Industrial internet of things automatic assembly system based on digital twinning

Abstract

The invention relates to the field of Internet of Things assembly, in particular to an industrial Internet of Things automatic assembly system based on digital twinning, which comprises a user platform, a service platform, a management platform, a sensing network platform and an object platform which are in communication connection in sequence, the object platform comprises a twin modeling module, a twin assembling module, a coordinate matching module and an adaptive feedback module which are in communication connection in sequence; an action tracking unit monitors assembly actions of a physical space in real time, an assembly deviation sequence is detected in a mode that a parameterized model is aligned with a triangular mesh model, and when deviation exceeds a set threshold value, an adaptive feedback unit starts a replanning process; the problem that in the prior art, the assembling action of an automatic assembling system cannot be accurately mapped between a virtual space layer and a physical space layer is solved, sudden disturbance is responded at the millisecond level, and the robustness of the assembling process is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet of Things assembly, and in particular to an industrial Internet of Things automatic assembly system based on digital twins. Background Art

[0002] With the rapid development of intelligent manufacturing technology, industrial automated assembly systems have become the core equipment of modern manufacturing. Traditional automated assembly systems mainly rely on pre-programmed logic control (PLC), fixed-path robotic arms and offline simulation technology to achieve assembly line assembly of standardized parts.

[0003] In recent years, digital twin technology has provided a new direction for the intelligent upgrade of assembly systems by building a real-time interactive closed loop between physical entities and virtual models. In existing technologies, digital twin systems usually establish a static mapping relationship based on CAD models and sensor data. However, in actual applications, due to the influence of environmental factors in the physical space layer, there are sudden disturbances, which make it impossible to accurately map the twin space layer and the physical space layer. Summary of the Invention

[0004] The main purpose of the present invention is to provide an industrial Internet of Things automatic assembly system based on digital twins, aiming to solve the problem in the existing technology that the assembly actions of the automatic assembly system cannot be accurately mapped between the virtual space layer and the physical space layer.

[0005] To achieve the above objectives, the present invention provides an industrial Internet of Things automatic assembly system based on digital twins, comprising a management platform, a sensor network platform, and an object platform that are communicatively connected in sequence. The management platform comprises: a twin modeling module for collecting physical space layer data and constructing a twin space layer; a twin assembly module for performing virtual assembly actions within the twin space layer; a coordinate matching module for synchronously mapping the assembly sequence of the virtual assembly action to the physical space layer; and an adaptive feedback module for tracking the assembly action within the physical space layer and detecting the assembly deviation sequence. Among them, the adaptive feedback module includes a motion tracking unit and an adaptive feedback unit that are communicatively connected. The motion tracking unit is used to track and detect the assembly deviation sequence of the assembly action in the physical space layer. The adaptive feedback unit is used to re-plan the assembly sequence of the virtual assembly action after the assembly deviation sequence reaches a set threshold, and then provide feedback on the assembly action in the physical space layer based on the re-planning of the assembly sequence.

[0006] Optionally, the twin modeling module includes: an acquisition unit for acquiring assembly system structural data; a modeling unit for using structural data to perform three-dimensional modeling to obtain a twin space layer; an extraction unit for extracting geometric features from structural data; structural data includes dimension data, position data, constraint data, and tolerance data; geometric features include feature points, feature lines, and feature surfaces.

[0007] Optionally, the twin assembly module includes a translation unit and a rotation unit that are communicatively connected, the translation unit performs a mobile assembly action through a translation matrix, and the rotation unit performs a rotation assembly action through a rotation matrix.

[0008] Optionally, the coordinate matching module includes: a path planning unit, used to plan the assembly path through a motion planning algorithm; an action decomposition unit, used to decompose the assembly task in the assembly path into discrete steps and bind each discrete step to the structural data; and a data conversion unit, used to convert the discrete steps into control signals within the physical space layer.

[0009] Optionally, the motion tracking unit includes a laser tracker, and the laser tracker adjusts the detection point in the physical space layer through reinforcement learning.

[0010] Optionally, the detection process of the assembly deviation sequence includes: using the parametric model as the assembly reference, aligning the triangular mesh model and the parametric model to the same coordinate system; then quantifying the deviation between the triangular mesh vertices and the parametric surface and calculating the shortest distance from each mesh vertex to the parametric surface; finally, comparing the shortest distance with the design error to obtain the assembly deviation sequence.

[0011] Optionally, the adaptive feedback unit includes: a feature stratification subunit, which is used to obtain a feature stratification sequence using a feature recognition method; an adaptive stratification subunit, which is used to obtain an adaptive stratification sequence using an adaptive stratification algorithm; and an assembly sequence stratification subunit, which is used to fuse the feature stratification sequence with the adaptive stratification sequence to obtain an assembly stratification sequence.

[0012] Optionally, the method of obtaining a feature hierarchical sequence using a feature recognition method includes: initializing geometric features and assembly variables; updating the triangular mesh model to obtain the shortest distance, determining whether the shortest distance exists in the deviation group, and if so, skipping it; otherwise, inserting the shortest distance into the deviation group; and finally arranging the deviation groups in descending order to obtain the feature hierarchical sequence.

[0013] Optionally, obtaining the adaptive layering sequence by using an adaptive layering algorithm includes: calculating the layering distance by using the minimum tilt angle of the triangle facets of each layer in the triangular mesh model to obtain the adaptive layering sequence.

[0014] Optionally, the adaptive feedback unit also includes a replanning subunit, which adopts a Q-learning-based reinforcement learning framework to dynamically adjust the replanning strategy of the assembly sequence, specifically including: constructing a Markov decision model, taking the current assembly deviation sequence and the physical space layer environment state as input states, and the assembly sequence adjustment action as the output action, and optimizing the action selection strategy through Q-value iterative update; designing a reward function, comprehensively calculating the instant reward based on the deviation reduction rate, path optimization efficiency and resource consumption indicators of the replanned assembly sequence, and guiding the algorithm to converge to the global optimal solution; simulating the execution results of different assembly actions in the twin space layer, dynamically adjusting the fusion weights of the feature layered sequence and the adaptive layered sequence through reinforcement learning, generating the optimal assembly layered sequence and feeding it back to the physical space layer for execution.

[0015] The beneficial effects that can be achieved by the present invention are as follows: The present invention ensures real-time transmission and processing of data from the physical layer to the virtual layer by integrating four modules of twin modeling, assembly, coordinate matching and adaptive feedback in the object platform; it also decomposes the virtual assembly action into discrete steps through the path planning unit of the motion planning algorithm, binds the structural data, and finally generates physical control signals through the data conversion unit. This process realizes the accurate mapping of the virtual assembly sequence to the physical execution instruction. In addition, the translation unit and the rotation unit use the translation matrix and the rotation matrix respectively to mathematically model the assembly action, ensuring that the movement and rotation actions in the virtual space can be converted into the motion trajectory of the physical robotic arm without error; it also monitors the assembly action in the physical space in real time through the motion tracking unit, and detects the assembly deviation sequence by aligning the parametric model with the triangular mesh model. When the deviation exceeds the set threshold, the adaptive feedback unit starts the re-planning process; it solves the problem in the prior art that the assembly action of the automatic assembly system cannot be accurately mapped between the virtual space layer and the physical space layer, realizes the response to sudden disturbances at the millisecond level, and significantly improves the robustness of the assembly process. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0017] Figure 1 Schematic diagram of the connection structure of the system of the present invention; Figure 2 Schematic diagram of the connection structure of the management platform of the present invention.

[0018] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] If there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or suggesting their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the meaning of "and / or" appearing throughout the text includes three parallel schemes. Taking "A and / or B" as an example, it includes scheme A, or scheme B, or a scheme in which A and B are satisfied at the same time. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in this field to implement. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0021] Example 1: As attached Figure 1 With attached Figure 2 As shown, this embodiment provides an industrial Internet of Things automatic assembly system based on digital twins, including a management platform, a sensor network platform, and an object platform that are sequentially connected in communication. The management platform includes: Twin modeling module, used to collect physical space layer data and build the twin space layer; Twin assembly module, used to perform virtual assembly actions within the twin space layer; The coordinate matching module is used to synchronously map the assembly sequence of virtual assembly actions to the physical space layer; Adaptive feedback module, used to track assembly actions in the physical space layer and detect assembly deviation sequences; Among them, the adaptive feedback module includes a motion tracking unit and an adaptive feedback unit that are communicatively connected. The motion tracking unit is used to track and detect the assembly deviation sequence of the assembly action in the physical space layer. The adaptive feedback unit is used to re-plan the assembly sequence of the virtual assembly action after the assembly deviation sequence reaches a set threshold, and then provide feedback on the assembly action in the physical space layer based on the re-planning of the assembly sequence.

[0022] It should be noted that the digital twin-based industrial Internet of Things automatic assembly system in this embodiment also includes a user platform and a service platform that are communicatively connected to each other, and the service platform is communicatively connected to the management platform, thereby forming a standard Internet of Things five-platform structure.

[0023] Among them, the physical entities of the user platform include various user terminals, such as mobile phones, computers, special terminals, etc., which realize user-end services through integration with user information system software.

[0024] The service platform is a functional platform for realizing service communication.

[0025] The management platform is a functional platform for realizing the operation and management of the Internet of Things system. In some embodiments, the management platform may include multiple management sub-platforms, which are respectively connected to the above-mentioned sensor network platforms, and each management sub-platform corresponds to a twin modeling module, a twin assembly module, a coordinate matching module and an adaptive feedback module.

[0026] The sensor network platform is a functional platform for realizing sensor communication.

[0027] The object platform is a functional platform for realizing perception control.

[0028] It should also be noted that traditional industrial automated assembly systems rely on pre-programmed logic controllers (PLCs) and fixed-path robotic arms, and the mapping relationship between their virtual simulation models and the actual physical environment is often static. This model is not very effective in the face of sudden disturbances in the physical space.

[0029] In some embodiments, sudden disturbances may occur as equipment vibration, part tolerance deviation, changes in ambient temperature and humidity, etc., making it impossible to dynamically adjust assembly actions, resulting in a disconnect between the virtual model and the actual assembly process, reduced assembly accuracy, or even failure.

[0030] In addition, although the existing digital twin technology can establish two-way interaction between physical and virtual, it still has shortcomings in real-time, adaptability and closed-loop control capabilities, especially the lack of rapid detection and dynamic replanning capabilities of assembly deviations.

[0031] Based on the above problems, the applicant proposed an industrial Internet of Things automatic assembly system based on digital twins. By integrating four modules of twin modeling, assembly, coordinate matching and adaptive feedback in the object platform, it ensures the real-time transmission and processing of data from the physical layer to the virtual layer.

[0032] The path planning unit, which utilizes a motion planning algorithm, decomposes the virtual assembly action into discrete steps, binds the structural data, and ultimately generates physical control signals through the data conversion unit. This process achieves a precise mapping of the virtual assembly sequence to physical execution instructions. Furthermore, the translation unit and rotation unit mathematically model the assembly action using translation and rotation matrices, respectively, ensuring that movement and rotation in virtual space can be accurately translated into the motion trajectory of the physical robot arm. This dynamic mapping mechanism breaks the "one-way static" relationship between virtual simulation and actual execution in traditional systems, achieving bidirectional real-time synchronization.

[0033] The motion tracking unit also monitors the assembly motion in the physical space in real time. The parametric model is aligned with the triangular mesh model to detect assembly deviation sequences. When the deviation exceeds a set threshold, the adaptive feedback unit initiates the following replanning process: Fusion of feature layering and adaptive layering: The feature layering sub-unit is used to identify the key deviation area, and the adaptive layering sub-unit is used to dynamically adjust the layering strategy according to the tilt angle of the triangle facet to generate an optimized assembly layering sequence. The key deviation area can be the maximum deviation point.

[0034] Reinforcement Learning-Driven Replanning Strategy: The replanning subunit utilizes a Q-learning-based reinforcement learning framework to construct a Markov decision model. This strategy is optimized through a reward function, simulating the outcomes of different assembly actions in virtual space. The layer weights are dynamically adjusted to ultimately generate the optimal assembly sequence, which is then fed back to the physical layer for execution. This closed-loop feedback mechanism enables the system to respond to sudden disturbances within milliseconds, significantly improving the robustness of the assembly process.

[0035] In some embodiments, the motion tracking unit includes a laser tracker.

[0036] In some embodiments, the input of the Markov decision model is the current deviation and the environment state, and the output is the adjustment action. In some embodiments, the reward function comprehensively considers deviation reduction rate, path efficiency, and resource consumption.

[0037] In some embodiments, the twin modeling module uses an acquisition unit to acquire real-time structural data (such as dimensions, positions, and tolerances) from the physical space. The modeling unit then constructs a high-precision 3D digital twin model based on this data. The extraction unit then further identifies geometric features (such as characteristic points, lines, and surfaces) to provide a precise geometric reference for subsequent virtual assembly. This modular design not only improves system scalability but also enables efficient collaboration among functional modules through standardized interfaces.

[0038] In some embodiments, the twin modeling module also includes an initialization submodule, a data communication submodule, a visualization module submodule, etc. For the initialization submodule, it is used for state initialization before the digital twin system is started, mainly including production line configuration initialization, virtual scene initialization, data configuration initialization, etc. Different twin devices, twin scenes, and production line original data can be configured according to the specific equipment of the assembly production line. For the data communication submodule, it is used for data communication between the physical production line and the virtual production line during system operation; it mainly includes real-time data acquisition, real-time data transmission, real-time data display, etc., with the help of multiple communication protocols to realize real-time data acquisition and processing of the system and integrate it into the system server, and process it through cloud computing and edge computing. For the visualization module submodule, it is used for user visual monitoring during the production and operation and maintenance of the production line; including production line virtual model, geometric motion simulation, production data display, etc. After the virtual model is established, the web platform is built with the help of front-end development, and the production data visualization is realized with the help of Echarts visualization charts.

[0039] In this embodiment, the twin modeling module includes: A collection unit, used for collecting assembly system structure data; A modeling unit, used to perform three-dimensional modeling using structural data to obtain a twin space layer; An extraction unit, used for extracting geometric features from structural data; Structural data includes dimension data, position data, constraint data, and tolerance data; Geometric features include feature points, feature lines, and feature surfaces.

[0040] In some embodiments, the acquisition unit may be a laser scanner, a three-dimensional coordinate measuring machine, an encoder, a strain gauge, etc.

[0041] In some embodiments, the dimensional data is the actual length or hole diameter of the assembly.

[0042] In some embodiments, the position data is a coordinate offset of the assembly reference plane.

[0043] In some embodiments, the constraint data is a range of motion limit for the joints of the manipulator.

[0044] In some embodiments, the tolerance data is a fit tolerance, a position tolerance, or the like.

[0045] For the modeling unit, based on the collected structural data, reverse engineering technology is used to reconstruct the 3D twin model, including: Data fusion: Compare the actual size data with the original CAD model and correct the model parameters through least squares fitting; Triangular mesh generation: Generates a high-density triangular mesh model based on point cloud data. In some embodiments, the resolution can reach 0.01mm. Parametric embedding: Tolerance data (such as hole-shaft fit tolerance) and constraints (such as motion range) are embedded into the model to form a virtual assembly space with tolerance boundaries.

[0046] For the extraction unit, in some embodiments, a hybrid algorithm (edge ​​detection, curvature analysis, cluster segmentation) is used to identify geometric features from the structural data: Feature points: such as the center of the assembly reference hole and the gripping point of the end effector; Characteristic lines: such as guide rail motion trajectory and bolt hole axis; Feature surface: such as gear meshing surface and assembly positioning plane.

[0047] Feature data is managed hierarchically according to priority, for example, high-precision fitting surfaces are marked as "key layers" and ordinary structural surfaces are marked as "auxiliary layers".

[0048] In this embodiment, the twin assembly module includes a translation unit and a rotation unit that are communicatively connected. The translation unit performs a mobile assembly action through a translation matrix, and the rotation unit performs a rotation assembly action through a rotation matrix.

[0049] Taking the virtual space layer built in Unity3D as an example, the position change of an assembly in three-dimensional space is described by the translation matrix. The translation matrix is ​​a special affine transformation matrix that only involves the position movement of the object without changing the object's shape, size or orientation.

[0050] In some embodiments, the translation matrix is ​​represented as a 4×4 matrix because Unity3D uses homogeneous coordinates to represent points in three-dimensional space. A homogeneous coordinate is a four-dimensional vector, where the first three components (x, y, z) represent the spatial position of the object, and the fourth component is used to represent scale or projection information; during a translation transformation, the fourth component remains 1. (Tx, Ty, Tz) represents the translation amount along the x, y, and z axes. Multiplying the translation matrix by the homogeneous coordinate vector of the object's original position yields the object's new position after the shift.

[0051] In some embodiments, Unity3D represents and applies the rotation of objects in three-dimensional space through rotation matrices. When a rotation matrix is ​​multiplied by a vector, it can change the direction of the vector without changing its magnitude, and maintain the chirality of the original coordinate system. Unity3D uses quaternions to store and calculate rotations, where a, b, c, and d are four real numbers, and i, j, and k are imaginary units. Its advantage is that it can avoid the universal joint deadlock problem in matrix operations and provide an efficient and non-singular way to represent three-dimensional rotations. In addition, quaternions can also be used to represent orientation, using four numbers to express direction or position.

[0052] In this embodiment, the coordinate matching module includes: Path planning unit, used to plan the assembly path through motion planning algorithm; Action decomposition unit, used to decompose the assembly task in the assembly path into discrete steps and bind each discrete step with structural data; The data conversion unit is used to convert discrete steps into control signals within the physical space layer.

[0053] In some embodiments, the path planning unit uses a motion planning algorithm, such as the improved RRT* algorithm or the artificial potential field method, to generate a high-precision, collision-free assembly path in the digital twin space.

[0054] The generation process specifically includes: when planning the path, considering constraints such as shortest time, lowest energy consumption and highest safety at the same time. For example, in automobile engine assembly, the system plans an obstacle avoidance path for the robot to bypass the cooling duct and verifies the feasibility of the path through Monte Carlo simulation; when a sudden disturbance occurs in the physical space (such as conveyor belt vibration causing a 2mm offset of the part), the path planning unit regenerates the local path based on real-time point cloud data to ensure consistency between virtual and physical space; the tolerance information in the structural data (such as the hole-shaft fit tolerance of ±0.01mm) is used as prior knowledge for path planning, and tolerance space is reserved in advance to avoid assembly interference.

[0055] In some embodiments, the motion decomposition unit divides the continuous path into millisecond-level motion sequences to ensure motion smoothness, and the motion sequences are displacement (mm) increments every 0.1 ms.

[0056] In some embodiments, the action decomposition unit associates the displacement and rotation angle of each step with the size, tolerance, and constraints of the part. Specifically, in the bearing press-fitting task, the decomposition unit binds the downward pressure (50N±2N) with the bearing inner diameter tolerance (H7 / g6) and dynamically adjusts the press-fitting speed.

[0057] In some embodiments, the data conversion unit converts discrete steps into control signals that can be executed by physical devices, and generates customized signals (such as PWM pulses, CAN bus instructions) for different actuators (such as servo motors and pneumatic clamps); it also uses timestamp alignment technology to ensure millisecond-level synchronization between virtual instructions and physical execution.

[0058] In some embodiments, the data conversion unit uses a Kalman filter to eliminate signal noise caused by electromagnetic interference or mechanical vibration, thereby improving control stability.

[0059] In this embodiment, the motion tracking unit includes a laser tracker, and the laser tracker adjusts the detection points in the physical space layer through reinforcement learning.

[0060] It is understandable that the laser tracker, as a high-precision measurement device, captures the three-dimensional coordinate data of the robot end effector, workpiece and fixture in the physical space layer in real time, while the reinforcement learning algorithm is embedded in the motion tracking unit.

[0061] In some embodiments, the reinforcement learning algorithm is a Deep Q Network, DQN.

[0062] In some embodiments, the reinforcement learning algorithm abstracts the physical environment state into a multidimensional vector, including real-time coordinate data from the laser tracker, assembly deviation sequences, environmental parameters, and historical detection point distributions. This allows the algorithm to dynamically adjust the location, density, and detection frequency of detection points. For example, if a high deviation rate is detected in a certain area, the number of detection points in that area can be increased, or the detection interval can be shortened, such as increasing the number of detection points from 5 to 10 or adjusting the detection interval from 100ms to 50ms.

[0063] In some embodiments, the reinforcement learning algorithm uses deviation detection accuracy, detection efficiency (number of effective detection points per unit time), and resource consumption (such as laser power and computing load) as optimization objectives to design a multi-objective reward function: ; in, Represents the comprehensive reward value, which is an immediate reward calculated based on the deviation detection accuracy, detection efficiency, and resource consumption. It is used to guide the reinforcement learning algorithm to converge to the global optimal solution. The weight coefficients of the three objectives of deviation detection accuracy, detection efficiency, and resource consumption in the reward function are used to balance the contribution of different objectives to the overall reward value; Indicates the deviation detection accuracy, reflecting the degree of consistency between the detected assembly deviation and the actual deviation; Indicates the detection efficiency, which is defined as the number of effective detection points per unit time, reflecting the execution speed of the detection process; Resource consumption, including laser power, computational load, and other resource usage, measures the system resource usage of the inspection process. This can be assessed through indicators such as laser tracker power consumption, and computing device CPU and memory usage.

[0064] In some embodiments, scattered detection points are randomly explored to cover the global area, gradually converging to the high deviation probability area.

[0065] In some embodiments, when the laser tracker detects that a bolt hole position has shifted by 0.2 mm due to vibration, the algorithm immediately adds three new detection points around the offset direction and increases the detection frequency of the area to 20 times per second.

[0066] In this embodiment, the detection process of the assembly deviation sequence includes: using the parametric model as the assembly reference, aligning the triangular mesh model and the parametric model to the same coordinate system; then quantifying the deviation between the triangular mesh vertices and the parametric surface and calculating the shortest distance from each mesh vertex to the parametric surface; finally, comparing the shortest distance with the design error to obtain the assembly deviation sequence.

[0067] Understandably, traditional detection methods rely heavily on manual visual inspection or fixed sensor sampling, which have the following limitations: The ability to capture complex curved surfaces or micro-deformations is insufficient. For example, when a robot assembles gears, micro-wear on the tooth surface (such as a 0.02mm dent) may be ignored. The static inspection point layout cannot adapt to the dynamic assembly process, resulting in missed detection of local deviations; Deviation data lacks structured description, making it difficult to support closed-loop feedback optimization.

[0068] In this embodiment, the parametric model is used as a theoretical benchmark, and the triangular mesh model generated by the actual scan is aligned to the same coordinate system through the ICP (Iterative Closest Point) algorithm. The alignment process adopts a multi-level optimization strategy: Coarse alignment: RANSAC algorithm based on feature points (such as hole centers and edge vertices) eliminates initial pose deviation; Fine alignment: Sub-micron registration (error < 0.005mm) is achieved by minimizing the Hausdorff distance from the global vertex to the surface.

[0069] In this embodiment, the kd-tree is used to accelerate the nearest point search, and the vertices are projected onto a parameterized surface. The deviations are assigned positive and negative signs according to whether the vertices are above or below the surface. Combined with the design error, the deviations are divided into three levels: acceptable, warning, and out-of-limit. This solves the problem that traditional voxelization or cross-sectional sampling methods cannot capture microscopic geometric fluctuations.

[0070] In some embodiments, the deviation data is integrated into a multi-dimensional sequence according to spatial distribution, severity, and time evolution characteristics.

[0071] In this embodiment, the adaptive feedback unit includes: A feature stratification subunit, used for obtaining a feature stratification sequence by a feature recognition method; An adaptive layering subunit, configured to obtain an adaptive layering sequence using an adaptive layering algorithm; The assembly sequence hierarchical subunit is used to fuse the feature hierarchical sequence with the adaptive hierarchical sequence to obtain the assembly hierarchical sequence.

[0072] In this embodiment, through the coordinated operation of the feature layering subunit, the adaptive layering subunit and the assembly sequence layering subunit, the core technical problems of the traditional assembly system in the face of complex deviations, such as a single feedback strategy, insufficient layering optimization and poor dynamic adaptability, are innovatively solved.

[0073] In some embodiments, highly significant features (such as assembly reference planes and mating hole edges) are extracted through edge detection, curvature analysis, and other methods, and their deviations are calculated. Then, based on the degree of impact of the deviations on the assembly function, the features are divided into key layers, important layers, and auxiliary layers, and differentiated correction weights are assigned.

[0074] During the assembly process, feature deviation changes are continuously monitored. When the deviation in a certain area suddenly increases, its layer priority is immediately increased and replanning is triggered.

[0075] In some embodiments, the adaptive layering subunit generates a dynamic layering strategy based on the geometric characteristics and physical constraints of the triangular mesh model. By analyzing the minimum tilt angle of the triangular facets, the layering step size is automatically adjusted to ensure the accuracy of the repair of complex surface areas. For example, in turbine blade repair, dense layering (0.01mm per layer) in the blade root transition zone can avoid stress concentration. In some embodiments, the assembly sequence layering subunit uses a reinforcement learning framework to intelligently fuse the feature layering sequence and the adaptive layering sequence, and dynamically adjusts the fusion weights of the two types of sequences based on the Q-learning algorithm to balance accuracy and efficiency; simulates the effects of different fusion strategies in the digital twin, and selects the solution that optimizes the comprehensive indicators (deviation reduction rate, time cost, resource consumption); converts the optimized assembly layering sequence into robot control instructions, such as first correcting the key layer deviation with a step size of 0.05mm, and then processing the secondary area with a step size of 0.1mm.

[0076] In this embodiment, the feature layering sequence obtained by the feature recognition method includes: initializing geometric features and assembly variables; updating the triangular mesh model to obtain the shortest distance, judging whether the shortest distance exists in the deviation group, if so, skipping it, otherwise inserting the shortest distance into the deviation group; finally, arranging the deviation groups in descending order to obtain the feature layering sequence.

[0077] In some embodiments, intelligent initialization of geometric features and assembly variables involves synchronously loading dimension data, position data, constraints (and other structural parameters) to build a priori knowledge base for feature recognition; and associating assembly process parameters with geometric features to form an iteratively updateable feature-variable matrix.

[0078] In some embodiments, a characteristic-variable matrix is ​​used to reflect the mapping relationship between expansion and compressive stress.

[0079] In this embodiment, obtaining the adaptive layering sequence by using the adaptive layering algorithm includes: calculating the layering distance by using the minimum tilt angle of the triangle facets of each layer in the triangular mesh model to obtain the adaptive layering sequence.

[0080] It's understandable that dynamic calculation of layer distances driven by the tilt angle of triangular facets systematically addresses the core technical challenges of traditional layered machining: poor geometric adaptability, a difficult balance between precision and efficiency, and unstable machining quality for complex curved surfaces. Specifically, in aircraft engine turbine blade repair, the curvature of the transition zone between the blade body and the blade root changes dramatically (the tilt angle suddenly increases from 5° to 45°). Using a fixed layer thickness of 0.1mm would create a stair-step effect (surface roughness Ra > 3.2μm) due to sparse layering in steep areas, while overly dense layering in flat areas would waste processing time.

[0081] Based on the above problems, this embodiment proposes to perform a topological analysis on each layer of the triangular mesh model, calculate the minimum tilt angle of each triangle facet, and then dynamically allocate the layer distance based on a preset tilt angle-layer thickness mapping table. The smaller the tilt angle, the denser the layering, and vice versa. The above process satisfies: ; in, is the minimum layer thickness; is the maximum layer thickness; is the tilt angle threshold; is the slope coefficient.

[0082] In this embodiment, the adaptive feedback unit further includes a replanning subunit, which adopts a Q-learning-based reinforcement learning framework to dynamically adjust the replanning strategy of the assembly sequence, specifically including: A Markov decision model is constructed, which takes the current assembly deviation sequence and the physical space layer environment state as input states, and the assembly sequence adjustment action as output action. The action selection strategy is optimized through Q value iterative update. Design a reward function to comprehensively calculate the immediate reward based on the deviation reduction rate of the replanned assembly sequence, path optimization efficiency, and resource consumption indicators, guiding the algorithm to converge to the global optimal solution; The execution results of different assembly actions are simulated in the twin space layer, and the fusion weights of the feature layered sequence and the adaptive layered sequence are dynamically adjusted through reinforcement learning to generate the optimal assembly layered sequence and feed it back to the physical space layer for execution.

[0083] It is understandable that by constructing a Markov decision model, a multi-objective reward function, and a dynamic optimization mechanism that integrates virtual and real elements, the core technical problems of traditional assembly systems in dynamic environments, such as low replanning efficiency, insufficient global optimization capabilities, and difficulty coordinating multiple constraints, have been systematically addressed. Traditional replanning methods often rely on pre-programmed rules or static optimization algorithms, making it difficult to cope with real-time disturbances in the assembly process and lack the ability to comprehensively balance multiple objectives. For example, if the vibration of the robot arm causes the alignment deviation of the battery cell tab to suddenly increase by 0.2mm, the traditional method requires interrupting the process and manually adjusting the path, which takes up to several minutes.

[0084] Based on the above problems, the applicant proposed to encode the current assembly deviation sequence, physical environment status and historical action records into a multi-dimensional state vector to build a dynamic decision-making basis; and allow the algorithm to select a variety of adjustment actions, including local path fine-tuning and global sequence rearrangement; and then continuously update the state-action value function through Q-table or deep Q-network (DQN), gradually converging to the optimal strategy.

[0085] In some embodiments, the current assembly deviation sequence is the cell tab deviation, and the cell tab deviation is 0.2 mm and the housing gap is 0.1 mm; In some embodiments, the tab lateral compensation is 0.15 mm In some embodiments, in 10 consecutive assemblies, the algorithm learns that the combined action of "prioritized compensation for tab deviation + reducing the pressing speed" can increase the overall reward value by 20%.

[0086] Specifically, for the modeling process of a specific robotic assembly system, comprehensive physical data of the robot is collected using a laser scanner, a 3D coordinate measuring machine, and embedded sensors (such as encoders and vibration sensors). More specifically, the laser scanner captures the geometric dimensions (such as connecting rod length and joint aperture) and surface topography of each robot component with micron-level accuracy; the encoder records the motion parameters (rotation angle and angular velocity) of each joint, while the vibration sensor monitors the mechanical disturbances in the assembly environment in real time. Simultaneously, tolerance data (such as fit tolerance of ±0.01mm and position tolerance of ±0.05mm) is imported from the design file to form a structured assembly constraint library. This data is transmitted to the twin modeling module via an IoT platform (such as the OPC UA protocol), where it undergoes noise filtering and coordinate alignment preprocessing to ensure data consistency.

[0087] Based on the collected data, the modeling unit fuses the original CAD model of the robot with the actual physical data. For example, if a scan reveals that the actual length of a connecting rod is 0.02mm shorter than the designed value, the system adjusts the corresponding parameters in the virtual model through reverse engineering and generates a high-fidelity triangular mesh model. Simultaneously, the extraction unit uses feature recognition algorithms (such as edge detection and curvature analysis) to annotate key geometric features: Feature points: such as the gripping point of the end effector and the center of the base mounting hole; Characteristic lines: such as the sliding track of the guide rail and the axis of the bolt hole; Feature surface: such as assembly reference plane and gear meshing contact surface.

[0088] The above features are combined with tolerance data to construct a parametric model, which serves as the geometric reference for virtual assembly.

[0089] Within the virtual environment, the system further establishes the kinematic and dynamic constraint models of the manipulator. For example, the six-axis manipulator's joint coordinate system is defined using Denavit-Hartenberg (DH) parameters, and the theoretical rotation range of each axis is calibrated based on actual motion data. Simultaneously, the tolerance chain analysis module simulates the cumulative error of multi-part assembly. For example, assuming the position tolerance of a bearing hole is ±0.03mm and the tolerance of its mating shaft is ±0.02mm, the system uses Monte Carlo simulation to predict the maximum radial runout error after assembly (e.g., 0.08mm) and optimizes the virtual assembly sequence accordingly.

[0090] Using 5G or industrial Ethernet, the sensor network platform synchronizes real-time data from the physical layer (such as joint angles, ambient temperature and humidity) with the virtual model. For example, when the robot performs a grasping action, encoder data drives the joint movement in the virtual model, while the laser tracker measures the actual position of the end effector in real time and compares it with the virtual trajectory. If a deviation is detected (such as a 0.1mm end offset due to vibration), the system immediately triggers the adaptive feedback module: Deviation quantification: Align the triangular mesh model and the parametric model to the same coordinate system, calculate the shortest distance from each mesh vertex to the theoretical surface, and generate an assembly deviation sequence; Dynamic compensation: Replan the virtual assembly path through reinforcement learning algorithms (such as Q-learning), simulate compensation actions in the twin space (such as adding a 0.1mm reverse translation), and send correction instructions to the physical robot arm.

[0091] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. The industrial Internet of Things automatic assembly system based on digital twin is characterized by: The system comprises a management platform, a sensor network platform and an object platform which are sequentially connected in communication, wherein the management platform comprises: Twin modeling module, used to collect physical space layer data and build the twin space layer; Twin assembly module, used to perform virtual assembly actions within the twin space layer; The coordinate matching module is used to synchronously map the assembly sequence of virtual assembly actions to the physical space layer; Adaptive feedback module, used to track assembly actions in the physical space layer and detect assembly deviation sequences; Among them, the adaptive feedback module includes a motion tracking unit and an adaptive feedback unit that are communicatively connected. The motion tracking unit is used to track and detect the assembly deviation sequence of the assembly action in the physical space layer. The adaptive feedback unit is used to re-plan the assembly sequence of the virtual assembly action after the assembly deviation sequence reaches a set threshold, and then provide feedback on the assembly action in the physical space layer based on the re-planning of the assembly sequence.

2. The industrial Internet of Things automatic assembly system based on digital twins according to claim 1 is characterized in that: The twin modeling module includes: A collection unit, used for collecting assembly system structure data; A modeling unit, used to perform three-dimensional modeling using structural data to obtain a twin space layer; An extraction unit, used for extracting geometric features from structural data; Structural data includes dimension data, position data, constraint data, and tolerance data; Geometric features include feature points, feature lines, and feature surfaces.

3. The industrial Internet of Things automatic assembly system based on digital twins according to claim 1 is characterized in that: The twin assembly module includes a translation unit and a rotation unit that are communicatively connected. The translation unit performs mobile assembly actions through a translation matrix, and the rotation unit performs rotational assembly actions through a rotation matrix.

4. The industrial Internet of Things automatic assembly system based on digital twins according to claim 2, characterized in that: The coordinate matching module includes: Path planning unit, used to plan the assembly path through motion planning algorithm; Action decomposition unit, used to decompose the assembly task in the assembly path into discrete steps and bind each discrete step with structural data; The data conversion unit is used to convert discrete steps into control signals within the physical space layer.

5. The industrial Internet of Things automatic assembly system based on digital twins according to claim 1, characterized in that: The motion tracking unit includes a laser tracker, which adjusts the detection points in the physical space layer through reinforcement learning.

6. The industrial Internet of Things automatic assembly system based on digital twins according to claim 5, characterized in that: The detection process of the assembly deviation sequence includes: using the parametric model as the assembly reference, aligning the triangular mesh model and the parametric model to the same coordinate system; then quantifying the deviation between the triangular mesh vertices and the parametric surface and calculating the shortest distance from each mesh vertex to the parametric surface; finally, comparing the shortest distance with the design error to obtain the assembly deviation sequence.

7. The industrial Internet of Things automatic assembly system based on digital twins according to claim 6, characterized in that: The adaptive feedback unit includes: A feature stratification subunit, used for obtaining a feature stratification sequence by a feature recognition method; An adaptive layering subunit, configured to obtain an adaptive layering sequence using an adaptive layering algorithm; The assembly sequence hierarchical subunit is used to fuse the feature hierarchical sequence with the adaptive hierarchical sequence to obtain the assembly hierarchical sequence.

8. The industrial Internet of Things automatic assembly system based on digital twins according to claim 7, characterized in that: The feature layering sequence obtained by the feature recognition method includes: initializing geometric features and assembly variables; updating the triangular mesh model to obtain the shortest distance, judging whether the shortest distance exists in the deviation group, if so, skipping it, otherwise inserting the shortest distance into the deviation group; and finally arranging the deviation groups in descending order to obtain the feature layering sequence.

9. The industrial Internet of Things automatic assembly system based on digital twins according to claim 7, characterized in that: The method of obtaining the adaptive layering sequence by using the adaptive layering algorithm includes: calculating the layering distance by the minimum tilt angle of the triangular facets of each layer in the triangular mesh model to obtain the adaptive layering sequence.

10. The industrial Internet of Things automatic assembly system based on digital twins according to claim 7, characterized in that: The adaptive feedback unit further includes a replanning subunit, which uses a Q-learning-based reinforcement learning framework to dynamically adjust the replanning strategy of the assembly sequence, specifically including: A Markov decision model is constructed, which takes the current assembly deviation sequence and the physical space layer environment state as input states, and the assembly sequence adjustment action as output action. The action selection strategy is optimized through Q value iterative update. Design a reward function to comprehensively calculate the immediate reward based on the deviation reduction rate of the replanned assembly sequence, path optimization efficiency, and resource consumption indicators, guiding the algorithm to converge to the global optimal solution; The execution results of different assembly actions are simulated in the twin space layer, and the fusion weights of the feature layered sequence and the adaptive layered sequence are dynamically adjusted through reinforcement learning to generate the optimal assembly layered sequence and feed it back to the physical space layer for execution.

Images