A digital leveling compensation system and method for a relative gravimeter
By combining deep echo state networks and the eagle search optimization algorithm, a multi-source sensor data modeling system was constructed, which solved the problem of inaccurate disturbance prediction of gravimeters in complex environments, realized high-precision leveling control and adaptive compensation, and improved the accuracy and stability of measurements.
Patent Information
- Application Number
- CN202511594738.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-03
- Publication Date
- 2026-11-17
AI Technical Summary
Existing digital leveling control methods for gravimeters are unable to accurately predict future disturbance trends in complex terrain or strong disturbance environments. They also lack sufficient model adjustment capabilities and adaptive adjustment mechanisms, leading to overcompensation or response lag, which affects measurement accuracy.
A multi-source sensor data modeling and prediction system is constructed by combining a deep echo state network with a vulture search optimization algorithm. Through multi-layer behavioral structure, dynamic switching of activation functions, multi-channel fusion and deep sparse adjustment, high-precision prediction of tilt angle disturbance trends is achieved, and a closed-loop leveling self-learning control mechanism is constructed.
It significantly improves the leveling control accuracy and adaptability of the gravimeter in complex environments, has rapid response and self-adaptation capabilities, reduces compensation errors, and improves the accuracy and stability of measurements.
Smart Images

Figure CN123113305A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of precision measurement and inertial navigation control technology, and in particular to a digital leveling compensation system and method for a relative gravimeter. Background Technology
[0002] In precision measurement fields such as gravity measurement, geophysical exploration, and high-precision navigation control, relative gravimeters are widely used as core equipment in critical scenarios such as oil exploration, mineral surveys, geodesy, and geodynamic monitoring. To ensure the accuracy and stability of gravity measurement data, relative gravimeters are typically equipped with a digital leveling control system to maintain the instrument's horizontal attitude in real time, eliminating measurement errors caused by instrument tilt. In practical applications, factors such as ground vibration, environmental interference, and changes in platform attitude can cause slight tilt disturbances in the instrument, thus affecting the accuracy of gravity measurements. Therefore, high-precision real-time compensation for the gravimeter's attitude has become an important research direction.
[0003] In existing technologies, digital leveling control methods generally employ compensation mechanisms based on PID controllers or fixed models, relying solely on current or limited historical attitude data to calculate leveling commands. This approach ignores the dynamic evolution of sensor data, making it difficult to accurately predict future disturbance trends. Furthermore, the model's adjustment capabilities are insufficient, easily leading to overcompensation or response lag issues in complex terrain or environments with strong disturbances. Simultaneously, traditional methods oversimplify sensor data processing, failing to fully exploit the nonlinear coupling relationships between multi-source data, thus limiting the system's ability to model complex disturbances. Moreover, existing leveling control systems often employ fixed structural parameters, lacking feedback-based adaptive adjustment mechanisms, and are unable to dynamically optimize the leveling strategy according to environmental changes.
[0004] Therefore, how to provide a digital leveling compensation system and method for a relative gravimeter is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a digital leveling compensation system and method for a relative gravimeter. This invention fully integrates neural dynamic modeling, intelligent evolutionary optimization, and inertial leveling control technologies. It describes in detail the complete process of modeling and predicting multi-source sensor data, disturbance compensation, error feedback, and adaptive model updating under the working state of the gravimeter. It has the advantages of fast compensation response, accurate disturbance prediction, strong self-learning ability, and adaptability to complex environments.
[0006] A digital leveling compensation method for a relative gravimeter according to an embodiment of the present invention includes the following steps: S1. Collect multi-source sensor data of the relative gravimeter in working condition, preprocess the multi-source sensor data, and construct the input sample set; S2. Construct a deep echo state network model, receive the generated input sample set, and perform modeling and prediction of tilt perturbation trends; S3. The vulture search optimization algorithm is introduced to optimize the structural parameters of the deep echo state network model. The vulture search algorithm is used to perform parameter combination search and construct a fitness function to perform search update. S4. Apply the optimized deep echo state network model to the input sensor data at the current moment and in the past, and output the predicted tilt perturbation trend value in the future period. S5. Calculate the leveling compensation command based on the tilt angle disturbance trend value, send it to the digital leveling controller, and drive the actuator to complete the leveling compensation operation. S6. Collect attitude feedback data after the digital leveling controller is executed, compare it with the tilt angle disturbance trend value, construct error feedback samples, and update the output connection weights of the deep echo state network model. S7. Update the search strategy of the Bald Eagle Search Optimization Algorithm based on the error feedback samples, make local adjustments and iterative optimizations to the structural parameter search space, and form a closed-loop leveling self-learning control mechanism in the process of multiple leveling compensations.
[0007] Optionally, the multi-source sensor data specifically includes tilt sensor data, ground vibration data, ambient temperature and air pressure data, target value of actual tilt disturbance trend, power status data, and leveling controller feedback data.
[0008] Optionally, the preprocessing of multi-source sensor data specifically includes timestamp alignment, missing value imputation, outlier removal, normalization, and format standardization.
[0009] Optionally, S2 specifically includes: S21. Set the input sample set to a standardized multi-source sensor data sequence. ,in, Indicates the number of sensor channels. Indicates the number of time steps. For the first The sensor data vector at time t, where R is the set of real numbers; S22. Divide the input sample set according to channels. There are 1 input sub-channels, denoted as follows: ,in And define the multi-channel fusion weight vector as ; S23, Calculate the first Input vector of the layered reservoir ; S24. The total number of layers in the deep echo state network is: Each layer of the storage pool contains There are 1 state node, and the initial state is a zero vector. ; S25. Set leakage rate parameters for each layer of the reservoir. And set the sparsity parameter. Control state transition matrix The proportion of non-zero connections; S26. Initialize the input weight matrix for each layer. With the state transition matrix The input weight matrix and state transition matrix are randomly generated using a Gaussian distribution and based on the sparsity parameter. Apply a sparse mask; S27. Specify the activation function type for each reservoir layer, and select parameters through the activation function. The activation function used is controlled, and the mapping relationship is defined as follows: Use Tanh; Use ReLU; Use ELU; : Use Swish; S28. Define the state update for each layer as follows: ; S29. Perform time-step processing on the state vectors of all layers. Concatenate the vectors to construct the network's global state output vector. ; S210. Construct a connection weight matrix and bias terms for the output, wherein the connection weight matrix receives the network global state output vector. As input, it is mapped to the tilt angle disturbance trend value according to the predetermined dimension. The bias term is used to perform linear offset correction on the tilt angle disturbance trend value to generate the output at the corresponding time. S211. Compare the tilt angle perturbation trend value with the corresponding true tilt angle perturbation trend target value, and conduct supervised training by constructing a loss function of the mean square error type. During the training process, adjust the output connection weight matrix and bias term parameters, while keeping the structure and state weights of the other layers of the reservoir unchanged.
[0010] Optionally, S3 specifically includes: S31. Construct a multi-layered behavioral structure for the bald eagle search optimization algorithm. Divide the initial optimization population into three sub-populations: a reconnaissance layer, a tracking layer, and a besieging layer. Each individual is encoded as a set of structural parameter combinations of a deep echo state network model, including: activation function selection parameters. Multi-channel fusion weight vector Skip Link Index Matrix sparsity parameter ; S32. Initialize the vulture algorithm population position vector and historical trajectory matrix. Set the maximum number of iteration rounds. Behavior switching threshold Local research probability and trajectory fusion coefficient ; S33. Constructing a multi-objective fitness function Simultaneously considering prediction error, sparsity of skip connections, and balance of fusion channels; S34. Calculate the fitness function value for all individuals, record the current best individual, and update the historical trajectory matrix. The optimal structural parameter solution for each round is retained; S35. Individuals in the reconnaissance layer execute a large-scale jump search strategy to expand the search space coverage. They adjust the jump amplitude and direction in combination with the current energy function. The current energy function is a search guidance function dynamically constructed based on the population's optimal fitness value, the individual's current fitness value, and the global average fitness value. S36, Tracking layer individuals are based on the optimal individual Current individual Trajectory memory vector Update location ; S37. Individuals in the siege layer perform small-step perturbation search around the globally optimal individual to finely adjust the local structural configuration of the deep echo state network. S38, ranking top in current fitness The individual triggers a local re-search mechanism, which further optimizes the combination of structural parameters based on micro-amplitude perturbations. Percentage threshold; S39. Connect the skip link index matrix With multi-channel fusion weight vector The network is constructed as a graph structure, and the nested features of the graph structure are used to adaptively constrain the network connection dependencies. S310. Combining graph structure and fitness function, perform joint structure-performance evaluation on the combination of structural parameters and update the individual fitness ranking. S311. Each behavior layer updates the search path based on the strategy weight and local experience value, and enters the next round of position iteration or switches the behavior layer state. S312. When the maximum number of iterations or the convergence threshold is reached, output the current optimal combination of structure parameters, including the optimal... , , , It is used to configure the deep echo state network model structure.
[0011] Optionally, S4 specifically includes: S41. The multi-source sensor data vector at the current moment and the multi-source sensor data from the k consecutive historical moments are combined to form an input sample sequence, which is then used as part of the input sample set and input into the deep echo state network model optimized by the Bald Eagle Search optimization algorithm. S42. Using the optimized deep echo state network model, multi-layer state propagation and update operations are performed sequentially on the input sample sequence. In each network layer, the state of the reserve pool node is updated under the combined action of the current input, the previous state, and the jump connection state. S43. A dynamic switching mechanism of activation function is adopted to complete the nonlinear transformation of the state of the reserve pool, and a multi-input channel fusion structure is introduced in each network layer. The feature weighted fusion of K input sub-channels is performed according to the multi-channel fusion weight vector. S44. Construct a jump connection matrix in each layer of the reserve pool, and guide the node state to receive the information flow of the previous time and the specific historical node state through the jump connection mechanism. S45. Introduce a deep sparse adjustment mechanism into the network structure, and use sparse masks to perform structured pruning operations on the input weight matrix and state transition matrix to control the network complexity. S46. Merge the final state outputs of all network layers to construct the global state output vector of the network. Perform a linear transformation on the global state output vector and the output connection weights to generate a tilt perturbation trend value vector, which represents the tilt perturbation trend value at each time point in the future.
[0012] Optionally, S5 specifically includes: S51. Based on the tilt perturbation trend value vector output by the optimized deep echo state network model, and combined with the current attitude state quantity and historical offset response, construct a sliding prediction window. S52. Based on the tilt perturbation trend value within the sliding prediction window, construct a compensation prediction structure that integrates multiple time windows, and extract the short-term window values respectively. With the mid-term window The perturbation information is obtained through the state uncertainty modulation factor. Comprehensive calculation of target tilt angle compensation adjustment amount ; S53. Constructing a dynamic control gain function Adjust the compensation sensitivity of the leveling response along different axes and calculate the three-dimensional leveling command vector. Anisotropic gain coefficients are defined based on the physical structure and inertial characteristics of the equipment, where, , and These represent the dynamic control gain function values for the X, Y, and Z axes at the current time t, respectively. S54, Transfer the three-dimensional leveling command vector Feedback status with the current digital leveling controller The combined input is processed according to the nonlinear control mapping relationship to construct a control command vector for driving the digital leveling controller. The control command vector is then subjected to amplitude compression processing. S55. Output the control command vector to the digital leveling controller, which drives the three-axis actuator to complete the tilt angle adjustment and compensation operation, thereby realizing real-time digital leveling based on the predicted compensation amount.
[0013] Optionally, S6 specifically includes: S61. Collect the attitude feedback data after the digital leveling controller is executed, and use it as the actual attitude response information at the current moment; S62. Extract the tilt angle disturbance trend value corresponding to the attitude feedback data, and construct a one-to-one correspondence between the tilt angle disturbance trend value and the attitude feedback data; S63. Compare the tilt angle disturbance trend value with the corresponding attitude feedback data to generate an error sequence, representing the prediction deviation at each time step; S64. Construct error feedback samples based on the error sequence, including the tilt perturbation trend value, attitude feedback data and error value at each time step; S65. Input the error feedback samples into the deep echo state network model and perform incremental updates on the output connection weights based on the error feedback mechanism. S66. By adjusting the updated output connection weights, the network's ability to predict tilt perturbation trends is improved, thereby enhancing the accuracy of its response to future perturbation changes.
[0014] A digital leveling and compensation system for a relative gravimeter according to an embodiment of the present invention includes the following modules: The data acquisition and preprocessing module is used to acquire multi-source sensor data and preprocess it to construct an input sample set. The Deep Echo State Network Building Module is used to build deep echo state network models. The structural parameter optimization module is used to introduce the vulture search optimization algorithm to search and optimize the structural parameters of the deep echo state network model; The tilt disturbance trend module is used to input current and historical sensor data and output tilt disturbance trend values; The leveling instruction generation module is used to construct a disturbance weight function and modulation coefficient based on the tilt angle disturbance trend value, generate a leveling instruction vector, and perform compensation operations. The feedback error construction module is used to collect attitude feedback data, compare it with the tilt angle disturbance trend value, and generate error feedback samples. The model and optimization update module is used to update the output connection weights based on error feedback samples and simultaneously adjust the local search strategy of the vulture search optimization algorithm.
[0015] The beneficial effects of this invention are: The relative gravimeter digital leveling compensation method proposed in this invention effectively solves the technical bottlenecks of traditional methods, such as low disturbance prediction accuracy, lag in control response, and lack of model adaptability. This invention employs a deep echo state network to dynamically model multi-source sensor data collected by the gravimeter during operation. Combined with improved structures such as dynamic switching of activation functions, multi-input channel fusion, skip-connectivity in the reservoir, and deep sparsity adjustment, it significantly improves the prediction accuracy and stability of tilt disturbance trends. Simultaneously, based on an improved vulture search optimization algorithm, it collaboratively searches and optimizes network structure parameters, introducing hierarchical population behavior mapping, memory-based trajectory guidance, and local graph structure re-search mechanisms to achieve efficient adaptive adjustment of structural parameters and dynamic enhancement of model performance.
[0016] The system constructs a complete error feedback mechanism. By collecting attitude feedback data and comparing it with disturbance prediction results, it continuously corrects the network output connection weights and adjusts the optimization strategy in reverse, forming a closed-loop leveling control process. This method possesses stronger disturbance prediction capabilities, leveling compensation accuracy, and environmental adaptability. Even in extreme application scenarios such as strong interference and complex terrain, it maintains stable and reliable leveling control performance, significantly improving the accuracy, safety, and efficiency of gravity measurement operations. This method is widely applicable to various scenarios such as oil exploration, mineral exploration, geodesy, and geophysical research, demonstrating good engineering feasibility and application value. Attached Figure Description
[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a digital leveling compensation method for a relative gravimeter proposed in this invention; Figure 2 This is a schematic diagram of the structure of a digital leveling and compensation system for a relative gravimeter proposed in this invention. Detailed Implementation
[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0019] refer to Figure 1 A digital leveling compensation method for a relative gravimeter includes the following steps: S1. Collect multi-source sensor data of the relative gravimeter in working condition, preprocess the multi-source sensor data, and construct the input sample set; S2. Construct a deep echo state network model, receive the generated input sample set, and perform modeling and prediction of tilt perturbation trends; S3. The vulture search optimization algorithm is introduced to optimize the structural parameters of the deep echo state network model. The vulture search algorithm is used to perform parameter combination search and construct a fitness function to perform search update. S4. Apply the optimized deep echo state network model to the input sensor data at the current moment and in the past, and output the predicted tilt perturbation trend value in the future period. S5. Calculate the leveling compensation command based on the tilt angle disturbance trend value, send it to the digital leveling controller, and drive the actuator to complete the leveling compensation operation. S6. Collect attitude feedback data after the digital leveling controller is executed, compare it with the tilt angle disturbance trend value, construct error feedback samples, and update the output connection weights of the deep echo state network model. S7. Update the search strategy of the Bald Eagle Search Optimization Algorithm based on the error feedback samples, make local adjustments and iterative optimizations to the structural parameter search space, and form a closed-loop leveling self-learning control mechanism in the process of multiple leveling compensations.
[0020] This invention, by constructing a deep echo state network model and introducing the vulture search optimization algorithm, achieves high-precision modeling and prediction of tilt disturbance trends under complex working conditions by a relative gravimeter, significantly improving the intelligence level of the leveling control system. This method fully utilizes multi-source sensor data, constructs high-quality input samples through preprocessing, and combines an improved deep neural structure to achieve dynamic response to disturbances. The vulture search optimization algorithm is used to intelligently optimize the model structure parameters, giving the network stronger adaptability and stability. Leveling compensation commands are generated based on predictions, and error correction is performed in conjunction with attitude feedback, constructing a complete error feedback mechanism and a model output connection weight update mechanism, effectively reducing compensation errors. Furthermore, through dynamic updates of the optimization algorithm strategy, a closed-loop self-learning process for leveling control is established, enhancing the system's robustness and generalization ability in complex scenarios. This invention effectively overcomes the problems of slow response, rigid parameters, and weak prediction capabilities of traditional methods, possessing advantages such as accurate prediction, timely compensation, strong adaptability, and wide applicability. It is suitable for various high-precision gravity measurement scenarios such as oil exploration, geodesy, national defense, and geophysics.
[0021] In this embodiment, the multi-source sensor data specifically includes tilt sensor data, ground vibration data, ambient temperature and air pressure data, target value of actual tilt disturbance trend, power status data, and leveling controller feedback data.
[0022] In this embodiment, the preprocessing of multi-source sensor data specifically includes timestamp alignment, missing value imputation, outlier removal, normalization, and format standardization. Specifically, timestamp alignment, missing value imputation, outlier removal, normalization, and format standardization refer to aligning the timestamps of data from different sensors according to a unified time base to ensure comparability of data from each sensor at the same time; then, imputing missing values in the data using linear interpolation; next, removing outliers from invalid data that significantly deviates from the normal range to eliminate erroneous or interfering data; subsequently, normalizing data of different dimensions to unify the data scale range and enhance the model's sensitivity to features; finally, format standardization is performed according to the requirements of the network input structure, including channel dimension arrangement and data structure reorganization, to construct standardized input samples that meet the processing requirements of deep learning models.
[0023] In this embodiment, S2 specifically includes: S21. Set the input sample set to a standardized multi-source sensor data sequence. ,in, Indicates the number of sensor channels. Indicates the number of time steps. For the first The sensor data vector at time t, where R is the set of real numbers; S22. Divide the input sample set according to channels. There are 1 input sub-channels, denoted as follows: ,in And define the multi-channel fusion weight vector as ; S23, Calculate the first Input vector of the layered reservoir :
[0024] ; in, For skip join indexes, indicate whether skip joins from the first... The output of one layer is used as part of the input of the current layer. This represents the fusion ratio of the k-th channel input in the l-th layer. This indicates whether layer l receives skip connection input from layer p. The output vector of the state of the p-th layer of the reservoir at time t; Input vector This demonstrates how the Deep Echo State Network (DENN) constructs input information at different levels. For the first layer, the input comes entirely from the original multi-channel data from the sensor. Weighted summations of the data from each channel are achieved using defined channel fusion weights, forming a unified input vector. This design ensures that the original sensor information directly drives the network's dynamic evolution, facilitating the capture of changes in lower-level features. From the second layer onwards, the input no longer directly relies on sensor data but instead fuses the output states of all previous layers. These historical output states are selectively introduced through skip connections and fusion weights to construct new input vectors. This mechanism enhances the model's context awareness and information retention capabilities, making it particularly suitable for modeling long-term dependencies and complex dynamic perturbations. Through this hierarchical input structure design, the feature extraction process from the perceptual layer to the semantic layer is realized, ensuring the complete preservation of the original input information and the deep fusion of multi-level features, providing richer and more robust input support for dynamic prediction and compensation.
[0025] S24. The total number of layers in the deep echo state network is: Each layer of the storage pool contains There are 1 state node, and the initial state is a zero vector. ; S25. Set leakage rate parameters for each layer of the reservoir. And set the sparsity parameter. Control state transition matrix The proportion of non-zero connections; S26. Initialize the input weight matrix for each layer. With the state transition matrix The input weight matrix and state transition matrix are randomly generated using a Gaussian distribution and based on the sparsity parameter. Apply a sparse mask; S27. Specify the activation function type for each reservoir layer, and select parameters through the activation function. The activation function used is controlled, and the mapping relationship is defined as follows: Use Tanh; Use ReLU; Use ELU; : Use Swish; S28. Define the state update for each layer as follows: : ; in, To represent the state vector of the l-th layer at the previous time step, () indicates that parameters are selected based on the activation function. The determined nonlinear activation function; This describes the dynamic update process of the reservoir state in each layer of a deep echo state network, specifically demonstrating the temporal evolution mechanism in which the current state is jointly determined by historical states and the current input. In this update mechanism, the previous state is retained through a decay factor, reflecting the system's ability to remember historical information. The current input, after weight mapping and activation function processing, introduces new dynamic changes to update the current state. By introducing a proportional adjustment factor between the old and new states, the state update rate is controlled, achieving a balance between rapid response to new changes and long-term memory retention. The introduction of a switchable activation function mechanism makes the state update flexible and adaptable in terms of nonlinear mapping, thereby enhancing the model's ability to model different perturbation patterns. This gives the deep echo state network good temporal memory, dynamic response capability, and nonlinear modeling capability, effectively supporting the fine modeling and prediction of tilt perturbation trends, and is a core step in achieving dynamic learning of complex attitude changes.
[0026] S29. Perform time-step processing on the state vectors of all layers. Concatenate the vectors to construct the network's global state output vector. ; S210. Construct a connection weight matrix and bias terms for the output, wherein the connection weight matrix receives the network global state output vector. As input, it is mapped to the tilt angle disturbance trend value according to the predetermined dimension. The bias term is used to perform linear offset correction on the tilt angle disturbance trend value to generate the output at the corresponding time. S211. Compare the tilt angle perturbation trend value with the corresponding true tilt angle perturbation trend target value, and conduct supervised training by constructing a loss function of the mean square error type. During the training process, adjust the output connection weight matrix and bias term parameters, while keeping the structure and state weights of the other layers of the reservoir unchanged.
[0027] This invention achieves high-precision prediction of attitude perturbation trends of relative gravimeters through an improved deep echo state network model. Several innovations in model structure and modeling mechanism significantly enhance the system's modeling capability and generalization performance. Specifically, efficient integration of information from different sensors is achieved through standardized input of multi-source sensor data, a multi-channel fusion mechanism, and adjustable fusion weights. A multi-layer reservoir structure is introduced, combined with a skip-connection mechanism and a state sparsity adjustment strategy, enhancing the model's ability to capture nonlinear and temporal coupling features. A dynamic switching mechanism of multiple activation functions allows the model to select the optimal nonlinear mapping function based on the state distribution at a specific level, improving expressive flexibility and training stability. A global state output vector is constructed by state concatenation, and output mapping weights are used to predict tilt perturbation trends. Training is supervised by a mean squared error loss function, resulting in higher fitting ability and convergence efficiency. The overall model structure has highly adjustable parameters, rich expressive power, controllable computational overhead, and adaptability to prediction needs in complex terrain and perturbation scenarios. As the core modeling component of the digital leveling compensation system, this network provides an accurate prior information basis for compensation command generation and feedback control.
[0028] In this embodiment, S3 specifically includes: S31. Construct a multi-layered behavioral structure for the bald eagle search optimization algorithm. Divide the initial optimization population into three sub-populations: a reconnaissance layer, a tracking layer, and a besieging layer. Each individual is encoded as a set of structural parameter combinations of a deep echo state network model, including: activation function selection parameters. Multi-channel fusion weight vector Skip Link Index Matrix sparsity parameter ; S32. Initialize the vulture algorithm population position vector and historical trajectory matrix. Set the maximum number of iteration rounds. Behavior switching threshold Local research probability and trajectory fusion coefficient ; S33. Constructing a multi-objective fitness function Simultaneously considering prediction error, sparsity of skip connections, and balance of fusion channels: ; in, These are the weighting coefficients for the error term. The sparse penalty weight coefficient, To balance the penalty weights for channel fusion, To monitor the target value, To predict the output, For the number of valid jump connections, The coefficient of variation for channel fusion. The total number of samples, This represents the theoretical maximum number of connections in the skip connection matrix. Multi-objective fitness function This study comprehensively considers the balance between prediction error, sparsity of skip connections, and channel fusion weights to evaluate the quality of deep echo state network structure parameter configuration. The first part measures the network's accuracy in predicting tilt perturbation trends by calculating the sum of squared errors between the true and predicted values, reflecting the model's performance in regression tasks. This is a core indicator in the optimization process, directly determining the effectiveness of the network output. The second part measures the sparsity of skip connections by calculating the difference between the proportion of non-zero connections in the connection matrix and the maximum possible number of connections, ensuring the network structure maintains necessary simplicity and information selectivity, avoiding redundant computation and overfitting risks caused by excessive connections. The third part focuses on the stability of input channel fusion, measuring the fusion balance of each channel through the coefficient of variation of the fusion weights, preventing certain sensor channels from excessively dominating the input, and improving the model's comprehensive perception ability of multi-source information. The fitness function, by introducing three indicators and corresponding weight factors, establishes an optimal balance between accuracy, structural sparsity, and channel fusion, guiding the Bald Eagle Search optimization algorithm to comprehensively evaluate the structural rationality and prediction effect of the network configuration. It is an important evaluation basis for achieving adaptive structure search and performance tuning.
[0029] S34. Calculate the fitness function value for all individuals, record the current best individual, and update the historical trajectory matrix. The optimal structural parameter solution for each round is retained; S35. Individuals in the reconnaissance layer execute a large-scale jump search strategy to expand the search space coverage. They adjust the jump amplitude and direction in combination with the current energy function. The current energy function is a search guidance function dynamically constructed based on the population's optimal fitness value, the individual's current fitness value, and the global average fitness value. S36, Tracking layer individuals are based on the optimal individual Current individual Trajectory memory vector Update location : ; in, Let be the position vector of the i-th individual in the t-th iteration. As a global guiding factor, As a self-preservation factor, For memory feedback factors; This describes the update strategy for individual positions in the Bald Eagle Search optimization algorithm, embodying the search behavior adjustment method under a multi-strategy collaborative guidance mechanism. In this strategy, the update of an individual's current position consists of three parts: first, guidance from the global optimal solution, pushing all individuals towards the optimal solution to achieve a global convergence trend; second, maintaining the individual's own historical state, ensuring that the individual possesses a certain inertia and continuity during the search process, avoiding excessive jumps in the search trajectory; and third, the introduction of a memory feedback factor, combined with the evolution direction of high-quality solutions in the historical trajectory matrix, providing the current individual with search guidance based on historical experience, thereby enhancing the algorithm's intelligence and diversity. The algorithm can enhance the detailed exploration of the local structural space while ensuring global search capability, and fully utilize the optimal trajectory information accumulated during the search process, effectively improving convergence speed and search accuracy. This mechanism is particularly suitable for optimizing high-dimensional, nonlinear, and sparse structural parameter configuration problems, and is the key strategy of this invention for achieving fine control and efficient search in the process of deep echo state network structural parameter tuning.
[0030] S37. The individuals in the besieging layer perform small-step perturbation search around the globally optimal individual to finely adjust the local structural configuration of the deep echo state network. Specifically, the small-step perturbation search means that during the search process, the individuals in the besieging layer take the current structural parameters of the globally optimal individual as the center and adjust the structural parameters in the vicinity of it with small parameter increments to gradually explore potentially better solutions. S38, ranking top in current fitness The individual triggers a local re-search mechanism, which further optimizes the combination of structural parameters based on micro-amplitude perturbations. Percentage threshold; S39. Connect the skip link index matrix With multi-channel fusion weight vector The network is constructed as a graph structure, and the nested features of the graph structure are used to adaptively constrain the network connection dependencies. S310. Combining graph structure and fitness function, perform joint structure-performance evaluation on the combination of structural parameters and update the individual fitness ranking. S311. Each behavior layer updates the search path based on the strategy weight and local experience value, and enters the next round of position iteration or switches the behavior layer state. S312. When the maximum number of iterations or the convergence threshold is reached, output the current optimal combination of structure parameters, including the optimal... , , , It is used to configure the deep echo state network model structure.
[0031] This invention addresses the challenge of optimizing the structural parameters of deep echo-state networks by designing a Bald Eagle Search optimization algorithm with hierarchical behavioral strategies and graph structure adaptability, significantly enhancing the model's adaptability and search efficiency. The algorithm constructs three subgroups—a reconnaissance layer, a tracking layer, and a besieging layer—assigning different search tasks and behavioral characteristics, achieving a combination of global exploration and local fine-tuning of the search space. Individual encoding encompasses key structural parameters such as activation function selection, multi-channel fusion weights, skip connection indices, and sparsity, enabling the optimization process to directly drive model performance evolution. In the fitness function design, three-dimensional objectives—fusion error accuracy, structural sparsity, and channel fusion balance—effectively avoid structural bias and model redundancy. Simultaneously, a historical trajectory memory mechanism and an energy function guidance strategy are introduced to improve search stability and diversity. Furthermore, by modeling the dependencies between skip connections and fusion channels through graph structure, the consistency and interpretability between parameter configuration and network topology are improved. Ultimately, the algorithm achieves the acquisition of the optimal combination of structural parameters within a finite number of iterations, significantly improving the prediction accuracy and training efficiency of deep echo-state networks, providing high-performance structural optimization support for gravimeter digital leveling systems.
[0032] In this embodiment, S4 specifically includes: S41. The multi-source sensor data vector at the current moment and the multi-source sensor data from the k consecutive historical moments are combined to form an input sample sequence, which is then used as part of the input sample set and input into the deep echo state network model optimized by the Bald Eagle Search optimization algorithm. S42. Using the optimized deep echo state network model, multi-level state propagation and update operations are sequentially performed on the input sample sequence. In each network layer, the state of the reservoir node is updated under the combined effect of the current input, the previous state, and the jump connection state. Specifically, the multi-level state propagation and update operations are sequentially performed on the input sample sequence to each layer of the deep echo state network reservoir. Starting from the first layer, the calculation is performed recursively according to the time step. Each layer determines the current state output based on the current input it receives, its own state in the previous time step, and the historical layer state passed in by the jump connection. This state is then saved as the input of the next layer or as part of the global state of the network. S43. A dynamic switching mechanism of activation function is adopted to complete the nonlinear transformation of the state of the reserve pool, and a multi-input channel fusion structure is introduced in each network layer. The feature weighted fusion of K input sub-channels is performed according to the multi-channel fusion weight vector. S44. Construct a jump connection matrix in each layer of the reservoir. Guide the node state to receive information flow from the previous time and the state of a specific historical node through a jump connection mechanism. Specifically, the construction of the jump connection matrix means that in each layer of the deep echo state network, a binary structure matrix is generated according to the jump connection index rule to identify whether the current layer node has established a connection with the state output of the historical layer. Each element of the matrix indicates whether the state output from a certain layer among the previous several layers is used as part of the current node input. S45. A deep sparsity adjustment mechanism is introduced into the network structure. The input weight matrix and state transition matrix are pruned in a structured manner using a sparse mask to control the network complexity. Specifically, the structured pruning operation refers to filtering the connection structure in the input weight matrix and state transition matrix according to a preset sparsity rate parameter during the construction of the deep echo state network. Only non-zero weight connections that meet the mask position are retained, and the weight values of other connection positions are set to zero, so that the network connection presents a structured and controllable sparse distribution. S46. Merge the final state outputs of all network layers to construct a global state output vector. Perform a linear transformation on the global state output vector and the output connection weights to generate a tilt perturbation trend value vector, representing the tilt perturbation trend value at each time step in the future. Specifically, the linear transformation on the global state output vector and the output connection weights refers to concatenating the state output vectors generated by the reservoirs of each layer of the deep echo state network at each time step to form a unified global state output vector. Then, a set of preset or trained output connection weight matrices is used to perform a weighted summation operation on this vector, and corresponding bias terms are added to obtain a vector representing the tilt perturbation trend value at each time step in the future.
[0033] This invention effectively improves the expressive power and generalization performance of tilt disturbance trend prediction by introducing several structural innovation mechanisms into the prediction process of a deep echo state network model. The process first constructs time-series samples using current and historical multi-source sensor data and inputs them into the model optimized by the Bald Eagle Search algorithm, ensuring an optimal balance between adaptability and accuracy in the network structure. A skip connection mechanism is introduced during state propagation to strengthen cross-layer information flow and historical dependency modeling; a dynamic activation function switching mechanism dynamically adapts to nonlinear transformation strategies based on network levels and input features, enhancing the flexibility of state updates. A channel fusion structure enables fine-grained weighted aggregation of multi-input sensor signals, allowing the model to more accurately capture disturbance characteristics from different channels; a deep sparse adjustment mechanism constrains network complexity through structural pruning, balancing model accuracy and operational efficiency. The network fuses the states of each layer to construct a global state vector, achieving accurate prediction of future tilt disturbance trends. This method significantly enhances the model's adaptability to environmental changes and dynamic disturbances while improving the accuracy of tilt disturbance prediction, making it suitable for digital leveling compensation control requirements under various complex operating conditions.
[0034] In this embodiment, S5 specifically includes: S51. Based on the tilt perturbation trend value vector output by the optimized deep echo state network model, and combined with the current attitude state quantity and historical offset response, construct a sliding prediction window. S52. Based on the tilt perturbation trend value within the sliding prediction window, construct a compensation prediction structure that integrates multiple time windows, and extract the short-term window values respectively. With the mid-term window The perturbation information is obtained through the state uncertainty modulation factor. Comprehensive calculation of target tilt angle compensation adjustment amount : ; in, For short-term forecast window length, For the overall predicted time length, The perturbation trend weighting function is defined as follows: , It is an exponential function. The state uncertainty modulation factor. Indicates the first The trend value of the tilt angle perturbation at any given time. Indicates the first The tilt angle disturbance trend value at time -1; Target tilt angle compensation adjustment amount This embodies the weighted fusion concept based on multi-scale predictive information. Its core lies in extracting the degree of influence of different time points on the current leveling operation from the tilt angle disturbance trend values over a future period, and dynamically modulating this information in conjunction with state uncertainty to generate a more representative and stable compensation command. The target tilt angle compensation adjustment amount... The numerator and denominator in the formula each form a weighted average structure, where the predicted value at each time step is weighted according to its corresponding uncertainty modulation factor and disturbance trend weight function. This weighting strategy enhances the sensitivity of short-term changes to immediate compensation response while retaining the corrective effect of medium-term trends on overall attitude changes, ensuring that the leveling operation is both fast-responding and stable. This compensation calculation method effectively avoids the accumulation of noise and errors that may be caused by predictions at a single moment by dynamically integrating prediction results from multiple time scales. This makes the leveling control more accurate and smooth, improving the system's compensation robustness and adjustment accuracy in complex dynamic environments, and is a key element in achieving high-performance digital leveling control.
[0035] S53. Constructing a dynamic control gain function Adjust the compensation sensitivity of the leveling response along different axes and calculate the three-dimensional leveling command vector. Anisotropic gain coefficients are defined based on the physical structure and inertial characteristics of the equipment, where, , and These represent the dynamic control gain function values for the X, Y, and Z axes at the current time t, respectively. S54, Transfer the three-dimensional leveling command vector Feedback status with the current digital leveling controller The input is combined and processed according to the nonlinear control mapping relationship to construct a control command vector for driving the digital leveling controller. The control command vector is subjected to amplitude compression processing. Specifically, the processing according to the nonlinear control mapping relationship refers to inputting the three-dimensional leveling command vector and the attitude feedback state into a preset nonlinear function model and performing a nonlinear transformation to generate a control command vector that can drive the digital leveling controller to perform compensation. S55. Output the control command vector to the digital leveling controller, which drives the three-axis actuator to complete the tilt angle adjustment and compensation operation, thereby realizing real-time digital leveling based on the predicted compensation amount.
[0036] This invention proposes a digital leveling compensation method based on a sliding prediction window and a multi-scale perturbation fusion mechanism, which significantly improves the real-time leveling response accuracy and stability of a relative gravimeter in dynamic environments. By constructing a sliding prediction window that includes the current attitude state and historical offset responses, the method captures the attitude evolution trend of the equipment in real time. Simultaneously, it integrates perturbation trend information within short-term and medium-term time windows and introduces a state uncertainty modulation factor, achieving adaptive identification and weighted compensation for different perturbation scales. This method effectively avoids the response delay and over-adjustment problems under a single-scale strategy, enhancing the stability and robustness of the leveling compensation calculation. In the compensation command generation stage, a dynamic control gain function is constructed, and the leveling sensitivity of the X, Y, and Z axes is adjusted according to the equipment inertia and physical structure, giving the three-dimensional command directional adjustment capability. Finally, the three-dimensional command vector and the controller feedback state are jointly modeled, and compressed control commands are formed through nonlinear mapping, precisely driving the three-axis actuators to complete real-time leveling adjustments. The overall solution features high response speed, high compensation accuracy, and anti-disturbance capability, providing stable and reliable tilt leveling support for gravity measurement equipment in complex scenarios such as oil exploration, geodesy, and military deployment.
[0037] In this embodiment, S6 specifically includes: S61. Collect the attitude feedback data after the digital leveling controller is executed, and use it as the actual attitude response information at the current moment; S62. Extract the tilt angle disturbance trend value corresponding to the attitude feedback data, and construct a one-to-one correspondence between the tilt angle disturbance trend value and the attitude feedback data; S63. Compare the tilt angle disturbance trend value with the corresponding attitude feedback data to generate an error sequence, representing the prediction deviation at each time step; S64. Construct error feedback samples based on the error sequence, including the tilt perturbation trend value, attitude feedback data and error value at each time step; S65. Input the error feedback samples into the deep echo state network model and perform incremental updates on the output connection weights based on the error feedback mechanism. S66. By adjusting the updated output connection weights, the network's ability to predict tilt perturbation trends is improved, thereby enhancing the accuracy of its response to future perturbation changes.
[0038] This invention proposes a feedback update mechanism based on the comparison of attitude feedback and tilt angle disturbance trend value errors after performing digital leveling compensation operations, aiming to improve the system's response accuracy and model adaptability. By collecting attitude feedback data after the digital leveling controller's execution and comparing it one-to-one with the previously predicted tilt angle disturbance trend values, an error sequence is constructed. This sequence accurately reflects the current model's prediction deviation of the disturbance trend in the actual scenario. The error feedback samples further constructed based on this error sequence not only cover the predicted and feedback values but also clearly indicate the degree of deviation at each time step, providing an accurate basis for online model adjustment. This invention inputs the error feedback samples into a deep echo state network model, updating only the output connection weights without changing the network structure parameters. This achieves dynamic optimization of the model's prediction performance while maintaining the original network stability. Through an incremental update strategy, the system can continuously learn the disturbance patterns under different operating conditions, effectively improving the model's prediction robustness and generalization ability for complex attitude disturbance scenarios, thereby enhancing the adaptability and compensation accuracy of the digital leveling control system in practical applications.
[0039] In this embodiment, S7 specifically includes: After completing a leveling compensation operation, error information between the actual attitude feedback data and the predicted tilt angle disturbance trend value is collected to construct an error feedback sample. Based on this feedback sample, the performance of the current deep echo state network model structural parameters is dynamically evaluated, and it is used as an important input for fitness evaluation to update the individual fitness ranking and historical trajectory matrix in the vulture search optimization algorithm. The search strategy is locally adjusted, including optimizing the behavior switching mechanism of the subgroup, search direction, and jump amplitude, so that the algorithm focuses on the parameter region with better performance for fine search. This process is continuously iterated after each leveling execution. Through multiple feedback corrections and structural updates, the model's prediction accuracy of tilt angle disturbance trends is gradually improved, and finally a closed-loop leveling self-learning control mechanism of "perception-prediction-control-feedback" is constructed.
[0040] refer to Figure 2 A digital leveling and compensation system for a relative gravimeter includes the following modules: The data acquisition and preprocessing module is used to acquire multi-source sensor data and preprocess it to construct an input sample set. The Deep Echo State Network Building Module is used to build deep echo state network models. The structural parameter optimization module is used to introduce the vulture search optimization algorithm to search and optimize the structural parameters of the deep echo state network model; The tilt disturbance trend module is used to input current and historical sensor data and output tilt disturbance trend values; The leveling instruction generation module is used to construct a disturbance weight function and modulation coefficient based on the tilt angle disturbance trend value, generate a leveling instruction vector, and perform compensation operations. The feedback error construction module is used to collect attitude feedback data, compare it with the tilt angle disturbance trend value, and generate error feedback samples. The model and optimization update module is used to update the output connection weights based on error feedback samples and simultaneously adjust the local search strategy of the vulture search optimization algorithm.
[0041] Example 1: To verify the feasibility of this invention in practice, it was applied to a key geophysical research institute. An area at an altitude of 4500 meters in the heart of a plateau was selected as the experimental site, and 10 relative gravity observation points were set up to conduct high-precision gravity measurements. Due to the complex terrain, strong winds, and dramatic diurnal temperature variations in this area, traditional gravimeters frequently experienced tilt drift and slow leveling system response during long-term operation, severely affecting the continuity and reliability of gravity measurement data.
[0042] To address the aforementioned issues, this invention proposes a digital leveling compensation system and method for relative gravimeters, which has been applied to an experimental gravimeter system. The method first constructs a standardized input sample set by collecting multi-source sensor data, including tilt angle, vibration, temperature, and humidity, during equipment operation. The constructed deep echo state network model integrates a multi-input channel mechanism and a skip connection strategy, and utilizes a dynamic activation function switching mechanism and a sparse adjustment mechanism during the prediction phase to accurately model tilt angle disturbance trends. Simultaneously, the model structure parameters are optimized through a multi-behavior hierarchical search using an improved vulture search optimization algorithm to adapt to nonlinear tilt angle fluctuation modeling in complex terrain environments.
[0043] After deployment at each observation point, the system predicts the tilt disturbance trend in real time and outputs compensation commands based on the trend to drive the digital leveling controller to adjust the tilt. Feedback data is then used to update the model output weights and optimization strategies in a closed-loop manner. Comparative tests with traditional leveling strategies reveal that, over a continuous 48-hour measurement period, the gravimeter system employing this invention significantly outperforms traditional methods in terms of prediction accuracy, response speed, and compensation error.
[0044] Table 1 Comparison of Prediction Accuracy in Geophysical Research Applications
[0045] Table 1 above shows the prediction accuracy of the method of the present invention and the traditional method at 10 typical gravity measurement points, from which the superior performance of the method of the present invention in practical applications can be clearly observed.
[0046] At all measurement points, the prediction accuracy of this invention is significantly higher than that of traditional methods, with an accuracy improvement generally exceeding 3%, and even approaching 7% at some individual measurement points. For example, at measurement point 2, the prediction accuracy of the traditional method is 89.75%, while the method of this invention reaches 96.16%, an improvement of 6.41 percentage points, demonstrating its powerful ability to model complex tilt disturbances. Accuracy improvements of over 6% were also achieved at measurement points 3 and 8. This indicates that the digital leveling compensation method of this invention, based on a deep echo state network and a vulture search optimization algorithm, can more accurately capture tilt disturbance trends and effectively output leveling compensation commands.
[0047] Furthermore, the accuracy of traditional methods at some measuring points, such as measuring points 5, 6, and 7, is generally low, hovering around 85%. However, after introducing the method of this invention, the prediction accuracy steadily improves to over 89%, indicating that this method also possesses stability and reliability in areas with complex terrain or strong external interference. Overall, this invention demonstrates stable and excellent prediction results under multiple measuring points and diverse terrain conditions. It is particularly suitable for the high-precision gravity measurement needs in complex environments such as plateaus and mountains in geophysical research, effectively supporting the scientific rigor and accuracy of subsequent gravity anomaly analysis and crustal movement research.
[0048] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method of digital levelling compensation for a relative gravimeter, characterized in that, Includes the following steps: S1. Collect multi-source sensor data of the relative gravimeter in working condition, preprocess the multi-source sensor data, and construct the input sample set; S2. Construct a deep echo state network model, receive the generated input sample set, and model and predict the tilt perturbation trend; S3. The vulture search optimization algorithm is introduced to optimize the structural parameters of the deep echo state network model. The vulture search algorithm is used to perform parameter combination search and construct a fitness function to perform search update. S4. Apply the optimized deep echo state network model to the input sensor data at the current moment and in the past, and output the predicted tilt perturbation trend value in the future period. S5. Calculate the leveling compensation command based on the tilt angle disturbance trend value, send it to the digital leveling controller, and drive the actuator to complete the leveling compensation operation. S6. Collect attitude feedback data after the digital leveling controller is executed, compare it with the tilt angle disturbance trend value, construct error feedback samples, and update the output connection weights of the deep echo state network model. S7. Update the search strategy of the Bald Eagle Search Optimization Algorithm based on the error feedback samples, make local adjustments and iterative optimizations to the structural parameter search space, and form a closed-loop leveling self-learning control mechanism in the process of multiple leveling compensations.
2. The method of claim 1, wherein the digital leveling compensation is performed by a computer program. The multi-source sensor data specifically includes tilt sensor data, ground vibration data, ambient temperature and air pressure data, target value of actual tilt disturbance trend, power status data, and feedback data from the leveling controller.
3. The method of claim 1, wherein the digital leveling compensation is performed by a computer program. The preprocessing of multi-source sensor data specifically includes timestamp alignment, missing value imputation, outlier removal, normalization, and format standardization.
4. The method of claim 1, wherein the digital leveling compensation is performed by a computer. S2 specifically includes: S21. Set the input sample set to a standardized multi-source sensor data sequence. ,in, Indicates the number of sensor channels. Indicates the number of time steps. For the first The sensor data vector at time t, where R is the set of real numbers; S22, dividing the input sample set into input sub-channels according to the channel, respectively denoted as , wherein , and defining a multi-channel fusion weight vector as ; S23, Calculate the first Input vector of the layered reservoir ; S24. The total number of layers in the deep echo state network is: Each layer of the storage pool contains There are 1 state node, and the initial state is a zero vector. ; S25. Set leakage rate parameters for each layer of the reservoir. And set the sparsity parameter. Control state transition matrix The proportion of non-zero connections; S26. Initialize the input weight matrix for each layer. With the state transition matrix The input weight matrix and state transition matrix are randomly generated using a Gaussian distribution and based on the sparsity parameter. Apply a sparse mask; S27. Specify the activation function type for each reservoir layer, and select parameters through the activation function. The activation function used is controlled, and the mapping relationship is defined as follows: Use Tanh; Use ReLU; Use ELU; : Use Swish; S28. Define the state update for each layer as follows: ; S29. Perform time-step processing on the state vectors of all layers. Concatenate the vectors to construct the network's global state output vector. ; S210. Construct a connection weight matrix and bias terms for the output, wherein the connection weight matrix receives the network global state output vector. As input, it is mapped to the tilt angle disturbance trend value according to the predetermined dimension. The bias term is used to perform linear offset correction on the tilt angle disturbance trend value to generate the output at the corresponding time. S211. Compare the tilt angle perturbation trend value with the corresponding true tilt angle perturbation trend target value, and conduct supervised training by constructing a loss function of the mean square error type. During the training process, adjust the output connection weight matrix and bias term parameters, while keeping the structure and state weights of the other layers of the reservoir unchanged.
5. The digital leveling compensation method for a relative gravimeter according to claim 1, characterized in that, S3 specifically includes: S31. Construct a multi-layered behavioral structure for the bald eagle search optimization algorithm. Divide the initial optimization population into three sub-populations: a reconnaissance layer, a tracking layer, and a besieging layer. Each individual is encoded as a combination of structural parameters of a deep echo state network model, including: activation function selection parameters. Multi-channel fusion weight vector Skip Link Index Matrix sparsity parameter ; S32. Initialize the vulture algorithm population position vector and historical trajectory matrix. Set the maximum number of iteration rounds. Behavior switching threshold Local research probability and trajectory fusion coefficient ; S33. Constructing a multi-objective fitness function Simultaneously considering prediction error, sparsity of skip connections, and balance of fusion channels; S34. Calculate the fitness function value for all individuals, record the current best individual, and update the historical trajectory matrix. The optimal structural parameter solution for each round is retained; S35. Individuals in the reconnaissance layer execute a large-scale jump search strategy to expand the search space coverage. They adjust the jump amplitude and direction in combination with the current energy function. The current energy function is a search guidance function dynamically constructed based on the population's optimal fitness value, the individual's current fitness value, and the global average fitness value. S36, Tracking layer individuals are based on the optimal individual Current individual Trajectory memory vector Update location ; S37. Individuals in the siege layer perform small-step perturbation search around the globally optimal individual to finely adjust the local structural configuration of the deep echo state network. S38, ranking top in current fitness The individual triggers a local re-search mechanism, which further optimizes the combination of structural parameters based on micro-amplitude perturbations. Percentage threshold; S39. Connect the skip link index matrix With multi-channel fusion weight vector The network is constructed as a graph structure, and the nested features of the graph structure are used to adaptively constrain the network connection dependencies. S310. Combining graph structure and fitness function, perform joint structure-performance evaluation on the combination of structural parameters and update the individual fitness ranking. S311. Each behavior layer updates the search path based on the strategy weight and local experience value, and enters the next round of position iteration or switches the behavior layer state. S312. When the maximum number of iterations or the convergence threshold is reached, output the current optimal combination of structure parameters, including the optimal... , , , It is used to configure the deep echo state network model structure.
6. The digital leveling compensation method for a relative gravimeter according to claim 4, characterized in that, S4 specifically includes: S41. The multi-source sensor data vector at the current moment and the multi-source sensor data from the k consecutive historical moments are combined to form an input sample sequence, which is then used as part of the input sample set and input into the deep echo state network model optimized by the Bald Eagle Search optimization algorithm. S42. Using the optimized deep echo state network model, multi-layer state propagation and update operations are performed sequentially on the input sample sequence. In each network layer, the state of the reserve pool node is updated under the combined action of the current input, the previous state, and the jump connection state. S43. A dynamic switching mechanism of activation function is adopted to complete the nonlinear transformation of the state of the reserve pool, and a multi-input channel fusion structure is introduced in each network layer. The feature weighted fusion of K input sub-channels is performed according to the multi-channel fusion weight vector. S44. Construct a jump connection matrix in each layer of the reserve pool, and guide the node state through the jump connection mechanism to receive the information flow of the previous time and the specific historical node state at the same time. S45. Introduce a deep sparse adjustment mechanism into the network structure, and use sparse masks to perform structured pruning operations on the input weight matrix and state transition matrix to control the network complexity. S46. Merge the final state outputs of all network layers to construct the global state output vector of the network. Perform a linear transformation on the global state output vector and the output connection weights to generate a tilt perturbation trend value vector, which represents the tilt perturbation trend value at each time point in the future.
7. The digital leveling compensation method for a relative gravimeter according to claim 1, characterized in that, S5 specifically includes: S51. Based on the tilt perturbation trend value vector output by the optimized deep echo state network model, and combined with the current attitude state quantity and historical offset response, construct a sliding prediction window. S52. Based on the tilt perturbation trend value within the sliding prediction window, construct a compensation prediction structure that integrates multiple time windows, and extract the short-term window values respectively. With the mid-term window The perturbation information is obtained through the state uncertainty modulation factor. Comprehensive calculation of target tilt angle compensation adjustment amount ; S53. Constructing a dynamic control gain function Adjust the compensation sensitivity of the leveling response along different axes and calculate the three-dimensional leveling command vector. Anisotropic gain coefficients are defined based on the physical structure and inertial characteristics of the equipment, where, , and These represent the dynamic control gain function values for the X, Y, and Z axes at the current time t, respectively. S54, Transfer the three-dimensional leveling command vector Feedback status with the current digital leveling controller The combined input is processed according to the nonlinear control mapping relationship to construct a control command vector for driving the digital leveling controller. The control command vector is then subjected to amplitude compression processing. S55. Output the control command vector to the digital leveling controller, which drives the three-axis actuator to complete the tilt angle adjustment and compensation operation, thereby realizing real-time digital leveling based on the predicted compensation amount.
8. The digital leveling compensation method for a relative gravimeter according to claim 1, characterized in that, S6 specifically includes: S61. Collect the attitude feedback data after the digital leveling controller is executed, and use it as the actual attitude response information at the current moment; S62. Extract the tilt angle disturbance trend value corresponding to the attitude feedback data, and construct a one-to-one correspondence between the tilt angle disturbance trend value and the attitude feedback data; S63. Compare the tilt angle disturbance trend value with the corresponding attitude feedback data to generate an error sequence, representing the prediction deviation at each time step; S64. Construct error feedback samples based on the error sequence, including the tilt perturbation trend value, attitude feedback data and error value at each time step; S65. Input the error feedback samples into the deep echo state network model and perform incremental updates on the output connection weights based on the error feedback mechanism. S66. Adjust the network's ability to predict tilt perturbation trends by updating the output connection weights.
9. A digital leveling compensation system for a relative gravimeter, applied to the digital leveling compensation method for a relative gravimeter as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to acquire multi-source sensor data and preprocess it to construct an input sample set. The Deep Echo State Network Building Module is used to build deep echo state network models. The structural parameter optimization module is used to introduce the vulture search optimization algorithm to search and optimize the structural parameters of the deep echo state network model; The tilt disturbance trend module is used to input current and historical sensor data and output tilt disturbance trend values; The leveling instruction generation module is used to construct a disturbance weight function and modulation coefficient based on the tilt angle disturbance trend value, generate a leveling instruction vector, and perform compensation operations. The feedback error construction module is used to collect attitude feedback data, compare it with the tilt angle disturbance trend value, and generate error feedback samples. The model and optimization update module is used to update the output connection weights based on error feedback samples and simultaneously adjust the local search strategy of the vulture search optimization algorithm.
