Grinding and polishing process monitoring and machining parameter self-adaptive regulation and control method and system

Abstract

The invention relates to the technical field of industrial robot precision machining, in particular to a grinding and polishing process monitoring and machining parameter self-adaption regulation and control method and system which are suitable for metal part surface grinding, composite material polishing, mold cavity precision finishing and other scenes. Real-time surface quality monitoring, machining force dynamic optimization, tool track intelligent adjustment and machining effect online evaluation in the grinding and polishing process can be achieved, and the problems that in traditional grinding, the surface quality consistency is poor, tool abrasion is fast, and the machining efficiency and precision are difficult to balance are solved.

Description

Technical Field

[0001] This invention relates to the field of precision machining technology for industrial robots, and in particular to a method and system for monitoring the grinding and polishing process and adaptively controlling the processing parameters. Background Technology

[0002] Automated grinding and polishing using industrial robots is a key process for improving product surface quality in high-end manufacturing. Its processing effect directly impacts the product's appearance, assembly performance, and lifespan. Currently, there are four major technical bottlenecks in grinding and polishing process control: First, surface quality inspection is lagging. Traditional methods rely on manual visual inspection or offline testing, which cannot assess key indicators such as surface roughness and gloss in real time, leading to batch quality problems. Second, the precision of processing force control is low. Conventional fixed-parameter control strategies are prone to over-grinding or under-grinding when faced with varying material hardness and surface curvature. Third, tool trajectory adaptability is poor. Traditional pre-programmed trajectories cannot dynamically adjust according to real-time surface conditions, resulting in uneven processing quality for complex curved surfaces. Fourth, tool wear monitoring is insufficient. The lack of real-time perception of the grinding head's wear status leads to a gradual decline in processing quality as the tool wears down.

[0003] In existing technologies, some solutions enhance surface inspection capabilities by adding vision systems, but lack closed-loop correlation with processing parameters; some solutions employ simple force control algorithms, which can only adapt to single working conditions; and some solutions rely on empirical parameter settings, making it difficult to achieve flexible production of multiple varieties. Therefore, there is an urgent need to break through traditional control modes and construct a grinding and polishing control system that integrates multi-source perception and intelligent decision-making. Summary of the Invention

[0004] This invention provides a method and system for monitoring and adaptively controlling processing parameters during the grinding and polishing process. Through a four-layer architecture of multi-sensor fusion perception, real-time surface quality assessment, intelligent optimization of processing parameters, and dynamic adjustment of tool trajectory, it achieves high-precision, high-efficiency, and high-consistency automated grinding and polishing. The core innovation lies in proposing four key algorithms: a real-time surface quality assessment algorithm based on multispectral vision, an adaptive control algorithm for dynamic matching of processing force and material removal, a trajectory compensation optimization algorithm considering tool wear, and a processing parameter decision algorithm for multi-objective collaborative optimization. These algorithms can significantly improve the quality and efficiency of grinding and polishing.

[0005] A first aspect of this invention provides a method for monitoring and adaptively controlling processing parameters during the grinding and polishing process, comprising the following steps:

[0006] Multi-source information acquisition during the polishing process: Deploy a vision-force-acoustic fusion sensing system to acquire workpiece surface images, processing contact forces, tool status, acoustic signals and robot motion parameters in real time;

[0007] Real-time surface quality assessment: Based on the collected data, the surface roughness and gloss are detected online through a multispectral vision-based real-time surface quality assessment algorithm;

[0008] Intelligent control of processing force: Based on the surface quality assessment results, an adaptive control algorithm for dynamic matching of processing force and material removal amount is activated to achieve precise control of processing force;

[0009] Tool trajectory and parameter optimization: Combining tool wear status, the motion trajectory and machining parameters are dynamically adjusted by using a trajectory compensation optimization algorithm that takes tool wear into account and a machining parameter decision algorithm that performs multi-objective collaborative optimization.

[0010] Processing effect verification and closed-loop optimization: Through full-process data collection and analysis, the algorithm model is iteratively optimized to continuously improve processing quality and efficiency.

[0011] A second aspect of the present invention provides a system for implementing the above method, comprising:

[0012] Multi-source sensing units: multispectral vision system (2 million pixels resolution, spectral range 400-1000nm), six-dimensional force sensor (range 0-100N, accuracy ±0.1%FS), acoustic sensor, speed sensor, temperature sensor;

[0013] Data processing unit: GPU acceleration module (visual data processing), real-time signal processing module (force and acoustic signals), industrial computer (data fusion and algorithm execution);

[0014] Intelligent decision-making unit: Deploys four core algorithms to achieve surface quality assessment, processing force control, trajectory optimization, and parameter decision-making;

[0015] Robot control unit: High-precision motion controller, supporting force / position hybrid control, control cycle ≤1ms;

[0016] Human-computer interaction unit: Real-time display of processing parameters, surface quality assessment results and system status, supporting parameter setting and manual intervention. Attached image description:

[0017] Figure 1 This is a schematic diagram illustrating the implementation process of the present invention.

[0018] Beneficial effects

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] Significantly improved surface quality: Surface roughness uniformity is improved by more than 50%, and the roughness Ra value can be stably controlled below 0.8μm;

[0021] Improved processing efficiency: While ensuring quality, processing efficiency is increased by 15-20%, and rework rate is reduced by more than 90%.

[0022] Extended tool life: Through intelligent wear compensation and parameter optimization, the grinding head's lifespan is extended by 30%, reducing consumable costs;

[0023] Enhanced flexibility: It can automatically adapt to different materials (metals, composites, plastics) and complex curved surface processing, reducing changeover and debugging time from 4 hours to 30 minutes. Detailed Implementation

[0024] Example 1: Real-time Surface Quality Assessment Algorithm Based on Multispectral Vision

[0025] To address the issues of lag and low accuracy in traditional surface quality detection, a surface quality assessment model integrating multispectral imaging and deep learning is proposed:

[0026] Algorithm architecture: Includes a multispectral image acquisition module (400-1000nm band), a feature extraction network (using an improved U-Net architecture to extract surface texture features), and a quality evaluation layer (outputting quantitative indicators such as roughness Ra and gloss Gs);

[0027] Key innovations:

[0028] A multispectral feature fusion mechanism is proposed, where images of different spectral bands correspond to different surface features (e.g., the 450nm band is sensitive to fine scratches, and the 800nm ​​band is sensitive to surface roughness).

[0029] A lightweight network structure was designed to reduce the processing time to less than 50ms while ensuring evaluation accuracy, thus meeting real-time requirements.

[0030] Performance indicators: Roughness assessment error ≤5%, gloss assessment error ≤3%, capable of identifying surface defects larger than 0.5μm, achieving fully online assessment.

[0031] Innovation Point 2: Adaptive Control Algorithm for Dynamic Matching of Processing Force and Material Removal Amount

[0032] To address the over- or under-grinding problem caused by the mismatch between processing force and material removal rate, an adaptive force control algorithm based on material properties is proposed:

[0033] Algorithm principle: Establish a dynamic mapping model of "material hardness-processing force-removal amount" and adjust the processing force parameter by real-time monitoring of the material removal rate;

[0034] Key innovations:

[0035] A variable gain adaptive PID control strategy is proposed, which dynamically adjusts the proportional coefficient (Kp), integral coefficient (Ki), and derivative coefficient (Kd) according to the material hardness and surface curvature.

[0036] A predictive control mechanism is introduced to predict the required processing force for the next moment based on the material removal trend in the previous 50ms, so as to achieve advance adjustment.

[0037] Performance indicators: processing force control accuracy ±0.2N, material removal amount control error ≤8%, surface flatness improved by more than 40%.

[0038] Innovation Point 3: Trajectory Compensation Optimization Algorithm Considering Tool Wear

[0039] To address the problem of decreased machining quality due to tool wear, a dynamic trajectory compensation algorithm integrating a tool wear model is proposed.

[0040] Algorithm improvements:

[0041] A grinding head wear estimation model was established, and the wear amount (Δd) was calculated in real time based on processing time, contact force and acoustic signal characteristics.

[0042] Based on traditional trajectory planning, wear compensation is added to dynamically adjust the distance between the tool and the workpiece (Δz=f(Δd, material hardness));

[0043] A segmented trajectory optimization strategy was designed, using denser trajectory points in wear-sensitive areas to ensure processing uniformity.

[0044] Key innovation: Achieves adaptive trajectory compensation throughout the entire tool lifecycle, maintaining stable processing quality without frequent tool replacements;

[0045] Performance indicators: Wear estimation error ≤0.02mm, surface quality consistency improved by 50% after compensation, and tool life extended by 30%.

[0046] Innovation Point 4: Processing Parameter Decision Algorithm for Multi-Objective Collaborative Optimization

[0047] To address the challenge of coordinating the optimization of processing efficiency, surface quality, and tool life, a multi-objective decision-making algorithm based on improved particle swarm optimization is proposed.

[0048] Optimization model:

[0049] Objective functions: Maximize machining efficiency (min T), minimize surface roughness (min Ra), and maximize tool life (max L);

[0050] Decision variables: machining speed (v), machining force (F), tool speed (n), trajectory spacing (s);

[0051] Constraints: Surface quality requirements (Ra≤Ra0), equipment capacity limitations (v≤vmax);

[0052] Algorithm Innovation:

[0053] An adaptive weighting factor is introduced to dynamically adjust the priority of each objective according to different processing stages;

[0054] Design hybrid mutation operators to improve the algorithm's ability to escape local optima;

[0055] Performance metrics: Multi-objective optimization convergence speed improved by 25%, processing efficiency improved by 15-20% while ensuring surface quality, and tool life extended by 20%.

[0056] Example 2: Application of a real-time surface quality assessment algorithm based on multispectral vision

[0057] To address the issues of lagging and low accuracy in traditional surface quality inspection, which makes it difficult to meet the needs of real-time processing control, this embodiment innovatively applies a "real-time surface quality assessment algorithm based on multispectral vision." By fusing multi-band spectral imaging with deep learning, it achieves online and accurate assessment of surface roughness, gloss, and minor defects, providing timely basis for subsequent parameter adjustments.

[0058] (I) Multispectral Image Acquisition and Preprocessing

[0059] The algorithm constructs a multispectral image acquisition system covering the 400-1000nm band: it synchronously acquires image data of the processed surface in different bands through a multi-channel spectral camera. Each band image corresponds to different surface feature information. The short-wave band (such as around 450nm) is sensitive to local defects such as fine scratches and stains on the surface and can clearly capture small morphological changes. The long-wave band (such as around 800nm) is more likely to reflect the overall surface roughness distribution and reflects the processing uniformity through the difference in texture features.

[0060] To address potential issues such as noise and uneven illumination in multispectral images, targeted preprocessing is employed: an adaptive median filtering algorithm is used to suppress random noise while preserving defect edge details; an illumination compensation algorithm is used to eliminate illumination differences during imaging at different spectral bands, ensuring the consistency of grayscale features for the same surface region across different spectral bands; and geometric correction is performed on the images to correct feature shifts caused by camera distortion, laying an accurate foundation for subsequent feature extraction.

[0061] (II) Improved U-Net Feature Extraction and Quality Assessment

[0062] An improved U-Net architecture is adopted as the feature extraction network: Based on the traditional U-Net encoder-decoder structure, an attention gating module is introduced, which enables the network to automatically focus on surface defect areas and areas with significant roughness changes during the feature extraction process, thereby enhancing the representation ability of key features; Multispectral image data of each band is used as the multi-channel input of the network, and the adaptive fusion of features of different bands is achieved through a cross-channel attention mechanism, making full use of the complementary information of each band.

[0063] To meet real-time evaluation requirements, the network was designed to be lightweight: depthwise separable convolutions were used to replace traditional convolutional layers, reducing the number of model parameters and computational cost while maintaining feature extraction accuracy; channel pruning was introduced to remove redundant feature channels in the network, further compressing the model size. The optimized network can complete feature extraction and processing of a single frame image in a very short time, meeting the real-time requirements of the processing.

[0064] A quality assessment layer is constructed at the network output: the high-dimensional feature vector output by the feature extraction network is input into the fully connected layer, and quantitative assessment indicators such as surface roughness and gloss are output through the regression task; simultaneously, the semantic segmentation branch is used to locate and identify surface defects, distinguishing different types of defects such as scratches, dents, and stains. The assessment results are fed back to the system control unit in real time, forming a closed loop for quality detection.

[0065] (III) Qualitative Explanation of Algorithm Enhancement

[0066] Multispectral visual surface quality assessment algorithms transform traditional offline inspection into online real-time evaluation, breaking the inefficient "processing-inspection-rework" model. This enables timely detection of surface quality issues during processing, preventing the generation of batches of defective products. The multi-band feature fusion mechanism enhances the comprehensiveness and accuracy of quality assessment, capturing both subtle local defects and reflecting overall surface characteristics. Lightweight network design ensures real-time assessment, saving valuable time for subsequent process parameter adjustments.

[0067] Efficiency Enhancement Principle: Real-time assessment reduces the time cost of post-weld inspection and rework, significantly improving processing efficiency. Precise quality assessment provides a reliable basis for process parameter optimization, making parameter adjustments more targeted and greatly improving surface quality consistency. Early identification of minute defects reduces scrap rates, minimizes material and energy waste, and simultaneously improves the final product quality level, enhancing product market competitiveness.

[0068] Example 2: Application of Adaptive Control Algorithm for Dynamic Matching of Processing Force and Material Removal Amount

[0069] To address the issues of over-grinding and under-grinding caused by the mismatch between processing force and material removal in traditional machining processes, this embodiment innovatively applies an "adaptive control algorithm for dynamic matching of processing force and material removal." By establishing a dynamic mapping relationship between material properties and processing parameters, it achieves real-time and precise adjustment of processing force, ensuring that the material removal is stable within the target range and improving the quality of the processed surface.

[0070] (I) Material property sensing and removal amount monitoring

[0071] The algorithm first achieves real-time perception of material properties: by integrating a hardness sensor and a material recognition module, it quickly detects key mechanical parameters such as the hardness and elastic modulus of the workpiece material before processing; during processing, it combines the real-time processing force signal collected by the force sensor and the tool displacement signal collected by the position sensor to invert the dynamic changes of the material properties, providing basic data for subsequent force control strategies.

[0072] A real-time material removal monitoring mechanism is established: This mechanism combines visual measurement with weight change detection to calculate the material removal rate per unit time in real time. The vision system calculates the volumetric removal by capturing the three-dimensional topographic changes of the processing area; the weight sensor acquires the mass removal through high-precision weighing. The two systems mutually verify each other to ensure the accuracy of the removal monitoring. The monitored removal amount is compared with the target removal amount to obtain the deviation value, which serves as the input signal for adjusting the processing force.

[0073] (II) Variable Gain Adaptive PID Control and Predictive Regulation

[0074] A variable-gain adaptive PID control strategy is designed to adjust the machining force: material hardness, surface curvature, and removal deviation are used as the basis for gain adjustment, dynamically optimizing the proportional coefficient (Kp), integral coefficient (Ki), and derivative coefficient (Kd) of the PID controller. For example, when machining high-hardness materials, the proportional coefficient is increased to improve the system response speed; when machining curved workpieces with large surface curvature variations, the derivative coefficient is adjusted to suppress machining force fluctuations; when the removal deviation is small, the integral coefficient is decreased to avoid overshoot.

[0075] A predictive control mechanism is introduced to enhance the forward-looking nature of adjustments: Based on material removal trend data over a previous period, a removal volume prediction model is constructed, and a time-series analysis algorithm is used to predict the change in removal volume at the next moment. Processing force parameters are adjusted in advance based on the prediction results to avoid increased deviations in removal volume due to system lag. The prediction model is dynamically updated in conjunction with material properties to ensure high prediction accuracy at different processing stages.

[0076] To prevent damage to the workpiece surface from sudden changes in machining force, a smooth transition mechanism for machining force is designed: during parameter adjustment, exponential functions or polynomial interpolation are used to achieve gradual changes in machining force, avoiding step changes and ensuring a smooth machining process. Simultaneously, upper and lower limits for machining force are set to prevent excessive machining force due to sensor malfunctions or algorithm anomalies, thus protecting the workpiece and machining tools.

[0077] (III) Qualitative Explanation of Algorithm Enhancement

[0078] The dynamic matching algorithm for processing force and material removal effectively solves the problems of over-grinding and under-grinding, significantly improving the smoothness of the processed surface. The variable gain adaptive PID control strategy enables the system to adapt to changes in different material properties and processing conditions, greatly improving the accuracy of processing force control and avoiding surface quality defects caused by improper processing force. Predictive control mechanism and smooth transition mechanism further enhance the stability of the processing process and reduce surface quality fluctuations.

[0079] Efficiency Enhancement Principle: Stable material removal control reduces subsequent grinding and correction processes, significantly improving processing efficiency. Improved surface smoothness reduces scrap rate and increases material utilization. Precise control of processing force avoids excessive tool wear, extends tool life, and reduces equipment maintenance costs. Simultaneously, improved processing stability reduces parameter debugging time and enhances the system's adaptability to various workpiece types.

[0080] Example 3: Application of trajectory compensation optimization algorithm considering tool wear

[0081] To address the issues of decreased machining quality and frequent tool replacements due to tool wear, this embodiment innovatively applies a "trajectory compensation optimization algorithm that considers tool wear." By estimating tool wear in real time and dynamically adjusting the machining trajectory, it achieves stable machining quality throughout the tool's entire lifecycle, reduces tool replacement frequency, and lowers production costs.

[0082] (I) Wear and tear estimation based on multi-source information fusion

[0083] A multi-source information fusion model for estimating grinding head wear is constructed: Wear is calculated by comprehensively utilizing processing time, contact force signals, and acoustic signal characteristics. Processing time is the fundamental factor for wear accumulation; a timer records the tool's working time to preliminarily estimate the wear trend. A contact force sensor collects the real-time contact force between the tool and the workpiece; characteristic fluctuations in contact force occur as wear intensifies. An acoustic sensor collects noise signals during the processing; wear causes changes in the contact state between the tool and the workpiece, thus altering the noise spectrum characteristics.

[0084] Features from various sources are extracted and fused: peak value, root mean square (RMS), and fluctuation frequency are extracted from the contact force signal; spectral centroid and peak energy are extracted from the acoustic signal. These features are combined with processing time parameters and input into a trained wear estimation model (such as a gradient boosting tree model) to output real-time wear (Δd). The model is continuously calibrated using historical wear data and actual measurement data to ensure estimation accuracy.

[0085] (II) Dynamic Trajectory Compensation and Segmented Optimization

[0086] This invention introduces wear compensation based on traditional trajectory planning: The distance adjustment between the tool and workpiece, i.e., the trajectory compensation, is calculated using the estimated wear amount (Δd) and the hardness parameter of the current material being processed, through the functional relationship Δz = f(Δd, material hardness). This compensation is then superimposed in real-time onto the z-axis coordinate of the original trajectory plan, dynamically correcting the actual machining position of the tool and offsetting the insufficient machining depth caused by wear.

[0087] A segmented trajectory optimization strategy is designed to improve machining uniformity: Based on the geometric features of the workpiece surface and machining requirements, the machining area is divided into wear-sensitive and non-sensitive areas. For wear-sensitive areas (such as planes and curved transitions with high precision requirements), a denser distribution of trajectory points is used, and the distance between adjacent trajectory points is shortened to ensure that the compensated trajectory can precisely fit the workpiece surface; for non-sensitive areas, the trajectory spacing is appropriately increased while ensuring quality, balancing machining efficiency and quality.

[0088] Establish a tool wear threshold early warning mechanism: Based on the tool's service life and machining quality requirements, set an upper limit threshold for wear. When the estimated wear reaches the threshold, the system issues an early warning signal, prompting the operator to prepare to replace the tool, thus preventing severe degradation of machining quality or tool breakage that could damage the workpiece due to excessive wear. This early warning mechanism balances maximizing tool utilization with ensuring machining quality stability.

[0089] (III) Qualitative Explanation of Algorithm Enhancement

[0090] The trajectory compensation algorithm considering tool wear enables adaptive trajectory adjustment throughout the tool's entire lifecycle, maintaining stable machining quality without frequent tool changes and significantly reducing tool change frequency and downtime. The wear estimation model based on multi-source information fusion improves the accuracy of wear detection, providing a reliable basis for compensation strategies. The segmented trajectory optimization strategy ensures machining uniformity while also considering machining efficiency, avoiding resource waste caused by over-compensation.

[0091] Efficiency Enhancement Principles: Extended tool life reduces tool procurement and replacement costs, while reduced downtime significantly improves equipment utilization. Stable processing quality reduces scrap and rework caused by wear, improving both processing efficiency and product qualification rate. A wear warning mechanism allows for more planned tool replacement, avoiding production interruptions due to sudden malfunctions and enhancing the overall stability and continuity of the production line.

[0092] Example 4: Application of a multi-objective collaborative optimization algorithm for determining processing parameters

[0093] To address the challenge of coordinating and optimizing processing efficiency, surface quality, and tool life, this embodiment innovatively applies a "multi-objective collaborative optimization processing parameter decision algorithm." By improving the particle swarm optimization algorithm, it achieves a global optimal balance of multiple objectives while satisfying constraints, and outputs the optimal combination of processing parameters.

[0094] (I) Construction of Multi-Objective Optimization Model

[0095] Define the objective function, decision variables, and constraints for multi-objective optimization: The objective function is set to maximize machining efficiency, minimize surface roughness, and maximize tool life. These three objectives are coupled and conflict to some extent, requiring optimization to achieve a synergistic balance. Decision variables include key process parameters such as machining speed, machining force, tool rotation speed, and trajectory spacing, which directly affect the degree to which each objective is achieved. Constraints include surface quality requirements (such as upper roughness limit), equipment capacity limitations (such as maximum machining speed and maximum output force), and machining safety requirements to ensure the feasibility of the optimization results.

[0096] To quantify the objective functions, correlation models between parameters and objectives were established: a linear relationship model between machining speed and machining efficiency was constructed through experimental design and data fitting; a nonlinear mapping model between machining force, tool rotation speed, and surface roughness was established based on the output results of the surface quality assessment algorithm; and a decay model between machining parameters and tool life was established by combining tool wear theory and experimental data. These correlation models provide accurate calculation basis for the objective functions of the optimization algorithm.

[0097] (II) Improved Particle Swarm Optimization Algorithm Design

[0098] An adaptive weighting factor is introduced to optimize the particle update strategy: the weight coefficients of each objective are dynamically adjusted according to the priority of the objectives at different processing stages. In the early stage of processing, the main objective is to rapidly improve efficiency, so the weight of processing efficiency is increased; in the middle stage of processing, quality and efficiency need to be considered, and the weights of each objective are balanced; in the later stage of processing, the focus is on ensuring surface quality and extending tool life, so the weights of these two objectives are increased. The adaptive adjustment of the weighting factor allows the algorithm to flexibly focus on key objectives according to the processing progress.

[0099] A hybrid mutation operator is designed to enhance the global search capability of the algorithm: Gaussian mutation and differential mutation are combined to mutate some particles during the particle swarm iteration process. Gaussian mutation enables particles to perform a fine-grained local search around the current optimal solution, improving convergence accuracy; differential mutation generates new particles through the difference information between particles, helping the algorithm escape local optima and explore a wider parameter space. The hybrid mutation operator balances the algorithm's convergence speed and optimization accuracy.

[0100] An elite retention strategy is employed to maintain the Pareto optimal solution set: during algorithm iteration, non-dominated solutions obtained in each iteration are added to the elite solution set, and redundant solutions are eliminated by crowding sorting, maintaining the diversity and uniformity of the solution set. Finally, the most suitable combination of processing parameters is selected from the elite solution set based on actual production needs (such as prioritizing efficiency or quality) and output to the control system for execution.

[0101] (III) Qualitative Explanation of Algorithm Enhancement

[0102] The multi-objective collaborative optimization algorithm effectively resolves the contradictions between processing efficiency, surface quality, and tool life, achieving a globally optimal balance among the three. The introduction of adaptive weighting factors and hybrid mutation operators improves the algorithm's convergence speed and optimization capability, enabling it to quickly find the optimal parameter combination that satisfies the multi-objective requirements. The elite retention strategy provides diverse options for production decisions, making parameter settings more closely aligned with actual production needs.

[0103] Efficiency Enhancement Principle: While ensuring surface quality, processing efficiency is significantly improved, and the processing cycle per piece is shortened. Extended tool life reduces tool consumption costs and equipment maintenance frequency. The application of optimal parameter combinations reduces energy and material waste, further lowering production costs. Simultaneously, the algorithm's rapid optimization capability significantly shortens the debugging time for new workpieces, enhancing the production line's flexibility and enabling rapid response to market order changes.

[0104] In summary, through the synergistic application of the four innovative algorithms mentioned above, a complete intelligent monitoring and adaptive control system for machining parameters has been constructed, realizing intelligent management and control of the entire machining process from surface quality assessment, machining force control, trajectory compensation to parameter optimization. This system effectively solves technical bottlenecks in traditional machining, such as lagging quality inspection, mismatch between machining force and material removal, tool wear affecting quality, and difficulties in multi-objective optimization. It significantly improves machining quality consistency and production efficiency, reduces energy consumption and production costs, and enhances the flexibility and stability of production lines, providing a practical technical solution for the intelligent upgrading of the manufacturing industry.

[0105] (II) Fuzzy PID Adaptive Regulation and Segmented Control

[0106] A fuzzy PID adaptive control strategy is designed to achieve precise parameter control: The characteristic deviations of the molten pool (such as area deviation and temperature deviation) and their rates of change are used as inputs to the fuzzy controller. Through fuzzy rule reasoning, the proportional, integral, and derivative parameter adjustments of the PID controller are obtained, dynamically optimizing the PID control output and thus adjusting key process parameters such as welding current, voltage, and welding speed. This control method combines the nonlinear processing capabilities of fuzzy control with the steady-state accuracy advantages of PID control, enabling rapid response to changes in the molten pool state and avoiding overshoot or lag in parameter adjustments.

[0107] Considering the differences in molten pool characteristics at different stages of the welding process, a segmented control logic is introduced: For the arc ignition stage, a low-current, slow-speed arc ignition strategy is adopted, gradually increasing parameters to ensure a smooth molten pool formation; in the stabilization welding stage, PID adaptive control is implemented based on molten pool characteristic deviations to maintain a stable molten pool state; in the arc termination stage, a gradual parameter decay rule is designed to avoid defects such as arc craters and cracks caused by a sudden drop in heat. The control parameters and switching conditions for each stage are optimized and determined using historical welding data to ensure a smooth transition throughout the entire welding process.

[0108] (III) Qualitative Explanation of Algorithm Enhancement

[0109] The molten pool feature-driven adaptive control algorithm achieves real-time and precise control of the molten pool state, effectively avoiding common defects such as burn-through due to an excessively large molten pool and incomplete penetration due to an excessively small molten pool. The fuzzy PID control strategy makes parameter adjustment more flexible and rapid, adapting to the welding requirements of different plate thicknesses and materials, and reducing the workload of manual parameter tuning. The segmented control logic ensures the stability of each stage of welding, especially improving the quality of arc initiation and termination, further reducing the welding defect rate.

[0110] Efficiency Enhancement Principle: Precise control of the molten pool state significantly improves welding quality consistency and greatly increases the CPK value. Adaptive parameter adjustment reduces rework caused by improper parameters, resulting in a significant improvement in welding efficiency. Simultaneously, precise parameter control avoids waste of energy and consumables, reducing both welding energy consumption and material consumption, thus saving production costs for enterprises.

[0111] Example 3: Application of Real-time Welding Defect Identification and Classification Algorithm

[0112] To address the shortcomings of traditional offline welding defect detection methods that make it difficult to detect problems in a timely manner, this embodiment innovatively applies a "real-time welding defect identification and classification algorithm." Through multi-source feature fusion and a lightweight deep learning model, it enables rapid identification, classification, and severity assessment of common defects during the welding process, providing a basis for timely parameter adjustments.

[0113] (I) Feature extraction of multi-source defects

[0114] The algorithm constructs a multi-source defect feature extraction system integrating infrared, vision, and electric arc sensors: It extracts temperature anomaly features from infrared thermal imager data, such as localized low-temperature points corresponding to pores and abrupt temperature gradient changes in crack areas; it extracts molten pool morphology anomaly features from high-speed visual images, such as irregular molten pool boundaries caused by incomplete melting and a sudden increase in molten pool area due to burn-through; and it extracts fluctuation anomaly features from voltage and current signals acquired by electric arc sensors, such as current abrupt changes during crack formation and high-frequency voltage fluctuations corresponding to pores. Through the complementarity of these multi-source features, the algorithm comprehensively captures the characteristic manifestations of different types of defects.

[0115] The extracted multi-source features are fused: a feature-level fusion method is adopted to concatenate the feature vectors of each modality into a unified high-dimensional feature vector, and the dimensionality is reduced by principal component analysis (PCA) to retain key feature information while reducing the amount of computation, thus providing efficient feature input for the subsequent defect identification model.

[0116] (II) Identification and Severity Assessment of Lightweight Defects

[0117] Real-time defect identification is achieved using an improved YOLOv5 algorithm: the original model is lightweighted by replacing traditional convolutional layers with depthwise separable convolutions, reducing the number of model parameters and computational cost; an attention mechanism module is introduced to enhance the model's ability to capture features of defect regions; and the loss function is optimized to improve the identification accuracy of minute defects. The improved model significantly improves inference speed while maintaining recognition accuracy, meeting the real-time requirements of the welding process.

[0118] Establish a defect severity assessment system: Based on factors such as defect size (area, length), location (center area / edge area of ​​weld), and diffusion trend (whether it continues to expand), set weight coefficients for each factor, and obtain a defect severity index through weighted summation, classifying defects into levels 1-5. Levels 1-2 are minor defects, which can be suppressed by subsequent parameter adjustments; levels 3-4 are moderate defects, requiring immediate adjustment of welding parameters; level 5 is a severe defect, requiring welding to be suspended for manual intervention to prevent defect expansion.

[0119] (III) Qualitative Explanation of Algorithm Enhancement

[0120] The real-time welding defect identification and classification algorithm transforms defect detection from offline post-weld inspection to real-time monitoring during welding, enabling timely defect detection and early warning, and effectively preventing the generation of batch defects. The lightweight model design ensures real-time identification, significantly reducing detection latency and providing valuable time for parameter adjustment. Multi-source feature fusion improves the accuracy of defect identification, particularly enhancing the ability to identify minute defects, reducing missed and false detection rates.

[0121] Efficiency Enhancement Principle: Real-time defect detection reduces post-weld inspection and rework time by over 80%, significantly improving production efficiency. The defect severity assessment system provides operators and the system with clear handling guidelines, avoiding over- or under-intervention, further enhancing the stability of the welding process. Reduced batch defects lower material waste and production costs while improving the overall product quality.

[0122] Example 4: Application of Intelligent Decision-Making Algorithm with Multi-Parameter Coupling Optimization

[0123] To address the problem of strong coupling of multiple welding parameters and the difficulty in achieving global optimization through manual adjustment, this embodiment innovatively applies an "intelligent decision-making algorithm for multi-parameter coupling optimization." By improving the deep reinforcement learning model and integrating expert knowledge, it achieves global optimization decision-making for welding parameters, balancing welding quality, efficiency, and energy consumption goals.

[0124] (I) Construction of Reinforcement Learning Models

[0125] The algorithm constructs a parameter optimization model based on deep reinforcement learning: the state space of the model is defined as a combination of workpiece features (material, plate thickness, joint type) and real-time process features (molten pool state, arc features, defect detection results); the action space is the adjustment range of key parameters such as welding current, voltage, welding speed, and wire feed speed; the reward function is designed as a multi-objective comprehensive evaluation function, which comprehensively considers three dimensions: welding quality (defect severity, weld formation quality), processing efficiency (welding speed, cycle time), and energy consumption (welding power consumption). The weight coefficients of each dimension are dynamically adjusted according to production needs to guide the model to optimize towards the target direction.

[0126] An improved Deep Deterministic Policy Gradient (DDPG) algorithm is adopted as the model training framework: an experience replay mechanism is introduced to store the experience data of the agent's interaction with the environment, and random sampling is used for training to improve sample utilization; a dual network structure of target network and evaluation network is designed to reduce parameter fluctuations during training and improve model stability; gradient pruning technique is used to prevent gradient explosion and ensure model convergence.

[0127] (II) Expert Knowledge Integration and Adaptive Exploration Strategies

[0128] To accelerate model convergence and improve the quality of initial decisions, a knowledge distillation mechanism is introduced to integrate expert experience: expert knowledge such as the parameter adjustment experience of senior welders and the recommended parameter ranges in process manuals are transformed into a rule base. In the early stage of model training, the actions of the agent are constrained, guiding the model to explore within a reasonable parameter space. During training, the knowledge distillation loss function is used to make the model output closer to the expert decision, internalizing the expert experience into the model's decision-making ability, and significantly shortening the model's training cycle.

[0129] The design employs an adaptive exploration strategy to balance exploration and utilization: In the early stages of model training, a higher exploration rate is used to encourage the agent to try different parameter combinations and explore a wider parameter space. As training progresses and the model performance reaches a certain threshold, the exploration rate is gradually reduced to increase the utilization of known optimal parameter combinations. Simultaneously, when welding conditions change significantly, the exploration rate is automatically increased to ensure the model can adapt to the new conditions and find the optimal parameter combination.

[0130] Efficiency Enhancement Principle: The multi-parameter coupled optimization algorithm achieves globally optimal decision-making for welding parameters, effectively solving the problem of difficult multi-variable coupled control. Expert knowledge fusion enables the model to converge quickly and possess good initial performance, while the adaptive exploration strategy ensures the model's adaptability and optimization capability under different working conditions. Guided by a multi-objective reward function, the algorithm achieves synergistic optimization of efficiency and energy consumption while ensuring welding quality. This significantly improves welding quality consistency, reduces energy and material consumption, greatly shortens the debugging time for new workpieces, and reduces operational difficulty, providing strong support for enterprises to achieve flexible and efficient production.

Claims

1. A method for monitoring and adaptively controlling processing parameters during the grinding and polishing process, characterized in that, include: Collect images of the workpiece surface, processing contact force, tool status, acoustic signals, and robot motion parameters during the grinding process; A real-time surface quality assessment algorithm based on multispectral vision is used to achieve online detection of surface roughness and gloss. An adaptive control algorithm based on dynamic matching of processing force and material removal amount is used to achieve precise control of processing force; A trajectory compensation optimization algorithm that takes tool wear into account is used to dynamically adjust the relative position of the tool and the workpiece; A multi-objective collaborative optimization algorithm for machining parameters is used to achieve collaborative optimization of machining efficiency, surface quality, and tool life.

2. The method according to claim 1, characterized in that, The real-time surface quality assessment algorithm based on multispectral vision adopts an improved U-Net architecture, integrates image features in the 400-1000nm band, has a roughness assessment error of ≤5%, and a processing time of ≤50ms.

3. The method according to claim 1, characterized in that, The adaptive control algorithm for dynamic matching of processing force and material removal amount adopts a variable gain adaptive PID control strategy, with a processing force control accuracy of ±0.2N and a material removal amount control error of ≤8%.

4. The method according to claim 1, characterized in that, The trajectory compensation optimization algorithm that takes tool wear into account calculates the wear amount based on processing time, contact force and acoustic signal characteristics. The wear amount estimation error is ≤0.02mm, and the surface quality consistency is improved by 50% after compensation.

5. The method according to claim 1, characterized in that, The multi-objective collaborative optimization processing parameter decision algorithm adopts an improved particle swarm optimization algorithm to achieve multi-objective optimization of processing efficiency, surface quality and tool life, thereby improving processing efficiency by 15-20%.

Images