Robot sand belt grinding process parameter autonomous decision method and system

Abstract

The application discloses a kind of robot abrasive belt grinding process parameter autonomous decision method and system.The method carries out unified time base processing to multi-source observation information, determines gate weight according to the local deviation degree of channel feature and cross-channel coherence relationship, obtains reliable grinding state feature;Tool state, contact state and task state representing the difference of abrasive belt cutting ability, local contact state and processing task are updated jointly;Material removal, surface quality, temperature rise, abrasive belt life and prediction uncertainty are predicted for candidate process parameter action;Accordingly, risk constraint process parameter scheduling model and robust safe action set are constructed, and joint decision is made on normal contact force, robot feed speed, abrasive belt linear velocity, contact posture and trajectory overlap rate, and safety correction and set projection are carried out on candidate action.The application can adapt to the change of abrasive belt state, curved surface contact condition and processing task, and reduce the risk of over-grinding, under-grinding, quality out-of-tolerance and execution constraint violation.

Description

Technical Field

[0001] This invention belongs to the field of robotic intelligent machining and flexible grinding technology, and more specifically, relates to a method and system for autonomous decision-making of process parameters in robotic belt grinding. Background Technology

[0002] Robotic belt grinding is widely used for precision finishing and surface uniform processing of complex free-form components such as aero-engine blades, integral bladed disks, orthopedic implants, automotive parts, and molds. This process belongs to flexible contact grinding, and the material removal result depends not only on nominal process parameters such as belt grit size, normal contact force, robot feed speed, and belt linear speed, but also on the combined influence of belt condition, contact geometry, and machining task.

[0003] Among these factors, belt wear, passivation, clogging, abrasive grain breakage, and the decline in effective cutting capability exhibit significant temporal evolution characteristics; the local curvature, contact posture, contact pressure distribution, contact spot migration, and trajectory overlap history of complex surfaces cause differences in the geometric domain; and workpiece material, initial allowance, target removal amount, and surface quality target cause differences in the task domain. Therefore, the same set of process parameters may produce significantly different material removal amounts, surface roughness, and contact temperature rise under different processing times, different surface regions, or different task conditions.

[0004] The selection of process parameters for existing robotic belt grinding typically relies on human experience, offline process experiments, or static material removal models. While existing technologies can improve machining quality in localized aspects such as offline programming, path planning, force control compensation, belt wear prediction, or detection feedback, they usually only model or compensate for single factors. It is difficult to uniformly express the temporal drift of belt capabilities, the geometric drift of complex surface contact states, and the task domain drift of materials and machining targets as a process state that can be updated online.

[0005] On the other hand, existing solutions often focus on optimizing material removal or a single quality index, neglecting the coupling and adjustment of parameters such as normal contact force, robot feed speed, sanding belt linear speed, contact posture, and trajectory overlap rate. Furthermore, under conditions of insufficient samples, new materials, new geometries, or sensor anomalies, the prediction results remain uncertain. If parameters are output solely based on the predicted mean or fixed empirical boundaries, issues such as over-grinding, under-grinding, out-of-tolerance surface quality, excessive temperature rise, contact impact, or violation of robot motion boundaries can easily occur.

[0006] Therefore, there is an urgent need for an autonomous decision-making technology that can integrate multi-source observation information, unify the representation tools, contact and task states, and explicitly introduce prediction uncertainties into the joint scheduling and safety correction process of multiple process parameters, so as to improve the adaptability, stability and consistency of robotic belt grinding of complex curved surfaces. Summary of the Invention

[0007] This invention aims to solve the problems in existing robotic belt grinding where it is difficult to uniformly perceive drift in the time domain, geometric domain, and task domain; multiple process parameters are difficult to jointly schedule online based on process status; and under conditions of prediction uncertainty, it is difficult to simultaneously ensure material removal quality, surface quality, temperature rise, belt life, and execution safety.

[0008] To achieve the above objectives, the present invention provides an autonomous decision-making method for process parameters in robotic belt grinding, comprising: S100. Acquire multi-source observation information, target processing requirements and process constraints during the robot belt grinding process, and synchronize the multi-source observation information according to a unified time reference to form time-series observation data; S200. Based on the local deviation of each sensor channel feature from the baseline feature and the cross-channel coherence relationship between different sensor channels, determine the gating weight of each sensor channel, and fuse the original timing features and smoothed baseline features according to the gating weight to obtain the grinding state features after reliability processing. S300. Based on the grinding state characteristics, the process state at the previous moment, the safety process parameter actions executed at the previous moment, and the current task information, jointly update the tool state, contact state, and task state to obtain the three-domain process state; S400. For candidate process parameter actions, predict the grinding response and prediction uncertainty based on the three-domain process state; S500. Based on the grinding response, the prediction uncertainty, the target machining requirements, and the process constraints, construct a risk-constrained process parameter scheduling model and a robust set of safe actions; S600. Generate candidate process parameter actions using the risk-constrained process parameter scheduling model, correct the candidate process parameter actions based on the prediction uncertainty, remaining life of the sand belt and process safety margin, and project the corrected candidate process parameter actions onto the robust safety action set to obtain safe process parameter actions. S700. Control the robot and belt grinding actuator to execute the safety process parameter actions, and continuously update the three-domain process status, the grinding response and the safety process parameter actions based on new observation information acquired during the execution process.

[0009] Further, in S200, time-series features are extracted from the observation sequences of at least two of the sensing channels, namely force / torque, temperature, acoustic emission, vibration, robot motion state, sanding belt linear velocity, spindle or motor current, visual measurement information and contour measurement information, within the sliding time window to obtain the corresponding channel features; the smooth baseline features are obtained by updating the channel features in time sequence.

[0010] Furthermore, S200 further includes: calculating the local deviation of the current channel feature relative to the corresponding smoothed baseline feature, and calculating the cross-channel coherence score based on the physical association between at least two sensing channels; inputting the local deviation, the cross-channel coherence score, and their interaction term into a gating function to obtain the gating weight of each sensing channel; and fusing the original temporal features and the smoothed baseline features based on the gating weight.

[0011] Furthermore, in step S300, the tool state characterizes the change in the effective cutting capability of the abrasive belt with processing time, the contact state characterizes the change in contact state caused by local surface geometry, contact posture and trajectory history, and the task state characterizes the task differences caused by workpiece material, local allowance and quality target.

[0012] Furthermore, the tool state includes at least one of the following: belt wear, passivation, blockage, abrasive breakage, local recovery, and effective cutting capability; the contact state includes at least one of the following: local curvature, contact posture, normal force, contact pressure distribution, contact spot change, and trajectory overlap history; the task state includes at least one of the following: workpiece material, local structural features, initial allowance, target removal amount, and quality preference.

[0013] Furthermore, in S300, the tool state is updated based on grinding time, executed safety process parameter actions, process observation information, and the observation reliability weight; the contact state is updated based on robot end-effector pose, normal force, local geometric information, and trajectory history; the task state is determined based on workpiece material, local allowance, surface quality target, and at least one of historical processing data, trial grinding data, or detection feedback; and the tool state, the contact state, and the task state are jointly corrected.

[0014] Furthermore, the grinding response in step S300 includes at least one of material removal depth, material removal error, surface roughness, contact temperature rise, peak contact force, remaining belt life, and constraint violation risk; the response prediction model simultaneously outputs the predicted value of the grinding response and the prediction uncertainty, and selects or weights multiple sub-models according to different tool states, contact states, or task states.

[0015] Further, the objective function of the risk-constrained process parameter scheduling model in step S500 includes at least two of the following: material removal error, surface quality error, temperature risk, processing time, belt life consumption, parameter change smoothness constraint, and prediction uncertainty risk. The robust safety action set is determined by at least two of the following constraints: material removal, surface quality, temperature, normal force, robot feed speed, belt linear speed, contact posture, trajectory overlap rate, robot reachability, and collision boundary. When the prediction uncertainty is higher than a preset uncertainty threshold, the remaining belt life is lower than a preset life threshold, or the process safety margin is lower than a preset margin threshold, the local compensation amplitude is reduced, the planning horizon is shortened, the process parameter change rate is limited, or the safety margin is increased, and the corrected candidate process parameter actions are projected onto the robust safety action set.

[0016] According to a second aspect of the present invention, an autonomous decision-making system for process parameters in robotic belt grinding is provided, implemented using the aforementioned autonomous decision-making method for process parameters in robotic belt grinding, comprising: The information acquisition unit is used to acquire multi-source observation information, target processing requirements, and technological constraints. A reliability preprocessing unit is used to determine the gating weights and obtain grinding state characteristics based on the degree of local deviation and cross-channel coherence. A three-domain process state estimation unit is used to jointly update tool state, contact state, and task state. The grinding response and uncertainty prediction unit is used to predict the grinding response and uncertainty for candidate process parameters. The risk constraint parameter scheduling unit is used to construct a risk constraint process parameter scheduling model and a robust set of safety actions; The safety correction and projection unit is used to correct the candidate process parameter actions based on the prediction uncertainty, the remaining life of the abrasive belt and the process safety margin, and project the corrected candidate process parameter actions onto the robust safety action set. And a process execution and rolling update unit, used to execute the safety process parameters obtained after projection and to perform rolling updates based on new observation information.

[0017] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for autonomous decision-making of process parameters for robotic belt grinding.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention combines the degree of local deviation with the cross-channel coherence relationship for multi-source observation gating, which can suppress the influence of isolated noise, abnormal disturbances and channel asynchrony on state estimation while retaining effective transient features with physical consistency, thereby improving the reliability of online grinding state identification.

[0019] (2) The present invention uses tool state, contact state and task state to form a three-domain process state, and transforms the time domain drift caused by belt wear and cutting capability changes, the geometric domain drift caused by complex curved surface contact geometry, and the task domain drift caused by material and processing target changes into a state that can be jointly updated online, providing a unified state basis for multi-parameter decision making.

[0020] (3) The response prediction module of the present invention outputs the grinding response and prediction uncertainty at the same time, and introduces the prediction uncertainty into the risk constraint scheduling and robust safety action set, so that the system can adaptively increase the safety margin and reduce the risk of quality deviation and execution boundary violation in areas with insufficient samples, new materials, new geometry or high risk areas.

[0021] (4) The present invention performs joint scheduling of normal contact force, robot feed speed, sanding belt linear speed, contact posture and trajectory overlap rate, and projects the corrected candidate process parameter actions to the robust safety action set, so as to meet the quality constraints and safety constraints while taking into account processing efficiency, parameter change smoothness and sanding belt life. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the autonomous decision-making method for robot belt grinding process parameters according to an embodiment of the present invention. Figure 2 This is a diagram illustrating the decision-making framework for three-domain process state estimation and risk constraint parameters according to an embodiment of the present invention. Figure 3 This is a diagram showing the composition and signal relationships of a robotic belt grinding platform according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the local contact relationship and combined process parameters in belt grinding according to an embodiment of the present invention; Figure 5 This is a unit diagram of the autonomous decision-making system for robot belt grinding process parameters according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, the technical features involved in the various embodiments described below can be combined with each other as long as they do not conflict with each other. Example 1

[0024] Please refer to Figures 1 to 4 This embodiment provides a method for autonomous decision-making of process parameters in robotic belt grinding. The method is executed by an autonomous process parameter decision-making device as the upper-level process parameter scheduling module, updating process parameters in units of trajectory segments or planning cycles. The lower-level motion control and force control are executed by the robot controller and the belt grinding unit controller.

[0025] like Figure 3 As shown, the robotic belt grinding platform includes an industrial robot 1, a belt grinding unit 2, a complex curved workpiece 3, a robot controller 4, a process parameter autonomous decision-making device 5, a multi-source observation component 6, and a contour / vision inspection module 7. The belt grinding unit 2 includes a contact wheel 21 and an abrasive belt 22; the multi-source observation component 6 includes a force / torque sensor 61, a temperature sensor 62, and an abrasive belt speed feedback module 63. The process parameter autonomous decision-making device 5 generates safe process parameter actions based on multi-source observation information and detection feedback, and sends these safe process parameter actions to the robot controller 4.

[0026] like Figure 4 As shown, the abrasive belt 22 is wound around the contact wheel 21 and forms flexible contact with the complex curved workpiece 3. The jointly adjusted process parameters include the normal contact force Fn, the robot feed speed vf, the abrasive belt linear speed vs, the contact posture θ, and the trajectory overlap rate ρ of adjacent grinding trajectories; the contact posture θ and the trajectory overlap rate ρ together affect the local contact spots and the distribution of material removal.

[0027] The method includes steps S100 to S700.

[0028] S100 acquires multi-source observation information, target processing requirements, and process constraints during the robot belt grinding process.

[0029] During robotic belt grinding, multi-source observation information related to the grinding state is collected, and the observation information from each sensor channel is synchronized according to a unified time base to form time-series observation data for state estimation. The multi-source observation information can be represented as: (1) In the formula, Represents the set of multi-source observations at time t; This represents the observation signal of the k-th sensing channel at time t; Indicates the number of sensor channels; Indicates the sensor channel number; Indicates the sampling time or the current control cycle.

[0030] Furthermore, the multi-source observation information at consecutive times is constructed into a sliding time window: (2) In the formula, This represents the multi-channel time-series observation data corresponding to the current time t; This represents the set of multi-source observations at the corresponding time. Indicates the length of the sliding time window; Indicates the starting sampling time of the current window.

[0031] In this embodiment, the target processing requirements may include the target material removal amount, allowable contour deviation, surface roughness target, local allowance distribution, and processing efficiency preference; the process constraints may include normal force range, temperature upper limit, speed range, attitude range, trajectory overlap rate range, robot reachability, collision boundary, sanding belt remaining life constraint, and parameter change rate constraint.

[0032] S200 performs reliability preprocessing on multi-source observation information and extracts effective features related to the grinding state.

[0033] Temporal features are extracted from the observation sequence of the k-th sensing channel within the current sliding time window to obtain the corresponding channel features: (3) In the formula, This represents the temporal characteristics of the k-th sensing channel; This represents the feature extraction function for the k-th sensing channel; This represents the observation sequence of the k-th sensing channel within the current sliding time window; Indicates the sensor channel number.

[0034] Channel gating weights are calculated based on the local deviation of channel features from the smoothed baseline features and the cross-channel coherence scores between different channels: (4) In the formula, This represents the gating weight of the k-th sensing channel; This indicates the degree of local deviation of the current feature of the k-th sensing channel relative to the baseline feature; Indicates the cross-channel coherence score; Indicates the gating weight coefficient; This represents the gating threshold for the k-th sensing channel; This represents the gated function.

[0035] The original temporal features and smoothed baseline features are weighted and fused according to the gating weights to obtain the grinding state features after reliability processing: (5) In the formula, This indicates the grinding state characteristics after reliability processing; This represents the original feature set composed of the features of each sensing channel; This represents the corresponding set of smoothed baseline features; This represents the gating matrix composed of the gating weights of each channel; This indicates element-wise multiplication.

[0036] When a transient signal exhibits both local significance and cross-channel consistency, its characteristic representation is preserved. When anomalies primarily occur in a single channel and lack cross-channel consistency, their weight is reduced or replaced with smoothed baseline features. Furthermore, a heavy-tailed residual model can be used to calculate the observation reliability weights, thereby reducing the impact of outlier observations on subsequent state updates.

[0037] S300 establishes and updates the three-domain process state, which consists of tool state, contact state, and task state.

[0038] The state of the three-domain process can be represented as follows: (6) In the formula, Represents the state of the three-domain process at time t; It indicates the tool status and is used to characterize changes in abrasive belt wear, passivation, clogging, abrasive grain breakage, and cutting capability; It represents the contact state and is used to characterize local curvature, contact posture, contact pressure distribution, contact spot changes, and trajectory overlap history; It indicates the task status and is used to characterize the workpiece material, local allowance, target removal amount, and quality preference.

[0039] The tool status is updated online based on grinding time, executed process parameters, process observation information, and observation reliability weights to characterize the gradual and sudden changes in the abrasive belt capability; the contact status is updated based on the robot end-effector pose, normal force, local geometric information, and trajectory history to characterize the impact of contact patch migration, curvature changes, and trajectory overlap on material removal; and the task status is determined based on workpiece material, local allowance, surface quality target, historical processing data, a small amount of trial grinding data, or detection feedback.

[0040] To incorporate the coupling relationship between changes in abrasive belt capability, contact geometry, and machining task into a unified state representation, the three-domain process state is jointly updated based on the reliability-processed grinding state characteristics, the previous process state, executed safety process parameters, and current task information: (7) In the formula, This represents the state update function for a three-domain process. This indicates the grinding state characteristics after reliability processing; This indicates the state of the three-domain process at the previous moment or in the previous planning cycle; This indicates the safety process parameter actions that were executed at the previous moment or in the previous planning cycle. This indicates current task information, partial margin information, or detection feedback information; This indicates the updated state of the three-domain process.

[0041] The state update function for a three-domain process can employ neural networks, probabilistic state-space models, Bayesian filters, graphical models, ensemble learning models, physical constraint models, or combinations thereof. This invention does not limit its specific implementation.

[0042] S400 predicts material removal response, mass response, and prediction uncertainty based on a three-domain process state.

[0043] For candidate process parameters, the response prediction model predicts the grinding response and uncertainty based on a three-domain process state: (8) In the formula, This represents the predicted grinding response; This indicates uncertainty in forecasting; This represents a response prediction model; Indicates the current state of the three-domain process; Indicates the action of candidate process parameters; This represents the model parameters of the grinding response prediction model.

[0044] In one implementation, the predicted grinding response is expressed as: (9) In the formula, Indicates the depth of material removal; Indicates surface roughness; Indicates the temperature rise upon contact; Indicates peak contact force; Indicates the remaining lifespan of the abrasive belt or lifespan-related indicators; This represents the transpose of a vector.

[0045] The prediction uncertainty is used to characterize the prediction risk caused by observation noise, operating condition randomness, insufficient samples, new material conditions, or new geometric conditions. The response prediction module can select or weight multiple sub-models according to different grinding states to adapt to the nonlinear grinding response under different wear stages of the abrasive belt, different materials, different contact postures, and different local curvature conditions.

[0046] S500 constructs a risk-constrained process parameter scheduling model and a robust set of safety actions.

[0047] Based on the target processing requirements, processing efficiency preferences, abrasive belt life preferences, and prediction uncertainties, a comprehensive cost function is constructed: (10) In the formula, This represents the comprehensive cost of process parameter action 'a' at time 't'. This indicates the material removal error term; Indicates the surface quality error term; Indicates temperature risk items; Indicates processing time or cycle time; This indicates the lifespan consumption of the sanding belt; This represents the constraint term for the smoothness of parameter changes; This indicates a risk item related to forecast uncertainty or constraint violation; Indicates the corresponding weight coefficient; Indicates the action of candidate process parameters; This indicates the current moment or the current planning period.

[0048] Risk constraints can employ conditional risk value constraints, opportunity constraints, safety margin constraints based on predictive uncertainty, or combinations thereof, to limit the risks of material removal errors exceeding limits, surface quality deviations exceeding tolerances, contact temperature rise exceeding limits, or execution boundary violations under high-uncertainty operating conditions. Further, a robust set of safety actions should be established: (11) In the formula, It represents the set of robust safety actions determined by quality constraints, safety constraints, and execution boundary constraints at the current moment or in the current planning period; Indicates the action space of optional process parameters; Indicates the action of candidate process parameters; Represents quality, safety, or enforcement boundary constraint functions; Indicates the material removal error constraint; Indicates surface quality constraints; Indicates temperature constraint; Indicates normal force constraint; This indicates the robot's feed speed constraint; Indicates the linear velocity constraint of the abrasive belt; Indicates contact posture constraints; Indicates trajectory overlap rate constraint; Indicates the constraint type index; This represents the set of constraint types corresponding to material removal, surface quality, temperature, normal force, robot feed speed, sanding belt linear speed, contact posture, and trajectory overlap rate.

[0049] The robust safety action set forms a safety margin based on the predicted mean and prediction uncertainty, ensuring that the generated process parameter actions meet quality constraints, safety constraints, and robot execution boundary constraints at a set confidence level.

[0050] S600 generates candidate process parameter actions and performs safety corrections on the candidate actions.

[0051] The candidate process parameter action can be expressed as: (12) In the formula, The vector representing the process parameter action at time t; Indicates the normal contact force; Indicates the robot's feed speed; Indicates the linear velocity of the abrasive belt; Indicates contact posture; Indicates the trajectory overlap rate; This represents the transpose of a vector.

[0052] The candidate process parameters are adjusted based on the prediction uncertainty, the remaining life of the abrasive belt, and the process safety margin. When the prediction uncertainty is high, the remaining life of the abrasive belt is low, or the constraint margin is small, the local compensation amplitude is reduced, the planning horizon is shortened, the parameter change rate is limited, or the safety margin is increased. When the prediction uncertainty is low and the constraint margin is large, the processing efficiency is improved or the local error compensation capability is enhanced while satisfying the constraints.

[0053] The corrected candidate process parameter actions are projected onto the robust safety action set to obtain the safety process parameter actions: (13) In the formula, This indicates the safety process parameters that are ultimately output to the robot controller or belt grinding actuator. This indicates the candidate process parameter action after adjustments for predicted uncertainty, remaining belt life, and process safety margin. This indicates a safety correction operation that projects candidate process parameter actions onto a robust set of safety actions.

[0054] The safety correction is not a simple, independent limitation of candidate process parameter actions, but rather a joint projection based on a robust set of safe actions determined by quality constraints, safety constraints, and robot execution boundary constraints. Furthermore, variations in normal force, velocity, attitude, and trajectory overlap between adjacent trajectory segments or adjacent control cycles are restricted to prevent abrupt changes in process parameters that could lead to contact impact, over-wearing, under-wearing, or surface quality fluctuations.

[0055] S700 executes the current safety process parameters and updates them on a rolling basis based on new observation information.

[0056] The robot controller and belt grinding actuator execute the current trajectory segment or control cycle according to the safe process parameters. During execution, new multi-source observation information is continuously collected, and the tool status, contact status, task status, response prediction results, and parameter scheduling results are updated based on the new process observation information and available online or offline detection feedback. S200 to S700 is repeated in a rolling time domain manner, so that the process parameters are continuously and adaptively adjusted according to changes in belt status, local contact status, and task requirements.

[0057] In a specific machining scenario, the robot performs segmented grinding on complex curved surface parts such as blades, covering different abrasive belt grit sizes, different abrasive belt wear stages, different local curvatures, different contact postures, and different trajectory overlap rates. The system uses a small amount of task data or historical machining data to determine the task status and quickly adapts to new materials, new local geometries, or new machining target conditions. For abrasive belts with extreme wear, tasks with strong extrapolation, or high-risk machining areas, machining risks are reduced by increasing safety margins, reducing feed rates, limiting normal forces, or shortening the planned field of view.

[0058] This embodiment is not limited to blade grinding, but can also be applied to robotic belt grinding, polishing or finishing processes of impellers, molds, aerospace structural parts, orthopedic implants, automotive curved parts, weld repair parts, castings and forgings and other complex curved parts. Example 2

[0059] Please refer to Figure 5 This embodiment provides an autonomous decision-making system for robotic belt grinding process parameters, including: The information acquisition unit 51 is used to acquire multi-source observation information, target processing requirements and process constraints during the robot belt grinding process, and form time-series observation data according to a unified time reference; The reliability preprocessing unit 52 is used to determine the gating weights based on the local deviation of the characteristics of each sensing channel and the cross-channel coherence relationship, and to fuse the original timing features and smoothed baseline features to obtain the grinding state features after reliability processing. The three-domain process state estimation unit 53 is used to jointly update the tool state, contact state and task state based on the grinding state characteristics, the process state at the previous moment, the safety process parameter actions executed at the previous moment and the current task information. Grinding response and uncertainty prediction unit 54 is used to predict one or more grinding responses and uncertainties based on three-domain process states and candidate process parameters, including material removal, surface quality, contact temperature rise, peak contact force and remaining belt life. The risk-constrained parameter scheduling unit 55 is used to construct a risk-constrained process parameter scheduling model and a robust safety action set based on grinding response, prediction uncertainty, target machining requirements and process constraints, and to generate candidate process parameter actions. The safety correction and projection unit 56 is used to correct the candidate process parameter actions based on the prediction uncertainty, the remaining life of the sand belt and the process safety margin, and project the corrected candidate process parameter actions onto the robust safety action set to obtain the safe process parameter actions. The process execution and rolling update unit 57 is used to output safe process parameter actions to the robot controller and the belt grinding actuator, and to continuously update the three-domain process status, grinding response and safe process parameter actions based on the new observation information obtained during the execution process.

[0060] Each of the above units can be implemented by one or more processors executing programs stored in memory, or by hardware circuits, programmable logic devices, or a combination of hardware and software. The specific functions of each unit can be found in Embodiment 1, and will not be repeated here. Example 3

[0061] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it can implement any step of the autonomous decision-making method for robot belt grinding process parameters described in Embodiment 1.

[0062] The computer-readable storage medium may include a read-only memory, random access memory, solid-state drive, portable hard drive, USB flash drive, magnetic disk, or optical disk, or other media capable of storing program code. For further descriptions of the computer-readable storage medium provided in this embodiment, please refer to the above method and system embodiments.

[0063] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention; all modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for autonomous decision-making of process parameters in robotic belt grinding, characterized in that, include: S100. Acquire multi-source observation information, target processing requirements and process constraints during the robot belt grinding process, and synchronize the multi-source observation information according to a unified time reference to form time-series observation data; S200. Based on the local deviation of each sensor channel feature from the baseline feature and the cross-channel coherence relationship between different sensor channels, determine the gating weight of each sensor channel, and fuse the original timing features and smoothed baseline features according to the gating weight to obtain the grinding state features after reliability processing. S300. Based on the grinding state characteristics, the process state at the previous moment, the safety process parameter actions executed at the previous moment, and the current task information, jointly update the tool state, contact state, and task state to obtain a three-domain process state; wherein, the tool state characterizes the change in the effective cutting capability of the abrasive belt with processing time, the contact state characterizes the change in contact state caused by local surface geometry, contact posture, and trajectory history, and the task state characterizes the task differences caused by workpiece material, local allowance, and quality target; S400. For candidate process parameter actions, predict the grinding response and prediction uncertainty based on the three-domain process state; S500. Based on the grinding response, the prediction uncertainty, the target machining requirements, and the process constraints, construct a risk-constrained process parameter scheduling model and a robust set of safe actions; S600. Generate candidate process parameter actions using the risk-constrained process parameter scheduling model, correct the candidate process parameter actions based on the prediction uncertainty, remaining life of the sanding belt, and process safety margin, and project the corrected candidate process parameter actions onto the robust safety action set to obtain safe process parameter actions; the safe process parameter actions include at least two of the following: normal contact force, robot feed speed, sanding belt linear speed, contact posture, and trajectory overlap rate. S700. Control the robot and belt grinding actuator to execute the safety process parameter actions, and continuously update the three-domain process status, the grinding response and the safety process parameter actions based on new observation information acquired during the execution process.

2. The autonomous decision-making method for robot belt grinding process parameters according to claim 1, characterized in that, In step S200, time-series features are extracted from the observation sequences of at least two of the following sensing channels within the sliding time window: force / torque, temperature, acoustic emission, vibration, robot motion state, sanding belt linear velocity, spindle or motor current, visual measurement information, and contour measurement information, to obtain the corresponding channel features; the smooth baseline features are obtained by updating the channel features in chronological order.

3. The autonomous decision-making method for robot belt grinding process parameters according to claim 2, characterized in that, The S200 further includes: calculating the local deviation of the current channel feature relative to the corresponding smoothed baseline feature, and calculating the cross-channel coherence score based on the physical association between at least two sensing channels; inputting the local deviation, the cross-channel coherence score, and their interaction term into a gating function to obtain the gating weight of each sensing channel; and fusing the original temporal features and the smoothed baseline features based on the gating weight.

4. The autonomous decision-making method for robot belt grinding process parameters according to claim 3, characterized in that, When the transient characteristics of a sensing channel simultaneously satisfy both the local significance condition and the cross-channel consistency condition, the corresponding gating weight is increased to preserve the transient characteristics; when anomalies occur in a single sensing channel and do not satisfy the cross-channel consistency condition, the corresponding gating weight is decreased; and a heavy-tailed residual model is used to determine the observation reliability weight to reduce the impact of outlier observations on the state update of the three-domain process.

5. The autonomous decision-making method for robot belt grinding process parameters according to claim 1, characterized in that, The tool state includes at least one of the following: belt wear, passivation, blockage, abrasive breakage, local recovery, and effective cutting capability; the contact state includes at least one of the following: local curvature, contact posture, normal force, contact pressure distribution, contact spot change, and trajectory overlap history; the task state includes at least one of the following: workpiece material, local structural features, initial allowance, target removal amount, and quality preference.

6. The autonomous decision-making method for robot belt grinding process parameters according to claim 5, characterized in that, In step S300, the tool state is updated based on grinding time, executed safety process parameter actions, process observation information, and the observation reliability weight; the contact state is updated based on robot end-effector pose, normal force, local geometric information, and trajectory history; the task state is determined based on workpiece material, local allowance, surface quality target, and at least one of historical processing data, trial grinding data, or detection feedback; and the tool state, the contact state, and the task state are jointly corrected.

7. The autonomous decision-making method for robot belt grinding process parameters according to claim 1, characterized in that, The grinding response includes at least one of the following: material removal depth, material removal error, surface roughness, contact temperature rise, peak contact force, remaining belt life, and constraint violation risk; the response prediction model simultaneously outputs the predicted value of the grinding response and the prediction uncertainty, and selects or weights multiple sub-models according to different tool states, contact states, or task states.

8. The autonomous decision-making method for robot belt grinding process parameters according to claim 7, characterized in that, The objective function of the risk-constrained process parameter scheduling model includes at least two of the following: material removal error, surface quality error, temperature risk, processing time, belt life consumption, parameter change smoothness constraint, and prediction uncertainty risk. The robust safety action set is determined by at least two of the following constraints: material removal, surface quality, temperature, normal force, robot feed speed, belt linear speed, contact posture, trajectory overlap rate, robot reachability, and collision boundary. When the prediction uncertainty is higher than a preset uncertainty threshold, the remaining belt life is lower than a preset life threshold, or the process safety margin is lower than a preset margin threshold, the local compensation amplitude is reduced, the planning horizon is shortened, the process parameter change rate is limited, or the safety margin is increased, and the corrected candidate process parameter actions are projected onto the robust safety action set.

9. An autonomous decision-making system for process parameters in robotic belt grinding, characterized in that, include: The information acquisition unit is used to acquire multi-source observation information, target processing requirements, and technological constraints. A reliability preprocessing unit is used to determine the gating weights and obtain grinding state characteristics based on the degree of local deviation and cross-channel coherence. A three-domain process state estimation unit is used to jointly update tool state, contact state, and task state. The grinding response and uncertainty prediction unit is used to predict the grinding response and uncertainty for candidate process parameters. The risk constraint parameter scheduling unit is used to construct a risk constraint process parameter scheduling model and a robust set of safety actions; The safety correction and projection unit is used to correct the candidate process parameter actions based on the prediction uncertainty, the remaining life of the abrasive belt and the process safety margin, and project the corrected candidate process parameter actions onto the robust safety action set. And a process execution and rolling update unit, used to execute the safety process parameters obtained after projection and to perform rolling updates based on new observation information.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the autonomous decision-making method for robot belt grinding process parameters as described in any one of claims 1-8.

Images