Multi-parameter cooperative control method and device for clamping system of precision boring and milling machine

Abstract

The invention relates to the technical field of automatic control, in particular to a multi-parameter cooperative control method and device for a clamping system of a precision boring and milling machine. According to the method, an industrial camera is used for collecting and processing a workpiece image and extracting visual features, and the visual features are fused with process parameters of a workpiece and geometric distribution parameters of a workpiece key area obtained from a process database; generating a comprehensive state vector; based on the vector, outputting a target clamping force adjustment coefficient by a pre-trained MLP decision model, and calculating a final target clamping force in combination with a safe clamping force range; in the main clamping stage, vibration, temperature and pressure signals are synchronously collected, multi-modal disturbance characteristics are extracted, observation vectors are constructed, the observation vectors are input into a preset fuzzy rule base for reasoning, and compensation activation factors are obtained; and finally, a current correction instruction is dynamically generated according to the clamping force deviation and the compensation factor, and an electro-hydraulic proportional valve is driven to achieve accurate compensation. According to the invention, the stability and the machining reliability of the clamping system are obviously improved.

Description

Technical Field

[0001] The present application relates to the field of automation control technology, and in particular to a multi-parameter collaborative control method and device for a clamping system of a precision boring and milling machine. Background Art

[0002] Precision boring and milling machines are widely used to machine high-hardness, high-value workpieces such as mold steel and titanium alloys. These machines generate enormous cutting forces and demand micron-level machining precision. Any micron-level workpiece slippage due to clamping failure can easily lead to catastrophic tool collisions and workpiece failure, resulting in significant economic losses. Current clamping control systems face several technical bottlenecks: 1) Clamping parameters are often set based on operator experience and lack a quantitative basis, resulting in poor clamping quality consistency and a frequent swing between "insufficient clamping force causing workpiece displacement" and "excessive clamping causing workpiece deformation." 2) Existing clamping systems generally employ single-variable feedback control, monitoring only clamping force or position deviation. These systems are unable to fully perceive multi-dimensional state information such as vibration and thermal deformation. This leads to a lag in response to sudden changes in cutting force or external disturbances, resulting in incomplete system perception and slow response, resulting in insufficient control accuracy. 3) Traditional control strategies are rigid and difficult to adapt to real-world variations such as geometric deviations between workpiece batches, fixture wear, or hydraulic fluctuations. Adapting to new, difficult-to-machine materials relies on trial-and-error procedures, significantly increasing machining costs and production lead times, and resulting in low overall equipment utilization. These deficiencies are the root cause of the current unstable quality, high scrap rates, and limited productivity gains in precision boring and milling machine processing. Therefore, a clamping control method that integrates multi-source sensing, intelligent decision-making, and dynamic collaboration is urgently needed to achieve precise and adaptive adjustment of clamping force and improve the stability and reliability of precision machining. Summary of the Invention

[0003] The present application provides a multi-parameter collaborative control method and device for the clamping system of a precision boring and milling machine, which solves the technical problems in the prior art caused by the technical bottleneck of the clamping control system, resulting in unstable processing quality, high scrap rate and low production efficiency of the precision boring and milling machine.

[0004] In view of the above problems, the present application provides a multi-parameter collaborative control method for a clamping system of a precision boring and milling machine, the method comprising: The workpiece is pre-clamped and fixed on the fixture. The workpiece surface image is captured by an industrial camera. The captured image is processed by SIFT feature matching algorithm and illumination normalization to form a composite image for feature extraction. Based on the composite image, the visual features of the workpiece surface structure and texture distribution are extracted through a pre-trained deep convolutional network, and the process parameters of the workpiece are obtained from the process database. At the same time, the geometric distribution parameters of the pre-calibrated key areas of the workpiece in the image are detected. The above visual features, process parameters and geometric distribution parameters are normalized and weighted to generate a comprehensive state vector for clamping decision-making. The process parameters include at least: material mechanical parameters and cutting process parameters; The integrated state vector is input into a pre-trained MLP decision model, which outputs the target clamping force. The MLP model learns its decision-making behavior from a clamping decision strategy network trained using the DDPG algorithm through knowledge distillation. The main clamping phase is the phase in which the clamping system applies and maintains the target clamping force during the stable machining process of the boring and milling machine. Based on the target clamping force obtained, the controller generates a corresponding target current command according to the preset pressure-current mapping relationship. This command drives the electro-hydraulic proportional valve to adjust the output force of the hydraulic cylinder, driving the clamping mechanism to complete the workpiece clamping. During the main clamping stage, the vibration signal, temperature signal and hydraulic oil pressure signal of the clamping system are collected simultaneously. The time domain features of the vibration signal and temperature signal are extracted in units of a preset time window to obtain the vibration time domain features and temperature time domain features. The average value of the hydraulic oil pressure in the window is calculated. Based on the known effective area of ​​the piston, the hydraulic oil pressure is converted into the current output force of the hydraulic cylinder. The current output force is input into the calibrated force transfer model to obtain the total effective clamping force on the workpiece, and the clamping force deviation between the target clamping force and the total effective clamping force is calculated. The extracted vibration time domain features and temperature time domain features are combined with the clamping force deviation to construct a normalized current multimodal disturbance observation vector in the form of [vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, clamping force deviation]; The current multimodal disturbance observation vector is input into the pre-built fuzzy rule model. After fuzzification, rule activation, implication operation and maximum aggregation, a synthetic fuzzy set of hydraulic compensation factors is generated. The center of gravity method is used for defuzzification within the domain of hydraulic compensation factors [0, 1]. The geometric center of the synthetic fuzzy set is calculated and the calculated result is used as the output value of the hydraulic compensation factor. The fuzzy rule model is constructed by a fuzzy decision tree. According to the clamping force deviation between the target clamping force and the total effective clamping force and the output value of the hydraulic compensation factor, the current correction value of the electro-hydraulic proportional valve is calculated and superimposed on the target current instruction. Together, they serve as the control signal of the electro-hydraulic proportional valve, thereby dynamically adjusting the valve core opening to generate the required hydraulic oil pressure and maintain a stable clamping force of the clamping system during processing. , in, is the current correction value, is the clamping force deviation, is the output value of the hydraulic compensation factor, For electricity The slope of the pressure-current characteristic curve of the hydraulic proportional valve.

[0005] The present application also provides a multi-parameter coordinated control device for a clamping system of a precision boring and milling machine, the device comprising: Composite image module, which is used to pre-clamp the workpiece on the fixture and collect the workpiece surface image through the industrial camera. The collected image is processed by SIFT feature matching algorithm and illumination normalization to form a composite image for feature extraction; A state vector module is used to extract visual features of the workpiece surface structure and texture distribution based on the composite image through a pre-trained deep convolutional network, obtain the workpiece process parameters from the process database, and simultaneously detect the geometric distribution parameters of the pre-calibrated key areas of the workpiece in the image. The module normalizes and weights the above visual features, process parameters, and geometric distribution parameters to generate a comprehensive state vector for clamping decision-making, where the process parameters include at least: material mechanical parameters and cutting process parameters; The MLP model module is used to input the comprehensive state vector into the pre-trained MLP decision model and output the target clamping force. The MLP model learns its decision-making behavior from the clamping decision strategy network trained based on the DDPG algorithm through knowledge distillation. The main clamping phase is the working stage during which the clamping system applies and maintains the target clamping force during the stable machining process of the boring and milling machine; The optimal clamping module is used to generate a corresponding target current command based on the obtained target clamping force and the preset pressure-current mapping relationship. This command drives the electro-hydraulic proportional valve to adjust the output force of the hydraulic cylinder, driving the clamping mechanism to complete the workpiece clamping. The disturbance information acquisition module is used to simultaneously collect the vibration signal, temperature signal and hydraulic oil pressure signal of the clamping system during the main clamping phase. The module extracts the time domain features of the vibration signal and temperature signal in units of a preset time window to obtain the vibration time domain features and temperature time domain features, and calculates the average value of the hydraulic oil pressure within the window. Based on the known effective area of ​​the piston, the hydraulic oil pressure is converted into the current output force of the hydraulic cylinder. The current output force is input into the calibrated force transfer model to obtain the total effective clamping force on the workpiece, and the clamping force deviation between the target clamping force and the total effective clamping force is calculated; Current disturbance observation vector module, which is used to combine the extracted vibration time domain features and temperature time domain features with the clamping force deviation to construct a normalized current multimodal disturbance observation vector in the form of [vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, clamping force deviation]; The hydraulic compensation factor module is used to input the current multimodal disturbance observation vector into the pre-built fuzzy rule model, generate a synthetic fuzzy set of the hydraulic compensation factor through fuzzification, rule activation, implication operation and maximum aggregation, and use the centroid method to perform defuzzification within the hydraulic compensation factor domain [0, 1]. The geometric center of the synthetic fuzzy set is calculated and the calculated result is used as the output value of the hydraulic compensation factor. The fuzzy rule model is constructed by a fuzzy decision tree. The compensation force module is used to calculate the current correction value of the electro-hydraulic proportional valve based on the clamping force deviation between the target clamping force and the total effective clamping force and the output value of the hydraulic compensation factor. The current correction value is added to the target current instruction and used as the control signal of the electro-hydraulic proportional valve to dynamically adjust the valve core opening, generate the required hydraulic oil pressure, and maintain a stable clamping force of the clamping system during processing. , in, is the current correction value, is the clamping force deviation, is the output value of the hydraulic compensation factor, For electricity The slope of the pressure-current characteristic curve of the hydraulic proportional valve.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application provides a multi-parameter collaborative control method and device for the clamping system of a precision boring and milling machine, using an industrial camera in real time Images of the workpiece after clamping are captured, stitched, and illuminated to produce a surface view of the workpiece. The camera captures the current, actual clamping state. Random factors that may vary with each clamping and cannot be retrieved from a fixed database directly affect the distribution and effectiveness of the clamping force. This step ensures a real-time perception foundation for subsequent intelligent decision-making. Integrating a comprehensive state vector enables the decision-making system to "see" the entire scene and make more global and forward-looking decisions. Using the MLP model, the target clamping force is output based on the comprehensive state vector, transforming the complex decision-making problem into a simple, normalized strategy selection problem. The target clamping force is converted into a current command to drive the hydraulic system to perform clamping, completing the closed loop from decision-making signal to physical execution. Multimodal disturbance monitoring and feature extraction enable real-time diagnosis of the machining process, capturing dynamic disturbances that are unexpected during the intelligent stage of the MLP model and occur suddenly during machining. Fuzzy rule-based real-time compensation decisions are made. The fuzzy rule base acts as a rapid reflex arc, responding to sudden disturbances in milliseconds, determining whether compensation is necessary and the strength of the compensation intention. The compensation intention u is multiplied by the force deviation ΔF to obtain the final compensation current value. Here, the fuzzy intelligent system modulates the output of PID according to the global state, avoiding the problems of "sluggishness" or "overshoot" of traditional PID under complex disturbances.

[0007] In summary, the present invention fundamentally improves the clamping quality and reliability of the clamping system of the precision boring and milling machine, deeply integrates multi-source information, realizes full-dimensional state perception during the operation of the clamping system, and can identify potential risks earlier rather than passively waiting for deviations to occur; adopts an offline / online two-level decision-making mechanism to improve the response speed to disturbances, shorten the production preparation cycle, and improve production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A flow chart of a multi-parameter collaborative control method for a clamping system of a precision boring and milling machine provided in an embodiment of the present application.

[0009] Figure 2 A schematic structural diagram of a multi-parameter collaborative control device for a precision boring and milling machine clamping system provided in an embodiment of the present application.

[0010] Explanation of the reference numerals: composite image module 10 , state vector module 20 , MLP model module 30 , optimal clamping module 40 , disturbance information acquisition module 50 , current disturbance observation vector module 60 , hydraulic compensation factor module 70 , compensation force module 80 . DETAILED DESCRIPTION

[0011] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0012] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0013] The embodiment of the present application provides a multi-parameter collaborative control method for a clamping system of a precision boring and milling machine. Figure 1 As shown, the method includes: S100: The workpiece is pre-clamped and fixed on the fixture, and the surface image of the workpiece is collected by an industrial camera. The collected image is processed by the SIFT feature matching algorithm and illumination normalization to form a composite image for feature extraction.

[0014] Specifically, this embodiment captures images after the workpiece is pre-clamped. Two industrial cameras (optionally 5 million pixels, resolution 2448×2048) are installed at a fixed distance from the worktable reference plane, such as 300±5mm. The two cameras are arranged above the workpiece clamping area, and the mechanical coordinate system of the workpiece clamping area (such as the worktable coordinate system) is used as a reference to ensure the consistency of image acquisition with the actual processing position. The optical axis of the first camera is at a 45° angle to the Z axis in the XZ plane, which can fully capture the side features of the workpiece in the positive X direction. The optical axis of the second camera is at a 45° angle to the Z axis in the YZ plane, which can fully capture the side features of the workpiece in the positive Y direction. The intersection of the optical axes of the two cameras is aligned with the geometric center of the workpiece clamping area (such as the workpiece center or the fixture center) to ensure that the coverage of the multi-view images overlaps and the field of view completely covers the maximum clamping size of the workpiece and the contact surface of the fixture. After the workpiece is temporarily fixed to the workbench with a pre-tightening force of 10-20N by a pneumatic clamp and stabilized, the dual cameras are triggered synchronously to capture multi-view images of the workpiece surface.

[0015] Furthermore, a composite image for feature extraction is formed, including: Two industrial cameras are placed above the workpiece clamping area. Using the mechanical coordinate system of the workpiece clamping area as a reference, the optical axis of the first camera is located in the XZ plane at a 45° angle to the Z axis, and the optical axis of the second camera is located in the YZ plane at a 45° angle to the Z axis. The intersection of the optical axes of the two cameras is aligned with the geometric center of the workpiece clamping area to obtain multi-angle images of the workpiece surface. The workpiece surface image captured by the first camera is image A, and the workpiece surface image captured by the second camera is image B. The key points of image A and image B are extracted using the SIFT algorithm and initially matched. The RANSAC algorithm is then combined with the homography model to estimate the optimal homography matrix from the initial matching point pairs. Mismatched points are removed at the same time, and the matrix is ​​used to perform perspective transformation on the images to obtain a spliced ​​composite image. The spliced ​​composite image is subjected to illumination normalization processing and converted from RGB color space to HSV color space. The hue, saturation and lightness channels are separated. The CLAHE algorithm is applied to the lightness channel to suppress the influence of uneven lighting. The highlight area is further compressed pixel by pixel through nonlinear mapping method to reduce the interference of mirror reflection on the metal surface of the workpiece. The processed lightness channel is merged with the original hue and saturation channels and converted back to RGB color space to obtain a normalized composite image for subsequent feature extraction.

[0016] Specifically, in this implementation, two industrial cameras are placed above the workpiece clamping area. When the workpiece is pre-clamped on the workbench, the first camera captures image A of the workpiece surface, and the second camera captures image B of the workpiece surface. The SIFT algorithm is used to detect key points in images A and B, that is, to locate the positions of significant features with scale invariance and rotation invariance in the image, and to generate a 128-dimensional feature vector for each detected key point. This feature vector is called a descriptor, which is used to characterize the gradient distribution characteristics of the local area around the key point, so that similar local structures in different images have similar descriptors. Subsequently, based on the nearest neighbor distance ratio method, the descriptor of each key point in image A is measured for similarity with the descriptors of all key points in image B. When the distance ratio between the nearest neighbor and the next nearest neighbor is less than the set threshold, it is determined to be a valid match, thereby obtaining a set of initial matching point pairs. Then the RANSAC algorithm is combined with the homography model. The RANSAC algorithm will repeatedly randomly select a portion of points from this set of initial matching point pairs, use these points to calculate a hypothetical homography matrix, and use this matrix to test all other matching points. According to the set threshold (such as reprojection error), it determines which points are inliers that conform to the model and which are outliers that do not conform. While eliminating the mismatched points, the optimal homography matrix is ​​determined, that is, the homography matrix with the most inliers is considered to be the optimal estimate. This method effectively eliminates the mismatched points and ensures the accuracy and robustness of the homography matrix. When using the optimal homography matrix H to perform perspective transformation on the image, with image A as the reference benchmark, the matrix needs to be applied to image B (i.e., the image to be transformed) according to the homography relationship X. B =HX AFirst, calculate the new positions of the coordinates of the top left, top right, bottom left and bottom right corners of image B after being transformed by the homography matrix H. These transformed points define a minimum bounding rectangle area, the size of which determines the size of the final output canvas, that is, the overall frame of the stitched image. For each target pixel coordinate (x', y') in the output canvas, convert it into homogeneous coordinate form , and use the inverse matrix of the homography To calculate its corresponding position (x, y) in image A, it means finding the mapping position of each pixel on the canvas in the original image through matrix operation. When (x, y) is within the range of image A, bilinear interpolation is used to estimate the pixel value at the non-integer coordinate position based on the color or grayscale value of the four nearest neighboring pixels; if (x, y) exceeds the boundary of image A, it is filled with zero value as the background. After completing the remapping of image B to the canvas, image A is superimposed on the same canvas, and the distance weight formula I is used in the overlapping area of ​​the two images. out =(1-λ)·I A +λ·I B Perform pixel fusion, where I out It represents the final color or grayscale value of a pixel position in the overlapping area of ​​the output image. A Represents the color or grayscale value of image A at the pixel position, I B represents the color or grayscale value of the same pixel in image B. λ is a weight coefficient between 0 and 1, representing the normalized distance from the current pixel to the boundary of the overlapping region. The closer a pixel is to image B, the more it tends to adopt the color value of image B; conversely, the closer it is to image A, the more it tends to adopt the color value of image A. This ensures a smooth transition between the two images in the overlapping region, resulting in a seamless composite image. For the composite image, the cvtColor function in the OpenCV library is first used to convert the spliced ​​RGB images to the HSV color space. This conversion process automatically decomposes the image into three independent channels: hue (H), saturation (S), and value (V), facilitating subsequent independent processing of the brightness component. Next, the Contrast-Limited Adaptive Histogram Equalization (CLAHE) algorithm is applied to the luminance V channel, dividing the image into 8×8 local grid areas. The grayscale histogram is calculated independently for each grid and histogram equalization is performed. At the same time, the contrast limit threshold is set to 2.0, and the histogram intervals exceeding the threshold are cropped. The cropped energy is redistributed evenly to other intervals, thereby effectively suppressing local over-exposure or under-exposure areas and improving lighting uniformity. Subsequently, gamma correction is applied to the V channel after CLAHE processing to further compress the brightness of the highlight area and reduce the impact of specular reflection on the metal surface; the correction process calculates the nonlinear mapping pixel by pixel: V out =255×(V in / 255) γ , where V in is the input pixel value, V out For output pixel values, a setting of γ = 0.8 helps reduce the dynamic range of bright areas. After processing, OpenCV's merge function is used to re-merge the corrected V channel with the original, unmodified H and S channels to form a new three-channel HSV image. This is then converted back to RGB color space using the cvtColor function to obtain a normalized composite image with uniform illumination distribution and color fidelity. This process completes the process from multi-view image acquisition, alignment and stitching, to illumination normalization, significantly improving the visual consistency and feature stability of images under different lighting and viewing angles, providing high-quality input for subsequent feature extraction.

[0017] S200: Based on the composite image, the visual features of the workpiece surface structure and texture distribution are extracted through a pre-trained deep convolutional network, and the process parameters of the workpiece are obtained from the process database. At the same time, the geometric distribution parameters of the pre-calibrated key areas of the workpiece in the image are detected, and the above visual features, process parameters and geometric distribution parameters are normalized and weightedly fused to generate a comprehensive state vector for clamping decision-making, wherein the process parameters include at least: material mechanical parameters and cutting process parameters.

[0018] Specifically, the system constructs a comprehensive state vector for clamping decisions based on the generated composite image and combined with knowledge from the process database. In this embodiment, a pre-trained deep convolutional neural network, such as a lightweight MobileNetV3 model, is first used to extract features from the composite image, outputting a set of high-dimensional vectors to characterize the structural distribution and texture characteristics of the workpiece surface. MobileNetV3 is supervised and trained on a large number of workpiece image datasets, enabling it to focus on visual patterns related to clamping stability. The pre-trained MobileNetV3 extracts the workpiece surface structure distribution, including at least: edges, pits, scratches, positioning reference surfaces (such as grinding planes, calibration grooves), and the contours of the area to be processed (such as the boundary of the milled surface). The extracted workpiece surface texture characteristics include at least: roughness, texture direction, and processing traces (residual features from the previous process). At the same time, the system retrieves material parameters related to the current workpiece (such as elastic modulus, stiffness coefficient, friction coefficient, etc.) and cutting process parameters (such as spindle speed, feed rate, cutting depth, tool type, etc.) from the process database. The system detects the geometric distribution parameters of the pre-calibrated key structural feature areas of the workpiece in the composite image. "Pre-calibration" refers to the registration and marking of functionally or structurally significant areas (such as locating holes, clamping avoidance grooves, and high-precision machined surfaces) in the image coordinate system based on the CAD model or process drawing during the workpiece design phase or system initialization. Finally, the visual feature vectors, process parameters, and geometric distribution parameters of key structural areas are normalized. Visual feature vectors can be normalized using the Z-score, process parameters can use the Min-Max or Z-score, and geometric distribution parameters can use the Min-Max standardization to eliminate dimensional differences. The preprocessed multi-source features are input into a hierarchical fusion network to generate a comprehensive state vector for clamping decision-making. First, the high-dimensional visual features are compressed and reduced in dimension through a nonlinear fully connected network to extract their compact representation; then, according to the preset weight coefficients, the reduced-dimensional visual features, material and process parameters, and geometric distribution parameters are multiplied by the corresponding fusion weight coefficients respectively, and then the weighted reduced-dimensional visual features, material and process parameters, and geometric distribution parameters are vector-spliced ​​in a predetermined order. The fusion weight reflects the relative importance of various parameters in the current processing scenario. For the high-precision processing scenario of precision boring and milling machines, the weight of the geometric distribution parameters can be increased to strengthen the guiding role of the position accuracy of the key structural area on the clamping layout. The weight coefficient distribution can be pre-set based on historical data; the splicing result is then nonlinearly transformed through a fully connected network to output a unified comprehensive state vector.

[0019] Furthermore, the geometric distribution parameters of the pre-calibrated key areas of the workpiece in the image are detected, including: Based on the CAD model of the workpiece, define the functionally significant areas as key areas, and record their theoretical positions, sizes, and contour shapes in the CAD model design coordinate system; Mapping the contour of the key area in the design coordinate system to the image coordinate system of the composite image to generate a corresponding binary template; The NCC matching algorithm is used to search for the best matching position of the template in the composite image and preliminarily locate the key areas of the image; Perform Canny edge detection on the key areas of the image to extract the contour of the workpiece on the composite image; The least squares method is used to fit the geometric transformation relationship between the designed contour and the detected contour. Based on this transformation relationship, the actual center coordinates, size and contour shape of the key area in the composite image are determined as geometric distribution parameters for subsequent state characterization.

[0020] Specifically, the workpiece's CAD model is analyzed to identify functionally significant key areas, including locating holes, clamping avoidance grooves, thin-walled areas, and high-precision machined surfaces. The location (coordinates), dimensions (length, width, diameter, etc.), and contour shape (circular, rectangular, irregular, etc.) of each key area are recorded in the workpiece's CAD model coordinate system. Based on the calibration results, the key area contours in the design coordinate system are converted to the image coordinate system, and a corresponding binary template is generated for subsequent matching. First, a projection transformation relationship is established from the design coordinate system to the image coordinate system using pre-calibrated camera intrinsic and extrinsic parameters. The camera intrinsic parameters include focal length, principal point coordinates, and distortion coefficients, while the extrinsic parameters are the rotation matrix and translation vector of the design coordinate system relative to the camera coordinate system. Next, the coordinates of the boundary vertices of the key regions in the CAD model are extracted, converted to homogeneous coordinates, and multiplied by the projection matrix for perspective transformation. This yields the initial projection coordinates of the key region outline in the image plane. These initial projection coordinates are then nonlinearly corrected based on the lens distortion model to eliminate the effects of radial and tangential distortion. Finally, the corrected outline coordinates are rasterized to generate a binary template image with the same resolution as the composite image. In this template image, pixel values ​​within the outline are set to 1, while those outside are set to 0, completing the precise mapping from the design space to the image space. This process can be implemented using the OpenCV projectPoints and fillPoly functions, ensuring that the template's geometric features remain accurate despite illumination variations and partial occlusion. After acquiring the composite image, the NCC (normalized cross-correlation) matching algorithm is used to search for the best matching position of the previously generated binary template in the actual composite image to preliminarily determine the approximate location of the key area. Canny edge detection is performed on the preliminarily located matching area to extract the true contour of the workpiece under actual imaging conditions and obtain a set of detection contour points. The design contour in the CAD model is transformed according to the preliminary positioning results so that it is in a similar position to the detection contour, and then a correspondence between the design contour points and the detection contour points is established in a local range. Based on the corresponding point pairs, the least squares method is used to solve the geometric transformation parameters that minimize the matching error, including translation, rotation angle, and scaling factor, to characterize the overall posture deviation of the key area. Furthermore, the degree of local shape deviation can be evaluated by analyzing the contour matching residual or fitting a non-rigid deformation model. Based on the detection contour point set, the actual center coordinates, size, and contour shape of the key area in the image coordinate system are directly fitted and calculated, and its relative spatial relationship with the surrounding structure is determined in combination with the transformation parameters, thereby comprehensively characterizing the geometric distribution characteristics of the workpiece in the clamped state.

[0021] S300: Input the comprehensive state vector into the pre-trained MLP decision model, output the target clamping force adjustment coefficient, retrieve the corresponding minimum and maximum process clamping force from the process database based on the workpiece material mechanical parameters and cutting process parameters in the vector, and convert the adjustment coefficient into the target clamping force through a linear mapping relationship. Among them, the MLP model learns its decision-making behavior from the clamping decision strategy network trained based on the DDPG algorithm through the knowledge distillation method. The main clamping stage is the working stage in which the clamping system applies and maintains the target clamping force during the stable processing of the boring and milling machine.

[0022] Specifically, the MLP model has been trained before the system is deployed. During the actual processing, only forward reasoning is performed, which does not affect the processing rhythm. The MLP model architecture adopts a three-layer fully connected structure. The number of nodes in the input layer is the same as the dimension of the comprehensive state vector. The hidden layer can choose to use the ReLU activation function with a negative slope coefficient of 0.01. The output layer has 1 node, corresponding to the target clamping force. In the process database of the intelligent manufacturing or CNC machining system, a process record will be established for each type of "workpiece material + processing technology" combination, including at least workpiece material, processing technology, tool type, cutting parameters, minimum safe clamping force, maximum safe clamping force, etc. During the simulation process, the corresponding [ , ] Parameter pair to ensure that the target clamping force is within a safe and effective range.

[0023] Furthermore, the MLP model learns its decision-making behavior from the clamping decision strategy network trained with the DDPG algorithm through knowledge distillation, including: A digital twin simulation environment for the clamping system of a boring and milling machine is constructed. The simulation environment includes a geometric data layer and a functional model layer. The geometric data layer includes the workpiece CAD geometric model and the fixture CAD geometric model. The functional model layer includes an assembly analysis module, a clamping mechanism dynamics model, a workpiece mechanical response model, a material contact characteristic model, a virtual visual perception module, and an intelligent clamping decision model. The assembly analysis module constructs the fixture-workpiece assembly by using the assembly constraint relationship between the workpiece CAD geometric model and the fixture CAD geometric model. It uses the geometric contact detection algorithm to identify the clamping contact points on the assembly and assign numbers to each contact point. It extracts the three-dimensional coordinates of each contact point and its surface normal vector. In combination with the preset fixture functional topology information, it assigns a force transmission path number to each contact point. The above contact point information is used to generate the structural configuration parameters of the clamping system. The dynamic model of the clamping mechanism simulates the dynamic response characteristics of the clamping system composed of a hydraulic cylinder and an electro-hydraulic proportional pressure valve in the main clamping stage. It takes the received target clamping force command as input and outputs the real-time total clamping force generated by the hydraulic cylinder piston rod. Based on the force transmission path number provided by the assembly analysis module and the finite element analysis of the fixture-workpiece assembly, the equivalent stiffness coefficient of each force transmission path is determined. The real-time total clamping force is distributed to each path according to the stiffness ratio to form the clamping force and its distribution of the real-time total clamping force. At the same time, the instantaneous power value of the clamping system is calculated based on the real-time total clamping force combined with the piston speed of the hydraulic cylinder, and time integration is performed in the main clamping stage to estimate the energy consumption of the clamping system during the main clamping process. The workpiece mechanical response model is constructed based on the clamping force distribution on the fixture-workpiece assembly. The clamping force is applied as a surface load to the corresponding contact points on the workpiece. At the same time, fixed constraints are applied in the assembly constraint relationship area. The finite element model is analyzed to obtain the stress distribution and elastic deformation field of the workpiece under the clamping force, and the maximum local deformation of the workpiece is extracted. The material contact characteristic model stores the contact point friction coefficient when the workpiece and fixture materials are paired. During the simulation process, the maximum total static friction resistance of the clamping system in the sliding direction is calculated, and the total cutting component of the estimated total cutting force in the sliding direction is obtained. The two are compared to obtain the actual safety factor, which is used to evaluate the risk of workpiece slippage. The estimated total cutting force comes from an external cutting force estimation model, which is called before each clamping decision. The cutting process parameters of the workpiece are called from the process database to estimate the total cutting force of the boring and milling machine. A virtual visual perception module is built in the digital twin simulation environment to generate visual features consistent with the actual extraction. It is used in the training phase of the clamping decision strategy to learn the ability to respond to changes in the surface state of the workpiece. The intelligent clamping decision model is built based on the DDPG algorithm. The process of DDPG training reinforcement learning policy network is as follows: S1: Randomly initialize the parameters of the Actor strategy network and the Critic evaluation network, and initialize their corresponding target networks; S2: For each time step (t), input the current integrated state vector S(t) to the Actor strategy network, which is based on The output target clamping force F(t) is compared and constrained with the minimum safe clamping force and the maximum safe clamping force allowed under the current working conditions retrieved from the process database before output, ensuring that the generated target clamping force instruction is always within the safe clamping force window. The target clamping force is converted into a target clamping force instruction and input into the digital twin simulation environment. S3: Based on this instruction, the digital twin simulation environment runs a simulation step. After the simulation is completed, it generates the next comprehensive state vector S(t+1). At the same time, based on the simulation results, it calculates the value of the multi-objective reward function and obtains the reward value R(t) of S(t). The simulation results include the maximum local deformation of the workpiece, the actual safety factor, and the energy consumption of the clamping system. S4: Take the four-tuple (S(t), F(t), R(t), S(t+1)) generated by this interaction as an experience sample and store it in the experience feedback Put the buffer, S5: Randomly sample small batches of quads from the experience replay buffer and first update the parameters of the critic evaluation network. To minimize its prediction error, the Actor parameter policy gradient is calculated based on the updated Critic network to guide the Actor policy network The direction of network parameter update is to gradually optimize the strategy to improve the long-term cumulative reward, and to synchronously update the Actor strategy network and Critic Evaluate the target network parameters of the network to improve training stability, S6: Take S(t+1) as the new current integrated state vector and repeat steps S2-S5 until the reinforcement learning policy network Converge and obtain a trained clamping decision strategy network; The MLP model is obtained by migrating the decision-making ability of the clamping decision strategy network trained and converged by the DDPG algorithm mentioned above through the knowledge distillation method.

[0024] Specifically, this embodiment constructs a digital twin simulation environment for the clamping system of a boring and milling machine to simulate the mechanical response of a workpiece under different clamping forces and support offline training of clamping decisions. The simulation environment includes a geometric data layer and a functional model layer. The geometric data layer includes a workpiece CAD geometric model and a fixture CAD geometric model. The functional model layer includes an assembly analysis module, a clamping mechanism dynamics model, a workpiece mechanical response model, a material contact characteristics model, a virtual visual perception module, and an intelligent clamping decision model. The modules and models are linked through clearly defined data interfaces to form a complete closed-loop system from geometric modeling, physical simulation, state perception to intelligent decision-making. The functions of each model and module in the functional model layer are as follows: The assembly analysis module first imports the workpiece CAD geometric model and the fixture CAD geometric model into the simulation environment, and completes the rigid body assembly in the virtual space based on the assembly constraint relationships in the process design, such as reference surface fitting, locating pin-hole matching, and stopper alignment. Then, an alignment bounding box is established on the fixture-workpiece assembly to quickly determine whether there is spatial overlap, and the mesh patches corresponding to the overlapping area are further analyzed, including constructing an oriented bounding box (OBB) for each mesh patch and performing collision detection between patch pairs; for the patch pairs where the OBB intersects, the minimum Euclidean distance is calculated, and candidate patch pairs with a minimum Euclidean distance less than a preset contact distance threshold (0.1mm) are retained; the surface normal vector angles of the candidate faces are analyzed, and only patch pairs with an angle less than a preset threshold (such as 10°) are retained to ensure that they are potential face-to-face contacts, and candidate face sets that may be in contact are screened out. The candidate contact areas are screened through this hierarchical collision detection strategy; then, the face-to-face nearest point search algorithm is used in the candidate contact area to identify the fitting area between the fixture end face and the clamped surface of the workpiece, the coordinates of the contact center point and its surface normal vector are extracted, and each contact center point, i.e., the contact point, is assigned a contact point. Numbering, all normal vectors are normalized into unit vectors, and ensure that the direction points to the inside of the workpiece (i.e., the direction of pressure); next, according to the preset fixture functional topology information, such as the hydraulic cylinder-grip driving relationship, lever transmission chain, etc., a force transmission path number is assigned to each contact point. This number is used to identify the actuator branch to which the contact point belongs. For example, all contact points on the gripper driven by the same hydraulic cylinder share the same path number, and contact points at different positioning levels (main positioning, auxiliary positioning) have different path priority numbers. The preset fixture functional topology information can be called from the fixture design database to call the predefined force transmission topology table or manually marked in the simulation configuration interface; finally, the identified contact point information is structured to generate a structural configuration parameter file for the clamping system, including the contact point number, contact point three-dimensional coordinates, contact point unit normal vector, the fixture element to which it belongs, and the force transmission path number.

[0025] The clamping mechanism dynamics model is used to simulate the dynamic response characteristics of the hydraulic cylinder and electro-hydraulic proportional pressure valve in the clamping system during the main clamping phase. Taking the received target clamping force command as input, the model simulates the pressure-flow characteristics of the hydraulic system and the dynamics of the actuator to output the real-time total clamping force generated by the hydraulic cylinder. Specifically, all contact points are divided into several force transmission subsystems based on the force transmission path numbers provided by the assembly analysis module. The equivalent stiffness coefficient of each force transmission path is determined based on the finite element analysis of the fixture-workpiece assembly. The output real-time total clamping force is then distributed to each path according to the stiffness ratio: ; in, For the The clamping force distributed along the force transmission path at time t, For the The equivalent stiffness of the force transmission path, is the combined stiffness of all force transmission paths, The total clamping force at time t output by the dynamic model is used. Within each path, the force distribution weight is set according to the local stiffness of the contact point. , accordingly Decomposed into the normal force at each contact point: ; in, is the normal force of the ith contact point at time t, For the The local force distribution weight coefficient of contact point i in the path is finally obtained as the time-varying directional force on each contact point on the workpiece surface The target clamping force instruction is converted into a control signal u through the controller model, and its value range is [0, 1], which corresponds to the 0-10V voltage or 4-20mA current in the real system. Estimate the energy consumption of the main clamping process, perform time integration on the instantaneous power value P(t)=F(t)·v(t) in the main clamping time interval, and use As a proxy for hydraulic system energy consumption, v(t) is the cylinder piston velocity, and dt is the time increment representing the integration. This dynamic model provides high-fidelity physical constraints for offline training of the clamping decision strategy, ensuring that the clamping process adheres to the physical limitations of the real actuator.

[0026] The workpiece mechanical response model is based on the geometric and contact relationship of the fixture-workpiece assembly, retrieves the mechanical parameters of the workpiece material from the process database, and receives the clamping force distribution output by the dynamic model of the clamping mechanism. This response model constructs a static finite element model that considers contact nonlinearity, applies the clamping force as a surface load to the corresponding contact area, and imposes fixed constraints on the contact area of ​​the positioning element, thereby accurately calculating the stress distribution and elastic deformation field of the workpiece under the action of the clamping force, and outputs the maximum local deformation as the key basis for evaluating the deformation control of the workpiece.

[0027] The material contact characteristic model is used to store the contact point friction coefficient under the pairing of workpiece and fixture materials. During the simulation of the clamping system, the contact point information in the structural configuration parameters and the real-time total clamping force output by the clamping mechanism dynamics model are called, and the current target clamping force of each contact point i is calculated according to the Coulomb friction law. The maximum static friction force under the condition of the clamping system is then calculated. Then, according to the current cutting process and tool path, the direction of the total cutting force vector is determined, thereby judging the most likely slip direction of the workpiece. For example, if the direction of the total cutting force vector is along the positive X-axis of the airport coordinate system, then the slip risk direction is along the negative X-axis. Based on this, the maximum total static friction resistance of the clamping system in the slip direction is obtained as follows: For the i-th contact point, the maximum static friction force is : , i=1,2,3,...,n; in, is the maximum static friction force of the i-th contact point at time t, is the contact point friction coefficient of the i-th contact point between the workpiece and the fixture material pairing, is the normal force of the i-th contact point at time t, and the maximum static friction force vector of all contact points is projected to the anti-slip direction. The sum of these projected forces is used to obtain the maximum total static friction resistance that the entire clamping system can provide in the slip direction. , then project the estimated tangential force onto the determined slip direction to obtain its component magnitude in that direction , calculate the actual safety factor, actual safety factor η= / , which is used to assess the risk of workpiece slippage.

[0028] The virtual visual perception module is designed to learn to respond to changes in the workpiece surface state during the training phase of the clamping decision strategy. This module is constructed within a digital twin simulation environment to generate visual features consistent with those used in the real-world deployment phase. Based on the fixture-workpiece assembly, this module combines preset material surface statistical parameters from the process database, such as surface roughness, texture orientation, and previous machining traces, with common workpiece defect patterns such as edges, pits, scratches, and positioning datums. Two virtual industrial cameras are configured, with their positions and parameters consistent with those used in the real-world deployment system. A physical rendering engine is used to generate synthetic images that incorporate the following perturbations: illumination variations, surface disturbances, and imaging noise. The resulting synthetic images visually simulate the workpiece surface images captured by real industrial cameras. The resulting synthetic images are fed into a pre-trained deep convolutional network identical to that used in the real-world deployment phase to extract high-dimensional feature vectors of the workpiece surface structure and texture distribution. These feature vectors serve as "simulated visual features" and contribute to the construction of the training state vector.

[0029] An external integrated cutting force estimation model is called before each clamping decision. The tangential force of the workpiece processed by the precision boring and milling machine is estimated by the mechanical power method. The details are as follows: the cutting process parameters of the current workpiece are called from the process database, including the axial cutting depth a p (mm), radial cutting depth ae (mm), feed speed V f (mm / min), spindle speed N (rmp), number of tool teeth Z, tool diameter D (mm), and unit cutting energy K corresponding to the workpiece material c (N / mm²) (e.g. 800–1100 N / mm² for aluminum alloys and 2500–4000 N / mm² for titanium alloys) to calculate the tangential force during workpiece machining. The material removal rate MRR is calculated as follows: MRR=a p ·a e ·V f (mm³ / min), calculate cutting power P c :P c =K c MRR / 60(W), calculate cutting speed V: V=π·D·N / 60000(m / s), calculate tangential cutting force : =P c / V(N), and then calculate the total cutting force of the workpiece. =K· (N), K is the empirical coefficient, the calculated Amplification for estimating total cutting forces during precision milling The total cutting force is a three-dimensional vector, including tangential force, radial force and axial force. The coefficient K is not a fixed value. It depends on the tool geometry parameters, cutting type and workpiece material. The K value range is [1.5, 2.5]. The estimated main cutting force is used as the benchmark input for workpiece slip risk assessment.

[0030] The intelligent clamping decision model is constructed based on the deep deterministic policy gradient (DDPG) algorithm, which takes the comprehensive state vector as input and outputs the target clamping force adjustment coefficient α∈[0, 1]. In the above simulation environment, a reinforcement learning training framework is constructed to define the comprehensive state vector S, action space and reward mechanism of reinforcement learning respectively. The comprehensive state vector S includes visual feature samples, geometric distribution parameter samples, material mechanics parameter samples and cutting process parameter samples. The observation value of the current state of the digital twin simulation environment from the DDPG algorithm reflects the instantaneous state of the twin environment. The sample parameters are normalized and weighted fused to form a comprehensive state vector S with n dimensions (n≥100), which comprehensively characterizes the key factors affecting the clamping effect under the current working conditions. The action space is designed as a single continuous control quantity, namely the target clamping force adjustment coefficient α, and the target clamping force is generated by mapping. Design a multi-objective reward function ,in, is the workpiece deformation penalty term, calling the finite element analysis results of the workpiece geometric model and taking the negative value of the maximum local deformation as the reward value. , where k d is the deformation penalty coefficient, Δd max The maximum local deformation is achieved to ensure that excessive deformation is suppressed first during the strategy learning process; is the energy consumption reward item of the clamping system, which is based on the actual clamping force integral output by the clamping mechanism dynamics model, and takes a negative value as the penalty item. , where k e is the energy consumption penalty coefficient, which guides the strategy to reduce resource usage under the premise of satisfying the constraints. The constraints include: the clamping force is sufficient to resist the estimated cutting force to prevent the workpiece from slipping, the maximum deformation of the workpiece does not exceed the process allowable value to avoid machining errors, and the clamping force is within the physically achievable range of the clamping system; is the slip risk penalty item. If the actual safety factor η is less than the preset safety threshold (such as 1.5), a negative reward is introduced to prevent relative slip during processing. 、 、 Dynamically adjust according to processing accuracy requirements.

[0031] Based on the above settings, the deep deterministic policy gradient (DDPG) algorithm is used to train the reinforcement learning policy network. The training process relies on the constructed digital twin simulation environment of the boring and milling machine clamping system. The reinforcement learning policy network receives the comprehensive state vector S as input and outputs the target clamping force. During the training process, the controller model converts the target clamping force into a control signal and inputs it into the digital twin simulation environment. Based on the assembly analysis module, the material mechanical property model and the clamping mechanism dynamics model, the stress distribution of the workpiece under the target clamping force is simulated by finite element analysis to obtain the maximum local maximum deformation, and the calculated workpiece slip risk and the calculated hydraulic system energy consumption are called at the same time. Subsequently, the system calculates the comprehensive reward value R of this decision according to the preset multi-objective reward function, and feeds the comprehensive reward value R of this decision and the next comprehensive state vector together to the DDPG algorithm to constitute the feedback signal of reinforcement learning. The training process is as follows: Build an Actor network (policy network) and a Critic network (value function network), configure the corresponding target network (TargetActor and TargetCritic), and randomly initialize the network parameters; establish an experience replay buffer to store the comprehensive state vector of the interaction between the DDPG algorithm and the twin digital environment. DDPG observes the current comprehensive state vector S(t) in each round of training. The Actor network outputs the action F(t), which is mapped and input into the simulation environment to trigger the clamping process simulation. The simulation environment returns an immediate reward R tTogether with the new comprehensive state vector S(t+1), the four-tuple (S(t), F(t), R(t), S(t+1)) is stored in the experience replay buffer. The current comprehensive state vector S(t) is the DDPG algorithm's observation of the current state of the digital twin simulation environment at time step t, i.e., the tth decision moment. Its data all originates from and reflects the instantaneous state of the twin environment in real time. This vector is constructed and fused from the twin environment through the following process: Visual feature sample: Based on the geometric model of the fixture-workpiece assembly in the twin environment, combined with the preset material surface properties in the process database, a synthetic image representing the current environmental state is generated through the virtual visual perception module and the physical rendering engine. The image is input into the pre-trained MobileNetV3 network to extract the feature vectors of the workpiece surface structure and texture distribution; Based on the geometric model of the fixture-workpiece assembly in the twin environment, the spatial coordinates, dimensional tolerances, shape and other geometric parameters of the key positioning areas of the workpiece are extracted; According to the material types of the workpiece and fixture defined in the twin environment, the corresponding material properties are called from the process database, including elastic modulus, Poisson's ratio, density, and the interface static friction coefficient of each contact point to form a material mechanical parameter sample; According to the currently simulated processing procedure, the cutting parameters, including ap, ae, Vf, N, Z, D and Kc, are called from the process database to form a cutting process parameter sample, the above four types of feature vectors are spliced, and the spliced ​​vectors are normalized. During each training iteration, the system randomly extracts a batch of historical interaction data from the experience replay buffer to update network parameters. First, the critic network (value function network) is optimized: based on the current state and the action taken, the critic network estimates the expected long-term reward of the action. Simultaneously, the target network predicts the optimal reward for the next state and combines it with the current immediate reward to calculate a target reward. By comparing the difference between its own predictions and the target reward, the critic network continuously adjusts its internal parameters to gradually reduce this prediction bias, thereby improving its evaluation accuracy. Subsequently, the actor network (policy network) is updated: based on the critic network's feedback on the action, the actor network adjusts its policy parameters in a direction that maximizes the expected reward. The system calculates the sensitivity of the reward to the action and the dependency of the action on the policy parameters. Using gradient ascent, the actor network's weights are updated, so that under similar states, the network is more likely to output a clamping force adjustment coefficient that yields a higher reward. In order to improve the stability of the training process, the system adopts a target network mechanism, that is, setting up independent TargetActor and TargetCritic networks, and gradually approaching the latest parameters of the main network at a smaller proportion through slow tracking, thereby effectively suppressing oscillations and divergence in the learning process.As the number of training rounds increases, the decision-making performance of the policy network tends to stabilize and reach a convergence state, obtaining the final clamping decision policy network.

[0032] Furthermore, the MLP model is obtained by transferring the decision-making ability of the clamping decision strategy network trained and converged by the DDPG algorithm through the knowledge distillation method, including: Run the trained clamping decision strategy network in the constructed digital twin simulation environment to generate multiple comprehensive state vectors S i , and collect the target clamping force adjustment coefficient α output by the clamping decision strategy network in this state i ; The state-action pair (S i , F i ), construct a distillation dataset D={(S i , F i )|i=1, 2, ..., M}, and perform normalization preprocessing on all state vectors; The MLP model structure is designed according to the dimension of the integrated state vector. The number of neurons in the input layer is consistent with the dimension of the integrated state vector. The hidden layer uses the ReLU activation function, and the output layer is a single neuron linear unit, which is used to predict the target clamping force. Using the distilled dataset D as training samples, a supervised learning approach is used to minimize the mean squared error between the MLP prediction output and the clamping decision policy network output, training the MLP model to imitate the policy decision-making ability of DDPG. The trained MLP model is quantized and compressed using INT8, and the compressed lightweight MLP model is deployed to the industrial computer of the boring and milling machine.

[0033] Specifically, a synthetic data distillation method driven by digital twins is used to run the clamping decision strategy network, i.e., the teacher model, which has been trained and converged by the DDPG algorithm, in a digital twin simulation environment. The teacher model traverses the geometric types of the fixture-workpiece assembly and combines them with the material mechanics parameters and cutting process parameters preset in the process database to generate a comprehensive state vector covering various processing scenarios in the simulation environment. The output α of the clamping decision strategy network in this state is then collected. i， Composition state-action pair (S i , F i ), construct a distillation dataset D={(S i , F i)|i=1, 2, ..., M}. All state vectors are pre-processed by normalization to ensure input distribution consistency. A multi-layer perceptron (MLP) is constructed as the student model. A supervised learning approach is used, with the distilled dataset D as training samples. The goal is to minimize the mean squared error between the MLP prediction output and the DDPG teacher model output. The MLP model is trained to mimic the strategic decision-making capabilities of DDPG. The trained MLP model is then quantized and compressed using INT8, converting floating-point weights and biases into 8-bit integer representations. This significantly reduces model storage space and inference computation while maintaining decision accuracy. The compressed lightweight model is then deployed on the machine tool industrial computer, establishing a state vector input interface and connecting it to the output port of the electro-hydraulic proportional pressure valve controller.

[0034] S400: Based on the obtained target clamping force, the controller generates a corresponding target current instruction according to a preset pressure-current mapping relationship. The instruction drives the electro-hydraulic proportional valve to adjust the output force of the hydraulic cylinder, driving the clamping mechanism to move and complete the workpiece clamping.

[0035] Specifically, this step is implemented to convert the target clamping force output by the previous sequence into actual executable mechanical control instructions. The clamping system can be configured with an independent industrial controller, which is a high-performance PLC or industrial PC. It realizes data exchange with the CNC controller of the precision boring and milling machine through a standard communication interface. The controller of the clamping system receives the target clamping force and calculates the required target hydraulic oil pressure based on the known cylinder piston area. Based on the target hydraulic oil pressure, the corresponding target current is found in the preset pressure-current mapping relationship through the table lookup method. The target current value range is usually 0-2A or 4-20mA. The controller writes the calculated target current into the register of the AO module, and the AO module converts it into a real current control signal output. The output terminal of the AO module is connected to the solenoid coil of the electro-hydraulic proportional valve, driving the valve core to move and adjust its valve port opening, thereby controlling the oil pressure and flow delivered by the hydraulic system to the hydraulic cylinder, driving the clamping assembly to complete the reliable clamping of the workpiece. The clamping assembly is a hydraulically driven linear motion unit, including a hydraulic cylinder, a piston rod, and a clamping plate or claw assembly linked to it, etc., which is used to convert hydraulic energy into a mechanical clamping force applied to the workpiece. The target clamping force is converted into control instructions for the electro-hydraulic proportional valve. The product manual can be consulted, which typically provides a detailed pressure-current characteristic curve. Alternatively, a table lookup method can be used. During the electro-hydraulic proportional valve testing phase, different currents are applied to the valve and its output pressure is measured to obtain a set of data points (Ii, Pi). Using a curve fitting algorithm such as the least squares method, this set of data points is fitted into a continuous pressure-current characteristic curve function P = f(I). This continuous function is then discretized to generate a pressure-current mapping lookup table. The clamping system must maintain a stable clamping force throughout the entire cutting process and cannot be released. Therefore, the main clamping phase also refers to the entire process from the completion of clamping, the start of cutting on the precision boring and milling machine, to the completion of the process. During this phase, the clamping system is in a dynamic load environment and requires real-time monitoring and compensation. Vibration signal: install a micro vibration sensor near the clamping base or pressure plate or use the vibration monitoring module provided by the precision boring and milling machine; temperature signal: arrange the temperature sensor at the hydraulic oil inlet / valve block oil circuit and near the workpiece clamping area; obtain the output force of the hydraulic cylinder in real time, install the pressure sensor at the outlet of the electro-hydraulic proportional valve (near the actuator side) or the hydraulic cylinder inlet to ensure that the measured pressure is the actual pressure acting on the cylinder.

[0036] S500: In the main clamping stage, the vibration signal, temperature signal and hydraulic oil pressure signal of the clamping system are collected simultaneously. The time domain features of the vibration signal and temperature signal are extracted in units of a preset time window to obtain the vibration time domain features and temperature time domain features. The average value of the hydraulic oil pressure in the window is calculated. Based on the known effective area of ​​the piston, the hydraulic oil pressure is converted into the current output force of the hydraulic cylinder. The current output force is input into the calibrated force transfer model to obtain the total effective clamping force on the workpiece, and the clamping force deviation between the target clamping force and the total effective clamping force is calculated.

[0037] Specifically, clamping force deviation is a result-based indicator, indicating only that the current total effective clamping force has deviated from the target clamping force, but it cannot reveal the cause or nature of the deviation. Vibration and temperature signals are process-based or causal indicators, containing rich information about the source and characteristics of disturbances. Vibration signals are extremely sensitive to high-frequency, transient disturbances, cutting shock, and spindle vibration. Temperature signals react strongly to low-frequency, slow disturbances (such as fixture thermal deformation, hydraulic oil viscosity changes, and valve body thermal drift). Simultaneous analysis of clamping force deviation, vibration, and temperature signals is necessary to comprehensively diagnose the system's "health" and enable intelligent compensation decisions. In most cases, the clamping system of a precision boring and milling machine typically utilizes a single-cylinder-driven, multi-point linkage clamping mechanism. The core structure is a centralized power source (a single hydraulic cylinder) that transmits force synchronously and proportionally to multiple clamping points through a mechanical force distribution mechanism. A lever-type force distribution mechanism is the most common type of force distribution mechanism. Through one or more rigid levers, the cylinder's output force is distributed to two or four clamping points according to the lever ratio. The force loss to be compensated primarily comes from friction and deformation in the force distribution mechanism. The total effective clamping force estimated by the model is the effective clamping force actually acting on the workpiece.

[0038] Furthermore, the calibrated force transfer model includes: Obtain the material, size, and hardness information of the workpiece, process a calibration block with the same properties, and install a force sensor at the bottom of the calibration block; The calibration block is installed in the fixture, and the controller sends N-level current signals. At each level, the hydraulic oil pressure P and the total effective clamping force monitored by the force sensor are synchronously collected. The hydraulic oil pressure P is combined with the known effective area A of the cylinder piston to calculate the current output force of the hydraulic cylinder in real time. , repeat the collection M times at each level, take the average value, and get a data set of N points , j = 1, 2, ..., N, where N ≥ 5 and M > 1; Directly fit the linear model between the current output force of the hydraulic cylinder and the total effective clamping force to establish the overall force transmission model. = + , j = 1, 2, ..., N, Use the least squares method to find the parameters and , is the total mechanical efficiency of the clamping system, <1,

[0039] is the total force loss of the clamping system; Will and Packaged as a force transfer model, , solidify the model parameters, and The force transfer model is written to the non-volatile memory of the controller and marked as a calibrated model. Specifically, this embodiment uses physical experimental measurements to determine specific parameters within a known model structure. First, a calibration fixture is developed to obtain information about the workpiece's material properties, dimensions, and hardness. A single, integrated calibration block with the same properties is then fabricated. This single calibration block can serve multiple production lines, reducing costs. A high-precision force sensor is installed between the calibration block and the machine tool's worktable (i.e., at the base of the workpiece). The total clamping force applied by the fixture through the indenter is equal to the worktable's support force on the workpiece. The force sensor measures this support force. For applications with diverse clamping forces, such as five-sided machining, this is mechanically represented by the sum of all clamping forces applied by the fixture system to the workpiece, effectively resisting the total cutting force in any direction. Because the purpose of calibration is to anchor it to reality, its task is to establish a mapping that ensures that when the controller commands a certain number of Newtons of cylinder pressure, it knows and guarantees the actual normal force acting on the workpiece. The total clamping force is the foundation of the spatial force system. Regardless of the direction of the cutting force, the workpiece's resistance to slip depends primarily on the friction on the fixture's contact surfaces. The magnitude of the friction force is proportional to the positive pressure on the contact surface (i.e., the normal clamping force). Therefore, the target clamping force is essentially a total, synthetic normal clamping force, and what the single sensor measures at the bottom is precisely this total, synthetic normal support force. The physical nature of the two is unified. Calibration compensates for "system internal friction" rather than "processing conditions." These internal frictions are independent of the processing direction. Regardless of whether three-sided or five-sided processing is performed, the inherent force transmission efficiency of the system remains basically unchanged. Therefore, it is reasonable to use a unidirectional calibration to measure and compensate for the inherent, unidirectional internal friction of the system. Secondly, perform multi-level calibration, install the calibration fixture into the real fixture, and the clamping system controller sends N-level current signals, that is, sends N different current signals, N≥5. The design principle of the N-level current signal follows the uniform distribution principle or the boundary point inclusion principle within the current signal value range. At each level, the hydraulic cylinder pressure P and the total effective clamping force measured by the force sensor are synchronously collected. The hydraulic cylinder pressure P is combined with the preset cylinder piston area A to calculate the current output of the hydraulic cylinder in real time. =P×A, repeat each level M times (M>1), take the average value, reduce the noise, and get a data set of N points , j = 1, 2, ..., N. Then, a force transmission model is established to directly fit the linear model between the cylinder output force and the total effective clamping force. = + , using the least squares method, find an optimal pair of and The value of makes the dependent variable in this line model With independent variables The sum of squared errors is minimized, and a set of optimal parameters is obtained and , is the total mechanical efficiency of the system, <1, represents the overall transmission efficiency from the cylinder output force to the workpiece effective force, is the total force loss of the system, which mainly represents the static friction that needs to be overcome. and Packaged as a force transfer model, Finally, solidify the model parameters and and The model is written into the non-volatile memory of the controller and marked as a calibrated model for recall during the operation phase.

[0040] S600: The extracted vibration time domain features and temperature time domain features are combined with the clamping force deviation to construct a normalized current multimodal disturbance observation vector in the form of [vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, clamping force deviation].

[0041] Specifically, during the main clamping phase, a unified high-precision clock serves as a synchronization reference, enabling multi-rate synchronous acquisition of hydraulic oil pressure, clamping system vibration, hydraulic oil temperature, and fixture temperature signals. All signals are precisely timestamped during acquisition and aligned based on a unified clock source, ensuring strict temporal synchronization of all physical quantities in subsequent analysis. For the hydraulic oil pressure signal, the arithmetic mean of the most recent 100ms of sampled data within a 100ms analysis period is calculated, converted into the current hydraulic cylinder output force, and the clamping force deviation is derived. Vibration signal processing: Using a sliding time window approach with a 100ms analysis period, the latest continuous 1s of raw vibration data are captured, and their root mean square (RMS) value and kurtosis are calculated. For the temperature signals, a linear regression method is used, using the same time window approach, to perform linear fitting on the hydraulic oil and fixture temperature signals, respectively. The resulting slopes are the hydraulic oil and fixture temperature change rates, respectively. For the hydraulic pressure signal, the output force at the last moment within the preset time window is calculated. Using the Min-Max normalization method, each feature is mapped to the interval [0, 1] to obtain the normalized multimodal disturbance observation vector [vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, and clamping force deviation]. Given the large dynamic range of vibration kurtosis features and the potential for extreme values, direct Min-Max normalization results in a loss of effective feature discrimination. Therefore, preprocessing is required before normalization. First, a natural logarithmic transformation (K' = ln(K + 1)) is applied to the raw vibration kurtosis values ​​to map them to a reasonable linear interval. Subsequently, the processed vibration kurtosis feature K' is subjected to Min-Max normalization along with the other features.

[0042] S700: Input the current multimodal disturbance observation vector into the pre-constructed fuzzy rule model, generate a synthetic fuzzy set of the hydraulic compensation factor through fuzzification, rule activation, implication operation and maximum aggregation, use the center of gravity method to perform defuzzification within the hydraulic compensation factor domain [0, 1], calculate the geometric center of the synthetic fuzzy set, and use the calculated result as the output value of the hydraulic compensation factor, wherein the fuzzy rule model is constructed by a fuzzy decision tree.

[0043] Furthermore, pre-built fuzzy rule models include: The historical operating time series data of a precision boring and milling machine was obtained. This data covered both normal and various fault conditions. A sliding window was used to extract features from this historical data, including vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, and the hydraulic oil pressure window mean. Based on the hydraulic oil window mean and the known effective piston area, the hydraulic cylinder output force was calculated and considered the total effective clamping force. This was then subtracted from the set clamping force to obtain the clamping force deviation. The sample values ​​of the five features (vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, and clamping force deviation) are combined in a fixed order to form a single multimodal disturbance observation sample vector. After completing the above operation on all historical data, all the obtained sample vectors are normalized to obtain a set of multimodal disturbance observation sample vectors for training. At the same time, experts label each multimodal disturbance observation sample vector in the sample vector set with its corresponding hydraulic compensation factor value based on the fault severity and compensation requirements. The labeling value range is [0, 1]. The set consisting of all sample vectors and their hydraulic compensation factor values ​​constitutes the training data set. A linguistic value is defined for each feature in the multimodal disturbance observation sample vector, including low, medium, and high. At the same time, a linguistic value is defined for the hydraulic compensation factor, including zero, small, medium, and large. Each of the above linguistic values ​​is configured with a corresponding membership function, which adopts a triangular or trapezoidal distribution. The parameters of each membership function are initialized to generate the corresponding initial fuzzy set for each of the above linguistic values, and the global parameters of the fuzzy decision tree are preset. The process of building a tree structure using fuzzy decision tree: S1: Starting from the root node, traverse the five features of the multimodal perturbation observation sample vector. For each feature, traverse each of its language values ​​and use each feature-language value combination as a candidate split condition to be evaluated. S2: For each candidate split condition, extract the sample values ​​of all sample vectors in the training data set on the feature, and substitute each sample value into the membership function corresponding to the language value to calculate the membership degree of the language value. Based on the membership degree calculated by the sample value, calculate the fuzzy information gain rate of the candidate split condition. S3: At the current node, compare the fuzzy information gain rates of all candidate splitting conditions, select the feature-language value combination with the largest gain rate as the optimal splitting condition for the current node, and generate two child nodes, one belonging to and one not belonging to. For each sample vector and its hydraulic compensation factor value in the training data set, calculate its membership to the language value of the optimal splitting condition based on its data on the feature specified by the optimal splitting condition. The sample vector and its hydraulic compensation factor value carry the weight of the membership size and enter the child node belonging to. At the same time, the sample vector and its hydraulic compensation factor value carry the weight of 1 minus the membership size and enter the child node not belonging to. The sample vector and its hydraulic compensation factor value remain unchanged during the allocation process to form a training data set for each child node. The weights carried are used in all subsequent recursive calculations. Each non-leaf child node is used as the new current node. The above splitting process is repeated recursively until the preset stop condition is met. The nodes that cannot continue to split become leaf nodes, and a fuzzy decision tree structure with unassigned output is constructed. S4: For each leaf node in the fuzzy decision tree structure, based on the cumulative path weights of all sample vectors in the node and from the root node to the node, the membership of each linguistic value of the hydraulic compensation factor is weightedly aggregated, and the linguistic value with the highest aggregated membership is selected as the output of the leaf node to complete the construction of the fuzzy decision tree; S5: Based on the constructed fuzzy decision tree, traverse each path from the root node to the leaf node, connect the splitting conditions of all non-leaf nodes on the path, that is, the feature-language value combination, through the AND logical operator to form the antecedent of the rule, that is, the IF part. The output language value assigned to the leaf node is used as the consequent of the rule, that is, the THEN part, to form a complete IF-THEN fuzzy rule. The rules corresponding to all paths together constitute the initial fuzzy rule set; The rule set combines the linguistic values ​​of all features in the multimodal disturbance observation sample vector, the linguistic values ​​of the hydraulic compensation factor, and their corresponding membership functions and initial parameters to form an initial fuzzy rule knowledge base; S6: Optimize the initial fuzzy rule base, encapsulate and store the optimized fuzzy rule set and the optimal parameters of the membership function corresponding to each language value in a structured data format to form an optimized fuzzy rule knowledge base. By introducing a fuzzy inference engine and combining it with the knowledge base, a fuzzy rule model that can be embedded in the control system is constructed.

[0044] Specifically, the constructed fuzzy rule model is used to solve the compensation force problem. The compensation control goal of the hydraulic clamping system is to maintain the stability of the clamping force. The fuzzy rule base is a set of fuzzy reasoning rules in the form of "IF-THEN", which describes the fuzzy mapping relationship between the input five-dimensional normalized features and the output hydraulic compensation factor u.

[0045] During the data acquisition and preprocessing stage, the historical operation time series data of the precision boring and milling machine are obtained. The data covers all important states of the equipment, including normal state and various fault states. The normal state includes stable operation data under different workpieces and different cutting parameters. Various fault states include direct disturbance sources such as tool wear, bearing failure, insufficient guide rail lubrication, hydraulic system leakage and insufficient clamping force. The historical operation time series data obtained at least includes original time series data such as vibration of the clamping system, hydraulic oil temperature, fixture temperature and hydraulic oil pressure. For the acquired historical operating time series data, consistent with the sliding time window method mentioned in the specific implementation of S600, the vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, and fixture temperature change rate are calculated for the data within each window, and the average hydraulic oil pressure within the window is calculated. Based on this average and the known effective piston area, the hydraulic oil output force is calculated and used as the total effective clamping force within the time period. This is subtracted from the set clamping force to obtain the clamping force deviation. The eigenvalues ​​of the five features (vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, and clamping force deviation) are combined into a five-dimensional eigenvector, which serves as the multimodal perturbation observation sample corresponding to the time window. The sliding window is used to traverse the entire time series to generate a set of multimodal perturbation observation sample vectors. The system-set clamping force is derived from manual experience and can be a fixed parameter preset in the machining program (G code) based on drawing requirements and historical experience. To label each sample vector with the output value of the hydraulic compensation factor u, the labeling rule is: for sample vectors in normal state, u=0, indicating that no compensation is required; for sample vectors in fault state, experts assign a value between 0 and 1 based on the severity of the fault and the compensation requirements. The normalized multimodal disturbance observation sample vectors and the corresponding hydraulic compensation factor u values ​​form a training data set.

[0046] In the initialization phase of the fuzzy rule model, a {low, medium, high} language value is defined for each sample feature in the sample vector, and a {zero, small, medium, large} language value is defined for the hydraulic compensation factor u. There are 5 features in the sample vector, and each feature has 3 language values ​​and 1 hydraulic compensation factor. 4 language values ​​are preset, which are the 15 feature language values ​​and 4 hydraulic compensation factors u mentioned above. Each tree is initialized with a separate membership function, typically a triangular or trapezoidal function. Membership function parameters are typically initialized using a uniform distribution strategy to cover the entire theoretical domain and provide a reasonable starting point for subsequent optimization. The global parameters of the fuzzy decision tree are preset, including the maximum depth, minimum number of node samples, minimum node split weight sum, minimum gain required for splitting, maximum number of leaf nodes, and fuzzy aggregation operators. Termination conditions for fuzzy decision tree learning are also set, including reaching the preset maximum depth, all samples belonging to the same class (or a very small output value variance), and the fuzzy information gain rate being less than a threshold.

[0047] In the fuzzy decision tree learning phase, a fuzzy decision tree (such as the fuzzy ID3 algorithm) is used to construct a tree structure and generate candidate splitting conditions. Based on the training data set, starting from the root node, each feature Xj and each of its linguistic values ​​Ak (such as the fuzzy linguistic values ​​of the vibration sample RMS: low, medium, and high) are traversed to generate each feature-linguistic value combination (Xj, Ak), such as (vibration RMS, high), (vibration RMS, medium), (vibration RMS, low), (hydraulic oil temperature change rate, high), etc. Each combination represents a potential splitting solution; the candidate splitting conditions are evaluated. For each candidate splitting condition (Xj, Ak), such as (vibration RMS, high), the following calculations are performed: a) Calculate the membership degree: Extract the specific numerical values ​​of all samples in the data set on the feature Xj, substitute each numerical value into the membership function corresponding to the linguistic value Ak, and calculate its membership degree to Ak. For example, calculate the degree to which each vibration RMS value belongs to "high"; b) Calculate the fuzzy information gain: Based on the category distribution of all samples at the current node, calculate the fuzzy information gain that would result from fuzzily dividing the sample into two child nodes ("belongs to Ak" and "does not belong to Ak") using the "degree to which the sample belongs to Ak" as the splitting criterion. Select the optimal splitting condition and compare the fuzzy information gain calculated for all candidate splitting conditions (Xj, Ak). Select the feature-language value combination with the highest gain as the optimal splitting condition for the current node. For example, (hydraulic oil temperature change rate, high) has the highest gain and becomes the splitting rule for the current node. Perform fuzzy splitting, simultaneously assigning each training sample at the current node to two new child nodes: The weight of the sample entering the "belongs" child node is equal to the degree to which its hydraulic oil temperature change rate value belongs to "high." The weight of the sample entering the "does not belong" child node is equal to 1 minus the aforementioned membership. All original feature values ​​of the sample remain unchanged during this process; only their weights are redistributed. The two newly generated child nodes of "belong" and "do not belong" are used as new current nodes respectively, and the steps are recursively repeated to generate candidate splitting conditions until fuzzy splitting is performed, and the best splitting condition is found for each new node until any stop condition preset in the initialization phase is met to form a leaf node. At this point, an unpruned optimized fuzzy decision tree is initially constructed. Among them, calculating the fuzzy information gain rate is a quantitative calculation process based on fuzzy set theory and information entropy, which aims to evaluate and select the most discriminative one from a large number of candidate splitting conditions. In the process of constructing a fuzzy decision tree, in order to evaluate the candidate splitting conditions and select the optimal splitting rule, for each condition to be evaluated (such as "vibration RMS is high"), the whole process is as follows: calculate the sample membership, and based on the membership function corresponding to the language value, calculate the degree to which each sample in the parent node satisfies the condition, that is, its membership μS 高(Xj) is used to quantify the degree of conformity of the sample to the fuzzy concept of "high vibration"; perform fuzzy splitting (soft assignment). This soft assignment process does not change any original values ​​of the sample vector and the hydraulic compensation factor. It only calculates two new weight values ​​for each sample to represent the degree of belonging or contribution of the sample to the newly generated "belong" and "do not belong" sub-nodes, with weight μS 高 (Xj) Assign the sample to the “belong” branch (yes) with a weight of 1-μS 高 (Xj) Assign the sample to the "does not belong" branch (no). This process implements soft partitioning, allowing samples to partially belong to multiple branches at the same time, reflecting the fuzzy logic's ability to model uncertainty and transition states; calculate the weighted fuzzy entropy of the child nodes, and calculate the fuzzy entropy of the two child nodes separately to measure the uncertainty of their output categories (such as the linguistic value of the hydraulic compensation factor). Subsequently, the fuzzy entropy of the two child nodes is weighted and summed, using the ratio of the total weight received by each child node to the total weight of the parent node as the weight coefficient, to obtain the expected fuzzy entropy under this splitting method; calculate the fuzzy information gain, and subtract the above-mentioned expected fuzzy entropy from the fuzzy entropy of the parent node to obtain the fuzzy information gain of the splitting condition. This value reflects the effectiveness of the condition in reducing output uncertainty. The larger the gain, the stronger its classification ability; calculate the splitting information. In order to avoid the algorithm's bias towards generating too many or unevenly distributed branches, splitting information is introduced, which is defined as the information entropy based on the weight ratio of each child node. The more evenly distributed the weights, the greater the split information value, indicating a higher data partitioning complexity. The fuzzy information gain ratio is calculated by dividing the fuzzy information gain by the split information. This ratio comprehensively considers the classification effect and partitioning complexity of the split, effectively balancing the gain size and model generalization ability. Ultimately, among all candidate split conditions, the one with the largest fuzzy information gain ratio is selected as the optimal splitting rule for the current node, used to generate the decision tree branch.

[0048] For each leaf node of the fuzzy decision tree, based on all the training samples that reach the node, the following operations are performed to determine its output language value: First, the cumulative path weight of each sample from the root node to the leaf node is calculated and normalized; secondly, for each candidate language value of the hydraulic compensation factor, its weighted aggregate membership at the node is calculated, that is, the sum of the product of the membership of each sample to the language value and its normalized weight; finally, the language value with the highest aggregate membership is selected as the final output of the leaf node to complete the construction of the fuzzy decision tree. The specific output is as follows: obtain the path weight of each sample, calculate the cumulative path weight of each sample vector from the root node to the leaf node, that is, the product of all split condition memberships on the path; normalize the weight, normalize the path weights of all samples so that their sum is 1; calculate the aggregate membership of each language value, for each candidate language value C of the hydraulic compensation factor k, calculate its weighted aggregate membership at the node , ,in, : hydraulic compensation factor value of the i-th sample vector For language value C k The membership degree, is the normalized path weight of the i-th sample vector, and N is the number of sample vectors; the language value with the highest membership is selected as the output, and the language value with the largest aggregated membership is selected as the final output of the leaf node.

[0049] Traversing each root-to-leaf path of the final decision tree and extracting a fuzzy rule generates a fuzzy rule IF ... THEN .... The splitting conditions on the path are connected through "AND" to form the rule antecedent (IF part). The output language value of the leaf node is determined by the average membership of the node sample to each output fuzzy set. The highest membership is selected as the rule consequent (THEN part), forming the initial rule base. For example, to deal with typical vibration disturbances, IF vibration energy IS high AND vibration impact IS high THEN hydraulic compensation factor IS large.

[0050] In the initial rule base, rule base optimization and parameter tuning are performed, and rule pruning is performed to remove rules with fewer samples covered than the threshold and with a confidence level lower than the set value to prevent overfitting. Rule merging is performed, and for rules with the same antecedent but different consequents, conflicts are resolved by weighted averaging or selecting the maximum membership based on the consequent membership. Membership function parameter optimization is performed using gradient descent or genetic algorithms with the goal of minimizing the prediction error of the training data. The parameters of all membership functions can be uniformly represented and then globally fine-tuned. The optimization process keeps the rule structure unchanged and only adjusts the shape of the membership function to improve the model accuracy and robustness.

[0051] Furthermore, the synthetic fuzzy set of hydraulic compensation factors includes: Substitute each eigenvalue in the current multimodal perturbation observation vector into the membership function of its corresponding {low, medium, high} language value, and calculate the membership degree of the eigenvalue on each language value; Traverse each rule in the optimized fuzzy rule set. If its antecedent is a single condition, the membership of the condition is the activation strength of the rule. If the antecedent contains multiple conditions connected by AND, take the minimum value of the membership of each condition as the activation strength of the rule. For each rule with activation strength greater than 0, obtain the membership function corresponding to the consequent linguistic value, perform min implication or multiplication implication on the membership function on the entire domain of the hydraulic compensation factor, and obtain the output fuzzy set of the activated rule, where the entire domain of the hydraulic compensation factor is [0, 1]. The output fuzzy sets of all activated rules are aggregated on the universe to obtain the composite fuzzy set of hydraulic compensation factors.

[0052] Specifically, according to the preset fuzzy rule base, the activation strength of the input variables is calculated, and the output fuzzy set of each rule is generated through the implication operation, and then the maximum method is used for synthesis operation to obtain the synthetic fuzzy set of u. Execute the fuzzy reasoning process: Fuzzification step, each eigenvalue in the current disturbance observation vector is substituted into its corresponding membership function, and the membership of the eigenvalue to each fuzzy language value is calculated. The input features include: clamping system vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, clamping force deviation, and the fuzzy language value of each feature is {low, medium, high}; rule activation strength calculation step, traverse each IF-THEN rule in the fuzzy rule base (for example: "IF vibration = high AND pressure = medium THEN u = large"), for the current rule, extract the membership corresponding to each fuzzy condition (such as "vibration = high", "pressure = medium") in its antecedent (IF part), and use the minimum operation to calculate the activation strength α of the rule r If α r > 0, the rule is activated and participates in subsequent reasoning; otherwise, it is skipped; the implication operation is performed to generate the rule output fuzzy subset. For each activated rule, the min-implication operation is performed on the consequent fuzzy set corresponding to the linguistic value specified by the consequent of the rule (such as "large"). The consequent fuzzy set is the membership function configured by the linguistic value (such as {zero, small, medium, large}) defined for the output variable u (hydraulic compensation factor) in the system initialization phase, and the initial parameters have been set; the implication operation is performed by comparing the membership function point by point with α on the output domain [0, 1] rTake the minimum value to achieve truncation and obtain the output fuzzy subset of this rule; perform maximum synthesis on the output fuzzy subsets of all activated rules in the domain [0, 1]. For each point in the output domain, take the maximum membership degree of all rules at this point to form the final synthesized fuzzy set, which comprehensively reflects the fuzzy inference results of all activated rules on the hydraulic compensation factor. All consequent fuzzy sets are artificially preset and initialized at the initial stage of system construction (for example, initialize "medium" as a trapezoidal function centered at 0.5), and are initialized synchronously with the membership functions of input features; during subsequent training, the parameters of these consequent fuzzy sets (such as vertices, base widths) can be optimized through data-driven methods (such as gradient descent, genetic algorithms) to improve the inference accuracy; the optimized consequent fuzzy sets are saved as part of the fuzzy rule base for real-time compensation decision-making during the online inference stage. After fuzzy inference, a synthesized fuzzy set of the hydraulic compensation factor is obtained. To convert it into an accurate control signal applicable to the hydraulic system, defuzzification is performed. Specifically, the centroid defuzzification method is used to perform discrete sampling at a preset step size (such as 0.01) in the domain [0, 1] of the output variable, and calculate the geometric center u of the synthesized fuzzy set. Its formula is: , where, is the k-th sampling point, is the membership degree of the synthesized fuzzy set at , and N is the total number of sampling points. The finally obtained accurate value , with a value range of [0, 1] as the final output of the hydraulic compensation factor, is used to adjust the opening degree of the proportional valve or servo valve in the hydraulic compensation circuit to achieve dynamic compensation of the clamping force.

[0053] S800: According to the clamping force deviation between the target clamping force and the total effective clamping force and the output value of the hydraulic compensation factor, calculate the current correction value of the electro-hydraulic proportional valve, and superimpose it on the target current command as the control signal of the electro-hydraulic proportional valve together, so as to dynamically adjust the spool opening degree, generate the required hydraulic oil pressure, and maintain the stable clamping force of the clamping system during processing. , where, is the current correction value, is the clamping force deviation, is the output value of the hydraulic compensation factor, is the slope of the pressure-current characteristic curve of the electro-hydraulic proportional valve.

[0054] Specifically, this embodiment is to use a simple method to solve a complex problem. It is to generate an additional "compensation force" to offset the disturbance. This compensation force can be directly generated by the hydraulic system. The clamping force deviation △F is the comprehensive final manifestation of various internal and external disturbances. Whether it is the change in viscosity characteristics caused by changes in hydraulic oil temperature, thermal deformation caused by changes in fixture temperature, or disturbances not reflected in the modeling force, such as workpiece positioning surface errors and friction fluctuations, the final effect will be concentrated on the deviation of the clamping force from its target value. Therefore, by observing △F, generalized disturbances including thermal effects can be indirectly captured. Based on the clamping force deviation △F and the hydraulic compensation factor u, a compensation force for offsetting the disturbance is generated and converted into a current instruction for the electro-hydraulic proportional valve. The specific implementation steps are as follows: Compensation force calculation, the clamping force deviation △F is regarded as the disturbance force that needs to be offset, and it is multiplied by the hydraulic compensation factor u to obtain a desired compensation force. , =(△F) u, u as an intelligent weight (between 0 and 1), determines the degree of compensation for △F under the current working conditions. u=1 means 100% compensation, and u=0 means no compensation. Current command conversion, the output force of the electro-hydraulic proportional valve is proportional to the input current. This proportional relationship is determined by the pressure-current curve of the valve, and its slope (proportional coefficient) is , the unit is N / mA or kN / mA, so in order to generate the desired compensation force The required current correction command △I is calculated by the following formula: ,parameter It is a fixed known parameter of the electro-hydraulic proportional valve. The product manual usually provides a detailed pressure-current characteristic curve or is calibrated through experiments. For example, a set of known current signals I is applied to the proportional valve, and the output force F generated is calculated. Then, a linear fit is performed on the data points (I, F). The slope of the fitted line is Instruction superposition, the calculated current correction instruction △I is superimposed on the system's original control instruction, that is, the target current instruction, and together they serve as the final control signal of the electro-hydraulic proportional valve, thereby dynamically adjusting the valve core opening, generating the required compensation force, and stabilizing the clamping force.

[0055] Based on the multi-parameter collaborative control method of a precision boring and milling machine clamping system in the aforementioned embodiment, and with the same inventive concept, the present application also provides a multi-parameter collaborative control device for a precision boring and milling machine clamping system, as shown in the attached Figure 2 , the device comprises: Composite image module 10, which is used to pre-clamp the workpiece on the fixture and collect the workpiece surface image through the industrial camera. The collected image is processed by SIFT feature matching algorithm and illumination normalization to form a composite image for feature extraction; A state vector module 20 is used to extract visual features of the workpiece surface structure and texture distribution based on the composite image through a pre-trained deep convolutional network, obtain process parameters of the workpiece from a process database, and simultaneously detect geometric distribution parameters of pre-calibrated key areas of the workpiece in the image. The module normalizes and weights the above visual features, process parameters, and geometric distribution parameters to generate a comprehensive state vector for clamping decision-making, wherein the process parameters include at least: material mechanical parameters and cutting process parameters; MLP model module 30: This module is used to input the comprehensive state vector into the pre-trained MLP decision model, output the target clamping force adjustment coefficient, retrieve the corresponding minimum and maximum process clamping force from the process database based on the workpiece material mechanical parameters and cutting process parameters in the vector, and convert the adjustment coefficient into the target clamping force through a linear mapping relationship. The MLP model learns its decision-making behavior from the clamping decision strategy network trained based on the DDPG algorithm through the knowledge distillation method. The main clamping stage is the working stage during the stable machining process of the boring and milling machine when the clamping system applies and maintains the target clamping force; Optimal clamping module 40: This module is used to generate a corresponding target current instruction based on the obtained target clamping force and the preset pressure-current mapping relationship. This instruction drives the electro-hydraulic proportional valve to adjust the output force of the hydraulic cylinder, driving the clamping mechanism to operate and complete the workpiece clamping; A disturbance information acquisition module 50 is used to simultaneously collect the vibration signal, temperature signal, and hydraulic oil pressure signal of the clamping system during the main clamping phase, perform time domain feature extraction on the vibration signal and temperature signal in units of a preset time window, obtain vibration time domain features and temperature time domain features, calculate the average value of the hydraulic oil pressure within the window, convert the hydraulic oil pressure into the current output force of the hydraulic cylinder based on the known effective area of ​​the piston, input the current output force into a calibrated force transfer model, obtain the total effective clamping force on the workpiece, and calculate the clamping force deviation between the target clamping force and the total effective clamping force; A current disturbance observation vector module 60 is used to combine the extracted vibration time domain features and temperature time domain features with the clamping force deviation to construct a normalized current multimodal disturbance observation vector in the form of [vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, clamping force deviation]; The hydraulic compensation factor module 70 is used to input the current multimodal disturbance observation vector into a pre-built fuzzy rule model, generate a synthetic fuzzy set of the hydraulic compensation factor through fuzzification, rule activation, implication operation and maximum aggregation, perform defuzzification using the centroid method within the hydraulic compensation factor domain [0, 1], calculate the geometric center of the synthetic fuzzy set, and use the calculated result as the output value of the hydraulic compensation factor, wherein the fuzzy rule model is constructed by a fuzzy decision tree; The compensation force module 80 is used to calculate the current correction value of the electro-hydraulic proportional valve based on the clamping force deviation between the target clamping force and the total effective clamping force and the output value of the hydraulic compensation factor. The current correction value is added to the target current command and used as the control signal of the electro-hydraulic proportional valve to dynamically adjust the valve core opening, generate the required hydraulic oil pressure, and maintain a stable clamping force of the clamping system during processing. , in, is the current correction value, is the clamping force deviation, is the output value of the hydraulic compensation factor, For electricity The slope of the pressure-current characteristic curve of the hydraulic proportional valve.

[0056] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-parameter collaborative control method for the clamping system of a precision boring and milling machine, characterized in that: include: The workpiece is pre-clamped and fixed on the fixture. The workpiece surface image is captured by an industrial camera. The captured image is processed by SIFT feature matching algorithm and illumination normalization to form a composite image for feature extraction. Based on the composite image, the visual features of the workpiece surface structure and texture distribution are extracted through a pre-trained deep convolutional network, and the process parameters of the workpiece are obtained from the process database. At the same time, the geometric distribution parameters of the pre-calibrated key areas of the workpiece in the image are detected. The above visual features, process parameters and geometric distribution parameters are normalized and weighted to generate a comprehensive state vector for clamping decision-making. The process parameters include at least: material mechanical parameters and cutting process parameters; The integrated state vector is input into a pre-trained MLP decision model, which outputs the target clamping force. The MLP model learns its decision-making behavior from a clamping decision strategy network trained using the DDPG algorithm through knowledge distillation. The main clamping phase is the phase in which the clamping system applies and maintains the target clamping force during the stable machining process of the boring and milling machine. Based on the target clamping force obtained, the controller generates a corresponding target current command according to the preset pressure-current mapping relationship. This command drives the electro-hydraulic proportional valve to adjust the output force of the hydraulic cylinder, driving the clamping mechanism to complete the workpiece clamping. During the main clamping phase, the vibration signal, temperature signal, and hydraulic oil pressure signal of the clamping system are collected simultaneously. Time domain feature extraction is performed on the vibration signal and temperature signal in units of a preset time window to obtain the vibration time domain features and temperature time domain features. The average value of the hydraulic oil pressure within the window is calculated. Based on the known effective area of ​​the piston, the hydraulic oil pressure is converted into the current output force of the hydraulic cylinder. The current output force is input into a calibrated force transfer model to obtain the total effective clamping force on the workpiece, and the clamping force deviation between the target clamping force and the total effective clamping force is calculated. The extracted vibration time domain features and temperature time domain features are combined with the clamping force deviation to construct a normalized current multimodal disturbance observation vector in the form of [vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, clamping force deviation]; The current multimodal disturbance observation vector is input into the pre-built fuzzy rule model. After fuzzification, rule activation, implication operation and maximum aggregation, a synthetic fuzzy set of hydraulic compensation factors is generated. The center of gravity method is used for defuzzification within the domain of hydraulic compensation factors [0, 1]. The geometric center of the synthetic fuzzy set is calculated and the calculated result is used as the output value of the hydraulic compensation factor. The fuzzy rule model is constructed by a fuzzy decision tree. According to the clamping force deviation between the target clamping force and the total effective clamping force and the output value of the hydraulic compensation factor, the current correction value of the electro-hydraulic proportional valve is calculated and superimposed on the target current instruction. Together, they serve as the control signal of the electro-hydraulic proportional valve, thereby dynamically adjusting the valve core opening to generate the required hydraulic oil pressure and maintain a stable clamping force of the clamping system during processing. ， in, is the current correction value, is the clamping force deviation, is the output value of the hydraulic compensation factor, For electricity The slope of the pressure-current characteristic curve of the hydraulic proportional valve.

2. The multi-parameter coordinated control method for the clamping system of a precision boring and milling machine according to claim 1, characterized in that: Forming a composite image for feature extraction, including: Two industrial cameras are placed above the workpiece clamping area. Using the mechanical coordinate system of the workpiece clamping area as a reference, the optical axis of the first camera is located in the XZ plane at a 45° angle to the Z axis, and the optical axis of the second camera is located in the YZ plane at a 45° angle to the Z axis. The intersection of the optical axes of the two cameras is aligned with the geometric center of the workpiece clamping area to obtain multi-angle images of the workpiece surface. The workpiece surface image captured by the first camera is image A, and the workpiece surface image captured by the second camera is image B. The key points of image A and image B are extracted using the SIFT algorithm and initially matched. The RANSAC algorithm is then combined with the homography model to estimate the optimal homography matrix from the initial matching point pairs. Mismatched points are removed at the same time, and the matrix is ​​used to perform perspective transformation on the images to obtain a spliced ​​composite image. The spliced ​​composite image is subjected to illumination normalization processing and converted from RGB color space to HSV color space. The hue, saturation and lightness channels are separated. The CLAHE algorithm is applied to the lightness channel to suppress the influence of uneven lighting. The highlight area is further compressed pixel by pixel through nonlinear mapping method to reduce the interference of mirror reflection on the metal surface of the workpiece. The processed lightness channel is merged with the original hue and saturation channels and converted back to RGB color space to obtain a normalized composite image for subsequent feature extraction.

3. The multi-parameter coordinated control method for the clamping system of a precision boring and milling machine according to claim 1, characterized in that: The MLP model learns its decision-making behavior from the clamping decision policy network trained with the DDPG algorithm through the knowledge distillation method, including: A digital twin simulation environment for the clamping system of a boring and milling machine is constructed. The simulation environment includes a geometric data layer and a functional model layer. The geometric data layer includes the workpiece CAD geometric model and the fixture CAD geometric model. The functional model layer includes an assembly analysis module, a clamping mechanism dynamics model, a workpiece mechanical response model, a material contact characteristic model, a virtual visual perception module, and an intelligent clamping decision model. The assembly analysis module constructs the fixture-workpiece assembly by using the assembly constraint relationship between the workpiece CAD geometric model and the fixture CAD geometric model. It uses the geometric contact detection algorithm to identify the clamping contact points on the assembly and assign numbers to each contact point. It extracts the three-dimensional coordinates of each contact point and its surface normal vector. In combination with the preset fixture functional topology information, it assigns a force transmission path number to each contact point. The above contact point information is used to generate the structural configuration parameters of the clamping system. The dynamic model of the clamping mechanism simulates the dynamic response characteristics of the clamping system composed of a hydraulic cylinder and an electro-hydraulic proportional pressure valve in the main clamping stage. It takes the received target clamping force command as input and outputs the real-time total clamping force generated by the hydraulic cylinder piston rod. Based on the force transmission path number provided by the assembly analysis module and the finite element analysis of the fixture-workpiece assembly, the equivalent stiffness coefficient of each force transmission path is determined. The real-time total clamping force is distributed to each path according to the stiffness ratio to form the clamping force and its distribution of the real-time total clamping force. At the same time, the instantaneous power value of the clamping system is calculated based on the real-time total clamping force combined with the piston speed of the hydraulic cylinder, and time integration is performed in the main clamping stage to estimate the energy consumption of the clamping system during the main clamping process. The workpiece mechanical response model is constructed based on the clamping force distribution on the fixture-workpiece assembly. The clamping force is applied as a surface load to the corresponding contact points on the workpiece. At the same time, fixed constraints are applied in the assembly constraint relationship area. The finite element model is analyzed to obtain the stress distribution and elastic deformation field of the workpiece under the clamping force, and the maximum local deformation of the workpiece is extracted. The material contact characteristic model stores the contact point friction coefficient when the workpiece and fixture materials are paired. During the simulation process, the maximum total static friction resistance of the clamping system in the sliding direction is calculated, and the total cutting component of the estimated total cutting force in the sliding direction is obtained. The two are compared to obtain the actual safety factor, which is used to evaluate the risk of workpiece slippage. The estimated total cutting force comes from an external cutting force estimation model, which is called before each clamping decision. The cutting process parameters of the workpiece are called from the process database to estimate the total cutting force of the boring and milling machine. A virtual visual perception module is built in the digital twin simulation environment to generate visual features consistent with the actual extraction. It is used in the training phase of the clamping decision strategy to learn the ability to respond to changes in the surface state of the workpiece. The intelligent clamping decision model is built based on the DDPG algorithm. The process of DDPG training reinforcement learning policy network is as follows: S1: Randomly initialize the parameters of the Actor strategy network and the Critic evaluation network, and initialize their corresponding target networks; S2: For each time step (t), input the current integrated state vector S(t) to the Actor strategy network, which is based on The output target clamping force F(t) is compared and constrained with the minimum safe clamping force and the maximum safe clamping force allowed under the current working conditions retrieved from the process database before output, ensuring that the generated target clamping force instruction is always within the safe clamping force window. The target clamping force is converted into a target clamping force instruction and input into the digital twin simulation environment. S3: Based on this instruction, the digital twin simulation environment runs a simulation step. After the simulation is completed, it generates the next comprehensive state vector S(t+1). At the same time, based on the simulation results, it calculates the value of the multi-objective reward function and obtains the reward value R(t) of S(t). The simulation results include the maximum local deformation of the workpiece, the actual safety factor, and the energy consumption of the clamping system. S4: Take the four-tuple (S(t), F(t), R(t), S(t+1)) generated by this interaction as an experience sample and store it in the experience feedback Put the buffer, S5: Randomly sample small batches of quads from the experience replay buffer and first update the parameters of the critic evaluation network. To minimize its prediction error, the Actor parameter policy gradient is calculated based on the updated Critic network to guide the Actor policy network The direction of network parameter update is to gradually optimize the strategy to improve the long-term cumulative reward, and to synchronously update the Actor strategy network and Critic Evaluate the target network parameters of the network to improve training stability, S6: Take S(t+1) as the new current integrated state vector and repeat steps S2-S5 until the reinforcement learning policy network Converge and obtain a trained clamping decision strategy network; The MLP model is obtained by migrating the decision-making ability of the clamping decision strategy network trained and converged by the DDPG algorithm mentioned above through the knowledge distillation method.

4. The multi-parameter coordinated control method for the clamping system of a precision boring and milling machine according to claim 3, characterized in that: The MLP model is obtained by transferring the decision-making capabilities of the clamping decision strategy network trained and converged by the DDPG algorithm through the knowledge distillation method, including: Run the trained clamping decision strategy network in the constructed digital twin simulation environment to generate multiple comprehensive state vectors S i , and collect the target clamping force adjustment coefficient α output by the clamping decision strategy network in this state i ; The state-action pair (S i , F i ), construct a distillation dataset D={(S i , α i )|i=1, 2, ..., M}, and perform normalization preprocessing on all state vectors; The MLP model structure is designed according to the dimension of the integrated state vector. The number of neurons in the input layer is consistent with the dimension of the integrated state vector. The hidden layer uses the ReLU activation function, and the output layer is a single neuron linear unit, which is used to predict the target clamping force. Using the distilled dataset D as training samples, a supervised learning approach is used to minimize the mean squared error between the MLP prediction output and the clamping decision policy network output, training the MLP model to imitate the policy decision-making ability of DDPG. The trained MLP model is quantized and compressed using INT8, and the compressed lightweight MLP model is deployed to the industrial computer of the boring and milling machine.

5. The multi-parameter coordinated control method for the clamping system of a precision boring and milling machine according to claim 1, characterized in that: Calibrated force transfer models, including: Obtain the material, size, and hardness information of the workpiece, process a calibration block with the same properties, and install a force sensor at the bottom of the calibration block; The calibration block is installed in the fixture, and the controller sends N-level current signals. At each level, the hydraulic oil pressure P and the total effective clamping force monitored by the force sensor are synchronously collected. The hydraulic oil pressure P is combined with the known effective area A of the cylinder piston to calculate the current output force of the hydraulic cylinder in real time. , repeat the collection M times at each level, take the average value, and get a data set of N points ; Directly fit the linear model between the current output force of the hydraulic cylinder and the total effective clamping force to establish the overall force transmission model. , j = 1, 2, ..., N, Use the least squares method to find the parameters and , is the total mechanical efficiency of the clamping system, < 1, is the total force loss of the clamping system; Will and Packaged as a force transfer model, , solidify the model parameters, and The force transmission model is written into the non-volatile memory of the controller and marked as a calibrated model.

6. The multi-parameter coordinated control method for the clamping system of a precision boring and milling machine according to claim 1, characterized in that: Pre-built fuzzy rule models, including: The historical operating time series data of a precision boring and milling machine was obtained. This data covered both normal and various fault conditions. A sliding window was used to extract features from this historical data, including vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, and the hydraulic oil pressure window mean. Based on the hydraulic oil window mean and the known effective piston area, the hydraulic cylinder output force was calculated and considered the total effective clamping force. This was then subtracted from the set clamping force to obtain the clamping force deviation. The sample values ​​of the five features (vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, and clamping force deviation) are combined in a fixed order to form a single multimodal disturbance observation sample vector. After completing the above operation on all historical data, all the obtained sample vectors are normalized to obtain a set of multimodal disturbance observation sample vectors for training. At the same time, experts label each multimodal disturbance observation sample vector in the sample vector set with its corresponding hydraulic compensation factor value based on the fault severity and compensation requirements. The labeling value range is [0, 1]. The set consisting of all sample vectors and their hydraulic compensation factor values ​​constitutes the training data set. A linguistic value is defined for each feature in the multimodal disturbance observation sample vector, and the linguistic values ​​include low, medium, and high. At the same time, the linguistic values ​​of the hydraulic compensation factor are defined, including zero, small, medium, and large. Each linguistic value of each of the above features and the linguistic value of each hydraulic compensation factor are configured with a corresponding membership function, which adopts a triangular or trapezoidal distribution. The parameters of each membership function are initialized to generate the corresponding initial fuzzy set for each of the above linguistic values, and the global parameters of the fuzzy decision tree are preset. The process of building a tree structure using fuzzy decision tree: S1: Starting from the root node, traverse the five features of the multimodal perturbation observation sample vector. For each feature, traverse each of its language values ​​and use each feature-language value combination as a candidate split condition to be evaluated. S2: For each candidate split condition, extract the sample values ​​of all sample vectors in the training data set on the feature, and substitute each sample value into the membership function corresponding to the language value to calculate the membership degree of the language value. Based on the membership degree calculated by the sample value, calculate the fuzzy information gain rate of the candidate split condition. S3: At the current node, compare the fuzzy information gain rates of all candidate splitting conditions, select the feature-language value combination with the largest gain rate as the optimal splitting condition for the current node, and generate two child nodes, one belonging to and one not belonging to. For each sample vector and its hydraulic compensation factor value in the training data set, calculate its membership to the language value of the optimal splitting condition based on its data on the feature specified by the optimal splitting condition. The sample vector and its hydraulic compensation factor value carry the weight of the membership size and enter the child node belonging to. At the same time, the sample vector and its hydraulic compensation factor value carry the weight of 1 minus the membership size and enter the child node not belonging to. The sample vector and its hydraulic compensation factor value remain unchanged during the allocation process to form a training data set for each child node. The weights carried are used in all subsequent recursive calculations. Each non-leaf child node is used as the new current node. The above splitting process is repeated recursively until the preset stop condition is met. The nodes that cannot continue to split become leaf nodes, and a fuzzy decision tree structure with unassigned output is constructed. S4: For each leaf node in the fuzzy decision tree structure, based on the cumulative path weights of all sample vectors in the node and from the root node to the node, the membership of each linguistic value of the hydraulic compensation factor is weightedly aggregated, and the linguistic value with the highest aggregated membership is selected as the output of the leaf node to complete the construction of the fuzzy decision tree; S5: Based on the constructed fuzzy decision tree, traverse each path from the root node to the leaf node, connect the splitting conditions of all non-leaf nodes on the path, that is, the feature-language value combination, through the AND logical operator to form the antecedent of the rule, that is, the IF part. The output language value assigned to the leaf node is used as the consequent of the rule, that is, the THEN part, to form a complete IF-THEN fuzzy rule. The rules corresponding to all paths together constitute the initial fuzzy rule set; The rule set combines the linguistic values ​​of all features in the multimodal disturbance observation sample vector, the linguistic values ​​of the hydraulic compensation factor, and their corresponding membership functions and initial parameters to form an initial fuzzy rule knowledge base; S6: Optimize the initial fuzzy rule base, encapsulate and store the optimized fuzzy rule set and the optimal parameters of the membership function corresponding to each language value in a structured data format to form an optimized fuzzy rule knowledge base. By introducing a fuzzy inference engine and combining it with the knowledge base, a fuzzy rule model that can be embedded in the control system is constructed.

7. The multi-parameter coordinated control method for the clamping system of a precision boring and milling machine according to claim 6, characterized in that: The synthetic fuzzy set of hydraulic compensation factors includes: Substitute each eigenvalue in the current multimodal perturbation observation vector into the membership function of its corresponding {low, medium, high} language value, and calculate the membership degree of the eigenvalue on each language value; Traverse each rule in the optimized fuzzy rule set. If its antecedent is a single condition, the membership of the condition is the activation strength of the rule. If the antecedent contains multiple conditions connected by AND, take the minimum value of the membership of each condition as the activation strength of the rule. For each rule with activation strength greater than 0, obtain the membership function corresponding to the consequent linguistic value, perform min implication or multiplication implication on the membership function on the entire domain of the hydraulic compensation factor, and obtain the output fuzzy set of the activated rule, where the entire domain of the hydraulic compensation factor is [0, 1]. The output fuzzy sets of all activated rules are aggregated on the universe to obtain the composite fuzzy set of hydraulic compensation factors.

8. A multi-parameter coordinated control device for the clamping system of a precision boring and milling machine, characterized in that: The device includes: Composite image module, which is used to pre-clamp the workpiece on the fixture and collect the workpiece surface image through the industrial camera. The collected image is processed by SIFT feature matching algorithm and illumination normalization to form a composite image for feature extraction; A state vector module is used to extract visual features of the workpiece surface structure and texture distribution based on the composite image through a pre-trained deep convolutional network, obtain the workpiece process parameters from the process database, and simultaneously detect the geometric distribution parameters of the pre-calibrated key areas of the workpiece in the image. The module normalizes and weights the above visual features, process parameters, and geometric distribution parameters to generate a comprehensive state vector for clamping decision-making, where the process parameters include at least: material mechanical parameters and cutting process parameters; The MLP model module is used to input the comprehensive state vector into the pre-trained MLP decision model and output the target clamping force. The MLP model learns its decision-making behavior from the clamping decision strategy network trained based on the DDPG algorithm through knowledge distillation. The main clamping phase is the working stage during which the clamping system applies and maintains the target clamping force during the stable machining process of the boring and milling machine; The optimal clamping module is used to generate a corresponding target current command based on the obtained target clamping force and the preset pressure-current mapping relationship. This command drives the electro-hydraulic proportional valve to adjust the output force of the hydraulic cylinder, driving the clamping mechanism to complete the workpiece clamping. The disturbance information acquisition module is used to simultaneously collect the vibration signal, temperature signal and hydraulic oil pressure signal of the clamping system during the main clamping phase. The module extracts the time domain features of the vibration signal and temperature signal in units of a preset time window to obtain the vibration time domain features and temperature time domain features, and calculates the average value of the hydraulic oil pressure within the window. Based on the known effective area of ​​the piston, the hydraulic oil pressure is converted into the current output force of the hydraulic cylinder. The current output force is input into the calibrated force transfer model to obtain the total effective clamping force on the workpiece, and the clamping force deviation between the target clamping force and the total effective clamping force is calculated; Current disturbance observation vector module, which is used to combine the extracted vibration time domain features and temperature time domain features with the clamping force deviation to construct a normalized current multimodal disturbance observation vector in the form of [vibration RMS, vibration kurtosis, hydraulic oil temperature change rate, fixture temperature change rate, clamping force deviation]; The hydraulic compensation factor module is used to input the current multimodal disturbance observation vector into the pre-built fuzzy rule model, generate a synthetic fuzzy set of the hydraulic compensation factor through fuzzification, rule activation, implication operation and maximum aggregation, and use the centroid method to perform defuzzification within the hydraulic compensation factor domain [0, 1]. The geometric center of the synthetic fuzzy set is calculated and the calculated result is used as the output value of the hydraulic compensation factor. The fuzzy rule model is constructed by a fuzzy decision tree. The compensation force module is used to calculate the current correction value of the electro-hydraulic proportional valve based on the clamping force deviation between the target clamping force and the total effective clamping force and the output value of the hydraulic compensation factor. The current correction value is added to the target current instruction and used as the control signal of the electro-hydraulic proportional valve to dynamically adjust the valve core opening, generate the required hydraulic oil pressure, and maintain a stable clamping force of the clamping system during processing. ， in, is the current correction value, is the clamping force deviation, is the output value of the hydraulic compensation factor, For electricity The slope of the pressure-current characteristic curve of the hydraulic proportional valve.

Images