Closed loop monitoring and adaptive correction method for the levelness of a seismograph

Abstract

This invention provides a closed-loop monitoring and adaptive correction method for seismograph levelness, comprising the following steps: S1. In-situ construction and solidification of a dual-benchmark system for seismograph levelness; S2. Full decoupling of multi-field interference and extraction of true horizontal dip angle; S3. Extraction and type identification of dip angle deviation features; S4. Singularity-free calculation and conflict verification of leveling parameters; S5. Synchronous leveling drive and real-time dynamic deviation correction; S6. Verification of correction effect and iterative update of dual benchmarks. This invention achieves real-time closed-loop monitoring and high-precision adaptive correction of seismograph levelness throughout all time periods, ensuring long-term compliance of the instrument with observation specifications. It fundamentally solves the industry pain point of existing technologies easily erasing effective precursor signals of earthquakes. Simultaneously, it achieves precise decoupling of multi-field coupling interference, singularity-free and highly secure leveling control, and full-module bidirectional collaborative optimization, comprehensively improving the accuracy, integrity, and long-term reliability of seismic observation data.

Description

Technical Field

[0001] This invention relates to the field of automated operation and maintenance technology for seismic observation instruments and seismic stations, and in particular to a closed-loop monitoring and adaptive correction method for the levelness of a seismograph. Background Technology

[0002] Broadband seismographs are core infrastructure for earthquake observation, crustal deformation monitoring, and earthquake precursor detection. The levelness of their pickup units directly determines the accuracy, validity, and comparability of earthquake observation data. Both domestic and international seismic station observation standards impose strict requirements on the long-term levelness of seismographs. Currently, the monitoring and correction of seismograph levelness suffers from the following technical deficiencies:

[0003] First, existing levelness monitoring and correction methods mostly rely on manual periodic inspections and calibrations, which cannot achieve real-time continuous monitoring throughout the entire period. When the seismograph experiences horizontal shifts due to factors such as changes in ambient temperature, uneven settlement of the observation pier, ground micro-deformation, and mechanical creep, it cannot be detected and corrected in a timely manner, which can easily lead to systematic distortion of long-term observation data. Some existing automated correction schemes are mostly open-loop control modes, which only perform a single correction action after detecting a deviation. There is no closed-loop verification, status feedback, or iterative optimization mechanism. The accuracy decays significantly during long-term operation, which cannot meet the needs of seismic stations for decades of continuous observation.

[0004] Second, seismic stations experience nonlinear and nonstationary interference from multiple coupled fields, such as temperature drift, ground micro-vibration, and spatial electromagnetic field interference. Existing levelness monitoring schemes often employ conventional linear processing methods such as integer-order filtering and wavelet threshold filtering, which cannot accurately fit and eliminate nonlinear coupling interference. This can easily lead to two extreme problems: first, environmental interference signals are misjudged as instrument level offset, causing frequent miscalibrations and interfering with normal seismic observations; second, interference components cannot be effectively eliminated, resulting in distorted levelness monitoring results that fail to meet the requirements of high-precision observation.

[0005] Third, existing automated correction schemes cannot effectively distinguish and separate the instrument's own horizontal offset from the slow crustal deformation signals of earthquake precursors. Both types of signals exhibit long-period, slow-changing characteristics in dip monitoring data. Existing schemes lack effective feature recognition and reasoning mechanisms, and generally misjudge earthquake precursor deformation signals as instrument horizontal deviations for correction, directly erasing the core effective data of earthquake precursor monitoring. This seriously affects the analysis of earthquake gestation processes and earthquake prediction research. This is a core industry pain point that has long remained unresolved in the field of automated horizontal correction of seismic stations.

[0006] Fourth, existing leveling parameter planning methods mostly use Euler angles for inverse kinematics calculation of rigid bodies. This method has an inherent gimbal lock singularity defect, which can lead to calculation failure and leveling loss of control when the instrument tilt angle approaches the critical value. At the same time, existing solutions do not fully consider the nonholonomic motion constraints of the rigid bearing platform, which can easily cause platform torsion, overturning, and excessive stroke of the execution unit during the leveling process. The stability, safety and reliability of the correction process are insufficient.

[0007] Fifth, the existing solutions' monitoring, filtering, identification, and leveling modules are all independent, unidirectional transmission structures with no bidirectional data interaction or collaborative optimization mechanism between modules. They cannot adjust the operating parameters of each module in real time according to the on-site working conditions, signal characteristics, and execution status, resulting in poor environmental adaptability and inability to guarantee long-term operational stability in complex station environments. Summary of the Invention

[0008] This invention provides a closed-loop monitoring and adaptive correction method for seismograph levelness. First, it constructs a dual-benchmark traceability system between the instrument's rigid body and the geoid, completing the initial state locking of the hardware system and the initial parameter calibration of the algorithm, providing a unified and traceable calculation and judgment benchmark. Then, it employs a fractional-order LMS adaptive filtering chaotic synchronization interference cancellation algorithm to fully decouple multi-field coupled nonlinear interference from the real-time acquired dip angle raw signal, eliminating environmental interference components and extracting the undisturbed true horizontal dip angle data. Finally, it uses a multi-scale permutation entropy-hierarchical confidence rule base inference algorithm to perform multi-scale nonlinear feature extraction and uncertainty inference on the dip angle deviation signal, accurately identifying all deviation types and achieving... The instrument's own horizontal deviation is completely separated from the precursory crustal deformation signal of the earthquake. Correction and blocking protection are implemented for the precursory signal, and only the correctable effective horizontal deviation is output. Then, an adaptive iterative algorithm of Lie group and Lie algebra for nonholonomic constrained rigid body motion is used to complete the planning of singularity-free leveling execution parameters and multi-dimensional mechanical conflict verification, and output stable and safe synchronous leveling control parameters. The leveling execution process adopts a stroke-tilt dual closed-loop dynamic correction mode, which executes, monitors and corrects simultaneously to ensure leveling accuracy and process stability. After the correction is completed, the correction effect is verified in all dimensions and the closed-loop iterative update of the dual benchmark system is performed. Finally, it returns to the real-time monitoring link to form a closed-loop monitoring and correction system with no breaks in the entire process.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A closed-loop monitoring and adaptive correction method for the levelness of a seismograph includes the following steps:

[0011] S1. Construct the initial mechanical reference, the initial inclination reference of the instrument rigid body, and the in-situ reference of the geoid for seismic instrument level monitoring, and solidify them into a dual reference system;

[0012] S2. Synchronously acquire the raw horizontal dip angle data of the seismograph, and use the chaotic synchronization interference cancellation algorithm of fractional LMS adaptive filtering to decouple the raw horizontal dip angle data to obtain the true horizontal dip angle data;

[0013] S3. Calculate horizontal tilt deviation data based on real horizontal tilt data and dual reference system. Use multi-scale permutation entropy-hierarchical confidence rule base inference algorithm to extract features and identify deviation types of horizontal tilt deviation data to obtain identification results. If the identification result is the instrument's own horizontal deviation or mixed interference deviation, obtain the correction enable command and effective horizontal deviation data.

[0014] S4. After receiving the correction enable command, based on the effective horizontal deviation data and the dual reference system, the Lie group Lie algebra adaptive iterative algorithm for nonholonomic constrained rigid body motion is adopted to complete the singularity-free solution and conflict verification of the leveling parameters, and obtain the leveling execution parameters.

[0015] S5. Based on the leveling execution parameters, execute the seismograph synchronous leveling drive, acquire real-time horizontal dip data during the leveling process, synchronously collect real-time travel data, perform dual closed-loop dynamic correction based on real-time horizontal dip data and real-time travel data, and obtain the final horizontal dip data and final travel data after leveling is completed.

[0016] S6. Compare the final horizontal tilt data and final travel data with the dual reference system, perform a full-dimensional verification of the correction effect, and after the verification is qualified, perform a closed-loop iterative update of the dual reference system. After the update is completed, return to S2 and enter the next levelness closed-loop monitoring and correction cycle.

[0017] In this manual, the construction of the dual-reference system in S1 is specifically as follows: The rigid bearing platform of the seismograph is rigidly connected to the observation pier through three sets of adaptive leveling execution units evenly distributed in a 120° circle. All adaptive leveling execution units are controlled to move to the midpoint of their mechanical stroke and locked. The stroke code value of each adaptive leveling execution unit is read to construct the initial mechanical reference. Multiple sets of initial tilt angle data of the rigid bearing platform in the locked state are collected and smoothed and filtered to construct the initial tilt angle reference of the instrument's rigid body. The geoid normal reference data at the location of the observation pier is obtained and combined with the rigid deformation parameters of the observation pier upon completion and acceptance to construct the in-situ geoid reference. The three types of references are solidified and stored in a non-volatile storage unit to form a dual-reference system.

[0018] In this specification, the processing procedure of the chaotic synchronization interference cancellation algorithm of the fractional-order LMS adaptive filter in S2 is as follows: Multi-field interference data in the same environment as the original horizontal tilt data are synchronously acquired. The nonlinear characteristics of the multi-field interference data are fitted by the Chen chaotic master system, and a chaotic slave system with adaptive control terms is constructed to achieve global synchronization between the master and slave systems, completing the distortion-free reconstruction of the nonlinear interference signal. Based on the definition of Caputo fractional calculus, a fractional-order LMS adaptive filtering framework is constructed. Using the reconstructed interference signal as a reference, adaptive interference cancellation is performed on the original horizontal tilt data to eliminate multi-field coupled interference components, obtaining distortion-free true horizontal tilt data.

[0019] In this specification, the feature extraction process of the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm in S3 is as follows: multi-scale coarse-graining processing is performed on the horizontal tilt deviation data to obtain coarse-grained sequences at different time scales; phase space reconstruction is performed on the coarse-grained sequences at each scale to mine the nonlinear dynamic features within the sequences; the probability of the occurrence of the permutation pattern of each sequence after reconstruction is statistically analyzed, the permutation entropy value corresponding to each scale is calculated, and normalization processing is performed on all permutation entropy values ​​to obtain the multi-scale permutation entropy feature vector used for deviation type identification.

[0020] In this manual, the deviation type identification process of the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm in S3 is as follows: the extracted multi-scale permutation entropy feature vector is input into the three-level hierarchical confidence rule base, the activation weight of each rule is calculated, and all activation rules are fused and calculated through the evidence reasoning algorithm to obtain the comprehensive confidence level corresponding to four types of results: no effective deviation, instrument self-level deviation, earthquake precursor deformation signal, and mixed interference deviation. The result with the highest comprehensive confidence level is selected as the final identification result.

[0021] In this specification, the solution process of the Lie group Lie algebra adaptive iterative algorithm for nonholonomic constrained rigid body motion in S4 is as follows: the three-dimensional rigid body motion of the rigid bearing platform is represented by a special Euclidean group, and the rotational motion of the rigid bearing platform is represented by a special orthogonal group. Based on the effective horizontal deviation data and the initial mechanical reference, the initial pose matrix and the target pose matrix of the rigid bearing platform are constructed. The deviation matrix of the initial pose and the target pose are mapped to the Lie algebra space through logarithmic mapping, and the nonlinear pose optimization problem is transformed into an iterative solution problem in linear space. Under the preset nonholonomic constraints, the optimal leveling execution parameters are obtained by adaptive iterative algorithm, with no singular solutions throughout the process.

[0022] In this specification, the conflict check of the leveling parameters in S4 is specifically as follows: for the leveling execution parameters obtained by calculation, three types of constraint checks are sequentially performed. The first type is the mechanical stroke range constraint check of the adaptive leveling execution unit, the second type is the non-torsion constraint check of the planar motion of the rigid bearing platform, and the third type is the synchronous motion speed constraint check of multiple groups of adaptive leveling execution units. When all constraint conditions are met, it is determined that the check passes, and the leveling execution parameters are obtained.

[0023] In this specification, the chaotic synchronization interference cancellation algorithm of fractional-order LMS adaptive filtering, the multi-scale permutation entropy-hierarchical belief rule base inference algorithm, and the Lie group and Lie algebra adaptive iterative algorithm for non-holonomic constrained rigid body motion are pairwise bidirectionally interactively and synergistically optimized. Among them, the deviation discrimination result is real-time fed back to the chaotic synchronization interference cancellation algorithm of fractional-order LMS adaptive filtering to correct the operating parameters of the algorithm; the leveling execution state is real-time fed back to the chaotic synchronization interference cancellation algorithm of fractional-order LMS adaptive filtering to correct the characteristic parameters of interference reconstruction; the leveling stroke margin data is real-time fed back to the multi-scale permutation entropy-hierarchical belief rule base inference algorithm to correct the rule weights of deviation discrimination.

[0024] In this specification, the execution process of the double closed-loop dynamic correction in S6 is specifically as follows: during the whole process of leveling drive execution, interference decoupling processing is synchronously performed, and the real-time horizontal inclination angle data after decoupling is obtained in real time. At the same time, the real-time stroke data fed back by each adaptive leveling execution unit is collected. Based on the deviation between the real-time stroke data and the target stroke corresponding to the leveling execution parameters, stroke following closed-loop correction is performed. Based on the deviation between the real-time horizontal inclination angle data and the initial inclination angle reference of the instrument rigid body, inclination accuracy closed-loop correction is performed. The two closed-loops run synchronously, correcting while executing until both the stroke following deviation and the inclination deviation meet the preset requirements, and it is determined that the leveling is completed.

[0025] In this specification, the full-dimensional check of the calibration effect in S6 is specifically as follows: first, the calibration accuracy check is performed, and the residual deviation between the final horizontal inclination angle data and the initial inclination angle reference of the instrument rigid body is compared. If the residual deviation is less than the preset level threshold, it is determined that the accuracy check is qualified, and the long-term stability check is entered. The long-term stability check is to continuously collect the decoupled horizontal inclination angle data for a set duration and calculate the degree of dispersion of the data. If the degree of dispersion is less than the preset stability threshold, it is determined that the check is qualified.

[0026] In summary, the present invention has at least the following beneficial effects:

[0027] This invention enables real-time closed-loop monitoring and adaptive correction of the seismograph's levelness throughout its entire lifecycle and at all times. It eliminates the need for manual inspection and intervention, promptly identifies and corrects the instrument's horizontal deviation, and ensures that the seismograph meets the levelness requirements of the seismic station's observation specifications over the long term. This effectively avoids systematic distortion of observation data caused by horizontal deviation and is suitable for the long-term continuous observation needs of seismic stations.

[0028] This invention enables accurate fitting, reconstruction, and cancellation of multi-field coupled nonlinear and non-stationary interferences at seismic stations, effectively eliminating interference from environmental factors such as temperature, vibration, and electromagnetic fields on levelness monitoring. It significantly improves the accuracy and anti-interference capability of levelness monitoring, and fundamentally avoids the problems of miscorrection and undercorrection caused by environmental interference.

[0029] This invention achieves accurate identification and complete separation of the instrument's own horizontal deviation from the precursor crustal deformation signal of an earthquake. It can effectively identify the slow deformation signal corresponding to the precursor of an earthquake and perform correction and lockout protection. It fundamentally solves the core industry pain point that existing automated correction schemes easily erase effective precursor data, completely preserve the effective signal of earthquake observation, and ensure the data integrity and effectiveness of earthquake precursor analysis and earthquake research.

[0030] This invention uses Lie groups and Lie algebras to represent and solve rigid body motion. It is globally singular and completely avoids the gimbaling defect inherent in conventional Euler angle solutions. At the same time, it uses multi-dimensional nonholonomic constraints to control the boundary of the leveling process, ensuring that the rigid bearing platform does not twist, overturn, or exceed the travel limit during the leveling process, which greatly improves the stability, safety and reliability of the correction process.

[0031] The three core algorithm modules of this invention achieve bidirectional interaction and fusion, breaking the independent module structure of the existing technology that transmits information in one direction. Each module can collaboratively optimize the operating parameters in real time according to the on-site working conditions, signal characteristics, identification results, and execution status, which greatly improves the environmental adaptability, all-scenario adaptability, and long-term operating stability of the method.

[0032] This invention constructs a dual-benchmark traceability system that can be updated in a closed loop, taking into account both the absolute geographical traceability of the geoid benchmark and the long-term operational adaptability of the instrument's rigid body benchmark. This ensures the consistency, continuity, and traceability of the seismograph's observation data throughout its entire lifecycle, providing stable benchmark support for long-term observation and data comparison analysis at seismic stations. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the closed-loop monitoring and adaptive correction method for the levelness of a seismograph.

[0034] Figure 2 This is a flowchart illustrating the process of multi-field interference cancellation and tilt angle extraction in fractional-order chaotic synchronous filtering.

[0035] Figure 3 This is a schematic diagram of the process for multi-scale permutation entropy-confidence rule base bias identification and precursor locking.

[0036] Figure 4 This is a flowchart illustrating the adaptive iterative balancing and conflict checking process for Lie groups and Lie algebras. Detailed Implementation

[0037] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0038] like Figure 1 As shown, this embodiment provides a closed-loop monitoring and adaptive correction method for the levelness of a seismograph, including:

[0039] S1. In-situ construction and solidification of the dual benchmark system for seismograph levelness; This step completes the initial state locking of the hardware system and the construction of the benchmark system, providing a unified calculation basis and boundary constraints for all subsequent monitoring, decoupling, identification, and leveling processes.

[0040] First, hardware system initialization and mechanical reference locking are completed. The closed-loop main control unit is electrically connected to the three-component tilt sensing unit, multi-field coupling sensing unit, adaptive leveling execution unit, and earthquake precursor signal identification unit via shielded twisted-pair cables. All units are mounted on the rigid support platform of the seismograph body. The rigid support platform is rigidly connected to the observation pier through three sets of adaptive leveling execution units evenly distributed in a 120° circle. The closed-loop main control unit sends a zero-reset drive command to the three sets of adaptive leveling execution units, controlling the telescopic shafts of all adaptive leveling execution units to move to the midpoint of their mechanical travel. This position provides the maximum bidirectional telescopic margin, minimizing the risk of overtravel during subsequent leveling. Simultaneously, the mechanical return error at the midpoint is minimized, ensuring the calibration accuracy of the initial reference. After the telescopic shafts reach their positions, the closed-loop main control unit locks all telescopic shafts, reads and records the telescopic shaft travel code values ​​of the three sets of adaptive leveling execution units at this time, and constructs the initial mechanical reference coding matrix.

[0041] ; This is the midpoint travel code value for the first group of adaptive leveling execution units; The midpoint travel code value for the second group of adaptive leveling execution units This is the midpoint travel code value for the third group of adaptive leveling execution units.

[0042] Subsequently, the initial tilt angle reference calibration of the instrument's rigid body was completed. The X-axis and Y-axis of the three-component tilt angle sensing unit were completely aligned with the geographic north and geographic east directions, respectively, maintaining complete coordinate system consistency with the seismograph's pickup axis. The Z-axis was completely aligned with the normal axis of the rigid bearing platform, ensuring coordinate system consistency between tilt angle measurement and seismic observation, and avoiding correction errors caused by coordinate system deviation. The closed-loop main control unit continuously acquired 1000 sets of X-axis, Y-axis, and Z-axis tilt angle values ​​output by the three-component tilt angle sensing unit at a sampling frequency of 100Hz. After arithmetic averaging and filtering to eliminate random noise, the initial tilt angle matrix of the instrument's rigid body was constructed.

[0043] ; This is the initial tilt angle value along the X-axis; This is the initial tilt angle value along the Y-axis; This is the initial tilt angle value along the Z-axis. The closed-loop main control unit synchronously acquires the ambient temperature value output by the multi-field coupled sensing unit at this time, and records it as the reference temperature. .

[0044] Then, the in-situ geoid benchmark was constructed. The closed-loop main control unit accessed the GNSS geodetic benchmark service system via a 4G / 5G wireless communication module to obtain the geoid normal benchmark data at the location of the observation tower. Combined with rigid deformation parameters such as elastic modulus, Poisson's ratio, and coefficient of thermal expansion extracted from the observation tower's construction completion acceptance report, the in-situ geoid benchmark matrix was constructed.

[0045] ; This is the horizontal reference value of the geoid in the X direction; This is the horizontal reference value of the geoid in the Y direction; This is the geoid normal reference value. This reference is an absolute geographic reference and is only updated synchronously when reference update data is received from the GNSS geodetic reference service system.

[0046] The closed-loop main control unit reads three types of basic data from the non-volatile storage unit to complete the initial parameter set calibration of the three core algorithms. The sources of the three types of basic data are: full-condition interference calibration test data completed on the observation pier after the seismograph was installed using a high and low temperature environment simulation device, a standard shaking table, and an electromagnetic field simulation device; the earthquake precursor deformation signal characteristic database released by the seismic network center; the design drawings and factory calibration parameters of the rigid bearing platform; and the factory calibration parameters of the adaptive leveling execution unit. Based on the three types of basic data, the initial parameter sets of the three algorithms are constructed. All parameters are calibrated through pre-training optimization and are solidified and stored in the non-volatile storage unit along with the dual-benchmark system. Subsequent closed-loop iterative updates are only completed synchronously with the dual-benchmark system after the calibration effect is verified to be qualified.

[0047] The initial parameter subset of the chaotic synchronization interference cancellation algorithm for fractional-order LMS adaptive filtering includes the fractional-order order, fixed control parameters of the Chen chaotic system, adaptive control gain coefficient, initial step size factor of fractional-order LMS filtering, and fractional-order differential term adjustment coefficient. All parameters were determined after pre-training and optimization using full-condition interference calibration test data.

[0048] The initial parameter subset of the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm includes the embedding dimension and delay time of the multi-scale permutation entropy, the total number of rules in the hierarchical confidence rule base, the rule weights, the premise attribute weight vectors, and the initial confidence of the rules. All parameters are determined after pre-training and optimization using the earthquake precursor deformation signal feature library and field calibration data.

[0049] The initial parameter subset of the Lie group Lie algebra adaptive iterative algorithm for nonholonomic constrained rigid body motion includes the adaptive iteration initial step size, Jacobian matrix initial parameters, nonholonomic constraint threshold, adaptive leveling execution unit travel boundary parameters, and iteration step size adjustment coefficient. All parameters are determined after calibration using the equipment factory calibration parameters and rigid platform design parameters.

[0050] After completing all benchmark construction and parameter calibration, the closed-loop main control unit will uniformly solidify and store the initial mechanical benchmark, the initial inclination benchmark of the instrument rigid body, the in-situ benchmark of the geoid, and the initial parameter set of the algorithm into a non-volatile storage unit, thus completing the construction of the full-process traceability benchmark.

[0051] The initial mechanical reference coding matrix, the initial inclination matrix of the instrument rigid body, the in-situ reference matrix of the geoid, and the initial parameter set of the algorithm are uniformly stored in a non-volatile storage unit to complete the in-situ construction of the dual reference system.

[0052] S2. Full Decoupling of Multi-Field Interference and Extraction of True Horizontal Tilt Angle: This step uses a fractional-order LMS adaptive filtering chaotic synchronous interference cancellation algorithm (FOCLMS-CS algorithm) to fully decouple multi-field coupled interference and extract true horizontal tilt angle data unaffected by environmental interference from the original tilt angle signal acquired by the three-component tilt angle sensing unit. This algorithm can accurately cancel non-stationary and nonlinear temperature, vibration, and electromagnetic multi-field coupled interference in seismic station scenarios, while preserving the low-frequency long-term trend characteristics of earthquake precursor signals and avoiding the false filtering of effective signals. The fractional-order chaotic synchronous filtering multi-field interference cancellation and tilt angle extraction process is as follows: Figure 2 As shown.

[0053] First, the FOCLMS-CS algorithm model was constructed, which consists of three core related parts.

[0054] The first part presents the fundamental definition of Caputo fractional calculus. A fractional adaptive filtering framework is constructed using the Caputo fractional derivative definition, which has the strongest engineering applicability. The initial conditions of this definition are in integer order, perfectly matching the initial state of hardware in practical engineering. The Caputo fractional derivative definition is as follows:

[0055] ; The fractional order is a real number ranging from 0 to 2, used to control the memory characteristics and convergence speed of fractional calculus. greater than The smallest positive integer when When the value range is between 0 and 1, The value is 1; is a gamma function used to implement the continuous field extension of fractional factorials; Let be a time-domain continuous function that needs to be differentiated; For function of Integer derivative; It is a time variable; It is the integral variable.

[0056] The second part is the Chen chaotic synchronization interference reconstruction system. Multi-field coupled interference from seismic stations exhibits strong nonlinearity and non-stationarity, making accurate reconstruction and cancellation impossible with conventional linear filtering methods. Therefore, a master-slave synchronization system is constructed using the Chen chaotic system. The master system is used to fit the nonlinear interference signal acquired by the multi-field coupled sensing units, while the slave system is used to reconstruct the interference signal without distortion, providing an accurate interference reference for subsequent adaptive cancellation. The definition of the Chen chaotic master system is as follows:

[0057] ; , , These are the three state variables of the Chen chaotic master system; , , , The fixed control parameters for the Chen chaotic system are 35, 3, 28, and 1, respectively, which can ensure that the system is always in a chaotic state. The multi-field interference data matrix output by the multi-field coupled sensing unit serves as the driving input for the chaotic main system. The multi-field coupled sensing unit consists of a temperature sensor, a triaxial vibration sensor, an electric field strength sensor, and a magnetic field strength sensor. The installation distance between all sensors and the three-component tilt angle sensing unit does not exceed 5cm to ensure synchronous measurement with the environment.

[0058] To achieve accurate reconstruction of the interference signal, a Chen chaotic slave system with adaptive control term is constructed as the response system of the master system, and its definition is as follows:

[0059] ; , , For Chen, chaos is derived from the three state variables of the system; , , This is an adaptive control term for the slave system, used to achieve chaotic synchronization between the master and slave systems. When the systems are synchronized, the state variables of the slave system can completely reproduce the nonlinear disturbance characteristics of the master system.

[0060] Define a synchronization error variable for the master-slave system to quantify the synchronization state of the master-slave system and drive the update of the adaptive control term:

[0061] ; , , The synchronization error variable of the master-slave system converges to 0 when the system achieves complete synchronization.

[0062] An adaptive control term is constructed based on the synchronization error variable to ensure that the master-slave system can quickly achieve global asymptotic synchronization. The control term is defined as follows:

[0063] ; , , This represents the adaptive control gain coefficient. When the master-slave system achieves chaotic synchronization, the state variables of the slave system can completely reconstruct the nonlinear disturbance signal of the master system, thus obtaining the reconstructed disturbance signal. .

[0064] The third part is the fractional-order LMS adaptive filtering weight update model. Based on the reconstructed interference signal, a fractional-order LMS adaptive filtering framework is constructed to achieve adaptive cancellation of the interference signal. This framework introduces a fractional-order differential term, which can retain the long-term memory characteristics of historical inputs, resulting in faster convergence speed and smaller steady-state error. Furthermore, by adjusting the fractional-order term, it can adapt to the filtering requirements under different interference scenarios. The filtering input-output model is as follows:

[0065] ; The original tilt matrix is ​​the output of the three-component tilt sensing unit, and the observation input is the filter model. is the true tilt angle matrix of the rigid body after decoupling, and is the expected output of the filtering model; This is the weight vector for the fractional-order LMS adaptive filter, used to control the cancellation weights of the reconstructed interference signal.

[0066] Define a filtering error signal to drive the adaptive update of the weight vector. The error signal is the difference between the original tilt angle signal and the filtered output signal.

[0067] ; The filtering error signal is such that, after filtering is completed, the error signal converges to an effective signal containing only the true tilt angle of the rigid body.

[0068] The weight update equation is constructed based on the Caputo fractional derivative, and a fractional differential term is introduced to realize the fusion of historical memory characteristics. The update equation is as follows: ; This is the initial step size factor for the fractional-order LMS adaptive filter, used to control the convergence rate of weight updates; This is the adjustment coefficient for the fractional derivative term, used to control the weight of historical memory characteristics.

[0069] Subsequently, the FOCLMS-CS algorithm model training process was completed. This training was performed after the seismograph was installed but before it was officially put into observation. The training results were stored in a non-volatile memory unit as initial parameters for the algorithm's formal application. The closed-loop master control unit read the full-condition interference calibration test data from the non-volatile memory unit, dividing the data into a training set (80%) and a validation set (20%). The training data covered a temperature range of -40℃ to 60℃, a vibration acceleration range of 0.01g to 2g, and a full range of electric and magnetic field interference conforming to the electromagnetic compatibility standards of seismic stations, containing over 1000 samples with different interference combinations. The closed-loop master control unit initialized the state variables and control parameters of the Chen chaotic master-slave system, initialized the weight vector, step size factor, fractional order, and adjustment coefficients of the fractional-order LMS adaptive filter, input the multi-field interference data matrix from the training set into the chaotic master system, used the original dip matrix from the training set as the observation input of the filtering model, and used the standard interference-free dip matrix from the training set as the expected output of the filtering model, performing iterative training. After each iteration, the closed-loop control unit calculates the mean square error (MSE) of the filtered error signal. Based on the MSE value, it optimizes and adjusts the fractional order, step size factor, adjustment coefficient, and adaptive control gain coefficient until the MSE value converges to below a preset threshold. The closed-loop control unit inputs the trained optimal parameters into the validation set for cross-validation. After verifying that the filtered MSE is less than 0.1 arcseconds, the optimal parameters are stored in a non-volatile memory unit, completing the algorithm model training.

[0070] The FOCLMS-CS algorithm model was then applied online in real time. The algorithm ran synchronously with the seismograph's sampling frequency of 100Hz. The closed-loop main control unit synchronously acquired the original dip matrix output from the three-component dip sensing unit and the multi-field interference data matrix output from the multi-field coupled sensing unit according to the preset sampling frequency. This multi-field interference data matrix was then input into the trained Chen chaotic master system, and the master-slave chaotic synchronization system was run to obtain the reconstructed interference signal. The closed-loop main control unit input the original dip matrix and the reconstructed interference signal into the fractional-order LMS adaptive filtering model, performed weight updates and interference cancellation, and obtained the decoupled true rigid body dip matrix. Then, a moving average filter with a sliding window length of 100 sampling points was applied to the true rigid body dip matrix to remove residual random noise, resulting in the true horizontal dip matrix.

[0071] ; This is the real-time tilt angle value of the X-axis after decoupling and filtering; This is the real-time tilt angle value of the Y-axis after decoupling and filtering; This is the real-time tilt angle value of the Z-axis after decoupling and filtering.

[0072] After extracting the true horizontal tilt matrix, input the matrix into the deviation identification process.

[0073] S3. Tilt Deviation Feature Extraction and Type Identification; This step employs the Multi-Scale Permutation Entropy-Hierarchical Confidence Rule Base Inference Algorithm (MPE-HBRB Algorithm) to accurately identify the type of horizontal deviation, achieving complete separation between the instrument's own horizontal deviation and the precursor deformation signal of the earthquake. This solves the core industry pain point of conventional methods misclassifying slowly changing precursor deformation as horizontal deviation, leading to the correction and erasure of precursor signals. This algorithm can achieve high accuracy in identifying deviation types in small-sample, high-interference seismic station environments, while ensuring zero omissions in the identification of earthquake precursor signals. The Multi-Scale Permutation Entropy-Confidence Rule Base deviation identification and precursor blocking process is as follows: Figure 3 As shown.

[0074] First, the MPE-HBRB algorithm model was constructed, which consists of two core related parts.

[0075] The first part is the multi-scale permutation entropy feature extraction model. The instrument's own horizontal deviation is mostly a high-frequency abrupt signal, while the earthquake precursor deformation signal is a low-frequency long-trend nonlinear signal. This algorithm uses multi-scale permutation entropy to extract multi-scale nonlinear features from the dip deviation signal, simultaneously covering both high-frequency detail features and low-frequency trend features, adapting to the differences in characteristics between the two types of signals. First, the input dip deviation time series is subjected to multi-scale coarse-grained processing. The coarse-grained sequence is defined as follows:

[0076] , ; The scaling factor, which takes a positive integer value, is used to control the coarsening time scale. In this algorithm, the scaling factor ranges from 1 to 20, covering the full frequency range of earthquake precursor signals and instrument bias signals. The input tilt angle deviation time series is calculated from the difference between the true horizontal tilt angle matrix and the constructed initial tilt angle matrix of the instrument rigid body; The total length of the tilt angle deviation time series; This is the floor function; scale factor The corresponding number A coarse-grained sequence value.

[0077] Phase space reconstruction is performed on the coarse-grained sequences at each scale to mine the nonlinear dynamic features within the sequences. The reconstructed sequences are defined as follows: , ; The embedding dimension is used to control the dimension of phase space reconstruction; This is the delay time, used to control the time delay of phase space reconstruction; scale factor The corresponding number A reconstructed phase space vector.

[0078] The elements within each reconstructed phase space vector are arranged in ascending order to obtain the corresponding permutation pattern. The probability of each permutation pattern is calculated, and the permutation entropy value corresponding to each scale is calculated to quantify the complexity of the sequence at that scale. The definition is as follows:

[0079] ; scale factor The corresponding number The probability of a certain permutation pattern occurring; Embedding dimension The corresponding factorial is the total number of permutation patterns; scale factor The corresponding permutation entropy value indicates that the higher the entropy value, the greater the complexity of the sequence at that scale.

[0080] The permutation entropy value is normalized to eliminate the influence of embedding dimension on the entropy value, resulting in a multi-scale permutation entropy feature vector, which serves as the input feature for the subsequent confidence rule base. The definition is as follows:

[0081] ; This is the normalized multi-scale permutation entropy feature vector.

[0082] The second part is the hierarchical confidence rule base inference model. The deviation signals from seismic stations exhibit strong uncertainty. This algorithm employs a hierarchical confidence rule base to achieve uncertain inference. Unlike conventional single-level confidence rule bases, the hierarchical structure divides the inference process into three levels: feature layer, type layer, and decision layer. This effectively reduces the number of rules, improves inference accuracy and speed, and perfectly adapts to the small-sample, high-uncertainty inference scenarios of seismic stations. The definition of a confidence rule is as follows:

[0083] : , , ; For the first Confidence rules; For the first The set of reference values ​​for the prerequisite attributes corresponding to each rule corresponds to the value range of the multi-scale permutation entropy feature vector; , , , The four result levels of the rule correspond to no effective deviation, instrument self-leveling deviation, geodetic frontal deformation signal, and mixed interference deviation, respectively. For the first The rule weight of each rule, with a value ranging from 0 to 1, is used to measure the importance of the rule; For the first The prerequisite attribute weight vector for each rule, with values ​​ranging from 0 to 1, is used to measure the importance of each input feature.

[0084] An evidence-based reasoning algorithm is used to perform fusion reasoning on the activated rules to obtain the confidence level of each outcome. First, the confidence level of the first outcome is calculated. The activation weight of a rule quantifies the degree of activation of the rule under the current input, and is defined as follows:

[0085] ; For the first The activation weight of each rule; For the first Rule number 1 The degree of matching of the prerequisite attributes corresponding to each input feature; This represents the total number of rules in the confidence rule base.

[0086] Based on activation weights, an evidence reasoning algorithm is used for fusion to obtain the comprehensive confidence score for each outcome level, as defined below:

[0087] ; For the first The overall confidence level corresponding to each outcome level The values ​​are 1, 2, 3, and 4, corresponding to four result levels; For the first Rule number 1 The initial confidence level corresponding to each outcome level.

[0088] Finally, the result level with the highest overall confidence level is selected as the inference output to obtain the bias type identifier. Simultaneously, the effective level bias value is calculated, defined as follows: ; It is an 8th-order Butterworth high-pass filter function with a cutoff frequency of 0.1Hz. It is used to extract the high-frequency instrument deviation component in the mixed interference deviation and filter out the low-frequency earthquake precursor signal component with a frequency lower than 0.1Hz to avoid the effective earthquake precursor signal being miscorrected.

[0089] The MPE-HBRB algorithm model training process is then completed. This training is performed after the seismograph is installed but before it is officially put into observation. The training results are stored in non-volatile storage units as initial parameters for the algorithm's formal application. The closed-loop main control unit first performs pre-training of the FOCLMS-CS algorithm in S2 to obtain the optimal decoupled model. Then, it uses this model to decouple the original dip angle data from the full-condition interference calibration test data to obtain interference-free true dip angle data. The difference between this true dip angle data and the instrument's rigid body initial dip angle reference is used as the dip angle deviation time series. Combined with standard precursor waveforms from the earthquake precursor deformation signal feature library, a training dataset for the MPE-HBRB algorithm is constructed. Each sample in the dataset contains: a multi-scale permutation entropy feature vector (calculated from the above dip angle deviation time series) and manually labeled deviation type labels (0~3). The earthquake precursor deformation signal samples include precursor deformation observation data of all earthquakes of magnitude 6 or above released by the China Earthquake Networks Center in the past 20 years, as well as artificially simulated slow deformation signals of different rates and amplitudes. The instrument horizontal deviation samples include instrument tilt data under different temperatures and vibrations, with no fewer than 500 samples in each category. The closed-loop main control unit initializes the number of rules, rule weights, premise attribute weights, and initial confidence levels of the hierarchical confidence rule base, and initializes the embedding dimension and delay time of the multi-scale permutation entropy. The multi-scale permutation entropy features of the training dataset are input into the confidence rule base, and evidence reasoning is performed to obtain the reasoning result. The closed-loop main control unit calculates the mean square error between the reasoning result and the sample label, and uses a constrained covariance evolution algorithm to optimize and adjust the rule weights, premise attribute weights, initial confidence levels, embedding dimension, and delay time. The optimization objective is to achieve a false negative rate of 0 for earthquake precursor signals and an overall false negative rate of less than 1%, until the mean square error value converges to below a preset threshold. The closed-loop main control unit performs cross-validation on the optimal parameters after training. Once the validation is successful, the optimal parameters are stored in a non-volatile storage unit, thus completing the training of the algorithm model.

[0090] Afterwards, the real-time online application of the MPE-HBRB algorithm model was completed, with the algorithm running synchronously with the FOCLMS-CS algorithm. The closed-loop main control unit receives the real horizontal tilt matrix, calculates the tilt deviation time series, performs multi-scale coarsening, phase space reconstruction, and permutation entropy calculation on the tilt deviation time series, and obtains the normalized multi-scale permutation entropy feature vector. The feature vector is input into the trained hierarchical confidence rule base, the rule activation weights are calculated, evidence reasoning fusion is performed, and the comprehensive confidence level corresponding to the four result levels is obtained. The result level with the highest comprehensive confidence level is selected, and the corresponding deviation type identifier is output. At the same time, the effective horizontal deviation value is calculated according to the definition formula, and a one-to-one full-type output is performed for the four types of deviation.

[0091] When the highest overall confidence level is When the deviation type is 0, corresponding to no valid deviation, the valid level deviation value is zero, a standby command is output, and the system directly returns to the S2 stage for continuous loop monitoring. When the result level with the highest overall confidence is... When the deviation type is identified as 1, corresponding to the instrument's own horizontal deviation, the effective horizontal deviation value is the difference between the true horizontal tilt matrix and the instrument's initial rigid body tilt matrix, and a correction enable command is output. When the result level with the highest overall confidence is... When the deviation type is identified as 2, corresponding to the precursor deformation signal relative to the geoid, the effective horizontal deviation value is zero matrix, a correction lockout command is output, and the correction operation is prohibited. Simultaneously, the absolute precursor deformation signal relative to the geoid reference is... The signal is sent to the earthquake precursor signal identification unit for storage and reporting. When the result level with the highest overall confidence is... When the deviation type is identified as 3, it corresponds to mixed interference deviation. The effective horizontal deviation value is the high-frequency instrument deviation component after high-pass filtering. The correction enable command is output, and the low-frequency component is sent to the earthquake precursor signal discrimination unit.

[0092] After completing the deviation type identification, the output deviation type identifier is fed back to the FOCLMS-CS algorithm of S2 in real time to complete the collaborative optimization of the filtering parameters. At the same time, the effective horizontal deviation value and the correction enable command are input to the leveling parameter planning stage to provide accurate target input and action boundary for subsequent leveling execution.

[0093] S4. Singularity-Free Solution and Conflict Verification of Leveling Parameters: This step employs the Lie group-Lie algebra adaptive iterative algorithm (LG-LA algorithm) for nonholonomically constrained rigid body motion to accurately plan the leveling execution parameters and verify mechanical conflicts. This solves the gimbaled singularity problem inherent in conventional Euler angle kinematics inverse kinematics solutions, while ensuring that the rigid bearing platform experiences no torsion, overturning, or excessive travel during the leveling process. This is the core execution foundation for achieving high-precision and high-safety adaptive correction in this scheme. Unlike conventional inverse kinematics methods, this approach uses Lie group-Lie algebras to represent rigid body motion, ensuring global singularity. It also introduces nonholonomic constraints and an adaptive iterative mechanism, balancing leveling accuracy, convergence speed, and operational safety. The Lie group-Lie algebra adaptive iterative leveling solution and conflict verification process is as follows: Figure 4 As shown.

[0094] First, the LG-LA algorithm model was constructed, which consists of three core related parts.

[0095] The first part is the rigid body motion representation model based on Lie group and Lie algebra. Conventional Euler angle representation methods suffer from gimbaled singularity, and the solution fails when the tilt angle approaches the critical value. This algorithm uses a special Euclidean group SE(3) to represent the three-dimensional rigid body motion of the rigid bearing platform, a special orthogonal group SO(3) to represent the rotational motion, and a so(3) Lie algebra to represent the tangent space of the rotational motion. It is globally singular and can adapt to the solution requirements of tilt angle deviations across the entire range. The initial pose of the rigid bearing platform is represented by the SE(3) group, defined as follows:

[0096] ; The initial pose matrix of the rigid bearing platform belongs to the SE(3) group; The initial rotation matrix of the rigid bearing platform belongs to the SO(3) group and is calculated from the effective horizontal deviation value; The initial translation vector of the rigid bearing platform is calculated from the initial mechanical reference encoding matrix; It is a 3-dimensional zero row vector.

[0097] The target pose of the rigid bearing platform is an absolutely horizontal state, and the corresponding target pose matrix is ​​defined as follows: ; The target pose matrix of the rigid bearing platform belongs to the SE(3) group; The target rotation matrix of the rigid bearing platform belongs to the SO(3) group, is a 3rd order identity matrix, and corresponds to the absolute horizontal state; The target translation vector of the rigid bearing platform is calculated from the target stroke encoding values ​​of the three sets of adaptive leveling execution units.

[0098] Define the pose error matrix between the initial pose and the target pose to quantify the deviation between the current pose and the target pose. The definition is as follows: ; Let be the pose error matrix, which belongs to the SE(3) group.

[0099] The pose error matrix on the SE(3) group is mapped to the se(3) Lie algebra space using a logarithmic mapping to obtain the Lie algebra vector. This transforms the nonlinear pose optimization problem into an iterative solution problem in a linear space, as defined below:

[0100] ; Let be the pose error vector in the se(3) Lie algebra space, which includes rotation error and translation error; For the logarithmic mapping from the SE(3) group to the se(3) Lie algebra; It is an isomorphic mapping operator from antisymmetric matrices to vectors.

[0101] The second part involves the construction of nonholonomic constraints. The rigid bearing platform is an equilateral triangular rigid plate. During the leveling process, it is necessary to ensure that the platform does not undergo torsional deformation, overturning, or exceed its travel limits. This algorithm constructs three types of nonholonomic constraints to provide boundary constraints for the iterative solution process, ensuring the stability and safety of the leveling process.

[0102] The first type is planar motion constraint, which ensures that the rigid bearing platform maintains planar motion and no torsional deformation during the leveling process, thus avoiding internal stress on the rigid bearing platform that could affect the seismograph's observation accuracy. The constraint formula is as follows: ; For the first The rotation matrix corresponding to the next iteration; This is the initial normal vector of the rigid bearing platform; For the first The change in the normal vector corresponding to each iteration is constrained to not exceed a preset threshold to prevent the platform from twisting.

[0103] The second type is stroke constraint, which ensures that the stroke of the three sets of adaptive leveling actuators is always within the mechanically permissible range to avoid equipment damage. The constraint formula is as follows: ; For the first The iteration corresponding to the first The travel code value of the group adaptive leveling execution unit; The minimum stroke code value for the adaptive leveling actuator is derived from the equipment's factory calibration parameters; The maximum stroke code value for the adaptive leveling actuator is derived from the equipment's factory calibration parameters.

[0104] The third type is synchronous motion constraint, which ensures the synchronous motion speed of the three sets of adaptive leveling execution units, avoids the overturning of the rigid bearing platform, and avoids generating additional vibration interference. The constraint formula is as follows:

[0105] ; For the first The iteration corresponding to the first The group adaptively adjusts the movement speed of the execution unit; The maximum permissible speed difference is fixed at 0.01 mm / s.

[0106] The third part is the adaptive iterative solution model. Based on the Lie algebra pose error vector, an adaptive iterative solution model is constructed under three types of nonholonomic constraints to iteratively solve for the optimal target travel code value. The iteration step size can be adjusted in real time according to the error changes, balancing convergence speed and solution stability. The iterative update formula is as follows:

[0107] ; For the first The Lie algebra pose vector corresponding to the next iteration; For the first The adaptive iteration step size corresponding to each iteration is used to control the iteration convergence speed and stability. For the first The Moore-Penrose generalized inverse of the Jacobian matrix corresponding to the next iteration is used to establish the mapping relationship between the Lie algebra pose vector and the run-code value of the adaptive leveling execution unit.

[0108] The formula for updating the adaptive iteration step size is as follows, which can be adjusted in real time according to error changes. When the error decreases rapidly, the step size is increased to improve the convergence speed; when the error oscillates, the step size is decreased to ensure stability.

[0109] ; This is the adjustment coefficient for the iteration step size, with a fixed value of 0.5; Let be the 2-norm of the vector. The iteration stops when the 2-norm of the pose error vector converges to below a preset threshold and satisfies all nonholonomic constraints, thus obtaining the optimal target travel coding matrix.

[0110] The LG-LA algorithm model training process is then completed. This training is performed after the seismograph is installed but before it is officially put into observation. The training results are stored in a non-volatile memory unit as initial parameters for the algorithm's formal application. The closed-loop control unit reads the rigid body motion parameters of the rigid bearing platform and the factory calibration parameters of the adaptive leveling execution unit from the non-volatile memory unit to construct a training dataset. This dataset contains the full range of dip angle deviations from 0 to 30 arcseconds, with each deviation gradient corresponding to more than 10 samples, covering the full travel range of three adaptive leveling execution units. The closed-loop control unit initializes the iteration step size, Jacobian matrix, and nonholonomic constraint threshold, as well as the Lie group Lie algebra mapping parameters. It inputs the dip angle deviation values ​​from the training dataset into the algorithm model and performs adaptive iterative solving to obtain the solved travel encoding values. The closed-loop control unit calculates the mean square error between the solution and the standard value, optimizing and adjusting the initial iteration step size, Jacobian matrix parameters, and nonholonomic constraint threshold until the mean square error converges below the preset threshold, and the number of iterations does not exceed 20, meeting real-time requirements. The closed-loop main control unit verifies the optimal parameters across the entire travel range after training. After verifying that the solution accuracy is better than 0.1 arcseconds, the optimal parameters are stored in a non-volatile storage unit, thus completing the algorithm model training.

[0111] Afterwards, the LG-LA algorithm model is triggered for online application. It is only started when a correction enable command is received; otherwise, it returns directly to S2 for continuous loop monitoring. The closed-loop main control unit receives the effective horizontal deviation value and the correction enable command. When the correction enable command is effective, the algorithm is started. Based on the effective horizontal deviation value, the initial rotation matrix and initial pose matrix of the rigid bearing platform are calculated, the target pose matrix is ​​constructed, the pose error matrix is ​​calculated, and the Lie algebra pose error vector is obtained through logarithmic mapping. Under three types of nonholonomic constraints, the closed-loop main control unit performs adaptive iterative solution. After each iteration, the pose vector, iteration step size, and Jacobian matrix are updated until the 2-norm of the pose error vector converges to below the preset threshold. After the iteration converges, the closed-loop main control unit converts the Lie algebra pose vector obtained by the solution into the target pose matrix on the SE(3) group through exponential mapping. The target travel coding matrix of the three adaptive leveling execution units is calculated, and the travel margin matrix is ​​calculated. The travel margin matrix is ​​the difference between the maximum travel coding value of the adaptive leveling execution unit and the target travel coding matrix.

[0112] The closed-loop main control unit performs conflict verification on the iteration results. When any coded value in the target stroke encoding matrix exceeds the mechanical stroke range, a stroke conflict is determined, a stroke conflict alarm signal is output, the leveling operation is terminated, and the system returns to S2 for continuous monitoring. Simultaneously, the alarm information is reported to the remote monitoring platform. When all coded values ​​in the target stroke encoding matrix are within the mechanical stroke range and all non-holonomic constraints are met, no stroke conflict is determined, the leveling execution parameter matrix and synchronous drive path parameters are output, and the synchronous drive mode and drive speed of the three sets of adaptive leveling execution units are planned synchronously.

[0113] After completing the leveling parameter planning, the output travel margin matrix is ​​fed back to the MPE-HBRB algorithm in S3 in real time to complete the collaborative optimization of rule weights. The rigid body motion acceleration during the leveling execution process is fed back to the FOCLMS-CS algorithm in S2 in real time to complete the collaborative optimization of disturbance reconstruction. At the same time, the leveling execution parameter matrix and the synchronous drive path parameters are input to the leveling execution stage.

[0114] S5. Synchronous Leveling Drive and Real-time Dynamic Deviation Correction: This step, based on the leveling execution parameter matrix and synchronous drive path parameters, reuses the FOCLMS-CS algorithm of S2 and the MPE-HBRB algorithm of S3 to achieve a dual closed-loop drive mode that executes, monitors, and corrects simultaneously. This can eliminate mechanical backlash errors, jamming, vibration interference, and other problems during the execution process in real time, ensuring that the leveling is completed in one go and improving the leveling efficiency and accuracy.

[0115] The closed-loop main control unit receives the leveling execution parameter matrix and synchronous drive path parameters, and synchronously sends drive commands to the three sets of adaptive leveling execution units. It controls all adaptive leveling execution units to move synchronously toward the target stroke encoding value according to the preset drive speed, ensuring that the rigid bearing platform always maintains planar motion during the leveling process, and avoiding torsional deformation and vibration interference.

[0116] Throughout the entire driving process, the closed-loop main control unit synchronously collects two types of real-time data at a sampling frequency of 100Hz. The first type is the FOCLMS-CS algorithm that repeatedly executes S2 to obtain the tilt angle matrix of the execution process after real-time decoupling, eliminating vibration interference and environmental interference generated during the leveling process, and obtaining distortion-free real-time tilt angle data. The second type is the current stroke encoding matrix fed back in real time by 3 sets of adaptive leveling execution units to obtain the real-time position data of the execution units.

[0117] The closed-loop master control unit synchronously calculates two types of real-time deviations: the first is the travel following deviation, which is the difference between the target travel encoding matrix and the current travel encoding matrix; the second is the real-time tilt angle deviation, which is the difference between the tilt angle matrix during execution and the initial tilt angle matrix of the instrument's rigid body. Based on these two types of real-time deviations, the closed-loop master control unit performs dual-closed-loop dynamic correction.

[0118] When the absolute value of the travel tracking deviation is less than or equal to the travel resolution, and the absolute value of the real-time tilt deviation is less than the preset levelness threshold, the operation is considered complete. All drives are stopped, the telescopic axes of all adaptive leveling execution units are locked, and the final tilt matrix and final travel encoding matrix are recorded and output to S6. The preset levelness threshold is 2 arcseconds, derived from the upper limit of levelness required by the seismic station observation specifications.

[0119] When the absolute value of the travel following deviation is greater than the travel resolution and the absolute value of the real-time tilt angle deviation is less than the preset levelness threshold, it is determined to be a travel following deviation. Travel closed-loop correction is performed, the number of drive pulses is adjusted according to the travel following deviation, and the adaptive leveling execution unit is controlled to move towards the target travel encoding value until the absolute value of the travel following deviation is less than or equal to the travel resolution.

[0120] When the absolute value of the travel follow deviation is less than or equal to the travel resolution, but the absolute value of the real-time tilt deviation is greater than or equal to the preset levelness threshold, it is determined to be a tilt closed-loop deviation. Tilt closed-loop correction is then performed. The real-time tilt deviation is used as the new effective levelness deviation value and substituted into the LG-LA algorithm of S4 to recalculate the target travel encoding value. Secondary drive correction is then performed until the absolute value of the real-time tilt deviation is less than the preset levelness threshold.

[0121] When the absolute value of the travel follow deviation is greater than the travel resolution and the absolute value of the real-time tilt angle deviation is greater than or equal to the preset levelness threshold, it is determined that the operation has lost synchronization. All drives are stopped immediately, all telescopic axes are locked, a loss of synchronization alarm signal is output, the leveling operation is terminated, and the system returns to S2 for continuous monitoring. At the same time, the alarm information is reported to the remote monitoring platform.

[0122] After the leveling process is completed, the closed-loop main control unit outputs the final tilt angle matrix and the final stroke encoding matrix, thus completing the entire leveling process.

[0123] S6. Verification of Correction Effect and Iterative Update of Dual Benchmarks: This step, based on the dual benchmark system and the final dip matrix and final travel coding matrix, completes the full-dimensional verification of the correction effect and the closed-loop iterative update of the dual benchmark system. At the same time, the final result is fed back to the three types of algorithms to complete the full life cycle iterative optimization of the algorithm parameters. Finally, the closed-loop iterative command is output and returned to S2, forming a closed-loop monitoring and correction system without any breaks in the entire process, ensuring the levelness accuracy and observation data quality of the seismograph throughout its entire life cycle.

[0124] The closed-loop main control unit first performs a full-dimensional verification of the correction effect, which is divided into two levels of verification:

[0125] The first stage is for calibration accuracy verification. The closed-loop main control unit calculates the residual levelness deviation after calibration. The residual levelness deviation is the difference between the final tilt angle matrix and the initial tilt angle matrix of the instrument's rigid body. If the absolute value of the residual levelness deviation is less than the preset levelness threshold, the calibration is deemed qualified, and the process proceeds to the stability verification stage. If the absolute value of the residual levelness deviation is greater than or equal to the preset levelness threshold, the calibration is deemed unqualified, a calibration failure alarm is output, and the process returns to S3 to re-perform deviation identification and parameter planning.

[0126] The second stage is long-term stability verification. For qualified calibration results, the closed-loop main control unit continuously collects decoupled tilt angle data for 30 minutes and calculates the standard deviation of the tilt angle data. If the standard deviation is less than or equal to 0.5 arcseconds, the stability is deemed qualified, and the system enters the dual-reference system iterative update stage; if the standard deviation is greater than 0.5 arcseconds, the stability is deemed unqualified, a mechanical loosening alarm is output, and the system returns to S1 to re-calibrate the dual-reference system in situ.

[0127] For results that meet both calibration and stability requirements, the closed-loop main control unit executes a closed-loop iterative update of the dual-benchmark system. During long-term operation, the seismograph's observation pier experiences slow settlement, the rigid bearing platform undergoes creep, and the adaptive leveling unit experiences mechanical wear. These factors all contribute to shifts in the initial benchmark. Iterative updates ensure that the benchmark always matches the actual condition of the equipment, while maintaining the stability of the geoid benchmark and ensuring the traceability of the observation data. The specific update process is as follows: First, the residual deviation of the corrected levelness is calculated. (Take the Euclidean norm of the X and Y components). If If the arcsecond value is reached, it is determined that the instrument's rigid body reference has drifted significantly, and an iterative update of the dual reference system is performed: the final stroke encoding matrix is ​​used as the new initial mechanical reference encoding matrix, replacing... The final tilt angle matrix is ​​used as the new initial tilt angle matrix for the instrument rigid body, replacing... .like For arcseconds, the dual-benchmark system remains unchanged; only the algorithm's operating parameters are updated. In-situ geoid benchmark. Unaffected by the above updates, it is only updated synchronously when GNSS reference update data is received.

[0128] The closed-loop master control unit constructs an updated dual-reference parameter set, which is then stored in a non-volatile storage unit as the sole traceability reference for the next closed-loop monitoring.

[0129] After completing the dual-reference system update, the closed-loop control unit inputs the updated dual-reference parameter set into the three types of algorithms, completing the full lifecycle iterative optimization of the initial parameters of the three types of algorithms to ensure the long-term operational accuracy and stability of the algorithms. After completing all verifications and updates, the closed-loop control unit outputs the closed-loop iteration command, returns to S2, and enters the next levelness closed-loop monitoring and correction cycle, realizing the seismograph's full lifecycle, all-time, uninterrupted levelness closed-loop monitoring and adaptive correction.

[0130] The three types of algorithms form a bidirectional interactive fusion closed loop, and the specific interaction process is as follows:

[0131] The bidirectional interaction between the FOCLMS-CS and MPE-HBRB algorithms: The true horizontal dip matrix output by the FOCLMS-CS algorithm is directly used as the input to the MPE-HBRB algorithm to calculate the multi-scale permutation entropy characteristics, directly determining the feature input and inference results of the MPE-HBRB algorithm; the deviation type identifier output by the MPE-HBRB algorithm is fed back to the FOCLMS-CS algorithm to correct the fractional order and step size factor in real time. When a precursor deformation signal before the leveling is identified, the fractional order and step size factor are reduced to slow down the convergence speed, improve the filtering stability, and avoid canceling effective earthquake precursor signals. When the instrument's own horizontal deviation is identified, the fractional order and step size factor are increased to speed up the convergence speed and improve the real-time performance of interference decoupling.

[0132] The bidirectional interaction between the FOCLMS-CS and LG-LA algorithms: The vibration interference amplitude in the reconstructed interference signal output by the FOCLMS-CS algorithm is directly used as the input of the LG-LA algorithm to correct the convergence step size of the adaptive iteration, which directly determines the iteration convergence speed and the leveling path planning accuracy of the LG-LA algorithm; the rigid body motion acceleration output by the LG-LA algorithm during the leveling process is fed back to the FOCLMS-CS algorithm to correct the initial state variables of the Chen chaotic master system in real time, and to reconstruct the mechanical vibration interference generated during the leveling process in real time, so as to avoid the vibration interference during the leveling process being misjudged as a valid tilt angle signal and improve the decoupling accuracy.

[0133] The MPE-HBRB and LG-LA algorithms interact bidirectionally: the effective horizontal deviation value output by the MPE-HBRB algorithm is directly used as input to the LG-LA algorithm to calculate the target pose matrix of the rigid bearing platform, directly determining the leveling parameter planning result of the LG-LA algorithm; the travel margin matrix output by the LG-LA algorithm is fed back to the MPE-HBRB algorithm to correct the rule weights of the confidence rule base in real time. When the travel margin is small, the rule weight corresponding to the instrument's own horizontal deviation is reduced, the judgment threshold is increased, and frequent leveling is avoided to prevent travel exceeding the limit, thus improving the system's operational safety.

[0134] In some embodiments, for the pre-training of the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm, if the earthquake precursor deformation signal feature database released by the seismic network center is not directly accessible to the public, this embodiment provides a method for constructing a standardized alternative training dataset: the alternative dataset contains four types of sample subsets. The first type is samples without effective bias, which collect dip deviation data of the seismograph under standard horizontal locking conditions and in a full range of operating environments for 72 consecutive hours, with a sample size of no less than 1000 sets; the second type is samples of the instrument's own horizontal bias, which simulate step-like and slope-like horizontal offsets in the range of 0.5 arcseconds to 30 arcseconds using a standard tilt table, and synchronously collect corresponding dip angle data, covering the full temperature range of -40℃ to 60℃. The first category consists of at least 50 samples corresponding to each dip gradient; the second category consists of at least 800 samples of precursory deformation signals, which use artificially simulated slow crustal deformation data with a deformation rate of 0.1 arcseconds / day to 10 arcseconds / day and a deformation duration of 3 days to 180 days, covering the typical low-frequency long-period characteristics of precursory signals; the third category consists of at least 800 samples of mixed interference bias samples, which combine the instrument horizontal bias samples, precursory deformation samples, and typical multi-field interference data from the stations; all samples are simultaneously labeled with corresponding multi-scale permutation entropy feature vectors and type labels, which can replace the database to complete the algorithm pre-training, ensuring that the false negative rate of precursory signals is 0 and the overall false negative rate is less than 1%.

[0135] In some embodiments, the typical values ​​of the initial parameter subsets of the three core algorithms in S1 are predefined as follows: In the chaotic synchronization interference cancellation algorithm of fractional-order LMS adaptive filtering, the fractional order... The initial value is 0.8, and the initial value of the adaptive control gain coefficient is [value missing]. Fractional LMS filter initial step size factor The initial value is 0.01, which is the adjustment coefficient for the fractional derivative term. The initial value is 0.005; in the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm, the embedding dimension is... The value is 3, representing the delay time. The value is 1, and the scale factor is... The value ranges from 1 to 20; in the Lie group Lie algebra adaptive iterative algorithm for nonholonomic constrained rigid body motion, the initial step size of the adaptive iteration is... The value is 0.1, which is the iteration step size adjustment coefficient. The value is fixed at 0.5, and the maximum number of iterations is set to 20.

[0136] In some embodiments, the adaptive update rule for the Chen chaotic system adaptive control gain coefficient in S2 is defined as follows: ,in , The gain update step size is fixed at 0.5. This represents the master-slave system synchronization error at the corresponding moment; when the synchronization error... When 100 consecutive sampling points are less than 0.001, it is determined that the master-slave system has achieved global synchronization, the current adaptive control gain coefficient is locked, and the distortion-free reconstruction of the nonlinear interference signal is completed.

[0137] In some embodiments, the installation and sampling synchronization requirements of the multi-field coupled sensing unit in S2 are as follows: the installation distance between the temperature sensor, the triaxial vibration sensor, the electric field strength sensor, the magnetic field strength sensor and the three-component tilt angle sensing unit shall not exceed 2cm, the sampling frequency of all sensing units shall be completely consistent with that of the three-component tilt angle sensing unit, which is 100Hz, and microsecond-level synchronous sampling shall be achieved through hardware triggering to ensure that the time axis of the interference data and the tilt angle data are completely aligned, thereby improving the accuracy of interference reconstruction and cancellation.

[0138] In some embodiments, the hierarchical structure and inference logic of the three-level hierarchical confidence rule base in S3 are specifically defined as follows: The first level is the feature layer rule base, the input of which is a multi-scale permutation entropy feature vector, which is divided into three feature subsets according to the scale factor: low frequency band (s=11~20), mid frequency band (s=6~10), and high frequency band (s=1~5). Three sets of parallel sub-rule bases are constructed for each subset, with 16 rules in each set. The output is the primary confidence level of the deviation type corresponding to the feature of the corresponding frequency band, realizing channel-specific inference of features at different time scales and avoiding mutual interference between high and low frequency features; The second level is the type layer rule base, the input of which is the three sets of primary confidence levels output by the feature layer. The reliability is assessed by constructing 12 confidence rules, corresponding to the type matching degree of four types of results: no effective bias, instrument self-level bias, earthquake precursor deformation signal, and mixed interference bias. The second-level confidence score of each type of bias is output to realize the mapping from feature inference results to type determination. The third level is the decision layer rule base, which takes the four types of second-level confidence scores output by the type layer as input and constructs eight confidence rules. Priority judgment weights are set for earthquake precursor deformation signals. When the second-level confidence score corresponding to the precursor deformation signal is greater than 0.3, the correction and blocking protection judgment is directly triggered. Finally, the comprehensive confidence score of the four types of results is output, and the result with the highest comprehensive confidence score is selected as the final discrimination result.

[0139] In some embodiments, the cutoff frequency setting of the 8th order Butterworth high-pass filter in S3 is based on the fact that the typical frequencies of earthquake precursor crustal deformation signals are all below 0.1Hz, while the effective signal frequencies of the instrument's own horizontal deviation are all above 0.1Hz. Therefore, setting the cutoff frequency to 0.1Hz can completely filter out the low-frequency earthquake precursor signal components in the mixed interference deviation, retaining only the high-frequency instrument deviation components for correction, thus avoiding the effective precursor signal being incorrectly corrected.

[0140] In some embodiments, the normal vector change rate of the rigid bearing platform in S4 is preset to a threshold of no more than 0.005° per sampling period. Corresponding to a sampling frequency of 100Hz, the deflection angle of the normal vector of the rigid bearing platform in a single sampling period is no more than 0.005°. This can completely avoid the platform from generating torsional deformation and internal stress during the leveling process, and ensure that the seismograph observation accuracy is not affected by the leveling action.

[0141] In some embodiments, the specific method for constructing the Jacobian matrix in S4 is as follows: For the three groups of adaptive leveling execution units evenly distributed in a 120° circle, their installation coordinates on the bottom surface of the rigid bearing platform are defined as follows: First group of units The second unit The third unit ,in The radius of the mounting circumference; the pose change of the rigid bearing platform can be decomposed into a rotation angle about the X-axis. Rotation angle about the Y-axis and the translation in the Z-axis direction. The stroke variation of the three sets of leveling execution units The mapping relationship with pose parameters is as follows:

[0142] ; ; ;

[0143] Based on the above mapping relationship, construct the Jacobian matrix. Establish Lie algebra pose vectors (including The linear mapping relationship between the travel variation of the leveling execution unit and the Jacobian matrix is ​​expressed in the following form:

[0144] Its Moore-Penrose generalized inverse matrix It can be solved by matrix singular value decomposition, and used for the conversion of Lie algebra pose vectors to run-length encoded values ​​during the iteration process, so as to realize the solution of singular leveling parameters.

[0145] In some embodiments, the specific judgment rules for the three types of constraint verification in S4 are as follows: The first type of mechanical travel range constraint verification requires that the calculated target travel code value be within the range of 10% to 90% of the maximum travel of the adaptive leveling execution unit to avoid backlash error and mechanical jamming at the travel boundary; the second type of planar motion without torsion constraint verification requires that the rotation angle of the rigid bearing platform around the Z-axis during the leveling process does not exceed 0.01°; the third type of synchronous motion speed constraint verification requires that the speed difference between any two sets of leveling execution units does not exceed 0.01 mm / s; when all three types of constraints are satisfied, the conflict verification is deemed to have passed.

[0146] In some embodiments, the control period of the dual closed-loop dynamic correction in S5 is consistent with the sampling period, both being 10ms. The stroke following closed loop adopts a PID control algorithm with a proportional coefficient of 5, an integral coefficient of 0.1, and a derivative coefficient of 0.05. The tilt angle accuracy closed loop adopts a fuzzy PID control algorithm, using the real-time tilt angle deviation and deviation change rate as input to adjust the PID parameters in real time, ensuring the speed and stability of the leveling process and avoiding overshoot.

[0147] In some embodiments, the minimum hardware requirements for the adaptive leveling execution unit in S5 are: a stroke resolution of not less than 0.001 mm, a repeatability of not less than 0.002 mm, and a corresponding tilt angle adjustment resolution of not less than 0.05 arcseconds, which can meet the 2 arcsecond levelness control accuracy requirement of the present invention; the measurement accuracy of the three-component tilt angle sensing unit is not less than 0.1 arcseconds, and the sampling frequency is not less than 100 Hz, to ensure the real-time performance and accuracy of tilt angle monitoring.

[0148] In some embodiments, the 30-minute long-term stability verification in S6 is set based on the fact that the typical fluctuation period of disturbances such as temperature drift and ground micro-vibration in the seismic station environment is less than 30 minutes. By continuously collecting tilt data for 30 minutes and calculating the standard deviation, the long-term stability of the instrument's levelness after calibration can be effectively verified. When the standard deviation is less than or equal to 0.5 arcseconds, it is determined that the instrument has no mechanical loosening or continuous deformation offset, and the calibration effect is stable and reliable.

[0149] In some embodiments, the iterative update of the dual reference system in S6 is set with a trigger threshold. The iterative update of the initial mechanical reference and the initial inclination reference of the instrument rigid body is only performed when the corrected residual deviation of the levelness is greater than 1 arcsecond. If the residual deviation is less than or equal to 1 arcsecond, only the algorithm running parameters are updated, and the dual reference system is not modified. This avoids reference drift caused by frequent iterations and ensures the long-term consistency and traceability of the observation data.

[0150] In some embodiments, the chaotic synchronization interference cancellation algorithm of fractional-order LMS adaptive filtering, the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm, and the Lie group Lie algebra adaptive iterative algorithm for nonholonomic constrained rigid body motion are mutually interactively optimized in a bidirectional manner, and the specific quantization rules are as follows:

[0151] (1) Bidirectional interaction between MPE-HBRB algorithm and FOCLMS-CS algorithm: When the MPE-HBRB algorithm identifies a precursory deformation signal, the fractional order of the FOCLMS-CS algorithm... Corrected to 0.5, step size factor The value was adjusted to 0.005 to slow down the convergence speed and avoid canceling out effective low-frequency precursor signals; when the discrimination result is due to the instrument's own horizontal deviation, Revised to 1.2 The value was adjusted to 0.02 to accelerate convergence and improve the real-time performance of interference decoupling; when the discrimination result shows no effective deviation, the parameter was restored to the initial calibration value.

[0152] (2) Bidirectional interaction between LG-LA algorithm and FOCLMS-CS algorithm: When the vibration disturbance amplitude output by FOCLMS-CS algorithm is greater than 0.1g, the initial step size of the adaptive iteration of LG-LA algorithm is corrected to 0.05, reducing the leveling driving speed and reducing the vibration disturbance during the leveling process; the leveling driving acceleration output by LG-LA algorithm is fed back to FOCLMS-CS algorithm in real time, and the state variables of Chen chaotic master system are updated synchronously, so as to include the vibration disturbance corresponding to the driving acceleration in the reconstruction range and avoid the leveling vibration being misjudged as tilt signal;

[0153] (3) Two-way interaction between LG-LA algorithm and MPE-HBRB algorithm: When the travel margin output by LG-LA algorithm is less than 20% of the maximum travel, the rule weight corresponding to the instrument's own horizontal deviation in MPE-HBRB algorithm is reduced by 30%, and the deviation judgment threshold is increased from 0.5 arcseconds to 2 arcseconds. The correction enable command is only output when the deviation exceeds 2 arcseconds to avoid frequent leveling leading to travel over-limit. When the travel margin is greater than 50% of the maximum travel, the rule weight and judgment threshold are restored to the initial calibration value.

[0154] In some embodiments, each adaptive leveling execution unit is a closed-loop vector-controlled servo electric cylinder. The cylinder body is rigidly connected to the observation pier via a spherical hinge, and the telescopic rod is connected to the rigid bearing platform via a spherical hinge. The electric cylinder has a built-in multi-turn absolute encoder, so the position information is not lost after power failure. The movement of the electric cylinder is driven by pulse commands received by the servo driver, with a stroke resolution of not less than 0.001 mm and a repeatability of not less than 0.002 mm. The three sets of electric cylinders are evenly distributed in a 120° circle with the geometric center of the rigid bearing platform as the center, and the installation circumference radius R is determined according to the size of the bearing platform.

[0155] In some embodiments, the three-component tilt sensing unit is a three-axis digital tilt sensor based on a MEMS capacitive accelerometer. The X-axis and Y-axis coincide with the geographic north and geographic east directions, respectively, and the Z-axis coincides with the geometric normal axis of the rigid bearing platform. It has a built-in 24-bit Δ-Σ analog-to-digital converter, with a single-axis measurement accuracy of not less than 0.05 arcseconds and an orthogonal axis coupling error of less than 0.1%. The sensor integrates a third-order low-pass filter circuit and a temperature compensation module, with zero drift of less than 0.2 arcseconds across the entire temperature range (-40℃ to +60℃). It is connected to the closed-loop main control unit via a shielded twisted-pair SPI interface with a sampling frequency of 100Hz and a hardware trigger synchronization accuracy better than 1 microsecond.

[0156] In some embodiments, the closed-loop main control unit is an embedded controller based on an ARM Cortex-R series real-time processor and an FPGA heterogeneous computing architecture; wherein the ARM core runs a Linux real-time operating system and is responsible for algorithm scheduling, network communication and data management; the FPGA logic unit is responsible for high-speed data sampling (100Hz three-channel synchronous sampling), PWM pulse synchronous driving (three-channel error less than 0.1 microseconds) and hardware-level chaotic synchronous calculation; it has a built-in non-volatile FRAM storage unit to store dual-reference system data; it maintains communication with the remote monitoring platform through a 4G / 5G wireless communication module and supports NTP network time synchronization.

[0157] In some embodiments, the multi-field coupling sensing unit is integrated on a multilayer PCB circuit board, with a center distance of no more than 2 cm from the three-component tilt sensing unit. Microsecond-level synchronous sampling is achieved through the same hardware trigger signal. The unit includes: a PT1000 platinum resistance temperature sensor (measurement range -55℃ to +125℃, accuracy ±0.1℃), a triaxial MEMS vibration sensor (±2g range, noise density 25μg / √Hz), and a three-dimensional electromagnetic field sensor (electric field measurement range 0.1V / m to 100V / m, magnetic field range 0.1μT to 100μT, resolution 0.01μT). All sensor outputs are converted into digital signals by a 16-bit synchronous sampling ADC and transmitted to the closed-loop main control unit FPGA module via the SPI bus, with a sampling frequency of 100Hz ±0.01%. When using a standardized alternative dataset for pre-training, the multi-field coupling sensing unit can be downgraded to a combination of a temperature sensor and a single-axis vibration sensor to adapt to station environments lacking a complete electromagnetic field sensor.

[0158] In some embodiments, the reference temperature recorded in S1 This is used to compensate for temperature drift in the output of the three-component tilt sensor unit. Specifically, during the S2 real-time decoupling process, the closed-loop main control unit synchronously acquires the current ambient temperature. Calculate temperature deviation According to the temperature drift coefficient specified by the sensor unit at the factory. (Unit: arcseconds / °C), perform pre-compensation on the original tilt angle data: After compensation This is then used as input to the FOCLMS-CS algorithm. Temperature drift coefficient. The initial parameters of the algorithm are stored in a subset of non-volatile memory units, with a typical value of 0.05 arcseconds / °C.

[0159] In some embodiments, the geoid in-situ reference matrix As an absolute geographical level reference, it is used for long-term stability verification and benchmark drift detection of the subsequent dual-benchmark system. Specifically, after each calibration effect verification is successful (S6), the closed-loop main control unit calculates the current instrument rigid body initial tilt angle benchmark. With geoid reference Deviation angle between ,like If three consecutive calibration cycles exceed the preset drift threshold (e.g., 1 arcsecond), it is determined that the observation pier has experienced uneven settlement or the rigid bearing platform has undergone creep, triggering the in-situ recalibration of the dual-reference system in S1. Simultaneously, of The component serves as the absolute reference benchmark for the "geoid deformation signal" in the S3 dip deviation calculation: when the discrimination result is At that time, the actual output precursor deformation signal is , rather than This is to ensure the traceability of precursor signals relative to an absolute geographical reference.

[0160] In some embodiments, the multi-field interference data matrix output by the multi-field coupling sensing unit The specific construction method is as follows: Assume the temperature sensor output is... (Unit: °C), the output vector of the triaxial vibration sensor is (Unit: g), the output of the electric field strength sensor is (Unit: V / m), the magnetic field strength sensor output is (Unit: μT). First, dimensionless normalization is performed on each physical quantity: , , , The normalization range is obtained from the statistical analysis of full-condition interference calibration test data (e.g., ; ; ; Then construct the fusion scalar. The weighting coefficient The typical value is determined through pre-training optimization. .final This serves as the common driving input for the three-dimensional chaotic system. If it is desired to preserve the independent influence of each disturbance component, a three-dimensional driving vector can also be constructed. At this time, the driving term of Chen's chaotic master system The multiplication of a vector with scalar coefficients is performed element-wise.

[0161] In some embodiments, Chen's chaos originates from the system's adaptive control gain coefficient. Instead of using a fixed value, it is updated adaptively in real time based on the synchronization error. The update rule is as follows: ;in The gain update step size is fixed at 0.5 (dimensionless). For the first One synchronization error variable. When there are 100 consecutive sampling points... When all values ​​are less than 0.001, the master-slave system is considered to have achieved global synchronization, and the current value is locked. If the value is not specified, updates will stop. This rule ensures that the chaotic synchronization system can converge quickly under different disturbance intensities and has small steady-state fluctuations.

[0162] In some embodiments, to avoid due to and The distortion of the effective level deviation value caused by the sum not being equal to 1 is corrected by normalizing the calculation formula as follows:

[0163] ;

[0164] in It is a very small positive number (e.g.) (), used to avoid the denominator being zero. When hour, This meets the expectation of no effective deviation. This normalization ensures that the effective level deviation output is dimensionlessly correct and its amplitude matches the actual physical deviation when the correction is enabled.

[0165] In some embodiments, when a step loss is detected and the leveling operation is terminated, the closed-loop main control unit does not directly output a permanent alarm, but instead executes an automatic recovery process: First, all adaptive leveling execution units are slowly returned to the midpoint of the mechanical stroke (the speed is reduced to 20% of the normal leveling speed), and the stroke code value is reread and compared with the initial mechanical reference. If the deviation is less than twice the travel resolution, the system is determined not to have mechanically jammed and automatically returns to S2 to re-identify the deviation and re-plan the balancing parameters, with a maximum of 3 retries. If execution still fails to synchronize after 3 retries, an "Unrecoverable execution failure" alarm is output and reported to the remote monitoring platform, requesting manual maintenance.

[0166] In some embodiments, the "dispersion of the computational data" specifically refers to the computation of continuously acquired data. Decoupled backslope angle data ( The sample standard deviation of )

[0167] , ;

[0168] in The mean is taken. This serves as the final indicator of dispersion. The preset stability threshold is 0.5 arcseconds, i.e. The stability is deemed acceptable at an arcsecond interval.

Claims

1. A closed-loop monitoring and adaptive correction method for the levelness of a seismograph, characterized in that, Includes the following steps: S1. Construct the initial mechanical reference, the initial inclination reference of the instrument rigid body, and the in-situ reference of the geoid for seismic instrument level monitoring, and solidify them into a dual reference system; S2. Synchronously acquire the raw horizontal dip angle data of the seismograph, and use the chaotic synchronization interference cancellation algorithm of fractional LMS adaptive filtering to decouple the raw horizontal dip angle data to obtain the true horizontal dip angle data; S3. Calculate horizontal tilt deviation data based on real horizontal tilt data and dual reference system. Use multi-scale permutation entropy-hierarchical confidence rule base inference algorithm to extract features and identify deviation types of horizontal tilt deviation data to obtain identification results. If the identification result is the instrument's own horizontal deviation or mixed interference deviation, obtain the correction enable command and effective horizontal deviation data. S4. After receiving the correction enable command, based on the effective horizontal deviation data and the dual reference system, the Lie group Lie algebra adaptive iterative algorithm for nonholonomic constrained rigid body motion is adopted to complete the singularity-free solution and conflict verification of the leveling parameters, and obtain the leveling execution parameters. S5. Based on the leveling execution parameters, execute the seismograph synchronous leveling drive, acquire real-time horizontal dip data during the leveling process, synchronously collect real-time travel data, perform dual closed-loop dynamic correction based on real-time horizontal dip data and real-time travel data, and obtain the final horizontal dip data and final travel data after leveling is completed. S6. Compare the final horizontal tilt data and final travel data with the dual reference system, perform a full-dimensional verification of the correction effect, and after the verification is qualified, perform a closed-loop iterative update of the dual reference system. After the update is completed, return to S2 and enter the next levelness closed-loop monitoring and correction cycle.

2. The closed-loop monitoring and adaptive correction method for seismograph levelness according to claim 1, characterized in that, The construction of the dual-reference system in S1 is as follows: the rigid bearing platform of the seismograph is rigidly connected to the observation pier through three sets of adaptive leveling actuators evenly distributed in a 120° circle. All adaptive leveling actuators are controlled to move to the midpoint of the mechanical stroke and locked. The stroke code value of each adaptive leveling actuator is read to construct the initial mechanical reference. Multiple sets of initial tilt angle data of the rigid bearing platform in the locked state are collected and constructed into the initial tilt angle reference of the instrument rigid body after smoothing and filtering. Obtain the geoid normal reference data at the location of the observation pier, and combine it with the rigid deformation parameters of the observation pier during the completion and acceptance to construct the in-situ geoid reference. The three types of references are permanently stored in non-volatile memory cells to form a dual-reference system.

3. The closed-loop monitoring and adaptive correction method for the levelness of a seismograph according to claim 1, characterized in that, The specific processing steps of the chaotic synchronization interference cancellation algorithm in the fractional-order LMS adaptive filtering in S2 are as follows: Multi-field interference data from the same environment as the original horizontal tilt data are synchronously acquired. The nonlinear characteristics of the multi-field interference data are fitted using the Chen chaotic master system. A chaotic slave system with adaptive control terms is constructed to achieve global synchronization between the master and slave systems, completing the distortion-free reconstruction of the nonlinear interference signal. Based on the definition of Caputo fractional calculus, a fractional-order LMS adaptive filtering framework is constructed. Using the reconstructed interference signal as a reference, adaptive interference cancellation is performed on the original horizontal tilt data, eliminating multi-field coupled interference components to obtain distortion-free true horizontal tilt data.

4. The closed-loop monitoring and adaptive correction method for the levelness of a seismograph according to claim 1, characterized in that, The feature extraction process of the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm in S3 is as follows: multi-scale coarse-graining processing is performed on the horizontal tilt deviation data to obtain coarse-grained sequences at different time scales; phase space reconstruction is performed on the coarse-grained sequences at each scale to mine the nonlinear dynamic features within the sequence; The probability of each sequence's permutation pattern after statistical reconstruction is calculated, the permutation entropy value corresponding to each scale is calculated, and all permutation entropy values ​​are normalized to obtain a multi-scale permutation entropy feature vector for identifying deviation types.

5. The closed-loop monitoring and adaptive correction method for seismograph levelness according to claim 4, characterized in that, The specific process for identifying the bias type in the S3 multi-scale permutation entropy-hierarchical confidence rule base inference algorithm is as follows: The extracted multi-scale permutation entropy feature vector is input into the three-level hierarchical confidence rule base, the activation weight of each rule is calculated, and all activation rules are fused and calculated through the evidence reasoning algorithm to obtain the comprehensive confidence level corresponding to four types of results: no effective bias, instrument self-level bias, earthquake precursor deformation signal, and mixed interference bias. The result with the highest comprehensive confidence level is selected as the final identification result.

6. The closed-loop monitoring and adaptive correction method for the levelness of a seismograph according to claim 1, characterized in that, The solution process of the Lie group and Lie algebra adaptive iterative algorithm for nonholonomic constrained rigid body motion in S4 is as follows: The three-dimensional rigid body motion of the rigid bearing platform is represented by a special Euclidean group, and the rotational motion of the rigid bearing platform is represented by a special orthogonal group. Based on the effective horizontal deviation data and the initial mechanical reference, the initial pose matrix and the target pose matrix of the rigid bearing platform with absolute horizontality are constructed. The deviation matrices of the initial pose and the target pose are mapped to the Lie algebra space through logarithmic mapping, transforming the nonlinear pose optimization problem into an iterative solution problem in linear space. Under the preset nonholonomic constraints, the optimal leveling execution parameters are obtained through an adaptive iterative algorithm.

7. The closed-loop monitoring and adaptive correction method for the levelness of a seismograph according to claim 1, characterized in that, The conflict verification of the leveling parameters in S4 is as follows: For the leveling execution parameters obtained by calculation, three types of constraint verification are performed in sequence. The first type is the mechanical stroke range constraint verification of the adaptive leveling execution unit, the second type is the planar motion non-torsional constraint verification of the rigid bearing platform, and the third type is the synchronous motion speed constraint verification of multiple sets of adaptive leveling execution units. When all constraint conditions are met, the verification is deemed to have passed, and the leveling execution parameters are obtained.

8. The closed-loop monitoring and adaptive correction method for the levelness of a seismograph according to claim 1, characterized in that, The fractional-order LMS adaptive filtering chaotic synchronization interference cancellation algorithm, the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm, and the Lie group Lie algebra adaptive iterative algorithm for nonholonomic constrained rigid body motion are mutually interactive and collaboratively optimized in pairs. Specifically, the deviation identification results are fed back to the fractional-order LMS adaptive filtering chaotic synchronization interference cancellation algorithm in real time to correct the algorithm's running parameters; the leveling execution status is fed back to the fractional-order LMS adaptive filtering chaotic synchronization interference cancellation algorithm in real time to correct the feature parameters of interference reconstruction; and the leveling travel margin data is fed back to the multi-scale permutation entropy-hierarchical confidence rule base inference algorithm in real time to correct the rule weights of deviation identification.

9. The closed-loop monitoring and adaptive correction method for the levelness of a seismograph according to claim 1, characterized in that, The execution process of the dual closed-loop dynamic correction in S5 is as follows: During the entire leveling drive execution, interference decoupling processing is performed synchronously, and real-time horizontal tilt angle data after decoupling is acquired in real time. At the same time, real-time travel data fed back by each adaptive leveling execution unit is collected. Based on the deviation between the real-time travel data and the target travel corresponding to the leveling execution parameters, travel following closed-loop correction is performed. Based on the deviation between the real-time horizontal tilt angle data and the instrument rigid body initial tilt angle reference, tilt angle accuracy closed-loop correction is performed. The two closed loops run synchronously, performing and correcting simultaneously, until the travel following deviation and tilt angle deviation both meet the preset requirements, at which point leveling is determined to be complete.

10. The closed-loop monitoring and adaptive correction method for the levelness of a seismograph according to claim 1, characterized in that, The full-dimensional verification of the correction effect in S6 is as follows: First, the correction accuracy verification is performed, and the residual deviation between the final horizontal tilt angle data and the initial tilt angle reference of the instrument rigid body is compared. If the residual deviation is less than the preset levelness threshold, the accuracy verification is deemed qualified, and the long-term stability verification is then performed. Long-term stability verification involves continuously collecting decoupled horizontal tilt angle data for a set duration, calculating the dispersion of the data, and determining that the verification is qualified if the dispersion is less than the preset stability threshold.

Images