A part deformation analysis system based on three-dimensional measurement data driving

Abstract

The present application relates to the technical field of three-dimensional measurement and digital twin assembly analysis, and particularly relates to a part deformation analysis system based on three-dimensional measurement data driving, comprising: an anisotropic feature decoupling unit: obtaining sparse point cloud data of a flexible target object, and generating a purified discrete feature set; a geometric hallucination monitoring unit: generating a hallucination entropy distribution map by using an implicit neural representation technology; an adaptive topological fidelity arbitration unit: generating a final deformation analysis report based on the hallucination entropy distribution map; a deformation logic iteration unit: dynamically adjusting weight parameters of a counter-deformation decoupling model and a calculation strategy of the geometric hallucination entropy according to the final deformation analysis report and downstream assembly feedback; the present application ensures the purity of input data and effectively avoids deformation omissions.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional measurement and digital twin assembly analysis technology, specifically to a part deformation analysis system driven by three-dimensional measurement data. Background Technology

[0002] Three-dimensional measurement technology is often used in the automated three-dimensional measurement and digital twin assembly analysis of flexible composite material skins for spacecraft. Its technical essence is to acquire point cloud data of flexible target objects and perform three-dimensional completion in order to output a three-dimensional reconstruction model for deformation assessment.

[0003] Existing 3D data processing methods typically rely on basic optical reflection models to directly filter and remove data in the spatial domain, or use implicit neural representation techniques to perform generative smoothing to complete unmeasured areas. However, in extremely fast production cycles and limited measurement environments, sensors often only acquire sparse point cloud data. At the same time, due to the anisotropic properties of composite materials, nonlinear optical artifacts are easily caused. Tiny real manufacturing defects and optical artifacts both exhibit highly similar local abrupt changes in 3D spatial geometry. If filtering is performed directly in the spatial domain, it is easy to mistakenly remove real tiny deformations as noise.

[0004] Furthermore, when dealing with measurement blind spots such as assembly edges, existing technologies often automatically interpolate to generate distorted data based on excessively smoothed fitting of prior experience in order to optimize the reconstruction integrity index. This excessive smoothing can mask the physical fragmentation features and abrupt change nodes. Currently, for high-precision business scenarios such as spacecraft skin assembly, a local feature skeleton with extremely high confidence is required to prevent assembly interference. Therefore, how to accurately remove optical artifacts and prevent topological distortion caused by excessive smoothing of algorithms under the interference of sparse point cloud data and the optical properties of anisotropic materials, so as to achieve high-fidelity part deformation analysis, has become a problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a part deformation analysis system driven by three-dimensional measurement data, and to solve the following technical problems:

[0006] This solution addresses the challenges of sparse point cloud data caused by extremely fast production cycles and constrained measurement environments, as well as nonlinear interference caused by the optical properties of anisotropic materials. It precisely removes nonlinear optical artifacts caused by material anisotropy, avoiding the misinterpretation of real, minute deformations as noise. Furthermore, by establishing a transparent monitoring system for data generation confidence, it prevents the blind trust in distorted data resulting from overly smoothed fitting by algorithms. This proactively ensures the fidelity of topological features in critical assembly areas, preventing sealing failures caused by algorithm overconfidence.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A part deformation analysis system driven by three-dimensional measurement data includes: an anisotropic feature decoupling unit: acquiring sparse point cloud data of a flexible target object and analyzing the surface texture features of the sparse point cloud data; identifying nonlinear optical artifacts caused by anisotropic material properties based on a preset material optical reflection model; and constructing an adversarial deformation decoupling model to separate the nonlinear optical artifacts from the real geometric deformation and generate a purified discrete feature set.

[0009] Geometric illusion monitoring unit: Utilizes implicit neural representation technology to perform generative 3D completion on the discrete feature set and constructs a candidate 3D reconstruction model; Simultaneously calculates the geometric illusion entropy during the reconstruction process, quantifies the degree of excessive smoothness and confidence decay of the candidate 3D reconstruction model in the unmeasured area, and generates an illusion entropy distribution map;

[0010] Adaptive Topology Fidelity Arbitration Unit: Based on the illusion entropy distribution map, detect the entropy status of key topological regions; if the entropy status is lower than a preset safety threshold, output the complete candidate 3D reconstruction model; if the entropy status is higher than the preset safety threshold, trigger the fidelity circuit breaker mechanism, abandon the generation of the complete model, and instead output a visually discrete but topologically reliable feature skeleton based on the discrete feature set to generate the final deformation analysis report.

[0011] Deformation logic iteration unit: Based on the final deformation analysis report and downstream assembly feedback, construct a deformation feature verification closed loop; quantify the false deformation misjudgment rate, and dynamically adjust the weight parameters of the adversarial deformation decoupling model and the calculation strategy of the geometric illusion entropy.

[0012] Preferably, the method of constructing an adversarial deformation decoupling model to separate the nonlinear optical artifacts from the real geometric deformation includes: extracting high-frequency noise signals from the sparse point cloud data and mapping them to the frequency domain feature space; loading a pre-set anisotropic material texture library and matching the texture response mode under the current illumination angle;

[0013] In the frequency domain feature space, a contrastive learning algorithm is used to distinguish between optical artifact signals that conform to texture response patterns and real deformation signals that conform to manufacturing defect characteristics; data points marked as optical artifact signals are suppressed while real deformation signals are retained, thereby generating the purified discrete feature set.

[0014] Preferably, the methods for calculating the geometric illusion entropy during the reconstruction process and quantifying the degree of oversmoothing include: defining reconstruction integrity index and deformation sensitivity index as mutually exclusive constraints;

[0015] In the derivation process of the implicit neural representation technology, the Euclidean distance variance between the generated data points and the discrete feature set is calculated; the Euclidean distance variance is combined with the local curvature change rate to obtain the geometric uncertainty coefficient; the geometric uncertainty coefficient is accumulated in the time dimension to generate the geometric illusion entropy, which is used to characterize the degree of fiction when the AI ​​model completes the unobserved region.

[0016] Preferably, the method for detecting the entropy state of the key topological region and triggering the fidelity circuit breaker mechanism includes: dividing the candidate 3D reconstruction model into multiple topological functional sub-regions, identifying the assembly edge region as the key topological region; and extracting the local average illusion entropy value within the key topological region.

[0017] Determine whether the local average illusion entropy value is in a high-risk range; if it is in a high-risk range, determine that there is a risk of topology locking failure, immediately block the smooth completion operation for this region, and lock the original discrete feature set as the output benchmark to preserve the physically real broken features.

[0018] Preferably, the method of outputting a visually discrete but topologically reliable feature skeleton based on the discrete feature set includes: constructing a non-manifold geometric topological mesh based on the discrete feature set while abandoning the reconstruction of the complete surface;

[0019] Key points are labeled on the non-manifold geometric topology mesh to generate a discrete point cloud set containing fracture edges and abrupt nodes; the discrete point cloud set is then mapped to a reliable feature skeleton that does not have visual continuity but mathematically satisfies GD&T constraints.

[0020] Preferably, the system further includes a gravity shape compensation module for handling the gravity masking effect: acquiring measured shape data of the flexible target object under gravity environment;

[0021] Based on the finite element physical simulation engine, the theoretical relaxation state of the flexible target object under weightlessness is simulated; the elastic deformation difference between the measured shape data and the theoretical relaxation state is calculated; the elastic deformation difference is injected as a correction factor into the adversarial deformation decoupling model to eliminate the steady-state interference of gravity on deformation analysis.

[0022] Preferably, the deformation logic iteration unit constructs a deformation feature verification closed loop by: collecting physical interference data during the actual assembly process as truth feedback;

[0023] The predicted deformation in the final deformation analysis report is compared with the physical interference data to calculate the false alarm rate and the missed detection rate.

[0024] Based on the false alarm rate and false alarm rate, a penalty factor for geometric illusion entropy is calculated; the penalty factor is then used to backpropagate and update the loss function of the implicit neural representation network, forcing the model to reduce its overconfidence in blurred regions in subsequent iterations.

[0025] Preferably, the data coverage of the sparse point cloud data is lower than a preset integrity threshold, and the flexible target object is a composite material skin structure with anisotropic reflective properties.

[0026] Preferably, the final deformation analysis report includes a three-dimensional model view and a confidence heatmap, wherein the confidence heatmap visually displays the distribution of the geometric illusion entropy on the surface of the part, and is used to indicate assembly risk areas.

[0027] Preferably, the system is deployed on an edge computing node, and the adaptive topology fidelity arbitration unit dynamically adjusts the sampling resolution of the implicit neural representation technology according to the production cycle requirements, so as to maintain a dynamic balance between computational efficiency and analysis accuracy.

[0028] The beneficial effects of this invention are:

[0029] 1) This invention extracts high-frequency noise signals from sparse point cloud data and maps them to the frequency domain. Combined with a pre-set anisotropic material texture library, it uses a contrastive learning algorithm to separate real deformations. This mechanism solves the problem that direct filtering in the spatial domain is prone to accidentally detecting small deformations. It accurately removes nonlinear optical artifacts caused by composite materials, ensures the purity of input data, and effectively avoids missing deformation detection.

[0030] 2) When using implicit neural representation technology to complete a 3D model, the system combines the distance variance between the generated points and the discrete feature set with the rate of curvature change to generate geometric illusion entropy. This mechanism effectively quantifies the degree of excessive smoothing and confidence decay of the model in the measurement blind zone, preventing downstream businesses from blindly trusting the distorted geometric artifacts speculated by the algorithm.

[0031] 3) This invention detects the state of key topological regions based on the illusion entropy distribution map. When the entropy value is in the high-risk range, it actively triggers the fidelity-preserving circuit breaker mechanism and abandons the generation of a complete smooth surface. Instead, the system outputs a topologically reliable discrete feature skeleton, which solves the problem of the algorithm's excessive smoothing masking the physical fragmentation features, ensures the fidelity of the assembly boundary, and avoids topology locking failure.

[0032] 4) This invention uses a physical simulation engine to simulate the relaxed state of the target object in a weightless state, calculates the difference in elastic deformation between the target object and the actual measured state under ground gravity, and injects it into the model as a correction factor. This mechanism eliminates the systematic interference of natural sagging on ground measurements, prevents the real small deformation from being masked by the gravity effect, and reflects the real space topology of the part.

[0033] 5) This invention collects physical interference feedback in actual assembly, compares the predicted deformation to calculate the ratio of missed detections and false alarms, and generates a penalty factor to update the loss function of the implicit neural representation network in reverse. This closed-loop iterative mechanism forces the model to reduce its overconfidence in unmeasured areas, continuously and dynamically calibrate under the normalization of incomplete data, and maintain a very high judgment accuracy.

[0034] 6) This invention is deployed on edge computing nodes and adaptively and dynamically adjusts the sampling resolution of the simulation according to the production line cycle requirements and remaining available time; it performs dimensionality reduction sampling in non-critical areas and tilts the core computing power to high-risk critical topology areas, which solves the problem of excessive time consumption caused by high-precision calculation across the entire domain and achieves the optimal balance between delivery timeliness and core accuracy. Attached Figure Description

[0035] The invention will now be further described with reference to the accompanying drawings.

[0036] Figure 1 This is a schematic diagram of a part deformation analysis system driven by three-dimensional measurement data, provided in an embodiment of this application. Detailed Implementation

[0037] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Please see Figure 1 This describes the operating mechanism of a part deformation analysis system driven by three-dimensional measurement data. Specifically, the system is deployed in the scenario of automated three-dimensional measurement and digital twin assembly analysis of flexible composite material skin for spacecraft. It is used to solve the problem of sparse point cloud data caused by extremely fast production cycles and limited measurement environments, as well as nonlinear interference caused by the optical properties of anisotropic materials. The system is collaboratively constructed by an anisotropic feature decoupling unit, a geometric illusion monitoring unit, an adaptive topology fidelity arbitration unit, and a deformation logic iteration unit.

[0039] This embodiment provides a mechanism for anisotropic feature decoupling; specifically, the anisotropic feature decoupling unit acquires sparse point cloud data of a flexible target object and analyzes the surface texture features of the sparse point cloud data; based on a preset material optical reflection model, it identifies nonlinear optical artifacts caused by anisotropic material properties;

[0040] However, when dealing with flexible composite materials with extremely uneven surfaces, tiny real manufacturing defects and optical artifacts both exhibit highly similar local abrupt changes in three-dimensional spatial geometry. If we rely solely on the basic optical reflection model to directly filter and eliminate them in the spatial domain, it is very easy to mistakenly identify the real tiny deformations as noise and kill them.

[0041] Therefore, this embodiment constructs an adversarial deformation decoupling model to separate nonlinear optical artifacts from real geometric deformations; furthermore, this separation process extracts high-frequency noise signals from sparse point cloud data through fast Fourier transform and maps them to the frequency domain feature space;

[0042] Load a pre-set anisotropic material texture library and match the texture response pattern under the current illumination angle; in the frequency domain feature space, use a contrastive learning algorithm to distinguish between optical artifact signals that conform to the texture response pattern and real deformation signals that conform to manufacturing defect characteristics.

[0043] When constructing the contrastive learning algorithm, the system constructs a feature extraction network containing multiple one-dimensional convolutional layers and fully connected layers. The frequency domain features of known optical artifact signals are extracted and output as positive sample set feature vectors through the feature extraction network. The frequency domain features of real small deformations verified by actual measurements are extracted and output as negative sample set feature vectors through the feature extraction network.

[0044] In the frequency domain feature space, the current high-frequency signal to be tested is input into the feature extraction network and the current feature vector is output. The cosine similarity between the current feature vector and the anchor points of the positive and negative samples is calculated.

[0045] If the similarity between the current signal and the positive sample anchor point is greater than its similarity with the negative sample anchor point, and the similarity between the frequency domain distribution and the texture response pattern is greater than the preset artifact threshold, then the data point marked as an optical artifact signal is suppressed.

[0046] If the similarity is less than or equal to the preset artifact threshold, the real deformation signal is retained, thereby generating a purified discrete feature set. Furthermore, the preset artifact threshold is obtained by conducting a standard illumination calibration experiment on the batch of composite materials in a laboratory environment and statistically analyzing the lower quartile of the similarity distribution of pure optical artifact samples.

[0047] The preset safety threshold is determined based on the statistical average of the maximum local entropy values ​​that did not occur in successful assembly cases of similar spacecraft in the past. This safety threshold serves as a unified benchmark for defining high-risk intervals in the underlying logic of the system, namely the high-risk interval threshold discussed later. This is used to globally determine whether the current geometric illusion entropy exceeds the limit in order to trigger the fidelity circuit breaker mechanism.

[0048] When extracting positive and negative sample anchor points, positive sample anchor points are obtained by average pooling the frequency domain features of optical artifacts collected by applying specific illumination to a standard defect-free panel; negative sample anchor points are obtained by average pooling the frequency domain features collected by a panel containing known small deformations under uniform diffuse illumination.

[0049] For example, during the scanning process of carbon fiber reinforced composite skin, due to environmental limitations, the sensor only captured about 60% of the sparse point cloud data. Because of the cross-woven texture on the surface, under specific light source illumination, a large number of high-frequency protrusions with an amplitude of 0.15 mm appeared in the point cloud matrix. After converting these high-frequency signals to the frequency domain, the system discovered the frequency components within them. If the match rate with the reflectance feature map of a specific lighting angle in the texture library reaches 94%, which is greater than the artifact threshold of 85%, then... It is identified as a nonlinear optical artifact and its coordinate shift in the spatial domain is suppressed;

[0050] And another frequency component If the matching degree is only 28% and there is a local low-frequency curvature change, it is judged as a real micro bending deformation defect and retained. The purpose of this mechanism is to accurately remove the nonlinear optical artifact interference caused by material anisotropy through frequency domain transformation and contrast learning, to ensure that the basic geometric data input to the subsequent analysis link does not contain false deformation, and at the same time achieve effective detection of fatal micro deformation defects without relying on excessive smoothing.

[0051] This embodiment provides a mechanism for monitoring geometric illusions; specifically, the geometric illusion monitoring unit uses implicit neural representation technology to perform generative 3D completion on the purified discrete feature set to construct a candidate 3D reconstruction model.

[0052] Specifically, a multilayer perceptron network is constructed, comprising an input layer, multiple hidden layers, a random deactivation layer, and an output layer, to convert spatial coordinates... As input, the signed distance field value of this coordinate is output to characterize the surface geometry;

[0053] Therefore, when the system infers the spatial coordinates of unmeasured areas using the network multilayer perceptron, it keeps the randomly deactivated layer in an active state, performs multiple forward propagations on the same input coordinates to construct multiple inference batches, extracts the surface spatial coordinate points generated by the coordinates in each inference, and calculates the Euclidean distance variance between the generated points in the multiple inference batches and the nearest real discrete feature points. The larger the variance, the higher the uncertainty of data generation, and the larger the corresponding geometric illusion entropy value.

[0054] Conversely, if the variance is extremely small and approximates the original discrete feature set, then the geometric illusion entropy value approaches zero.

[0055] For example, when performing three-dimensional completion on the assembly edge of a spacecraft skin, since this area is in the measurement blind zone, implicit neural representation technology automatically interpolates and generates a visually perfect continuous edge in order to optimize the reconstruction integrity index.

[0056] However, when the system extracted the inferred feature variance of the edge region, it calculated that its local geometric illusion entropy value reached 0.88, with a normalized full score of 1.0. The system also marked the region as a dark red high-risk area in the illusion entropy distribution map. This indicates that the smooth edge is very likely a distorted data that has been over-smoothed and fitted by the algorithm based on prior experience, and its confidence has been severely weakened.

[0057] The purpose of this step is to establish a transparent monitoring system for data generation confidence, quantify the boundary between the measured physical truth and the algorithm-predicted data in the 3D reconstruction model, and prevent the downstream assembly process from introducing mathematically perfect but physically non-existent geometric artifacts.

[0058] This embodiment provides an adaptive topology fidelity arbitration mechanism. Specifically, the adaptive topology fidelity arbitration unit detects the entropy status of key topological regions based on the illusion entropy distribution map. If the entropy status is lower than a preset safety threshold, it indicates that the confidence of the reconstructed data is within a controllable range, and the system outputs a complete candidate 3D reconstruction model normally. If the entropy status is higher than the preset safety threshold, the fidelity circuit breaker mechanism is triggered, the generation of a complete model is abandoned, and instead, a visually discrete but topologically reliable feature skeleton is output based on a discrete feature set to generate the final deformation analysis report.

[0059] For example, the safety threshold for geometric illusion entropy is set to 0.65. When the system scans a flexible skin panel that has undergone natural deformation due to Earth's gravity, it detects that the entropy value around its core assembly docking hole is as high as 0.92. At this time, if the complete candidate 3D reconstruction model continues to be output, it will cause a serious topology locking failure. That is, the virtual model that appears to fit smoothly under Earth's gravity will have tiny gaps that cannot be physically closed when actually assembled in the weightless state of space.

[0060] Therefore, the system immediately triggers the fidelity circuit breaker mechanism, forcibly interrupts the smooth completion process of the area, directly extracts the original discrete point cloud coordinates around the hole, generates a feature skeleton that is visually broken and discontinuous, but is absolutely realistic in terms of geometric dimensions and tolerance GD&T constraints, and submits the report to the digital twin assembly center.

[0061] The purpose of this mechanism is to make a mandatory system logic judgment between pursuing visually perfect 3D reconstruction and tolerating data fragmentation in order to preserve physical truth. By actively sacrificing the model integrity of unmeasured areas, it ensures the fidelity of the topological features of key assembly areas and avoids the risk of serious sealing failure caused by algorithm overconfidence.

[0062] This embodiment provides a deformation logic iteration mechanism; specifically, the deformation logic iteration unit constructs a deformation feature verification closure based on the final deformation analysis report and downstream assembly feedback; quantifies the false deformation misjudgment rate, and dynamically adjusts the weight parameters of the adversarial deformation decoupling model and the calculation strategy of geometric illusion entropy;

[0063] If downstream assembly feedback indicates that a certain batch of parts has physical interference, but the system fails to issue a warning in the report, it is determined that a missed detection has occurred; if the report warns that deformation hinders assembly, but the actual physical assembly closes smoothly, it is determined to be a false alarm; the system calculates the loss gradient based on the frequency of missed detections and false alarms, and updates the hyperparameters of each analysis unit in reverse.

[0064] For example, after the trial assembly of 500 spacecraft skin panels, the torque sensor of the assembly robotic arm reported four edge physical interference events. After receiving the feedback, the system traced back and found that in the final deformation analysis report corresponding to these four events, the calculation strategy of geometric illusion entropy was not sensitive enough to high-frequency small abrupt changes, and its false deformation misjudgment rate was 2.5%.

[0065] The system then increases the weight of retaining real deformation features in the adversarial deformation decoupling model and decreases the safety threshold for triggering the fidelity-preserving circuit breaker mechanism from 0.65 to 0.58, forcing the system to abandon high-risk smooth completion earlier in subsequent batches. The purpose of this step is to use the real physical assembly results to verify the theoretical deduction of the algorithm, and to continuously calibrate the decision boundary of 3D data processing through closed-loop feedback, so that the system can maintain a very high deformation judgment accuracy even in the dual crisis of data incompleteness and inconsistency in form.

[0066] In a preferred embodiment of the present invention, this embodiment provides a mechanism for calculating the geometric illusion entropy during the reconstruction process and quantifying the degree of excessive smoothness; specifically, when performing three-dimensional completion using implicit neural representation technology, if confidence is evaluated based solely on the coordinate deviation at a single moment, it is very easy to mistakenly identify the tiny abrupt changes that the algorithm forcibly smooths out in order to smooth the surface as high-confidence data when dealing with flexible target objects with complex curvature.

[0067] Therefore, in this embodiment, the reconstruction integrity index and the deformation sensitivity index are defined as mutually exclusive constraints; in the derivation of implicit neural representation technology, the system calculates the Euclidean distance variance between the generated data points and the discrete feature set;

[0068] The geometric uncertainty coefficient is obtained by combining the Euclidean distance variance with the local curvature change rate. The geometric uncertainty coefficient is accumulated over time to generate the geometric illusion entropy, which is used to characterize the degree of distortion when the AI ​​model completes the unobserved area.

[0069] In micro-data extrapolation, it is assumed that the implicit neural representation network exists in the region. Generate a sequence of data points ,in , For positive integers, the system calculates each data point in the sequence. The variance of the Euclidean distance between the nearest true discrete feature point and the Euclidean distance. ;

[0070] Simultaneously, calculate the sequence Local curvature change rate Specifically, the local curvature change rate By analyzing the sequence The network obtains the surface by fitting a local quadratic surface to adjacent data points and calculating the derivative of the local differential gradient of the principal curvature of the fitted surface; if the network forcibly generates a smooth surface in pursuit of reconstruction integrity, it will lead to... Increase and The anomaly approaches 0; at this point, the geometric uncertainty coefficient is calculated. The system uses a weighted formula:

[0071] ;

[0072] in, To prevent division by zero anomalies, a preset small constant is used, such as taking... To ensure the rate of change of local curvature The abnormal decrease approaches The system will not crash or overflow. and These are preset weighting coefficients; used to adjust the proportions of the Euclidean distance variance and the reciprocal of the local curvature change rate in the uncertainty assessment; simultaneously... and It has a built-in dimension conversion factor, which normalizes and unifies the physical dimensions on both sides of the formula and between each term, ensuring the mathematical rigor of numerical accumulation and the validity of the equation.

[0073] Specifically, the preset weighting coefficients and The result was obtained by optimizing a grid search algorithm on a small-scale calibration dataset, with the objective of maximizing the variance of the geometric uncertainty coefficient in the known defect region; if Within the normal range but The system will also determine if the abnormal decrease approaches 0. The value increased;

[0074] The system will perform multiple simulation batches, such as to Moment The values ​​are integrated and summed to obtain the final geometric illusion entropy. If, during the accumulation process, at a certain moment... Since the value cannot be calculated due to the extreme sparseness of the local point cloud, the system extracts values ​​from adjacent regions. The mean value is injected into the accumulation formula as a penalty compensation term to prevent interruption or distortion in the entropy value calculation;

[0075] For example, during the reconstruction of a flexible target object, the sensor's data coverage of a certain stressed bending area was only 40%; when the implicit neural representation network extrapolated this area, it generated an extremely smooth transition surface; during calculation, the system discovered that the variance of the Euclidean distance between the generated point and the real feature point was... It reached 0.45 square millimeters, and the local curvature change rate was... It is almost zero; based on this, the system calculates an extremely high geometric uncertainty coefficient, and after 5 simulation cycles, it accumulates a geometric illusion entropy as high as 0.89.

[0076] The purpose of this mechanism is to accurately quantify the degree of distortion caused by excessive smoothing of the model in the data blind zone through the joint calculation of variance and curvature and the accumulation of time dimension, so as to provide a reliable quantitative indicator for subsequent fidelity decisions.

[0077] This embodiment provides a mechanism for detecting the entropy status of key topological regions and triggering a fidelity-based circuit breaker mechanism. Specifically, if a globally unified entropy monitoring and circuit breaker standard is used for the entire candidate 3D reconstruction model, it will not only consume huge computing resources, but also cause non-critical regions, such as the non-assembly area at the center of the flexible target object, to be frequently blocked due to slight smoothing and completion, thus reducing the overall analysis efficiency.

[0078] Therefore, in this embodiment, the candidate 3D reconstruction model is divided into multiple topological functional sub-regions, and the assembly edge region is identified as the key topological region. The local average illusion entropy value in the key topological region is extracted. It is determined whether the local average illusion entropy value is in a high-risk range. If it is in a high-risk range, it is determined that there is a risk of topological locking failure. The smooth completion operation of the region is immediately blocked, and the original discrete feature set is locked as the output benchmark to preserve the physically real broken features.

[0079] In the logical deduction, the system analyzes the CAD design semantics of the flexible target object and divides the model into a central load-bearing area. and assembly docking area System extraction The geometric illusion entropy within is calculated, and its local average value is calculated. ;

[0080] Set the high-risk interval threshold as For example, 0.65, the high-risk interval threshold. This refers to the aforementioned preset safety threshold, which is obtained by combining the statistical mean of the aforementioned successful historical assembly cases with the lower limit of the geometric illusion entropy that triggers the critical state of topological interference in the Monte Carlo simulation experiment. For example, the smaller of the two values ​​is taken to ensure the most stringent safety redundancy.

[0081] The high-risk range refers to the range greater than this threshold. The numerical range; if Then the circuit breaker signal will be triggered. Immediately freeze the implicit neural network in The weight update and coordinate output of the region directly use the purified discrete feature set as the final geometric expression of the region.

[0082] like ,but This allows the network to perform smooth completion; if semantic gaps in the CAD design prevent automatic identification of assembly edge regions, the system activates an exception fallback mechanism, marking the physical boundary with the highest curvature gradient as a critical topology region by default and applying the strictest applicable rules. Threshold, such as 0.50;

[0083] For example, in the digital twin analysis of a flexible target object, the system divides the flange edges around the flexible target object used for bolt connections into key topological regions. During the reconstruction process, the system extracted The local average hallucination entropy value soared to 0.82, entering the high-risk range;

[0084] The system determined that if the implicit neural network were allowed to continue completing these sparse point clouds with natural drooping deformations into a perfectly smooth flange surface, it would cause a fatal topology locking failure during docking; therefore, the system instantly issued a fuse-breaking command to cut off the connection. The generative completion of the region forcibly preserves the original discrete points of the flange edge as presented in the actual measurement; the purpose of this mechanism is to achieve asymmetric allocation of computing resources and precise interception of risks, ensuring that at the most critical assembly boundary, the system can resolutely execute the decision to sacrifice integrity for fidelity.

[0085] This embodiment provides a mechanism for outputting a visually discrete but topologically reliable feature skeleton based on a discrete feature set. Specifically, after triggering the fidelity-based circuit breaker and abandoning the reconstruction of the complete surface, if the system directly outputs scattered, topologically unrelated original point clouds to the digital twin assembly hub, the downstream geometric dimension and tolerance GD&T analysis software will be unable to resolve its surface normals and assembly constraints, resulting in a break in the assembly simulation process. Therefore, this embodiment constructs a non-manifold geometric topology mesh based on a discrete feature set while abandoning the reconstruction of the complete surface.

[0086] Keypoint annotation is performed on a non-manifold geometric topology mesh to generate a discrete point cloud set containing fracture edges and abrupt nodes. This discrete point cloud set is then mapped to a reliable feature skeleton that lacks visual continuity but mathematically satisfies GD&T constraints. During the data structure transformation, it is assumed that the locked discrete feature set contains a node set. The system utilizes a variant of the triangulation algorithm in Build a grid on top Due to data incompleteness, the system allows the generation of non-manifold meshes, that is, it allows the existence of topological singularities where an edge connects only one face or is suspended, in order to truly reflect the physical fragmentation state of the data.

[0087] System traversal The topological connectivity matrix is ​​used to mark nodes with a degree of 1 as break edge points. Nodes whose angle between the normal vectors of adjacent faces is greater than a preset angle threshold are marked as mutation nodes. The preset angle threshold is determined based on the empirical value of the maximum theoretical bending angle allowed for the local contour of the part in the downstream GD&T specification; the system extracts... and By combining the benchmark reference system in the GD&T specification, it is encapsulated as a feature skeleton object. ;

[0088] Specifically, during the encapsulation process, the system extracts the fracture edge points based on the predefined geometric reference planes of the CAD model. and mutation nodes Perform spatial coordinate system alignment; based on the tolerance zone definitions for profile, straightness and flatness in the GD&T specification, expand each discrete node into a topological unit with a three-dimensional tolerance boundary;

[0089] Finally, these topological units with tolerance constraints are aggregated to form a rigid data structure, namely the feature skeleton object, which, although lacking visual continuity in space, has a clearly defined normal vector and tolerance boundary at each node mathematically. In extreme cases, the node set Extreme sparsity makes it impossible to construct even non-manifold meshes, so the system degenerates into executing a bounding box strategy, only computing... The smallest oriented bounding box (OBB) is used to extract the principal axis direction as the output of the degraded feature skeleton to ensure that the data flow is uninterrupted.

[0090] For example, in the welded area of ​​the flange edge of the aforementioned flexible target object, the system obtained a discrete feature set consisting of 150 isolated points; instead of forcibly stitching them together with a smooth surface, the system constructed a non-manifold topological mesh containing a large number of topological holes, and marked 42 fracture edge points and 12 abrupt change nodes at the fracture point of the mesh.

[0091] The system outputs a feature skeleton that is presented in a discrete state in a 3D view to the assembly center. Although the skeleton does not have visual continuity, each node contains absolutely real physical coordinates and normal constraints, enabling the assembly robot arm to accurately calculate the docking interference based on these real fracture boundaries. The purpose of this step is to solve the analytical compatibility problem of residual defect clouds in downstream engineering software, and to provide mathematically rigorous and topologically reliable rigid skeleton support for extremely high-risk flexible assembly without faking any geometric surfaces.

[0092] In a preferred embodiment of the present invention, this embodiment provides a gravity shape compensation mechanism for handling the gravity masking effect; specifically, if only the measured point cloud data is used for anisotropic feature decoupling and deformation analysis, since the flexible target object is inevitably affected by gravity when measured on the ground, its physical form on Earth is fundamentally different from its working form in the weightless state of space.

[0093] This natural sagging caused by gravity is easily misjudged by the system as a manufacturing defect, or it may mask the real minute deformation, leading to a distortion of the baseline of the adversarial deformation decoupling model when dealing with high-risk steady states.

[0094] Therefore, this embodiment introduces a gravity shape compensation module to obtain the measured shape data of the flexible target object under gravity environment, simulate the theoretical relaxation shape of the flexible target object under weightlessness based on the finite element physical simulation engine, calculate the elastic deformation difference between the measured shape data and the theoretical relaxation shape, and inject the elastic deformation difference as a correction factor into the adversarial deformation decoupling model to eliminate the steady-state interference of gravity on deformation analysis.

[0095] In microscopic data extrapolation, the system acquires a set of measured morphological points containing three-dimensional coordinates. ;Will Importing the finite element physical simulation engine, the system performs a set of measured shape points. Poisson surface reconstruction and mesh generation are performed to generate a shell element mesh model suitable for finite element analysis;

[0096] The structural mechanical parameters of the composite material, such as Young's modulus and Poisson's ratio, are set. The Young's modulus and Poisson's ratio are obtained by reading the measured calibration values ​​in the material mechanical property quality inspection report that comes with the flexible target object.

[0097] The system extracts the assembly hole nodes of the flexible target object, sets them as fixed constraint boundary conditions, and applies a reverse gravity field vector, such as... ;

[0098] Specifically, during the iterative calculation process, the system establishes the global stiffness matrix of the flexible target object based on the shell element mesh model and constraint boundary conditions, and constructs an external load vector containing the reverse gravity vector; based on the force equilibrium state, the system uses a numerical iterative method to solve the displacement vector matrix of each node;

[0099] After the solution is obtained, the displacement components of each node in the displacement vector matrix are superimposed onto the original measured point set. On the spatial coordinates, the theoretical relaxation morphological point set is calculated. ;

[0100] System traversal and For each corresponding spatial node, calculate the displacement vector difference between the nodes to generate the elastic deformation difference matrix. The system will The inverse offset, converted into spatial coordinates, is directly superimposed as a correction factor into the input layer of the adversarial deformation decoupling model to counteract the large-scale low-frequency deformation caused by gravity.

[0101] If the finite element physics simulation engine fails to converge when processing extremely sparse or topologically incomplete nonmanifold meshes, resulting in no output... The system then triggers an anomaly fallback mechanism, downgrading to use a pre-set empirical drooping surface equation based on historical measurement data of the flexible target object for approximate compensation, or truncating abnormal displacement values ​​in the difference matrix that exceed the material physical limits, in order to prevent erroneous correction factors from contaminating subsequent decoupling analysis links.

[0102] For example, in a ground-based automated measurement scenario of a flexible target object, a 2-meter-long flexible target object sags naturally by up to 3 millimeters in the central area due to gravity. Without intervention, the system would mark it as a severe dent deformation defect. Through the gravity shape compensation module of this embodiment, finite element simulation calculates that the elastic deformation difference in this area is 3.1 millimeters upward.

[0103] The system injects this difference as a correction factor into the adversarial deformation decoupling model, which precisely offsets the macroscopic sag caused by gravity. This allows a real, existing, high-frequency micro-crease protrusion of only 0.2 mm, which was originally hidden in the sag area, to be successfully peeled off and preserved, thus avoiding missed detection caused by the gravity masking effect.

[0104] The purpose of this mechanism is to eliminate the systematic steady-state interference of gravity on the deformation analysis of flexible target objects, and to ensure that the deformation features extracted under the gravity environment on the ground can truly reflect the physical topology of the parts under the weightless state in space, so as to provide an absolutely reliable benchmark data base for subsequent reconstruction deviation monitoring and fidelity arbitration.

[0105] This embodiment provides a mechanism for constructing a closed loop for deformation feature verification. Specifically, although the system monitors and melts down the smooth completion of high-risk areas through geometric illusion entropy, if the loss function of the implicit neural representation network remains static and cannot learn from real physical assembly failures, the system will be unable to cope with the cumulative deviations caused by material batch changes or process fine-tuning, causing the calculation of geometric illusion entropy to gradually deviate from the actual physical boundary.

[0106] Therefore, in this embodiment, physical interference data during the actual assembly process is collected as truth feedback through the deformation logic iteration unit; the predicted deformation amount in the final deformation analysis report is compared with the physical interference data to calculate the false alarm rate and false alarm rate; based on the false alarm rate and false alarm rate, the penalty factor of geometric illusion entropy is calculated; the penalty factor is used to backpropagate and update the loss function of the implicit neural representation network, forcing the model to reduce its overconfidence in the fuzzy region in subsequent iterations;

[0107] In data flow and closed-loop simulation, the system extracts the physical interference depth scalar value from the torque sensor of the downstream assembly robot arm during the actual assembly process. The system will use the predicted interference depth from the final deformation analysis report. and Perform point-by-point comparison; if and If it is, it is determined to be a missed detection event and is included in the overall missed detection rate. ;like and If it is, it is determined to be a false alarm event and is included in the global false alarm rate. ;

[0108] The system sets asymmetric weighting coefficients. That is, the weight of missed detections, and That is, false alarm weight, and The penalty factor for geometric illusion entropy can be calculated using the formula. The calculation formula is as follows:

[0109] ;

[0110] Specifically, asymmetric weighting coefficients and Based on the engineering risk cost matrix, the cost of space assembly interference repair caused by missed detection is far greater than the cost of ground retesting caused by false alarms. The system determines this based on historical cost assessment data. and The scale order; in the next training iteration of the implicit neural representation network, the system will use the original reconstruction mean squared error loss function. Modify to a joint loss function:

[0111] ;

[0112] in, The mean square error between the generated surface points and the real discrete feature points is calculated. To generate the local variance term of the data point set, it is calculated as the sum of the coordinate variances of the generated points within their local neighborhoods; the system utilizes Calculate the gradient and backpropagate to update the network parameters;

[0113] If the collected physical interference data exhibits outliers due to a momentary sensor malfunction, such as the feedback interference depth exceeding the absolute physical limit of the flexible target object's thickness, the system will mark this batch of feedback data as invalid and discard it, forcibly maintaining the penalty factor from the previous iteration. This remains unchanged to prevent incorrect truth values ​​from causing the network loss function to fail to converge;

[0114] For example, in the digital twin docking assembly of flexible target objects, the system predicts that a certain flange edge region can be perfectly and smoothly closed, i.e. However, during actual space simulation assembly, the robotic arm reported a physical interference of 0.5 millimeters, i.e. The system marks this event as a fatal missed detection and calculates the missed detection rate for the current batch. Surge;

[0115] The system then calculates a significantly increased penalty factor. The weights are injected into the joint loss function of the implicit neural representation network for backpropagation. When processing the next batch of similar edges, the loss value of the network increases sharply when trying to generate an overly smooth surface due to the amplification effect of the penalty factor. This forces the model to output a higher feature variance, which increases the geometric illusion entropy of the region and prompts the adaptive topological fidelity arbitration unit to trigger the fidelity circuit breaker mechanism earlier and output a discrete feature skeleton.

[0116] The purpose of this mechanism is to build a closed-loop evolutionary capability by utilizing real physical assembly feedback. By dynamically adjusting the network's penalty weights through quantifying the false deformation misjudgment rate, the implicit neural representation network is forced to continuously reduce its overconfidence in unmeasured fuzzy areas in subsequent iterations. In this way, it continuously converges to the safest system survival boundary between pursuing visual reconstruction and tolerating data fragmentation.

[0117] In a preferred embodiment of the present invention, this embodiment provides a mechanism for triggering analysis logic based on specific materials and data coverage. Specifically, if the system indiscriminately initiates computationally expensive contrastive learning deformation decoupling and implicit neural completion for all types of input data, it will result in a serious waste of computing resources when processing conventional rigid or data-complete flexible target objects, and is very likely to cause overfitting. Therefore, this embodiment limits the data coverage of sparse point cloud data to be lower than a preset integrity threshold, and the flexible target object is a composite material skin structure with anisotropic reflective properties.

[0118] During the data admission phase, the system calculates the effective projected area of ​​the input point cloud. Theoretical surface area of ​​the target object's CAD model The ratio is used to calculate the data coverage. Simultaneously, the system parses the metadata of the input data stream and extracts the material attribute tags of the target object. ;

[0119] If judgment Less than a preset integrity threshold, for example, 60%, and Including markings indicating anisotropic reflective properties such as carbon fiber cross-weaving, the system fully activates the aforementioned anisotropic feature decoupling unit and geometric illusion monitoring unit, entering a high-risk incomplete data processing mode; if the system detects... Greater than or equal to the integrity threshold, or When the material is isotropic, such as a common aluminum alloy with a uniform surface coating, the system triggers a resource degradation fallback mechanism, bypassing the complex contrastive learning decoupling network and frequency domain conversion, and directly using basic spatial domain median filtering and conventional polynomial surface fitting algorithms for fast processing, so as to release the computing power of edge nodes.

[0120] For example, at the automated scanning station of the spacecraft assembly line, due to the limited field of view and partial obstruction of the sensor, the point cloud data coverage of the currently input flexible target object is calculated to be only 55%; at the same time, the system recognizes that the carbon fiber texture on the surface of the flexible target object corresponds to strong anisotropic reflective properties.

[0121] Based on these dual triggering conditions, the system immediately initiates the deep decoupling and geometric illusion entropy quantization process to cope with the extreme working conditions of extremely scarce data and severe optical artifact interference. The purpose of this mechanism is to achieve on-demand allocation of algorithm computing power through pre-construction coverage and material dual verification, ensuring that the system's high-order deformation decoupling and topology fidelity capabilities are accurately focused on specific complex scenarios with high risk and high interference.

[0122] This embodiment provides a mechanism for generating a multidimensional visualization deformation analysis report. Specifically, if the system only outputs a numerical geometric illusion entropy matrix or discrete feature skeleton coordinates to the digital twin assembly center, the downstream automated assembly robot or quality inspector will find it difficult to intuitively and globally assess the comprehensive assembly risk of the flexible target object in three-dimensional space, and will easily miss local fatal defects from a macroscopic perspective.

[0123] Therefore, this embodiment sets the final deformation analysis report to include a 3D model view and a confidence heatmap. The confidence heatmap visually displays the distribution of geometric illusion entropy on the part surface, used to indicate assembly risk areas. During the report rendering stage, the system uses the geometric illusion entropy values ​​calculated in the previous step. Mapped to the preset RGB pseudo-color space;

[0124] For entropy Regions approaching 0, i.e., areas with sufficient raw measured data and extremely high confidence in the extrapolation, are rendered by the system in cool tones, such as dark blue; as... An increase in the value indicates an increase in the component inferred by the algorithm, leading to a greater risk of distortion due to over-smoothing, and a smoother transition of rendered colors to warmer tones in the color space; for high-risk intervals that trigger the fidelity-based circuit breaker mechanism, such as If so, it will be forcibly rendered as a bright warning color, such as dark red;

[0125] The system fuses the aforementioned color matrix with candidate 3D reconstruction models or discrete feature skeletons using texture mapping to generate a global confidence heatmap. If, during the rendering process, it is found that the point cloud data in a certain local area is in an absolute vacuum state—that is, even the adjacent reference points used for interpolation do not exist—this results in a local entropy value... If a division-by-zero error occurs or the calculation fails, the system will activate the blind zone warning fallback mechanism, forcibly marking the area as pure black on the heat map and sending a high-priority physical retest command to the main control bus. The system will refuse to provide any mathematical confidence endorsement for the pure black area.

[0126] For example, on the digital twin assembly hub screen of the flexible target object, the final deformation analysis report is presented as a three-dimensional model view that supports multi-view rotation; the quality control engineer can intuitively see through the confidence heatmap that the large load-bearing area in the center of the flexible target object is a safe dark blue, but at the edge of the critical docking flange in the upper right corner, due to severe data incompleteness and extremely high uncertainty of algorithm completion, this area is a bright dark red, clearly indicating an extremely high risk of topology lock failure, prompting the assembly system to perform interference calculations based on the discrete skeleton of this red area rather than trusting the smooth surface;

[0127] The purpose of this step is to transform abstract, multi-dimensional algorithmic uncertainties into visual maps that conform to human intuition and industrial vision standards, providing global and intuitive risk guidance for assembly decisions of complex flexible parts.

[0128] This embodiment provides a dynamic balancing mechanism for computing resources based on production cycle time. Specifically, under the extremely fast production cycle time of spacecraft assembly lines, if the implicit neural representation technology always maintains a fixed high sampling resolution for full-domain 3D inference, it will cause the analysis time of a single flexible target object to exceed the standard, leading to the stagnation of the entire assembly line.

[0129] Conversely, if low resolution is used globally in pursuit of timeliness, it will blur the subtle real deformations of critical assembly edges, leading to fatal missed detections.

[0130] Therefore, in this embodiment, the system is deployed on edge computing nodes. The adaptive topology fidelity arbitration unit dynamically adjusts the sampling resolution of the implicit neural representation technology according to the production cycle requirements to maintain a dynamic balance between computational efficiency and analytical accuracy. During the dynamic scheduling process, the edge computing nodes obtain the rated cycle time of the current workstation in real time. With time already consumed Calculate the remaining available time ;

[0131] Simultaneously, the system evaluates the initial geometric illusion entropy distribution matrix of the region to be reconstructed; if With sufficient power, the system will increase the global sampling resolution to the highest level. To obtain the finest reconstructed surface;

[0132] like When the system approaches a preset time threshold, it triggers an asymmetric dimensionality reduction sampling strategy: the preset time threshold is defined as the rated cycle time of the current workstation. 15%; in non-critical, smooth areas with low geometric illusion entropy, such as the center of a flexible target object, apply extremely low resolution. In critical topological regions where the entropy value is in the medium warning range or has not yet triggered a circuit breaker, high resolution is forcibly maintained. Detailed simulations are performed, while for high-risk intervals that have triggered the fidelity-guaranteed circuit breaker mechanism, the simulations are stopped directly to release computing power;

[0133] If an edge computing node experiences a sudden computing bottleneck, such as hardware overheating leading to severe frequency throttling, and If the system runs out of resources, it will trigger a rapid blocking fallback mechanism, instantly stopping the implicit neural completion process of all non-critical areas and only outputting the original discrete feature set of the critical topological areas to ensure that a deformation analysis report with a minimum usable state is forcibly output before the timeout deadline.

[0134] For example, the spacecraft assembly line requires that the digital twin deformation analysis of each flexible target object must be completed within 45 seconds; when the analysis process reaches the 30th second, the edge computing node finds that the remaining 15 seconds are insufficient to support the global high-precision reconstruction of the entire panel.

[0135] The system immediately adjusts its strategy dynamically according to the production cycle requirements, widening the spacing between implicit neural sampling points in the central area of ​​the flexible target object from 0.1 mm to 2.0 mm, while allocating all the saved edge computing power to the high-risk flange edge that appears in deep red on the heat map, maintaining its high-resolution simulation of 0.05 mm.

[0136] The purpose of this mechanism is to ensure that, under the constraints of strict industrial production time and limited edge hardware computing power, the system can always achieve the optimal balance between delivering analysis reports on time and maintaining the accuracy of core assembly topology through asymmetric resolution dynamic allocation.

[0137] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A part deformation analysis system driven by three-dimensional measurement data, characterized in that, include: Anisotropic feature decoupling unit: acquires sparse point cloud data of flexible target object, and analyzes the surface texture features of the sparse point cloud data; Based on a pre-defined material optical reflection model, nonlinear optical artifacts caused by anisotropic material properties are identified. An adversarial deformation decoupling model is constructed to separate the nonlinear optical artifacts from the real geometric deformation, generating a purified discrete feature set; Geometric illusion monitoring unit: Utilizes implicit neural representation technology to perform generative 3D completion on the discrete feature set and constructs a candidate 3D reconstruction model; Simultaneously calculates the geometric illusion entropy during the reconstruction process, quantifies the degree of excessive smoothness and confidence decay of the candidate 3D reconstruction model in the unmeasured area, and generates an illusion entropy distribution map; Adaptive topology fidelity arbitration unit: Based on the illusion entropy distribution map, detect the entropy status of key topological regions; If the entropy value is lower than the preset safety threshold, the complete candidate 3D reconstruction model is output; if the entropy value is higher than the preset safety threshold, the fidelity circuit breaker mechanism is triggered, the generation of the complete model is abandoned, and instead, a visually discrete but topologically reliable feature skeleton is output based on the discrete feature set to generate the final deformation analysis report. Deformation logic iteration unit: Based on the final deformation analysis report and downstream assembly feedback, construct a deformation feature verification closed loop; Quantify the false deformation misjudgment rate and dynamically adjust the weight parameters of the adversarial deformation decoupling model and the calculation strategy of the geometric illusion entropy; The method of constructing the adversarial deformation decoupling model to separate the nonlinear optical artifacts from the real geometric deformation includes: extracting high-frequency noise signals from the sparse point cloud data and mapping them to the frequency domain feature space; loading a pre-set anisotropic material texture library and matching the texture response mode under the current illumination angle. In the frequency domain feature space, a contrastive learning algorithm is used to distinguish between optical artifact signals that conform to texture response patterns and real deformation signals that conform to manufacturing defect characteristics; data points marked as optical artifact signals are suppressed while real deformation signals are retained, thereby generating the purified discrete feature set.

2. The part deformation analysis system based on three-dimensional measurement data as described in claim 1, characterized in that, The geometric illusion entropy in the computational reconstruction process, and the way to quantify the degree of excessive smoothness, includes: defining reconstruction integrity index and deformation sensitivity index as mutually exclusive constraints; In the derivation process of the implicit neural representation technology, the Euclidean distance variance between the generated data points and the discrete feature set is calculated; the Euclidean distance variance is combined with the local curvature change rate to obtain the geometric uncertainty coefficient; the geometric uncertainty coefficient is accumulated in the time dimension to generate the geometric illusion entropy, which is used to characterize the degree of fiction when the AI ​​model completes the unobserved region.

3. The part deformation analysis system based on three-dimensional measurement data as described in claim 2, characterized in that, The method for detecting the entropy state of the key topological region and triggering the fidelity circuit breaker mechanism includes: dividing the candidate 3D reconstruction model into multiple topological functional sub-regions, identifying the assembly edge region as the key topological region; and extracting the local average illusion entropy value within the key topological region. Determine whether the local average illusion entropy value is in a high-risk range; if it is in a high-risk range, determine that there is a risk of topology locking failure, immediately block the smooth completion operation for this region, and lock the original discrete feature set as the output benchmark to preserve the physically real broken features.

4. The part deformation analysis system based on three-dimensional measurement data as described in claim 3, characterized in that, The method of outputting a visually discrete but topologically reliable feature skeleton based on the discrete feature set includes: constructing a non-manifold geometric topological mesh based on the discrete feature set while abandoning the reconstruction of the complete surface; Key points are labeled on the non-manifold geometric topology mesh to generate a discrete point cloud set containing fracture edges and abrupt nodes; the discrete point cloud set is then mapped to a reliable feature skeleton that does not have visual continuity but mathematically satisfies GD&T constraints.

5. The part deformation analysis system based on three-dimensional measurement data as described in claim 1, characterized in that, The system also includes a gravity shape compensation module for handling gravity masking effects: acquiring measured shape data of the flexible target object under gravity conditions; Based on the finite element physical simulation engine, the theoretical relaxation state of the flexible target object under weightlessness is simulated. Calculate the difference in elastic deformation between the measured morphological data and the theoretical relaxation morphology; The elastic deformation difference is injected as a correction factor into the antagonistic deformation decoupling model to eliminate the steady-state interference of gravity on deformation analysis.

6. The part deformation analysis system based on three-dimensional measurement data as described in claim 1, characterized in that, The deformation logic iteration unit constructs a deformation feature verification closed loop by collecting physical interference data from the actual assembly process as truth feedback. The predicted deformation in the final deformation analysis report is compared with the physical interference data to calculate the false alarm rate and the missed detection rate. Based on the false alarm rate and false alarm rate, a penalty factor for geometric illusion entropy is calculated; the penalty factor is then used to backpropagate and update the loss function of the implicit neural representation network, forcing the model to reduce its overconfidence in blurred regions in subsequent iterations.

7. The part deformation analysis system based on three-dimensional measurement data as described in claim 1, characterized in that, The data coverage of the sparse point cloud data is lower than the preset integrity threshold, and the flexible target object is a composite material skin structure with anisotropic reflective properties.

8. The part deformation analysis system based on three-dimensional measurement data as described in claim 1, characterized in that, The final deformation analysis report includes a three-dimensional model view and a confidence heatmap, wherein the confidence heatmap visually displays the distribution of the geometric illusion entropy on the surface of the part, and is used to indicate assembly risk areas.

9. A part deformation analysis system based on three-dimensional measurement data as described in any one of claims 1 to 8, characterized in that, The system is deployed on edge computing nodes. The adaptive topology fidelity arbitration unit dynamically adjusts the sampling resolution of the implicit neural representation technology according to the production cycle requirements, so as to maintain a dynamic balance between computational efficiency and analysis accuracy.

Images