Distillation ensemble learning soil parameter inversion method oriented to extraterrestrial celestial body safe sampling
Through the machine-soil coupled contact mechanical model and the Monte Carlo method combined with the lightweight Stacking integrated learning model, the accuracy and resource limitation problems of soil parameter identification in extraterrestrial celestial sampling are solved, efficient and accurate soil parameter inversion is achieved, and the task capability of the robotic arm is improved.
Patent Information
- Application Number
- CN202510960344.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-11
- Publication Date
- 2025-09-05
AI Technical Summary
Traditional soil identification methods are unable to obtain a large amount of actual test data in extraterrestrial celestial sampling due to scarce samples, resulting in poor accuracy in soil parameter identification and limited on-site computing resources, making it difficult to train large-scale computing models.
The machine-soil coupled contact mechanics model is used to generate the shovel digging working condition with the Monte Carlo method, and the integrated learning model of fusion knowledge distillation is established. The lightweight Stacking integrated learning model is trained through the six-dimensional force sensor to real-time data acquisition and training of the lightweight Stacking integrated learning model to achieve real-time inversion of soil parameters.
It improves the accuracy and efficiency of soil parameter identification, reduces the storage space requirements of the model, and enhances the task capability and task success rate of the robotic arm in unknown environments.
Smart Images
Figure CN120597722A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence. Background Art
[0002] As humankind continues to deepen its exploration of space, sampling extraterrestrial bodies has become an inevitable trend in the future development of aerospace. Identification of soil mechanical parameters (soil parameters for short) is one of the key technologies for sampling extraterrestrial bodies. The soil of extraterrestrial bodies such as the Moon and Mars is significantly different from that of the Earth. The complexity of their physical properties and mechanical behavior makes soil parameter inversion a technical challenge. For sampling missions, understanding and mastering the mechanical properties of soil can effectively enhance the ability of the robotic arm to perform tasks in unknown soil environments, improve the success rate of missions, and extend the service life of the robotic arm.
[0003] However, traditional soil identification methods typically rely on extensive experimental data and complex computational models. Due to insufficient prior knowledge of extraterrestrial soils and the scarcity of returned samples, it is impossible to obtain extensive experimental data from actual tests, resulting in poor accuracy in identifying soil parameters. Furthermore, the computing resources and storage space of onboard computers are severely limited, making it difficult to train and deploy large-scale computational models. Therefore, efficiently extracting soil mechanical parameters within the constraints of limited resources and computing power has become a pressing technical challenge. Summary of the Invention
[0004] The present invention aims to solve the problem of poor accuracy in identifying soil parameters caused by the scarcity of samples and the inability to obtain a large amount of experimental data from actual tests in traditional soil identification methods. A distillation ensemble learning soil parameter inversion method for safe sampling of extraterrestrial objects is now provided.
[0005] A distillation ensemble learning soil parameter inversion method for safe sampling of extraterrestrial objects is provided, which includes the following contents:
[0006] Step 1: Establish a machine-soil coupled contact mechanics model;
[0007] Step 2: Using the Monte Carlo method to randomly generate multiple groups of excavation conditions for each type of soil, the machine-soil coupled contact mechanics model obtains the excavation resistance generated by each type of soil based on the historical soil parameters of each type of soil and each group of excavation conditions. The soil parameters include the soil internal friction angle, the excavation-soil internal friction angle, and the soil density.
[0008] Step 3: Take each set of excavation conditions and the excavation resistance generated by the corresponding soil type as a set of input data, and take the soil parameters of the corresponding soil type as a set of output data to form a data set to train the fusion knowledge distillation ensemble learning model, and obtain the trained fusion knowledge distillation ensemble learning model;
[0009] Step 4: The excavation resistance and excavation conditions of the soil to be predicted are collected in real time and input into the trained fusion knowledge distillation integrated learning model to obtain the corresponding soil parameters.
[0010] Preferably, in step 1, the machine-soil coupled contact mechanics model:
[0011] F ED =F E +F V +F B ,
[0012] Where, F E F represents the digging resistance generated by the excavated star soil. V represents the inertial resistance related to the digging rate, F B represents the digging resistance associated with the bucket edge effect; F E =F EP +F q , F q =qHN q b, ψ represents the angle between the tangent line of the bucket surface point corresponding to the microelement dz and the vertical direction, δ represents the friction angle between the bucket and the lunar soil, β represents the angle between the slope of the lunar soil surface and the bucket, H is the excavation depth, represents the internal friction angle of the soil, q represents the equivalent uniformly distributed load generated by the excavated soil, V e represents the digging rate, θ * is the angle between the actual velocity direction of the soil layer and the horizontal direction, L α is the horizontal projection length of the shovel, γ is the soil density, α is the shovel digging angle, L is the shovel digging distance, b is the width of the bucket, z is the width of the bucket, N q is the influence coefficient of the additional distributed load on the soil surface, g is the local gravity acceleration, K b is the lateral excavation resistance correction coefficient of the shovel body at depth z, and the soil parameters are: δ and γ, shovel excavation conditions are: α, V e , H and L.
[0013] Preferably, in step 4, the digging resistance of the soil to be predicted is collected by using a six-dimensional force sensor mounted at the connection between the robotic arm and the bucket.
[0014] Preferably, in step 4, the shoveling and digging conditions of the soil to be predicted are collected, specifically:
[0015] The forward kinematics model estimates the end position of the robotic arm based on the joint angle information of the robotic arm at each moment, and combines the relative position relationship between the robotic arm and the soil surface to be predicted to estimate the shoveling condition of the soil to be predicted at each moment.
[0016] Preferably, a Monte Carlo method is used to randomly generate multiple groups of shoveling conditions from a preset shoveling condition range.
[0017] Preferably, the preset shoveling working condition range is that the shoveling angle α ranges from 30° to 75°, and the shoveling rate V e The range is 0.001m / s-0.005m / s, the shovel digging distance L range is 0.1m-0.5m, and the digging depth H range is 0.01-0.03m.
[0018] Beneficial effects of this application:
[0019] An accurate machine-soil coupled contact mechanics model was established. Mechanical equations were used to determine the relationship between excavation resistance and three soil parameters and four operating parameters. This clearly demonstrates the different effects of each parameter and enhances the reliability of the inversion results. Furthermore, the Monte Carlo method was used to randomly simulate the operating parameters. Combined with prior knowledge of terrestrial soil, the excavation resistance was calculated based on the machine-soil coupled contact mechanics equations, and a data set was established. This reduced the model training's reliance on large amounts of measured data and addressed the challenge of insufficient prior knowledge for extraterrestrial sampling. Therefore, compared to existing technologies using the machine-soil coupled contact mechanics model and the Monte Carlo method, the accuracy of identifying soil parameters is higher.
[0020] Furthermore, this method utilizes an ensemble learning model to enhance prediction accuracy. To address the limited storage space of onboard computers, this method incorporates knowledge distillation technology to significantly reduce the model size while simultaneously increasing the speed of soil parameter inversion, providing the necessary guarantees for real-time feedback control of the robotic arm's coupled machine-soil excavation. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 Flowchart of the distillation ensemble learning soil parameter inversion method for safe sampling of extraterrestrial objects;
[0022] Figure 2 This is the framework diagram of the Stacking integrated learning model;
[0023] Figure 3 Schematic diagram of the knowledge distillation algorithm. DETAILED DESCRIPTION
[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other in the absence of conflict.
[0025] Example:
[0026] A distillation ensemble learning soil parameter inversion method for safe sampling of extraterrestrial objects is provided, which includes the following contents:
[0027] Step 1: Establish a machine-soil coupled contact mechanics model;
[0028] Step 2: Using the Monte Carlo method to randomly generate multiple groups of excavation conditions for each type of soil, the machine-soil coupled contact mechanics model obtains the excavation resistance generated by each type of soil based on the historical soil parameters of each type of soil and each group of excavation conditions. The soil parameters include the soil internal friction angle, the excavation-soil internal friction angle, and the soil density.
[0029] Step 3: Take each set of excavation conditions and the excavation resistance generated by the corresponding soil type as a set of input data, and take the soil parameters of the corresponding soil type as a set of output data to form a data set to train the fusion knowledge distillation ensemble learning model, and obtain the trained fusion knowledge distillation ensemble learning model;
[0030] Step 4: The excavation resistance and excavation conditions of the soil to be predicted are collected in real time and input into the trained fusion knowledge distillation integrated learning model to obtain the corresponding soil parameters.
[0031] Specifically, the established machine-soil coupled contact mechanics model can quickly calculate the digging resistance, thereby improving the speed of obtaining soil parameters.
[0032] In the inversion of mechanical parameters of extraterrestrial regolith, stacking ensemble learning and knowledge distillation are key components of model intelligence and lightweighting. This process consists of two phases: first, training a high-precision stacking ensemble learning model that fuses multiple models; then, using knowledge distillation, compressing this model into a lightweight student model suitable for onboard computing deployment. The principles and implementation process of these two modules are detailed below.
[0033] 1. Structure and training process of stacking ensemble learning model
[0034] Stacking is a hierarchical ensemble learning method consisting of a base learner layer and a meta-learner layer. This method introduces four complementary models into the base learner layer: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Support Vector Regression (SVR). These models are used to model the complex nonlinear mapping relationship between working condition parameters and digging resistance to soil parameters from different perspectives. In the meta-learner layer, a Lightweight Gradient Boosting Machine (LightGBM) is used as a fusion layer to relearn the predictions of each base model to further improve the generalization capability of the overall model.
[0035] During training, the machine-soil coupled contact mechanics model established earlier was first used, combined with a large number of randomly generated working condition parameter samples using the Monte Carlo method, to calculate the corresponding excavation resistance. The working condition parameters and resistance were then used as input, and soil mechanical parameters (including internal friction angle, excavator-soil internal friction angle, and soil density) were used as labels to construct a standard supervised learning dataset.
[0036] During training, each base model is first trained using K-fold cross-validation, generating predictions on the validation set. These outputs are treated as new features and fed into the meta-learner for secondary modeling. This enables the stacking model to extract information from multiple sub-models and integrate predictions at a higher level, significantly improving inversion accuracy and robustness.
[0037] After training is completed, the Stacking model becomes a "teacher model" with superior performance but large size, providing a knowledge source for subsequent model distillation and compression.
[0038] 2. Knowledge Distillation and Construction of Lightweight Student Model
[0039] Given the limited computing resources of onboard computers, it's impractical to directly deploy a full stacking model. Therefore, this method introduces knowledge distillation technology to compress the complex teacher model into a computationally less expensive student model, enabling efficient deployment and online inference.
[0040] Specifically, the XGBoost model was selected as the student model. During distillation training, the input remains the combination of working parameters and digging resistance, but the labels are no longer the original soil parameters, but rather the teacher model's predictions based on these inputs (i.e., "soft labels"). In this way, the student model does not directly learn the real data labels, but instead learns the teacher model's predictions, attempting to approximate its outputs and thus inherit its modeling capabilities.
[0041] The entire distillation process is essentially a regression task, using mean squared error (MSE) as the loss function to measure the distance between the student model output and the teacher model output. Through continuous training, the student model gradually acquires the teacher model's understanding of the input-output mapping and expresses this relationship in a more concise structure.
[0042] Ultimately, the resulting student model has the following advantages: small size, fast inference speed, and low computing resource usage. It can be deployed on the rover's onboard computing platform to achieve real-time inversion of field-collected data and output the three key mechanical parameters of extraterrestrial soil.
[0043] The trained student model is deployed on the onboard computer and launched with the probe.
[0044] After the probe lands on an extraterrestrial body, it performs a sampling operation, using a multi-degree-of-freedom robotic arm with a sampling shovel at the end to dig and sample the shallow surface of the extraterrestrial body.
[0045] The stacking ensemble learning algorithm integrates multiple powerful learners to overcome the limitations of a single model, improve inversion accuracy, and leverage the strengths of different models. Knowledge distillation reduces model size and improves operational efficiency, enabling deployment and execution with limited computing power and resources. While ensuring model prediction accuracy, it also achieves model lightweighting, improving computational efficiency and real-time performance, adapting to limited computing power and small sample sizes. This lightweight model enables the rover to perform real-time inversion of soil mechanical parameters on extraterrestrial bodies without relying on ground support, enhancing mission autonomy and efficiency.
[0046] Further definition, in step 1, the machine-soil coupled contact mechanics model:
[0047] F ED =F E +F V +F B ,
[0048] Where, F E F represents the digging resistance generated by the excavated star soil. V represents the inertial resistance related to the digging rate, F B represents the digging resistance associated with the bucket edge effect; FE =F EP +F q , F q =qHN q b, ψ represents the angle between the tangent line of the bucket surface point corresponding to the microelement dz and the vertical direction, δ represents the friction angle between the bucket and the lunar soil, β represents the angle between the slope of the lunar soil surface and the bucket, H is the excavation depth, represents the internal friction angle of the soil, q represents the equivalent uniformly distributed load generated by the excavated soil, V e represents the digging rate, θ * is the angle between the actual velocity direction of the soil layer and the horizontal direction, L α is the horizontal projection length of the shovel, γ is the soil density, α is the shovel digging angle, L is the shovel digging distance, b is the width of the bucket, z is the width of the bucket, N q is the influence coefficient of the additional distributed load on the soil surface, g is the local gravity acceleration, K b is the lateral excavation resistance correction coefficient of the shovel body at depth z, and the soil parameters are: δ and γ, shovel excavation conditions are: α, V e , H and L.
[0049] Specifically, for step 1: the digging rate, edge effect, and trembling shape are integrated into the model to construct a general calculation expression for the digging resistance of the curved digging body, namely, the machine-soil coupled contact mechanics model.
[0050] For step 2: By constructing a complete sampling shovel-soil coupled contact mechanics model, it can be obtained that the digging resistance F at the end of the robotic arm is only related to the soil internal friction angle Shovel-soil internal friction angle δ, soil density γ, shovel entry angle α, excavation rate V e , the shovel depth H, the shovel distance L. Among them, the soil internal friction angle The shovel-soil internal friction angle δ and soil density γ are soil parameters related to soil properties, the shovel angle α, the shovel rate V e , digging depth H, and digging distance L are working condition parameters related to the digging action. Therefore, these eight quantities are selected as parameters to construct the data set.
[0051]
[0052] In order to truly reflect the operation scenario of the detector in an unknown environment, the Monte Carlo method is used to perform multi-dimensional random simulation of the shoveling working condition. The shoveling angle α ranges from 30° to 75°, and the shoveling rate V eThe range is 0.001m / s-0.005m / s to adapt to the speed limit conditions of the robotic arm joint, the shoveling distance L range is 0.1m-0.5m as the effective working stroke of the corresponding sampling arm, and the excavation depth H range is 0.01-0.03m.
[0053] Combined with the prior knowledge of extraterrestrial bodies and soils (remote sensing, field measurements, etc.), the soil parameter space is established. For example, for the mechanical parameters of lunar soil, based on the research information, seven types of lunar soil parameter spaces are established as shown in the following table:
[0054] Table 1 Lunar soil parameter space
[0055]
[0056] Through the uniform distribution random sampling strategy, 10,000 groups of working condition parameter combinations are generated for each soil type.
[0057] For each combination of working condition parameters and soil parameters, the corresponding digging resistance is calculated in combination with the shovel-soil coupled contact mechanics model.
[0058] The shovel angle α and the shovel rate V e , digging depth H, digging distance L and digging resistance F are taken as sample sets, and the soil internal friction angle The shovel-soil internal friction angle δ and soil weight γ are used as feature sets to construct a dataset for subsequent learning algorithms.
[0059] It is further defined that in step 4, the digging resistance of the soil to be predicted is collected using a six-dimensional force sensor mounted at the connection between the robotic arm and the bucket.
[0060] Further defined, in step 4, the shoveling and digging conditions of the soil to be predicted are collected, specifically:
[0061] The forward kinematics model estimates the end position of the robotic arm based on the joint angle information of the robotic arm at each moment, and combines the relative position relationship between the robotic arm and the soil surface to be predicted to estimate the shoveling condition of the soil to be predicted at each moment.
[0062] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should therefore be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.
Claims
1. A distillation ensemble learning soil parameter inversion method for safe sampling of extraterrestrial objects, characterized by: The method includes the following: Step 1: Establish a machine-soil coupled contact mechanics model; Step 2: Using the Monte Carlo method to randomly generate multiple groups of excavation conditions for each type of soil, the machine-soil coupled contact mechanics model obtains the excavation resistance generated by each type of soil based on the historical soil parameters of each type of soil and each group of excavation conditions. The soil parameters include the soil internal friction angle, the excavation-soil internal friction angle, and the soil density. Step 3: Take each set of excavation conditions and the excavation resistance generated by the corresponding soil type as a set of input data, and take the soil parameters of the corresponding soil type as a set of output data to form a data set to train the fusion knowledge distillation ensemble learning model, and obtain the trained fusion knowledge distillation ensemble learning model; Step 4: The excavation resistance and excavation conditions of the soil to be predicted are collected in real time and input into the trained fusion knowledge distillation integrated learning model to obtain the corresponding soil parameters.
2. The distillation ensemble learning soil parameter inversion method for extraterrestrial object safe sampling according to claim 1 is characterized in that: In step 1, the machine-soil coupled contact mechanics model: F ED =F E +F V +F B , Where, F E F represents the digging resistance generated by the excavated star soil. V represents the inertial resistance related to the digging rate, F B represents the digging resistance associated with the bucket edge effect; F E =F EP +F q , F q =qHN q b, ψ represents the angle between the tangent line of the bucket surface point corresponding to the microelement dz and the vertical direction, δ represents the friction angle between the bucket and the lunar soil, β represents the angle between the slope of the lunar soil surface and the bucket, H is the excavation depth, represents the internal friction angle of the soil, q represents the equivalent uniformly distributed load generated by the excavated soil, V e represents the digging rate, θ * is the angle between the actual velocity direction of the soil layer and the horizontal direction, L α is the horizontal projection length of the shovel, γ is the soil density, α is the shovel digging angle, L is the shovel digging distance, b is the width of the bucket, z is the width of the bucket, N q is the influence coefficient of the additional distributed load on the soil surface, g is the local gravity acceleration, K b is the lateral excavation resistance correction coefficient of the shovel body at depth z, and the soil parameters are: δ and γ, shovel excavation conditions are: α, V e , H and L.
3. The distillation ensemble learning soil parameter inversion method for extraterrestrial object safe sampling according to claim 2 is characterized in that: In step 4, the digging resistance of the soil to be predicted is collected using a six-dimensional force sensor mounted at the connection between the robotic arm and the bucket.
4. The distillation ensemble learning soil parameter inversion method for safe sampling of extraterrestrial objects according to claim 2 or 3 is characterized in that: In step 4, the excavation conditions of the soil to be predicted are collected, specifically: The forward kinematics model estimates the end position of the robotic arm based on the joint angle information of the robotic arm at each moment, and combines the relative position relationship between the robotic arm and the soil surface to be predicted to estimate the shoveling condition of the soil to be predicted at each moment.
5. The distillation ensemble learning soil parameter inversion method for extraterrestrial object safe sampling according to claim 2 is characterized in that: The Monte Carlo method is used to randomly generate multiple groups of shoveling conditions from the preset shoveling condition range.
6. The distillation ensemble learning soil parameter inversion method for extraterrestrial object safe sampling according to claim 5 is characterized in that: The preset shoveling working condition range is shoveling angle α range of 30° to 75°, shoveling rate V e The range is 0.001m / s-0.005m / s, the shovel digging distance L range is 0.1m-0.5m, and the digging depth H range is 0.01-0.03m.