Intelligent assembly control method for computer case, transparent panel and metal frame

Abstract

The application provides a computer case and a smart assembly control method for a transparent panel and a metal frame, comprising: converting the relative position offset through a pre-established fastening force mapping model to determine the initial fastening force parameter corresponding to each fastening point; extracting the stress field distribution characteristics around each fastening point from the internal stress tensor distribution data to obtain a stress distribution characteristic vector set corresponding to each fastening point; determining the actual stress scalar value of the first assembly component at each fastening point through a stress calculation model based on elasticity mechanics according to the target fastening force parameter; performing grading processing on the actual stress scalar value according to a pre-set stress grading interval standard to map to obtain buffer structure parameter combination data corresponding to each fastening point; and generating an assembly execution instruction set for driving the assembly equipment to perform fastening and buffer installation actions according to the buffer structure parameter combination data and the target fastening force parameter.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an intelligent assembly control method for computer chassis, transparent panels, and metal frames. Background Technology

[0002] As computer cases evolve towards high-end and panoramic designs, transparent panels have become a core design element for showcasing internal hardware. The assembly quality of the transparent panel and the metal frame directly determines the case's appearance, structural strength, and user experience.

[0003] However, current assembly technologies generally employ a model of independent manufacturing followed by manual assembly. The transparent panel undergoes cutting, edge grinding, and chemical strengthening processes on one production line, while the metal frame is formed on another line through stamping, bending, and welding. The processing techniques for these two types of parts are completely different, and their dimensional tolerance systems are incompatible. When the transparent panel is installed onto the metal frame via multiple fastening points, the processing errors of the panel, the frame, and the installation and positioning errors of each fastening point are superimposed, forming a complex cumulative effect of dimensional chains. This cumulative effect further exacerbates the uneven stress distribution within the transparent panel during assembly. Assembly workers or robots rely on preset torque values ​​for fastening, but the actual fastening force generated by the same torque value under different thread conditions may vary significantly. This uneven fastening force at each fastening point creates localized stress concentrations within the panel.

[0004] More importantly, current technology completely lacks methods for monitoring the real-time stress distribution of transparent panels during assembly. Assemblers cannot know the actual impact of each fastening point on the internal stress field of the panel. This "blind" assembly results in some chassis having significant residual stress inside the transparent panels at the time of manufacture. For example, when a tempered glass panel is fixed with six screws, if the tightening force of one screw is too great, a stress concentration zone will form around that point. The local stress may exceed the strength limit of the glass, leading to delayed breakage triggered by external factors such as transportation vibration or temperature changes.

[0005] Existing buffer structures have also failed to effectively alleviate this problem. Rubber pads or foam strips are typically used as a buffer layer between the transparent panel and the metal frame, but the thickness and hardness of this layer are often standardized, failing to consider the actual stress distribution of the panel under different fastening points. In areas of high stress, a uniformly standardized buffer layer cannot provide sufficient stress dispersion space, and may even exacerbate the risk of glass breakage due to localized pressure concentration. Furthermore, assembly quality inspection relies primarily on manual visual inspection of whether the joints are uniform and the panel is flat. This inspection method is highly subjective, lacks quantification, and completely fails to detect the internal stress state of the transparent panel, making it impossible to promptly identify and correct assembly defects on the production line.

[0006] Therefore, how to predict and compensate for the assembly accuracy deviation caused by the accumulation of tolerances of multiple parts before assembly, how to monitor the stress distribution of the transparent panel in real time during the assembly process and dynamically adjust the fastening force of each fastening point, and how to intelligently arrange the buffer structure according to the actual stress distribution of the panel have become key issues to achieve high-precision and high-reliability assembly of transparent panels and metal frames. Summary of the Invention

[0007] This invention provides an intelligent assembly control method for a transparent panel and a metal frame, mainly including:

[0008] A three-dimensional point cloud dataset of surface topography and dimensional deviation of the first and second assembly components at multiple measurement locations is acquired. Features are extracted from this dataset using a convolutional neural network-based error prediction model to obtain a cumulative predicted assembly error value. Based on the cumulative predicted assembly error value and pre-stored component mating surface design reference position data, the relative positional offset of the first and second assembly components at each fastening point is determined. A pre-established fastening force mapping model is used to perform torque conversion on the relative positional offset to determine the initial fastening force parameter corresponding to each fastening point. The initial fastening force parameter drives a virtual assembly simulation model to perform a virtual simulation of the assembly loading process, obtaining internal stress tensor distribution data of the first assembly component after being subjected to force in each region. Based on the internal stress tensor distribution data, the values ​​of each fastening point are extracted. The stress field distribution characteristics of the surrounding area are analyzed to obtain a set of stress distribution feature vectors corresponding to each fastening point. It is then determined whether the maximum principal stress component of any fastening point in the set of stress distribution feature vectors exceeds a preset stress concentration safety threshold. If the maximum principal stress component exceeds the preset stress concentration safety threshold, the initial fastening force parameter is iteratively adjusted based on a stress equalization optimization strategy to obtain a target fastening force parameter. Based on the target fastening force parameter, the actual stress scalar value of the first assembly component at each fastening point is determined using a stress calculation model based on elasticity mechanics. The actual stress scalar value is then graded according to a preset stress grading interval standard to map and obtain buffer structure parameter combination data corresponding to each fastening point. Finally, based on the buffer structure parameter combination data and the target fastening force parameter, a set of assembly execution instructions is generated to drive the assembly equipment to perform fastening and buffer component installation actions. Furthermore, the step of extracting features from the three-dimensional point cloud data set of surface morphology and dimensional deviation using a convolutional neural network-based error prediction model to obtain the cumulative predicted value of assembly error includes: constructing a multi-dimensional spatial deviation matrix based on the spatial coordinate information and deviation information contained in the three-dimensional point cloud data set of surface morphology and dimensional deviation; using a convolutional neural network to extract local features from the multi-dimensional spatial deviation matrix to obtain an initial feature tensor containing local morphology features; extracting significant feature values ​​from the initial feature tensor by performing a max pooling operation to obtain a dimensionality-reduced global feature vector; if the feature value in the global feature vector is greater than a preset interference threshold, determining that there is contact surface interference and extracting the interference region features from the global feature vector; determining a multi-measurement point deviation trend sequence based on the interference region features; and using a fully connected layer to perform a nonlinear mapping on the multi-measurement point deviation trend sequence to integrate the multi-measurement point deviation superposition effect in the multi-measurement point deviation trend sequence to obtain the cumulative predicted value of assembly error.Further, determining the relative positional offset of the first assembly component and the second assembly component at each fastening point based on the cumulative predicted value of the assembly error and the pre-stored component mating surface design reference position data includes: obtaining the relative positional offset of the first assembly component and the second assembly component at each fastening point along the assembly axial and radial directions; constructing a fastening point spatial offset matrix based on the relative positional offset; using a support vector regression model to extract features from the fastening point spatial offset matrix to obtain the initial torque compensation coefficient corresponding to each fastening point; if the initial torque compensation coefficient is greater than a preset compensation threshold, determining that the corresponding fastening point has an overload risk and extracting the local force feature vector corresponding to the fastening point with overload risk; based on the... The stress concentration index of the contact surface between the first assembly component and the second assembly component at the corresponding fastening point is calculated using the local stress characteristic vector. A target fastening torque adjustment value is obtained by linearly scaling the contact surface stress concentration index. The final assembly execution torque sequence for each fastening point is determined based on the target fastening torque adjustment value and preset standard fastening torque data. Further, the step of converting the relative position offset into torque using a pre-established fastening force mapping model to determine the initial fastening force parameters corresponding to each fastening point includes: analyzing the relative position offset using a pre-established deformation analytical model containing a stiffness damping matrix to obtain elastic deformation parameters; and determining the elastic deformation parameters based on the preset contact surface roughness... A stress-friction calculation function is constructed; the torsional shear stress and thread friction coefficient during the assembly process are calculated using the stress-friction calculation function; the initial tightening force is determined based on the torsional shear stress and the thread friction coefficient; if the initial tightening force is greater than a preset yield strength threshold, it is determined that there is a risk of plastic deformation, and the corresponding preload attenuation rate is calculated based on the difference between the initial tightening force and the yield strength threshold; the dynamic compensation torque is obtained by multiplying the preload attenuation rate and a preset benchmark compensation coefficient; the initial tightening force is corrected based on the dynamic compensation torque to determine the final tightening force as the initial tightening force parameter; furthermore, the virtual assembly simulation model is driven based on the initial tightening force parameter. A virtual simulation of the assembly loading process is performed to obtain the internal stress tensor distribution data of the first assembly component after being subjected to force in various regions. This includes: extracting the principal stress and shear stress of each grid point within the first assembly component based on the internal stress tensor distribution data; performing dimensionality reduction processing on the principal stress and shear stress using a principal component analysis algorithm to obtain the principal stress eigenvector; calculating the spatial gradient value of the principal stress eigenvector between adjacent grid points to obtain the stress gradient distribution matrix; if an element in the stress gradient distribution matrix is ​​greater than a preset stress concentration threshold, the corresponding grid point is determined to be in a stress concentration state, and the set of grid points in the stress concentration state is extracted as the target stress area; and the nodal forces of each grid point within the target stress area are extracted.The fatigue life prediction value of the target stress area is determined by cyclic loading calculation of the nodal force using a pre-established fatigue damage accumulation model. If the fatigue life prediction value is less than a preset life safety threshold, fatigue fracture risk is determined and the coordinates of the structural weak point corresponding to the target stress area are output. Further, the initial fastening force parameter is iteratively adjusted based on a stress equilibrium optimization strategy to obtain the target fastening force parameter, including: extracting the initial fastening force parameter of the fastening point and the adjacent force parameters of adjacent fastening points to obtain a force parameter set; calculating the force difference value between the fastening point and the adjacent fastening points based on the force parameter set; and obtaining the spatial positional relationship between the fastening point and the adjacent fastening points. Based on the spatial positional relationship data, a local stiffness matrix is ​​constructed; a displacement solution function with the objective of minimizing deformation potential energy is constructed using the local stiffness matrix and the force difference value; the deformation displacement of the fastening point is calculated by solving the displacement solution function using a gradient descent algorithm; residual stress distribution data is calculated based on the product relationship between the deformation displacement and the local stiffness matrix; a stress equilibrium objective function is constructed based on the residual stress distribution data and the force difference value; the initial fastening force parameter is iteratively adjusted based on the stress equilibrium objective function until the residual stress distribution data meets the preset convergence condition to obtain the target fastening force parameter; furthermore, the step is based on the preset stress grading interval... The actual stress scalar value is graded and mapped to obtain buffer structure parameter combination data corresponding to each fastening point. This includes: comparing the actual stress scalar value with a preset stress grading interval standard to determine the stress level corresponding to each fastening point; extracting the buffer material properties and thickness dimension corresponding to each fastening point based on the stress level; concatenating the buffer material properties and thickness dimension to obtain buffer structure parameter combination data; querying a preset buffer property mapping table to obtain the stiffness coefficient and deformation limit corresponding to the buffer structure parameter combination data; and using the stiffness coefficient and deformation limit as input features to a pre-trained linear regression model for prediction. The compensation allowance corresponding to each fastening point is determined; it is determined whether the compensation allowance is greater than a preset allowance threshold; if the compensation allowance is greater than the allowance threshold, the product of the compensation allowance and a preset attenuation ratio is determined as the target compensation value; the pre-tightening torque of each fastening point is calculated based on the target compensation value; further, the step of generating an assembly execution instruction set for driving the assembly equipment to perform fastening and buffer component installation actions based on the buffer structure parameter combination data and the target fastening force parameter includes: extracting the material hardness identifier from the buffer structure parameter combination data and the torque scalar value from the target fastening force parameter; querying a preset assembly action matching table based on the material hardness identifier and the torque scalar value to obtain the initial assembly action sequence;A pre-trained decision tree model is used to parse the execution timestamps of each action node in the initial assembly action sequence to obtain an action time interval sequence. If the maximum time interval in the action time interval sequence is greater than a preset interval threshold, the initial assembly action sequence is split to obtain a first assembly action sub-sequence and a second assembly action sub-sequence. The first assembly action sub-sequence and the second assembly action sub-sequence are encoded according to a preset equipment control protocol code to obtain an assembly execution instruction set. The assembly equipment is driven to perform fastening actions and buffer installation actions through the assembly execution instruction set, and real-time displacement data of the assembly equipment during the fastening actions and buffer installation actions is collected. The actual installation stroke value is calculated based on the real-time displacement data. If the actual installation stroke value does not match a preset stroke standard value, a shutdown alarm signal is generated.

[0009] A computer chassis is manufactured using an intelligent assembly control method for a transparent panel and a metal frame. It includes a metal frame and a transparent panel, wherein the transparent panel is mounted on the metal frame via multiple fastening points, and a buffer structure is provided at the mating surface between the transparent panel and the metal frame.

[0010] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0011] To address the problem of uneven stress distribution at fastening points caused by surface morphology deviations of components in complex assembly structures, this invention first uses a convolutional neural network to analyze the deviation superposition effect of 3D point cloud data from multiple measurement points, predicting the cumulative assembly error and determining the relative positional offset of each fastening point. Then, a mechanical relationship model is used to convert the positional deviation into initial fastening force parameters. Subsequently, a virtual simulation model is driven to obtain the stress field distribution characteristics around each fastening point. When stress concentration exceeds a safety threshold, an optimization strategy that minimizes the force difference between adjacent fastening points is used to iteratively adjust the fastening parameters. Finally, based on the optimized stress values, buffer parameters are automatically matched according to a grading standard, and assembly instructions are generated. This invention, by combining geometric deviation measurement, stress simulation analysis, and fastening process optimization, achieves proactive control of stress distribution during assembly, effectively avoiding structural fatigue and deformation risks caused by stress concentration, and improving assembly quality and structural reliability. Attached Figure Description

[0012] Figure 1 This is a flowchart of the intelligent assembly control method for the transparent panel and the metal frame of the present invention;

[0013] Figure 2 This is a schematic diagram of the module frame in S103 of the intelligent assembly control method for the transparent panel and the metal frame of the present invention.

[0014] Figure 3This is a schematic diagram of the module frame in S105 of the intelligent assembly control method for the transparent panel and the metal frame of the present invention.

[0015] Figure 4 This is a schematic diagram of the module frame in S106 of the intelligent assembly control method for the transparent panel and the metal frame of the present invention.

[0016] Figure 5 This is a three-dimensional structural diagram of the computer chassis of the present invention.

[0017] Figure 5 In the middle, 1-transparent panel, 2-metal frame. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0019] like Figure 5 As shown, the computer case is made by an intelligent assembly control method of a transparent panel and a metal frame, including a metal frame 2 and a transparent panel 1. The transparent panel 1 is installed on the metal frame through multiple fastening points, and a buffer structure is provided at the mating surface between the transparent panel 1 and the metal frame 2.

[0020] like Figures 1-4 The intelligent assembly control method for the transparent panel and the metal frame in this embodiment may specifically include:

[0021] S101. Obtain a three-dimensional point cloud data set of surface topography and dimensional deviation of the first and second assembly components at multiple measurement positions. This three-dimensional point cloud data set includes spatial coordinate information for each measurement position and deviation information relative to the design dimensions. A convolutional neural network-based error prediction model is used to extract geometric features and deviation trends from the three-dimensional point cloud data set, obtaining a cumulative assembly error prediction value to characterize the superposition effect of deviations at multiple measurement points. The convolutional neural network-based error prediction model is a custom prediction model oriented towards the point cloud deviation field. It organizes the spatial coordinates x, y, z of the measurement positions and the corresponding deviation d into an N*4 dimensional input tensor, where N is the number of measurement points for a single component (example value is 2048). The network consists of three three-dimensional convolutional layers with kernel sizes of 3*3*3, 3*3*3, and 1*1*1, and channel numbers of 64, 128, and 256, respectively, used to extract local topography features and neighborhood deviation gradients. These are then processed by a full-processing network. Local max pooling aggregates global features, and then two fully connected layers map the features to the error values. The number of hidden layer nodes is 128 and 32, and the activation function is ReLU. The model output is a one-dimensional cumulative prediction value e of the assembly error, in millimeters, with a value range of 0 to 0.5. The loss function is mean squared error. During training, the measured deviation of the mating surface after assembly is used as the supervision label. The learning rate is initially set to 0.001 and decays to 0.5 times the original value every 20 rounds, thereby realizing the quantitative characterization of the superposition effect of deviations at multiple measurement points.

[0022] A three-dimensional point cloud dataset of surface topography and dimensional deviations of the first and second assembly components at multiple measurement locations is obtained. Based on the spatial coordinate information and deviation information contained in the three-dimensional point cloud dataset, a multi-dimensional spatial deviation matrix is ​​constructed. Specifically, the point cloud data at each measurement location is divided into an M-row, N-column two-dimensional plane according to a spatial grid. Each grid cell stores the deviation values ​​of four channels: X-coordinate deviation, Y-coordinate deviation, Z-coordinate deviation, and normal deviation, forming an M×N×4 three-dimensional matrix structure. A convolutional neural network with three convolutional layers is used to extract local features from the multi-dimensional spatial deviation matrix. The first convolutional layer uses 32 3×3 convolutional kernels, the second convolutional layer uses 64 3×3 convolutional kernels, and the third convolutional layer uses 128 3×3 convolutional kernels. Each convolutional layer is followed by a ReLU activation function to obtain an initial feature tensor containing local topography features. By performing a 2×2 window max pooling operation on the initial feature tensor, significant eigenvalues ​​are extracted to obtain a dimensionality-reduced global feature vector. If the eigenvalues ​​in the global feature vector are greater than the interference threshold of 0.05 mm, contact surface interference is determined, and the position indices of eigenvalues ​​greater than the interference threshold and their neighboring eigenvalues ​​are extracted from the global feature vector to form interference region features. The interference region features are arranged in spatial order according to the measurement positions to form a temporal structure, and the deviation change trend of each measurement point is recorded to obtain a multi-measurement point deviation trend sequence. The multi-measurement point deviation trend sequence is nonlinearly mapped using a network containing two fully connected layers. The first fully connected layer contains 256 neurons and uses the ReLU activation function, and the second fully connected layer contains one neuron and uses a linear activation function. The deviation values ​​of each measurement point in the sequence are multiplied by their weight coefficients and then summed to integrate the multi-measurement point deviation superposition effect in the multi-measurement point deviation trend sequence, thus obtaining the cumulative predicted value of assembly error.

[0023] In one implementation, a three-dimensional point cloud dataset of surface topography and dimensional deviations of the first and second assembly components at multiple measurement locations is first acquired. During measurement, 8 to 12 sampling areas are evenly distributed along the mating surfaces, with at least 50,000 points collected in each area, covering all critical mating surfaces of the workpiece, and the positional accuracy of a single point is controlled within 0.005 mm. The collected data includes topographic information such as contour undulations, ripples, and micro-scratches, as well as the dimensional deviations of each point relative to the nominal model. The deviation magnitude is typically within ±0.02 mm, providing a realistic and complete physical state input for subsequent assembly error prediction.

[0024] The three-dimensional point cloud dataset of surface topography and dimensional deviation includes spatial coordinate information of each measurement location and deviation information relative to the design size. In one embodiment, a high-precision three-dimensional laser scanning device is used to perform dense scanning on the mating surface of the first and second assembly components to obtain the original surface topography point cloud. The obtained original point cloud is spatially registered with the theoretical computer-aided design model of the component. After registration, the normal distance from each discrete measurement point in the original point cloud to the surface of the theoretical model is calculated, and this normal distance is the deviation information relative to the design size. In the constructed dataset, each data point contains three-dimensional spatial coordinates and a scalar value, which is used to characterize the magnitude of the processing error at that location. After obtaining the above dataset, the three-dimensional point cloud dataset of surface topography and dimensional deviation is processed using an error prediction model based on a convolutional neural network.

[0025] The error prediction model based on convolutional neural networks involves multi-step spatial data processing when extracting geometric features and deviation trends. Addressing the disordered and irregular distribution of the original 3D point cloud in space, the model first voxels the point cloud containing deviations: using the bounding box of the mating surface as a reference, the space is uniformly divided into cubic units with a side length of 2mm. A hash index is used to map each measurement point to its corresponding voxel. Voxels with sparse measurement points (less than 3) are merged with adjacent voxels to avoid gaps. The initial features of each voxel consist of five channels: the number of measurement points, the mean deviation, the variance of the deviation, the maximum deviation along the normal, and the centroid offset of the measurement point within the voxel, used to simultaneously characterize roughness, waviness, and local curvature changes. The 3D convolutional kernel slides across the voxel mesh with a stride of 1. The first layer uses 3*3*3 convolutions to capture the spatial gradient of the deviation between adjacent voxels, g=(d2-d1) / s, where d1 and d2 are the average deviations between adjacent voxels, and s is the voxel center distance. This is used to determine whether the deviation increases or decreases along the x, y, and z directions. After three layers of convolution and two max pooling operations, the receptive field expands to approximately 20mm. The model gradually aggregates from single-point deviations to the overall morphological undulations of the mating surface, thereby locating high-gradient regions that may cause assembly interference or uneven gaps.

[0026] In one implementation, after extracting local geometric features and deviation trends, the first fully connected layer within the model performs dimensionality reduction and nonlinear transformation on the local feature vectors distributed at different spatial locations, mapping the original 512-dimensional features into a 128-dimensional intermediate representation. Subsequently, the second fully connected layer uses an attention-weighted mechanism to globally fuse these intermediate representations, automatically focusing on key areas of the mating surface by calculating the correlation weights between features at different spatial locations. This fusion process can identify the contact state between the first and second assembly components on the mating surface, including three modes: local point contact, edge line contact, or large-area surface contact, and outputs a 64-dimensional global feature representation for subsequent error accumulation calculation.

[0027] The cumulative predicted assembly error is the core indicator output by this method, used to characterize the superposition effect of deviations at multiple measuring points. When two parts are fitted together, there are multiple microscopic protrusions and depressions on their respective surfaces. The height deviation of a single point is usually in the range of 0.005 to 0.05 mm. Under the action of assembly clamping force of 50 to 500 Newtons, they will interfere with each other, squeeze or avoid each other. When the height of a certain measuring point protrusion exceeds 0.02 mm, it will cause the entire part to tilt slightly by 0.001 to 0.01 degrees, which will amplify the assembly gap on the opposite side by 1.5 to 3 times. Based on the global features extracted above, this method completes the comprehensive calculation in the following steps: First, the deviation values ​​of each measuring point and the spatial coordinates are input into the contact mechanics solution module to simulate the transmission path of the clamping force along the normal direction; then, the part pose is iteratively updated according to the torque balance condition until the pose change between two adjacent times is less than 0.0001 mm; finally, the residual gap and interference of each measuring point are accumulated to obtain the cumulative predicted assembly error reflecting the combined effect of all local deviations.

[0028] In one possible implementation, the cumulative predicted assembly error value is a comprehensive quantitative index obtained by weighted summation of deviations from multiple measuring points. Its output can include two dimensions: the out-of-tolerance in the critical dimension direction and the deviation in the geometric tolerance direction. The critical dimension out-of-tolerance ranges from 0.00 to 0.50 mm, reflecting the actual deviation of the mating surfaces in the dimensional chain such as length and diameter after assembly. The geometric tolerance deviation ranges from 0.00 to 0.10 mm, reflecting the cumulative deviation of geometric features such as flatness, coaxiality, and perpendicularity. In specific implementation, the deviation data collected from each measuring point are first weighted and summed using weighting coefficients w1 to wn to obtain a comprehensive cumulative value S. Then, based on the assembly feature type, S is mapped to the corresponding dimensional out-of-tolerance or geometric tolerance deviation. The mapping relationship can use S*k1 to output the dimensional dimension and S*k2 to output the geometric dimension, where k1 and k2 are conversion coefficients obtained through calibration. This allows for the simultaneous provision of two types of interpretable quantities under the same predicted value.

[0029] For example, the predicted value could be the overall coaxiality error after the first and second assembly components are assembled, or the maximum peak and valley value of the end face runout. This predicted value can intuitively reflect the assembly quality level of the two components combined.

[0030] For example, consider the assembly scenario of a high-pressure compressor rotor disk in an aero-engine. The first assembly component is the preceding stage rotor disk, and the second assembly component is the following stage rotor disk. Using the method described above, the model accurately extracts the machining waviness characteristics and deviation distribution trends of the end face teeth based on the three-dimensional point cloud data set of the surface morphology and dimensional deviations of the end face teeth of the two-stage disks. This allows for the prediction of the cumulative concentricity error of the rotor after the two-stage disks are joined. Assembly personnel can predict the assembly quality before actual physical assembly, guiding subsequent selection or repair work and improving the first-pass yield of complex mechanical products.

[0031] S102. The cumulative predicted assembly error is organized in a multi-dimensional vector form, including three components: cumulative expansion / contraction along the assembly axis, cumulative misalignment in the radial direction, and angular deflection around the assembly axis. Each component is arranged in the order of the preceding process nodes, forming a data sequence with a length consistent with the number of fasteners. For the mating surface between the first and second assembly components, the three-dimensional coordinates of each fastener in its nominal state are retrieved from the pre-stored design reference position data, including the nominal axial position Za, nominal radial radius Rb, and nominal circumferential angle Ac. The component corresponding to the k-th fastener in the cumulative predicted value is mapped to the local coordinate system of that fastener according to its geometric position. Specifically, the fasteners are first indexed and aligned according to their distribution order on the mating surface, and then the radial component is proportionally distributed according to the actual radius of the fastener from the assembly axis to avoid fasteners near the axis bearing excessive radial deviation. The axial offset dZk = Zk - Za and the radial offset dRk = Rk - Rb at the fastening point are then obtained using the following formulas, where Zk and Rk are the actual predicted coordinates after superimposing the cumulative predicted components. Taking a flange mating surface with a typical set of 8 fastening points as an example, if the axial component in the cumulative predicted value is 0.12 mm, the radial component is 0.08 mm, and the angular component is 0.03 degrees, then after mapping and difference calculation, dZk at each fastening point is distributed in the range of 0.09 to 0.15 mm, and dRk is distributed in the range of 0.05 to 0.11 mm, forming the offset dataset required for subsequent adjustments.

[0032] The relative positional offsets of the first and second assembly components at each fastening point along the assembly axial and radial directions are obtained. A spatial offset matrix of fastening points is constructed based on these offsets. The matrix has an n×3 dimension, where n is the number of fastening points, and the three columns correspond to the axial offset, radial X-axis offset, and radial Y-axis offset, respectively. This spatial offset matrix is ​​input into a support vector regression model, which uses a radial basis function kernel for nonlinear mapping and outputs the initial torque compensation coefficient for each fastening point. This coefficient ranges from 0 to 1.5, representing the compensation multiple relative to the standard fastening torque. The support vector regression model is trained using historical assembly data, with training samples containing actual assembly stress distribution data measured under different offset combinations. The initial torque compensation coefficient is obtained. If the initial torque compensation coefficient is greater than 0.85, the corresponding fastening point is considered to have an overload risk. In this case, the local stress feature vector corresponding to the fastening point with overload risk is extracted. This feature vector contains four parameters: peak normal stress at the contact surface, tangential stress component, stress gradient, and stress distribution uniformity index. Based on the local stress characteristic vector, the stress concentration index of the contact surface between the first and second assembly components at the corresponding fastening points is calculated using the von Mises equivalent stress criterion. This index is defined as Si = σmax / σavg, where σmax is the local maximum stress and σavg is the average stress of the contact surface. The contact surface stress concentration index is converted into a target fastening torque adjustment value through linear scaling. The scaling formula is ΔT = T0 × k × (Si - 1.0), where T0 is the standard fastening torque, k is a scaling factor with a value of 0.3, and the torque is reduced when Si is greater than 1.0. Based on the adjusted target fastening torque and the preset standard fastening torque data, the final assembly execution torque sequence for each fastening point is determined.

[0033] Taking the mating surface between the spindle box and the bed of a precision machine tool as an example, there are 12 fastening points evenly distributed on the flange surface, numbered 1 to 12. Their theoretical three-dimensional coordinates are represented by a global coordinate system with the geometric center of the mating surface as the origin and the assembly spindle as the z-axis. For example, the coordinates of point 1 are x=150.0, y=0.0, and z=20.0, all in millimeters. The remaining points are taken at 30-degree intervals on a circle with a radius of 150.0 millimeters. The cumulative predicted value of assembly error output in the previous step is essentially still the critical dimension deviation and form and position deviation. In this step, it is equivalently mapped to six-degree-of-freedom components according to rigid body kinematics, denoted as the rotation angles about the x, y, and z axes rx=0.012, ry=0.008, and rz=0.003, in degrees, and the translations along the three axes tx=0.02, ty=-0.01, and tz=0.05, in millimeters. Based on these six components, a fourth-order homogeneous transformation matrix T is constructed, where the rotation sub-blocks are synthesized from the three-axis small rotation angles in the order zyx, and the translation columns are filled with tx, ty, and tz. The theoretical coordinates of each fixed point are extended into a four-dimensional homogeneous column vector P, and matrix multiplication with T is performed P'=T*P to obtain the expected coordinates P' after deflection. Taking point 4 as an example, its theoretical coordinates are x=0.0, y=150.0, and z=20.0. After substituting them, they are dominated by rx and tz, and the solutions are x'=0.0, y'=149.98, and z'=20.08. The expected difference vector d = P' - P is obtained by subtracting the theoretical difference. Its projection component on the z-axis is the relative offset dz = 0.08 mm along the assembly axis. On the plane perpendicular to the assembly axis, dx and dy are combined into the radial offset dr, using dr = sqrt(dx*dx + dy*dy). This combination method is suitable for the assembly coordinate system defined by the orthogonal x and y axes in this example. Substituting dx = 0.0 and dy = -0.02, we get dr = 0.02 mm.

[0034] S103. Obtain the relative position offset at each fastening point. Convert the relative position offset into torque using a pre-established fastening force mapping model based on mechanical relationship fitting to determine the initial fastening force parameters corresponding to each fastening point. The fastening force mapping model takes the relative position offset d of each fastening point in the assembly reference coordinate system as input and the corresponding initial fastening torque T as output, where d is in millimeters and T is in Newton-meters, with a value range of 8 to 45 Newton-meters. This model is obtained by least-squares fitting of the offset and measured torque data of the same type of connector under different pre-tightening states, using a piecewise linear form T = k * d + T0, where k is the fastening stiffness coefficient, reflecting the torque change corresponding to a unit offset, with a value range of 15 to 25 Newton-meters per millimeter; T0 is the basic pre-tightening torque of the fastening point, determined by the connector specifications and material, typically taken as 10 Newton-meters. When d is less than 0.05 mm, it is considered to be in place, and T0 is output directly; when d exceeds 0.8 mm, it is truncated to the upper limit of 45 Nm to avoid damage from overtightening. Substituting the offset into each fastening point sequentially yields the initial fastening force parameter for that point, which can be used by subsequent tightening actuators.

[0035] Based on the relative offset data at each fastening point obtained in step S102, the deformation is analyzed using a pre-established analytical model. This analytical model employs a finite element stiffness matrix K and a Rayleigh damping matrix C. The stiffness matrix K is obtained by discretizing the material at each fastening point using a finite element mesh, based on the elastic modulus and geometric dimensions of the material. The damping matrix C = αM + βK, where α is 0.1 and β is 0.001. The model input is the relative offset vector d, and the output is the strain vector ε at each fastening point. Elastic deformation parameters, including axial strain εa and radial strain εr, are calculated based on the strain vector ε and the material's elastic modulus E. A stress-friction calculation model is then constructed, with the elastic deformation parameters εa and εr as inputs and a pre-defined contact surface roughness Ra. The roughness Ra is measured using a surface profilometer and ranges from 0.8 micrometers to 3.2 micrometers. The model first calculates the axial stress σa = E × εa based on εa. Then, it obtains the thread friction coefficient μ from the roughness-friction coefficient lookup table based on the contact surface roughness Ra. When Ra is 0.8 μm to 1.6 μm, μ is taken as 0.12 to 0.15; when Ra is 1.6 μm to 3.2 μm, μ is taken as 0.15 to 0.18. Next, it calculates the torsional shear stress τ = G × γ based on εr, where G is the shear modulus and γ is the shear strain. Finally, it calculates the initial tightening torque T0 = τ × A × d / 2 + μ × F × d / 2 based on the torsional shear stress τ, the thread friction coefficient μ, and the nominal bolt diameter d, where A is the effective contact area of ​​the thread and F is the axial preload. The initial tightening torque T0 is obtained and converted into the equivalent stress σeq = T0 / (π × d³ / 16) for comparison with the material yield strength. If the equivalent stress σeq is greater than the preset yield strength threshold σy, which is determined based on the fastener material grade (640 MPa for 8.8 grade bolts and 900 MPa for 10.9 grade bolts), then a risk of plastic deformation is identified. The stress excess ratio η = (σeq - σy) / σy is calculated. The preload attenuation rate λ is obtained by referring to the preload attenuation curve based on the stress excess ratio η. When η is 0 to 0.1, λ is 0.05 to 0.15; when η is 0.1 to 0.2, λ is 0.15 to 0.30. The dynamic compensation torque ΔT = λ × Tb is obtained by multiplying the preload attenuation rate λ with the preset baseline compensation torque Tb. The baseline compensation torque Tb is set according to the bolt specification: 20 N·m for M8 bolts, 40 N·m for M10 bolts, and 70 N·m for M12 bolts. The initial tightening torque T0 is added to the dynamic compensation torque ΔT to obtain the final tightening torque Tf = T0 + ΔT. This final tightening torque is the final assembly execution torque of each fastening point.

[0036] In one implementation, the system calls the relative position offsets of each fastening point determined in S102, and organizes them into a two-dimensional array according to the fastening point number. Each row contains axial deviation and radial deviation components. The axial deviation is usually in the range of -0.05mm to +0.05mm, and the radial deviation is usually in the range of -0.03mm to +0.03mm. Fastening points outside this range are marked as abnormal points and recorded separately. The organized array is directly sent to the subsequent mechanical compensation module as the input vector for bolt preload correction and contact surface deformation calculation, avoiding repeated data acquisition and secondary conversion.

[0037] The fastening force mapping model takes the three-dimensional offsets dx, dy, and dz as inputs and outputs the fastening torques Tx, Ty, and Tz in the X, Y, and Z directions. At the microscopic level, the mating surface between the cast iron bed and the spindle box exhibits contact deformation caused by surface roughness and waviness. The model uses a piecewise stiffness form T = k1*d + k2*d^n to describe this nonlinear relationship, where k1 is the normal contact stiffness coefficient, k2 is the tangential contact stiffness coefficient, and n ranges from 1.3 to 1.5, reflecting the stiffness hardening trend as the compression increases. The elastic modulus E of cast iron is taken as 110 GPa as the initial fitting benchmark for k1. In the elastic segment with compression less than 5 micrometers, Hertzian contact theory is used to solve for k1; in the rough contact segment with compression from 5 to 30 micrometers, fractal geometry theory is used to correct k2 and the exponent n to cover the abrupt change in stiffness from peak contact to surface contact. The above coefficients were obtained by regression analysis of 1200 sets of assembly bench test samples, which covered 3 bed specifications and 4 preload conditions. The coefficient of determination for the fit was not less than 0.95. The regression results were solidified into a lookup table plus interpolation form for real-time access.

[0038] During the model building phase, finite element simulation software was used to apply a preload gradient of 0.5 to 2.5 kN to the bolt assembly process of the spindle box and bed, with each step size of 0.2 kN. The nodal displacements of eight fastening points in the X, Y, and Z directions were extracted. On the assembly workshop side, a triaxial eddy current displacement sensor and a piezomagnetic force sensor were placed at each fastening point, with a sampling frequency of 1 kHz, covering 60 assembly conditions. Approximately 12,000 sets of displacement-preload paired samples were obtained and merged with the simulation data to form a training set. The sample set was fitted using multivariate nonlinear regression with a Gaussian kernel. The input was a 3D offset vector d=[dx,dy,dz], and the output was a 3D compensation force vector F=[Fx,Fy,Fz]. The fitting yielded a 3*3 mapping matrix K, i.e., F=K*d, where the diagonal elements of K reflect the principal stiffness and the off-diagonal elements reflect the coupling stiffness. The fitting residuals were controlled within 3% through 10-fold cross-validation. After obtaining the measured offset vector, substitute it into K to obtain the compensation force, and then convert it into tightening torque T=F*L according to the wrench lever arm L, and send it to the electric tightening gun for execution.

[0039] After obtaining the axial position deviation Δz and radial position deviation Δr of each fastening point, the system calculates the required fastening force through the following process. First, a local stiffness matrix K is established, where the axial stiffness coefficient kz is calculated based on the elastic modulus of the bolt material and the contact area, and the radial stiffness coefficient kr is determined by the friction coefficient of the contact surface and the preload condition. For M8 bolts, kz is set to 150 N / mm, and kr is set to 85 N / mm. When an axial deviation of 0.3 mm and a radial deviation of 0.15 mm are detected at a fastening point, the required normal clamping force Fz is equal to kz multiplied by Δz, resulting in 125 N, and the shear force Fr is equal to kr multiplied by Δr, resulting in 12.75 N. Based on the calculation results, the system adjusts the tightening torque at that point to generate sufficient preload to eliminate the position deviation.

[0040] For fastening point number four on the machine tool spindle box flange, its axial offset is 0.08 mm. The fastening force mapping model, based on the local contact stiffness curve of the area where this point is located, determines that a normal preload of 8500 N is required to ensure complete contact between the mating surfaces. This value is obtained by multiplying the contact stiffness of 106250 N / mm by the offset of 0.08 mm. The model uses the thread tightening torque relationship T=KFd for torque conversion, where T is the tightening torque, F is the preload of 8500 N, d is the nominal bolt diameter of 0.016 m, and K is the comprehensive friction coefficient of 0.15. Substituting the parameters into the formula, the required tightening torque for this fastening point is calculated to be 20.4 Nm. The model repeats the above calculation process for all fastening points on the spindle box flange, ultimately generating a complete mapping table containing the axial offset, preload, and tightening torque for each point.

[0041] After completing the torque conversion, a set of tightening control parameters is generated for each tightening point. In addition to the target torque T0, the contact torque Tc and the slope of the elastic segment k also need to be given: Tc is used to overcome the initial friction between the threaded pair and the support surface. The value is obtained by first screwing in at a low speed before tightening, and recording the steady-state torque Tf when the bolt head does not contact the workpiece. Let Tc = 1.2 * Tf, which usually falls in the range of 0.5 to 3.0 Nm; k reflects the rate of torque increase with the rotation angle after entering the contact phase. It is calculated by taking two points from the contact point to 0.7 * T0 on the tightening curve, and using k = (T2 - T1) / (θ2 - θ1) in Nm per degree. For common bolts, k falls between 0.3 and 1.5, which is used for subsequent out-of-tolerance judgment and anti-slip thread protection. The aforementioned fields T0, Tc, k, as well as speed gear and timeout threshold, are encapsulated into a JSON message in a fixed order and sent to the tightening shaft controller or smart wrench via industrial Ethernet or EtherCAT bus as control instructions for performing tightening actions at the fastening point.

[0042] S104. Based on the initial fastening force parameters, drive the virtual assembly simulation model to perform virtual simulation of the assembly loading process, and obtain the internal stress tensor distribution data of the first assembly component after being subjected to force in each region. The specific process is as follows: The initial fastening force parameters obtained in S103 are loaded onto the corresponding assembly interface nodes of the virtual assembly simulation model in the form of load boundary conditions, and the target force value is applied in a linearly increasing manner from 0 to a preset time period according to the preset loading curve, while keeping the normal degree of freedom of the fixed constraint surface of the first assembly component locked; then, the finite element solver is called to perform quasi-static solution on the model, calculate the displacement field and strain field of each element node under load step by step, and derive the Cauchy stress tensor at each integration point based on the material elastoplastic constitutive relation to obtain six-component data containing three normal stress components and three shear stress components. To facilitate subsequent steps, the stress tensor data of all elements are collected according to the N pre-defined region labels of the first assembly component. Within each region, the average value is weighted by the element volume, forming a stress tensor distribution dataset indexed by the region number and containing six-component stress tensors. This dataset is stored in structured data format for use in step S105. The value of N ranges from 8 to 32, with the specific value determined based on the geometric complexity of the component.

[0043] The initial fastening force parameters are obtained, and the virtual assembly simulation model is driven by these parameters to perform assembly loading simulation on the first assembly component, obtaining the internal stress tensor distribution data of the first assembly component after being subjected to force. Based on the internal stress tensor distribution data, the principal stress and shear stress of each grid point in the first assembly component are extracted. The multidimensional stress sample matrix composed of the principal stress components and shear stress components is reduced in dimensionality using a principal component analysis algorithm, retaining the first two principal components with a cumulative variance contribution rate of not less than 95%, to obtain the principal stress principal component vectors used to characterize the dominant force direction of each grid point. Based on the principal stress principal component vectors, the difference between the vectors of adjacent grid points is calculated and divided by the grid spacing to obtain the stress gradient distribution matrix G, where the element g = |v1-v2| / d, v1 and v2 are the principal component vectors of adjacent grid points, and d is the grid spacing. If an element in G is greater than the preset stress gradient concentration threshold of 0.35 MPa / mm, the corresponding grid point is determined to be in a stress concentration state. All grid points in the stress concentration state are extracted as the target stress zone. This threshold is only used for gradient abrupt change determination and is independent of the stress concentration safety threshold used for subsequent comparison of the maximum principal stress amplitude. The target stress zone is obtained, and the nodal forces of each grid point within the target stress zone are extracted. These nodal forces are input as cyclic loads to a pre-established fatigue damage accumulation model based on Miner's linear cumulative damage rule. The allowable number of cycles Ni corresponding to each load amplitude is obtained through the SN curve. The cumulative damage degree D = Σni / Ni is obtained by summing the ratio of the actual number of cycles Ni at each stage to Ni. The equivalent number of cycles when D = 1 is taken as the predicted fatigue life value of the target stress zone. If the predicted fatigue life is less than the preset life safety threshold of 100,000 cycles, it is determined that there is a risk of fatigue fracture. The grid point with the largest D value is located in the target stress area, and its coordinates in the first assembly component coordinate system are output as the coordinates of the structural weak point. The corresponding predicted fatigue life value and the peak information in the stress gradient distribution matrix are also transmitted to the subsequent fastening force optimization stage.

[0044] The system parses the received initial tightening force parameters and transforms them into dynamic boundary loading conditions required for the virtual assembly simulation model. For the virtual assembly simulation model of the machine tool spindle box, the system converts the set torque parameters into a preload sequence applied to the virtual bolt cross-section, setting the loading time to 2.0 seconds and dividing it into 100 incremental steps. Subsequently, the system calls the built-in nonlinear finite element solver engine and uses the Newton-Raphson iterative algorithm to perform virtual simulation of the assembly loading process. In each incremental step, the solver engine calculates the displacement changes of the contact surface and internal mesh nodes based on the spindle box material's set elastic modulus of 160 GPa and Poisson's ratio of 0.28. As the virtual load is gradually applied, the system tracks the deformation state of the three-dimensional voxel mesh inside the spindle box in real time and uses the generalized Hooke's law to convert the strain data of each node into stress components. After the loading process simulation is complete, the system traverses all mesh elements of the spindle box model, extracts and generates the internal stress tensor distribution data of the first assembly component after being subjected to force in each region. This distribution data includes the values ​​of normal stress and shear stress in three orthogonal directions for key areas such as the bearing hole periphery of the spindle box and the flange mating surface. For example, the system calculation shows that the maximum first principal stress in the front bearing housing area of ​​the spindle box reaches 45.6 MPa, and the maximum shear stress is concentrated in the flange root area, with a specific value of 12.3 MPa.

[0045] S105. For the internal stress tensor distribution data, extract the stress field distribution features in the neighborhood area centered on each fastening point. The stress field distribution features include the principal stress direction and principal stress amplitude, as well as the stress gradient attenuation trend, to obtain the stress distribution feature vector set corresponding to each fastening point.

[0046] The set of stress distribution feature vectors is obtained, and the K-means clustering algorithm is used to cluster the set of stress distribution feature vectors to obtain multiple feature vector clusters containing different fastening points and the cluster centers corresponding to each feature vector cluster. For each cluster center, the Euclidean distance between each feature vector in the set of stress distribution feature vectors and the cluster center of its respective feature vector cluster is calculated to obtain the deviation distance value corresponding to each fastening point. According to the deviation distance value, if the deviation distance value is greater than a preset deviation threshold of 0.35, it is determined that there is abnormal stress at the corresponding fastening point, and the fastening point with abnormal stress is extracted as the fastening point to be inspected. The fastening point to be inspected is obtained, and the node load data of the fastening point to be inspected within the neighborhood is extracted. The node load data includes the load amplitude Fi of each finite element node within the neighborhood and the distance ri from the node to the center of the fastening point, where i is the node number in the neighborhood, ranging from 1 to n, and n is the total number of neighborhood nodes. Using the load amplitude Fc at the center of the fastening point as a reference, the neighborhood is divided into a core area and an outer expansion area. The core area includes nodes whose ri is less than or equal to 0.3 times the neighborhood radius R, while the outer expansion area includes nodes whose ri is greater than 0.3 times R. The load amplitudes of the nodes in both areas are summed to obtain Fcore and Fout, respectively. The load concentration C = Fcore / (Fcore + Fout). Taking a neighborhood radius R of 8 mm and n = 24 nodes as an example, if the sum of the load amplitudes Fcore of the 6 nodes in the core area is 420 N, and the sum of the load amplitudes Fout of the 18 nodes in the outer expansion area is 180 N, then C = 0.7. If the load concentration is greater than the preset concentration threshold of 0.6, the fastening state of the fastening point under inspection is determined to be loose, and the corresponding feature vector in the stress distribution feature vector set is marked as a loose label, serving as the input basis for correcting the stress distribution feature vector set in subsequent steps.

[0047] After obtaining the internal stress tensor distribution data of the first assembly component under stress in various regions, the system needs to further refine the local features of this massive amount of data.

[0048] Specifically, the system focuses on the fastening connection area, which is most complex in terms of stress and prone to concentrated deformation during the assembly process. Through spatial geometric positioning and data filtering, it defines a precise analysis range for subsequent feature extraction.

[0049] For example, the system first identifies the coordinates of all fastening points based on the three-dimensional geometric topology information of the virtual assembly simulation model of the machine tool spindle box. These fastening points typically correspond to the center of the flange bolt holes on the spindle box housing or the midpoint of the axis of the bearing housing preload bolts. Subsequently, the system sets a spatial envelope of a preset size as the neighborhood range with each fastening point as the geometric center.

[0050] For example, for the bolt holes on the front flange of the spindle box, the system uses its center point as the fastening point and defines a spherical region with a radius of 30 mm as the neighborhood. The system traverses the internal stress tensor distribution data and extracts the stress tensor data of all three-dimensional voxel mesh nodes falling within this spherical neighborhood, forming a local stress dataset for that specific fastening point. Based on the obtained local stress dataset, the system further extracts the stress field distribution characteristics within the neighborhood centered on each fastening point. These stress field distribution characteristics are key indicators describing the stress state of a local area, including the principal stress direction and magnitude, as well as the stress gradient decay trend.

[0051] In one possible implementation, the system analyzes the stress tensor matrix of each mesh node within its neighborhood to obtain the principal stress directions and magnitudes. The stress tensor matrix contains the normal and shear stress components of the node along the three orthogonal coordinate axes. The system determines the principal stress state of the node by calling the matrix operation module to solve for the eigenvalues ​​and eigenvectors of the stress tensor matrix. Specifically, the three calculated eigenvalues ​​represent the magnitudes of the maximum, intermediate, and minimum principal stresses of the node, while the corresponding eigenvectors indicate the directions of the principal stresses in three-dimensional space.

[0052] For example, in the neighborhood of the fastening point of the spindle box bearing housing, the system calculates the maximum principal stress amplitude of a node through matrix decomposition, and the corresponding principal stress direction vector points radially outward from the bearing hole. The system summarizes the principal stress amplitudes and directions of all nodes in the neighborhood to construct a principal stress field distribution map around the fastening point, thus clearly reflecting the transmission path and main force direction of the preload within the spindle box material. Furthermore, to comprehensively characterize the diffusion behavior of the fastening force within the material, the system also needs to calculate the stress gradient decay trend. The stress gradient decay trend reflects the spatial rate of change of stress amplitude as the distance from the center of the fastening point increases. Using the fastening point as the origin, the system extracts the equivalent stress values ​​of each node along multiple preset spatial ray directions. Subsequently, the system calculates the ratio of the stress difference between adjacent nodes to the spatial distance, obtaining the local stress gradient at each location. By fitting the stress gradient data along these ray directions, the system can generate a continuous curve reflecting the stress decay with distance.

[0053] For example, around the fastening point of the spindle box flange mating surface, the system calculation found that the stress gradient decreases faster along the direction perpendicular to the flange surface, while the stress gradient decreases slower along the radial direction of the flange surface. The quantitative data of this decrease trend can accurately characterize the contact stiffness distribution at the mating surface.

[0054] Understandably, after extracting the aforementioned features, the system structurally combines the principal stress direction vector, the statistical characteristics of the principal stress amplitude, and the fitting parameters of the stress gradient attenuation trend corresponding to each fastening point. Through this data fusion process, the system ultimately obtains a set of stress distribution feature vectors corresponding to each fastening point. Each feature vector in this set completely and independently describes the local mechanical response characteristics of a specific fastening connection, thus transforming the complex spatial stress field into a structured numerical vector.

[0055] S106. Determine whether the maximum principal stress component of any fastening point in the stress distribution feature vector set exceeds a preset stress concentration safety threshold. If the maximum principal stress component exceeds the preset stress concentration safety threshold, then iteratively adjust the initial fastening force parameter based on a stress equalization optimization strategy aimed at minimizing the force difference between adjacent fastening points to obtain the target fastening force parameter.

[0056] The maximum principal stress component of any fastening point in the stress distribution feature vector set is obtained. If the maximum principal stress component exceeds a preset stress concentration safety threshold, for example, a threshold set to 0.6 to 0.8 times the material yield strength, the initial fastening force parameter of the fastening point and the adjacent force parameters of adjacent fastening points are extracted to obtain a force parameter set. Based on the force parameter set, the force difference value between the fastening point and the adjacent fastening points is calculated, and the spatial positional relationship data between the fastening point and the adjacent fastening points is obtained. Based on the spatial positional relationship data, a local stiffness matrix K is constructed, and the matrix elements are jointly determined by the fastening point spacing, the connection direction cosine, and the equivalent elastic modulus of the connected components. Using the local stiffness matrix K and the force difference value, a displacement solution function E = 0.5 * u transpose * K * u F transpose * u is constructed with the objective of minimizing deformation potential energy. Here, u is the deformation displacement to be determined, and F is the equivalent nodal load vector mapped from the force difference value. The gradient descent algorithm is used to solve the displacement solution function to obtain the deformation displacement u of the fastened point. The residual stress distribution data is then calculated using the residual stress sigma = K * u. Based on this, a stress equilibrium objective function J is constructed. J consists of two linearly superimposed parts: the first part is the sum of squares of the force differences between adjacent fastened points, used to suppress sudden changes in local force, with a weighting coefficient ranging from 0.4 to 0.6; the second part is a penalty term for exceeding the maximum principal stress limit, accumulated by the square of the excess when the residual stress exceeds the safety threshold, with a weighting coefficient ranging from 0.4 to 0.6. The sum of the weights of the two parts is normalized to 1. Based on the stress equilibrium objective function J, the gradient descent algorithm is used to iteratively adjust the initial fastening force parameter. After each iteration, the displacement solution function is re-executed to obtain the updated residual stress distribution until the change in J between two adjacent iterations is less than 0.001 or the maximum principal stress falls back to within the safety threshold, thus obtaining the target fastening force parameter.

[0057] After obtaining the set of stress distribution feature vectors corresponding to each fastening point, the system further performs a safety assessment on these feature vectors. Specifically, it determines whether the maximum principal stress component of any fastening point in the set of stress distribution feature vectors exceeds a preset stress concentration safety threshold. The stress concentration safety threshold is a critical value preset based on the physical and mechanical properties of the machine tool spindle box housing material and the assembly process specifications. Typically, the system extracts the yield strength of the spindle box material and calculates this threshold in conjunction with a preset safety factor to ensure that the material around the fastening point does not undergo plastic deformation or microcracks during actual assembly and operation. The system traverses the feature vector set, extracts the maximum principal stress amplitude component in the feature vector of each fastening point, and compares it with the safety threshold one by one.

[0058] For example, in the assembly analysis of the front flange of the machine tool spindle box, the system detected that the maximum principal stress component around the bolt hole numbered 5 reached 78.4 MPa, while the system's preset stress concentration safety threshold for cast iron materials is 60.0 MPa. The measured value is 18.4 MPa higher than the threshold, exceeding the limit by approximately 30.7%. This local stress exceedance is usually caused by excessive initial preload or minor geometric irregularities on the flange mating surface. If not adjusted, it can easily lead to local warping of the flange or seal failure during subsequent machining vibrations. If the maximum principal stress component exceeds the preset stress concentration safety threshold, the system triggers a stress equalization optimization strategy based on minimizing the force difference between adjacent fastening points, and iteratively adjusts the initial fastening force parameter. The initial fastening force parameter refers to the initial preload value set for each fastening bolt at the beginning of the assembly simulation. In this example, the initial preload of bolt number 5 is set to 28.5 kN. The core logic of the stress equilibrium optimization strategy lies in globally coordinating the stress state of each fastening point to disperse localized excessive stress to surrounding areas with stronger load-bearing capacity, while ensuring that the clamping force distribution across the entire mating surface is as uniform as possible. The iterative convergence condition is set as follows: the maximum principal stress falls below 60.0 MPa and the difference in preload between adjacent bolts does not exceed 2.0 kN. Minimizing the force difference between adjacent fastening points is the key optimization objective of this strategy, aiming to constrain the difference in preload between adjacent bolts in spatial location and prevent additional bending moments and shear stresses caused by significant differences in the forces on adjacent bolts.

[0059] Specifically, when executing the stress equalization optimization strategy, the system first constructs a comprehensive objective function containing multiple constraints and penalty terms. The first part of this objective function is the sum of squares of the force differences between adjacent fastening points. Based on the three-dimensional topology of the spindle box, the system establishes a spatial adjacency matrix between each fastening point and calculates the preload difference between all pairs of adjacent fastening points accordingly. The second part of the objective function is a maximum principal stress over-limit penalty term, used to impose a large numerical penalty on fastening points that exceed the safety threshold. The system uses gradient descent or heuristic search algorithms to solve this objective function. In each iteration, the algorithm calculates the partial derivatives or fitness changes of the objective function with respect to each fastening force parameter, thereby determining the adjustment direction and step size of the preload for each fastening point.

[0060] In one possible implementation, regarding the excessive stress in the hole of bolt #5, the algorithm calculates during the first iteration that the initial tightening force parameter of bolt #5 needs to be appropriately reduced. However, simply reducing the force at this point would result in insufficient overall clamping force in that area of ​​the flange. Therefore, based on the goal of minimizing the force difference between adjacent tightening points, the algorithm simultaneously calculates the force margin of its adjacent bolts #4 and #6 while reducing the force of bolt #5, and appropriately increases the tightening force parameters of these two adjacent bolts. Through this coordinated adjustment, the system, in the process of peak smoothing and valley filling, gradually reduces the stress peak around bolt #5 back to within the safe threshold, while the force difference between bolts #4, #5, and #6 is compressed to a minimum.

[0061] It's important to note that this iterative adjustment isn't a one-time event, but rather a dynamic, cyclical process of approximating the optimal solution. After each update of the fastening force parameters, the system invokes the underlying mechanical solution engine to recalculate the new stress distribution feature vectors for each fastening point under the current parameter configuration, and then performs threshold checks and objective function evaluations again. As the number of iterations increases, the force parameters for each fastening point are continuously fine-tuned, and the stress distribution across the entire spindle box assembly surface gradually becomes smoother and more uniform.

[0062] Understandably, the system continuously monitors the convergence status during the iteration process. When, after multiple iterations, the maximum principal stress components of all fastening points are strictly below the preset stress concentration safety threshold, and the change in the sum of the force differences between adjacent fastening points in the objective function is less than the preset convergence accuracy, the system determines that the iteration process has ended. At this point, the set of parameters fixed after multiple rounds of optimization and updates is output as the target fastening force parameters. This set of target fastening force parameters not only avoids the risk of local stress concentration but also achieves force balance at the assembly joint surface.

[0063] S107. Based on the target fastening force parameter, determine the actual stress scalar value of the first assembly component at each fastening point along the force direction using a stress calculation model based on elasticity mechanics.

[0064] The target fastening force parameter, the elastic modulus of the first assembly component, and the contact surface geometry of the first assembly component at each fastening point are obtained. The contact surface geometry includes the contact surface diameter, contact surface thickness, and effective bearing area around the fastening point. Using the target fastening force parameter as the external load input and the elastic modulus of the material as the material property input, a stress-strain mapping relationship is established based on Hooke's law sigma=E*epsilon, where sigma is stress, E is the elastic modulus of the material, and epsilon is strain. This constructs an elastic mechanical stress calculation model that outputs the stress distribution at the fastening points. Using this stress calculation model, the force direction vector of the first assembly component at each fastening point is extracted. Combining this with the ratio of the projected component of the fastening force along the force direction to the effective bearing area of ​​the contact surface, the actual stress scalar value along the force direction at each fastening point is calculated. The contact surface geometry is imported into the finite element preprocessing module. The geometric model around the fastening points is meshed into hexahedral or tetrahedral elements. The actual stress scalar values ​​are applied as boundary conditions to the corresponding nodes. The solver calculates the local stress tensor matrix corresponding to each fastening point. Based on the local stress tensor matrix, the principal stress components sigma1, sigma2, and sigma3 along the three orthogonal directions are extracted. The equivalent stress value sigma_v of each fastening point is calculated using the von Mises criterion. Obtain the material yield strength value sigma_s of the first assembly component and the pre-calibrated material stress-strain curve. Determine whether the equivalent stress value is greater than the material yield strength value. If sigma_v is greater than sigma_s, the difference between sigma_v and sigma_s is used as the overstress force in the plastic stage. Substitute this into the plastic constitutive model based on the Lamberg-Osgood relation, and combine the hardening index and strength coefficient of the yield segment in the stress-strain curve to solve for the corresponding plastic strain. Multiply this by the characteristic length of the bearing direction of the fastening point to obtain the plastic deformation. If sigma_v is not greater than sigma_s, the safety margin value of each fastening point is determined by the ratio of sigma_s to sigma_v. When the ratio is greater than or equal to 1.5, it is considered safe; when it is between 1.0 and 1.5, it is considered that a warning is needed.

[0065] After receiving the target fastening force parameter, the system directly calls the stress calculation model based on elasticity mechanics, which has been constructed in step S107, to quantify the three-dimensional stress state of the first assembly component at each fastening point. Taking the front flange of the machine tool spindle box as the first assembly component, the system reads the target fastening force parameter of fastening point 7 as 12500.5 Newtons, and writes it into the nodal force vector of the model as a boundary load condition. The material constitutive parameters directly use the elastic modulus and Poisson's ratio already loaded in S107, without repeating the setting, to ensure that the parameter source is singular. In terms of the calculation process, the system calls the finite element method to assemble the global stiffness matrix containing fastening point 7 and its local subdomain, solves the linear algebraic equation system composed of nodal force vectors and stiffness matrix, and obtains the three-dimensional displacement vector of each node in the subdomain; then, the geometric equations are used to complete the conversion from displacement to strain, and the complete three-dimensional strain tensor is output. The system extracts the normal strain component along the principal axis as a microstrain of 755.2 μs. Then, using the generalized Hooke's law under the three-dimensional constitutive relation, it simultaneously calculates the six independent components of the stress tensor and extracts the normal stress component corresponding to the axial direction. Through the above matrix operations and mechanical transformations, the system finally determines the actual stress scalar value of the seventh fastening point along the principal axis as 158.6 MPa, and stores this scalar value in the database for subsequent fatigue life prediction.

[0066] S108. Obtain the actual stress scalar value, classify the actual stress scalar value according to the preset stress grading interval standard divided by stress level, and map the buffer structure parameter combination data corresponding to each fastening point according to the stress level, including the material properties and thickness dimension of the buffer component.

[0067] The actual stress scalar value is obtained and compared with a preset stress grading interval standard to classify it into three stress levels: light load, medium load, and heavy load. The stress level corresponding to each fastening point is then determined. Based on the stress level, the material grade and thickness of the buffer component corresponding to each fastening point are extracted from the fastening point configuration library. These two parameters are concatenated into a string-like buffer structure parameter combination, with the material grade first and the thickness second, serving as a unique index key. Using this buffer structure parameter combination as the key, a preset buffer component attribute mapping table is queried. This mapping table was entered after compression tests on buffer components with different material grades and thickness combinations. The corresponding stiffness coefficient k and deformation limit d are output, where k reflects the load-bearing capacity per unit deformation, and d reflects the maximum compressibility. The three-dimensional input feature vector is composed of k, d, and the actual stress scalar value obtained above, and fed into a pre-trained linear regression model for prediction. This model uses k, d, and stress values ​​from historical assembly data as independent variables and the measured deformation compensation amount as the dependent variable. The least squares method is used to complete parameter fitting, and the compensation margin c corresponding to each fastening point is output. A preset attenuation ratio r is obtained, with a value ranging from 0.6 to 0.9. It is determined whether the compensation margin c is greater than a preset margin threshold of 0.15 mm. If c is greater than the threshold, c*r is used as the target compensation value c1; otherwise, c1 is directly set to c. Finally, the pre-tightening torque T of each fastening point is calculated according to the formula T=F0+K1*c1, where F0 is the reference torque corresponding to the target fastening force obtained above, and K1 is the conversion coefficient from deformation of the buffer component to torque, which is given along with k and d in the mapping table. Thus, the buffer compensation amount is superimposed on the original fastening force to obtain the final applied pre-tightening torque.

[0068] In one implementation, the system first performs the operation of acquiring the actual stress scalar value. This actual stress scalar value characterizes the internal mechanical response intensity borne by the front flange of the machine tool spindle box at a specific fastening point, and serves as the basic data input for subsequent structural protection design and assembly parameter adjustment.

[0069] The system has a preset stress grading interval standard, determined based on the metal material parameters of the front flange of the spindle box. These material parameters include yield strength σs, tensile strength σb, and lattice slip critical stress τc. The stress grading intervals are divided according to the stress range of the flange from zero load to plastic deformation, with 5 to 8 intervals, each continuous and with non-overlapping boundaries. Taking 45# steel as an example, when σs is 355MPa, σb is 600MPa, and τc is 180MPa, the grading intervals are successively 0 to 71MPa, 71 to 142MPa, 142 to 213MPa, 213 to 284MPa, 284 to 355MPa, 355 to 426MPa, 426 to 497MPa, and 497 to 600MPa.

[0070] For example, the system defines the stress range of 0 to 50 MPa as the first-level low-stress range, the range of 50 to 120 MPa as the second-level medium-stress range, and the range above 120 MPa as the third-level high-stress range. This grading mechanism discretizes continuous mechanical variables into grade labels with clear engineering guidance significance. Further, the system grades the actual stress scalar values ​​according to the aforementioned stress grading range standards. In this process, the system uses numerical comparison logic to compare the read actual stress scalar values ​​sequentially with the upper and lower thresholds of each grading range. When the scalar value is determined to fall into a specific range, the system assigns a stress level label to the corresponding fastening point. This process realizes the conversion from continuous physical quantities to discrete state quantities, providing a standardized index for subsequent parameter matching. Based on the results of the above grading process, the system then enters the parameter matching stage, mapping the stress level to obtain the buffer structure parameter combination data corresponding to each fastening point, including the buffer material properties and thickness dimension.

[0071] The material properties of buffer components characterize the physicochemical properties of gaskets or isolation rings that absorb and disperse localized stress concentration at fastening points. In the assembly environment of the machine tool spindle box front flange, buffer components are subjected to high-frequency vibrations ranging from 800Hz to 2500Hz and come into contact with emulsified cutting fluid containing extreme pressure additives. Therefore, the material properties encompass three key indicators: Shore hardness A, dynamic damping coefficient tanδ, and oil corrosion resistance grade. The selectable range for Shore hardness A is 60 to 95, the selectable range for dynamic damping coefficient tanδ is 0.05 to 0.35, and the oil corrosion resistance grade is classified into levels 1 to 5 according to the NBR standard. The system internally maintains a material property mapping matrix M. The row index corresponds to the stress level L, and the column index corresponds to the three material indicators mentioned above. When looking up a table, the stress level L of the fastening point is used as input, and the combination of the three parameters in that row is directly read as the output. For the first-level low-stress range, i.e., fastening points with equivalent stress below 45MPa, the matrix outputs a conventional friction-reducing formula based on polytetrafluoroethylene, with a Shore A hardness of 65, tanδ of 0.08, and an oil corrosion resistance grade of 2. For the third-level high-stress range, i.e., fastening points with equivalent stress above 120MPa, the matrix outputs a composite formula of high-molecular polyurethane elastomer and aramid fiber reinforced mesh, with a Shore A hardness of 90, tanδ of 0.28, and an oil corrosion resistance grade of 4, utilizing its high damping characteristics to dissipate excessive local strain energy.

[0072] In this embodiment, the thickness dimension refers to the geometric dimension parameter of the buffer component in the fastening axis, ranging from 5 mm to 30 mm. The thickness setting directly affects the compression stroke and stiffness performance of the buffer component. The system uses a nonlinear interpolation algorithm to determine the thickness dimension, and makes fine adjustments based on the relative position of the actual stress scalar value within the same stress level. Specifically, when the stress level increases from level 1 to level 5, the mapped reference thickness gradually increases from 8 mm to 25 mm to provide a longer deformation buffer distance and prevent microcracks from forming in the flange metal substrate due to rigid contact. The thickness value h is determined by the formula h = h0 + k × Δσ, where h0 is the reference thickness for that stress level, k is an adjustment coefficient ranging from 0.02 to 0.08, and Δσ is the stress deviation value.

[0073] For example, taking the seventh fastening point of the front flange of the machine tool spindle box as an example, assume that the system initially obtained an actual stress scalar value of 135.4 MPa at this point. During the classification process, the system determines that this value is greater than the threshold of 120 MPa, thus classifying it as a level three high-stress interval. Subsequently, the system triggers a mapping mechanism to retrieve and extract the formulation parameters of the polymer polyurethane composite material in the material property dimension, setting its Shore hardness to 85A and its dynamic damping coefficient to 0.45. Simultaneously, in the thickness dimension, based on the position of 135.4 MPa within the level three interval, the system calculates the required axial thickness of the buffer component to be 2.5 mm. Finally, the system structurally encapsulates these discrete parameters to generate buffer structure parameter combination data for the seventh fastening point. This combination data precisely defines the material and dimensions of the buffer component to be installed at the seventh fastening point, thereby achieving targeted mitigation of localized high stress at this location at the physical level.

[0074] The mapping process described above relies on a multi-dimensional relational mapping table pre-stored in the system's underlying database. This mapping table is generated based on over 3000 sets of bench fatigue test data, covering stress conditions ranging from 50MPa to 450MPa. The least squares method is used to fit and establish the correspondence between stress level σ and buffer structure parameters K, E, and H, where K is the stiffness coefficient, E is the elastic modulus, and H is the structural height. The mapping table achieves a prediction accuracy of over 92%, ensuring that the system can output assembly guidance data that meets actual working conditions within 0.5 seconds, including key parameters such as buffer component type, installation position, and preload torque.

[0075] S109. Generate a set of assembly execution instructions for driving the assembly equipment to perform fastening and buffer installation actions based on the buffer structure parameter combination data and the target fastening force parameter.

[0076] Obtain the combined data of the buffer structure parameters and the target fastening force parameters. Extract the material hardness identifier from the combined data of the buffer structure parameters and the torque scalar value from the target fastening force parameters. Based on the material hardness identifier and the torque scalar value, query a preset assembly action matching table to obtain an initial assembly action sequence. This assembly action matching table contains three fields: a hardness range column, a torque range column, and an action sequence column. The hardness range column records the numerical range of material hardness, the torque range column records the numerical range of torque, and the action sequence column records the corresponding assembly action identifier code sequence. During matching, the corresponding action sequence is determined based on the range in which the material hardness identifier and the torque scalar value fall. Obtain the execution timestamp of each action node in the initial assembly action sequence, and calculate the difference between adjacent timestamps to obtain the action time interval sequence. The system determines whether the maximum time interval in the action time interval sequence is greater than a threshold of 3.5 seconds. If the maximum time interval is greater than the threshold, the initial assembly action sequence is split at the position corresponding to the maximum time interval to obtain a first assembly action sub-sequence and a second assembly action sub-sequence. The first assembly action sub-sequence is assigned to the first assembly station for execution, and the second assembly action sub-sequence is assigned to the second assembly station for execution. A preset equipment control protocol code is obtained. The equipment control protocol code uses a hexadecimal encoding format and includes three parts: equipment address code, function code, and data code. Each action identifier code in the first and second assembly action sub-sequences is mapped and encoded according to the equipment control protocol code, converting the action identifier code into a corresponding combination of function code and data code to obtain an assembly execution instruction set. The assembly equipment is driven to perform fastening actions and buffer installation actions through the assembly execution instruction set. A laser displacement sensor installed at the execution end of the assembly equipment collects real-time displacement data of the assembly equipment during the fastening actions and buffer installation actions at a sampling frequency of 200 Hz. The actual installation travel value is calculated based on the difference between the starting position coordinate value and the ending position coordinate value in the real-time displacement data. It is then determined whether the actual installation travel value is within the error range of ±0.5 mm of the preset travel standard value. If the actual installation travel value exceeds the error range, a shutdown alarm signal is generated.

[0077] After acquiring the combination data of the buffer structure parameters and the target tightening force parameters, the system immediately starts the instruction compilation engine to generate a set of assembly execution instructions. Taking the eighth tightening point as an example, the system first parses the thickness dimension value in the buffer structure parameters and substitutes it into the spatial coordinate calculation process of the robotic arm. Assuming that the thickness of the buffer component at the eighth tightening point is 2.5 mm and the absolute vertical coordinate of the flange reference plane is 150.0 mm, the system calculates the target vertical coordinate of the robotic arm pressing down as 152.5 mm through coordinate addition, and generates spatial motion control instructions containing three-dimensional coordinates and attitude quaternions accordingly. Subsequently, the system extracts the target tightening force parameters, such as the preset final tightening torque value of 78.5 Nm, and uses the inverse kinematics algorithm to convert it into the drive current threshold of the servo tightening shaft. The system divides the torque value T by the product of the servo motor torque constant Kt and the gear ratio R, where T is 78.5 Nm, Kt is 1.5, and R is 20, and calculates the target peak current command as 2.62 Amps. To prevent overshoot during tightening, the system further introduces a piecewise linear torque interpolation algorithm, expanding the single current threshold into a piecewise torque control curve command that includes an initial contact current of 0.5 amps, a constant-speed screwing-in current of 1.2 amps, and a final locking current of 2.62 amps. The system uses a linear ramp transition between the current values ​​at each stage to achieve smooth loading. Finally, the system serializes and packages the above spatial motion control commands and piecewise torque control curve commands according to the data frame format of the industrial real-time communication protocol. The device address code of the eighth tightening point is added to the header of the data frame, and a cyclic redundancy check code is added to the tail, thereby constructing a set of assembly execution commands that can directly drive a six-axis assembly robot and an end-effector servo tightening gun to perform precise installation and tightening actions.

Claims

1. A method for intelligent assembly control of a transparent panel and a metal frame, characterized in that, include: A set of three-dimensional point cloud data of surface morphology and dimensional deviation of the first assembly component and the second assembly component at multiple measurement positions is obtained. The feature of the three-dimensional point cloud data of surface morphology and dimensional deviation is extracted by the error prediction model based on convolutional neural network to obtain the cumulative predicted value of assembly error. Based on the cumulative predicted value of assembly errors and the pre-stored design reference position data of component mating surfaces, the relative positional offset of the first assembly component and the second assembly component at each fastening point is determined. The relative positional offset is converted into torque using a pre-established fastening force mapping model to determine the initial fastening force parameter corresponding to each fastening point. The initial fastening force parameter drives a virtual assembly simulation model to perform a virtual simulation of the assembly loading process, obtaining the internal stress tensor distribution data of the first assembly component after being subjected to force in each region. For the internal stress tensor distribution data, the stress field distribution characteristics around each fastening point are extracted to obtain a set of stress distribution feature vectors corresponding to each fastening point. It is determined whether the maximum principal stress component of any fastening point in the set of stress distribution feature vectors exceeds a preset stress concentration safety threshold. If the maximum principal stress component exceeds the preset stress concentration safety threshold, the initial fastening force parameter is iteratively adjusted based on a stress balance optimization strategy to obtain a target fastening force parameter. Based on the target fastening force parameter, the actual stress scalar value of the first assembly component at each fastening point is determined using a stress calculation model based on elasticity mechanics. The actual stress scalar value is graded according to the preset stress grading interval standard, and the buffer structure parameter combination data corresponding to each fastening point is mapped to obtain the buffer structure parameter combination data. Based on the buffer structure parameter combination data and the target fastening force parameter, an assembly execution command set is generated to drive the assembly equipment to perform fastening and buffer component installation actions.

2. The intelligent assembly control method for the transparent panel and metal frame as described in claim 1, characterized in that, The step of extracting features from the three-dimensional point cloud data set of surface morphology and dimensional deviation using a convolutional neural network-based error prediction model to obtain the cumulative predicted value of assembly error includes: constructing a multi-dimensional spatial deviation matrix based on the spatial coordinate information and deviation information contained in the three-dimensional point cloud data set of surface morphology and dimensional deviation; using a convolutional neural network to extract local features from the multi-dimensional spatial deviation matrix to obtain an initial feature tensor containing local morphology features; extracting significant feature values ​​from the initial feature tensor by performing a max pooling operation to obtain a dimensionality-reduced global feature vector; if the feature value in the global feature vector is greater than a preset interference threshold, determining that there is contact surface interference and extracting the interference region features from the global feature vector; determining a multi-measurement point deviation trend sequence based on the interference region features; and using a fully connected layer to perform a nonlinear mapping on the multi-measurement point deviation trend sequence to integrate the multi-measurement point deviation superposition effect in the multi-measurement point deviation trend sequence to obtain the cumulative predicted value of assembly error.

3. The intelligent assembly control method for the transparent panel and metal frame as described in claim 1, characterized in that, The step of determining the relative positional offset of the first assembly component and the second assembly component at each fastening point based on the cumulative predicted value of assembly error and the pre-stored component mating surface design reference position data includes: obtaining the relative positional offset of the first assembly component and the second assembly component at each fastening point along the assembly axial and radial directions; constructing a fastening point spatial offset matrix based on the relative positional offset; using a support vector regression model to extract features from the fastening point spatial offset matrix to obtain the initial torque compensation coefficient corresponding to each fastening point; if the initial torque compensation coefficient is greater than a preset compensation threshold, determining that the corresponding fastening point has an overload risk and extracting the local force feature vector corresponding to the fastening point with overload risk; calculating the stress concentration index of the contact surface between the first assembly component and the second assembly component at the corresponding fastening point based on the local force feature vector; obtaining the target fastening torque adjustment value by linearly scaling the contact surface stress concentration index; and determining the final assembly execution torque sequence for each fastening point based on the target fastening torque adjustment value and the preset standard fastening torque data.

4. The intelligent assembly control method for the transparent panel and metal frame as described in claim 1, characterized in that, The step of converting the relative position offset into torque using a pre-established fastening force mapping model to determine the initial fastening force parameters for each fastening point includes: analyzing the relative position offset using a pre-established deformation analytical model containing a stiffness damping matrix to obtain elastic deformation parameters; constructing a stress-friction calculation function based on the elastic deformation parameters and a preset contact surface roughness; calculating the torsional shear stress and thread friction coefficient during assembly using the stress-friction calculation function; determining the initial fastening force based on the torsional shear stress and the thread friction coefficient; if the initial fastening force is greater than a preset yield strength threshold, determining that there is a risk of plastic deformation and calculating the corresponding preload attenuation rate based on the difference between the initial fastening force and the yield strength threshold; obtaining a dynamic compensation torque by multiplying the preload attenuation rate and a preset benchmark compensation coefficient; and correcting the initial fastening force based on the dynamic compensation torque to determine the final fastening force as the initial fastening force parameter.

5. The intelligent assembly control method for the transparent panel and metal frame as described in any one of claims 1-4, characterized in that, The virtual simulation of the assembly loading process driven by the initial fastening force parameters to obtain the internal stress tensor distribution data of the first assembly component after being subjected to force in various regions includes: extracting the principal stress and shear stress of each grid point in the first assembly component based on the internal stress tensor distribution data; obtaining the principal stress feature vector by performing dimensionality reduction processing on the principal stress and shear stress using a principal component analysis algorithm; calculating the spatial gradient value of the principal stress feature vector between adjacent grid points to obtain a stress gradient distribution matrix; if the elements in the stress gradient distribution matrix are greater than a preset stress concentration threshold, determining that the corresponding grid point is in a stress concentration state and extracting the set of grid points in the stress concentration state as the target stress area; extracting the nodal forces of each grid point in the target stress area; performing cyclic loading calculation on the nodal forces using a pre-established fatigue damage accumulation model to determine the fatigue life prediction value of the target stress area; if the fatigue life prediction value is less than a preset life safety threshold, determining that there is a risk of fatigue fracture and outputting the coordinates of the structural weak point corresponding to the target stress area.

6. The intelligent assembly control method for a transparent panel and a metal frame as described in any one of claims 1-4, characterized in that, The step of iteratively adjusting the initial fastening force parameters based on the stress equilibrium optimization strategy to obtain the target fastening force parameters includes: extracting the initial fastening force parameters of the fastening point and the adjacent force parameters of adjacent fastening points to obtain a force parameter set; calculating the force difference value between the fastening point and the adjacent fastening points based on the force parameter set; obtaining the spatial positional relationship data between the fastening point and the adjacent fastening points; constructing a local stiffness matrix based on the spatial positional relationship data; constructing a displacement solution function with the objective of minimizing deformation potential energy using the local stiffness matrix and the force difference value; calculating the deformation displacement of the fastening point by solving the displacement solution function using the gradient descent algorithm; calculating the residual stress distribution data based on the product relationship between the deformation displacement and the local stiffness matrix; constructing a stress equilibrium objective function based on the residual stress distribution data and the force difference value; and iteratively adjusting the initial fastening force parameters based on the stress equilibrium objective function until the residual stress distribution data meets the preset convergence condition to obtain the target fastening force parameters.

7. The intelligent assembly control method for a transparent panel and a metal frame as described in any one of claims 1-4, characterized in that, The step of classifying the actual stress scalar value according to a preset stress grading interval standard to map the buffer structure parameter combination data corresponding to each fastening point includes: comparing the actual stress scalar value according to the preset stress grading interval standard to determine the stress level corresponding to each fastening point; extracting the buffer material properties and thickness dimension corresponding to each fastening point according to the stress level; splicing the buffer material properties and thickness dimension to obtain the buffer structure parameter combination data; querying a preset buffer property mapping table according to the buffer structure parameter combination data to obtain the stiffness coefficient and deformation limit corresponding to the buffer structure parameter combination data; using the stiffness coefficient and deformation limit as input features to a pre-trained linear regression model for prediction to obtain the compensation margin corresponding to each fastening point; determining whether the compensation margin is greater than a preset margin threshold; if the compensation margin is greater than the margin threshold, determining the product of the compensation margin and a preset attenuation ratio as the target compensation value; and calculating the pre-tightening torque of each fastening point according to the target compensation value.

8. The intelligent assembly control method for a transparent panel and a metal frame as described in any one of claims 1-4, characterized in that, The step of generating an assembly execution instruction set for driving the assembly equipment to perform fastening and buffer installation actions based on the buffer structure parameter combination data and the target fastening force parameter includes: extracting the material hardness identifier from the buffer structure parameter combination data and the torque scalar value from the target fastening force parameter; querying a preset assembly action matching table based on the material hardness identifier and the torque scalar value to obtain an initial assembly action sequence; parsing the execution timestamps of each action node in the initial assembly action sequence using a pre-trained decision tree model to obtain an action time interval sequence; if the maximum time interval in the action time interval sequence is greater than a preset interval threshold, splitting the initial assembly action sequence to obtain a first assembly action sub-sequence and a second assembly action sub-sequence; encoding the first assembly action sub-sequence and the second assembly action sub-sequence according to a preset equipment control protocol code to obtain an assembly execution instruction set; driving the assembly equipment to perform fastening and buffer installation actions using the assembly execution instruction set and collecting real-time displacement data of the assembly equipment when performing the fastening and buffer installation actions; calculating the actual installation stroke value based on the real-time displacement data; and generating a shutdown alarm signal if the actual installation stroke value does not match a preset stroke standard value.

9. A computer case, characterized in that, The computer chassis is manufactured by the intelligent assembly control method of transparent panel and metal frame as described in any one of claims 1-8, including metal frame and transparent panel, the transparent panel is installed on metal frame through multiple fastening points, and a buffer structure is provided at the mating surface of transparent panel and metal frame.

Images