A machine vision-based gating self-adaptive grinding control system

Abstract

The application discloses a casting gate self-adaptive grinding control system based on machine vision, which comprises a three-dimensional vision acquisition unit, a grinding load acquisition unit, a grinding execution mechanism and a controller, the controller is electrically connected with the three-dimensional vision acquisition unit, the grinding load acquisition unit and the grinding execution mechanism respectively; the three-dimensional vision acquisition unit is used for acquiring three-dimensional point cloud containing a gate area and a casting stable reference, the casting stable reference is a geometric reference which can be repeatedly positioned before and after gate grinding and does not disappear with the removal of the gate; the grinding load acquisition unit is used for acquiring the grinding load during the short grinding stroke of the grinding execution mechanism. The application can correct the local material removal capacity online, limit the accumulation of grinding error and reduce the risk of overcutting of the gate root.

Description

Technical Field

[0001] This invention relates to the field of casting processing technology, and in particular to an adaptive grinding control system for casting gates based on machine vision. Background Technology

[0002] Casting gates typically need to be removed using robots, gantry grinding machines, or multi-axis special-purpose machines. Due to casting shrinkage, pouring fluctuations, and differences in pre-cutting processes, the gate height, root profile, and local allowance vary between different castings; clamping deviations can also cause positional errors between the offline model and the actual workpiece. When using a fixed trajectory, variations in local allowance can easily lead to under-grinding or over-cutting.

[0003] Existing casting gate grinding solutions typically utilize visual inspection to determine the area to be processed, or adjust the feed rate based on grinding force, spindle current, and surface condition. However, the grinding load is simultaneously affected by material hardness, contact area, grinding wheel sharpness, and clamping rigidity. Simply maintaining the grinding force or adjusting parameters for the next pass based on whether the machining meets the standards makes it difficult to directly obtain the actual volume of material removed in a short stroke. Furthermore, directly attributing load variations caused by wheel dulling or clamping looseness to the material removal parameters will cause subsequent control parameters to continuously deviate from their intended range.

[0004] Therefore, there is a need in the field for a control system that can establish a comparable remaining material state across strokes during processing, correct local material removal capacity based on the actual removal results of each short stroke, and generate subsequent trajectories accordingly, while also constraining the risk of overcutting at the gate root and interference from abnormal sources. Summary of the Invention

[0005] The purpose of this invention is to provide a machine vision-based adaptive grinding control system for casting gates, which solves the problems that fixed trajectories cannot balance removal efficiency and body protection when the local allowance and material removal capacity of castings fluctuate, as well as the problems that tool or clamping abnormalities may contaminate the adaptive parameters.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An adaptive grinding control system for casting gates based on machine vision includes a three-dimensional vision acquisition unit, a grinding load acquisition unit, a grinding execution mechanism, and a controller. The controller is electrically connected to the three-dimensional vision acquisition unit, the grinding load acquisition unit, and the grinding execution mechanism, respectively. The three-dimensional vision acquisition unit is used to acquire a three-dimensional point cloud containing the gate area and the casting stability reference. The casting stability reference is a geometric reference that can be repeatedly positioned before and after the gate is ground and does not disappear when the gate is removed. The grinding load acquisition unit is used to acquire the grinding load during the short grinding stroke performed by the grinding actuator. The controller is configured to: A local coordinate system is established based on the casting stability reference. A first three-dimensional point cloud is registered in the local coordinate system to determine the target cut surface. A first residual volume field is established based on the spatial relationship between the first three-dimensional point cloud and the target cut surface. Based on the first remaining volume field and the local removal coefficient corresponding to the spatial region in the first remaining volume field, a first short grinding stroke is generated and the grinding actuator is controlled to execute the first short grinding stroke. The grinding load obtained during the first short grinding stroke is converted into grinding power, and the grinding power is integrated over time to obtain the grinding energy corresponding to the first short grinding stroke. After the first short grinding stroke is completed, the three-dimensional vision acquisition unit is controlled to acquire the second three-dimensional point cloud, and the second three-dimensional point cloud is registered based on the same casting stability reference, and a second residual volume field is established. Calculate the volume difference between the first residual volume field and the second residual volume field within the spatial region covered by the first short grinding stroke, use the ratio of the volume difference to the grinding energy as the observed value of the local removal coefficient, and update the local removal coefficient corresponding to the spatial region based on the observed value of the local removal coefficient. Based on the second residual volume field and the updated local removal coefficient, at least one of the following is determined: trajectory length, feed rate, normal force, and endpoint of the next short grinding stroke. The next short grinding stroke is constrained by the spatial boundary of the target cut surface. After the next short grinding stroke is completed, the residual volume field establishment, grinding energy acquisition, and local removal coefficient update are repeated.

[0007] Preferably, the controller is further configured to: Based on the root curvature abrupt change, surface normal change, and point cloud confidence of the gate region, a root protection strip of variable width is formed at the edge of the target cut surface; At least a portion of the root protection strip is designated as a no-entry zone to prevent grinding tools from entering, and the area adjacent to the no-entry zone is designated as a tangential sweep zone to restrict the movement of the grinding tool along the tangential direction of the target cutting surface. The width of the root protection strip is increased when the point cloud confidence decreases.

[0008] Preferably, the controller controls the grinding actuator to switch between the rough grinding stage and the fine grinding stage based on the remaining height of the gate, the distance between the gate and the root protection strip, the confidence level of the local removal result, and the cumulative grinding energy; when the remaining height of the gate is 0.10 to 0.30 mm, and the difference between two consecutive remaining heights of the gate is not greater than a preset stable threshold, the controller controls the grinding actuator to terminate the grinding.

[0009] Preferably, the controller sets the trajectory length of the short grinding stroke to 10-30 mm, sets the acquisition time interval of adjacent three-dimensional point clouds to 0.5-3.0 s, and controls the grinding actuator to apply a normal force of 35-80 N in the rough grinding stage and a normal force of 15-40 N in the fine grinding stage.

[0010] Preferably, the controller updates the local removal coefficient using an exponential forgetting method, with an exponential forgetting factor of 0.80 to 0.98, and limits the single-cycle change of the updated local removal coefficient relative to the original local removal coefficient to 10% to 30%.

[0011] Preferably, the machine vision-based adaptive grinding control system for casting gates also includes an image acquisition unit for acquiring images of grinding sparks, and the controller is further configured to: Extract the harmonic features of the grinding power and the image features of the grinding spark image, and map the anomalies represented by the harmonic features and the image features to the workpiece coordinate system, the grinding wheel phase coordinate system and the fixture coordinate system, respectively. Based on the positional repeatability of the anomaly in the workpiece coordinate system, the phase repeatability in the grinding wheel phase coordinate system, and the positional repeatability in the fixture coordinate system, the anomaly source is determined to be a workpiece defect, a grinding wheel condition abnormality, or a clamping abnormality; and the volume difference and grinding energy corresponding to the short grinding strokes determined to be the grinding wheel condition abnormality or the clamping abnormality are prevented from participating in the update of the local removal coefficient.

[0012] Preferably, when the source of the anomaly is a workpiece defect, the controller controls the grinding actuator to perform detour grinding with reduced depth of cut at the abnormal location; when the source of the anomaly is an abnormal grinding wheel condition, the controller suspends the update of the local removal coefficient and controls the grinding actuator to perform grinding wheel dressing; when the source of the anomaly is a clamping abnormality, the controller controls the grinding actuator to retract and lock the tool.

[0013] Preferably, the controller writes a unified clock timestamp to the 3D point cloud, the grinding load, and the short grinding stroke, so that each 3D point cloud retest result corresponds to its previous short grinding stroke and the grinding load of that short grinding stroke; the 3D vision acquisition unit is a structured light camera, a line laser scanner, or a multi-view camera, and the grinding load acquisition unit is configured to acquire spindle power, spindle current, or spindle torque.

[0014] Preferably, the controller maintains the local removal coefficient before the update and controls the grinding actuator to reacquire the three-dimensional point cloud after lifting the tool when the grinding energy is lower than the energy threshold required for effective contact, the volume difference is negative, the three-dimensional point cloud registration is invalid, the spatial area covered by the first short grinding stroke is occluded, or the grinding load record is incomplete.

[0015] Preferably, the controller includes a processor and a memory, the memory storing program instructions for the processor to perform three-dimensional point cloud registration, residual volume field establishment, grinding energy calculation, local removal coefficient update, and short grinding stroke generation; the grinding actuator is an industrial robot, a gantry three-axis grinding machine, or a five-axis grinding machine.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: This invention discloses a machine vision-based adaptive grinding control system for casting gates. It utilizes a stable casting reference that does not disappear with gate removal to register adjacent point clouds, ensuring that volume changes primarily correspond to local material removal, thus reducing interference from overall workpiece pose changes on volume calculation. The grinding energy of the same short grinding stroke is paired with the remaining volume difference before and after that stroke to form spatially assigned local removal coefficient observations. This allows the combined influence of material, contact, and tool conditions on removal capability to be corrected online through the processing results. The next short stroke is generated using the current remaining volume field and the updated local removal coefficients, limiting a single processing operation to between two re-measurements, thus confining the impact of parameter deviations to a shorter spatial and temporal range. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an adaptive grinding control system for casting gates based on machine vision according to the present invention. Figure 2 This is a schematic diagram of the remaining volume field of the casting gate and the root protection zone in this invention; Figure 3 This is a flowchart of the short grinding stroke control in this invention; Figure 4 This is a schematic diagram of the three-coordinate discrimination of anomaly sources in this invention; Figure 5 This is a flowchart of the grinding wheel condition processing in this invention; In the figure: 1. 3D vision acquisition unit; 2. Grinding load acquisition unit; 3. Controller; 4. Grinding actuator; 5. Casting; 6. Grinding tool; 21. Gate; 22. Casting body; 23. Target cut surface; 24. Remaining volume field; 25. Root protection zone; 251. No-entry zone; 252. Tangential sweep zone; 26. Stable datum. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be understood that the implementation of the present invention is not limited to the following embodiments, and any modifications and / or alterations made to the present invention will fall within the protection scope of the present invention.

[0019] Unless otherwise specified, the methods described in the following embodiments are conventional methods in the art. Unless otherwise specified, the components or devices in the following embodiments are general standard parts or components known to those skilled in the art, and their structures and principles can be obtained by those skilled in the art through technical manuals or conventional experimental methods.

[0020] The stable datum referred to in this invention is a geometric datum on the casting that remains repeatably positioned before and after grinding and does not disappear with the removal of the gate. It can be a stable plane, hole, boss, or contour edge on the casting body, or a combination of two or more of these geometric elements. The remaining volume field is the volume and distribution of the material to be removed above the target cut surface in a spatial grid. Grinding energy is the integral of the net grinding power over time within a short grinding stroke. The local removal coefficient is a characterization of the amount of material removed per unit grinding energy within the corresponding spatial region. A short grinding stroke is a trajectory segment executed between two available three-dimensional re-measurements, whose control parameters can be independently updated based on the current intermediate state.

[0021] like Figure 1 As shown, an embodiment of the present invention discloses a machine vision-based adaptive grinding control system for casting gates, including a three-dimensional vision acquisition unit 1, a grinding load acquisition unit 2, a controller 3, and a grinding execution mechanism 4. The casting 5 is fixed on a fixture, and the grinding execution mechanism 4 drives the grinding tool 6 to move relative to the casting 5. The three-dimensional vision acquisition unit 1, the grinding load acquisition unit 2, and the grinding execution mechanism 4 are electrically connected to the controller 3.

[0022] The 3D vision acquisition unit 1 employs a structured light camera, a line laser scanner, or a multi-view camera. For workstations with significant dust impact, the controller 3 lifts the grinding tool 6 away from the workpiece after a short stroke, clears the measurement field of view using an air curtain, and then triggers the 3D vision acquisition unit 1 again. The lateral sampling interval for the point cloud is 0.05–0.30 mm. The controller 3 writes a unified clock timestamp to the point cloud and load data, ensuring that each retest result, previously executed short stroke, and the load record of that short stroke correspond.

[0023] The grinding load acquisition unit 2 directly acquires the spindle power, or it can acquire the spindle current or spindle torque and convert it into power according to the spindle calibration relationship. If a current signal is used, the controller 3 obtains the spindle output power based on the no-load current, bus voltage, drive efficiency, and speed correction; if a torque signal is used, the spindle output power is obtained based on the product of torque and angular velocity. The contact force is obtained by a force sensor installed on the grinding head or wrist, or it can be obtained from the displacement and stiffness relationship of the grinding head flexible mechanism.

[0024] Controller 3 is implemented by a processor and memory in an industrial computer, edge controller, or equipment control cabinet. The memory stores point clouds, target cut surfaces, voxel indexes, local removal coefficients, short-stroke parameters, abnormal states, and program instructions. The processor reads data from the 3D vision acquisition unit 1 and the grinding load acquisition unit 2 according to a unified timestamp, performs registration, volume calculation, energy integration, coefficient updates, and trajectory generation, and sends motion commands to the grinding actuator 4. The grinding actuator 4 is an industrial robot, a gantry three-axis machine, or a five-axis grinding machine.

[0025] like Figure 2 As shown, the gate 21 is located on the casting body 22, and the controller 3 selects a stable datum 26 around the gate 21. During the initial measurement, the controller 3 separates the stable datum 26 and the gate 21 from the point cloud and establishes a local coordinate system based on the geometric center, principal direction, and surface normal of the stable datum 26. In subsequent measurements, the rigid transformation is calculated using only the point set near the stable datum 26, and then the point cloud of the gate area is transformed to the same local coordinate system. In this way, the gradual disappearance of the gate 21 will not change the registration basis, and the clamping repetition error will not directly become a change in material volume.

[0026] The target cut surface 23 is given by the digital model of the casting in the local coordinate system, or it can be obtained by fitting the continuous surface of the casting body 22 around the gate. The controller 3 establishes a voxel mesh above the target cut surface 23, and assigns the volume to be removed to voxels that fall within the gate entity and are located above the target cut surface 23, thus obtaining the remaining volume field 24. Each voxel stores the spatial index, remaining volume, point cloud confidence, and local removal coefficient. The mesh size is selected based on the point cloud spacing, so that a voxel is supported by multiple effective measurement points. Voxels below the target cut surface 23 do not belong to the volume to be removed and are used as hard boundaries for trajectory planning.

[0027] Point cloud confidence is determined based on one or more of the following: local point density, incident angle, dispersion of adjacent normals, and consistency of multiple sampling. If the number of valid points on the stable reference 26 is lower than the registration requirement, or the registration residual exceeds the acceptable range set by the device, the controller 3 does not generate a new grinding stroke, but instead re-cleans the field of view and acquires a point cloud. This process avoids using two incomparable volume fields for material removal updates.

[0028] The controller 3 calculates the surface curvature and the normal change of adjacent points along the transition area between the gate 21 and the casting body 22, and forms a root profile at the position where the curvature or normal change reaches the root identification condition. The controller 3 establishes a root protection zone 25 with a width of 0.5 to 2.0 mm along the root profile. The root protection zone 25 includes a restricted area 251 near the casting body 22 and a tangential sweep area 252 located outside the restricted area 251.

[0029] Within the restricted zone 251, the envelope of the grinding tool 6 is prohibited from entering; within the tangential sweep zone 252, the trajectory direction is restricted to the tangential direction along the target cut surface 23, and the normal feed is limited. When the local point density decreases, the normal dispersion increases, or the measurement incident angle is unfavorable, the controller 3 reduces the point cloud confidence at the corresponding location and increases the guard band width within the range of 0.5–2.0 mm. The reason for this setting is that the root geometry and point cloud quality jointly determine the confidence level of the target cut boundary; directly converting a lower confidence level into a larger spatial interval can limit the intrusion of the grinding tool 6 into the casting body 22 without changing the closed-loop core.

[0030] like Figure 3 As shown, after completing point cloud registration and establishing the remaining volume field, controller 3 selects a starting position from the grid region with a larger remaining volume and constructs candidate trajectories along adjacent voxels to be removed. For the candidate trajectories, controller 3 determines at least one of the following based on the remaining volume and local removal coefficient of each region: trajectory length, feed rate, normal force, and endpoint. When the local removal coefficient is large, the input energy per unit region is reduced without crossing the target cut surface 23 and the root protection zone 25; when the local removal coefficient is small, the feed rate is reduced or the normal force is adjusted to ensure that the region obtains energy matching its remaining volume.

[0031] Controller 3 performs grinding tool envelope checks, actuator motion boundary checks, and root protection distance checks on the candidate trajectory. Once the checks are passed, it issues a short grinding stroke to the grinding actuator 4. The short grinding stroke length is 10–30 mm. The time interval between two 3D point cloud acquisitions is 0.5–3.0 s. The shorter trajectory and retest interval ensure that an inaccurate local removal coefficient only affects a limited area and can be corrected before the next stroke.

[0032] When machining a specific material and grinding tool combination for the first time, the initial value of the local removal coefficient is taken from a calibrated combination of the same material, type of grinding tool, and similar grinding parameters. Alternatively, initial observations can be obtained by performing a low-load, short grinding stroke in a region far from the root guard zone 25. For grids where initial observations have not yet been obtained, conservative initial values ​​are used, along with shorter trajectories and lower normal forces. For a spatial grid, controller 3 calculates the target grinding energy by dividing the remaining volume of the grid by the local removal coefficient, and then combines this with the allowable normal force and feed rate of the equipment to form a candidate trajectory. This relationship allows the same remaining volume to obtain different energy configurations in regions with different removal capacities.

[0033] Grinding is divided into rough grinding and fine grinding stages. The normal force in the rough grinding stage is 35–80 N, and the normal force in the fine grinding stage is 15–40 N. Controller 3 switches stages based on the remaining height of the gate, the distance from the current trajectory to the root guard strip 25, the confidence level of the local removal result, and the cumulative grinding energy. When approaching the target cut surface 23 or the root guard strip 25, controller 3 enters the fine grinding stage, shortening the single trajectory or reducing the normal force, thereby reducing the single state change.

[0034] When the grinding actuator 4 performs a short grinding stroke, the grinding load acquisition unit 2 sends the spindle power, current, or torque to the controller 3 in a continuous sampling manner. The controller 3 converts the sampled load into grinding power P(t) and subtracts the no-load power P0(t) from it to obtain the net grinding power. When the net grinding power is less than zero, it is counted as zero. The grinding energy Ei corresponding to the i-th short grinding stroke is calculated by the following formula: ; Where ti0 is the moment when the grinding tool 6 and the gate 21 begin to make effective contact, ti1 is the moment when the effective contact ends, and t is time. The start and end times of contact can be determined by the normal force threshold, the power increment threshold, or a combination of both.

[0035] After the short grinding stroke ends, the grinding tool 6 is lifted away from the workpiece, and the 3D vision acquisition unit 1 re-acquires the point cloud. The controller 3 completes registration with the same stable reference 26 and establishes a new residual volume field according to the same target cut surface 23 and mesh index. Lifting the tool ensures that the grinding tool 6 does not obstruct the gate area, and the unified stable reference and mesh index allow the difference between two adjacent volume fields to be calculated on a volume-by-volume basis.

[0036] Controller 3 determines the coverage area Ωi based on the tool envelope and trajectory position of the i-th short grinding stroke, and calculates the difference between two adjacent residual volume fields within this area. The removed volume ΔVi of the i-th short grinding stroke is obtained by subtracting the sum of the residual volumes of the same voxels after the stroke from the sum of the residual volumes of each voxel before the stroke. The local removal coefficient observation value kobs,i is calculated by the following formula: ; The unit of this observation is mm. 3 / J is used to characterize the amount of material removed per unit grinding energy within the covered area. When a short stroke covers multiple spatial grids, kobs,i can be weighted according to tool dwell time, normal force, or trajectory sweep area, and the weighted results can be written into the corresponding grids. In this way, the local removal coefficient has a clear spatial assignment, rather than a fixed value shared by the entire casting.

[0037] In this embodiment, controller 3 updates the local removal coefficients using an exponential forgetting method: ; Where j is the spatial grid index covered by the short grinding stroke, kold,j is the local removal coefficient before the update, know,j is the local removal coefficient after the update, and λ is the exponential forgetting factor of 0.80 to 0.98. Controller 3 also limits the single-cycle variation of know,j relative to kold,j to 10% to 30%. Exponential forgetting allows historical valid results to participate in the update together with the current results, while the variation limit avoids isolated load fluctuations or local point cloud errors causing abrupt changes in parameters for the next stroke.

[0038] If Ei is below the energy threshold required for effective contact, ΔVi is negative, point cloud registration is invalid, the coverage area is severely occluded, or the load record is incomplete, then this short stroke will not produce valid observations, and controller 3 will maintain the original local removal coefficients. This boundary processing ensures that the local removal coefficients are updated only by samples that correspond to the time, space, and processing object.

[0039] Controller 3 uses the updated local removal coefficient and the current remaining volume field as inputs for the next short grinding stroke, and performs candidate trajectory generation and spatial boundary checks again. Thus, each cycle operates in the sequence of current remaining material state, short-stroke energy input, actual volume change, local removal capability correction, and the next short stroke. Compared to adjusting the feed rate solely based on grinding force deviation, this sequence incorporates the post-processed material volume change into the control loop, enabling the differentiation of removal capability differences in different regions under the same load.

[0040] When the remaining height of the gate enters the range of 0.10–0.30 mm, controller 3 maintains the fine grinding parameters and compares the remaining height of two consecutive measurements. When the difference between the two consecutive remaining heights is not greater than the preset stability threshold, and the trajectory does not intrude into the root protection zone 25, controller 3 ends the grinding. If the remaining height enters the target range but the results of two consecutive measurements are still unstable, a short grinding stroke with low normal force is executed, or the point cloud is re-acquired. Using the two consecutive stability conditions can reduce the possibility of premature stopping caused by a single point cloud disturbance.

[0041] like Figure 4As shown, controller 3 extracts the fundamental frequency and its harmonic amplitude, impact component, or periodic variation from the grinding power time series, and extracts the brightness, area, direction distribution, and time-varying characteristics from the grinding spark image. Controller 3 correlates the detected abnormal moments with the pose of the grinding tool 6, the grinding wheel rotation angle, and the fixture position, and records the abnormal positions in the workpiece coordinate system, the grinding wheel phase coordinate system, and the fixture coordinate system, respectively.

[0042] The grinding spark images are acquired by an industrial camera positioned to the side of the grinding area. The industrial camera can be integrated into the 3D vision acquisition unit 1 or used as a stand-alone image acquisition device electrically connected to the controller 3. The field of view of the industrial camera covers the contact area between the grinding tool 6 and the gate 21, and its acquisition timestamp is aligned with the grinding power timestamp to map the spark features to the same short grinding stroke and grinding wheel phase.

[0043] If an anomaly repeatedly occurs at the same location in the workpiece coordinate system, but its phase relative to the grinding wheel and the fixture position are not fixed, the anomaly source is determined to be a workpiece defect. If the anomaly repeats at the same phase as the grinding wheel rotates, but its position relative to the workpiece changes, it is determined to be a grinding wheel state anomaly. If the anomaly corresponds to changes in the fixture position or clamping posture, it is determined to be a clamping anomaly. Anomaly samples that cannot be clearly classified are considered invalid samples, do not participate in the local removal coefficient update, and trigger tool lifting retesting.

[0044] like Figure 5 As shown, when the anomaly source is a workpiece defect, controller 3 reduces the cutting depth at the anomaly location and generates a detour trajectory, subsequently processing it again with a shorter stroke. When the anomaly source is an abnormal grinding wheel condition, controller 3 pauses the local removal coefficient update and causes the grinding actuator 4 to perform grinding wheel dressing, after which the no-load power and grinding wheel condition baseline are regained. When the anomaly source is a clamping anomaly, controller 3 causes the grinding tool 6 to retract from the casting 5 and locks the grinding actuator 4, resuming operation only after re-clamping and stabilization benchmark calibration are completed. The volume difference and grinding energy corresponding to grinding wheel condition anomalies and clamping anomalies are not used to update the local removal coefficient; therefore, non-material factors are not fixed as spatial material parameters.

[0045] To verify the closed-loop identification, root protection, and anomaly source separation effects of the machine vision-based adaptive grinding control system for casting gates, four sets of experiments were set up under the condition that the object specifications, equipment capabilities, environmental conditions, and measurement rules remained consistent, with each set completing 5 independent records. The short grinding stroke length of each set was 10–30 mm, the normal force in the rough grinding stage was 35–80 N, and the normal force in the fine grinding stage was 15–40 N. Set A used conventional grinding equipment and fixed parameters as a control; Set B enabled a combination of a stable reference constraint residual volume field, a local removal coefficient jointly updated by grinding energy and volume difference, and a short grinding stroke generated by rolling the residual volume field and local removal coefficient; Set C, based on Set B, enabled a variable width root protection zone driven by curvature, normal change, and point cloud confidence; Set D, based on Set C, enabled power harmonics, spark images, and three-coordinate repeatability anomaly source discrimination and contaminated sample blocking.

[0046] The experiment used volume removal prediction accuracy, grinding wheel wear stage profile pass rate, and casting body overcut rate as evaluation indicators. Specifically, volume removal prediction accuracy was used to evaluate the ability of the local removal coefficient to characterize the actual removed volume; the grinding wheel wear stage profile pass rate was used to evaluate the system's ability to maintain the target profile when the grinding wheel condition changes; and the casting body overcut rate was the percentage of the number of castings whose gate roots crossed the protection boundary and cut into the casting body out of the total number of castings inspected. Table 1 lists the measured results of each experiment.

[0047] Table 1. Item-by-item test results for each group Compared to Group A, Group B's average prediction accuracy for volume removal improved from 72.0% to 89.0%, indicating that the core combination of a stable benchmark, energy-volume difference pairing within the same short stroke, and rolling updates of spatial local removal coefficients can improve the prediction accuracy of the actual removed volume. Compared to Group B, Group C's average profile pass rate during the grinding wheel wear stage improved from 89.0% to 94.0%, indicating that the variable width root protection zone can maintain the trajectory boundary when point cloud quality and root geometry change. Compared to Group C, Group D's average body overcut rate decreased from 2.8% to 1.2%, indicating that the three-coordinate anomaly source discrimination and contaminated sample blocking can prevent grinding wheel or clamping anomalies from being written into local material removal parameters and reduce body overcut caused by erroneous updates.

[0048] In summary, this machine vision-based adaptive grinding control system for casting gates maintains a comparable spatial frame for continuous point clouds through a stable benchmark. It interlocks the load integration time window, trajectory coverage area, and remeasured volume changes via short-stroke operations. The system then updates the spatial local removal coefficients with energy-volume observations and immediately applies the updated results to subsequent constrained trajectories, forming a complete material removal closed loop. A root protection zone provides a safe interval at geometric boundaries that varies with point cloud quality, while anomaly source identification and invalid sample blocking maintain the physical attribution of updated data. This system can adapt to changes in casting allowance, material condition, and grinding wheel condition, while limiting the accumulation of under-grinding, over-grinding, and parameter drift.

[0049] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A machine vision-based adaptive grinding control system for casting gates, characterized in that: The machine vision-based adaptive grinding control system for casting gates includes a three-dimensional vision acquisition unit, a grinding load acquisition unit, a grinding actuator, and a controller. The controller is electrically connected to the three-dimensional vision acquisition unit, the grinding load acquisition unit, and the grinding actuator, respectively. The three-dimensional vision acquisition unit is used to acquire a three-dimensional point cloud containing the gate area and the casting stability reference. The casting stability reference is a geometric reference that can be repeatedly positioned before and after the gate is ground and does not disappear when the gate is removed. The grinding load acquisition unit is used to acquire the grinding load during the short grinding stroke performed by the grinding actuator. The controller is configured to: A local coordinate system is established based on the casting stability reference. A first three-dimensional point cloud is registered in the local coordinate system to determine the target cut surface. A first residual volume field is established based on the spatial relationship between the first three-dimensional point cloud and the target cut surface. Based on the first remaining volume field and the local removal coefficient corresponding to the spatial region in the first remaining volume field, a first short grinding stroke is generated and the grinding actuator is controlled to execute the first short grinding stroke. The grinding load obtained during the first short grinding stroke is converted into grinding power, and the grinding power is integrated over time to obtain the grinding energy corresponding to the first short grinding stroke. After the first short grinding stroke is completed, the three-dimensional vision acquisition unit is controlled to acquire the second three-dimensional point cloud, and the second three-dimensional point cloud is registered based on the same casting stability reference, and a second residual volume field is established. Calculate the volume difference between the first residual volume field and the second residual volume field within the spatial region covered by the first short grinding stroke, use the ratio of the volume difference to the grinding energy as the observed value of the local removal coefficient, and update the local removal coefficient corresponding to the spatial region based on the observed value of the local removal coefficient. Based on the second residual volume field and the updated local removal coefficient, at least one of the following is determined: trajectory length, feed rate, normal force, and endpoint of the next short grinding stroke. The next short grinding stroke is constrained by the spatial boundary of the target cut surface. After the next short grinding stroke is completed, the residual volume field establishment, grinding energy acquisition, and local removal coefficient update are repeated.

2. The machine vision-based adaptive grinding control system for casting gates according to claim 1, characterized in that: The controller is also configured to: Based on the root curvature abrupt change, surface normal change, and point cloud confidence of the gate region, a root protection strip of variable width is formed at the edge of the target cut surface; At least a portion of the root protection strip is designated as a no-entry zone to prevent grinding tools from entering, and the area adjacent to the no-entry zone is designated as a tangential sweep zone to restrict the movement of the grinding tool along the tangential direction of the target cutting surface. The width of the root protection strip is increased when the point cloud confidence decreases.

3. The machine vision-based adaptive grinding control system for casting gates according to claim 2, characterized in that: The controller controls the grinding actuator to switch between the rough grinding stage and the fine grinding stage based on the remaining height of the gate, the distance between the gate and the root protection strip, the confidence level of the local removal result, and the cumulative grinding energy. When the remaining height of the gate is 0.10 to 0.30 mm, and the difference between two consecutive remaining heights of the gate is not greater than a preset stable threshold, the controller controls the grinding actuator to terminate the grinding.

4. The machine vision-based adaptive grinding control system for casting gates according to claim 1, characterized in that: The controller sets the trajectory length of the short grinding stroke to 10–30 mm, sets the acquisition time interval of adjacent three-dimensional point clouds to 0.5–3.0 s, and controls the grinding actuator to apply a normal force of 35–80 N in the rough grinding stage and a normal force of 15–40 N in the fine grinding stage.

5. The machine vision-based adaptive grinding control system for casting gates according to claim 1, characterized in that: The controller updates the local removal coefficient using an exponential forgetting method, with an exponential forgetting factor of 0.80 to 0.98, and limits the single-cycle change of the updated local removal coefficient relative to the original local removal coefficient to 10% to 30%.

6. The machine vision-based adaptive grinding control system for casting gates according to claim 1, characterized in that: The machine vision-based adaptive grinding control system for casting gates also includes an image acquisition unit for acquiring images of grinding sparks, and the controller is further configured to: Extract the harmonic features of the grinding power and the image features of the grinding spark image, and map the anomalies represented by the harmonic features and the image features to the workpiece coordinate system, the grinding wheel phase coordinate system and the fixture coordinate system, respectively. Based on the positional repeatability of the anomaly in the workpiece coordinate system, the phase repeatability in the grinding wheel phase coordinate system, and the positional repeatability in the fixture coordinate system, the anomaly source is determined to be a workpiece defect, a grinding wheel condition abnormality, or a clamping abnormality; and the volume difference and grinding energy corresponding to the short grinding strokes determined to be the grinding wheel condition abnormality or the clamping abnormality are prevented from participating in the update of the local removal coefficient.

7. The machine vision-based adaptive grinding control system for casting gates according to claim 6, characterized in that: When the source of the anomaly is a workpiece defect, the controller controls the grinding actuator to perform detour grinding with reduced depth of cut at the abnormal location; when the source of the anomaly is an abnormal grinding wheel condition, the controller suspends the update of the local removal coefficient and controls the grinding actuator to perform grinding wheel dressing; when the source of the anomaly is a clamping abnormality, the controller controls the grinding actuator to retract and lock the tool.

8. The machine vision-based adaptive grinding control system for casting gates according to claim 1, characterized in that: The controller writes a unified clock timestamp to the 3D point cloud, the grinding load, and the short grinding stroke, so that each 3D point cloud retest result corresponds to its previous short grinding stroke and the grinding load of that short grinding stroke; the 3D vision acquisition unit is a structured light camera, a line laser scanner, or a multi-view camera, and the grinding load acquisition unit is configured to acquire spindle power, spindle current, or spindle torque.

9. The machine vision-based adaptive grinding control system for casting gates according to claim 1, characterized in that: The controller maintains the local removal coefficient before the update and controls the grinding actuator to reacquire the three-dimensional point cloud after lifting the tool when the grinding energy is lower than the energy threshold required for effective contact, the volume difference is negative, the three-dimensional point cloud registration is invalid, the spatial area covered by the first short grinding stroke is occluded, or the grinding load record is incomplete.

Images