Robotic blade constant force polishing impedance control method with fusion wear compensation

Abstract

The application relates to the technical field of robot polishing, and particularly discloses a robot blade constant force polishing impedance control method fusing wear compensation, which comprises the following steps: acquiring polishing state data, dividing the working surface of a polishing tool into multiple independent grid units, establishing a contact area geometric model according to the blade surface curvature of a polishing contact point and the effective radius of the polishing tool, determining a target grid unit in the contact area, and distributing the total normal force measured in real time to the target grid unit. In the actual polishing operation of the robot, the three-dimensional wear morphology is reconstructed in real time by dividing the surface of the polishing tool into grid units and combining model error correction, dynamic three-dimensional geometric compensation can be performed on the tool center point coordinates when the polishing track is executed, and the size deviation caused by non-uniform wear of the tool is reduced; meanwhile, the cutting ability attenuation caused by abrasive passivation is quantified as an equivalent cutting force feature.

Description

Technical Field

[0001] This invention relates to the field of robotic polishing technology, and in particular to a method for controlling the impedance of constant force polishing of robotic blades that integrates wear compensation. Background Technology

[0002] In the automated grinding process of complex curved surface parts such as aero-engine blades, industrial robots typically employ a constant force impedance control method to maintain the normal contact force between the grinding tool and the workpiece surface. Existing control systems mainly rely on pre-set empirical process parameters and the initial theoretical geometry of the grinding tool, combined with force data fed back from sensors, to adjust the robot's motion trajectory and impedance parameters.

[0003] In actual grinding operations, the surface of grinding tools will experience non-uniform physical wear. As the processing progresses, the effective geometry of the grinding tool changes, and the surface abrasive grains gradually become dull, causing the actual contact area, local cutting pressure distribution, and material removal capacity to continuously deviate from the initial settings.

[0004] Because existing impedance control methods fail to perceive and quantify the dynamic degradation characteristics of this geometry and cutting state, the system still outputs constant control commands according to the theoretical state, which makes the actual contact state between the grinding tool and the blade surface inconsistent with expectations. This results in deviations in the profile accuracy of the processed blade, and technical problems such as out-of-tolerance surface dimensions and unstable processing quality. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a constant-force grinding impedance control method for robotic blades that integrates wear compensation, in order to ensure the stability of material removal and the machining accuracy of the overall curved surface profile during blade processing.

[0006] To achieve the above objectives, a first aspect of the present invention proposes a robot blade constant force grinding impedance control method integrating wear compensation, used to control a robot to perform grinding operations, comprising: acquiring grinding state data; dividing the working surface of the grinding tool into multiple independent grid cells; establishing a contact area geometric model based on the blade surface curvature at the grinding contact point and the effective radius of the grinding tool; determining the target grid cell within the contact area; and distributing the real-time measured total normal force to the target grid cell; constructing a physical wear basis model based on contact pressure and sliding velocity for the target grid cell to calculate the nominal wear amount, and combining the Gaussian process error based on the working surface partitioning. The model is modified to obtain the final wear amount of each target mesh cell through uncertainty-weighted fusion, and the three-dimensional wear morphology of the grinding tool is reconstructed in real time. Geometric compensation of the grinding tool center point is performed based on the three-dimensional wear morphology, and a multi-dimensional composite cost function including contact force safety and wear uniformity is constructed. The optimal combination of process parameters corresponding to the current contact area is generated through a preset optimization algorithm. The equivalent cutting force is calculated based on the abrasive passivation coefficient of the target mesh cell, and the force error compensation value is output in combination with the pre-trained network model to generate the final target contact force. The optimal combination of process parameters and the final target contact force are constrained by a preset saturation function and then input into the robot's impedance controller.

[0007] To achieve the above objectives, a second aspect of the present invention proposes a constant-force grinding impedance control system for robot blades with integrated wear compensation, used to control a robot to perform grinding operations. The system includes: a data acquisition and mesh allocation module, used to acquire grinding state data, divide the working surface of the grinding tool into multiple independent mesh units, establish a geometric model of the contact area based on the blade surface curvature at the grinding contact point and the effective radius of the grinding tool, determine the target mesh unit within the contact area, and allocate the real-time measured total normal force to the target mesh unit; and a wear prediction and morphology reconstruction module, used to construct a physical wear model based on contact pressure and sliding velocity for the target mesh unit to calculate the nominal wear amount, combined with a Gaussian process based on the partitioning of the working surface. An error correction model obtains the final wear amount of each target mesh cell through uncertainty-weighted fusion, and reconstructs the three-dimensional wear morphology of the grinding tool in real time. A geometric compensation and parameter optimization module performs geometric compensation of the grinding tool center point based on the three-dimensional wear morphology, and constructs a multi-dimensional composite cost function that includes contact force safety and wear uniformity. It generates the optimal combination of process parameters corresponding to the current contact area through a preset optimization algorithm. An impedance control execution module calculates the equivalent cutting force based on the abrasive passivation coefficient of the target mesh cell, and outputs the force error compensation value in combination with the pre-trained network model to generate the final target contact force. The optimal combination of process parameters and the final target contact force are then input into the robot's impedance controller after being constrained by a preset saturation function.

[0008] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein when the computer program is executed by the processor, it implements the above-described method for controlling the impedance of constant force grinding of robot blades with integrated wear compensation.

[0009] In the actual grinding operation of the robot, the technical solution of this invention divides the surface of the grinding tool into grid cells and combines model error correction to reconstruct the three-dimensional wear morphology in real time. This enables dynamic three-dimensional geometric compensation of the tool center point coordinates when executing the grinding trajectory, reducing dimensional deviations caused by non-uniform tool wear. At the same time, this solution quantifies the attenuation of cutting ability caused by abrasive passivation into equivalent cutting force characteristics, and optimizes process parameters based on a cost function that includes contact force boundaries and wear uniformity constraints. This allows the robot's impedance controller to dynamically output control commands and target contact forces that match the current actual force characteristics under the condition that the geometry and cutting state of the grinding tool continue to decline, ensuring the stability of material removal and the machining accuracy of the overall curved surface profile during blade processing. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the constant force grinding impedance control method for robot blades with integrated wear compensation provided by the present invention. Figure 2 This is a schematic diagram illustrating the implementation of the wear compensation-integrated constant force grinding impedance control system for robot blades provided by the present invention. Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0011] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0012] The following describes, with reference to the accompanying drawings, an electronic device for a robot blade constant force grinding impedance control system and method with integrated wear compensation according to an embodiment of the present invention. Example

[0013] like Figure 1 As shown, this embodiment provides a constant force grinding impedance control method for robot blades that integrates wear compensation, used to control the robot to perform grinding operations.

[0014] In the aerospace manufacturing field, for the processing of complex spatial curved surface parts such as titanium alloy blades for aero engines, a six-degree-of-freedom or seven-degree-of-freedom industrial robot equipped with a constant force floating spindle or directly equipped with a six-dimensional force sensor is usually used to form an automated grinding system.

[0015] The control method in this embodiment relies on this type of underlying hardware platform and is executed through real-time control software in a host industrial control computer. The underlying hardware platform collects the robot's joint positions, poses, spindle current, and force and torque signals fed back from the six-dimensional force sensor in real time, and transmits them to the main control computer via a high-speed industrial bus such as EtherCAT. This control method aims to solve the problems of effective geometric dimension changes and cutting capability degradation caused by uneven material wear during long-term contact operations of grinding tools, thereby ensuring the contour accuracy of the blade surface and the stability of the material removal rate.

[0016] Specifically, the control method, in terms of both time and logical hierarchy, progresses according to the system's operating cycle, sequentially including core operational stages such as data acquisition and mapping, wear reconstruction and prediction, strategy optimization and compensation, and impedance control execution. To clearly demonstrate the implementation path of this technical solution, the overall solution is broken down into several sequential steps with strict causal logic for detailed explanation below.

[0017] Step 1: Grinding status data acquisition, mesh generation, and contact area geometric mapping.

[0018] Specifically, in the first cycle of the control system initialization and in each subsequent real-time interpolation cycle, the system first performs the following basic steps: acquiring grinding status data, dividing the working surface of the grinding tool into multiple independent grid units, establishing a geometric model of the contact area based on the curvature of the blade surface at the grinding contact point and the effective radius of the grinding tool, determining the target grid unit within the contact area, and distributing the real-time measured total normal force to the target grid unit.

[0019] It is also important to note that the surface of grinding tools, such as belt wheels, rubber contact wheels, or bonded abrasive wheels, is a continuous cylindrical or toroidal geometry. To digitally model the microscopic wear evolution, this continuous surface must be discretized. The working surface of the grinding tool is divided into a three-dimensional mesh unit sequence along the circumference at preset equal arc length intervals and along the axial direction at preset equal distance intervals. In practical engineering applications, the preset equal distance interval along the axial direction can be set to 0.5 mm, and the circumferential direction can be divided into 360 units. This level of discretization allows for control of the dimension of the system matrix operations while ensuring sufficient spatial resolution, avoiding timing timeouts in the control cycle. Each tiny region after partitioning is defined as a mesh unit, which exists in computer memory as elements of a two-dimensional feature matrix.

[0020] Optionally, at the instant the grinding tool contacts the blade, due to the local elastic properties of the tool matrix or blade material, the contact between the two is not an ideal line or point contact, but rather a tiny contact surface is formed. Based on the blade surface curvature and the effective radius of the grinding tool at the current grinding contact point, the contact deformation is calculated using Hertzian contact theory, generating an elliptical geometric model of the contact area. Hertzian contact theory is mainly used to describe the local deformation of two elastic bodies with curvature under compression. By extracting the principal curvature of the blade surface at the current interpolation point and the current effective radius of the grinding tool, the system calculates the downward deformation at the contact center, and then derives an elliptical geometric model of the contact area characterized by the major and minor axes on the tangential plane. This model accurately depicts the spatial boundary of the actual physical interaction at the current moment.

[0021] Furthermore, the system maps the elliptical geometric model of the contact area to the local coordinate system of the working surface of the grinding tool, filters out the target mesh cells within the contact range, and distributes the total normal force measured in real time by the six-dimensional force sensor to each target mesh cell according to the contact area ratio to generate instantaneous contact pressure. The target mesh cells specifically refer to those mesh cells whose spatial coordinates fall within the boundary of the aforementioned elliptical geometric model of the contact area during the current sampling period. Since the six-dimensional force sensor can only feed back the macroscopic total normal force acting on the entire grinding tool, the system needs to calculate the effective projected area of ​​each target mesh cell within the contact ellipse using a geometric intersection algorithm, and use the area ratio as a weighting coefficient to decompose the macroscopic total force into microscopic distributed forces, which are then divided by the area of ​​a single mesh cell to obtain the instantaneous contact pressure. This process achieves a high-fidelity mapping from macroscopic mechanical characteristics to a microscopic mesh physical field.

[0022] Step 2: Construct a physical wear model to calculate nominal wear.

[0023] Specifically, after the microscopic distribution of force is completed, the system enters the wear evolution assessment stage. At this time, it is necessary to construct a physical wear model based on contact pressure and sliding velocity for the target mesh element to calculate the nominal wear amount. The nominal wear amount refers to the theoretical characterization value of the volume loss of the grinding wheel surface calculated solely based on the principle of frictional work under ideal conditions, without considering random defects inside the material, non-uniform distribution of cutting fluid, and external uncertain disturbances such as high-frequency vibration of the system.

[0024] For example, this step specifically includes establishing an independent wear model in the form of a modified Preston equation for each of the target mesh cells. The classic Preston equation assumes that the amount of material wear is proportional to the product of contact pressure and relative sliding velocity. However, in constant-force grinding applications of titanium alloys, due to the nonlinear abrasive fragmentation mechanism under high stress, the simple linear proportional relationship is difficult to apply.

[0025] Therefore, the initial wear coefficient of the target mesh element is obtained through offline calibration experiments. Offline calibration experiments refer to the testing and grinding of a test block with standard hardness and roughness using brand-new grinding tools of the same batch and material, applied with a constant gradient contact force and multiple stepped spindle speeds, before the actual blade is machined. Subsequently, a laser 3D profile scanner is used to measure the volume loss on the tool surface, and the initial parameters reflecting the inherent wear resistance characteristics of the tool matrix are obtained through back-fitting using the least squares method.

[0026] Based on the initial wear coefficient, the product of the instantaneous contact pressure and instantaneous friction velocity of the target mesh element is integrated along the time dimension, and a nonlinear mapping is performed in conjunction with the material removal pressure index to output the nominal wear amount of the target mesh element. In this step, to quantify this nonlinear accumulation process, the system background calls the following defined physical wear integral formula algorithm: In the formula: Defined as coordinate index OK The target grid cell of the column at the current control time Nominal wear and tear; Defined as coordinate index OK The initial wear coefficient corresponding to the target mesh cell in the column; Defined as coordinate index OK The target grid cell of the column at the historical evolution variable time The instantaneous contact pressure it withstands; Defined as coordinate index OK The target grid cell of the column at the historical evolution variable time The surface linear velocity compared to the instantaneous friction velocity of the workpiece; Defined as the material removal pressure index associated with the physical properties of the blade material to be processed, it is used to regulate the nonlinear gain effect of the pressure-velocity coupling term on wear.

[0027] Through this integral calculation, the control system can track the accumulated frictional work process of each tiny grid cell in the continuous time domain, thereby generating a basic loss depth field distribution at the theoretical level.

[0028] Step 3: Reconstruct the Gaussian process error based on partitioning and the final wear amount.

[0029] Specifically, while the physical model explains the basic mechanical wear laws, it cannot cover the random variations in actual grinding conditions. Therefore, the control method combines a Gaussian process error correction model based on the partitioning of the working surface, and obtains the final wear amount of each target mesh cell through uncertainty-weighted fusion, thereby reconstructing the three-dimensional wear morphology of the grinding tool in real time.

[0030] It is also important to note that the working surface of the grinding tool is divided axially into five wear characteristic zones: the left edge zone, the left transition zone, the center zone, the right transition zone, and the right edge zone. In actual blade grinding, the tool's center zone is typically used for large areas of the blade base and the gently sloping back, where the stress is relatively uniform. The edge zone, however, is frequently used for machining sharp-angled areas with drastic curvature changes, such as the leading and trailing edges of the blade, which are highly susceptible to stress concentration, abnormal abrasive fragmentation, and softening of the bonding agent due to high grinding temperatures. Isolating the working surface according to its physical characteristics helps prevent data from different failure mechanisms from interfering with each other during model training.

[0031] Optionally, an independent Gaussian process error correction model is established for each wear characteristic partition, outputting the physical model prediction error and prediction standard deviation. Gaussian process regression is a non-parametric machine learning method based on Bayesian inference. Its advantage lies not only in predicting the specific nonlinear error expectation value, but also in simultaneously outputting the variance characterizing the current prediction confidence. The system extracts measured wear deviations from historical processing data as a training set and calculates the covariance between each dimension of features using a preset kernel function. During online inference, for the currently input working condition features, the model calculates the corresponding physical model prediction error and simultaneously outputs the prediction standard deviation reflecting uncertainty.

[0032] Subsequently, the reciprocal of the prediction standard deviation is calculated as the dynamic fusion weight for the corresponding partition. The nominal wear amount and the physical model prediction error are then weighted and summed based on the dynamic fusion weight to obtain the final wear amount for each target grid cell. A larger prediction standard deviation means a lower confidence level of the machine learning model in predicting the current state. In this case, its reciprocal becomes smaller, automatically reducing the proportion of the error correction term in the total result, and the system will place greater trust in the physical model based on mechanical mechanisms; conversely, a smaller standard deviation indicates a lower confidence level. To clarify the data fusion mechanism, the system executes the following uncertainty-weighted fusion algorithm: In the formula: Defined as coordinate index OK The target grid cell of the column at the current control time Final wear amount after weighted correction; As defined above, this corresponds to the nominal wear amount; Defined as the first to which this grid cell belongs Each wear characteristic zone at the current control moment The standard deviation of the prediction output by the Gaussian process model; Defined as a positive real-number minimal constant used to prevent the denominator from having a zero value; Defined as the physical model prediction error output by the Gaussian process model for this grid cell.

[0033] Through the above steps, the system calculates the thickness loss distributed on all grid nodes on the surface of the grinding tool in real time, and combines it in the form of a three-dimensional point cloud matrix to successfully reconstruct the three-dimensional wear morphology of the grinding tool in real time, providing a high-precision data source for subsequent trajectory intervention.

[0034] Step 4: Execute the triple geometric compensation mechanism based on the three-dimensional wear morphology.

[0035] Specifically, after determining the amount of degradation of the tool's geometric surface, the control system performs geometric compensation of the grinding tool's center point based on the three-dimensional wear morphology. The compensation system adopts a triple geometric compensation mechanism to eliminate undercutting and trajectory deviation caused by the reduction in tool radius.

[0036] For example, these three mechanisms progress layer by layer on a time scale. The first layer is real-time partitioned dynamic compensation: using the final wear amount along the unit normal vector of the blade surface at the current grinding contact point, the coordinates of the initial calibration tool center point of the grinding tool are dynamically updated in real time. In robot kinematic control, the tool center point (TCP) is the core reference origin for trajectory planning. If a specific mesh on the tool surface wears 0.2mm, but the TCP is not modified, the robot will feed 0.2mm less during actual contact. The system extracts the deepest final wear amount of the current force-contact mesh and adds a corresponding feed offset along the blade normal towards the workpiece, so that the actual physical surface re-fits the theoretically planned surface. This step is calculated in real time within each millisecond-level interpolation cycle.

[0037] The second step is three-dimensional morphology calibration between passes: After a single grinding pass, a scanning device generates the measured three-dimensional wear morphology, compares it with the predicted values, and updates the hyperparameters of the Gaussian process error correction model. A pass refers to a complete traversal scan of the entire blade surface or a specified area by the robot. During blade replacement or tool retraction intervals, a fixed laser line scan profiler located on the side of the grinding unit rapidly rescans the grinding tool to capture the true morphology information. The system calculates the residual between the measured and predicted values ​​and calibrates the kernel function scaling parameters and noise variance parameters of the Gaussian process model using gradient descent or marginal likelihood maximization algorithms to ensure higher fidelity in the prediction model for the next pass.

[0038] The third layer involves local correction of critical areas between passes: After a single grinding pass, the critical areas of the blade are scanned to generate local geometric errors, which are then superimposed on the compensation baseline for the next pass. The leading edge, trailing edge, and root fillet of the blade are called critical areas. These areas have extremely large curvature gradients, and the robot's follow-up response delay and the small deformation of the grinding tool can easily accumulate errors due to insufficient machining allowance. The system uses an on-machine probe or optical tracking device to evaluate the measured allowance in these areas and generate an independent local allowance deviation map. When performing the grinding task in the same area for the next time, the robot will use this local geometric error value as feedforward compensation for the underlying space, forcing the tool to further approximate the theoretical surface.

[0039] Step 5: Construction of multidimensional composite cost function and optimization of process parameters through reinforcement learning.

[0040] Specifically, geometric compensation solves the problem of accurate positioning, but as the tool becomes dull, the mechanical and motion parameters of impedance control must also be adaptively adjusted to maintain the same material removal rate. Therefore, the control method constructs a multi-dimensional composite cost function that includes contact force safety and wear uniformity, and generates the optimal combination of process parameters corresponding to the current contact area through a preset optimization algorithm.

[0041] It is important to note that a five-dimensional composite cost function is constructed, which is formed by multiplying the material removal rate error cost term, the contour accuracy error cost term, the contact force safety penalty term, the process parameter change rate penalty term, and the wear uniformity cost term by their respective weighting coefficients and then superimposing them. These five cost terms cover all the constraints that the grinding process needs to consider. The material removal rate error cost term ensures the consistency of processing efficiency; the contour accuracy error cost term constrains the deviation of the follower normal feed; the contact force safety penalty term sets a rigid safety threshold for the force, and once it exceeds the plastic yield or burn threshold of the titanium alloy, its cost value increases exponentially to absolutely prohibit the generation of dangerous parameters; the process parameter change rate penalty term, like mechanical damping, prevents newly generated parameters from jumping too much with the parameters of the previous cycle, causing system resonance.

[0042] The wear uniformity cost term, specifically defined in the aforementioned features, is used to suppress non-uniform wear of the grinding tool. Without restriction, the system might over-rely on a single part of the tool for heavy cutting, causing that area to quickly become unusable while other areas remain underutilized. The specific calculation method is as follows: extract the maximum and average wear amounts from all target mesh cells of the grinding tool at the current moment; calculate the square of the difference between the maximum and average wear amounts; multiply the square of the difference by a preset wear uniformity penalty coefficient to obtain the output value of the wear uniformity cost term. This evaluation mechanism based on the fusion of range and variance prompts the optimization algorithm to actively guide the robot to utilize less worn mesh areas for contact when planning subsequent motion and resistance forces, thereby extending the global lifespan of the grinding tool.

[0043] Optionally, the system employs a reinforcement learning optimization algorithm to search for strategies under the constraints of the five-dimensional composite cost function, outputting the optimal combination of process parameters corresponding to the current contact area. This optimal combination of process parameters includes equivalent cutting force, impedance stiffness, impedance damping, and feed rate. The reinforcement learning optimization algorithm possesses the long-term benefit evaluation capability of a Markov decision process. The system considers the current mesh wear state and blade local curvature as environmental state inputs, and feeds back the negative value of the five-dimensional composite cost function as an immediate reward signal to the agent network. Through tens of thousands of interactive iterations with the internal simulation dynamics model, the agent explores a multi-dimensional continuous action vector that minimizes the overall cost function under the current state, i.e., outputting a completely new set of feed rate and impedance parameter configurations.

[0044] Step 6: Extraction and compensation of equivalent cutting force based on abrasive passivation coefficient.

[0045] Specifically, the optimal parameters obtained through the above optimization include a key mechanical reference vector. The control method calculates the equivalent cutting force based on the abrasive passivation coefficient of the target mesh element. Besides the macroscopic dimensional changes caused by volume loss of material on the surface of the grinding tool, the sharp edges of the abrasive grains adhering to the surface gradually become rounded and blunt with continuous cutting, losing their microscopic sharpness. This is known as the abrasive passivation effect.

[0046] For example, the passivation rate coefficient of the target mesh cell is obtained, and the product of the instantaneous contact pressure and instantaneous friction velocity of the cell is integrated over time. The abrasive passivation coefficient at the current moment is calculated using an exponential decay function. Passivation is an irreversible degradation process dominated by cutting heat and alternating frictional stress, and its rate is directly related to the external energy input. Subsequently, the preset initial target contact force is divided by the exponential power function of the abrasive passivation coefficient to obtain the equivalent cutting force after compensating for the abrasive passivation effect, where the base of the exponential power function is the abrasive passivation coefficient, and the exponent is the removal rate force exponent related to material properties. To intuitively present the core mathematical logic of this step, the following exponential compensation algorithm formula is specifically executed: In the formula: Defined as coordinate index OK The target grid cell of the column at the current control time The corresponding equivalent cutting force; Defined as the preset initial target contact force constant obtained based on the test calibration of a brand new, unworn grinding tool; Defined as coordinate index OK The target grid cell of the column at the current control time The wear passivation coefficient, calculated and evaluated based on integrals, is a real number between zero and one. Defined as a removal rate force exponential constant related to the material properties of the target workpiece. Since the abrasive passivation coefficient is a value that gradually decreases from 1 to near 0 over time, its negative power amplification operation causes the equivalent cutting force to adaptively increase as the tool becomes duller. This aligns with real industrial physics: the duller the tool, the greater the normal force required to press the abrasive grains back into the metal lattice for shear fracture, thus objectively maintaining a stable material removal depth.

[0047] Step 7: Partition evaluation of network model feedforward error compensation and final target contact force generation.

[0048] Specifically, although the above steps calculate the theoretically required equivalent cutting force, the underlying impedance system, when tracking a complex free-form surface, is subject to unpredictable disturbances caused by the robot's joint flexibility, reducer backlash, and dynamic forced vibration of the spindle, resulting in a dynamic static error in the actual following force. Therefore, the control method combines a pre-trained network model to output force error compensation values ​​to generate the final target contact force.

[0049] For example, the pre-trained network model is a partitioned evaluation network, which possesses high-dimensional nonlinear mapping characteristics and can handle coupled multi-source feedback signals. The working surface of the grinding tool is divided into an edge contact area, a transition contact area, and a center contact area, and an independent partitioned evaluation network model is constructed for each contact area. The equivalent cutting force, actual contact force, spindle speed, feed rate, and blade surface curvature of the current contact area are used as input feature vectors and input into the corresponding partitioned evaluation network model. By inputting the deviation between the actual force and the expected equivalent force, combined with kinematic constraint parameters, the partitioned network can identify whether the system is currently in an underdamped oscillation or an overdamped hysteresis state.

[0050] Subsequently, the feedforward force error compensation value for that region is output through the partitioned evaluation network model, and the feedforward force error compensation value is superimposed and corrected with the equivalent cutting force to generate the final target contact force. By introducing this data-driven feedforward compensation component, the impedance controller can adjust the output bias in advance without waiting for error accumulation when facing external abrupt curvature changes (such as the large twist angle transition section of the blade), significantly improving the speed and stability of contact force tracking.

[0051] Step 8: Issue boundary constraints and impedance controllers for the trapezoidal saturation function.

[0052] Specifically, after all parameters and target control quantities have been reconstructed and compensated, directly issuing the original command may exceed the safe bandwidth threshold of the underlying servo driver, causing the motor to stall or the system to shut down for protection. Therefore, the optimal combination of process parameters and the final target contact force are input into the robot's impedance controller after being constrained by a preset saturation function.

[0053] It is important to note that the saturation function mentioned is a trapezoidal saturation function specifically designed for grinding. Traditional hard-truncation limiting functions can cause discontinuities in the first derivative when cutting off over-limit commands, and sudden step spikes in the input signal can easily trigger high-frequency chatter at the end effector of the robotic arm. Therefore, a trapezoidal saturation function specifically designed for grinding, constructed based on the superposition and combination of sinusoidal trigonometric functions, is used. This special function, which incorporates sinusoidal smoothing characteristics, does not instantly flatten the command as it approaches the set boundary, but rather gently suppresses it through gradually increasing curve curvature, maintaining the continuous differentiability of the signal trajectory.

[0054] Optionally, based on preset upper and lower limits of the process parameters corresponding to the current contact area, the parameters in the optimal process parameter combination and the final target contact force are normalized. Because the dimensions and orders of magnitude of stiffness, damping, and mechanical characteristics differ greatly (stiffness is measured in Newtons per meter, typically on the order of tens of thousands; feed rate is measured in millimeters per second, with relatively small values), directly applying a uniform limiting operation lacks numerical stability. The normalization operation maps them to a unified dimensionless interval.

[0055] Finally, the normalized parameters are subjected to nonlinear smoothing boundary limiting constraints and rate of change constraints using the aforementioned grinding-specific trapezoidal saturation function to suppress abrupt changes in process parameters. The constrained parameters are then denormalized and used as control commands input to the robot's impedance controller. Upon receiving the updated stiffness, damping, and TCP bias corrected for topographic deviations, the robot's underlying motion controller adjusts the control law using second-order mass-spring-damped dynamics equations. This allows the robot's end flange to exhibit compliant environmental adaptability, robustly completing high-quality machining of complex surfaces.

[0056] In summary, considering the shortcomings of existing technologies such as reliance on static constant parameters, lack of wear condition perception, and difficulty in handling complex curvature abrupt changes, the technical solution described in this embodiment possesses significant technological advantages. Specifically, existing conventional grinding impedance control systems inevitably face dimensional undercut defects caused by the reduction in tool geometry and slippage vibration marks caused by abrasive passivation during long-term operation.

[0057] This method constructs a fully closed-loop self-sensing and self-decision-making control system by fine-grained meshing of the tool surface, integrating macroscopic machine learning error correction techniques with a microscopic physical evolution model, and superimposing five-dimensional cost constraints and nonlinear feedforward force compensation. This data- and mechanism-driven architecture effectively isolates the risk of dimensional deviations caused by non-uniform tool wear, and can dynamically stabilize the actual equivalent material removal depth throughout the entire lifespan of the tool. From a systems theory perspective, it practically guarantees the surface consistency of high-value aerospace titanium alloy blades, demonstrating high industrial application value and economic benefits.

[0058] Example 2: In the manufacturing of high-end aero-engine components, the automated grinding of complex curved blades made of titanium alloys with low thermal conductivity and high chemical reactivity generates a large amount of grinding heat instantaneously in the contact area due to the intense friction between the high-speed grinding tool and the blade substrate surface. Because titanium alloys have extremely low spatial thermal conductivity, transient heat accumulation is highly likely to occur in the processing area. If the control system cannot identify and intervene in this heat evolution online, severe physical damage to the blade surface, such as thermal phase change layers, microscopic tearing cracks, and residual tensile stress concentration, can easily occur, directly leading to a significant reduction in the fatigue life of the parts or even scrapping them.

[0059] This embodiment, based on the backbone architecture of the constant-force grinding impedance control method for robot blades established in Embodiment 1 above, provides an extended implementation scheme that deeply integrates thermo-mechanical coupling adaptive constraints and regulation. This embodiment directly embeds an analytical thermo-mechanical coupling dynamic adaptive regulation mechanism without external temperature sensors, aiming to ensure that the grinding area is always within the material phase transformation safety boundary.

[0060] Specifically, the thermo-coupling adaptive constraint method described in this embodiment, in terms of the control flow time sequence, begins after the optimal combination of process parameters corresponding to the current contact area is generated through a preset optimization algorithm, and before the optimal combination of process parameters and the final target contact force are input into the robot's impedance controller after being constrained by a preset saturation function. The entire thermo-coupling control process is driven by the real-time interrupt control kernel at the bottom layer of the industrial control computer, and is implemented by breaking it down into multiple discrete calculation steps with strong causal logic and data transmission relationships.

[0061] Specifically, the control system first executes an analytical calculation step for transient frictional heat generation power based on the equivalent cutting force transformation. The system obtains the frictional force component separated from the equivalent cutting force, and defines the product of the frictional force component and the instantaneous sliding velocity of the corresponding mesh cell as the transient frictional heat generation power, which is used to characterize the heat input of the current contact area. In grinding operations, the normal contact force of the grinding tool is mainly consumed to overcome the indentation resistance caused by the elastic-plastic deformation of the material, while the tangential frictional force is the core source of the cutting heat in the machining area.

[0062] The control system utilizes a preset Coulomb friction matrix model or dynamic shear slip equation to accurately extract the tangential friction force component from the equivalent cutting force calculated in Example 1 above. Subsequently, the system extracts the instantaneous sliding velocity of the target mesh element relative to the blade surface within the current interpolation cycle, and uses the product of the two as the total heat energy generated by mechanical power dissipation at the current contact interface per unit time. Thus, real-time high-fidelity analytical reconstruction of the contact heat input is achieved without the need for high-latency external hardware such as infrared thermometers or thermocouples.

[0063] Specifically, the system synchronously or sequentially executes the step of constructing an effective heat dissipation volume index by combining contact geometry and local curvature. The system extracts the contact ellipse area from the geometric model of the contact region and extracts the blade surface curvature at the current grinding contact point. The product of the contact ellipse area, the reciprocal of the blade surface curvature, and a preset spatial thermal conductivity constant is used as the proper noun. After superimposing a smoothing constant, a natural logarithmic nonlinear mapping calculation is performed to generate an effective heat dissipation volume index, which characterizes the heat dissipation capacity of the blade substrate in the current contact region. From a heat transfer perspective, the transient temperature rise in the grinding region depends not only on the amount of heat input but also on the geometric volume of the metal substrate around the contact point that can absorb and conduct heat. The reciprocal of the blade surface curvature physically uniquely corresponds to the equivalent contact radius of the blade in the current grinding posture.

[0064] When the robot grinds the leading edge, trailing edge, or thin-walled tip of a blade, the curvature value increases sharply, causing its reciprocal, the equivalent contact radius, to decrease significantly. This objectively means that the metal material in the processing area is extremely thin, resulting in a substantial reduction in the overall heat dissipation volume. This solution constructs a geometric feature that dynamically characterizes the local three-dimensional heat dissipation capacity of the workpiece under the current processing posture by multiplying the contact ellipse area (representing the two-dimensional heated boundary) with the reciprocal of the curvature (representing the three-dimensional heat conduction depth). To match this geometric feature with the logarithmic temperature gradient field in the heat conduction control law, the system uses it as the base of the logarithmic function for nonlinear spatial mapping, ultimately transforming it into the effective heat dissipation volume index.

[0065] It is also important to note that after quantitatively analyzing the heat input and heat dissipation capacity, the control system performs a dynamic assessment of thermal stress over-limit risk and a safety threshold determination step. The system divides the transient frictional heat generation power by the effective heat dissipation volume index to generate a thermal stress over-limit risk assessment value; then it determines whether the thermal stress over-limit risk assessment value is greater than a preset material phase change thermal safety threshold. The thermal stress over-limit risk assessment value physically corresponds to the heat flux density gradient within the current processing micro-region. When the heat generation power is extremely high and the local heat dissipation volume index is extremely low, this assessment value will exhibit a significant nonlinear step. The material phase change thermal safety threshold is a constant derived from the reverse force-thermal coupling based on the critical temperature of the metal phase change of the material to be processed. For titanium alloy parts, this threshold is set at the thermal risk boundary corresponding to the critical temperature that triggers the transformation of its microstructure from the alpha phase to the beta phase.

[0066] To accurately represent the real-time computational rules of the aforementioned physical variables within the digital control core, the control system executes the following defined thermal coupling risk assessment formula algorithm in the background: In the formula: Defined as the current control time The thermal stress over-limit risk assessment value calculated and output by the system is a dimensionless real-time control variable; Defined as coordinate index OK The target grid cell of the column at the current control time The frictional force component obtained by separation and extraction is measured in Newtons. Defined as coordinate index OK The target grid cell of the column at the current control time The instantaneous friction speed is expressed in meters per second. Defined as the current control time The area of ​​the contact ellipse, calculated using Hertzian contact theory, is in square millimeters. Defined as the current control time The reciprocal of the curvature of the blade surface at the grinding contact point is the only characteristic of the local radius of curvature, and its unit is millimeters. Defined as a preset spatial thermal conductivity constant, it is used to participate in the calculation of the proper terms in the natural logarithm mapping operation to ensure numerical stability. It is a positive real constant obtained by calibration through the thermal diffusivity coefficient of titanium alloy.

[0067] This formula combines heat generation power with geometric heat dissipation capacity in an analytical manner, enabling the control system to make accurate quantitative predictions of the temperature rise risk at the processing interface within each microsecond-level interruption cycle.

[0068] Specifically, the control system executes branch control based on the judgment result. If the thermal stress over-limit risk assessment value is not greater than the material phase change thermal safety threshold, it indicates that the current area has good heat dissipation conditions or low heat generation, and the processing is safe. In this case, the system maintains the optimal process parameter combination and enters the subsequent saturation function constraint step without changing the control state. If the thermal stress over-limit risk assessment value is greater than the material phase change thermal safety threshold, it is determined that there is a risk of instantaneous burn damage at the current contact interface. The system immediately switches to the limit protection branch, and the feed rate and impedance damping extracted from the optimal process parameter combination are uniquely defined as the initial optimal feed rate and initial optimal impedance damping, respectively, and the thermodynamic reconstruction mechanism is activated.

[0069] Specifically, in the thermal reconstruction mechanism, the system first executes a feed rate reduction correction step based on a negative exponential decay function. The thermal reconstruction mechanism specifically involves: using a preset negative exponential decay function to downward correct the initial optimal feed rate, generating a thermal boundary-limited feed rate to increase the dwell time of the grinding tool in the current contact area for heat dissipation. Conventional control methods typically choose to accelerate when facing high temperatures, but in thin-walled blade grinding, acceleration leads to a further sharp increase in cutting power.

[0070] This solution employs a speed reduction strategy. By introducing a negative exponential decay function, the robot's spatial dragging speed is forcibly reduced when thermal risk increases. Although this operation lengthens the local processing time, the reduced feed rate directly leads to a significant decrease in the amount of material removed per unit time and the power generated by friction. Simultaneously, it extends the residence time of the external cutting fluid or the grinding tool's own adaptive air cooling system in the current area, allowing accumulated heat to be rapidly conducted and dissipated through the workpiece substrate to the unprocessed area, thereby achieving effective control of the processed surface temperature.

[0071] Optionally, the system synchronously executes an impedance damping adaptive increase step based on a logarithmic rigid compensation function. Specifically, this includes: using a preset logarithmic rigid compensation function to upward correct the initial optimal impedance damping, generating thermal boundary-limited impedance damping to suppress mechanical chatter caused by feed rate reduction. In robot impedance control, the dynamic contact behavior between the robotic arm end effector and the workpiece greatly depends on the drag feed speed. When the feed rate is forcibly reduced in the above steps for heat dissipation, the robotic arm end effector easily slides from a relatively smooth continuous cutting state into a stick-slip friction state where the contact surfaces are mutually sticky. This sudden speed drop easily excites high-frequency chatter at the end effector due to insufficient low-frequency joint stiffness, leaving permanent macroscopic vibration marks on the blade surface. This solution, while reducing the speed, significantly increases the damping coefficient of the impedance loop through a logarithmic rigid compensation function. Utilizing the reconstructed high damping characteristics, it forcibly absorbs and dissipates the dynamic torque oscillation energy caused by deceleration, maintaining the dynamic stability of the contact interface.

[0072] To clarify the rules for digital modification of process parameters in the aforementioned thermal reconfiguration mechanism, the control system internally executes a set of adaptive reconfiguration formula algorithms for process parameters as defined below: In the formula: Defined as the current control time The thermal boundary-limited feed rate, generated after thermal reconstruction calculation, is used as the desired linear velocity command to be finally issued to the robot's motion axis, and its unit is millimeters per second; Defined as the current control time The initial optimization feed rate, extracted from the original optimal combination of process parameters, is expressed in millimeters per second. Defined as a preset negative exponential decay constant, it is used to control the rate of decrease in feed rate as the thermal risk exceeds the limit. As defined above, this is the thermal stress over-limit risk assessment value; Defined as the preset safety threshold for the phase change heat of the material, which is a dimensionless control setting constant; Defined as the current control time The thermal boundary-limited impedance damping generated after thermal reconstruction calculation is used as the target damping parameter finally sent to the impedance controller, and its unit is Newton-second per meter. Defined as the current control time The initial optimization impedance damping extracted from the original optimal process parameter combination is expressed in Newton-seconds per meter. Defined as a preset logarithmic rigid amplification constant, it is used to control the compensation strength of damping gain as the thermal risk exceeds the limit.

[0073] Through the cross-operation of this set of formulas, the system performs a linkage transformation between the decoupled kinematic and dynamic commands, completing the full-process mapping of thermal safety constraints at the impedance control layer.

[0074] Specifically, after calculating the two key control parameters mentioned above, the control system executes the command replacement and loop closure steps. The system replaces the initial optimal feed rate with the thermally boundary-limited feed rate and replaces the initial optimal impedance damping with the thermally boundary-limited impedance damping, completing the reconstruction and update of the optimal process parameter combination. The reconstructed and updated optimal process parameter combination is then input into the saturation function for constraint. This step re-injects the process vector, after thermal risk correction, into the system's main operating loop. After subsequent trapezoidal saturation boundary constraints, it is sent to the underlying servo drive device to drive the robot's end effector.

[0075] In summary, considering the shortcomings of existing technologies such as reliance on offline process testing, blindly increasing coolant flow rate, or adding expensive external temperature sensors for controlling thermal damage in robotic grinding, the thermo-coupled adaptive constraint control scheme provided in this embodiment demonstrates high technical innovation and promising industrial applicability. Traditional control systems are purely mechanical closed loops. When faced with extreme regions where thermal conductivity volumes change abruptly, such as the leading and trailing edges of blades, the system lacks thermal situational awareness and continues to maintain a large constant cutting force in one direction, often resulting in transient large-area burns on titanium alloy blades.

[0076] This embodiment does not require modification or addition to the robot's external hardware shell. Utilizing the equivalent cutting force mechanical indices and contact elliptical space geometric indices already present in Embodiment 1, it performs in-depth low-level data feature mining and utilization, decoupling and deriving highly timely transient heat generation power and heat dissipation volume characteristics. By implementing a special speed-reduction and drag-increasing thermodynamic reconstruction action over time, it ensures effective heat dissipation to the depths of the substrate while significantly suppressing end-effector chatter easily triggered during low-speed machining using high damping characteristics.

[0077] This solution enables the entire robotic automated grinding system to achieve intelligent and proactive intervention capabilities for online prevention and control of thermophysical damage to material surfaces while ensuring the geometric accuracy of the contour. This has significant practical application value for ensuring high metallographic consistency and high fatigue resistance processing quality of aero-blades.

[0078] Example 3: In the automated grinding process of complex curved titanium alloy blades for aero-engines, in addition to the macroscopic geometric volume wear of the grinding tool caused by long-term friction, namely the shedding of abrasive grains and the consumption of the matrix, there is another complex physical phenomenon that seriously weakens the processing efficiency.

[0079] Titanium alloys possess extremely strong chemical affinity and ductility. Under the high temperature and pressure interface generated by constant-force grinding, the tiny high-temperature metal chips cut and detached from the blade surface easily fuse together, firmly filling the chip space on the working surface of the grinding tool. This phenomenon leads to severe chip adhesion and clogging on the grinding tool surface. Once clogging occurs, the grinding tool does not experience substantial geometric wear, but its effective cutting depth and micro-sharpness are drastically lost. The physical mechanism of the contact interface instantly deteriorates from normal micro-cutting fracture to macro-viscous sliding friction. At this point, the high-frequency dynamic fluctuations of the cutting force will show a dramatic increase, triggering high-frequency vibration damage on the blade surface.

[0080] This embodiment, based on the backbone architecture of the constant-force grinding impedance control method for robot blades established in the above embodiments, further provides an implementation scheme to deeply address the non-physical degradation of the grinding wheel surface. Specifically, this embodiment provides an active sensing and online micro-excitation obstacle removal mechanism for grinding debris adhesion and blockage. Without adding any external visual inspection equipment or additional hardware costs, it achieves online obstacle removal by relying on in-depth mining of underlying mechanical time-series data and adaptive reconstruction of the control loop.

[0081] Specifically, the active sensing and online micro-vibration obstacle clearing step described in this embodiment has a strictly defined position within the temporal logic flow of the entire control system. This step is configured to be executed after the output force error compensation value of the pre-trained network model is combined to generate the final target contact force, and before the optimal process parameter combination and the final target contact force are input into the robot's impedance controller after being constrained by a preset saturation function. This timing arrangement aims to fully utilize the mechanical and process parameter benchmarks that have already been computed and converged in the preceding steps, and to perform targeted post-processing interventions on this basis, ensuring unidirectional data flow transmission and logical decoupling.

[0082] Specifically, the control system first performs an online extraction step of high-frequency mechanical wave characteristics. This involves extracting real-time normal force data continuously collected by the robot from a six-dimensional force sensor within a preset time window. The six-dimensional force sensor of an industrial robot typically provides real-time feedback on the contact reaction force experienced by the end effector at a high sampling rate on the kilohertz level. The control system allocates a fixed-length circular data buffer in its underlying memory to store this preset time window, for example, the historical normal force time series over the past 500 milliseconds.

[0083] Subsequently, the system uses a preset bandpass filter to filter out the steady-state force characteristics corresponding to the preset low-frequency passband, and then generates a normal force fluctuation time sequence corresponding to the preset high-frequency stopband. In normal constant-force grinding operations, the impedance controller mainly responds to the macroscopic undulations of the blade surface by adjusting low-frequency servo motion, so the steady-state cutting force is mainly concentrated in the low-frequency range. However, when grinding debris adheres and clogs, the surface of the grinding tool loses its cutting ability and continuously slips on the blade surface and instantly re-engages. This high-frequency stick-slip alternating friction mechanism will generate significant high-frequency torque oscillations.

[0084] The aforementioned preset bandpass filter employs a specific digital signal processing structure to suppress low-frequency signal components that represent the desired constant contact force, and to extract dynamic alternating force signals that overflow to the high-frequency monitoring band due to the deterioration of the cutting state, i.e., the high-frequency stopband range in terms of the position control loop, thereby generating a time sequence of normal force fluctuations that purely reflects contact instability.

[0085] It is also important to note that after acquiring the pure high-frequency fluctuation characteristics, the system needs to further exclude legitimate fluctuations caused by normal physical wear of the grinding tool. Therefore, variable decoupling and clogging deterioration feature quantification steps are performed. The system performs variance statistics on the time series of normal force fluctuations and divides the obtained fluctuation variance value by the spatial statistical mean of the nominal wear of all target mesh cells in the current contact area (i.e., the equivalent nominal wear of the area) to generate a clogging deterioration index that characterizes the degree of non-wear loss of the cutting ability of the working surface of the grinding tool. By introducing the spatial statistical mean, the system achieves physical alignment of the microscopic mesh wear data and the macroscopic mechanical fluctuation variance on a spatial scale.

[0086] Variance is a core physical quantity in statistics and signal processing, used to measure the degree of data dispersion and the intensity of signal alternating energy. Calculating the variance of the normal force fluctuation time series allows for precise quantification of the severity of high-frequency slippage and vibration at the contact surface within the current time window. However, an increase in the fluctuation variance value could be caused by cutting passivation due to normal wear of abrasive particles, or by debris blockage. To decouple these two factors, the system creatively incorporates the nominal wear amount of the target mesh cell, derived from the physical mechanism in the preceding steps, as a penalty denominator in the calculation.

[0087] By dividing by the nominal wear amount, the system constructs a relative state evaluation logic: if the nominal wear amount of the current mesh cell is already large, then the large torque fluctuation is a normal degradation decay; however, if the bottom layer reports an abnormally large fluctuation variance value in the early stage of processing or when the nominal wear amount is extremely small, it means that the loss of cutting ability is not due to physical wear, but rather to a serious non-wear state deterioration, i.e., wear debris clogging.

[0088] To precisely constrain the above decoupling quantization process, the control system executes the following defined congestion exacerbation index evaluation formula algorithm in the background: In the formula: The congestion deterioration index represents the congestion deterioration index calculated by the system within the time window of the current control moment, which is a dimensionless relative state assessment characteristic scalar. The variance value of the fluctuation is obtained after statistical calculation of the normal force fluctuation time series (or time series), which is used to quantify the energy of the high-frequency stick-slip oscillation of the contact interface within a unit time window; This represents the spatial average of the nominal wear amount transmitted to the calculation module at the current moment, corresponding to all target mesh cells within the current contact area, i.e., the equivalent nominal wear amount of the area, and its unit is millimeters; This represents a preset constant-type minimum offset, used to ensure the numerical stability of the entire evaluation division operation when the nominal wear amount approaches zero during numerator and denominator operations.

[0089] The values ​​calculated using the above formula provide an independent characteristic index that can resist macroscopic physical wear interference and directly and objectively reflect the loss of cutting ability on the surface of the grinding tool due to the adhesion of grinding debris.

[0090] Optionally, the system then proceeds to a state determination and control branch decision-making step. The control system determines whether the clogging deterioration index is greater than a preset critical adhesion clogging threshold. This preset critical adhesion clogging threshold is a safety boundary parameter calibrated during offline process testing by analyzing sample data of the clogging deterioration index corresponding to the critical point of macroscopic vibration marks appearing on the polished surface. Based on the determination result, the system implements a strict binary tree branch control flow.

[0091] Specifically, if the clogging deterioration index is not greater than the critical adhesion clogging threshold, it indicates that the chip space on the surface of the current grinding tool is still sufficient, the abrasive grain exposure height can maintain normal micro-cutting operations, and the high-frequency fluctuations of the cutting force are within the background noise range allowed by healthy machining. Under this safe operating condition, the control system maintains the optimal combination of process parameters and the final target contact force unchanged, and directly proceeds to the subsequent constraint steps. This means that the system does not apply any additional interference to the impedance parameters and contact force references carefully optimized by the preceding modules, ensuring the continuity of conventional constant force grinding operations and the smoothness of trajectory tracking.

[0092] It is also important to note that if the clogging deterioration index exceeds the critical adhesion clogging threshold, the system determines that the current contact area is clogged with abrasive debris and activates the online micro-vibration clearing mechanism. This determination means that the tool surface is largely covered by high-temperature molten titanium alloy abrasive debris. Continuing with conventional grinding will not only fail to effectively remove the material but will also generate a large amount of useless frictional heat, exacerbating the concentration of thermal stress on the workpiece surface and leading to tool failure. The activation of the online micro-vibration clearing mechanism signifies that the system temporarily interrupts the conventional constant force position follow-up strategy and enters an active excitation mode with the primary goal of disrupting the interface adhesion state.

[0093] For example, the online micro-excitation obstacle clearing mechanism is first manifested at the physical execution level as the reconstruction of the target value of the normal contact force. Specifically, the online micro-excitation obstacle clearing mechanism generates an internally sinusoidal periodic oscillating bias force with a preset obstacle clearing amplitude and a preset obstacle clearing frequency. This preset obstacle clearing frequency is typically set to a high-frequency band that avoids the mechanical resonance point of the robot as a whole and is sufficient to induce localized wear fatigue fracture. The sinusoidal periodic oscillating bias force is then time-domain superimposed with the final target contact force to generate a micro-excitation reconstructed contact force completely independent of the final target contact force.

[0094] By introducing a sinusoidal periodic oscillating bias force, the system forcibly transforms the originally intended stable and constant contact force command into a dynamic control command containing alternating impact loads. This alternating impact load, acting on the surface of the grinding tool filled with abrasive debris, generates a high-frequency compression and release cycle effect within the micro-contact area between the tool and the titanium alloy workpiece. Under the continuous action of this alternating stress, due to the significant difference in elastic modulus and thermal expansion coefficient between the adhering abrasive debris and the tool matrix binder, micro-fatigue damage rapidly accumulates at the lattice interface of the weld area, generating micro-cracks. This, in turn, forces the adhered metal abrasive debris to detach from the grinding tool surface, achieving in-situ online self-cleaning.

[0095] To ensure the determinism of system control commands and the rigor of data flow, the control system executes the following defined micro-excitation dynamics formula algorithm for contact force reconstruction: In the formula: The micro-excitation reconstructed contact force command, which includes alternating impact characteristics and is calculated and generated by the system at the current control moment, is expressed in Newtons. This represents the final target contact force reference value output by the preceding module before the obstacle clearing mechanism is activated. It represents the expected macroscopic normal force of conventional grinding, and its unit is Newton. The preset obstacle clearing amplitude constant, which is pre-set and stored in the controller register, is used to define the relative intensity of the peak and trough of the alternating force shock wave, and its unit is Newton; The preset obstacle clearing frequency constant is used to determine the density of the impact alternation cycle per unit time, and its unit is Hertz; This represents a continuous time variable representing the operation of a real-time control system.

[0096] At the software kernel level, this formula smoothly and seamlessly transforms a constant static expectation into a dynamic excitation force expectation, and the two achieve complete isolation of physical variables in terms of mathematical expression and memory usage.

[0097] Specifically, to ensure that the aforementioned alternating impact force commands can be effectively transmitted to the machining interface by the underlying mechanical structure, the system synchronously executes a matching reconstruction step for the impedance controller's dynamic parameters. Simultaneously, the impedance stiffness in the optimal combination of process parameters is extracted, and this extracted impedance stiffness is uniquely defined as the conventional optimization impedance stiffness. In the conventional constant force grinding mode, the reinforcement learning optimization algorithm typically outputs a relatively small conventional optimization impedance stiffness, causing the robot's end effector to exhibit compliant characteristics similar to a soft spring. This allows it to adaptively yield to workpiece surface errors, ensuring a constant contact force.

[0098] However, in the micro-excitation obstacle clearing mode, if the robotic arm end effector remains in this low-stiffness compliant state, the high-frequency oscillation component in the micro-excitation reconstruction contact force generated by the system will be largely absorbed and buffered by the flexible deformation caused by the low stiffness. This prevents the grinding tool from generating actual high-frequency striking motions in physical space, significantly reducing the obstacle clearing effect. Therefore, the system uses a preset stiffness multiplication factor to linearly amplify and correct the conventional optimization impedance stiffness, generating a high-stiffness obstacle clearing impedance stiffness. By significantly increasing the position stiffness term in the impedance loop, the robot end effector is instantly locked into a high-stiffness dynamic response mode, allowing the alternating peaks in the mechanical commands to be transformed into rigid displacement impacts that directly act on the adhered grinding debris.

[0099] Optionally, after completing the dual reconstruction of mechanical objectives and dynamic parameters, the system performs variable substitution and loop closure operations. The conventional optimization impedance stiffness in the optimal process parameter combination is removed and replaced with the high-rigidity obstacle-clearing impedance stiffness to form a reconstructed process parameter combination. This update the set of instruction vectors to be issued in the memory data structure, ensuring that the impedance controller has a matching high-frequency response dynamic basis when receiving the excitation force command. Finally, the micro-excitation reconstruction contact force and the reconstructed process parameter combination are input to the saturation function for constraint. After smoothing, limiting, and slope restriction using a dedicated trapezoidal saturation function, the command is safely issued to the servo driver for execution.

[0100] After several seconds of high-frequency vibration clearing, once the blockage deterioration index falls back to within the safe threshold, the system will automatically cancel the aforementioned bias force and stiffness multiplication, smoothly restore to the optimal normal state output by the preceding module, and continue to perform high-quality grinding.

[0101] In summary, considering the current situation where addressing grinding tool slag clogging typically relies on manual judgment followed by machine shutdown and belt replacement, or the addition of expensive peripheral equipment such as high-pressure cutting fluid flushing and laser vision monitoring, the technical solution described in this embodiment offers the following advantages. Conventional systems often struggle to effectively intervene when faced with non-wearing deterioration, easily leading to a continuous decline in machining quality or even workpiece scrap.

[0102] This embodiment requires no additional hardware. By deeply mining and decoupling the high-frequency unsteady energy in the existing six-dimensional force sensor signal, a highly robust blockage deterioration index is constructed, effectively eliminating interference from conventional physical wear. After confirming blockage, the solution overcomes the limitation of traditional impedance control systems that only optimize contact force stability. It introduces an online micro-excitation clearing action, which is a combination of high-frequency bias force and high stiffness parameters. The alternating contact stress induced in situ causes fatigue fracture and detachment of the high-temperature fused titanium alloy wear debris.

[0103] This solution endows the robotic automated grinding system with advanced closed-loop intervention capabilities for self-diagnosis and self-cleaning. While effectively extending the overall service life of the grinding tools, it significantly reduces the risk of surface vibration damage to blades caused by grinding debris slippage, demonstrating outstanding engineering application value in improving the processing stability of high-end complex curved surface parts.

[0104] Example 4: like Figure 2 As shown, in order to realize the constant force grinding impedance control method for robot blades with integrated wear compensation as described in the above-mentioned method embodiments, this embodiment provides a high-fidelity physical system architecture.

[0105] Specifically, referring to the actual processing and manufacturing production line configuration, a wear-compensated robotic blade constant-force grinding impedance control system is mainly deployed in an automated workstation consisting of a multi-degree-of-freedom industrial robot, an end-effector grinding actuator, multi-dimensional force sensors, and a main control industrial computer. The industrial robot is a six-degree-of-freedom serial articulated robot, used to carry the end-effector grinding tool through complex trajectory movements in three-dimensional space. The end-effector grinding actuator uses a constant-force floating electric spindle or a belt sander contact wheel, its working surface designed for direct contact with the gently curved surface or high-curvature edge of the aero-engine titanium alloy blade.

[0106] The multi-dimensional force sensor is a six-dimensional force / torque sensor integrated between the robot's end flange and the grinding spindle to capture high-frequency changing mechanical contact force signals during the grinding operation in real time. The main control industrial computer, as the core computing device, is equipped with a high-speed central processing unit, large-capacity random access memory, and non-volatile solid-state drive. It runs a real-time operating system with microsecond-level real-time interpolation response capability, used to build and run multiple functional modules that interact with each other at the software level.

[0107] Specifically, the control system precisely decomposes and integrates the following four core virtual functional modules in terms of spatial structure, software algorithm architecture, and data communication flow. The modules communicate with each other through an internal high-speed data bus and a shared memory buffer, enabling sub-millisecond data transfer: Specifically, the system includes a data acquisition and grid allocation module. At the hardware implementation level, this module primarily runs in the real-time microkernel of the main control industrial computer and achieves clock synchronization with the robot's joint encoders, spindle speed sensors, and the six-dimensional force sensor via high-speed digital input / output interfaces and industrial buses (such as EtherCAT). This module is used to acquire grinding status data, divide the working surface of the grinding tool into multiple independent grid units, establish a geometric model of the contact area based on the blade surface curvature at the grinding contact point and the effective radius of the grinding tool, determine the target grid units within the contact area, and allocate the real-time measured total normal force to the target grid units.

[0108] In actual operation, this module allocates a multi-dimensional feature matrix space in the random access memory of the industrial computer, discretizes the geometric outer surface of the grinding tool, such as a sanding belt wheel or a contact rubber wheel, and establishes a one-to-one corresponding digital mesh cell file. When the robot drives the grinding tool to press against the blade surface, the module dynamically calculates the blade surface curvature at the current contact point by reading the spatial pose calculated from the robot's forward kinematics in real time and combining it with the blade digital geometric model (CAD model) pre-imported into the main control computer.

[0109] Simultaneously, based on the current physical radius of the grinding tool, the underlying algorithm performs spatial geometric intersection calculations according to the elastic Hertzian contact equation, thereby accurately delineating the spatial boundaries of elastic deformation under pressure within the virtual mesh matrix. These delineated specific meshes are defined as the target mesh units. Since the six-dimensional force sensor feeds back a concentrated total force, this module utilizes a built-in area-weighted integral operator to discretely distribute the total normal force to each target mesh unit according to the area ratio, thus converting the macroscopic sensor electrical signal into the instantaneous surface pressure field of the microscopic contact mesh.

[0110] Specifically, the system includes a wear prediction and morphology reconstruction module. This module is physically associated with a high-speed central processing unit or dedicated graphics processing unit of an industrial computer. It is used to construct a physical wear model based on contact pressure and friction velocity for the target mesh cells to calculate the nominal wear amount. Combined with a Gaussian process error correction model based on the partitioning of the working surface, the final wear amount of each target mesh cell is obtained through uncertainty-weighted fusion, and the three-dimensional wear morphology of the grinding tool is reconstructed in real time.

[0111] To ensure full transparency, it's important to note that the physical wear model within this module is stored in the non-volatile memory of an industrial computer as low-level mathematical operators. It dynamically calls upon the instantaneous pressure of each target grid cell output from the aforementioned data acquisition and grid allocation module, and calculates the friction velocity by combining this with the rotational speed feedback from the current spindle servo motor encoder. Following the principle of energy dissipation, it performs integral waveform superposition on the time axis to calculate the theoretical nominal wear amount of each local grid cell.

[0112] Meanwhile, due to the physical differences in the intensity of cutting and shearing experienced by different regions along the axial direction of the grinding tool, such as the edge and center regions, this module divides the mesh archive into five wear characteristic zones according to the axial span. Each zone independently runs a set of Gaussian process error correction models stored in the industrial computer's memory. These models read the deviation residual data generated by historical grinding samples uploaded by external measuring equipment in real time, and calculate the current theoretical prediction error and uncertainty variance. This module uses the reciprocal of the variance as a dynamic uncertainty weighting factor, and fuses the nominal wear amount with the Gaussian prediction error in the arithmetic logic unit, updating the three-dimensional mesh thickness vector stored in memory in real time. This allows the three-dimensional wear morphology point cloud matrix of the micro-undulations on the surface of the grinding tool to be reconstructed in real time in the monitoring memory of the main control computer.

[0113] Specifically, the system includes a geometric compensation and parameter optimization module. This module is closely connected to the robot's advanced motion planning core and trajectory interpolation algorithm. It is used to perform geometric compensation of the grinding tool's center point based on the three-dimensional wear morphology, and to construct a multi-dimensional composite cost function that includes contact force safety and wear uniformity. The module then generates the optimal combination of process parameters corresponding to the current contact area through a preset optimization algorithm.

[0114] In real-world scenarios, as the grinding tool wears down, its physical radius inevitably shrinks, either overall or locally. If the robot continues to move along the initially planned theoretical trajectory, the machining process will inevitably lead to loss of contact force or undercutting of the blade dimensions. This module extracts the deepest mesh loss value output by the aforementioned wear morphology reconstruction module, converts it into an offset vector, and superimposes it onto the robot's tool center point (TCP) coordinate system in real time. This allows the robot's trajectory interpolator to automatically perform feed depth compensation towards the blade matrix, thereby ensuring that the actual physical boundary of the grinding tool always remains in contact with the theoretical machining profile of the blade.

[0115] Based on this geometric correction, to adaptively adjust the dynamic performance during the machining process, this module constructs a five-dimensional mathematical evaluation space within the processor, incorporating material removal rate deviation, contact force safety limit, process parameter change rate penalty, and the squared term of mesh uniformity range into the composite cost function. Subsequently, a reinforcement learning optimization algorithm running within the industrial computer performs multi-generational searches under the aforementioned multiple boundary constraints, outputting online a set of optimal process parameter combinations including equivalent cutting force, impedance stiffness, impedance damping, and feed rate, and writing it to a shared memory buffer for subsequent control loop calls.

[0116] Specifically, the system includes an impedance control execution module. The hardware control terminal of this module directly acts on the servo drivers of each joint and the spindle motor driver of the robot through a high-speed industrial bus. It is used to calculate the equivalent cutting force based on the abrasive passivation coefficient of the target mesh cell, and combine it with the pre-trained network model to output the force error compensation value to generate the final target contact force. The optimal process parameter combination and the final target contact force are then constrained by a preset saturation function and input into the robot's impedance controller.

[0117] In the specific physical operation closed loop, this module first uses time-domain integration to track the microscopic self-sharpening degradation of each target mesh element caused by material adhesion and metallographic friction, and calculates the passivation coefficient reflecting the degree of abrasive grain dulling. The arithmetic unit divides the reference force obtained from process optimization by the nonlinear power of this passivation coefficient, thereby amplifying the output equivalent cutting force to physically overcome the increase in cutting resistance caused by tool dulling.

[0118] Meanwhile, to eliminate the adverse effects of macroscopic structural disturbances such as nonlinear friction in robot joint reducers and gravitational deformation of connecting rods on force tracking accuracy, this module invokes a pre-trained partitioned evaluation network model deployed in an industrial computer. This network model quickly predicts the static mechanical error under the current dynamic environment by reading the actual torque, feed rate, and surface curvature fed back from sensors, outputting a feedforward force error compensation value. This value is then linearly superimposed with the equivalent cutting force in an adder to generate the final target contact force.

[0119] To prevent drastic adjustments to control quantities from exceeding hardware electrical safety limits, this module is configured with a dedicated trapezoidal saturation function based on trigonometric functions at the software level. All reconfigured process parameters and force commands, after normalization, trapezoidal limiting, and denormalization smoothing, are written in real-time to the impedance control algorithm register of the robot joint controller at a megahertz-level control bus rate. The joint actuators then dynamically adjust the motor current accordingly, enabling the robot end effector to exhibit compliant force-following behavior consistent with the complex surface deformation characteristics of the blade, smoothly completing the high-precision grinding operation of the entire blade.

[0120] Example 5: Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0121] like Figure 3 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0122] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0123] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0124] The memory 103 stores a computer program corresponding to the wear compensation-integrated constant force grinding impedance control method for robot blades in the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0125] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 3 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0126] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for controlling the impedance of a robot blade in constant force grinding with wear compensation, used to control a robot to perform grinding operations, characterized in that, include: Acquire grinding status data, divide the working surface of the grinding tool into multiple independent grid units, establish a geometric model of the contact area based on the blade surface curvature of the grinding contact point and the effective radius of the grinding tool, determine the target grid unit in the contact area, and distribute the real-time measured total normal force to the target grid unit. A physical wear model based on contact pressure and friction velocity is constructed for the target mesh cell to calculate the nominal wear amount. Combined with a Gaussian process error correction model based on the working surface partition, the final wear amount of each target mesh cell is obtained through uncertainty weighted fusion, and the three-dimensional wear morphology of the grinding tool is reconstructed in real time. Geometric compensation of the grinding tool center point is performed according to the three-dimensional wear morphology, and a multi-dimensional composite cost function including contact force safety and wear uniformity is constructed. The optimal combination of process parameters corresponding to the current contact area is generated through a preset optimization algorithm. The equivalent cutting force is calculated based on the abrasive passivation coefficient of the target mesh unit. The force error compensation value is then combined with the pre-trained network model to generate the final target contact force. The optimal process parameter combination and the final target contact force are then input into the robot's impedance controller after being constrained by a preset saturation function.

2. The method according to claim 1, characterized in that, The process of dividing the working surface of the grinding tool into multiple independent grid units, establishing a geometric model of the contact area based on the blade surface curvature at the grinding contact point and the effective radius of the grinding tool, determining the target grid units within the contact area, and allocating the real-time measured total normal force to the target grid units includes: dividing the working surface of the grinding tool into a three-dimensional grid unit sequence along the circumferential direction at preset equal arc length intervals and along the axial direction at preset equal distance intervals; calculating the contact deformation based on the blade surface curvature at the current grinding contact point and the effective radius of the grinding tool, combined with Hertzian contact theory, to generate an elliptical geometric model of the contact area; mapping the elliptical geometric model of the contact area to the local coordinate system of the working surface of the grinding tool, selecting the target grid units within the contact range, and allocating the real-time measured total normal force of the six-dimensional force sensor to each target grid unit according to the contact area ratio to generate instantaneous contact pressure.

3. The method according to claim 1, characterized in that, The step of constructing a physical wear model based on contact pressure and friction velocity for the target mesh cell to calculate the nominal wear amount includes: establishing an independent wear model in the form of a modified Preston equation for each target mesh cell; obtaining the initial wear coefficient of the target mesh cell through offline calibration experiments; and, based on the initial wear coefficient, integrating the product of the instantaneous contact pressure and instantaneous friction velocity of the target mesh cell along the time dimension, and performing a nonlinear mapping in conjunction with the material removal pressure index to output the nominal wear amount of the target mesh cell.

4. The method according to claim 1, characterized in that, The method of combining the Gaussian process error correction model based on the working surface partitioning to obtain the final wear amount of each target mesh cell through uncertainty weighted fusion includes: dividing the working surface of the grinding tool into five wear characteristic partitions along the axial direction: left edge region, left transition region, center region, right transition region, and right edge region; establishing an independent Gaussian process error correction model for each wear characteristic partition, outputting the physical model prediction error and prediction standard deviation; calculating the reciprocal of the prediction standard deviation as the dynamic fusion weight of the corresponding partition; and weighting and summing the nominal wear amount and the physical model prediction error based on the dynamic fusion weight to obtain the final wear amount of each target mesh cell.

5. The method according to claim 1, characterized in that, The process of performing geometric compensation for the center point of the grinding tool based on the three-dimensional wear morphology employs a triple geometric compensation mechanism, specifically including: the first layer is real-time dynamic compensation in partitions: using the final wear amount along the unit normal vector of the blade surface at the current grinding contact point, the coordinates of the initial calibration tool center point of the grinding tool are dynamically superimposed and updated in real time; the second layer is three-dimensional morphology calibration between passes: after a single grinding pass, a measured three-dimensional wear morphology is generated using a scanning device, and the hyperparameters of the Gaussian process error correction model are updated after comparing with the predicted values; the third layer is local correction of key areas between passes: after a single grinding pass, the key areas of the blade are scanned to generate local geometric errors, and these local geometric errors are superimposed on the compensation reference for the next pass.

6. The method according to claim 1, characterized in that, The process involves constructing a multidimensional composite cost function that includes contact force safety and wear uniformity. An optimal combination of process parameters corresponding to the current contact area is generated using a pre-defined optimization algorithm. This includes: constructing a five-dimensional composite cost function, which is composed of material removal rate error cost, contour accuracy error cost, contact force safety penalty, process parameter change rate penalty, and wear uniformity cost multiplied by their respective weighting coefficients and then superimposed; and employing a reinforcement learning optimization algorithm to perform strategy search under the constraints of the five-dimensional composite cost function, outputting the optimal combination of process parameters corresponding to the current contact area. The optimal combination of process parameters includes equivalent cutting force, impedance stiffness, impedance damping, and feed rate.

7. The method according to claim 6, characterized in that, The wear uniformity cost term is used to suppress non-uniform wear of the grinding tool. It is calculated as follows: extract the maximum wear amount and the average wear amount of all target grid cells of the grinding tool at the current moment; calculate the square of the difference between the maximum wear amount and the average wear amount; multiply the square of the difference by a preset wear uniformity penalty coefficient to obtain the output value of the wear uniformity cost term.

8. The method according to claim 1, characterized in that, The step of calculating the equivalent cutting force based on the abrasive passivation coefficient of the target mesh cell includes: obtaining the passivation rate coefficient of the target mesh cell, integrating the product of the instantaneous contact pressure and the instantaneous sliding velocity of the cell over time, and calculating the abrasive passivation coefficient at the current moment using an exponential decay function; dividing the preset initial target contact force by the exponential power function of the abrasive passivation coefficient to obtain the equivalent cutting force after compensating for the abrasive passivation effect, wherein the base of the exponential power function is the abrasive passivation coefficient, and the exponent is a removal rate force exponent related to material properties.

9. The method according to claim 1, characterized in that, The pre-trained network model is a partitioned evaluation network. The step of combining the output force error compensation value of the pre-trained network model to generate the final target contact force includes: dividing the working surface of the grinding tool into an edge contact area, a transition contact area, and a center contact area, and constructing an independent partitioned evaluation network model for each contact area; using the equivalent cutting force, actual contact force, spindle speed, feed rate, and blade surface curvature of the current contact area as input feature vectors, and inputting them into the partitioned evaluation network model of the corresponding area; outputting the feedforward force error compensation value of the area through the partitioned evaluation network model, and superimposing and correcting the feedforward force error compensation value with the equivalent cutting force to generate the final target contact force.

10. The method according to claim 1, characterized in that, The saturation function is a grinding-specific trapezoidal saturation function. The process of inputting the optimal process parameter combination and the final target contact force into the robot's impedance controller after being constrained by the preset saturation function includes: using a grinding-specific trapezoidal saturation function constructed based on the superposition and combination of sinusoidal trigonometric functions; normalizing each parameter in the optimal process parameter combination and the final target contact force according to preset upper and lower limits of the process parameters corresponding to the current contact area; using the grinding-specific trapezoidal saturation function to apply nonlinear smooth boundary amplitude constraints and rate of change constraints to the normalized parameters to suppress abrupt changes in process parameters, and then inversely normalizing the constrained parameters as control commands input to the robot's impedance controller.

Images