matrix transpose + process noise covariance. The third step is observation update: innovation = actual observation vector - observation function (predicted state), Kalman gain = prediction covariance × observation matrix transpose × (observation matrix × prediction covariance × observation matrix transpose + inverse matrix of observation noise covariance), state update = predicted state + Kalman gain × innovation. The attitude estimate is taken from the updated quaternion part, and the innovation sequence records the innovation vector of each recursion. Quaternion right-multiplication matrix maps triaxial angular velocity to quaternion change rate, with time intervals ranging from 0.01 to 0.02 seconds to match the 200 to 100 Hz sampling frequency of the inertial sensor; the initial diagonal elements of the process noise covariance matrix are on the order of 10^-4 to 10^-6 to characterize the uncertainty of gyroscope zero-bias drift and angular velocity random walk; the observation vector consists of normalized accelerometer and electronic compass data, and the observation function converts the predicted quaternion into theoretical gravity direction and geomagnetic direction, which are then compared with the measured values ​​to generate innovation; the Kalman gain automatically balances the confidence of model prediction and observation correction based on the relative proportion of predicted covariance and observation noise covariance. The innovation sequence provides input to the wavelet neural network in S23 for learning the nonlinear temporal pattern of the innovation, and also provides basic data for the adaptive adjustment of noise covariance in S24 and the detection of innovation anomalies in S25.

[0026] S23. Perform wavelet inference estimation on the innovation sequence to generate innovation prediction values ​​and prediction confidence. In this step: First, take the innovation sequence from several consecutive past moments as the input vector of the wavelet neural network; Second, the hidden layer activation uses wavelet basis functions, in the form of a cosine function (1.75 × input value) × e^-input value^2 / 2; Third, the innovation prediction value = sum of the outputs of each hidden layer node × their output weights + output bias; Fourth, the prediction confidence = the prediction variance of the additional output of the output layer, obtained by maximizing the likelihood function of the training data. The predicted innovation value plays a dual role: it serves as a reference signal for adjusting the noise covariance of the extended Kalman filter in S24, and as a backup value for the innovation in S25 under abnormal conditions. The wavelet basis function adopts a cosine-modulated Gaussian function, whose scaling and translation characteristics enable it to simultaneously capture local mutations and global trends in the innovation sequence, outperforming the traditional sigmoid activation function in handling non-stationary signals. The input window length is set to 5 to 15 time steps to cover the short-term memory requirements of dynamic attitude changes. The additional output prediction variance of the output layer is trained by maximum likelihood, so that the network not only provides point predictions but also quantifies the uncertainty of the prediction. The larger the prediction variance, the lower the predictability of the innovation at that moment, which may correspond to environmental mutations or sensor anomalies. The wavelet neural network uses the innovation sequence for online incremental training during normal system operation, with a training learning rate of 0.001 to 0.01 to smoothly update the weights. The first role of the predicted innovation value is to provide deviation information from the measured innovation in S24 to drive the process noise covariance correction. The second role is to provide the optimal estimated replacement value based on historical patterns in S25 when the measured innovation is abnormal.

[0027] S24. Based on the innovation sequence and the innovation prediction value, the adjusted process noise covariance is obtained through covariance adaptive processing. In this step, the first step is: actual innovation covariance = sum of the outer products of the innovation vectors at the most recent time points / window length; the second step is: theoretical innovation covariance = observation matrix × prediction covariance × transpose of observation matrix + observation noise covariance; the third step is: if the norm of the difference between the actual covariance and the theoretical covariance exceeds the threshold, then the preliminary adjusted process noise = Kalman gain × actual covariance × Kalman gain transpose; the fourth step is: final process noise = preliminary adjusted process noise + correction coefficient × (measured innovation - predicted innovation value) × (measured innovation - predicted innovation value) transpose. The statistical window length of the actual covariance of the innovation is set to 10 to 30 recursive periods to balance real-time performance and statistical stability. The covariance deviation threshold is set to 0.3 to 0.5 times the theoretical covariance trace as the sensitivity threshold for triggering adjustment. The Kalman gain matrix is ​​generated by the observation update step of S22 and transmitted in real time. The initial adjustment feeds back the innovation statistics to the process noise covariance. The correction coefficient is set to 0.1 to 0.3 so that the influence of the wavelet neural network prediction deviation on the process noise is a smooth and gradual adjustment rather than an abrupt change. The innovation prediction value of the wavelet neural network plays the first role. When the deviation between the predicted innovation and the measured innovation increases, it indicates that the temporal pattern of the innovation has changed and the process noise needs to be increased accordingly to increase the confidence of the filter in the observation. Conversely, the process noise is reduced to make the filter more dependent on the state model.

[0028] S25. Based on the prediction confidence, innovation sequence, and predicted innovation value, perform anomalous secondary feedback correction to obtain the corrected attitude estimate. In this step, the first step is: Normalized innovation deviation XX_pc = (Absolute value of measured innovation SC_jd - Absolute value of predicted innovation YC_jd) / Prediction standard deviation BC_yc; the second step is: If the normalized innovation deviation XX_pc ≤ the first threshold YZ_dy, then the corrected attitude estimate is directly taken from the original extended Kalman filter state update result; the third step is: If the normalized innovation deviation XX_pc is greater than the first threshold YZ_dy and ≤ the second threshold YZ_de, then it is determined to be a minor issue. If the state is severely abnormal, the state correction value XZ_zt = predicted state YC_zt + Kalman gain KM_zy × innovation prediction value YC_xx. The corrected attitude estimate is the corrected quaternion XZ_sy. Step 4: If the normalized innovation deviation XX_pc is greater than the second threshold YZ_de, it is judged as severely abnormal. The current innovation sequence XL_xx is input into the wavelet neural network for single-step incremental learning to update the network weights. The updated network is used to obtain the innovation prediction value YC_xx again, and the state is corrected again. At the same time, the sensor is marked as suspicious. The corrected attitude estimate is the quaternion XZ_sye after the second correction. The first threshold is set at 2 to 3 times the prediction standard deviation, and the second threshold is set at 5 to 6 times the prediction standard deviation, corresponding to the critical values ​​of approximately 95% and approximately 99.9% confidence intervals in statistics, respectively. Normalized innovation bias unifies the innovation components of different dimensions into dimensionless bias, so that the anomaly judgment is not affected by the dimensions of each channel. In the case of mild anomaly, the predicted innovation value replaces the contaminated innovation to directly correct the state. The predicted innovation value is generated by the wavelet neural network in S23 based on the historical normal pattern, which can effectively filter out the instantaneous pulse interference of the sensor without interrupting the filtering recursion. In the case of severe anomaly, the wavelet neural network is triggered to learn incrementally in one step, with a learning rate of 0.005 to 0.02 to quickly adapt to the current mutation mode. After the update, the network re-predicts and corrects the state again, while marking the sensor as suspicious for subsequent investigation. This secondary feedback mechanism enables the wavelet neural network to be repaired online in the case of anomaly rather than degraded. Experimental data description: A certain type of carbon fiber centering rod was used as the test object. An inertial measurement unit (IMU) equipped with a microelectromechanical system (MEMS) was used to collect attitude data in a field environment. Comparative experiments were conducted 30 times each using standard extended Kalman filtering, novelty replacement only without incremental learning, and the complete method of this invention. In 42 samples where the sensor experienced a slight anomaly due to a momentary impact, the standard extended Kalman filtering, directly using the abnormal novelty, caused an average jump of 0.8 degrees in the pitch angle estimate, and the average plane offset of the measurement point after projection onto the 2-meter centering rod reached 28 millimeters. The novelty replacement only method reduced the attitude jump to 0.3 degrees and the offset to 10 millimeters. The complete method of this invention further reduces the offset to within 6 mm and ensures uninterrupted filtering recursion. In 11 samples with severe anomalies, the standard extended Kalman filter experienced filter divergence and required manual restart. The method of this invention triggers incremental learning of the wavelet neural network to re-predict and correct the state, with a filter divergence rate of less than 1%. After the second correction, the mean attitude estimation error recovered to 0.08 degrees, consistent with the normal level before the anomaly occurred. The combined false alarm rate and false alarm rate were optimal when the first threshold was 2 times the standard deviation and the second threshold was 5 times the standard deviation. When the learning rate was 0.01, the incremental learning convergence speed and stability were balanced. This verifies that the dual-threshold hierarchical judgment and the secondary feedback closed loop effectively realize the hierarchical fault tolerance of sensor anomalies and the online self-repair of the wavelet neural network.

[0029] S26. Encapsulate the corrected attitude estimate and the preprocessed sensor dataset for error estimation to obtain the attitude error estimate. In this step: First step: Pitch angle = arctangent function (2 × (first quaternion component × second quaternion component + third quaternion component × fourth quaternion component) / (1 - 2 × (square of second quaternion component + square of third quaternion component)); Second step: Roll angle = arcsine function (2 × (first quaternion component × third quaternion component - fourth quaternion component × second quaternion component)); Third step: Attitude error pitch angle = pitch angle - 0, attitude error roll angle = roll angle - 0, ideal vertical state corresponds to both pitch angle and roll angle being 0; Fourth step: Extract the temperature value of the current time frame from the preprocessed sensor dataset, attitude error estimate = {attitude error pitch angle, attitude error roll angle, temperature value, timestamp}. The quaternion-to-Euler angle conversion uses standard aeronautical sequence to map four-dimensional quaternions into intuitive physical angles. The pitch angle reflects the forward and backward tilt of the component, and the roll angle reflects the left and right tilt of the component. Under ideal vertical conditions, both angles are 0. The ratio of the numerator and denominator in the arctangent function is automatically processed by quaternion normalization constraints, and the output angle range is -90 to 90 degrees. The arcsine function output range is also -90 to 90 degrees. The temperature value is extracted from the same time frame to ensure that the attitude error is strictly synchronized with the current temperature.

[0030] This invention addresses the problem that the process noise covariance matrix of extended Kalman filters is typically set to a fixed empirical value, which fails to adaptively match the drift of noise statistical characteristics caused by changes in the field environment. Filtering accuracy significantly decreases under non-stationary conditions. In practice, the gyroscope vibration noise of the total station centering rod increases under sudden wind changes. The fixed process noise covariance fails to expand in time, leading to over-reliance on the state model and neglect of observation correction. The attitude estimation error deteriorates from 0.05 degrees at rest to 0.3 degrees, and the plane offset of the measurement point exceeds 6 mm at a centering rod height of 2 meters. This invention uses a wavelet neural network to predict the current news value using historical news sequences as input. It corrects the process noise covariance by utilizing the deviation between the predicted and measured news. Simultaneously, it combines statistical matching of the actual and theoretical news covariance for dual adjustment. This not only enables real-time adaptive optimization of filter parameters with environmental changes but also allows the same predicted news to directly correct the state when an anomaly is detected, replacing the contaminated news. This achieves the dual effect of optimizing parameters and compensating for the state.

[0031] By using normalized innovation bias as an anomaly detection index, the deviation between the measured and predicted innovations is divided by the prediction standard deviation to achieve dimensionlessness. A dual-threshold hierarchical judgment is constructed, with the first threshold set at 2 to 3 times the standard deviation and the second threshold at 5 to 6 times the standard deviation. In cases of mild anomalies, the predicted innovation is used instead of the contaminated innovation for state correction to avoid filter divergence. In cases of severe anomalies, the wavelet neural network is further triggered to perform incremental learning, repairing the network weights online before re-predicting and correcting. This not only achieves hierarchical fault tolerance for anomalies but also continuously strengthens the wavelet neural network through anomaly-driven incremental learning. The ability to adapt to abrupt change patterns solves the problem of pulse-like abnormal observations caused by instantaneous impacts or electromagnetic interference during field operations. Traditional extended Kalman filters lack anomaly detection mechanisms, and directly using them for state updates leads to spike errors or even filter divergence in attitude estimation. In practice, an accidental collision between the bottom of the centering rod and a rock generates an instantaneous acceleration impact, and the abnormal information directly enters the observation update, causing the pitch angle estimate to jump by 1.2 degrees in a single recursive cycle. After projection onto the centering rod, the measurement point shifts by as much as 42 millimeters, and the entire station's measurement results need to be discarded and remeasured.

[0032] Specifically, the above-mentioned scheme can address the issues in practice where attitude calculation using standard Euler angles or direction cosine methods is sensitive to linearization errors, and where the lack of an online correction mechanism for gyroscope bias estimation leads to long-term integral drift. In practice, the linearization assumption of traditional complementary filtering fails when the component tilts at angles exceeding 30 degrees, resulting in a nonlinear error of approximately 1.5 degrees in roll angle estimation. In special cases where the gyroscope bias temperature drift rate reaches 0.02 degrees per minute under high-temperature conditions, the fixed process noise of the standard extended Kalman filter cannot track the rapidly changing time-varying bias, leading to accumulated attitude errors within ten minutes. To address the issue of continuous offset of the centering rod measurement point due to a deviation exceeding 0.5 degrees, this invention incorporates the three-axis zero bias of the gyroscope into online estimation using quaternion state vectors. Time updates accurately describe the attitude kinematics using quaternion right-multiplication matrices. The adaptively adjusted process noise covariance introduces appropriate uncertainty to the state corresponding to the gyroscope's zero bias, ensuring continuous convergence of the zero bias estimation. This not only eliminates Euler angle singularities and linearization errors but also converts the corrected quaternions into pitch and roll angle errors and encapsulates them in relation to temperature, providing an accurate attitude error benchmark and synchronous temperature input for the thermodynamic deformation model.

[0033] S3. The attitude error estimate is subjected to deformation fusion processing to obtain the preliminary deformation, residual prediction and theoretical deformation. Based on the preliminary deformation, residual prediction and theoretical deformation, the deformation prediction is obtained through feedback correction and encapsulation processing. Existing technology cannot capture the nonlinear effects of composite materials. The fixed material parameters do not undergo online correction with aging and environment, which leads to the accumulation of prediction deviation. The deformation prediction only outputs a definite value and lacks uncertainty quantification, which makes it impossible for subsequent fusion to distinguish the prediction credibility. As a result, the system cannot self-calibrate with the time of use, and the long-term measurement accuracy is prone to continuous degradation. To solve the above problems, the specific steps are as follows: S31. Based on the attitude error estimate, obtain the theoretical deformation amount through thermodynamic deformation processing; the first step in this step is to extract the current temperature value, temperature gradient value, and preset parameters such as component length, material linear expansion coefficient, reference temperature, and cross-sectional diameter from the attitude error estimate; the second step is: theoretical axial deformation amount = linear expansion coefficient × component length × (current temperature - reference temperature); the third step is: theoretical lateral offset amount = linear expansion coefficient × component length squared × temperature gradient / (2 × cross-sectional diameter); the fourth step is: theoretical deformation amount = {theoretical axial deformation amount, theoretical lateral offset amount}; The thermoelastic beam model assumes that the component is a uniformly materialed rod with a linear expansion coefficient of 0.5 × 10⁻⁶ to 2 × 10⁻⁶ per degree Celsius, typical for carbon fiber composites. The component length is typically 2 to 3 meters, and the reference temperature is the standard calibration temperature of 20 degrees Celsius. The axial deformation reflects the length expansion and contraction of the component under a uniform temperature field, while the lateral offset reflects the bending deformation when there is a temperature difference between the upper and lower ends of the component. The temperature gradient is obtained by dividing the difference between the two temperature sensors at the upper and lower ends of the component by the distance between them. The cross-sectional diameter is typically 30 to 50 millimeters, representing the outer diameter of the component.

[0034] S32. Based on the attitude error estimate, perform random forest residual prediction to obtain the residual prediction value. In this step, the first step is to extract temperature, temperature gradient and timestamp from the attitude error estimate as the input feature vector of the random forest model; the second step is to output the prediction value of the deformation residual independently by each decision tree; the third step is to calculate the residual prediction value as the arithmetic mean of the prediction values ​​of all decision trees, i.e., the sum of the prediction values ​​of each tree / the total number of decision trees. Random forest ensemble inference reduces the risk of overfitting of a single tree by averaging the output of multiple decision trees. The number of decision trees is between 50 and 200. During training, each tree randomly draws a subset of samples with replacement from historical data and splits on the feature subset to ensure diversity among trees. The timestamps in the input features are used to capture long-term effects such as material aging and creep, and the temperature gradient is used to capture nonlinear bending caused by non-uniform temperature fields. The random forest model is trained using component calibration experimental data during the offline stage of the system. The training label is the residual of the measured total deformation minus the theoretical deformation of S31. The residual prediction value captures the complex nonlinear deformation components not modeled by the thermodynamic model and is obtained by adding the theoretical deformation of S33 to obtain the preliminary deformation value.

[0035] S33. Perform preliminary fusion processing on the theoretical deformation and the predicted residual value to obtain the preliminary deformation; in this step, the first step is: preliminary axial deformation = theoretical axial deformation + predicted axial residual value; the second step is: preliminary lateral offset = theoretical lateral offset + predicted lateral residual value; the third step is: preliminary deformation = {preliminary axial deformation, preliminary lateral offset}. The fusion method directly adds the basic deformation of the thermodynamic physical model to the nonlinear residuals learned by the random forest. The physical model provides the main trend of deformation and clear engineering interpretability, while the random forest residuals supplement the model bias caused by complex factors such as anisotropy and viscoelasticity of the resin matrix in carbon fiber composites. When the random forest is sufficiently trained and the residual prediction is accurate, the initial deformation will approach the true deformation value. When the random forest has a large prediction bias due to insufficient training data, the physical model still ensures that the deformation is within a reasonable range and does not produce outrageous results.

[0036] S34. Adaptive correction is performed based on the residual prediction value and theoretical deformation to obtain the corrected linear expansion coefficient, which is then fed back to S31. In this step, the first step is: expansion coefficient correction amount = axial residual prediction value / (component length × (current temperature - reference temperature)); the second step is: corrected linear expansion coefficient = original linear expansion coefficient + learning rate × expansion coefficient correction amount; the third step is: the corrected linear expansion coefficient is fed back to S31 to replace the original linear expansion coefficient, and the theoretical deformation of the next cycle is calculated. Exponential smoothing updates control the step size of each correction with a learning rate between 0.05 and 0.2, allowing the linear expansion coefficient to smoothly and gradually approach the true value, avoiding drastic jumps in the coefficient caused by single residual abnormalities. Correction is paused when the temperature difference from the reference temperature is less than 2 degrees Celsius to avoid amplifying noise due to an excessively small denominator. The corrected linear expansion coefficient is used for the theoretical calculation of both axial deformation and lateral displacement, as they share the same material thermophysical properties. This adaptive correction enables the thermodynamic model to continuously self-calibrate as the component is used over time and the environment changes, gradually reducing the dependence on random forest residual predictions.

[0037] S35. Based on the initial deformation, perform uncertainty assessment and generate deformation uncertainty; the first step in this step is to calculate the sum of squares of the differences between the predicted values ​​of each decision tree and the mean of the random forest residual prediction; the second step is to calculate the deformation uncertainty as the square root of (sum of squares / (number of decision trees - 1)). The deformation uncertainty measures the degree of divergence among decision trees within the random forest in predicting the residuals. The greater the dispersion of predictions among trees, the higher the uncertainty, indicating a lower confidence level in the deformation prediction under the current input features. The number of decision trees is usually between 50 and 200. The more trees there are, the more stable the uncertainty estimation becomes, but the computational cost increases linearly. This uncertainty is directly input into S36 and associated with the initial deformation amount for encapsulation. It is also passed to subsequent steps as the source of uncertainty for the deformation prediction component in the confidence weight calculation. This drives the automatic reduction of the deformation prediction contribution ratio at high uncertainty moments during multi-source fusion, thereby improving the robustness of the final compensation amount.

[0038] S36. Encapsulate the deformation uncertainty and preliminary deformation amount into predicted values ​​to obtain the deformation prediction value. The first step in this step is to extract the preliminary axial deformation and preliminary lateral offset from the preliminary deformation amount, extract the uncertainty value from the deformation uncertainty, and extract the current temperature value from the attitude error estimate. The second step is: Deformation prediction value = {Axial deformation amount is taken from the preliminary axial deformation amount, lateral offset is taken from the preliminary lateral offset, prediction uncertainty is taken from the deformation uncertainty, and temperature is taken from the current temperature of the attitude error estimate}. The deformation prediction values ​​encapsulate the deformation components, prediction uncertainties, and associated temperatures driven by the thermodynamic physical model and random forest data into a standardized data structure. The axial deformation is used to correct the thermal expansion error in the length direction of the component, and the lateral offset is used to correct the lateral displacement of the measurement point caused by the bending of the component. The prediction uncertainty is passed to the fourth step of the parent process and merged with the attitude estimation uncertainty to calculate the confidence weight of each source. The temperature data is used in the fourth step of the parent process to verify the consistency between the deformation prediction and the current operating conditions.

[0039] This invention solves the problem that existing technologies cannot capture the nonlinear thermal deformation behavior of carbon fiber composites caused by fiber orientation, resin matrix viscoelasticity, and manufacturing residual stress. In practice, a carbon fiber centering rod predicted an axial deformation of 0.18 mm under a 30°C temperature difference using a pure physical model, but the measured value reached 0.31 mm due to the combined effect of moisture absorption and swelling of the resin matrix. The model underestimated the deformation by about 42%, resulting in a centimeter-level deviation in the elevation direction of the measurement point. This invention uses a random forest with temperature, temperature gradient, and timestamp as inputs to learn the complex nonlinear relationships in the residuals of the physical model. The data-driven residuals are added to the basic values ​​of the physical model to obtain the initial deformation. This not only improves the deformation prediction accuracy to more than 90% of the measured value compared to the pure physical model, but also back-calculates the residuals into a linear expansion coefficient correction amount to exponentially smooth and update the physical model parameters, thereby ensuring that the physical model continuously self-calibrates with usage time and environmental changes.

[0040] By using exponential smoothing updates with a learning rate of 0.05 to 0.2, the expansion coefficient correction from each random forest residual back-calculation is progressively integrated into the current coefficient. When the temperature difference is less than 2 degrees Celsius, the correction is paused to avoid amplifying noise due to an excessively small denominator. This not only ensures that the linear expansion coefficient continuously tracks the true thermophysical properties of the material, but also gradually improves the accuracy of the theoretical deformation base value calculation with each correction iteration, gradually reducing the dependence on data-driven residual prediction. This can solve the problem that the material thermophysical parameters are based on factory nominal values ​​and are not corrected online as the components age and are exposed to the environment, resulting in the deformation prediction deviation gradually increasing over time. In practice, after two years of field use, the resin matrix of the carbon fiber centering rod developed microcracks, and the equivalent linear expansion coefficient drifted from 1.2 × 10^-6 degrees Celsius at the factory to approximately 1.8 × 10^-6 degrees Celsius. The fixed parameter model under-predicted axial deformation by about 0.1 mm at a length of two meters, and the cumulative offset of the plane measurement points after projection onto the centering rod exceeded 3 mm.

[0041] Specifically, the above scheme can solve the problem in practice that deformation prediction only provides a definite value without the quantification of uncertainty, which leads to the inability to distinguish between high-confidence and low-confidence predictions during subsequent fusion compensation. All prediction results are treated with equal weight, resulting in low accuracy. In practice, under conditions of rapid temperature gradient changes, random forests suffer from sparse training data, resulting in a dispersion of 0.05 mm in the predicted values ​​of each decision tree. However, the lack of uncertainty labels ensures that low-confidence predictions and high-confidence predictions have the same weight during fusion, especially under steady-state conditions. This is particularly relevant in special cases where components experience local temperature jumps due to alternating direct sunlight and shadows. In this scenario, the lag in the temperature sensor response leads to a mismatch between the input features and the actual temperature field, significantly increasing the discrepancy between the predicted values ​​of the random forest trees. Without an uncertainty label, unreliable predictions at that moment will directly participate in the fusion compensation. This invention uses random forest prediction variance estimation to quantify the dispersion of deformation predictions by using the sample standard deviation of the predicted values ​​between decision trees. This serves as a confidence index for deformation predictions and is associated with and encapsulated with the initial deformation amount. This not only gives the deformation predictions a confidence label but also automatically reduces the contribution of deformation predictions at high uncertainty moments during inverse variance weighted fusion, thereby ensuring the robustness of the fusion compensation amount.

[0042] S4. Based on the deformation prediction value, the confidence weight is obtained by weighting with the inverse variance of the redundancy error. The weighted detection, transformation and encapsulation are performed based on the confidence weight to obtain the final compensation amount. The existing technology cannot dynamically allocate weights according to the actual estimated quality of each source, resulting in loss of fusion accuracy. It lacks an automatic identification and isolation mechanism for abnormal sources, which allows sensor failures to directly contaminate the fusion results. The fusion results have no self-consistency verification and cannot detect errors when multiple sources have systematic deviations at the same time. When there are large-area sensor anomalies, there is no independent backup channel to ensure the compensation safety. To solve the above problems, the specific steps are as follows: S41. Based on the deformation prediction value, a multi-source attitude error set is obtained through redundant error processing. In this step, the timestamp in the deformation prediction value is used as an index to extract the attitude error estimate value calculated independently by S2 at the same time from each tilt sensor channel. Then, all the extracted attitude error estimates are combined into a multi-source attitude error set. Two to four tilt sensors are typically redundantly installed on the centering rod. Each sensor independently collects acceleration and angular velocity data and performs extended Kalman filtering and wavelet neural network calculations in S2, outputting its own attitude error pitch and roll angles. The timestamps carried in the deformation prediction values ​​are consistent with the timestamps of the attitude errors of each sensor, ensuring that the collected multi-source data are strictly synchronized in the time dimension. The multi-source redundancy configuration ensures that other sensors can still provide normal attitude error estimates when a single sensor fails or is subject to local interference, providing a multi-source input basis for the single-source compensation calculation in S42 and the anomaly detection and isolation in S44.

[0043] S42. Based on the multi-source attitude error set, perform compensation and conversion processing to obtain the single-source compensation set. The first step in this process is to extract the pitch angle and roll angle of each sensor from the multi-source attitude error set. The second step is to calculate the horizontal X-axis compensation amount as: component height × tangent function (pitch angle of attitude error) and the horizontal Y-axis compensation amount as: component height × tangent function (roll angle of attitude error). The third step is to extract the lateral offset from the deformation prediction value. The total X-axis compensation amount is calculated as: horizontal X-axis compensation amount + X component of the lateral offset; the total Y-axis compensation amount is calculated as: horizontal Y-axis compensation amount + Y component of the lateral offset. The fourth step is to calculate the compensation vector for each source as: {total X-axis compensation amount, total Y-axis compensation amount}. The compensation vectors from all sources constitute the single-source compensation set. Geometric projection compensation uses the tangent function to convert the tilt angle of the centering rod into the horizontal displacement of the measurement point on the ground. The component height is usually taken as 2 to 3 meters as the standard length of the centering rod. When the attitude error angle is less than 2 degrees, the tangent function is approximately equal to the angle itself. When the angle is larger, accurate calculation is required to avoid linear approximation error. The lateral offset in the deformation prediction value comes from the thermal deformation estimate of S36, reflecting the additional lateral displacement caused by the bending deformation of the carbon fiber rod under the temperature gradient at the measurement point. It is vector-added with the attitude error projection in the horizontal plane. The compensation vector of each source in the single-source compensation set independently reflects the complete correction value of the sensor and the corresponding deformation prediction channel for the horizontal deviation of the measurement point.

[0044] S43. Perform inverse variance weighting on the single-source compensation set to obtain the confidence weights; the first step in this step is to extract the attitude angle variance from the extended Kalman filter covariance matrix from each attitude error estimate and extract the deformation uncertainty from the deformation prediction value; the second step is to calculate the total uncertainty as the square root of (attitude angle variance + square of deformation uncertainty); the third step is to calculate the confidence weight of a source as (1 / square of the total uncertainty of that source) / the sum of (1 / square of the total uncertainty of all sources); Inverse variance weighting uses the inverse square of the total uncertainty of each source as the basis for weight allocation. The lower the uncertainty of a source, the greater its weight, ensuring that high-precision sources dominate in the fusion. The attitude angle variance is extracted from the diagonal elements of the covariance matrix of the extended Kalman filter in S22, reflecting the convergence of the filter estimate of the sensor at the current moment. It is usually on the order of 10^-6 to 10^-6 square radians. The deformation uncertainty is passed from S35, reflecting the degree of divergence of the deformation predictions of each decision tree in the random forest. It is usually on the order of 0.01 to 0.1 millimeters. The two are combined into the total uncertainty after unifying the dimensions.

[0045] S44. Perform weighted detection based on the single-source compensation set and confidence weights to obtain the fused compensation amount and anomaly label. In this step: First, the fused compensation vector BC_rh = the sum of the confidence weights of each source Q_zx × the corresponding source compensation vector BC_dy. Second, calculate the Mahalanobis distance MS_jl from each source compensation vector BC_dy to the fused compensation vector BC_rh. The Mahalanobis distance MS_jl = ((the transpose of the source compensation vector BC_dy - the fused compensation vector BC_rh) × the inverse matrix DY_nj of the single-source compensation set covariance matrix × (the square root of the source compensation vector BC_dy - the fused compensation vector BC_rh)). Third, if the Mahalanobis distance MS_jl of a source is greater than the Mahalanobis distance threshold MS_yz, then the anomaly label for that source is set to 1 and removed from the single-source compensation set. The first step is then re-executed to calculate the fused compensation vector BC_rh; otherwise, the anomaly label is set to 0. Fourth, the final fused compensation amount BC_rhz and the anomaly labels BJ_y of each source constitute the output. The weighted average is used to perform a convex combination of the compensation vectors of each source with confidence weights. Sources with high confidence contribute more and sources with low confidence contribute less. The Mahalanobis distance normalizes the bias of each source through the covariance matrix of the single-source compensation set, eliminating misjudgments caused by scale correlation between the compensation quantities of each source. The Mahalanobis distance threshold of 2 to 3 corresponds to the boundary of the ellipsoid with a confidence of about 95% to 99%. The covariance matrix of the single-source compensation set is estimated by the statistical covariance of the historical compensation vectors of each source and is updated frame by frame during online operation. After removing abnormal sources, the weighted fusion is re-weighted to ensure that the final fused compensation quantity is not contaminated by a single faulty sensor, providing a reliable center estimate for the back projection verification in sub-step 4.5. Experimental data description: A certain type of carbon fiber centering rod was used as the test object. Three tilt sensors were redundantly installed on it and independently ran S2 to calculate the attitude error. An instantaneous pulse with an amplitude of 0.3 m / s² was injected into the accelerometer channel of sensor B to simulate the sensor being subjected to an impact anomaly. Three methods were compared and tested 30 times each: simple averaging fusion, inverse variance weighted fusion without anomaly detection, and the inverse variance weighted fusion plus Mahalanobis distance detection of the present invention. Among them, under the abnormal state, the X-axis component of the single-source compensation vector of sensor B deviated from the mean of sensors A and C by 0.42 mm. The simple averaging fusion caused a fusion compensation deviation of 0.14 mm due to the three sources being equally weighted and the abnormal source not being isolated. The residual deviation of the measurement point reached 0.28 mm. Although the inverse variance weighted fusion without anomaly detection reduces the contribution of sensor B by weighting the uncertainty, the anomaly source still participates in the fusion, resulting in a compensation deviation of 0.08 mm and a residual deviation of 0.16 mm. The method of this invention calculates the Mahalanobis distances of the three sources to be 0.8, 4.2, and 1.1, respectively. Sensor B exceeds the threshold of 2.5 and is automatically marked as an anomaly and removed. After re-weighting and fusing only sensors A and C, the compensation deviation is reduced to 0.03 mm, and the residual deviation is controlled within 0.06 mm. This verifies that the combination of weighted fusion and Mahalanobis distance detection achieves the dual effect of high-precision automatic source highlighting and active isolation of fault sources.

[0046] S45. Perform back projection consistency verification based on the fused compensation amount and anomaly markers to obtain the verified compensation amount. In this step: First, estimate the pitch angle of the attitude error = arctangent function (X-axis component of the fused compensation amount / component height); estimate the roll angle of the attitude error = arctangent function (Y-axis component of the fused compensation amount / component height); Second, after removing sources marked as 1 from the multi-source attitude error set, the mean attitude error of the normal sources = (sum of pitch angles of attitude errors of each normal source / number of normal sources, sum of roll angles of attitude errors of each normal source / number of normal sources); Third, deviation = ((estimated pitch angle of attitude error - mean pitch angle of normal source attitude error)^2 + (estimated roll angle of attitude error - mean roll angle of normal source attitude error)^2) square root; Fourth, if the deviation is greater than the back verification threshold, it is marked as verification failure, and re-acquisition or downgrading to use the historical model for compensation is triggered; otherwise, the verified compensation amount is directly taken as the fused compensation amount. The back-projection consistency verification uses the component height to back-calculate the horizontal plane compensation amount back to the attitude angle space, and performs a cross-comparison with the average attitude error of each source that is not marked as abnormal in the angle space. The component height is 2 to 3 meters, and the back-projection verification threshold is 0.05 to 0.1 degrees, corresponding to a measurement point deviation of about 2 to 3 millimeters at a height of 2 meters. This verification can capture hidden faults that Mahalanobis distance detection fails to detect, such as two abnormal sources canceling each other out, causing the fused compensation amount to appear reasonable in numerical terms, but after back-projection to the angle space, it deviates significantly from the average value of the normal source. When the verification fails, the degradation strategy prioritizes the deformation amount predicted by the S31 pure physical model as the compensation benchmark, suspends random forest residual prediction and inverse variance weighted fusion, and ensures that when large-area sensor anomalies or drastic environmental changes cause loss of multi-source consistency, the system can still output a safe and reliable compensation amount instead of erroneous correction.

[0047] S46. Convert and encapsulate the verified compensation amount to obtain the final compensation amount. The first step in this step is to extract the X-axis compensation component and Y-axis compensation component in the horizontal plane from the verified compensation amount, and extract the axial deformation amount as the height compensation component from the deformation prediction value. The second step is to calculate the final compensation amount as follows: The X-axis compensation amount is taken from the X component of the verified compensation amount, the Y-axis compensation amount is taken from the Y component of the verified compensation amount, and the Z-axis compensation amount is taken from the axial deformation amount. The compensation values ​​for the X and Y axes in the horizontal plane are directly taken from the verified compensation values. These verified compensation values ​​are the optimal horizontal compensation values ​​after triple screening through inverse variance weighted fusion, Mahalanobis distance anomaly detection, and back projection consistency verification. The Z-axis compensation value is taken from the axial deformation value, which is provided by the axial deformation component in the deformation prediction value output by S36, reflecting the length expansion and contraction of the component due to thermal expansion. If the original measurement coordinates are in the NE-H coordinate system, then the X-axis corresponds to the northward offset correction, the Y-axis corresponds to the eastward offset correction, and the Z-axis corresponds to the elevation direction correction.

[0048] This invention addresses the problem that when multi-source attitude error estimation results are fused, simple averaging or equal weighting fails to dynamically allocate weights based on the actual estimation quality of each source, leading to the dilution of the contribution of high-precision sources by low-precision sources. In practice, the three tilt sensors installed on the centering pole have different installation positions, with the sensor closer to the bottom of the pole experiencing higher noise due to ground vibration. Simple averaging fusion treats low-quality sources and high-quality sources with equal weights, resulting in the attitude error estimate being skewed by approximately 0.04 degrees by the low-quality sources. This invention uses inverse variance weighting with the inverse square of the total uncertainty of each source as the weight. The total uncertainty is obtained by combining the extended Kalman filter covariance and deformation prediction uncertainty. Low-noise sources automatically receive high weights, while high-noise sources are automatically suppressed. This not only makes the fusion result approach the optimal estimation accuracy but also uses Mahalanobis distance detection to normalize the deviation of each source using covariance, ensuring that anomaly detection is not affected by the differences in the absolute values ​​of each source. This ensures that the fusion weights and anomaly judgment work collaboratively within the same uncertainty framework.

[0049] By normalizing the source biases using the covariance matrix of the single-source compensation set through Mahalanobis distance, and setting the threshold to 2 to 3 to correspond to approximately 95% to 99% confidence ellipsoid boundaries, abnormal sources are automatically marked as 1 and removed before re-weighted fusion. This not only achieves automatic isolation of fault sources, but also enables back-projection verification to back-calculate the fused compensation amount to the attitude angle space and compare it with the mean of normal sources, further capturing hidden anomalies that Mahalanobis distance fails to detect. This ensures double verification of the output compensation amount and solves the problem that traditional fusion methods lack the ability to automatically identify and isolate abnormal sources when a single sensor in a redundant sensor system fails or is subject to local interference, leading to abnormal data directly contaminating the fusion results. For example, if a tilt sensor experiences intermittent data jumps due to poor cable contact, its output attitude error pitch angle jumps by 0.5 degrees. If it directly participates in fusion, it will result in a final compensation amount deviation of 0.17 mm and a measurement point offset of approximately 0.35 mm after centering rod projection.

[0050] Specifically, the above scheme can solve the problem in practice that fusion compensation is directly output after a single calculation, lacking a verification mechanism for the rationality of the fusion result itself. When multiple sources simultaneously exhibit deviations in the same direction, the weighted fusion result may appear numerically reasonable but is actually incorrect. In practice, two tilt sensors experienced approximately 0.2-degree offset errors in the same direction due to synchronous loosening of their fixing bolts; the Mahalanobis distance detection failed to trigger rejection due to the proximity of two abnormal sources; and the fusion compensation deviation reached 0.07 mm. Furthermore, a sudden change in ambient temperature caused two of the three sensors to simultaneously exhibit transient deviations in the same direction but with different amplitudes due to differences in thermal conduction. In special cases, the third normal source is mistakenly identified as an outlier by Mahalanobis distance. If directly adopted, normal information will be completely lost. The reverse projection verification of this invention calculates the fusion compensation amount back to the attitude error angle space through the arctangent function and component height, and compares it with the mean attitude error of the normal source after removing the outlier source. The deviation threshold is set to 0.05 to 0.1 degrees. When the verification fails, the pure physical model is used for compensation to ensure the safety bottom line. This not only provides a secondary verification of the fusion result in the spatial domain, but also enables the physical model of S3 to serve as an independent backup channel when the sensor has a large area of ​​anomalies, thereby ensuring that the system still outputs reliable compensation amount under extreme abnormal conditions.

[0051] S5. Based on the final compensation amount, the measurement correction index is smoothed to obtain the updated sensor parameters. The existing technology lacks real-time residual comparison with the control point, making the residual deviation unknown. The sensor parameters and model coefficients drift during long-term use, but there is no online self-calibration mechanism. The independent adjustment of parameters in each link lacks unified residual feedback coordination, which may lead to conflicting correction directions. To solve the above problems, the specific implementation steps are as follows: S51. Based on the final compensation amount, the measurement correction process is used to obtain the corrected measurement result. In this step, the first step is to extract the X-axis compensation amount, Y-axis compensation amount, and Z-axis compensation amount from the final compensation amount; the second step is to calculate the corrected X-coordinate as follows: corrected X-coordinate = original X-coordinate + final compensation amount X component, corrected Y-coordinate = original Y-coordinate + final compensation amount Y component, and corrected Z-coordinate = original Z-coordinate + final compensation amount Z component; the third step is to calculate the corrected measurement result as follows: corrected X-coordinate, corrected Y-coordinate, corrected Z-coordinate. The direct correction applies the final compensation amount component by component to the original measurement coordinates. The compensation amounts for the X and Y axes are derived from the compensation amounts after horizontal plane verification in S4 through redundancy fusion and anomaly detection, while the compensation amount for the Z axis is derived from the axial thermal deformation prediction value in S3. The original measurement coordinates are output in real time by the total station or global navigation satellite system receiver. The corrected measurement results eliminate the system deviations caused by the tilt of the centering rod attitude and thermal deformation.

[0052] S52. Perform residual statistical processing based on the corrected measurement results and generate a residual statistical report. The first step in this process is to extract the corrected 3D coordinates from the corrected measurement results and obtain the true 3D coordinates of the corresponding control points from the known control point database. The second step is to calculate the residual vector as: Residual Vector = Corrected Coordinates - Control Point Coordinates. The third step is to obtain the mean of each component of the residual vector using the mean formula and the standard deviation of each component using the standard deviation formula. The fourth step is to calculate the residual statistical report as: {Residual Mean, Residual Standard Deviation, Number of Control Points Participating in the Statistics}. Known control points are usually fixed points pre-deployed within the survey area and have undergone high-level adjustment calculations. Their coordinate accuracy is 1 to 2 orders of magnitude higher than that of real-time measurement points, and they can be used as true value references. The mean residual reflects the systematic bias of the compensated measurement results. If the mean residual deviates significantly from zero, it indicates that there is an unmodeled constant bias in the compensation chain from step one to step four. The standard deviation of the residual reflects the degree of random dispersion of the compensated measurement results. The smaller the standard deviation, the higher the stability of point-by-point compensation. The number of control points is usually 3 to 5 to meet the statistical degrees of freedom requirements.

[0053] S53. Perform exponential smoothing correction on the residual statistical report to obtain the updated sensor parameters. In this step, the residual mean is first extracted from the residual statistical report. Step 1: Updated sensor zero bias = original sensor zero bias + updated gain × residual mean. Step 2: Adjust the observation noise covariance of the extended Kalman filter in S2 according to the residual standard deviation, and correct the linear expansion coefficient in S3 according to the residual mean. All updated parameters constitute the updated sensor parameters and are fed back to S1. Exponential smoothing online correction uses the update gain to control the magnitude of each correction. The update gain is set to 0.1 to 0.3 to make the zero-bias parameter smoothly and asymptotically approach the true value, avoiding drastic parameter jumps caused by single residual abnormalities. The residual mean reflects the system bias remaining after compensation. When the residual mean continuously deviates from zero, it indicates that the sensor zero bias has drifted and needs to be compensated online. The residual standard deviation reflects the random dispersion of the measurement results after compensation. When the standard deviation continuously increases, it indicates that the noise statistics of the extended Kalman filter no longer match the current environment, and the diagonal elements of the observation noise covariance need to be increased accordingly. The correction of the linear expansion coefficient forms a dual correction with S34. S34 is driven by the random forest residual for micro-correction, and S53 is driven by the final measurement residual for macro-calibration. The updated sensor parameters are fed back to S12 so that the sensor calibration compensation for the next operation adopts the latest zero bias value. At the same time, they are fed back to S22 and S31 so that the initialization parameters of the extended Kalman filter and thermodynamic deformation model can be continuously self-optimized as the system runs, forming a complete process from the original sensor data to the final measurement results and back to the sensor calibration.

[0054] This invention achieves real-time quantitative evaluation of compensation effects, directional tracing of system residual deviations, and multi-parameter coordinated closed-loop updates through a progressive design that directly corrects the applied compensation amount to the measurement coordinates, compares the residual mean and standard deviation with the control points in real time, and uses exponential smoothing to correct sensor zero bias and model parameters online and feeds them back to S1 to S3. This enables the final compensation amount to be directly converted into measurement results, and allows the updated parameters to become the initial calibration values ​​for the next operation. Furthermore, it ensures that the filter noise parameters and material physical coefficients continuously evolve with the residual feedback, guaranteeing the quantifiable accuracy, self-calibrating parameters, and long-term stability of the entire surveying and mapping compensation and correction system from the compensation execution and feedback calibration stages.

[0055] Example 2: Due to the inability of existing technologies to adapt to changes in the field environment, leading to decreased accuracy under non-stationary conditions, and the lack of online correction for gyroscope bias estimation, which easily results in cumulative drift and inaccurate processing results, please refer to [link to relevant documentation]. Figure 2 The diagram shown is a structural block diagram of a compensation and correction system for a surveying and mapping auxiliary component provided in this embodiment. The system includes a preprocessing module, an attitude error module, a deformation prediction module, a compensation module, and an update parameter module. The preprocessing module collects raw multi-source sensor data of the target component, and generates a preprocessed sensor dataset by cleaning, normalizing, aligning and encapsulating the raw multi-source sensor data. The attitude error module performs attitude integral inference estimation based on the preprocessed sensor dataset, generates innovation prediction values ​​and prediction confidence, and encapsulates the innovation prediction values ​​and prediction confidence by covariance adaptive estimation to obtain the attitude error estimate. The deformation prediction module performs deformation fusion processing on the attitude error estimate to obtain the preliminary deformation, residual prediction and theoretical deformation. Based on the preliminary deformation, residual prediction and theoretical deformation, the deformation prediction value is obtained through feedback correction and encapsulation processing. The compensation module calculates the confidence weight by weighting the predicted deformation value using the inverse variance of the redundancy error. Based on the confidence weight, it performs weighted detection, transformation, and encapsulation to obtain the final compensation amount. The parameter update module performs measurement correction exponential smoothing based on the final compensation amount to obtain the updated sensor parameters.

[0056] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0057] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A compensation and correction method for a surveying auxiliary component, characterized in that, The steps of this method are as follows: collect the original multi-source sensor data of the target component, and generate a preprocessed sensor dataset by cleaning, normalizing, aligning and encapsulating the original multi-source sensor data. Attitude integral inference estimation is performed based on the preprocessed sensor dataset to generate innovation prediction values ​​and prediction confidence. Covariance adaptive estimation is then performed on the innovation prediction values ​​and prediction confidence to obtain the attitude error estimate. The attitude error estimate is subjected to deformation fusion processing to obtain the preliminary deformation, residual prediction and theoretical deformation. Based on the preliminary deformation, residual prediction and theoretical deformation, the deformation prediction is obtained through feedback correction and encapsulation processing. The confidence weight is obtained by weighting the predicted deformation value with the inverse variance of the redundancy error. The final compensation amount is obtained by weighted detection, transformation and encapsulation based on the confidence weight. The sensor parameters are updated by performing measurement correction exponential smoothing based on the final compensation amount. Among them, the covariance adaptive estimation of the innovation prediction value and the prediction confidence is encapsulated, including: obtaining the adjusted process noise covariance by adaptive covariance processing based on the innovation sequence and the innovation prediction value; Anomaly secondary feedback correction is performed based on prediction confidence, innovation sequence, and innovation prediction value to obtain the corrected attitude estimate; Error estimation is encapsulated by combining the corrected attitude estimate and the preprocessed sensor dataset to obtain the attitude error estimate.

2. The compensation and correction method for a surveying auxiliary component according to claim 1, characterized in that, The process involves cleaning, normalizing, aligning, and encapsulating the raw multi-source sensor data, including: cleaning the raw multi-source sensor data using timestamps to generate synchronized cleaned data. Calibration compensation data is obtained by compensating calibration parameters based on synchronous cleaning data; The calibration compensation data is subjected to multimodal normalization, alignment, and encapsulation to generate a preprocessed sensor dataset.

3. The compensation and correction method for a surveying auxiliary component according to claim 1, characterized in that, Attitude integral inference estimation based on preprocessed sensor dataset includes: generating initial attitude quaternions by attitude integral initialization based on preprocessed sensor dataset. Quaternion recursion is performed based on the initial attitude quaternion and the preprocessed sensor dataset to obtain the attitude estimate and the innovation sequence. Wavelet inference estimation is performed on the innovation sequence to generate innovation prediction values ​​and prediction confidence.

4. The compensation and correction method for a surveying auxiliary component according to claim 1, characterized in that, The attitude error estimate is subjected to deformation fusion processing, including: obtaining the theoretical deformation amount by thermodynamic deformation processing based on the attitude error estimate; Random forest residual prediction is performed based on the attitude error estimate to obtain the residual prediction value; The theoretical deformation and the predicted residual value are initially fused to obtain the preliminary deformation.

5. The compensation and correction method for a surveying auxiliary component according to claim 1, characterized in that, The process involves feedback correction and encapsulation based on the preliminary deformation, residual prediction, and theoretical deformation, including: adaptive correction based on the residual prediction and theoretical deformation to obtain the corrected linear expansion coefficient, and adjusting the attitude error estimate based on the corrected linear expansion coefficient. Uncertainty assessment is performed based on the initial deformation amount to generate deformation uncertainty; The deformation uncertainty and the initial deformation amount are encapsulated to obtain the deformation prediction value.

6. The compensation and correction method for a surveying auxiliary component according to claim 1, characterized in that, The deformation prediction values ​​are weighted by the inverse variance of the redundancy error, including: obtaining a multi-source attitude error set by processing the redundancy error based on the deformation prediction values; The compensation calculation is performed based on the multi-source attitude error set to obtain the single-source compensation set; The confidence weights are obtained by applying inverse variance weighting to the single-source compensation set.

7. The compensation and correction method for a surveying auxiliary component according to claim 1, characterized in that, Weighted detection transformation and encapsulation based on confidence weights include: performing weighted detection based on single-source compensation set and confidence weights to obtain fused compensation amount and anomaly label; Based on the fusion compensation amount and anomaly markers, back projection consistency verification is performed to obtain the verification compensation amount; The compensation amount after verification is converted and encapsulated to obtain the final compensation amount.

8. The compensation and correction method for a surveying auxiliary component according to claim 1, characterized in that, The measurement correction index smoothing process based on the final compensation amount includes: obtaining the corrected measurement result by performing measurement correction processing based on the final compensation amount; Based on the corrected measurement results, residual statistical processing is performed to generate a residual statistical report; The residual statistical report is exponentially smoothed to obtain the updated sensor parameters.

9. The compensation and correction method for a surveying auxiliary component according to claim 1, characterized in that, The raw multi-source sensor data includes any one or more of the following: triaxial acceleration data, triaxial angular velocity data, biaxial tilt data, temperature data, and air pressure data.

10. A system for applying a compensation and correction method to a surveying auxiliary component according to any one of claims 1-9, characterized in that, The system includes: a preprocessing module, which collects raw multi-source sensor data of the target component, and generates a preprocessed sensor dataset by cleaning, normalizing, aligning and encapsulating the raw multi-source sensor data; The attitude error module performs attitude integral inference estimation based on the preprocessed sensor dataset, generates innovation prediction values ​​and prediction confidence, and encapsulates the innovation prediction values ​​and prediction confidence by covariance adaptive estimation to obtain the attitude error estimate. The deformation prediction module performs deformation fusion processing on the attitude error estimate to obtain the preliminary deformation, residual prediction and theoretical deformation. Based on the preliminary deformation, residual prediction and theoretical deformation, the deformation prediction value is obtained through feedback correction and encapsulation processing. The compensation module calculates the confidence weight by weighting the predicted deformation value using the inverse variance of the redundancy error. Based on the confidence weight, it performs weighted detection, transformation, and encapsulation to obtain the final compensation amount. The parameter update module performs measurement correction exponential smoothing based on the final compensation amount to obtain the updated sensor parameters.

Images