Multi-target assembly precision control method and system for liquid crystal display module

Abstract

The invention provides a multi-target assembly precision control method and system for a liquid crystal display module, and relates to the technical field of assembly control, and the method comprises the steps: obtaining the three-dimensional CAD data and assembly process requirements of the liquid crystal display module, and building a three-dimensional grid structure of an assembly space through a layering A * algorithm; acquiring images and force sensor data to realize real-time identification of the assembly state; carrying out assembly track planning; calculating a stress distribution state and establishing an assembly precision evaluation index system; and the robot actuator is accurately adjusted to apply compensation force. According to the method, high-precision assembly of the liquid crystal display module can be achieved, and the assembly efficiency and precision are improved.

Description

Technical Field

[0001] This invention relates to assembly control technology, and more particularly to a multi-objective assembly accuracy control method and system for liquid crystal display modules. Background Technology

[0002] With the rapid development of smart display technology, LCD modules, as core components of smart terminal devices, directly affect the display effect, reliability, and lifespan of the product due to their assembly precision. LCD modules typically consist of multiple precision components such as a display panel, backlight module, and driving circuitry, requiring micron-level precision control during assembly. Traditional LCD module assembly mainly relies on manual operation or simple automated equipment, which is insufficient to meet the current demands for high-precision and high-efficiency production.

[0003] In recent years, with the advancement of machine vision and intelligent control technologies, automated assembly systems have been gradually applied to the production of LCD display modules. Existing technologies typically use industrial cameras for visual positioning and inspection, combined with robotic actuators to complete assembly actions, and force sensors to monitor assembly forces to ensure assembly quality.

[0004] Existing assembly systems lack the ability to effectively organize and express assembly space, and cannot describe complex assembly environments at multiple levels and scales. This results in low path planning efficiency under complex assembly conditions and difficulty in coping with dynamically changing assembly conditions.

[0005] Existing technologies have limited ability to perceive the assembly status in real time, making it difficult to effectively integrate visual data and force sensing data. This makes it impossible to accurately identify the real-time status of the LCD module during the assembly process, thus affecting assembly accuracy and efficiency.

[0006] Existing assembly systems lack the ability to accurately predict and control stress distribution during the assembly process, making it difficult to assess the trend of stress changes over time. This can easily lead to excessive residual stress inside the LCD module, affecting the long-term reliability and display performance of the product. Summary of the Invention

[0007] This invention provides a multi-target assembly accuracy control method and system for liquid crystal display modules, which can solve the problems in the prior art.

[0008] A first aspect of the present invention provides a multi-target assembly accuracy control method for liquid crystal display modules, comprising:

[0009] Obtain the 3D CAD data and assembly process requirements of the LCD display module, and establish a 3D mesh structure of the assembly space through the hierarchical A* algorithm to achieve multi-level division of the assembly space;

[0010] Based on the spatial layering characteristics of the three-dimensional mesh structure, multiple high-definition industrial camera image data and force sensor data of the liquid crystal display module assembly station are collected and mapped to the corresponding mesh level to realize real-time identification of the current assembly status of the liquid crystal display module.

[0011] Based on the real-time identification results of the assembly status, the hierarchical A* algorithm is used to plan the assembly trajectory. The identified assembly space is discretized into a state space, and the corresponding assembly actions are discretized into an action space. The optimal assembly trajectory is obtained through multiple rounds of iterative optimization based on the state space and the action space.

[0012] Based on the optimal assembly trajectory and real-time force sensing data, the hierarchical A* algorithm is used to calculate the stress distribution state at different grid levels. The change trend of assembly stress over time is predicted through inter-layer data transmission. The stress distribution state and the change trend are used as the basis to establish an assembly accuracy evaluation index system.

[0013] Based on the assembly accuracy evaluation index system, the hierarchical A* algorithm is used to precisely adjust the compensation force applied by the robot actuator, thereby achieving high-precision assembly of the LCD display module.

[0014] Based on the spatial layering characteristics of the three-dimensional mesh structure, multiple high-definition industrial camera image data and force sensor data from the liquid crystal display module assembly station are collected and mapped to the corresponding mesh levels to achieve real-time identification of the current assembly status of the liquid crystal display module, including:

[0015] The three-dimensional mesh structure is divided into multi-level spatial partitions to generate multi-level mesh nodes. A mesh subdivision threshold is determined based on the assembly feature complexity of the multi-level mesh nodes. The multi-level mesh nodes are then adaptively subdivided according to the mesh subdivision threshold.

[0016] Multiple high-definition industrial cameras and force sensors are arranged at the LCD module assembly station to collect image data and force data of the LCD module assembly process. The image data of the assembly process is preprocessed to obtain image feature data, and the force data of the assembly process is extracted to obtain force feature data.

[0017] The image feature data is mapped to the corresponding grid level of the multi-level grid node to obtain the image data spatial mapping result; the force feature data is mapped to the corresponding grid level of the multi-level grid node to obtain the force data spatial mapping result.

[0018] A multimodal feature vector is constructed based on the spatial mapping results of the image data and the spatial mapping results of the force data. The assembly state probability distribution of the liquid crystal display module is calculated based on the multimodal feature vector. The current assembly state of the liquid crystal display module is identified based on the assembly state probability distribution.

[0019] Based on the real-time assembly state recognition results, the hierarchical A* algorithm is used for assembly trajectory planning. The recognized assembly space is discretized into a state space, and the corresponding assembly actions are discretized into an action space, including:

[0020] Based on the results of the real-time assembly status recognition, the current assembly status information of the liquid crystal display module is obtained. The assembly status information includes spatial position data, posture data and assembly constraint data. A status evaluation function is constructed based on the assembly status information.

[0021] The assembly state information is processed in layers using the layered A* algorithm to obtain a multi-layer search structure. In each layer of the multi-layer search structure, a state search criterion is established based on the state evaluation function. The evaluation cost of each state in the current search layer is calculated according to the state search criterion.

[0022] The optimal expansion state is selected based on the evaluation cost, and the optimal expansion state has the minimum evaluation cost. The optimal expansion state is then used as the starting state for the next layer of search.

[0023] The assembly space in the assembly state information is discretized into a state space according to a preset accuracy requirement. The state nodes in the state space are prioritized according to the evaluation cost. The state nodes adjacent to the initial state are selected as candidate expansion nodes. The assembly actions corresponding to the state space are discretized into an action space.

[0024] The assembly state information is processed hierarchically using the hierarchical A* algorithm to obtain a multi-layer search structure. In each layer of the multi-layer search structure, state search criteria are established based on the state evaluation function, including:

[0025] The assembly state information is divided into multiple state levels according to the resolution using the hierarchical A* algorithm, and a multi-level search structure is constructed. The state vector in each state level records the position and attitude information of the assembled parts. The search range is limited by the set of state variables based on the position and attitude information.

[0026] A chaotic mapping evaluation function is constructed based on the set of state variables. The position and attitude information is nonlinearly mapped and transformed according to the chaotic control parameters in the chaotic mapping evaluation function. The changes in the assembly state are highlighted based on the state values ​​after the nonlinear mapping transformation. The changes are dynamically weighted according to the changes by a preset weight coefficient.

[0027] The weighted change features are dynamically amplified using the chaotic control parameters. The differences between state features are continuously amplified based on the chaotic control parameters. The evaluation accuracy of the hierarchical A* algorithm is improved based on the amplified differences. A state search criterion is constructed based on the amplified differences and the corresponding state features.

[0028] The optimal assembly trajectory, obtained through multiple rounds of iterative optimization based on the state space and the action space, includes:

[0029] A basic framework for assembly trajectory is constructed, comprising a state space and an action space. The basic framework includes multiple state nodes and corresponding action instructions. An initial comprehensive evaluation value of the basic framework is calculated based on the state nodes and the action instructions.

[0030] The initial state transition probability between adjacent state nodes is calculated based on the initial comprehensive evaluation value. The initial state transition probability decreases as the initial comprehensive evaluation value increases. The assembly trajectory basic framework is optimized in the first round based on the initial state transition probability to obtain the first round assembly trajectory.

[0031] Calculate the first round comprehensive evaluation value of the first round assembly trajectory, determine the first round trajectory update amount based on the first round comprehensive evaluation value, and superimpose the first round trajectory update amount with the basic framework of the assembly trajectory to obtain the second round assembly trajectory;

[0032] The second round of comprehensive evaluation value is calculated based on the second round of assembly trajectory. The initial state transition probability is updated according to the second round of comprehensive evaluation value to obtain the state transition probability. The second round of assembly trajectory is optimized based on the state transition probability to obtain the third round of assembly trajectory.

[0033] Determine whether the difference between the second-round assembly trajectory and the third-round assembly trajectory is less than a preset error threshold, and whether the second-round comprehensive evaluation value meets the optimization objective. When the preset error threshold and the optimization objective are met, iterative convergence is determined, and the second-round assembly trajectory is output as the optimal assembly trajectory.

[0034] Based on the optimal assembly trajectory and real-time force sensing data, the layered A* algorithm is used to calculate the stress distribution at different mesh levels, including:

[0035] A multi-layer mesh system is constructed based on the hierarchical A* algorithm. The multi-layer mesh system contains multiple mesh levels. Adjacent mesh levels are connected through a biomimetic interface. The biomimetic interface dynamically adjusts the interlayer connection strength according to the stress distribution state during the assembly process, and the stress distribution between the mesh levels is transferred based on the interlayer connection strength.

[0036] The local stress distribution during the assembly process is calculated at each grid level, and the stress transmission path under the optimal assembly trajectory is determined based on the local stress distribution. The local stress value of the grid level is monitored, and when the local stress value exceeds a preset stress threshold, the bionic interface undergoes reversible deformation to absorb assembly overload stress through the reversible deformation.

[0037] The local stress distribution is reconstructed based on the degree of deformation of the bionic interface. The original stress distribution is transformed into an optimized stress distribution through dynamic adjustment of the bionic interface. The maximum stress value, minimum stress value, and average stress value in the optimized stress distribution are calculated. The uniformity of the local stress distribution is evaluated based on the maximum stress value, the minimum stress value, and the average stress value. When the uniformity of the local stress distribution meets a preset uniformity threshold, the final stress distribution state is output.

[0038] Predicting the trend of assembly stress over time through interlayer data transfer, and establishing an assembly accuracy evaluation index system based on the stress distribution and the trend of change, including:

[0039] Interlayer data records the stress transfer state between adjacent assembly layers, and the time-domain distribution of assembly stress is realized through the stress transfer state;

[0040] Based on the time-domain distribution, data transmission is predicted, and the stress change trend is analyzed according to the transmission law of the assembly stress. The stress transmission state and the stress change trend are used to construct an assembly accuracy evaluation index system.

[0041] A second aspect of the present invention provides a multi-target assembly accuracy control system for liquid crystal display modules, comprising:

[0042] The first unit is used to acquire the 3D CAD data and assembly process requirements of the LCD display module, and to establish a 3D mesh structure of the assembly space through the hierarchical A* algorithm to realize the multi-level division of the assembly space.

[0043] The second unit is used to collect multiple high-definition industrial camera image data and force sensor data of the liquid crystal display module assembly station based on the spatial layering characteristics of the three-dimensional mesh structure, and map them to the corresponding mesh level to realize real-time identification of the current assembly status of the liquid crystal display module.

[0044] The third unit is used to plan the assembly trajectory based on the real-time identification result of the assembly state and the hierarchical A* algorithm. The identified assembly space is discretized into a state space and the corresponding assembly actions are discretized into an action space. The optimal assembly trajectory is obtained through multiple rounds of iterative optimization based on the state space and the action space.

[0045] The fourth unit is used to calculate the stress distribution state at different grid levels using the hierarchical A* algorithm based on the optimal assembly trajectory and real-time force sensing data, predict the trend of assembly stress over time through inter-layer data transmission, and establish an assembly accuracy evaluation index system based on the stress distribution state and the trend of change.

[0046] The fifth unit is used to precisely adjust the compensation force applied by the robot actuator based on the assembly accuracy evaluation index system and the hierarchical A* algorithm to achieve high-precision assembly of the liquid crystal display module.

[0047] A third aspect of the present invention provides an electronic device, comprising:

[0048] processor;

[0049] Memory used to store processor-executable instructions;

[0050] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0051] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0052] The beneficial effects of this application are as follows:

[0053] This invention establishes a three-dimensional mesh structure of the assembly space through a hierarchical A* algorithm, realizing multi-level division of the assembly space, making the assembly state recognition more accurate, and improving the efficiency and accuracy of assembly trajectory planning, which can adapt to the fine operation requirements in complex assembly environments.

[0054] This invention combines multi-sensor fusion technology to map high-definition industrial camera image data and force sensor data to corresponding grid layers, enabling real-time and accurate identification of the assembly status of the liquid crystal display module. It also predicts the trend of assembly stress changes through inter-layer data transmission, effectively avoiding excessive stress or positional deviation problems during the assembly process.

[0055] This invention constructs an assembly accuracy evaluation index system based on the hierarchical A* algorithm. By calculating the stress distribution state at different grid levels, it can accurately adjust the compensation force applied by the robot actuator, effectively solving the problems of difficult accuracy control, low efficiency, and poor adaptability in traditional assembly methods, and significantly improving the assembly accuracy and yield of LCD display modules. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the multi-target assembly accuracy control method for liquid crystal display modules according to an embodiment of the present invention.

[0057] Figure 2 This is a flowchart of the assembly trajectory planning based on the hierarchical A algorithm according to an embodiment of the present invention;

[0058] Figure 3 This is a flowchart of stress distribution optimization based on the hierarchical A algorithm in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0060] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0061] Figure 1 This is a flowchart illustrating a multi-target assembly accuracy control method for liquid crystal display modules according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] Obtain the 3D CAD data and assembly process requirements of the LCD display module, and establish a 3D mesh structure of the assembly space through the hierarchical A* algorithm to achieve multi-level division of the assembly space;

[0063] Based on the spatial layering characteristics of the three-dimensional mesh structure, multiple high-definition industrial camera image data and force sensor data of the liquid crystal display module assembly station are collected and mapped to the corresponding mesh level to realize real-time identification of the current assembly status of the liquid crystal display module.

[0064] Based on the real-time identification results of the assembly status, the hierarchical A* algorithm is used to plan the assembly trajectory. The identified assembly space is discretized into a state space, and the corresponding assembly actions are discretized into an action space. The optimal assembly trajectory is obtained through multiple rounds of iterative optimization based on the state space and the action space.

[0065] Based on the optimal assembly trajectory and real-time force sensing data, the hierarchical A* algorithm is used to calculate the stress distribution state at different grid levels. The change trend of assembly stress over time is predicted through inter-layer data transmission. The stress distribution state and the change trend are used as the basis to establish an assembly accuracy evaluation index system.

[0066] Based on the assembly accuracy evaluation index system, the hierarchical A* algorithm is used to precisely adjust the compensation force applied by the robot actuator, thereby achieving high-precision assembly of the LCD display module.

[0067] In one optional implementation, based on the spatial layering characteristics of the three-dimensional mesh structure, multiple high-definition industrial camera image data and force sensor data of the liquid crystal display module assembly station are collected and mapped to the corresponding mesh level to achieve real-time identification of the current assembly status of the liquid crystal display module, including:

[0068] The three-dimensional mesh structure is divided into multi-level spatial partitions to generate multi-level mesh nodes. A mesh subdivision threshold is determined based on the assembly feature complexity of the multi-level mesh nodes. The multi-level mesh nodes are then adaptively subdivided according to the mesh subdivision threshold.

[0069] Multiple high-definition industrial cameras and force sensors are arranged at the LCD module assembly station to collect image data and force data of the LCD module assembly process. The image data of the assembly process is preprocessed to obtain image feature data, and the force data of the assembly process is extracted to obtain force feature data.

[0070] The image feature data is mapped to the corresponding grid level of the multi-level grid node to obtain the image data spatial mapping result; the force feature data is mapped to the corresponding grid level of the multi-level grid node to obtain the force data spatial mapping result.

[0071] A multimodal feature vector is constructed based on the spatial mapping results of the image data and the spatial mapping results of the force data. The assembly state probability distribution of the liquid crystal display module is calculated based on the multimodal feature vector. The current assembly state of the liquid crystal display module is identified based on the assembly state probability distribution.

[0072] When performing multi-level spatial partitioning of the 3D mesh structure, the assembly space is divided into multiple mesh levels, each with different mesh density and resolution. Specifically, an octree structure is used for spatial partitioning, progressively subdividing from a coarse-grained top mesh to a fine-grained bottom mesh. For the assembly scenario of a 55-inch LCD display module, the assembly space dimensions are 1200mm × 800mm × 300mm. The top mesh size is set to 300mm × 200mm × 75mm, forming a 4×4×4 mesh structure. The second mesh layer has dimensions of 75mm × 50mm × 18.75mm, subdivided into four parts in each dimension relative to the top mesh. The third mesh layer has dimensions of 18.75mm × 12.5mm × 4.69mm. The fourth mesh layer has dimensions of 4.69mm × 3.13mm × 1.17mm. The bottommost grid size is 1.17mm × 0.78mm × 0.29mm, which is sufficient to capture the fine structural features in the assembly of the LCD module, such as clips, screw holes, and positioning pins.

[0073] The mesh subdivision threshold is determined based on the assembly feature complexity of multi-level mesh nodes. Assembly feature complexity is comprehensively evaluated through three dimensions: geometric complexity, visual feature richness, and mechanical behavior complexity. Geometric complexity is calculated based on local curvature, edge density, and surface irregularity, with a value range of 0 to 100. Visual feature richness is calculated based on texture variation, color contrast, and feature point density, with a value range of 0 to 100. Mechanical behavior complexity is calculated based on contact stiffness variation, friction coefficient distribution, and stress concentration, with a value range of 0 to 100. For the assembly area of ​​the LCD module back panel and front frame, the geometric complexity is 85, the visual feature richness is 72, the mechanical behavior complexity is 78, and the overall complexity is 78.5. The mesh subdivision threshold is set as follows: when the overall complexity is greater than 70, subdivision proceeds to the fifth level; when the overall complexity is between 40 and 70, subdivision proceeds to the fourth level; and when the overall complexity is less than 40, subdivision proceeds to the third level.

[0074] Adaptive subdivision of multi-level mesh nodes is performed based on a mesh subdivision threshold. The adaptive subdivision process dynamically adjusts the mesh density according to the complexity of local areas, using finer meshes for complex areas and coarser meshes for simpler areas to balance computational efficiency and recognition accuracy. The adaptive subdivision algorithm employs a four-step iterative process: evaluating the overall complexity of the current mesh node; comparing the complexity with the subdivision threshold; subdividing nodes exceeding the threshold; and recursively evaluating newly generated child nodes. For the snap-fit ​​connection area of ​​the LCD module, due to its complex geometry and multiple contact points, the overall complexity is 88, exceeding the threshold of 70. Therefore, it is subdivided to the fifth layer, with a mesh size of 1.17mm × 0.78mm × 0.29mm. For the flat display area, the overall complexity is only 35, so it is subdivided only to the third layer, with a mesh size of 18.75mm × 12.5mm × 4.69mm. After adaptive subdivision, the total number of mesh nodes in the entire assembly space is approximately 25,000, a significant reduction compared to the 120,000 nodes in uniform subdivision, thus improving computational efficiency.

[0075] Multiple high-definition industrial cameras and force sensors are deployed at the LCD module assembly station. The camera arrangement employs a multi-view coverage strategy to ensure no blind spots in the assembly area. Specifically, two top-view cameras are mounted at the top, with a resolution of 4096×3072 pixels, a focal length of 12mm, and a field of view of 60°, covering the entire assembly area; four side-view cameras are mounted on the sides, with a resolution of 2048×1536 pixels, a focal length of 8mm, and a field of view of 75°, observing the assembly process from four different sides; and two macro cameras are mounted at close range, with a resolution of 1920×1080 pixels, a focal length of 35mm, and a field of view of 30°, used to observe details at key connection points. All cameras operate at a frame rate of 30 frames per second, an exposure time of 10 milliseconds, and an aperture of F4.0. Force sensors, including one six-dimensional force / torque sensor and four single-axis force sensors, are positioned at the end of the assembly tools and on the assembly station base. A six-dimensional force / torque sensor is installed at the end of the robotic arm, with a measurement range of Fx, Fy: ±100N, Fz: ±200N, Mx, My, Mz: ±10Nm, and a sampling frequency of 1000Hz. It is used to measure contact forces and torques during the assembly process. Four single-axis force sensors are installed at the four corners of the assembly station, with a measurement range of 0-500N and a sampling frequency of 500Hz. They are used to monitor the distribution of overall support forces.

[0076] The assembly process of the LCD module is used to collect image and force data. To improve image quality, a ring-shaped LED light source with a color temperature of 5000K, adjustable brightness, and an illuminance of 1000 lux is used at the assembly station. The acquired raw image data is in RGB format, with 3 channels per pixel and 8 bits per channel. For 8 cameras, approximately 5.6 GB of raw image data is generated per second. During force data acquisition, a six-dimensional force / torque sensor generates 6000 data points per second (1000Hz × 6D), and four single-axis force sensors generate 2000 data points per second (500Hz × 4). The raw force data includes sensor readings, timestamps, and sensor status information.

[0077] Image feature data is obtained by preprocessing the assembly process image data. Image preprocessing includes four steps: distortion correction, noise removal, illumination equalization, and image registration. Distortion correction uses Zhang's calibration method, obtaining camera intrinsic parameters and distortion coefficients through a 12×9 checkerboard calibration board, achieving a distortion correction accuracy better than 0.1 pixels. Noise removal employs a bilateral filtering algorithm, preserving edge features while removing Gaussian noise. The filter kernel size is 5×5, with a spatial domain standard deviation of 1.5 and a value domain standard deviation of 25. Illumination equalization uses adaptive histogram equalization, dividing the image into 8×8 blocks. Each block undergoes independent histogram equalization, and the results are then merged using bilinear interpolation to improve local contrast. Image registration uses a feature point matching method, extracting SIFT feature points, removing outlier matches using the RANSAC algorithm, calculating the transformation matrix, and aligning multiple camera views to a unified coordinate system. After preprocessing, image feature data is extracted, including edge features, corner features, texture features, and color features. Edge features were extracted using the Canny operator with thresholds set to 50 and 150; corner features were extracted using the Harris corner detection algorithm with a response threshold set to 0.01; texture features were extracted using Local Binary Pattern (LBP) with a circular neighborhood of radius 2 and 8 sampling points; color features were extracted using HSV color histograms, with the H channel divided into 18 bins and the S and V channels each divided into 3 bins.

[0078] Force feature data is obtained by extracting features from the force data during the assembly process. Force feature extraction includes three steps: filtering, feature calculation, and feature normalization. Filtering employs a combination of low-pass filtering and median filtering. The low-pass filter cutoff frequency is 50Hz to remove high-frequency noise; the median filter window size is 5 to remove impulse noise. Feature calculation is performed in both the time and frequency domains. Time-domain features include mean, standard deviation, maximum, minimum, peak, and root mean square (RMS) values. Frequency-domain features are obtained through Fast Fourier Transform (FFT) and include the dominant frequency, power spectral density, and band energy distribution. For the snap-fit ​​assembly process of the LCD module, typical force feature data includes: a mean force of 12.5N, a standard deviation of 2.8N, a maximum force of 35.6N, a minimum force of 0.2N, a peak force of 35.6N, and an RMS value of 12.8N in the Fz direction. The frequency-domain features show a dominant frequency of 2-5Hz, reflecting the periodic pressing action during assembly. Feature normalization uses the min-max normalization method to map each feature value to the [0,1] interval, which facilitates subsequent processing.

[0079] Image feature data is mapped to the corresponding grid levels of multi-level grid nodes to obtain the spatial mapping result of image data. The mapping process first performs camera calibration and 3D reconstruction to establish the correspondence between pixel coordinates and 3D spatial coordinates. Camera calibration uses Zhang's calibration method, with a reprojection error of less than 0.2 pixels; 3D reconstruction uses a multi-view stereo vision algorithm with a depth accuracy of ±0.5mm. For each feature point detected in the image, its 3D spatial coordinates are calculated, and then its corresponding grid node is determined. In the specific implementation, an octree spatial index structure is used, with a query complexity of O(log n), where n is the total number of grid nodes. For example, for a feature point detected on the front frame of an LCD display module, its image coordinates are (1024, 768), and the spatial coordinates obtained through 3D reconstruction are (450mm, 320mm, 25mm). This point is located in the fourth-level grid node (96, 102, 5). After the image feature mapping is completed, each grid node is associated with a set of image feature descriptors, including the number of feature points, edge density, corner density, main direction distribution, and color distribution. For grid nodes in complex regions, the average number of associated feature points is 120-150; for grid nodes in simple regions, the average number of associated feature points is 20-40.

[0080] Force feature data is mapped to the corresponding mesh level of multi-level mesh nodes to obtain the force data spatial mapping result. Force data spatial mapping is based on contact point location estimation and a force transmission model. The contact point location is calculated through inversion of force / torque sensor data. The basic principle is that when the end effector of the robotic arm is subjected to an external force, the force and torque measured by the six-dimensional force / torque sensor satisfy a specific relationship, through which the point of force application can be calculated. The force transmission model considers both rigid body transmission and elastic deformation. Rigid body transmission is applicable to rigid components, while elastic deformation is applicable to flexible components such as circuit boards and flexible screens. For each estimated contact point, its corresponding mesh node is determined, and the force feature data is associated with that node. During the assembly of the LCD module, when the back panel contacts the front frame, four main contact points are detected, with spatial coordinates of (320mm, 240mm, 15mm), (320mm, 560mm, 15mm), (880mm, 240mm, 15mm), and (880mm, 560mm, 15mm), corresponding to four snap-fit ​​positions. The normal contact force at each contact point is approximately 15 N, and the tangential friction force is approximately 3 N. These force data are mapped to the corresponding fifth-layer mesh nodes to form a force characteristic distribution map.

[0081] Multimodal feature vectors are constructed based on the spatial mapping results of image data and force data. These vectors fuse visual and force features to form a comprehensive description of the assembly state. An early fusion strategy is employed, directly connecting different modal features at the feature level. For each grid node, a feature vector is constructed, containing a visual feature sub-vector and a force feature sub-vector for that node. The visual feature sub-vector contains 64-dimensional edge features, 32-dimensional corner features, 128-dimensional texture features, and 54-dimensional color features, totaling 278 dimensions; the force feature sub-vector contains 18-dimensional temporal features and 24-dimensional frequency domain features, totaling 42 dimensions. The fused feature vector has a dimension of 320. To handle scale differences between different modal features, Z-score standardization is used, subtracting the mean from each feature and dividing by the standard deviation to ensure that the mean of each feature distribution is 0 and the standard deviation is 1. To reduce feature dimensionality and improve computational efficiency, principal component analysis (PCA) is used for dimensionality reduction, retaining the principal components needed to explain 95% of the variance. Typically, 320-dimensional features are reduced to 80-100 dimensions.

[0082] The assembly state probability distribution of the LCD module is calculated based on multimodal feature vectors. The assembly state is defined as a set of discrete states, including 10 key states such as "initial alignment," "partial insertion," "full insertion," and "locking." A Gaussian Mixture Model (GMM) is used to calculate the state probability distribution, establishing a probability model for each assembly state. Model training utilizes a large amount of labeled data containing multimodal feature vectors under different assembly states. For a 55-inch LCD module assembly, the training dataset contains 5000 samples, covering 10 assembly states, with approximately 500 samples per state. The GMM model uses 5-8 Gaussian components for each state, and the covariance matrix is ​​in diagonal form to simplify computational complexity. The expectation-maximization (EM) algorithm is used for training, with 100 iterations and a convergence threshold of 1e-6. After training, the model parameters include the weights, mean vector, and covariance matrix of each Gaussian component. For newly observed multimodal feature vectors, the posterior probability of their belonging to each assembly state is calculated, forming a probability distribution. For example, the feature vector of a certain observation yields the following probability distribution: "Initial alignment" 0.05, "Partial insertion" 0.15, "Full insertion" 0.75, "Locked and fixed" 0.03, and other states 0.02.

[0083] The system identifies the current assembly state of the LCD module based on the probability distribution of assembly states. The state recognition uses the maximum a posteriori probability criterion, selecting the state with the highest probability as the current recognition result. To improve recognition stability, a temporal smoothing strategy is introduced, considering the state probability distribution across multiple consecutive frames and eliminating instantaneous fluctuations through weighted averaging or median filtering. The temporal window size is set to 5 frames, approximately 0.17 seconds. To handle ambiguity during state transitions, a probability threshold of 0.65 is set. When the highest probability falls below the threshold, the current state is marked as a "transitional state," awaiting more definitive observations. For LCD module assembly, the state recognition accuracy reaches 96.5%, with an average recognition delay of 0.2 seconds, meeting the requirements for real-time assembly monitoring. Typical assembly states reliably detected by the system include: when all four clips are fully inserted and the force sensor detects a characteristic "click," it is identified as a "fully inserted" state; when only two of the four clips are fully inserted, it is identified as a "partially inserted" state; when excessive clip deformation or contact force exceeding 40N is detected, it is identified as an "assembly abnormality" state and triggers an early warning.

[0084] Through the above method, this invention achieves real-time identification of the assembly status of liquid crystal display modules based on a three-dimensional mesh structure, providing reliable status perception capabilities for automated assembly and significantly improving assembly efficiency and success rate. Experimental verification shows that in the assembly scenario of a 55-inch liquid crystal display module, the method achieves a recognition accuracy of 96.5%, improves the assembly success rate by 18%, and increases assembly efficiency by 32%.

[0085] In one optional implementation, based on the real-time identification result of the assembly state, the hierarchical A* algorithm is used for assembly trajectory planning. The identified assembly space is discretized into a state space, and the corresponding assembly actions are discretized into an action space, including:

[0086] Based on the results of the real-time assembly status recognition, the current assembly status information of the liquid crystal display module is obtained. The assembly status information includes spatial position data, posture data and assembly constraint data. A status evaluation function is constructed based on the assembly status information.

[0087] The assembly state information is processed in layers using the layered A* algorithm to obtain a multi-layer search structure. In each layer of the multi-layer search structure, a state search criterion is established based on the state evaluation function. The evaluation cost of each state in the current search layer is calculated according to the state search criterion.

[0088] The optimal expansion state is selected based on the evaluation cost, and the optimal expansion state has the minimum evaluation cost. The optimal expansion state is then used as the starting state for the next layer of search.

[0089] The assembly space in the assembly state information is discretized into a state space according to a preset accuracy requirement. The state nodes in the state space are prioritized according to the evaluation cost. The state nodes adjacent to the initial state are selected as candidate expansion nodes. The assembly actions corresponding to the state space are discretized into an action space.

[0090] like Figure 2 As shown, the method includes:

[0091] Based on the real-time assembly status recognition results, the current assembly status information of the LCD module is obtained. This assembly status information mainly includes three types of data: spatial position data, attitude data, and assembly constraint data. Spatial position data is represented using a three-dimensional Cartesian coordinate system (x, y, z). For assembling a 55-inch LCD module, the accuracy requirement for spatial position data is ±0.05mm, and the sampling frequency is 100Hz. The initial position of a typical LCD module backplate before assembly is (500mm, 300mm, 50mm), relative to the robotic arm base coordinate system. Attitude data is represented using Euler angles (α, β, γ) or quaternions (q0, q1, q2, q3). The accuracy requirement for Euler angles is ±0.1°, and for quaternions, it is ±0.001. The initial attitude angle of the LCD module backplate is (0°, 0°, 180°), indicating that the backplate is horizontally placed and rotated 180°. Assembly constraint data includes geometric constraints, force constraints, and velocity constraints. Geometric constraints define the allowable range of motion of the assembled parts, such as a 0.3mm clearance between the back panel and the front frame of the LCD module. Force constraints define the allowable range of contact forces during assembly, such as a maximum allowable contact force of 20N when assembling the back panel and the front frame. Velocity constraints define the speed limits during assembly, such as a maximum linear velocity of 50mm / s during the initial approach phase and a maximum linear velocity reduced to 5mm / s during the precise insertion phase.

[0092] A state evaluation function is constructed based on the assembly state information. This function evaluates the quality of the current assembly state, providing heuristic information for the hierarchical A* algorithm. The state evaluation function comprehensively considers four aspects: distance cost, attitude cost, constraint violation cost, and operational difficulty cost. Distance cost reflects the Euclidean distance between the current and target positions; for LCD module assembly, distance cost accounts for 40% of the total evaluation cost. Attitude cost reflects the difference between the current and target attitudes, obtained by calculating the weighted sum of Euler angles; attitude cost accounts for 30% of the total evaluation cost. Constraint violation cost reflects the degree to which the current state violates assembly constraints. For example, there is a parallelism requirement between the optical film and polarizer of the LCD module; for every 0.1° increase in deviation, the constraint violation cost increases by 5 units, accounting for 20% of the total evaluation cost. Operational difficulty cost reflects the difficulty of transitioning from the current state to the target state; considering factors such as spatial confinement, visual obstruction, and force control difficulty, operational difficulty cost accounts for 10% of the total evaluation cost.

[0093] A hierarchical A* algorithm is used to process assembly state information hierarchically, resulting in a multi-layered search structure. This hierarchical processing is based on spatial discretization at different precision levels, typically consisting of 3 to 5 layers. For LCD module assembly, a 4-layer search structure is employed: a coarse search layer, an intermediate search layer, a fine search layer, and a precise insertion layer. The coarse search layer has a spatial resolution of 20mm and an angular resolution of 10°, primarily used for planning large-scale movement paths. The intermediate search layer has a spatial resolution of 5mm and an angular resolution of 5°, used for obstacle avoidance and path refinement. The fine search layer has a spatial resolution of 1mm and an angular resolution of 1°, used for accurately approaching the assembly position. The precise insertion layer has a spatial resolution of 0.1mm and an angular resolution of 0.1°, used to achieve final precise assembly. Connections between layers are established through a state mapping function, with the search results from the upper layer serving as the starting point and constraint region for the lower layer's search.

[0094] In each layer of the multi-layered search structure, state search criteria are established based on the state evaluation function. These criteria define how to select and expand nodes at the current layer. The search criteria include expansion priority, pruning conditions, and termination conditions. Expansion priority is based on the evaluation cost of a node; nodes with lower evaluation costs are expanded first. Pruning conditions are used to reduce the search space and improve search efficiency, including boundary pruning, collision detection pruning, and evaluation threshold pruning. For example, a node is pruned when its minimum distance to an obstacle is less than 5mm; a node is pruned when its evaluation cost exceeds 150% of the current optimal path's evaluation cost. Termination conditions define the criteria for ending the search at the current layer, including finding the target state, reaching the maximum search depth, or timeout. For LCD module assembly, the maximum search depth for the coarse search layer is 100, with a search time limit of 2 seconds; the maximum search depth for the intermediate search layer is 200, with a search time limit of 5 seconds; the maximum search depth for the fine search layer is 300, with a search time limit of 10 seconds; and the maximum search depth for the precise insertion layer is 100, with a search time limit of 5 seconds.

[0095] The evaluation cost of each state in the current search layer is calculated according to the state search criterion. The evaluation cost consists of two parts: the actual cost *g* from the initial state to the current state and the estimated cost *h* from the current state to the target state. The actual cost *g* is calculated by accumulating the transformation costs of each segment on the path, taking into account distance changes, attitude changes, and the required operational difficulty. The estimated cost *h* is calculated using a heuristic function, which must be permissible, i.e., it will not overestimate the cost of the target. For LCD module assembly, the estimated cost *h* is calculated using weighted Manhattan distance, with the weighting coefficients dynamically adjusted according to the movement difficulty along different axes. For example, the weight for the z-axis direction (vertical lifting) is 1.5, the weight for movement in the xy plane is 1.0, and the weight for rotation operations is 2.0. The calculated total evaluation cost *f* is equal to the sum of the *g* and *h* values; the smaller the *f* value, the better the state.

[0096] The optimal expansion state is selected based on the evaluation cost; the optimal expansion state is the state with the lowest evaluation cost in the current search layer. In the hierarchical A* algorithm, an open list is maintained, storing the state nodes to be expanded in order of evaluation cost. Each time, the node with the lowest evaluation cost is selected from the open list as the current expansion node. For LCD module assembly, assuming there are three candidate expansion nodes in the coarse search layer with evaluation costs of 85, 72, and 93 respectively, the node with an evaluation cost of 72 is selected as the optimal expansion state. If multiple nodes have the same evaluation cost, the node with the smaller g-value is preferred, i.e., the node closer to the initial state is preferred.

[0097] The optimal extended state is used as the starting state for the next layer of search. After the current layer's search is completed, the search results are passed to the next layer as its starting state and constraints. During this process, state mapping is required to map the discrete state of the current layer to a more refined discrete representation in the next layer. For example, when the optimal path point (500mm, 300mm, 50mm) found by the coarse search layer is mapped to the intermediate search layer, a spherical search region with a radius of 30mm centered on that point is generated. The intermediate search layer then performs a more refined search within this region with a resolution of 5mm. Simultaneously, path constraints, such as obstacle avoidance information, velocity constraints, and force constraints, are also passed to the next layer.

[0098] The assembly space in the assembly status information is discretized into a state space according to preset accuracy requirements. The discretization process involves spatial sampling based on the accuracy requirements of different search layers. For LCD module assembly, the working space range is 1000mm × 800mm × 500mm. In the coarse search layer, a spatial resolution of 20mm is used, generating 50 × 40 × 25 = 50,000 spatial points; in the intermediate search layer, a spatial resolution of 5mm is used within 30mm of the coarse search result, typically generating several hundred to several thousand spatial points; in the fine search layer, a spatial resolution of 1mm is used within 10mm of the intermediate search result, generating several hundred spatial points; and in the precise insertion layer, a spatial resolution of 0.1mm is used within 2mm of the fine search result, generating several hundred spatial points. The attitude space is also discretized according to the corresponding angular resolution. For LCD modules with complex shapes, the geometric constraints of the assembly components, such as slot shape and snap-fit ​​position, also need to be considered, as these constraints further limit the effective state space.

[0099] The state nodes in the state space are prioritized based on their evaluation cost, f. Nodes with smaller f values ​​have higher priority. The sorting is implemented using a binary heap to ensure that the node with the smallest f value is retrieved each time, resulting in a time complexity of O(log n). To improve search efficiency, heuristic strategies can be used to dynamically adjust node priorities. For example, if a search in a certain direction is found to be more promising, the h value of nodes in that direction can be appropriately reduced to increase their priority; when the search gets stuck in a local optimum, exploration can be temporarily increased, giving higher priority to some nodes with high f values. For the precise insertion operation of the LCD module, priorities can be dynamically adjusted based on force sensor feedback; when an increase in contact force is detected, directions that reduce the contact force are prioritized.

[0100] The adjacent state nodes are selected as candidate expansion nodes. An adjacent state node is a state that can be reached from the initial state through a single basic operation. Basic operations typically include translation along a coordinate axis and rotation around a coordinate axis. For LCD module assembly, in the coarse search layer, adjacent states are defined as nodes with a Manhattan distance of 1, i.e., adjacent grid points; in the fine search layer, adjacent states can be defined more flexibly, including diagonal movements and compound movements. For example, for the position (500mm, 300mm, 50mm), its 6 connected adjacent nodes are (520mm, 300mm, 50mm), (480mm, 300mm, 50mm), (500mm, 320mm, 50mm), (500mm, 280mm, 50mm), (500mm, 300mm, 70mm), and (500mm, 300mm, 30mm). Collision detection is required when generating candidate expansion nodes to exclude nodes that cause collisions. Collision detection employs a combination of hierarchical bounding boxes and precise mesh models. The coarse search layer uses a simplified geometric model, while the fine search layer uses a precise geometric model.

[0101] The assembly actions corresponding to the state space are discretized into an action space, which consists of a series of basic assembly operations. Each operation corresponds to a state transition in the state space. Basic assembly operations include translation, rotation, and compound operations. A translation operation is defined as moving a specific distance along a specific direction, such as "moving 20mm to the right" or "lifting 30mm upwards." A rotation operation is defined as rotating a specific angle around a specific axis, such as "rotating 90° clockwise around the Z-axis" or "rotating 45° counterclockwise around the X-axis." A compound operation is a combination of translation and rotation, such as a spiral insertion operation. For LCD module assembly, 30 basic operations are defined, including 6 single-axis translations, 6 single-axis rotations, and 18 compound operations. Each operation is associated with execution parameters, such as speed, acceleration, and force control parameters. For example, the precise insertion operation parameters for the LCD module backplate are: approach speed 5mm / s, insertion speed 2mm / s, maximum contact force 15N, and compliance stiffness 0.5mm / N. The discretization granularity of the action space corresponds to that of the state space, ensuring that there are clear action instructions from one state node to the adjacent state node.

[0102] Through the above method, this invention realizes assembly trajectory planning for liquid crystal display modules based on the hierarchical A* algorithm, discretizing the assembly space into a state space and the assembly actions into an action space. This method has achieved significant results in practical applications. Compared with traditional methods, planning time is reduced by 65%, from an average of 12 seconds to 4.2 seconds; planned path length is reduced by 28%; assembly success rate is increased by 15%, from 83% to 98%; and assembly efficiency is improved by 45%, with the average assembly time reduced from 68 seconds to 37 seconds. This method is particularly suitable for the automated assembly of precision electronic products, such as liquid crystal display modules, flexible circuit boards, and precision optical components.

[0103] In one optional implementation, the assembly state information is processed hierarchically using the hierarchical A* algorithm to obtain a multi-layer search structure. In each layer of the multi-layer search structure, state search criteria are established based on the state evaluation function, including:

[0104] The assembly state information is divided into multiple state levels according to the resolution using the hierarchical A* algorithm, and a multi-level search structure is constructed. The state vector in each state level records the position and attitude information of the assembled parts. The search range is limited by the set of state variables based on the position and attitude information.

[0105] A chaotic mapping evaluation function is constructed based on the set of state variables. The position and attitude information is nonlinearly mapped and transformed according to the chaotic control parameters in the chaotic mapping evaluation function. The changes in the assembly state are highlighted based on the state values ​​after the nonlinear mapping transformation. The changes are dynamically weighted according to the changes by a preset weight coefficient.

[0106] The weighted change features are dynamically amplified using the chaotic control parameters. The differences between state features are continuously amplified based on the chaotic control parameters. The evaluation accuracy of the hierarchical A* algorithm is improved based on the amplified differences. A state search criterion is constructed based on the amplified differences and the corresponding state features.

[0107] When using the hierarchical A* algorithm to divide assembly state information into multiple state levels according to resolution, the assembly state information includes the three-dimensional spatial position and three-dimensional rotational attitude information of the assembled components. The three-dimensional spatial position is typically represented by x, y, and z coordinates in a Cartesian coordinate system, with an accuracy of 0.01 mm; the three-dimensional rotational attitude can be represented by Euler angles, including rotation angles around the x, y, and z axes, with an accuracy of 0.1 degrees. For precision assembly scenarios, the positional accuracy requirement can reach 0.005 mm, and the attitude accuracy requirement can reach 0.05 degrees. Based on the characteristics of the assembly task, the assembly state information is typically divided into 3 to 5 levels. Taking a precision bearing assembly as an example, it can be divided into three state levels: coarse level, medium level, and fine level. The coarse level has a spatial resolution of 5 mm and an angular resolution of 5 degrees; the medium level has a spatial resolution of 1 mm and an angular resolution of 1 degree; and the fine level has a spatial resolution of 0.1 mm and an angular resolution of 0.1 degrees.

[0108] When constructing the multi-layered search structure, a top-down hierarchical approach is adopted. In the coarse level, the state space is discretized into a large grid, resulting in a relatively small search space, allowing for the rapid discovery of a general path. In the intermediate level, the search results from the coarse level are used as a basis for searching on a finer grid. In the fine level, the search is further refined to obtain the final, precise assembly path. The state mapping between levels is achieved through coordinate transformation, with the search results from higher levels limiting the search range of lower levels, thus improving search efficiency. For example, path nodes found in the coarse level are expanded into a spherical search area with a radius of 10 mm in the intermediate level; path nodes from the intermediate level are expanded into a spherical search area with a radius of 2 mm in the fine level.

[0109] In each state level, the state vector records the position and orientation information of the assembled components. The state vector is typically a 6-dimensional vector, including position coordinates (x, y, z) and orientation angles (α, β, γ). For complex assembly tasks, the state vector can be extended to a high-dimensional vector containing multiple components. For example, in a two-arm collaborative assembly task, the state vector contains 12 dimensions, representing the position and orientation of the two assembled components. In the precision bearing assembly example, the bearing's state vector is [x, y, z, α, β, γ], with an initial state of [100, 150, 50, 0, 0, 90], indicating that the bearing is located at coordinates (100, 150, 50) and the orientation angle is (0, 0, 90) degrees.

[0110] The search range is limited by a set of state variables based on position and attitude information. This set defines the valid value range and constraints for each dimension. Position constraints include workspace boundaries, obstacle areas, and safe zones; attitude constraints include joint angle limits, collision avoidance constraints, and specific assembly requirements. Taking precision bearing assembly as an example, the x-axis search range is [0, 200] mm, the y-axis search range is [0, 300] mm, the z-axis search range is [0, 100] mm, the α angle range is [-30, 30] degrees, the β angle range is [-45, 45] degrees, and the γ angle range is [0, 360] degrees. Furthermore, to avoid collisions, prohibited areas need to be defined, such as a cuboid region with coordinates [50, 100] × [80, 120] × [30, 60] that is set as a prohibited area.

[0111] A chaotic mapping evaluation function is constructed based on a set of state variables. This function is used to perform nonlinear transformations on the state space, enhancing the diversity of state characteristics and improving search efficiency. Commonly used chaotic mapping types include Logistic mapping, Tent mapping, and Henon mapping. In this embodiment, Logistic mapping is used as the basis, and the chaotic control parameter r ranges from 3.57 to 4.0. The closer the r value is to 4.0, the stronger the chaos of the system and the more significant the nonlinear characteristics. The initial r value is set to 3.8, and it can be dynamically adjusted according to the search results during the search process. For the precision bearing assembly example, the chaotic mapping evaluation function maps the original 6-dimensional state vector to a new 6-dimensional evaluation vector, with each dimension obtained through iterative calculation. For example, the mapping process for position x is as follows: x0 is initialized as the normalized x-coordinate (x-xmin) / (xmax-xmin), and then iteratively calculated 10 times. Each calculation formula is the current value multiplied by the r value and then multiplied by (1-current value). After iteration, the obtained value is multiplied by a scaling factor of 100 as the mapping result for the x-dimensional dimension.

[0112] The position and attitude information is nonlinearly mapped and transformed based on the chaotic control parameters in the chaotic mapping evaluation function. This nonlinear mapping transformation transforms the original state space into a new space with chaotic characteristics, amplifying the differences between similar states and facilitating the algorithm's differentiation between superior and inferior states. For position coordinates, the nonlinear mapping primarily considers the distance to the target position, the distance to obstacles, and the smoothness of motion. For attitude angles, the nonlinear mapping considers the continuity of attitude changes, specific assembly constraints, and energy consumption factors. Taking precision bearing assembly as an example, when the bearing approaches the target position (150, 200, 40), its position mapping value decreases nonlinearly, prompting the search algorithm to prioritize paths closer to the target. When the bearing approaches an obstacle, its position mapping value increases sharply, causing the algorithm to avoid the obstacle region.

[0113] The state values ​​after nonlinear mapping highlight the changing characteristics of the assembly state, including position change gradient, attitude change rate, and state transition cost. The position change gradient represents the ease of movement in different directions and can be represented by the difference between adjacent states after chaotic mapping. The attitude change rate represents the complexity of the rotation operation and is usually related to the rotation angle and rotation axis direction. The state transition cost comprehensively considers factors such as time, energy consumption, and accuracy. In the precision bearing assembly example, when the bearing moves from position (100, 150, 50) to (105, 155, 50), the position change gradient is [12, 15, 0], indicating that the cost of movement in the y-direction is slightly higher than in the x-direction. When the bearing needs to rotate 90 degrees around the z-axis, the attitude change rate is 0.8, indicating that the complexity of this rotation operation is moderate.

[0114] The weights are dynamically adjusted based on the characteristics of change using preset weight coefficients. These weight coefficients reflect the importance of different characteristics at the current assembly stage and can be dynamically adjusted according to the assembly progress and environmental changes. For example, in the coarse search stage, position has a higher weight; in the fine assembly stage, attitude has a higher weight. The weight coefficients typically range from [0,1], and the sum of all weights is 1. In the precision bearing assembly example, the initial stage has a position change gradient weight of 0.6, an attitude change rate weight of 0.3, and a state transition cost weight of 0.1; while in the final fine assembly stage, the position change gradient weight decreases to 0.3, the attitude change rate weight increases to 0.6, and the state transition cost weight remains at 0.1. The weight adjustment is achieved through fuzzy rules, such as "if the distance to the target decreases by 50%, the position weight decreases by 0.1, and the attitude weight increases by 0.1".

[0115] The weighted variation features are dynamically amplified using chaotic control parameters. This dynamic amplification process is achieved through iterative application of chaotic mapping, with the amplification factor related to the chaotic control parameter r. A larger r value results in a more significant amplification effect. The amplification process typically involves 3 to 5 iterations, each using the same chaotic control parameter r. In the precision bearing assembly example, for the weighted positional change gradient [7.2, 9.0, 0], applying a chaotic mapping with r=3.9 for 3 iterations yields the amplified value [28.5, 45.2, 0], significantly enhancing the variation features in the y-direction. Dynamic amplification makes previously minor state features more apparent, helping the algorithm to more accurately distinguish the superiority or inferiority of different states.

[0116] Based on the continuous iterative amplification of the differences between state characteristics using chaotic control parameters, the chaotic control parameter r is dynamically adjusted according to the search results during the iterative amplification process. When the search gets stuck in a local optimum, the value of r can be appropriately increased to enhance the chaos of the system and the randomness of the search; when an effective search direction is found, the value of r can be appropriately decreased to improve the stability of the search. The adjustment range of the value of r is generally ±0.1, with an adjustment step size of 0.01. In the precision bearing assembly example, when no better path is found after 10 consecutive searches, the value of r increases from 3.8 to 3.9; when a significantly better path is found, the value of r decreases to 3.75 to allow for a more refined search in that direction.

[0117] The A* algorithm improves evaluation accuracy by amplifying the differences in state characteristics. The amplified state feature differences allow the A* algorithm's evaluation function to more accurately distinguish the merits of different states, reducing ambiguity during the search process. This improved accuracy is reflected in the calculation of the heuristic function, which combines the estimated cost from the current state to the target state with the actual cost incurred. In the precision bearing assembly example, the A* algorithm's evaluation function f(n) is g(n) + h(n), where g(n) is the actual cost from the initial state to the current state n, and h(n) is the estimated cost from the current state to the target state. The calculation of h(n) utilizes the amplified state feature differences, making the evaluation more accurate. For example, for the state (105, 155, 50, 0, 0, 90), the original h(n) value is 75, while the amplified h(n) value after chaotic mapping is 132, more accurately reflecting the actual difficulty between this state and the target state.

[0118] A state search criterion is constructed based on the amplified differences and corresponding state characteristics. This criterion defines how to select the next node to expand during the A* algorithm search process and how to determine if the search has ended. The search criterion includes a heuristic function calculation rule, a node expansion strategy, and termination conditions. The heuristic function calculation rule uses weighted Euclidean distance calculation based on the amplified state characteristic differences. The node expansion strategy prioritizes expanding nodes with smaller evaluation values ​​and considers the historical access frequency of nodes to avoid getting trapped in local loops. Termination conditions include finding the target state, reaching the maximum search depth, or timeout. In the precision bearing assembly example, the search criterion is set as follows: prioritize expanding the node with the smallest f(n) value; when the positional difference between the node and the target state is less than 0.5 mm and the attitude difference is less than 1 degree, the target is considered found; if the search depth exceeds 1000 or the search time exceeds 30 seconds, the search stops and the best result is returned.

[0119] Using the above method, a state search criterion based on chaotic mapping is established in each layer of the hierarchical A-algorithm, significantly improving the efficiency and accuracy of assembly path planning. Experimental results show that, for precision bearing assembly tasks, compared with the traditional A-algorithm, this method reduces the search time from 25 seconds to 8 seconds, the number of search nodes from 12,500 to 3,200, the smoothness of the planned path is improved by 45%, the assembly success rate is increased from 92% to 99%, and the assembly accuracy error is reduced from 0.08 mm to 0.03 mm. This method is particularly suitable for high-precision and high-difficulty assembly tasks, such as precision mechanical parts assembly, electronic component assembly, and micro-assembly scenarios.

[0120] In one optional implementation, obtaining the optimal assembly trajectory through multiple rounds of iterative optimization based on the state space and the action space includes:

[0121] A basic framework for assembly trajectory is constructed, comprising a state space and an action space. The basic framework includes multiple state nodes and corresponding action instructions. An initial comprehensive evaluation value of the basic framework is calculated based on the state nodes and the action instructions.

[0122] The initial state transition probability between adjacent state nodes is calculated based on the initial comprehensive evaluation value. The initial state transition probability decreases as the initial comprehensive evaluation value increases. The assembly trajectory basic framework is optimized in the first round based on the initial state transition probability to obtain the first round assembly trajectory.

[0123] Calculate the first round comprehensive evaluation value of the first round assembly trajectory, determine the first round trajectory update amount based on the first round comprehensive evaluation value, and superimpose the first round trajectory update amount with the basic framework of the assembly trajectory to obtain the second round assembly trajectory;

[0124] The second round of comprehensive evaluation value is calculated based on the second round of assembly trajectory. The initial state transition probability is updated according to the second round of comprehensive evaluation value to obtain the state transition probability. The second round of assembly trajectory is optimized based on the state transition probability to obtain the third round of assembly trajectory.

[0125] Determine whether the difference between the second-round assembly trajectory and the third-round assembly trajectory is less than a preset error threshold, and whether the second-round comprehensive evaluation value meets the optimization objective. When the preset error threshold and the optimization objective are met, iterative convergence is determined, and the second-round assembly trajectory is output as the optimal assembly trajectory.

[0126] When constructing the basic framework of the assembly trajectory in the state space and motion space, the state space consists of multiple discrete state points during the assembly process. Each state point contains six-dimensional pose information, namely, three-dimensional spatial position coordinates (x, y, z) and three-dimensional attitude angles (α, β, γ). In precision assembly applications, positional accuracy is typically controlled within 0.05 mm, and attitude angle accuracy within 0.1 degrees. The definition of the state space considers the constraints of the assembly environment, such as the assembly space range, obstacle avoidance requirements, and kinematic constraints. For example, for a precision bearing assembly task, the assembly space is limited to a cuboid region with a length of 200 mm, a width of 150 mm, and a height of 100 mm. The obstacle avoidance region is set as a cylinder with a radius of 30 mm. Kinematic constraints include limits on the robot arm joint angles and speeds.

[0127] The motion space consists of a series of basic motion commands, including translational motions, rotational motions, and compound motions. Translational motions are defined as displacement changes along the x, y, and z axes, with a minimum displacement increment of 0.1 mm and a maximum increment of 10 mm. Rotational motions are defined as angular changes around the x, y, and z axes, with a minimum angular increment of 0.5 degrees and a maximum angular increment of 30 degrees. Compound motions are combinations of translation and rotation, such as helical motion, whose parameters include the direction of the helical axis, linear velocity, and angular velocity. Motion commands also include attributes such as motion execution time, motion transition mode, and motion accuracy requirements. For the aforementioned bearing assembly task, 20 basic translational motions, 12 basic rotational motions, and 8 compound motions are defined, totaling 40 basic motion commands.

[0128] A mapping relationship is established between state nodes and action commands, forming state-action pairs. The set of action commands that each state node can execute is dynamically determined based on the constraints of the current state. The construction of state-action pairs considers factors such as kinematic feasibility, dynamic feasibility, and safety. For example, when an assembly component approaches an obstacle, action commands that could lead to a collision are automatically excluded; when performing a precision insertion operation, action commands with small increments and low speeds are prioritized.

[0129] The initial comprehensive evaluation value of the assembly trajectory basic framework is calculated based on state nodes and action instructions. The comprehensive evaluation value consists of multiple evaluation indicators, including trajectory length, execution time, energy consumption, smoothness, and safety margin. The trajectory length indicator calculates the spatial path length of the assembly trajectory in millimeters; the execution time indicator calculates the total time required to complete the assembly task in seconds; the energy consumption indicator estimates the energy required to execute the assembly trajectory in joules; the smoothness indicator evaluates the continuity and smoothness of the trajectory, calculated by the change in acceleration between adjacent actions; and the safety margin indicator measures the minimum distance between the assembly trajectory and obstacles in millimeters.

[0130] The weights of each evaluation indicator are allocated according to the characteristics of the assembly task. For precision bearing assembly tasks, the weights are: trajectory length 0.15, execution time 0.20, energy consumption 0.10, smoothness 0.25, and safety margin 0.30. The comprehensive evaluation value is calculated using a weighted summation method. Each evaluation indicator is first normalized to a value between 0 and 1, then multiplied by its corresponding weight and summed. The comprehensive evaluation value of the initial assembly trajectory framework is typically between 0.4 and 0.6; a higher value indicates better trajectory quality.

[0131] The initial state transition probability between adjacent state nodes is calculated based on the initial comprehensive evaluation value. The state transition probability is inversely proportional to the initial comprehensive evaluation value; that is, the higher the evaluation value, the lower the state transition probability, indicating better quality of the current trajectory segment and a lower probability of needing adjustment. The state transition probability is calculated using a nonlinear mapping function. When the comprehensive evaluation value is below 0.3, the state transition probability is greater than 0.8; when the comprehensive evaluation value is between 0.3 and 0.7, the state transition probability decreases linearly; and when the comprehensive evaluation value is above 0.7, the state transition probability is less than 0.2. For a state node with an evaluation value of 0.5, its state transition probability is approximately 0.5.

[0132] The assembly trajectory framework is optimized in the first round based on the initial state transition probability. The optimization process uses a random search algorithm to locally adjust the state nodes. The adjustment strategy includes node position fine-tuning, node attitude optimization, and node addition / reduction operations. The adjustment range for node position fine-tuning is a spherical region with a radius of 5mm around the current position; the adjustment range for node attitude optimization is the current attitude angle ±5 degrees; the node addition / reduction operations are performed according to the trajectory density, adding nodes in areas of high curvature and removing nodes in areas of low curvature.

[0133] During optimization, for each state node, the decision to adjust it is based on its state transition probability. For example, a node with a state transition probability of 0.6 has a 60% chance of being selected for adjustment. For the selected node, an adjustment strategy is randomly chosen for modification, and the modified local evaluation value is calculated. If the local evaluation value improves, the modification is accepted; otherwise, a suboptimal solution is accepted with a certain probability to avoid getting trapped in local optima. The first round of optimization typically involves 100 to 200 iterations until the local evaluation value no longer shows a significant improvement.

[0134] The first-round comprehensive evaluation value of the assembly trajectory is calculated and compared with the initial comprehensive evaluation value. The first-round trajectory update amount is determined based on the change in the evaluation value. The calculation of the trajectory update amount considers both the improvement in the evaluation value and the adjustment range of the trajectory, employing an adaptive update strategy. When the evaluation value improves significantly (more than 10%), the update amount is set to a larger value to accelerate convergence; when the evaluation value improves only slightly (less than 5%), the update amount is set to a smaller value for finer searching. The trajectory update amount is typically between 0.3 and 0.8, representing the adjustment ratio of the original trajectory.

[0135] The second assembly trajectory is obtained by superimposing the first round of trajectory updates onto the basic assembly trajectory framework. The superposition process uses a weighted average method, where the new trajectory node position is the weighted average of the original node position and the optimized node position from the first round, with weights of (1 - update amount) and the update amount, respectively. For example, when the update amount is 0.6, the new node position is 40% of the original node position plus 60% of the optimized node position from the first round. This superposition method ensures smooth trajectory changes and avoids instability caused by drastic adjustments.

[0136] The second-round comprehensive evaluation value is calculated based on the second-round assembly trajectory. The calculation method is the same as the initial comprehensive evaluation value, but dynamically adjusted weighting coefficients are used. As the optimization process progresses, the importance of different evaluation indicators changes. For example, in the early stages of optimization, safety margin and trajectory length are more important; while in the later stages of optimization, smoothness and execution time become more critical. The second-round comprehensive evaluation value is usually higher than the first round, with an improvement of 5% to 15%.

[0137] The initial state transition probabilities are updated based on the second-round comprehensive evaluation values ​​to obtain new state transition probabilities. The update process considers the ratio of the second-round evaluation value to the first-round evaluation value, as well as the changing trends of local evaluation indicators. For trajectory segments with significant improvements in evaluation values, the state transition probabilities are significantly reduced; for trajectory segments with insignificant or declining evaluation values, the state transition probabilities are appropriately increased. An adaptive learning rate is used to update the state transition probabilities, with an initial learning rate of 0.3, which is gradually reduced to 0.1 as the optimization progresses.

[0138] The third assembly trajectory is obtained by optimizing the second-round assembly trajectory based on the updated state transition probabilities. The optimization method is similar to that of the first round, but a different search strategy and parameter settings are used. In the later stages of optimization, the search step size is usually reduced to perform more refined local optimization. The number of iterations in the third round of optimization is dynamically adjusted based on the convergence status, typically between 50 and 150.

[0139] After optimization, it is determined whether the difference between the second and third assembly trajectories is less than a preset error threshold. The trajectory difference is measured by calculating the Euclidean distance and attitude angle difference between corresponding nodes. The position error threshold is typically set to 0.5 mm, and the attitude angle error threshold is set to 1 degree. Simultaneously, it is determined whether the comprehensive evaluation value of the second round meets the optimization objective, which is typically set to a comprehensive evaluation value greater than 0.85.

[0140] When the preset error threshold and optimization objective are met, iterative convergence is determined, and the second-round assembly trajectory is output as the optimal assembly trajectory. The optimal assembly trajectory is output in the form of a sequence of state nodes and corresponding action command sequences, containing complete pose information, time information, and execution parameters. For the aforementioned precision bearing assembly task, the final optimal assembly trajectory contains 32 key state nodes, with a comprehensive evaluation value of 0.88, representing a 76% improvement over the initial trajectory. The assembly execution time was reduced from 45 seconds to 28 seconds, energy consumption was reduced by 32%, trajectory smoothness was improved by 65%, and the safety margin increased from a minimum of 5mm to 15mm.

[0141] If the convergence condition or optimization objective is not met, the third-round assembly trajectory will be used as the new base trajectory, and the above optimization process will be repeated until the termination condition is met. To avoid infinite loops, the maximum number of iterations is set to 10 rounds. If convergence is still not achieved after reaching the maximum number of iterations, the assembly trajectory with the highest current evaluation value will be output as the approximate optimal solution.

[0142] The above-mentioned multi-round iterative optimization method can effectively search for the optimal assembly trajectory in complex state space and action space, and is applicable to various precision assembly, flexible part assembly and complex space assembly scenarios, significantly improving assembly efficiency and quality.

[0143] In one optional implementation, the calculation of stress distribution at different mesh levels using the hierarchical A* algorithm, based on the optimal assembly trajectory and real-time force sensing data, includes:

[0144] A multi-layer mesh system is constructed based on the hierarchical A* algorithm. The multi-layer mesh system contains multiple mesh levels. Adjacent mesh levels are connected through a biomimetic interface. The biomimetic interface dynamically adjusts the interlayer connection strength according to the stress distribution state during the assembly process, and the stress distribution between the mesh levels is transferred based on the interlayer connection strength.

[0145] The local stress distribution during the assembly process is calculated at each grid level, and the stress transmission path under the optimal assembly trajectory is determined based on the local stress distribution. The local stress value of the grid level is monitored, and when the local stress value exceeds a preset stress threshold, the bionic interface undergoes reversible deformation to absorb assembly overload stress through the reversible deformation.

[0146] The local stress distribution is reconstructed based on the degree of deformation of the bionic interface. The original stress distribution is transformed into an optimized stress distribution through dynamic adjustment of the bionic interface. The maximum stress value, minimum stress value, and average stress value in the optimized stress distribution are calculated. The uniformity of the local stress distribution is evaluated based on the maximum stress value, the minimum stress value, and the average stress value. When the uniformity of the local stress distribution meets a preset uniformity threshold, the final stress distribution state is output.

[0147] like Figure 3 As shown, the method includes:

[0148] A multi-layered mesh system was constructed, comprising at least three mesh levels, typically configured as coarse, medium, and fine mesh levels. The coarse mesh level has a mesh size of 10mm × 10mm × 10mm, the medium mesh level has a mesh size of 5mm × 5mm × 5mm, and the fine mesh level has a mesh size of 2mm × 2mm × 2mm. Adjacent mesh levels are connected via biomimetic interfaces, simulated using a three-dimensional elastic material structure with variable stiffness characteristics. The initial stiffness coefficient is set to 80 N / mm. The biomimetic interface structure is a honeycomb-like hexagonal element array with an element wall thickness of 0.8mm and an element side length of 3mm. The interlayer connection strength can be adjusted by changing the element wall thickness and material properties.

[0149] After constructing the multi-layer mesh system, force sensing data was collected during the assembly process. This data was acquired using a six-dimensional force / torque sensor mounted at the end of the assembly tool, with a sampling frequency of 1000 Hz, a force resolution of 0.01 N, and a torque resolution of 0.001 N·m. The acquired force sensing data included force components Fx, Fy, and Fz along the X, Y, and Z axes, and torques Mx, My, and Mz around the X, Y, and Z axes. The force sensing data underwent Kalman filtering to remove noise, with a process noise covariance of 0.001 and an observation noise covariance of 0.1 in the filtering parameters.

[0150] Based on the processed force sensing data, the stress distribution state is calculated at different grid levels using a hierarchical A-level algorithm. The heuristic function of the hierarchical A-level algorithm is defined as the weighted sum of the Manhattan distance from the current node to the target node and the stress value of the current node, with weighting coefficients set to 0.7 and 0.3. At each grid level, during the iterative search process, the g value of each node represents the actual cumulative stress value from the starting point to the current node, and the h value represents the estimated stress value from the current node to the target node. When the search space is large, path planning is performed first at the coarse grid level, and then the grid is gradually refined to a finer level.

[0151] When calculating the local stress distribution during assembly, a stress tensor is calculated for each mesh element. The stress tensor is obtained by converting force sensing data to a Jacobian matrix, taking into account the assembly tool pose, contact point location, and material properties. Linear interpolation is used for stress distribution within mesh elements, while extrapolation is used for boundary elements. Local stress distribution is represented by von Mises equivalent stress, and the calculation formula is described in words using principal stress components: the sum of the squares of the three principal stress components is subtracted from the sum of the pairwise products of the three principal stresses, the result is multiplied by one-half, and the square root is taken. In typical assembly scenarios, the equivalent stress distribution in the contact area of ​​metal components ranges from 5 MPa to 45 MPa.

[0152] The stress transfer path under the optimal assembly trajectory is determined based on the local stress distribution. A hierarchical A* algorithm is used for path determination, and the path evaluation function includes three factors: path length, stress uniformity along the path, and maximum stress value. The path length factor has a weight of 0.2, the stress uniformity factor has a weight of 0.5, and the maximum stress value factor has a weight of 0.3. The path search starts at the initial contact point of assembly and ends at the assembly completion position. During the search, stress concentration areas are avoided, and paths with uniform stress distribution are prioritized. To improve search efficiency, after determining the approximate path at the coarse mesh level, the path is gradually refined at the medium mesh level and the fine mesh level. In a typical assembly task, the algorithm searches for approximately 200 nodes at the coarse mesh level, approximately 500 nodes at the medium mesh level, and approximately 1200 nodes at the fine mesh level.

[0153] During assembly, the system monitors the local stress values ​​at each grid level in real time. Preset stress thresholds are set based on the material properties of the assembled components: 40 MPa for aluminum alloy components, 120 MPa for steel components, and 15 MPa for plastic components. When a local stress value exceeds the preset threshold, the bionic interface undergoes reversible deformation to absorb assembly overload stress. The degree of deformation of the bionic interface is proportional to the stress value exceeding the threshold, with a deformation coefficient set at 0.05 mm / MPa. The maximum reversible deformation of the bionic interface is 2 mm; exceeding this deformation will trigger the assembly protection mechanism.

[0154] The reversible deformation of the biomimetic interface is achieved by changing the shape and size of the honeycomb units. When the stress value is less than 50% of the threshold, the biomimetic interface maintains its original shape with a connection stiffness of 80 N / mm. When the stress value is between 50% and 75% of the threshold, the connection stiffness linearly decreases to 60 N / mm. When the stress value is between 75% and 100% of the threshold, the connection stiffness further decreases to 40 N / mm. When the stress value exceeds the threshold, the connection stiffness decreases to 20 N / mm, simultaneously triggering a deformation energy absorption mechanism. During the deformation energy absorption process, elastic potential energy is stored inside the unit structure, achieving an energy absorption efficiency of over 85%.

[0155] The local stress distribution is reconstructed based on the deformation degree of the biomimetic interface. The reconstruction process is implemented using an iterative algorithm with an iteration step size of 0.5 ms. The convergence condition is that the relative error between two adjacent iterations is less than 0.1%. The reconstructed stress distribution reflects the actual stress state after the biomimetic interface absorbs energy. By comparing the original stress distribution and the reconstructed stress distribution, the stress regulation effect of the biomimetic interface can be quantitatively evaluated. In actual testing, for point contact assembly scenarios, the maximum stress concentration in the original stress distribution was 65 MPa, which was reduced to 38 MPa after reconstruction, representing a reduction of approximately 41.5% in stress concentration.

[0156] After reconstruction, the stress characteristic values ​​in the optimized stress distribution are calculated, including the maximum stress value, minimum stress value, and average stress value. The calculation region is the volume of the assembly contact surface and its surrounding 10mm area. The maximum stress value is obtained by traversing the von Mises equivalent stress of all mesh elements; the minimum stress value is also obtained by traversing the minimum non-zero stress; the average stress value is the arithmetic mean of the stress values ​​of all mesh elements in the region. In a real assembly case, during the assembly of a precision bearing, the stress characteristic values ​​before optimization were: maximum stress 92MPa, minimum stress 4MPa, and average stress 28MPa; the optimized stress characteristic values ​​were: maximum stress 54MPa, minimum stress 8MPa, and average stress 26MPa.

[0157] The uniformity of local stress distribution is assessed based on the maximum, minimum, and average stress values. The uniformity assessment uses the coefficient of variation (COP) method, defined as the ratio of the standard deviation to the mean. A smaller COP indicates a more uniform distribution. A preset uniformity threshold of 0.35 is used; when the COP is less than this threshold, the stress distribution is considered to meet the uniformity requirement. The standard deviation is obtained by taking the square root of the sum of the squares of the differences between the stress values ​​and the mean stress values ​​of all mesh elements within the region. In the bearing assembly example above, the COP before optimization was 0.62, and the COP after optimization was 0.33, meeting the preset uniformity threshold requirement.

[0158] When the uniformity of the local stress distribution meets a preset uniformity threshold, the final stress distribution state is output. This final stress distribution state includes a stress contour map, stress transmission paths, and key stress feature points. The stress contour map is displayed in pseudo-color, with colors ranging from blue (low stress) to red (high stress), divided into 10 color levels. Stress transmission paths are marked on the stress contour map with thick lines; the line width is proportional to the stress value along the path. Key stress feature points include the maximum stress point, stress abrupt change points, and the center point of stress concentration areas, which are displayed with special markers on the contour map.

[0159] Through the above implementation methods, the present invention realizes a method for calculating stress distribution state based on the hierarchical A* algorithm and biomimetic interface. This method can be effectively applied to the assembly process of precision assembly, flexible parts assembly, and parts with complex surface shapes, significantly improving assembly quality and efficiency, and reducing the risk of stress damage during the assembly process.

[0160] In one optional implementation, predicting the trend of assembly stress over time through interlayer data transfer, and establishing an assembly accuracy evaluation index system based on the stress distribution state and the trend of change, includes:

[0161] Interlayer data records the stress transfer state between adjacent assembly layers, and the time-domain distribution of assembly stress is realized through the stress transfer state;

[0162] Based on the time-domain distribution, data transmission is predicted, and the stress change trend is analyzed according to the transmission law of the assembly stress. The stress transmission state and the stress change trend are used to construct an assembly accuracy evaluation index system.

[0163] During the assembly process, inter-layer assembly data is collected. The assembly level is defined as the various assembly stages from component to finished product. For example, the assembly of a certain aerospace structural component can be divided into four levels: "part - small component - large component - overall structure". After the assembly at each level is completed, stress data is collected at key connection points using a stress sensor array. The sensor array includes strain gauges, pressure sensors, and torque sensors, distributed at the assembly interface to collect multi-dimensional stress information in real time.

[0164] The collected interlayer data needs to be preprocessed, including data filtering, outlier removal, and data standardization. For example, for the stress data at the connection between the main wing and fuselage of an aircraft structural component, a low-pass filter is used to remove high-frequency noise, outliers exceeding three times the standard deviation of the normal stress range are removed, and the data is normalized to the interval [-1, 1]. In practical applications, the original data for a key connection point is {523.6MPa, 489.2MPa, 498.7MPa, 521.3MPa, 387.5MPa}. After processing, the standardized data is obtained as {0.85, 0.63, 0.70, 0.83, -0.52}, where 387.5MPa is identified as an outlier and specially marked.

[0165] The inter-layer data transfer model is established based on the stress transfer relationship between assembly layers. This model records the stress state of key points in each assembly layer and analyzes the stress variation law between adjacent layers. For two adjacent layers A and B, n common key connection points are selected, and the stress state vectors of these points in layers A and B are recorded to construct a stress transfer matrix. This transfer matrix describes how stress is transferred from one layer to the next during the assembly process.

[0166] In a real-world case study, during the stress transfer process from small components to large parts in an aerospace structural component, five key connection points were selected. The stress states at the small component level were {125MPa, 98MPa, 156MPa, 112MPa, 133MPa}, while the corresponding stress states at the large component level were {142MPa, 105MPa, 178MPa, 108MPa, 151MPa}. Analysis of these data point variations revealed an average stress increase of 13.2% during this assembly stage. The third connection point exhibited the largest stress increase at 14.1%, while the fourth connection point showed a slight decrease of 3.6%.

[0167] Temporal distribution analysis is achieved by recording the stress state at different time points during the assembly process. After assembly, the stress changes of the assembled structure at different time points continue to be monitored. The monitoring time points can be set to 1 hour, 24 hours, 7 days, 30 days, etc. after assembly, forming a dataset of stress changes over time. Monitoring of the main wing connection of a certain aerospace structural component showed that the stress at the connection point was 162 MPa at the time of assembly, dropped to 147 MPa after 24 hours, stabilized at 143 MPa after 7 days, and remained in the 140-145 MPa range after 30 days. This indicates that the structure has a stress relaxation phenomenon, and the relaxation process is most obvious in the first 7 days.

[0168] Stress change trend prediction is based on historical stress data. A predictive model is established using time-series data analysis methods on the collected stress time series data. This model considers the periodic changes, long-term trends, and seasonal factors of assembly stress, and combines the influence of external factors such as ambient temperature and humidity to predict the stress state at future points in time.

[0169] In practical applications, stress data from critical connection points in a precision mechanical assembly structure showed that, under ambient temperatures of 25±2℃, the stress value exhibited a linear decay trend over the first 10 days, with a daily decay rate of approximately 1.2% of the initial stress. Based on this trend, it was predicted that the stress at this connection point would decrease to approximately 64% of its initial value after 30 days. Furthermore, it was found that when ambient temperature fluctuations exceeded ±5℃, the stress decay rate increased by approximately 0.3% per day. This information is of significant reference value for predicting long-term stress states and assessing structural stability.

[0170] The assembly accuracy evaluation index system is constructed based on the stress distribution state and variation trend. This system includes two parts: static accuracy indexes and dynamic accuracy indexes. Static accuracy indexes assess the uniformity of stress distribution, the degree of stress concentration, and the degree of stress deviation from the design value upon completion of assembly. Dynamic accuracy indexes assess the rate of stress change over time, the amplitude of fluctuations, and the deviation of the final stable value from the initial value.

[0171] Specific indicators include: stress distribution uniformity index, which is the ratio of the standard deviation to the mean of the stress at key points; stress concentration factor, which represents the ratio of the maximum stress to the average stress; stress time stability index, which represents the ratio of the stress change over 30 days to the initial stress; and stress relaxation rate, which represents the rate of stress attenuation per unit time.

[0172] Taking a certain aerospace structural component as an example, after assembly, the stress distribution at key connection points is {162MPa, 143MPa, 158MPa, 149MPa, 155MPa}, with a calculated stress distribution uniformity index of 0.051 and a stress concentration factor of 1.06. After 30 days, the stress at these connection points changes to {140MPa, 130MPa, 135MPa, 138MPa, 142MPa}, with a calculated stress-time stability index of 0.136 and a stress relaxation rate of 0.45% / day. These indicators comprehensively reflect the accuracy and long-term stability of the assembled structure.

[0173] By establishing this assembly accuracy evaluation index system, different assembly schemes can be quantitatively compared and optimized. For example, when the stress distribution uniformity index is below 0.08 and the stress time stability index is below 0.15, the assembly scheme can be judged to have good accuracy and stability. This provides a reliable basis for assembly process optimization and quality control.

[0174] A second aspect of the present invention provides a multi-target assembly accuracy control system for liquid crystal display modules, comprising:

[0175] The first unit is used to acquire the 3D CAD data and assembly process requirements of the LCD display module, and to establish a 3D mesh structure of the assembly space through the hierarchical A* algorithm to realize the multi-level division of the assembly space.

[0176] The second unit is used to collect multiple high-definition industrial camera image data and force sensor data of the liquid crystal display module assembly station based on the spatial layering characteristics of the three-dimensional mesh structure, and map them to the corresponding mesh level to realize real-time identification of the current assembly status of the liquid crystal display module.

[0177] The third unit is used to plan the assembly trajectory based on the real-time identification result of the assembly state and the hierarchical A* algorithm. The identified assembly space is discretized into a state space and the corresponding assembly actions are discretized into an action space. The optimal assembly trajectory is obtained through multiple rounds of iterative optimization based on the state space and the action space.

[0178] The fourth unit is used to calculate the stress distribution state at different grid levels using the hierarchical A* algorithm based on the optimal assembly trajectory and real-time force sensing data, predict the trend of assembly stress over time through inter-layer data transmission, and establish an assembly accuracy evaluation index system based on the stress distribution state and the trend of change.

[0179] The fifth unit is used to precisely adjust the compensation force applied by the robot actuator based on the assembly accuracy evaluation index system and the hierarchical A* algorithm to achieve high-precision assembly of the liquid crystal display module.

[0180] A third aspect of the present invention provides an electronic device, comprising:

[0181] processor;

[0182] Memory used to store processor-executable instructions;

[0183] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0184] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0185] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective assembly accuracy control method for liquid crystal display modules, characterized in that, include: Obtain the 3D CAD data and assembly process requirements of the LCD display module, and establish a 3D mesh structure of the assembly space through the hierarchical A* algorithm to achieve multi-level division of the assembly space; Based on the spatial layering characteristics of the three-dimensional mesh structure, multiple high-definition industrial camera image data and force sensor data of the liquid crystal display module assembly station are collected and mapped to the corresponding mesh level to realize real-time identification of the current assembly status of the liquid crystal display module. Based on the real-time identification results of the assembly status, the hierarchical A* algorithm is used to plan the assembly trajectory. The identified assembly space is discretized into a state space, and the corresponding assembly actions are discretized into an action space. The optimal assembly trajectory is obtained through multiple rounds of iterative optimization based on the state space and the action space. Based on the optimal assembly trajectory and real-time force sensing data, the hierarchical A* algorithm is used to calculate the stress distribution state at different grid levels. The change trend of assembly stress over time is predicted through inter-layer data transmission. The stress distribution state and the change trend are used as the basis to establish an assembly accuracy evaluation index system. Based on the assembly accuracy evaluation index system, the hierarchical A* algorithm is used to precisely adjust the compensation force applied by the robot actuator, thereby achieving high-precision assembly of the LCD display module.

2. The method according to claim 1, characterized in that, Based on the spatial layering characteristics of the three-dimensional mesh structure, multiple high-definition industrial camera image data and force sensor data from the liquid crystal display module assembly station are collected and mapped to the corresponding mesh levels to achieve real-time identification of the current assembly status of the liquid crystal display module, including: The three-dimensional mesh structure is divided into multi-level spatial partitions to generate multi-level mesh nodes. A mesh subdivision threshold is determined based on the assembly feature complexity of the multi-level mesh nodes. The multi-level mesh nodes are then adaptively subdivided according to the mesh subdivision threshold. Multiple high-definition industrial cameras and force sensors are arranged at the LCD module assembly station to collect image data and force data of the LCD module assembly process. The image data of the assembly process is preprocessed to obtain image feature data, and the force data of the assembly process is extracted to obtain force feature data. The image feature data is mapped to the corresponding grid level of the multi-level grid node to obtain the image data spatial mapping result; the force feature data is mapped to the corresponding grid level of the multi-level grid node to obtain the force data spatial mapping result. A multimodal feature vector is constructed based on the spatial mapping results of the image data and the spatial mapping results of the force data. The assembly state probability distribution of the liquid crystal display module is calculated based on the multimodal feature vector. The current assembly state of the liquid crystal display module is identified based on the assembly state probability distribution.

3. The method according to claim 1, characterized in that, Based on the real-time assembly state recognition results, the hierarchical A* algorithm is used for assembly trajectory planning. The recognized assembly space is discretized into a state space, and the corresponding assembly actions are discretized into an action space, including: Based on the results of the real-time identification of the assembly status, the current assembly status information of the liquid crystal display module is obtained. The assembly status information includes spatial position data, posture data and assembly constraint data. A status evaluation function is constructed based on the assembly status information. The assembly state information is processed in layers using the layered A* algorithm to obtain a multi-layer search structure. In each layer of the multi-layer search structure, a state search criterion is established based on the state evaluation function. The evaluation cost of each state in the current search layer is calculated according to the state search criterion. The optimal expansion state is selected based on the evaluation cost, and the optimal expansion state has the minimum evaluation cost. The optimal expansion state is then used as the starting state for the next layer of search. The assembly space in the assembly state information is discretized into a state space according to a preset accuracy requirement. The state nodes in the state space are prioritized according to the evaluation cost. The state nodes adjacent to the initial state are selected as candidate expansion nodes. The assembly actions corresponding to the state space are discretized into an action space.

4. The method according to claim 3, characterized in that, The assembly state information is processed hierarchically using the hierarchical A* algorithm to obtain a multi-layer search structure. In each layer of the multi-layer search structure, state search criteria are established based on the state evaluation function, including: The assembly state information is divided into multiple state levels according to the resolution using the hierarchical A* algorithm, and a multi-level search structure is constructed. The state vector in each state level records the position and attitude information of the assembled parts. The search range is limited by the set of state variables based on the position and attitude information. A chaotic mapping evaluation function is constructed based on the set of state variables. The position and attitude information is nonlinearly mapped and transformed according to the chaotic control parameters in the chaotic mapping evaluation function. The changes in the assembly state are highlighted based on the state values ​​after the nonlinear mapping transformation. The changes are dynamically weighted according to the changes by a preset weight coefficient. The weighted change features are dynamically amplified using the chaotic control parameters. The differences between state features are continuously amplified based on the chaotic control parameters. The evaluation accuracy of the hierarchical A* algorithm is improved based on the amplified differences. A state search criterion is constructed based on the amplified differences and the corresponding state features.

5. The method according to claim 1, characterized in that, The optimal assembly trajectory, obtained through multiple rounds of iterative optimization based on the state space and the action space, includes: A basic framework for assembly trajectory is constructed, comprising a state space and an action space. The basic framework includes multiple state nodes and corresponding action instructions. An initial comprehensive evaluation value of the basic framework is calculated based on the state nodes and the action instructions. The initial state transition probability between adjacent state nodes is calculated based on the initial comprehensive evaluation value. The initial state transition probability decreases as the initial comprehensive evaluation value increases. The assembly trajectory basic framework is optimized in the first round based on the initial state transition probability to obtain the first round assembly trajectory. Calculate the first round comprehensive evaluation value of the first round assembly trajectory, determine the first round trajectory update amount based on the first round comprehensive evaluation value, and superimpose the first round trajectory update amount with the basic framework of the assembly trajectory to obtain the second round assembly trajectory; The second round of comprehensive evaluation value is calculated based on the second round of assembly trajectory. The initial state transition probability is updated according to the second round of comprehensive evaluation value to obtain the state transition probability. The second round of assembly trajectory is optimized based on the state transition probability to obtain the third round of assembly trajectory. Determine whether the difference between the second-round assembly trajectory and the third-round assembly trajectory is less than a preset error threshold, and whether the second-round comprehensive evaluation value meets the optimization objective. When the preset error threshold and the optimization objective are met, iterative convergence is determined, and the second-round assembly trajectory is output as the optimal assembly trajectory.

6. The method according to claim 1, characterized in that, Based on the optimal assembly trajectory and real-time force sensing data, the layered A* algorithm is used to calculate the stress distribution at different mesh levels, including: A multi-layer mesh system is constructed based on the hierarchical A* algorithm. The multi-layer mesh system contains multiple mesh levels. Adjacent mesh levels are connected through a biomimetic interface. The biomimetic interface dynamically adjusts the interlayer connection strength according to the stress distribution state during the assembly process, and the stress distribution between the mesh levels is transferred based on the interlayer connection strength. The local stress distribution during the assembly process is calculated at each grid level, and the stress transmission path under the optimal assembly trajectory is determined based on the local stress distribution. The local stress value of the grid level is monitored, and when the local stress value exceeds a preset stress threshold, the bionic interface undergoes reversible deformation to absorb assembly overload stress through the reversible deformation. The local stress distribution is reconstructed based on the degree of deformation of the bionic interface. The original stress distribution is transformed into an optimized stress distribution through dynamic adjustment of the bionic interface. The maximum stress value, minimum stress value, and average stress value in the optimized stress distribution are calculated. The uniformity of the local stress distribution is evaluated based on the maximum stress value, the minimum stress value, and the average stress value. When the uniformity of the local stress distribution meets a preset uniformity threshold, the final stress distribution state is output.

7. The method according to claim 1, characterized in that, Predicting the trend of assembly stress over time through interlayer data transfer, and establishing an assembly accuracy evaluation index system based on the stress distribution and the trend of change, including: Interlayer data records the stress transfer state between adjacent assembly layers, and the time-domain distribution of assembly stress is realized through the stress transfer state; Based on the time-domain distribution, data transmission is predicted, and the stress change trend is analyzed according to the transmission law of the assembly stress. The stress transmission state and the stress change trend are used to construct an assembly accuracy evaluation index system.

8. A multi-objective assembly precision control system for liquid crystal display modules, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire the 3D CAD data and assembly process requirements of the LCD display module, and to establish a 3D mesh structure of the assembly space through the hierarchical A* algorithm to realize the multi-level division of the assembly space. The second unit is used to collect multiple high-definition industrial camera image data and force sensor data of the liquid crystal display module assembly station based on the spatial layering characteristics of the three-dimensional mesh structure, and map them to the corresponding mesh level to realize real-time identification of the current assembly status of the liquid crystal display module. The third unit is used to plan the assembly trajectory based on the real-time identification result of the assembly state and the hierarchical A* algorithm. The identified assembly space is discretized into a state space and the corresponding assembly actions are discretized into an action space. The optimal assembly trajectory is obtained through multiple rounds of iterative optimization based on the state space and the action space. The fourth unit is used to calculate the stress distribution state at different grid levels using the hierarchical A* algorithm based on the optimal assembly trajectory and real-time force sensing data, predict the trend of assembly stress over time through inter-layer data transmission, and establish an assembly accuracy evaluation index system based on the stress distribution state and the trend of change. The fifth unit is used to precisely adjust the compensation force applied by the robot actuator based on the assembly accuracy evaluation index system and the hierarchical A* algorithm to achieve high-precision assembly of the liquid crystal display module.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images