A complex spacecraft target geometry modeling method and device based on a modular implicit neural representation

Abstract

This invention discloses a method and apparatus for geometric modeling of complex spacecraft targets based on modular implicit neural representation. By dividing the overall structure into multiple independent functional sub-modules according to the functional attributes and geometric features of the spacecraft, and constructing a signed distance function (SDF) model for each sub-module, high-precision implicit modeling of local geometry is achieved. After independent training of the SDF, by applying pose parameter transformations and coordinate mappings to each module, the complex overall shape of the original spacecraft can be accurately reconstructed. This method not only enables continuous, differentiable, and real-time geometric representation of spacecraft during dynamic structural changes, but also allows for local updates of only the corresponding sub-modules when components are added or replaced, improving modeling efficiency and engineering practicality. This invention is particularly suitable for the geometric description and process modeling needs of complex spacecraft structures with multiple module combinations, multiple deployment mechanisms, and multiple auxiliary modules.

Description

Technical Field

[0001] This invention relates to the field of spacecraft on-orbit servicing and space target modeling, specifically to a method and apparatus for geometric modeling of complex spacecraft targets based on modular implicit neural representation. It is particularly suitable for space target modeling, state perception, and on-orbit rendezvous and servicing missions with features of multi-component combination, multi-mechanism deployability, and multi-module separability. Background Technology

[0002] With the rapid development of complex spacecraft such as space stations, on-orbit refueling platforms, space telescopes, and large combined structures, their structural forms are becoming increasingly diverse, and the number of external auxiliary modules is constantly increasing, including various functional units such as solar panels, communication antennas, robotic arms, experimental modules, and cargo spacecraft. These modules not only have complex shapes during on-orbit operation but also frequently undergo dynamic configuration changes such as unfolding, folding, rotating, separating, and replacing, posing a significant challenge to the high-precision, real-time mathematical description of the overall spacecraft shape. In mission scenarios such as on-orbit inspection, close-range maintenance, contact repair, autonomous assembly of combined structures, and rendezvous and docking, service spacecraft must rely on accurate target spacecraft geometric models to achieve accurate path planning, reliable collision avoidance, and safe and stable mission operations. Therefore, constructing a geometric modeling method that can accurately describe complex geometry and update in real time with configuration changes has become crucial to ensuring the safety and autonomy of on-orbit missions.

[0003] Existing target geometry modeling methods mainly include explicit representations based on polyhedral models, simplified geometric approximations based on ellipsoids or capsules, and methods that estimate the range field using point clouds or depth maps. However, these methods all have significant shortcomings. Explicit polyhedral models require reconstruction or complex topological modifications during module deployment or compartment separation, making it difficult to meet the real-time requirements of on-orbit servicing missions. Modeling methods based on parametric simplified geometry cannot accurately approximate the shape of real spacecraft with concave and convex details, sharp corner structures, or large-scale differences, often resulting in large collision detection errors or overly conservative safety boundaries. While holistic neural implicit representations can describe complex geometry in a continuously differentiable form, they rely on a single network to model the complete structure, limiting their expressive power for spacecraft with numerous modules and complex combinations. When any local structure changes, the entire system needs to be retrained, failing to meet the requirements of flexibility and engineering practicality.

[0004] Signed distance function (SDF) models based on implicit neural representations have attracted attention in recent years. Through asymmetric losses, they can achieve over-approximation of surfaces, providing a feasible path for constructing safe envelopes. However, these studies generally employ monolithic network structures, which cannot effectively handle the modularity of spacecraft or flexibly reflect changes in position, attitude, and scale among different modules. When simultaneously approximating multiple components with vastly different scales and geometric complexities, the gradient propagation and expressive power of a single network are significantly limited, leading to increased training difficulty, gradient degradation, uncontrollable local errors, and the inability to independently update local modules. This hinders their application in real spacecraft geometric modeling.

[0005] Given the increasing diversity and frequent changes in the geometry of complex spacecraft, a new geometric modeling paradigm is urgently needed. This paradigm should be able to naturally decompose the overall structure of a spacecraft into multiple independent functional modules and represent each module implicitly using neural networks. It should be able to unify and integrate these modules through attitude transformation, position mapping, and scale adjustment, thereby forming a dynamically updatable overall implicit geometric model. Furthermore, it should provide continuously differentiable SDF values ​​and gradients, providing a mathematical foundation for on-orbit path planning, collision avoidance, obstacle function construction, and control algorithms. Simultaneously, it should avoid repetitive training of large-scale models, enabling flexible module-level replacement, expansion, and local correction, and possess the ability to approximate complex concave and convex structures and multi-scale details, making it truly suitable for practical tasks such as on-orbit inspection, maintenance interaction, space construction, and component docking or separation.

[0006] Against this backdrop, there is an urgent need to propose a geometric modeling method for complex spacecraft targets based on modular implicit neural representations. This method should simultaneously meet the requirements of high-fidelity geometric fitting, independent module-level management, dynamic configuration updates, and engineering practicality, thereby overcoming the shortcomings of existing modeling techniques in terms of adaptability to complex geometry and matching with mission requirements. This invention addresses this core technological bottleneck by proposing an innovative modular implicit neural representation framework to achieve accurate, flexible, and scalable mathematical descriptions of complex spacecraft structures. Summary of the Invention

[0007] This invention aims to address the shortcomings of existing spacecraft geometric modeling methods, such as insufficient expressive power, high update costs, and difficulty in providing continuously differentiable geometric information when dealing with complex structures, multi-module assemblies, and dynamic configuration changes. It proposes a method and apparatus for geometric modeling complex spacecraft targets based on modular implicit neural representation. This method decomposes a complex spacecraft into multiple functionally and geometrically independent sub-modules, trains a signed distance function (SDF) network for each sub-module, and introduces pose parameterization transformation and smoothing fusion operators to construct a time-varying, scalable, and replaceable overall implicit geometric model. Furthermore, considering the inclusion of gradient and higher-order derivative regularization terms in the loss function, this invention systematically designs an efficient gradient calculation and propagation mechanism, providing stable and reliable geometric derivative information for subsequent SDF-based collision avoidance constraint construction and controller design.

[0008] The technical solution of this invention is as follows: A method for geometric modeling of complex spacecraft targets based on modular implicit neural representation, the method comprising the following steps:

[0009] The first step is to divide the overall structure into several independent functional modules based on the functional composition and geometric characteristics of the target spacecraft.

[0010] The second step is to construct the local geometric data of the modules: generate local point cloud datasets for each module based on its real or simulated geometric model.

[0011] The third step is to establish a modular neural SDF representation: train a neural signature distance function SDF network for each module separately. The network is trained using a loss function that includes an approximation strategy, so that the zero-level set of the modules forms an outer envelope, resulting in a conservative geometric approximation for security control.

[0012] The fourth step is to establish module-level pose transformation: define the pose transformation function of each module relative to the spacecraft's global coordinate system;

[0013] The fifth step is to fuse the SDFs of each module to construct the overall implicit surface: the Smooth-Min function is used to continuously and differentiably fuse the SDFs of each module to obtain the overall implicit geometric expression of the target spacecraft, which is used to describe the dynamic changes of the overall shape and configuration of the spacecraft over time.

[0014] Step 6, Dynamic Update of Complex Structures: When the spacecraft undergoes configuration changes, the overall geometric model is updated in real time by updating the attitude parameters of the corresponding modules or removing the corresponding modules.

[0015] Furthermore, the neural network The following loss function is used for training:

[0016]

[0017] in: This represents the position error loss term. This represents the Eikonal regularization term. Represents the curvature regularization term; These are the weight parameters corresponding to each item.

[0018] Furthermore, the position error loss term An asymmetric penalty strategy is employed to cause the SDF prediction to be biased towards the outside and over-approximated:

[0019]

[0020] in, These are spatial sampling points in the local coordinate system. The signed distance value from the spatial sampling point to the real surface of the k-th module is determined by adjusting... By controlling the ratio of positive to negative errors, a safe outer envelope is formed by the zero-level set.

[0021] Furthermore, Eikonal regularization terms as follows:

[0022]

[0023] Make the gradient of the neural network function satisfy the SDF property and maintain geometric realism.

[0024] Furthermore, the curvature regularization term as follows:

[0025]

[0026] The curvature regularization term keeps the second derivative in the normal direction zero to avoid excessive smoothing that could lead to geometric deformation. This represents the Hessian matrix.

[0027] Furthermore, in the fifth step, the overall implicit geometric expression of the target spacecraft is:

[0028]

[0029] in, It is a smooth minimum operator used to fuse the SDF of each module while maintaining differentiability, so that the overall implicit geometry has sufficient smoothness and consistency at the module connection.

[0030] Secondly, the present invention also provides a complex spacecraft target geometry modeling device based on modular implicit neural representation, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the aforementioned complex spacecraft target geometry modeling method based on modular implicit neural representation.

[0031] Thirdly, the present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the aforementioned method for geometric modeling of complex spacecraft targets based on modular implicit neural representation.

[0032] Fourthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for geometric modeling of complex spacecraft targets based on modular implicit neural representation.

[0033] The beneficial effects of this invention are:

[0034] Through the above technical solution, this invention forms a complex spacecraft geometric modeling method that combines high-fidelity geometric fitting capability, module-level maintainability, and dynamic update capability. Compared with existing global INR or explicit geometric models, the advantages of this invention are:

[0035] Firstly, the modular SDF network enables independent training and replacement of each component, avoiding the need to retrain the entire model due to changes in local structure, thus significantly reducing update costs.

[0036] Secondly, through parameterized pose transformation and smooth fusion operators, a unified expression of the attitude, position and scale of multi-module assemblies is achieved, which can adapt to various on-orbit configuration change scenarios.

[0037] Third, by explicitly designing loss terms related to gradients and higher-order derivatives and combining them with analytical and automatic differentiation calculation mechanisms, stable and reliable SDF derivative information was obtained, providing a solid geometric foundation for subsequent control and planning.

[0038] In summary, this invention systematically constructs a new paradigm for geometric modeling of complex spacecraft targets based on modular implicit neural representation from three levels: structural design, loss construction, and gradient calculation. This effectively overcomes many limitations of traditional methods in complex geometric representation and engineering applications. Attached Figure Description

[0039] Figure 1 A flowchart of a modular implicit neural representation method for complex spacecraft;

[0040] Figure 2This method includes a schematic diagram of the modular breakdown of the Chinese space station;

[0041] Figure 3 A schematic diagram of the individual implicit neural representations of the split modules;

[0042] Figure 4 An implicit representation of the Chinese space station obtained through attitude transformation isomorphism;

[0043] Figure 5 An implicit representation of the Chinese space station after the cargo spacecraft has detached;

[0044] Figure 6 A schematic diagram representing the motion of a solar panel;

[0045] Figure 7 A schematic diagram illustrating the docking process of a cargo spacecraft;

[0046] Figure 8 This is a structural diagram of a complex spacecraft target geometry modeling device based on modular implicit neural representation, according to the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0048] In a preferred embodiment of the present invention, such as Figure 1 As shown, a method for geometric modeling of complex spacecraft targets based on modular implicit neural representation is presented. This method is used for high-precision, scalable, and dynamically updatable geometric modeling of target spacecraft with complex shapes and multiple functional components. The overall process includes: target module partitioning and local coordinate system establishment, generation of module local SDF data, construction and training of module neural SDF network, calculation of gradient and higher-order derivatives, module assembly and implicit geometric fusion based on pose transformation, and dynamic updating of complex configurations.

[0049] In terms of target module division and coordinate system establishment, such as Figure 2As shown, based on the structural composition and functional partitioning of the target spacecraft, its overall geometry is first divided into multiple functional modules, such as the main cabin module, solar panel module, robotic arm module, antenna and equipment module, and cargo spacecraft segment module. Each module is considered a geometrically and functionally independent sub-body, with relatively stable structural features within its local area. For each module, a local coordinate system aligned with its principal axis of inertia or structural features is established, with the origin selected as the module's centroid, geometric center, or mounting base location. This modular division allows each sub-module to be modeled by an independent implicit neural network, simplifying the representational burden of a single network and providing a structural basis for subsequent local updates and replacements.

[0050] Regarding the generation of local SDF data for each module, point cloud datasets are generated offline in their local coordinate system based on the CAD model, mesh model, or high-precision measurement data of the target spacecraft. Figure 3 As shown. Specifically, for the k-th module, its SDF training dataset is constructed.

[0051] (1)

[0052] in, These are spatial sampling points in the local coordinate system. It represents the signed distance value from the spatial sampling point to the real surface of the k-th module (the point is positive when it is outside, negative when it is inside, or vice versa). Let be the number of sampling points for the k-th module. The sampling area not only covers the module surface but also the surrounding operational domain of a certain thickness to ensure that the network has sufficient interpolation and extrapolation capabilities near the surface. The sampling method can employ a strategy combining uniform grid sampling and dense sampling near the surface to balance global coverage and local accuracy.

[0053] In terms of constructing and training modular neural SDF networks, this invention employs a multilayer perceptron (MLP) as a function approximator for each module's SDF. For the k-th module, its implicit surface is output by the neural network. The description states that the network input consists of points in a local coordinate system. The output is a scalar SDF value. The neural network structure of each module can be the same, or it can be adjusted appropriately according to geometric complexity. Taking a typical structure as an example, the network consists of several fully connected layers stacked together, each layer containing several hidden units. The activation function is a differentiable nonlinear function, such as... Smooth variants of ReLU or SIREN-like sinusoidal activations, etc. To obtain an SDF with conservative over-approximation properties, this invention constructs the following comprehensive loss function for each module:

[0054] (2)

[0055] in, For the spatial location error term, a conservative over-approximation of SDF is achieved through asymmetric weighting design, making the predicted value more sensitive to positive errors, thereby pushing the zero level set outward and constructing a safe outer envelope; For Eikonal regularization terms, through constraints Approaching 1 enhances the physical consistency and geometric fidelity of SDF; For the second derivative regularization term based on the normal direction, by controlling The size of the parameters is adjusted to avoid excessive smoothing in high curvature regions while suppressing numerical oscillations. The three types of sub-losses mentioned above are respectively determined by the weighting parameters. Weighting is applied to balance the relationship between fitting accuracy, geometric consistency, and numerical stability.

[0056] Position error loss term An asymmetric penalty strategy is employed to make the SDF prediction biased towards the outside and over-approximate:

[0057]

[0058] By adjusting By controlling the ratio of positive to negative errors, a safe outer envelope is formed by the zero-level set.

[0059] Eikonal regularization terms as follows:

[0060]

[0061] Make its gradient satisfy the SDF property to maintain geometric realism.

[0062] Curvature regularization term as follows:

[0063]

[0064] Keep the second derivative in the normal direction zero to avoid excessive smoothing that could lead to geometric deformation.

[0065] Since the aforementioned loss function explicitly involves the gradient of the network output with respect to the input coordinates and a Hessian-based combination term, this embodiment of the invention further provides an efficient gradient calculation method. For example... Figure 4 As shown, let the network be... The output of the layer is The inter-layer mapping of the network can then be represented as:

[0066] (3)

[0067] in, This is the weight matrix. For bias vectors, Let be an element-wise differentiable activation function. Using the chain rule, we can obtain the output with respect to the input coordinates. Gradient recurrence relation:

[0068] (4)

[0069] in, The first derivative of the activation function. Records the current layer's output to input The Jacobian information. By recursively applying the formulas at each layer, the network output can ultimately be obtained. about gradient Furthermore, a Hessian-related regularization term is constructed based on this. In practical engineering implementation, this invention combines the automatic differentiation function of the deep learning framework, embedding the above analytical recursive formula into the backpropagation process, thereby efficiently calculating the loss term involving the first and second derivatives and ensuring the numerical stability of the training process.

[0070] After achieving module-level implicit representation, this invention further reconstructs and dynamically updates the overall geometry of the target spacecraft through pose transformation and smooth fusion. For example... Figure 4 As shown, let the global coordinate system be the target spacecraft's body coordinate system or the mission reference coordinate system. For any spatial point at time t... The transformation of its mapping to the local coordinate system of the k-th module is denoted as:

[0071] (5)

[0072] Where x represents the input point, This is a scale factor used to match components of different sizes; The quaternion describes the attitude of this component relative to the target spacecraft's own system. for The corresponding rotation matrix; This is the translation vector of the module relative to the target spacecraft's own system. By adjusting... and This can represent various configuration changes such as the deployment and folding of solar panels, the pointing adjustment of antennas, the multi-joint movement of robotic arms, and the docking and separation of cargo spacecraft.

[0073] Based on this, the present invention uses a smooth minimum operator to fuse the SDF of each module to construct the overall implicit surface function of the target spacecraft:

[0074] (6)

[0075] in, To achieve a smooth approximate minimum function, for example, a Log-Sum-Exp form can be used to fuse the SDFs of each module in a continuously differentiable manner. By letting... This allows us to obtain the overall surface of the target spacecraft at time t, where the points represent whether the points are located on the outer or inner side of the surface. Because... , as well as All are differentiable functions, and are global implicit surface functions. right The gradient can also be obtained through the chain rule and automatic differentiation, thus providing continuously differentiable geometric information for subsequent applications such as path planning, collision detection, and safety control based on control obstacle functions.

[0076] In a specific application embodiment of the present invention, for a complex target consisting of a space station assembly and a cargo spacecraft, the various core modules of the space station and one cargo spacecraft segment are first treated as six independent modules, and corresponding neural SDF networks are constructed and offline training is completed for each. During normal operation, the pose parameters of the cargo spacecraft module are set, it is attached to the docking port position of the space station, and the overall geometry of the system is reconstructed through the aforementioned smooth fusion method. Figure 5 As shown, when the cargo spacecraft is preparing to separate, its pose parameters only need to be updated in a time series along the separation trajectory to obtain the overall implicit geometric changes of the combined structure during the transition from "docked state" to "separated state" in real time, without requiring any retraining of the networks of other modules. When the solar panels or antennas adjust their attitude, the new configuration can also be expressed by updating the quaternions and position vectors of the corresponding modules, realizing a mathematical and continuous description of the configuration changes of complex spacecraft combinations. Figure 6 As shown, this can be used to represent the motion process of a solar panel evolving over time; such as Figure 7 As shown, this can be used to represent the docking and separation process of a cargo spacecraft relative to the main body of the space station.

[0077] Corresponding to the aforementioned embodiment of a method for geometric modeling complex spacecraft targets based on modular implicit neural representation, the present invention also provides an embodiment of a device for geometric modeling complex spacecraft targets based on modular implicit neural representation.

[0078] See Figure 8 The present invention provides a complex spacecraft target geometry modeling device based on modular implicit neural representation, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement a complex spacecraft target geometry modeling method based on modular implicit neural representation in the above embodiment.

[0079] The embodiment of the complex spacecraft target geometry modeling device based on modular implicit neural representation provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 8 The diagram shown is a hardware structure diagram of any data processing-capable device, which is part of a complex spacecraft target geometry modeling device based on modular implicit neural representation provided by the present invention. (Except for...) Figure 8 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0080] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0081] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0082] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements a method for geometric modeling of complex spacecraft targets based on modular implicit neural representation as described in the above embodiments.

[0083] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0084] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for geometric modeling of complex spacecraft targets based on modular implicit neural representation.

[0085] The number of network layers, hidden units, activation function forms, specific implementations of the smooth minimum operator, numerical calculation details of gradients and Hessians, and module partitioning strategies described in the embodiments of this invention can all be adjusted and optimized according to specific task requirements and engineering constraints, and do not constitute a limitation on the scope of protection of this invention. Contents not described in detail in this specification are prior art known to those skilled in the art, and any equivalent substitutions and modifications made without departing from the concept of this invention should fall within the scope of protection of this invention.

Claims

1. A method for geometric modeling of complex spacecraft targets based on modular implicit neural representation, characterized in that, The method Includes the following steps: The first step is to divide the overall structure into several independent functional modules based on the functional composition and geometric characteristics of the target spacecraft. The second step is to construct the local geometric data of the modules: generate local point cloud datasets for each module based on its real or simulated geometric model. The third step is to establish a modular neural SDF representation: train a neural signature distance function SDF network for each module. The network is trained using a loss function that includes an approximation strategy, so that the zero-level set of the module forms an outer envelope, and a conservative geometric approximation for security control is obtained. The fourth step is to establish module-level pose transformation: define the pose transformation function of each module relative to the spacecraft's global coordinate system; The fifth step is to fuse the SDFs of each module to construct the overall implicit surface: the Smooth-Min function is used to continuously and differentiably fuse the SDFs of each module to obtain the overall implicit geometric expression of the target spacecraft, which is used to describe the dynamic changes of the overall shape and configuration of the spacecraft over time. Step 6, Dynamic Update of Complex Structures: When the spacecraft undergoes configuration changes, the overall geometric model is updated in real time by updating the attitude parameters of the corresponding modules or removing the corresponding modules.

2. The method according to claim 1, characterized in that: The loss function of the neural signature distance function SDF network includes a position error loss term, an Eikonal regularization term, and a curvature regularization term.

3. The method according to claim 2, characterized in that: The position error loss term employs an asymmetric penalty strategy to make the SDF prediction biased to the outside and too close, and controls the ratio of positive to negative errors to make the zero level set form a safe outer envelope.

4. The method according to claim 2, characterized in that: The Eikonal regularization term ensures that the gradient of the neural network function satisfies the SDF property, preserving geometric realism.

5. The method according to claim 2, characterized in that: The curvature regularization term keeps the second derivative in the normal direction zero to avoid excessive smoothing that could lead to geometric deformation.

6. The method according to claim 2, characterized in that: In the fifth step, the target spacecraft implements implicit geometric expression based on the smooth minimum operator. While maintaining differentiability, the SDF of each module is fused to make the overall implicit geometry have sufficient smoothness and consistency at the module connection.

7. A complex spacecraft target geometry modeling device based on modular implicit neural representation, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a method for geometric modeling of complex spacecraft targets based on modular implicit neural representation as described in any one of claims 1-6.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for geometric modeling of complex spacecraft targets based on modular implicit neural representation as described in any one of claims 1-6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for geometric modeling of complex spacecraft targets based on modular implicit neural representation as described in any one of claims 1-6.

Images