Laser trimming path real-time compensation method and system fused with online monitoring

Abstract

The invention relates to the technical field of precise grinding of aero-engine blades, and discloses a laser trimming path real-time compensation method and system.The laser trimming path real-time compensation method fusing online monitoring comprises the steps that monitoring data of a multi-source heterogeneous sensor for grinding wheel laser trimming are collected; planning a laser trimming path; carrying out data preprocessing and space-time alignment; carrying out multi-modal feature extraction and feature selection; multi-modal feature fusion is conducted, and the current machining state of grinding wheel laser finishing is evaluated; the machining state of grinding wheel laser finishing is predicted; performing multi-target compensation optimization on the basic path planning result; performing path correction and smoothing processing; executing the compensation path instruction, monitoring the execution process in real time, evaluating the execution effect, and performing self-adaptive feedback; by continuously monitoring the execution effect and conducting self-adaptive adjustment, intelligent control over the finishing process is achieved.

Description

Technical Field

[0001] This invention relates to the field of precision grinding technology for aero-engine blades, and more specifically, to a method and system for real-time compensation of laser dressing paths that integrates online monitoring. Background Technology

[0002] As a core component of the engine, the geometric accuracy and surface quality of aero-engine blades directly affect the engine's performance, efficiency, and safety. In the precision grinding process of blades, the shape accuracy and surface condition of the grinding wheel are key factors determining the machining quality. With the aerospace industry's ever-increasing demands for blade machining precision, traditional grinding wheel dressing methods are no longer sufficient to meet the requirements of high-precision and high-efficiency machining. Laser dressing technology, as an advanced grinding wheel dressing method, has advantages such as non-contact operation, high precision, and strong controllability, and has received widespread attention and application in the field of aero-engine blade grinding.

[0003] However, existing laser dressing technologies face numerous challenges in practical applications. The inhomogeneity of the grinding wheel material leads to differences in laser absorption rate and removal threshold at different locations, making it difficult to achieve uniform dressing using uniform process parameters. Dynamic disturbances during processing, such as machine tool vibration, thermal deformation, and fluctuations in material properties, can affect the stability of dressing accuracy. The heat accumulation effect during laser dressing may cause localized overheating of the grinding wheel, leading to a decrease in bond strength, changes in material structure, and even thermal cracks, severely impacting the performance and lifespan of the grinding wheel.

[0004] Most current laser dressing systems employ offline path planning and open-loop control, lacking the ability to monitor and dynamically adjust the dressing process in real time. This approach cannot respond promptly to changes in the processing status, making it difficult to guarantee the consistency and stability of dressing quality. The limitations of existing technology are particularly pronounced when dealing with complex-shaped grinding wheels or high-precision dressing requirements, urgently necessitating the development of an intelligent laser dressing method capable of real-time sensing of the processing status and dynamic adjustment of the dressing strategy. Summary of the Invention

[0005] This invention provides a real-time compensation method and system for laser dressing path that integrates online monitoring, solving the technical problem in related technologies where it is difficult to achieve stable and high-precision dressing when faced with multiple factors such as unevenness of grinding wheel material, dynamic disturbances in the processing process, and heat accumulation effects.

[0006] This invention provides a real-time compensation method for laser trimming paths that integrates online monitoring, comprising the following steps:

[0007] Data from multi-source heterogeneous sensors were collected for laser dressing of grinding wheels. The laser dressing path was planned to obtain the basic path planning results. Data preprocessing and spatiotemporal alignment were performed to obtain standard multi-source monitoring data.

[0008] Based on the standard multi-source monitoring dataset, multimodal feature extraction and feature selection are performed to obtain a multimodal key feature set;

[0009] Based on the multimodal key feature set, multimodal feature fusion is performed, and the current processing status of the grinding wheel laser dressing is evaluated to obtain a comprehensive evaluation result of the current processing status;

[0010] Based on the comprehensive evaluation results of the current processing status, the processing status of laser dressing of grinding wheels is predicted, and the comprehensive prediction results are output.

[0011] Based on the comprehensive evaluation and prediction results of the current processing status, the basic path planning results are optimized by multi-objective compensation to obtain the comprehensive path compensation amount.

[0012] Based on the comprehensive path compensation amount and the basic path planning results, path correction and smoothing are performed to obtain the compensation path instructions.

[0013] The system executes compensation path instructions, monitors the execution process in real time, evaluates the execution effect, provides adaptive feedback, and obtains complete data records and trimming reports of the processing process.

[0014] In a preferred embodiment, the data collected by the multi-source heterogeneous sensor for laser dressing of the grinding wheel includes:

[0015] Configure a multi-sensor monitoring system, including a laser displacement sensor, a visual imaging system, an acoustic emission sensor, and an infrared temperature sensor;

[0016] For the laser displacement sensor, a calibration curve between the distance measurement value and the actual distance is established using standard gauge blocks. For the visual imaging system, a calibration board is used to calibrate the camera intrinsic parameters, and a rotation matrix and translation vector between the camera coordinate system and the machining coordinate system are established through hand-eye calibration. For the acoustic emission sensor, its frequency response characteristics and sensitivity parameters are tested using a known sound source. For the infrared temperature sensor, temperature calibration is performed based on the emissivity characteristics of the material being tested. Thus, the calibration parameters of the multi-sensor monitoring system are obtained.

[0017] Multi-source heterogeneous sensor monitoring data are collected synchronously according to their respective sampling frequencies and timestamped during data collection, outputting multi-source raw monitoring data.

[0018] In a preferred embodiment, the data preprocessing and spatiotemporal alignment include:

[0019] The monitoring data from multi-source heterogeneous sensors are filtered, denoised, and preliminarily feature-extracted to output preprocessed monitoring data.

[0020] Based on the calibration parameters of the multi-sensor monitoring system and the preprocessed monitoring data, the spatiotemporal alignment of the multi-source monitoring data is performed, a unified time grid is established, the timestamps of each sensor are compensated for time delay, the spatial positions measured by different sensors are mapped to a unified processing coordinate system, and a standard multi-source monitoring dataset is generated.

[0021] In a preferred embodiment, the multimodal feature extraction includes:

[0022] Geometric deviation features are extracted from topographic point cloud data. The point cloud data is registered with the target trimming surface. The normal distance from each measurement point in the point cloud to the target surface is calculated to generate the shape deviation field distribution. The global features of the deviation are extracted, and the spatial distribution features of the deviation are extracted.

[0023] Surface quality features are extracted from visual image data, texture feature parameters are extracted using the gray-level co-occurrence matrix method, the contour boundary of the grinding wheel is extracted by edge detection to assess boundary integrity, and burn areas are identified by color space analysis.

[0024] Material removal process features are extracted from acoustic emission signals, the main frequency component and its time domain variation are extracted, the cumulative energy and energy change rate of the signal within each time window are calculated, and acoustic emission events are identified and event features are statistically analyzed by setting amplitude thresholds.

[0025] Extract heat accumulation effect features from temperature data, extract the standard deviation of spatial statistical features, extract the temporal evolution features of temperature, and extract the features of the heat-affected zone.

[0026] In a preferred embodiment, the multimodal feature fusion includes:

[0027] Establish a feature fusion rule base based on expert knowledge;

[0028] A lightweight model is used for feature fusion and state assessment. The state assessment model adopts gradient boosting decision tree or support vector machine to output shape deviation score, uniformity removal score, thermal accumulation level, processing stability score, and surface quality level.

[0029] In a preferred embodiment, predicting the processing state of the laser dressing of the grinding wheel includes:

[0030] A simplified physical model for material removal is established, and an energy threshold criterion is adopted. When the energy density received per unit area of ​​the laser-acted region exceeds the material removal threshold, the material is removed. The removal depth is proportional to the energy density, and the amount of material removed is predicted.

[0031] A simplified physical model of heat conduction is established. The surface of the grinding wheel is divided into multiple lumped elements using a lumped parameter model. The heat balance equations of each element are established, and the temperature field evolution is predicted by numerically solving the heat balance equations of each element.

[0032] A simplified physical model of thermal deformation is established, the thermal strain at each location is calculated based on the temperature field distribution, the stress-strain relationship is established based on the material's elastic modulus and Poisson's ratio to solve for the thermal stress distribution, and the thermal stress and thermal deformation of the grinding wheel are predicted.

[0033] Using data-driven models to predict state evolution trends;

[0034] By integrating the prediction results from the physical model and the data-driven model, a comprehensive prediction result is obtained.

[0035] In a preferred embodiment, the multi-objective compensation optimization of the basic path planning results includes:

[0036] A multi-objective compensation optimization problem is established, with optimization objectives including minimizing shape deviation, maximizing removal uniformity, controlling temperature level, and maintaining execution smoothness. Compensation variables include path point position adjustment, scanning speed adjustment, and laser power adjustment.

[0037] A hierarchical iterative strategy is employed to solve the collaborative optimization problem;

[0038] Conduct a feasibility check and implement constraint measures for the compensation plan;

[0039] Calculate the priority and execution strategy for compensation.

[0040] In a preferred embodiment, the path correction and smoothing process includes:

[0041] Generate a sequence of compensated path point coordinates. For each path point in the path, add its position adjustment amount to the original coordinates to obtain the new compensated coordinates. The position adjustment is performed along the normal direction of the grinding wheel surface.

[0042] The compensation path is smoothed, and the B-spline fitting method is used to fit the compensated path points as control points or data points to generate a B-spline curve.

[0043] Perform dynamic constraint checks and optimizations on the path, and calculate the velocity, acceleration, and jerk of each axis when executing the path;

[0044] The trimming path is segmented by using segmented trajectory generation and look-ahead buffering.

[0045] In a preferred embodiment, the adaptive feedback includes:

[0046] Continuously collect monitoring data and evaluate the execution effect, compare the current actual state with the expected state, and calculate the state tracking error;

[0047] An adaptive feedback control adjustment and compensation strategy is implemented, the feedback control gain is set, and incremental adjustments are made based on the state tracking error. The adjustment methods include incremental adjustment of the compensation amount, fine-tuning of process parameters, and updating of prediction model parameters.

[0048] Achieve closed-loop data flow and system self-learning, forming a complete closed loop from monitoring data acquisition to execution control;

[0049] Establish an anomaly detection and graded anomaly handling mechanism. When an anomaly is detected, the graded anomaly handling mechanism is activated, including: Level 1 anomalies are handled by parameter fine-tuning and enhanced monitoring; Level 2 anomalies are handled by immediately reducing the laser power to a safe level; and Level 3 anomalies are handled by immediately suspending processing and activating the safety protection program.

[0050] In a preferred embodiment, a real-time compensation system for laser trimming paths integrating online monitoring is used to perform the above-described real-time compensation method for laser trimming paths integrating online monitoring, comprising:

[0051] The multi-sensor monitoring module collects monitoring data from multiple heterogeneous sensors during laser dressing of grinding wheels, plans the laser dressing path, and obtains the basic path planning results; it also performs data preprocessing and spatiotemporal alignment to obtain standard multi-source monitoring data.

[0052] The feature extraction module, based on the standard multi-source monitoring dataset, performs multimodal feature extraction and feature selection to obtain a multimodal key feature set;

[0053] The status assessment module, based on the multimodal key feature set, performs multimodal feature fusion and assesses the current processing status of the grinding wheel laser dressing, obtaining a comprehensive assessment result of the current processing status;

[0054] The status prediction module predicts the processing status of the grinding wheel laser dressing based on the comprehensive evaluation results of the current processing status and outputs the comprehensive prediction results.

[0055] The compensation optimization module performs multi-objective compensation optimization on the basic path planning results based on the comprehensive evaluation results and comprehensive prediction results of the current processing status, and obtains the comprehensive path compensation amount.

[0056] The path generation module performs path correction and smoothing based on the comprehensive path compensation amount and the basic path planning results to obtain the compensated path instructions.

[0057] The execution control module executes compensation path instructions, monitors the execution process in real time and evaluates the execution effect, provides adaptive feedback, and obtains complete data records and trimming reports of the processing process.

[0058] The beneficial effects of this invention are as follows:

[0059] This invention constructs a multi-sensor collaborative monitoring system to acquire multi-dimensional information such as grinding wheel surface morphology, material removal status, and thermal accumulation effects in real time. It establishes a lightweight feature fusion mechanism and a hybrid prediction strategy, achieving comprehensive perception and accurate prediction of the dressing process. Employing a coupled sensing collaborative optimization algorithm to calculate compensation amounts, it can simultaneously optimize multiple objectives such as shape accuracy, removal uniformity, temperature control, and execution smoothness, effectively solving the problem of conflicting optimization objectives in traditional methods. Through adaptive real-time path compensation, the system can dynamically respond to grinding wheel material inhomogeneity and processing disturbances, improving the stability and consistency of dressing accuracy.

[0060] The closed-loop feedback control mechanism established in this invention achieves intelligent control of the dressing process by continuously monitoring the execution effect and making adaptive adjustments. The system possesses self-learning capabilities, enabling it to mine processing patterns from historical data and continuously optimize the state assessment model and compensation strategy, thus improving its adaptability to new grinding wheel types and process parameters. Through a cross-process feedback mechanism, the actual performance of the grinding wheel in subsequent grinding applications is fed back to the dressing system, extending the dressing quality evaluation from geometric accuracy to service performance. This ensures that the dressed grinding wheel meets the actual requirements of high-precision grinding, providing reliable technical support for the precision manufacturing of aero-engine blades. Attached Figure Description

[0061] Figure 1 This is a flowchart of the main process of a real-time compensation method for laser trimming path that integrates online monitoring in this invention;

[0062] Figure 2 This is a detailed flowchart of a real-time compensation method for laser trimming path that integrates online monitoring, as described in this invention.

[0063] Figure 3 This is a block diagram of a real-time compensation system for laser trimming path that integrates online monitoring, as described in this invention. Detailed Implementation

[0064] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0065] At least one embodiment of the present invention discloses a real-time compensation method for laser trimming paths that integrates online monitoring, such as... Figures 1 to 2 As shown, it includes the following steps:

[0066] Step 1: Collect monitoring data from multi-source heterogeneous sensors for laser dressing of grinding wheels, plan the laser dressing path, and obtain the basic path planning results; perform data preprocessing and spatiotemporal alignment to obtain standard multi-source monitoring data;

[0067] Step 1.1: Configure the multi-sensor monitoring system and complete the calibration;

[0068] Based on the geometric characteristics of the grinding wheel to be dressed, the structural layout of the laser dressing equipment, and monitoring requirements, a multi-sensor monitoring system is configured. This system includes a laser displacement sensor, a visual imaging system, an acoustic emission sensor, and an infrared temperature sensor. The laser displacement sensor is installed at a suitable location within the dressing work area, with its measuring beam maintaining a predetermined spatial distance from the laser dressing beam to avoid mutual interference. It is used for non-contact measurement of the distance from the grinding wheel surface to the sensor, collecting surface topography data. The visual imaging system includes an industrial camera and an illumination device. The camera is equipped with a macro lens, covering the laser dressing area and its surroundings, acquiring image information of the grinding wheel surface during the dressing process. The illumination device uses an adjustable-intensity LED light source to ensure clear images at different processing stages. The acoustic emission sensor is fixed to the grinding wheel spindle or support structure via coupling agent or mechanical clamping, acquiring acoustic signals generated during material removal in real time. The infrared temperature sensor monitors the temperature distribution in and around the laser-affected area in a non-contact manner, with its response wavelength range covering the main radiation bands of the measured material at the processing temperature.

[0069] Before the repair task was carried out, the multi-sensor monitoring system was calibrated. For the laser displacement sensor, a calibration curve between the distance measurement value and the actual distance was established using standard gauge blocks to obtain the measurement gain and zero-point deviation parameters. For the vision imaging system, a calibration board was used to calibrate the camera's intrinsic parameters to obtain parameters such as focal length, principal point position, and distortion coefficient. Hand-eye calibration was performed using the calibration board or known feature points to establish the rotation matrix and translation vector between the camera coordinate system and the machining coordinate system. For the acoustic emission sensor, its frequency response characteristics and sensitivity parameters were tested using a known sound source. For the infrared temperature sensor, temperature calibration was performed based on the emissivity characteristics of the measured material to establish the correspondence between the radiation signal and the actual temperature.

[0070] Simultaneously, the data acquisition delay of each sensor is recorded to obtain the calibration parameters of the multi-sensor monitoring system, including sensor response time, signal conditioning time, and data transmission time, providing parameter basis for subsequent time synchronization.

[0071] Step 1.2: Perform laser trimming and synchronize monitoring data from multi-source heterogeneous sensors;

[0072] Based on the initial geometric model of the grinding wheel and the target dressing profile, an offline laser dressing path is planned. The path planning employs a geometric approximation method, discretizing the target dressing profile into a sequence of path points. Process parameters for each path point, including laser power, scanning speed, and defocusing amount, are determined according to the grinding wheel material properties and laser performance. A path optimization algorithm sorts and connects the path points to generate a continuous dressing trajectory, ensuring both geometric accuracy and execution efficiency. The basic path planning results are obtained, including the path point coordinate sequence and process parameters such as laser power, scanning speed, and defocusing amount for each path point. The planned path is then sent to the motion control system to initiate the laser dressing task. A multi-axis motion mechanism drives the laser head to move along the planned path. The laser outputs a laser beam with a set power. After being focused by the optical system, the laser beam acts on the rotating grinding wheel surface. Through the interaction between the laser energy and the grinding wheel material, material is removed, achieving the dressing of the grinding wheel surface.

[0073] During the finishing process, based on the calibration parameters of the multi-sensor monitoring system output in step 1.1, each sensor synchronously collects monitoring data according to its own sampling frequency. The laser displacement sensor continuously measures the distance from the grinding wheel surface to the sensor at a set frequency. Considering that the grinding wheel is rotating, it is necessary to combine the grinding wheel speed and the measurement time to map the measured value to the spatial position of the grinding wheel surface, forming surface topography point cloud data. The visual imaging system acquires image sequences at a set frame rate. Each frame corresponds to a collection time and camera pose. The image data includes texture information, boundary shape, color features, etc. of the grinding wheel surface. The acoustic emission sensor continuously records acoustic signals at a high sampling rate. The signal amplitude and spectral characteristics reflect the material removal mechanism, removal rate, and processing stability. The infrared temperature sensor periodically collects temperature distribution data to obtain temperature field information of the laser-affected area and the heat-affected zone.

[0074] Each sensor timestamps its data during acquisition, based on the system's unified clock, providing a reference for subsequent time synchronization. The acquired raw data is transmitted to edge computing nodes or the central processing unit via data acquisition cards or network interfaces, entering the preprocessing stage and outputting multi-source raw monitoring data.

[0075] Step 1.3: Filter and denoise the raw monitoring data and perform preliminary feature extraction;

[0076] Based on the multi-source raw monitoring data output in step 1.2, filtering and denoising processes are performed to eliminate measurement noise and environmental interference. For the distance measurement sequence acquired by the laser displacement sensor, median filtering is used to remove outliers, followed by low-pass filtering to smooth the data. The filter cutoff frequency is set according to the grinding wheel rotation frequency and surface roughness characteristics, preserving the true surface morphology information while removing noise. For visual images, if noise interference exists, Gaussian filtering or bilateral filtering is used for noise reduction to maintain edge sharpness. For acoustic emission signals, a bandpass filter is used to filter out low-frequency mechanical vibrations and high-frequency electromagnetic interference, retaining the characteristic frequency bands related to the material removal process. For temperature data, time-domain smoothing filtering is used to eliminate sensor random noise.

[0077] To reduce data transmission volume and subsequent processing burden, preliminary feature extraction is performed on some sensor data at edge computing nodes. For acoustic emission signals, a sliding time window is used for short-time Fourier transform to extract feature parameters such as spectral energy distribution, dominant frequency components, and root mean square value of the signal within each time window, converting the high-frequency raw signal into a sequence of low-frequency feature parameters. For visual images, a rapid quality assessment is performed, calculating the overall brightness, contrast, and sharpness indicators of the image. If the quality of a frame does not meet the requirements, it is marked as an invalid frame to avoid wasting computational resources in subsequent processing. For temperature data, statistical features such as peak temperature, average temperature, and temperature gradient at each sampling time are extracted, and preprocessed monitoring data is output.

[0078] Step 1.4: Spatiotemporal alignment of multi-source monitoring data;

[0079] Based on the multi-sensor monitoring system calibration parameters output in step 1.1 and the preprocessed monitoring data output in step 1.3, spatiotemporal alignment of multi-source monitoring data is achieved. Time alignment maps data collected by different sensors at different times onto a unified time axis. Considering the different sampling frequencies and time delays of each sensor, based on the calibration parameters of each sensor monitoring system output in step 1.1, time delay compensation is performed on the timestamps of each sensor, subtracting the calibration time delay of that sensor to obtain the actual processing time reflected by the data.

[0080] A unified time grid is established, with the grid interval determined based on the system's required update cycle. For sensor data with a sampling frequency higher than the update frequency, data aggregation is performed within each time grid interval, such as taking the average, median, or maximum / minimum value, compressing multiple sampling points into a single representative value. For sensor data with a sampling frequency lower than the update frequency, linear interpolation or spline interpolation is used to generate interpolated data at each time grid point. After time alignment, all types of sensor data have corresponding data values ​​at each time grid point.

[0081] Spatial alignment maps the spatial positions measured by different sensors to a unified machining coordinate system. Based on the vision system calibration parameters output in step 1.1, the image coordinates are converted into machining coordinate system coordinates. Using the camera intrinsic parameter model and hand-eye calibration matrix, the spatial position of each pixel in the image is calculated. Based on the laser displacement sensor installation pose parameters output in step 1.1, the measured distance is converted into the coordinates of points on the grinding wheel surface in the machining coordinate system. Combined with the grinding wheel rotation angle, the circumferential measurement sequence is unfolded into a spatial point cloud. For sensors that perform indirect spatial measurements such as acoustic emission and temperature, the location range of their monitoring area in the machining coordinate system is determined based on their measurement principle and installation position, and the collected data is associated with the corresponding spatial region.

[0082] After spatiotemporal alignment, a standard multi-source monitoring dataset is generated. Each data unit in the dataset contains information such as time label, spatial coordinates, measurement value, and sensor type, providing standardized input for subsequent feature extraction and fusion processing.

[0083] In some embodiments, since some sensors may experience short-term data loss or quality degradation due to processing splashes, strong light interference, etc., an adaptive fusion weight adjustment method based on data confidence assessment can be adopted. The purpose is to dynamically adjust the contribution of a sensor in the fusion when the sensor data quality fluctuates, so as to avoid low-quality data affecting the overall evaluation accuracy. Specifically, a quality evaluation index is established for the data of each sensor, including signal-to-noise ratio, measurement consistency, and continuity with historical data. Data confidence is calculated based on the quality evaluation index. When the confidence of a sensor's data decreases, its weight in subsequent feature fusion is automatically reduced. When the data quality recovers, its weight is gradually restored. This adaptive adjustment mechanism is achieved by online monitoring of the statistical characteristics of each sensor's data. For example, for a laser displacement sensor, if the variance of the measured value suddenly increases by more than a set multiple, it is judged that there may be interference, and the confidence is reduced. For visual images, if the number of feature point matches in multiple consecutive frames is lower than a threshold, it is judged that the image quality has deteriorated, and the confidence is reduced.

[0084] Step 2: Based on the standard multi-source monitoring dataset, perform multimodal feature extraction and feature selection to obtain a multimodal key feature set;

[0085] Step 2.1: Extract geometric deviation features from the topographic point cloud data;

[0086] Based on the point cloud data of the grinding wheel surface topography collected by the laser displacement sensor in the standard multi-source monitoring dataset output in step 1, geometric deviation features are extracted. The point cloud data is registered with the target trimming surface using an iterative nearest-point algorithm or a feature-based registration method to solve for the overall pose transformation parameters of the point cloud, aligning it spatially with the target surface. The normal distance from each measurement point in the point cloud to the target surface is calculated; this distance represents the shape deviation at that location. A shape deviation field distribution is generated to describe the deviation at various locations across the entire grinding wheel surface. Statistical analysis is performed on the deviation field to extract global features of the deviation, including parameters such as maximum positive deviation, maximum negative deviation, mean deviation, standard deviation of deviation, skewness, and kurtosis of the deviation distribution.

[0087] Further extraction of the spatial distribution characteristics of the deviation is performed. The grinding wheel surface is divided into multiple regions, and the average deviation and deviation gradient of each region are calculated to identify areas of concentrated deviation. Through frequency domain analysis, Fourier transforms are performed on the deviation field along the circumferential and axial directions to extract the spatial frequency components of the deviation. Low-frequency components reflect overall shape deviations, while high-frequency components reflect surface roughness and local unevenness. Morphological analysis of the deviation field is conducted to identify the geometric features of the deviation, such as the presence of specific shape defects like steps, grooves, and protrusions. These features are crucial for determining whether abnormal removal has occurred during the finishing process.

[0088] Extract local curvature features. Based on point cloud data, a local surface fitting method is used to calculate the principal curvature and average curvature at various locations on the grinding wheel surface. Curvature features reflect the degree of surface curvature. For grinding wheels with complex profiles, whether the curvature distribution meets the design requirements is an important indicator for evaluating shape accuracy. Output geometric deviation feature parameters.

[0089] Step 2.2: Extract surface quality features from visual image data;

[0090] Based on the image sequence acquired by the centralized visual imaging system from the standard multi-source monitoring dataset output in step 1, the quality features of the grinding wheel surface are extracted. Image preprocessing, including distortion correction and illumination normalization, ensures the comparability of images from different times and under different illumination conditions. Texture analysis is performed on the trimmed grinding wheel surface image, using the gray-level co-occurrence matrix method to extract texture feature parameters, including contrast, entropy, energy, and correlation. These parameters reflect the uniformity of the surface's microstructure. Edge detection is performed on the image, using the Canny or Sobel operator to extract the grinding wheel contour boundary. The integrity of the trimmed boundary is evaluated by the sharpness and continuity of the boundary.

[0091] Surface burn areas are identified through color space analysis. During laser finishing, excessive energy or prolonged exposure time can lead to localized overheating and burns, which manifest as different color characteristics in the image. The image is converted from RGB to HSV or Lab color space, and a feature threshold for the burn area is set within the color space. Threshold segmentation is used to identify potential burn areas, and the area percentage and distribution of these burn areas are calculated.

[0092] Image segmentation techniques are used to identify material removal boundaries. During the trimming process, the removed and unremoved areas exhibit obvious texture or color differences in the image. Semantic segmentation methods based on region growing or deep learning are used to segment the image into different regions, extracting the area, shape, and boundary features of the removed areas. These features reflect the progress and uniformity of material removal, and surface quality feature parameters are output.

[0093] Step 2.3: Extract material removal process features from acoustic emission signals;

[0094] Based on the signals collected by the acoustic emission sensors in the standard multi-source monitoring dataset output from Step 1 and their preliminary feature extraction results, further characteristic parameters reflecting the material removal process are extracted. The time-frequency analysis results of the acoustic emission signals are used to extract the dominant frequency components and their time-domain variations. The removal mechanisms of grinding wheel materials include multiple modes such as brittle fracture, thermal melting, and thermal stress spalling. Different removal mechanisms produce acoustic emission signals with different spectral characteristics. Signals generated by brittle fracture have a higher dominant frequency and pulse characteristics, while signals generated by thermal melting have a lower dominant frequency and are more continuous. By analyzing the distribution and evolution of the dominant frequency, the current material removal mechanism is identified.

[0095] Extract the energy characteristics of the acoustic emission signal. Calculate the cumulative energy and rate of change of the signal within each time window; the energy level reflects the speed of material removal. Compare the acoustic emission energy at different spatial locations or time periods to assess the uniformity of material removal. If the energy in certain areas or time periods deviates from the average level, it indicates that the removal status at these locations is abnormal.

[0096] Extract acoustic emission event features. By setting amplitude thresholds, identify sudden events in the acoustic emission signals and statistically analyze the frequency, duration, and amplitude distribution of these events. Sudden events may correspond to discrete removal behaviors such as the shedding of large abrasive particles or the breakage of binders. Event features help assess the stability of the processing and output characteristic parameters of the material removal process.

[0097] Step 2.4: Extract heat accumulation effect features from temperature data;

[0098] Based on the temperature distribution data collected by the infrared temperature sensor in the standard multi-source monitoring dataset output from step 1, the characteristics of thermal accumulation effect are extracted. For the temperature field at each sampling time, spatial statistical features are extracted, including peak temperature, average temperature, and the standard deviation of the temperature distribution. The peak temperature reflects the highest temperature near the heat source. If the peak temperature exceeds the tolerance temperature of the grinding wheel binder or the phase transition temperature of the material, it may lead to a decrease in binder strength or changes in material microstructure. The average temperature reflects the overall level of thermal accumulation.

[0099] Extract the temporal evolution characteristics of temperature. For the time series of peak and average temperatures, calculate their rate of change and trend. A continuous rise in temperature indicates increasing heat accumulation, while a stable temperature indicates that heat input and heat dissipation have reached equilibrium. By fitting the temporal evolution curve of temperature, predict the temperature level at future times.

[0100] Extracting the characteristics of the heat-affected zone (HAZ). Based on the spatial distribution of the temperature field, the HAZ is delineated using temperature threshold or gradient threshold methods. The area and geometry of the HAZ are calculated. The size of the HAZ reflects the degree of heat diffusion and is of great significance for assessing thermal deformation and residual stress distribution.

[0101] Analyze the temperature gradient distribution. Calculate the spatial gradient of the temperature field; the magnitude of the temperature gradient reflects the level of thermal stress. An excessively large temperature gradient may lead to internal cracks or deformation in the material. Identify regions with abnormal temperature gradients as early warning signals of potential quality risks and output characteristic parameters of thermal accumulation effects.

[0102] Step 2.5: Perform feature selection and construct a multimodal key feature set;

[0103] The geometric deviation feature parameters output from step 2.1, the surface quality feature parameters output from step 2.2, the material removal process feature parameters output from step 2.3, and the thermal accumulation effect feature parameters output from step 2.4 are summarized to form a preliminary multimodal feature set. The preliminary multimodal feature set contains feature parameters from multiple dimensions such as geometric deviation features, surface quality features, material removal features, and thermal state features, and the total number of features may reach dozens or even hundreds.

[0104] To improve the efficiency of subsequent processing and avoid interference from redundant features, feature selection is performed. Initial screening is conducted based on the physical meaning of the features, removing features that are clearly irrelevant to the processing state or are redundant. The correlation between each feature and processing state indicators, such as shape accuracy, surface quality, and processing efficiency, is calculated. Correlation coefficients and mutual information are used to evaluate the feature correlation. Features with high correlation are retained, while those with low correlation are removed.

[0105] Further, feature importance assessment methods are employed. A regression or classification model is established between processing states and features, and the model is trained using methods such as random forest and gradient boosting tree. During training, these methods calculate the importance score of each feature, rank the features according to their importance scores, and select the features with the highest importance ranking as key features.

[0106] Simultaneously, redundancy between features should be considered. If two features are highly correlated, it indicates they contain similar information, and only one needs to be retained. A stepwise forward selection or stepwise backward elimination method should be employed to minimize the number of features while ensuring model performance.

[0107] Finally, a multimodal key feature set is constructed, containing feature parameters that best characterize the current processing state. The number of features is controlled within a reasonable range, ensuring both information sufficiency and computational efficiency. Each feature in the feature set is normalized to eliminate dimensional differences, facilitating subsequent fusion calculations.

[0108] In some embodiments, since the characterization ability of various features to the processing state may differ under different processing stages or different grinding wheel materials, a dynamic feature weight adjustment method based on processing context can be adopted. The purpose is to dynamically adjust the weight of various features in the state assessment according to contextual information such as the current processing stage, grinding wheel type, and process parameters, thereby improving the pertinence and accuracy of the assessment. Specifically, the processing task is divided into roughing stage, semi-finishing stage, and finishing stage. In the roughing stage, the shape deviation feature has a higher weight because the focus is mainly on the rapid approximation of the overall shape. In the finishing stage, the surface quality feature and thermal state feature have higher weights because the focus is mainly on surface details and avoiding thermal damage. The weight of acoustic emission features is adjusted according to the different grinding wheel materials. For brittle materials such as corundum grinding wheels, the acoustic emission feature has a higher weight; for tough materials such as resin-bonded grinding wheels, the thermal state feature has a higher weight. This dynamic weight adjustment is achieved by establishing a mapping relationship between context and weight. The mapping relationship is determined based on expert experience or historical data statistics. In practical applications, the corresponding weight configuration is queried according to the current context.

[0109] Step 3: Based on the multimodal key feature set, perform multimodal feature fusion and evaluate the current processing status of the grinding wheel laser dressing to obtain a comprehensive evaluation result of the current processing status;

[0110] Step 3.1: Establish a feature fusion rule base based on expert knowledge;

[0111] To achieve rapid feature fusion and avoid the problems of large training data requirements and long computation time caused by complex deep neural networks, a feature fusion rule base based on expert knowledge and physical mechanisms is established based on the multimodal key feature set output in step 2. The feature fusion rule base contains multiple fusion rules, each describing how to fuse multimodal features under specific conditions.

[0112] The fusion rule takes the form of: if the current processing state satisfies condition A, then fusion strategy B is adopted. Condition A can be a judgment of the processing stage, a threshold judgment of feature values, a pattern judgment of feature combination, etc.; fusion strategy B includes the weight allocation of various features, the method of feature combination, the calculation method of evaluation output, etc.

[0113] For example, in Rule 1, if the maximum value of the geometric deviation is greater than the set threshold and the standard deviation of the deviation is greater than the set threshold, it indicates that the current shape deviation is obvious and unevenly distributed. In this case, the weight of the geometric deviation feature is set to the highest value, the weight of the material removal feature is second, and the weights of the surface quality feature and the thermal state feature are set to the lowest value. The fusion output focuses on a detailed description of the shape deviation field.

[0114] Rule 2: If the maximum value of the geometric deviation is less than the set threshold and the peak temperature is greater than the set threshold, it means that the shape is close to the target but the heat accumulation is serious. At this time, the weight of the thermal state feature is set to the highest value, the weight of the surface quality feature is second, and the weight of the geometric deviation feature is reduced. The fusion output focuses on the heat accumulation level and the risk assessment of surface burn.

[0115] Rule 3: If the standard deviation of the acoustic emission signal energy is greater than the set threshold, it indicates that the material removal is uneven or the processing is unstable. In this case, the weight of the material removal feature is increased, and the fusion output adds evaluation of removal uniformity and processing stability indicators.

[0116] The rules in the feature fusion rule base are established based on the domain experts' experience and understanding of the physical process of laser trimming, covering various typical processing state modes. In practical applications, the condition part of each rule is checked one by one according to the current feature value. If the condition is met, the corresponding selected fusion strategy is applied.

[0117] Step 3.2: Use a lightweight model for feature fusion and state evaluation;

[0118] Based on the selected fusion strategy output in step 3.1, feature fusion calculation is performed. Feature fusion adopts either weighted summation or weighted nonlinear mapping. For linear fusion, the fused feature vector is calculated as a weighted sum of the feature vectors of each modality, with the weights determined by the fusion rule base. For nonlinear fusion, a lightweight mapping model is established using a shallow neural network structure. This network includes an input layer, two hidden layers, and an output layer. The input layer receives an 84-dimensional multimodal fusion feature vector. The first hidden layer has 64 neurons, and the second hidden layer has 32 neurons, both using the ReLU activation function. The output layer generates a 16-dimensional state representation vector. The network maps the multimodal features to a unified fusion feature space through forward propagation. The complexity of the mapping model is controlled within a reasonable range to ensure that the computation time meets real-time requirements.

[0119] The processing status is evaluated based on the fused feature vector. The status evaluation model employs lightweight machine learning models such as gradient boosting decision trees or support vector machines. The model's input is the fused feature vector, and the output consists of multiple evaluation metrics for the processing status, including shape deviation score, uniformity removal score, thermal accumulation level, processing stability score, and surface quality level.

[0120] Shape deviation scoring comprehensively considers the maximum value, standard deviation, and spatial distribution characteristics of the deviation. A scoring function maps the deviation characteristics to a scoring space; the lower the score, the closer the shape is to the target. Removal uniformity scoring is calculated based on the spatial distribution uniformity of acoustic emission energy and the regularity of the removed boundaries in the visual image; a higher score indicates more uniform removal. Thermal accumulation level is divided into multiple levels based on temperature peaks and the extent of the heat-affected zone; a higher level indicates more severe thermal accumulation. Processing stability scoring is based on the characteristics of acoustic emission events and the temporal fluctuations of each characteristic; a higher score indicates more stable processing. Surface quality level comprehensively considers surface texture uniformity, the proportion of burned areas, and boundary integrity, and is divided into excellent, good, medium, and poor levels.

[0121] Specifically, the shape deviation score comprehensively considers the maximum value, standard deviation, and spatial distribution characteristics of the deviation. A scoring function maps the deviation characteristics to a scoring space. Lower scores indicate that the shape is closer to the target. 0-20 points are excellent (maximum deviation <0.1mm), 21-40 points are good (maximum deviation 0.1-0.2mm), 41-60 points are average (maximum deviation 0.2-0.3mm), and above 61 points are poor or very poor (maximum deviation >0.3mm). The removal uniformity score is calculated based on the spatial distribution uniformity of acoustic emission energy and the regularity of the removed boundaries in the visual image. Higher scores indicate more uniform removal. Above 80 points is excellent (energy distribution standard deviation <10%), 60-79 points are good, 40-59 points are average, and below 40 points are poor. The heat accumulation level is divided into multiple grades based on the peak temperature and the extent of the heat-affected zone (HAZ). Higher grades indicate more severe heat accumulation: Grade 1 is slight (temperature <200℃, HAZ <0.5mm), Grade 2 is moderate (temperature 200-400℃), Grade 3 is severe (temperature 400-600℃), and Grade 4 is extremely severe (temperature >600℃, HAZ >1.5mm). The processing stability score is based on acoustic emission event characteristics and the temporal fluctuations of each characteristic. Higher scores indicate more stable processing: 80 points or above is excellent (amplitude fluctuation <15%), 60-79 points is good, 40-59 points is average, and below 40 points is poor. The surface quality level comprehensively considers surface texture uniformity, the proportion of burned areas, and boundary integrity, and is divided into excellent, good, medium, and poor grades. Excellent requires texture uniformity >90% and no burned areas; good allows burned areas <5%; medium allows burned areas 5-15%; and poor allows burned areas >15%.

[0122] The state assessment model was trained offline. Training data came from historical processing experiments, including data samples with different grinding wheels, process parameters, and processing stages. Each sample contained an extracted feature vector and a state assessment result obtained through manual annotation or measurement. A supervised learning method was used to train the model, employing a mini-batch gradient descent algorithm with a batch size of 64 and a learning rate of 0.001. The Adam optimizer was used for parameter updates, and the training run consisted of 200 epochs. Cross-validation was used to ensure the model's generalization performance, enabling it to learn the mapping relationship between features and states and output a state assessment metric.

[0123] Step 3.3: Calculate the confidence level of the state assessment and output the assessment results;

[0124] The confidence level of the state evaluation metrics output in step 3.2 is calculated. The confidence level reflects the reliability of the evaluation results. Several factors are considered in the confidence level calculation. First, the quality of the input features is considered. If some sensor data is missing or has high noise, resulting in incomplete or inaccurate extracted features, the confidence level decreases. Second, the similarity between the current state and the training data distribution is considered. If the current state features are within the coverage of the training data, it indicates that the model's evaluation ability in that state has been sufficiently trained, and the confidence level is high. If the current state exceeds the training data range, it is considered model extrapolation, and the confidence level decreases. Third, the consistency between multiple evaluation metrics is considered. If the state judgments given by different metrics contradict each other, it indicates uncertainty, and the confidence level decreases.

[0125] The confidence level is calculated using a comprehensive scoring method. A confidence score function is established, taking the quantified values ​​of the above factors as input and outputting the confidence score value. The confidence score value is normalized to between 0 and 1, with the value closer to 1 indicating a higher confidence level.

[0126] The system outputs a comprehensive evaluation result of the current processing status, including the numerical values ​​or levels of each evaluation indicator, confidence levels, values ​​of key features, and identifiers of abnormal signals. The evaluation results are transmitted to the subsequent status prediction and compensation calculation modules in structured data format. Simultaneously, the evaluation results can be displayed in real time through a human-machine interface, providing operators with visual monitoring of the processing status.

[0127] Step 3.4: Implement incremental learning to update the state evaluation model;

[0128] To improve the continuous adaptability of the state assessment model, an incremental learning mechanism is implemented. During actual processing, new data samples are accumulated, including feature vectors and actual processing results. After processing, the dressed grinding wheel is precisely measured to obtain true shape accuracy, surface quality, and other indicators. These true results are then correlated with the feature vectors at that time to form new training samples.

[0129] The state evaluation model is updated periodically or when the number of accumulated new samples reaches a set threshold. An incremental learning algorithm is used to add new samples to the training set, retraining or fine-tuning the state evaluation model to enable it to learn new processing patterns and state features. The incremental learning process retains existing model knowledge, avoiding catastrophic forgetting, while simultaneously absorbing new knowledge to improve the model's performance in new scenarios.

[0130] After the state evaluation model is updated, validation tests are conducted to ensure that the performance of the updated model is no less than that of the original model. If the validation passes, the new model is deployed to the online application; if the validation fails, the original state evaluation model is retained and used, the reasons for the update failure are analyzed, the incremental learning strategy is adjusted, and the updated state evaluation model is output.

[0131] In some embodiments, the initial accuracy of the state assessment model is limited due to insufficient initial training data for certain special grinding wheel materials or novel processing techniques. A rapid model adaptation method based on transfer learning can be adopted. The aim is to utilize the knowledge of existing models trained in similar scenarios to quickly construct an assessment model suitable for new scenarios, reducing data requirements and training time in new scenarios. Specifically, an existing scenario with similar material properties or processing mechanisms to the new scenario is selected as the source domain. This source domain already has a well-trained state assessment model. The parameters of the feature extraction layer and some fusion layers of the source domain model are transferred to the new scenario model, keeping these layer parameters unchanged or only fine-tuning them. Only the output layer and some decision layers of the model are trained using a small amount of data from the new scenario. Through transfer learning, the new model can inherit the general feature representation capabilities learned by the source model. Only targeted adjustments are needed for the specificities of the new scenario, reducing the training data requirements. For example, when migrating from a corundum grinding wheel scenario to a silicon carbide grinding wheel scenario, the physical mechanisms of material removal are common, and relevant knowledge can be transferred. Only adjustments need to be made for differences in material hardness, brittleness, etc.

[0132] Step 4: Based on the comprehensive evaluation results of the current processing status, predict the processing status of the grinding wheel laser dressing and output the comprehensive prediction results;

[0133] Step 4.1: Establish a simplified physical model for material removal;

[0134] A simplified physical model describing the interaction between the laser and the grinding wheel material is established to predict the amount of material removed. The inputs to the simplified physical model include parameters such as laser power, spot diameter, scanning speed, laser absorptivity of the material, and material removal threshold. An energy threshold criterion is used: when the energy density received per unit area of ​​the laser-interacting region exceeds the material removal threshold, the material is removed; the removal depth is directly proportional to the energy density.

[0135] The energy distribution within the laser spot is calculated, assuming a Gaussian distribution. Based on the laser power and spot diameter, the energy density at the spot center and edges is calculated. The effect of scanning speed is considered; the faster the scanning speed, the less energy is received per unit area. The material removal depth at each position along the scanning path is calculated. The relationship between the removal depth and the local energy density and material removal threshold is determined through experimental calibration or empirical formulas.

[0136] Based on the comprehensive evaluation results of the current processing state output in step 3, the parameters of the simplified physical model are adjusted. According to the shape deviation score and removal uniformity score, the removal threshold parameter in the model is adjusted: in areas with low removal uniformity scores, the removal threshold and removal rate coefficient for that area are corrected based on the deviation between the actual removal state and the expected result; according to the heat accumulation level, the laser absorptivity parameter of the material is adjusted, as the material properties change in high-temperature regions, and the absorptivity is adjusted accordingly; according to the processing stability score, the random perturbation parameter of the model is adjusted, increasing the uncertainty range of the removal amount in areas with poor stability, making the model prediction closer to the actual processing state.

[0137] Step 4.2: Establish a simplified physical model of heat conduction;

[0138] A simplified heat conduction model describing laser heating and thermal diffusion is established to predict the temperature field evolution. A lumped parameter model is used, dividing the grinding wheel surface into multiple lumped cells, with the temperature of each cell represented by an average temperature value. Cells within the laser-affected region receive laser energy input; the energy input rate is the laser power multiplied by the proportion of energy received by that cell. Heat is exchanged between cells via thermal conduction; the heat conduction rate is proportional to the temperature difference between adjacent cells and the thermal conductivity coefficient. Heat is dissipated between the cells and the environment through convection and radiation; the heat dissipation rate is related to the temperature difference between the cell and the environment.

[0139] The heat balance equations for each unit are established, where the rate of change of unit temperature equals the energy input rate minus the heat conduction output rate minus the heat dissipation rate, divided by the unit's heat capacity. The heat balance equations for each unit are numerically solved using explicit or implicit time integration methods to obtain the temperature values ​​of each unit at each moment, thus forming the spatiotemporal evolution of the temperature field.

[0140] The thermal properties parameters in the model, such as thermal conductivity, specific heat capacity, and convective heat transfer coefficient, are obtained from tables or determined experimentally based on the type of grinding wheel material. Considering the influence of temperature on material properties, such as the decrease in binder strength and changes in thermal conductivity at high temperatures, a temperature-related parameter correction function is introduced to improve model accuracy and output predicted results of temperature field evolution.

[0141] Step 4.3: Establish a simplified physical model of thermal deformation;

[0142] Based on the temperature field evolution prediction results output in step 4.2, a simplified model of thermal deformation is established to predict the thermal stress and thermal deformation of the grinding wheel. Using thermal stress theory, temperature changes cause thermal strain in the material; when this thermal strain is constrained, thermal stress is generated, leading to material deformation.

[0143] The grinding wheel structure is simplified, treating it as an axisymmetric body, and deformation calculations are performed using beam theory or shell theory. Based on the temperature field distribution, the thermal strain at each location is calculated; the thermal strain equals the coefficient of linear expansion multiplied by the temperature change. Based on the material's elastic modulus and Poisson's ratio, a stress-strain relationship is established, and the thermal stress distribution is solved. Based on the thermal stress and boundary conditions, the deformation field is solved using the energy method or finite difference method to obtain the displacement at each location of the grinding wheel.

[0144] Thermal deformation causes changes in the surface shape of the grinding wheel. The predicted displacement is superimposed on the current shape to update the shape deviation field, which serves as the predicted value of the shape deviation at future moments, and the thermal deformation prediction result is output.

[0145] Step 4.4: Use a data-driven approach to predict the state evolution trend;

[0146] A time-series prediction model based on historical data is established to predict the evolution trend of future states. Long Short-Term Memory (LSTM) networks are used as the main time-series prediction method. This network structure includes an input layer, an LSTM layer, and an output layer. The input layer receives historical state feature sequences with a time window length of 10 time steps, and each time step has a feature dimension of 84. The LSTM layer adopts a two-layer structure, with each layer containing 128 hidden units. A dropout rate of 0.2 is set between layers to prevent overfitting. The LSTM units learn long-term dependencies through forgetting gates, input gates, and output gates, effectively capturing the temporal patterns of state evolution. The output layer performs feature transformation through two fully connected layers: the first layer has 64 neurons, and the second layer has 32 neurons. Finally, it outputs the predicted state feature values ​​for the next 5 time steps, with an output dimension of 5×84. The model training process uses a historical processed dataset, employing a time-series sliding window method to construct training samples. The batch size is set to 32, the learning rate is 0.001, and the Adam optimizer is used for parameter optimization. The model converges after 200 training epochs.

[0147] By performing time-series modeling on historical processing data, patterns in state evolution under different processing conditions can be extracted. For example, during continuous trimming, temperature initially rises and then stabilizes, while shape deviation gradually decreases and converges. By learning these patterns, the time-series model can predict the possible future state based on current and historical states.

[0148] The time-series prediction model takes into account the influence of processing conditions and inputs future process parameters such as laser power and scanning speed as exogenous variables into the model, so that the prediction can reflect the impact of changes in process parameters on state evolution.

[0149] Step 4.5: Integrate the prediction results from the physical model and the data-driven model;

[0150] Based on the material removal amount output in step 4.1, the temperature field evolution prediction results output in step 4.2, the thermal deformation prediction results output in step 4.3, and the future state evolution trend output in step 4.4, a fusion prediction is performed to obtain a comprehensive prediction result. The simplified physical model can provide the basic trend of state evolution based on physical mechanisms, but its prediction accuracy is limited due to model simplification and parameter uncertainty. The data-driven model can learn complex patterns in historical data, but it is highly dependent on training data and has limited generalization ability. Both have their advantages and limitations; fusion can leverage their respective strengths and compensate for their weaknesses.

[0151] A dynamic weight determination method is established: The basic weights are determined based on the reliability of the physical model input parameters. When process parameters such as laser power and scanning speed are stable and sensor data quality is high, the physical model weights are set to 0.6-0.8. Weights are adjusted based on the similarity between the current state and historical training data. Feature space distance is used to measure similarity; when the similarity is higher than 0.8, the data-driven model weights are increased by 0.1-0.2. Weights are fine-tuned based on the consistency of the prediction results of the two models. When the prediction deviation is less than 10%, the current weights are maintained; when the deviation is greater than 20%, the weights of the models with larger deviations are reduced by 0.1-0.3. Finally, weights are dynamically adjusted based on historical prediction accuracy. The prediction errors of each model in recent times are statistically analyzed, and the weights of models with higher prediction accuracy are increased accordingly.

[0152] A weighted fusion strategy is adopted, which involves a weighted average of the predictions from the physical model and the data-driven model, with the weights dynamically determined according to the method described above. Another fusion method is hierarchical fusion, where the physical model prediction serves as the base prediction, and the data-driven model prediction serves as a correction. The physical model provides the main evolutionary trend of the state, while the data-driven model predicts the prediction error of the physical model. The error prediction value is then superimposed on the physical model prediction result to obtain the final comprehensive prediction result.

[0153] The comprehensive prediction results include state parameters such as shape deviation prediction, temperature field prediction, and removal rate prediction at each time point within the future time window. At the same time, the uncertainty assessment of the prediction is output, and the uncertainty reflects the reliability of the prediction results.

[0154] Step 4.6: Implement rolling window prediction and dynamic updates;

[0155] A rolling window prediction strategy is adopted, which predicts the state evolution within a fixed time window in each update cycle. The length of the time window is determined based on the time requirements for compensation calculation and execution, and must be long enough to support forward-looking compensation, but short enough to ensure prediction accuracy.

[0156] As processing progresses, deviations may arise between the actual and predicted states. These deviations can be caused by factors such as model errors, disturbances, and adjustments to process parameters. A dynamic update mechanism is employed, comparing newly collected actual monitoring data with previous predictions at each update cycle to calculate the prediction error. If the prediction error exceeds an acceptable range, the source of the error is analyzed, model parameters are adjusted or weights are fused, and the prediction is repeated.

[0157] Through scrolling windows and dynamic updates, the forecasting process can continuously self-correct, adapt to dynamic changes in the processing, and maintain the accuracy and timeliness of the forecasts.

[0158] In some embodiments, due to individual differences and uncertainties in parameters of the simplified physical model, such as the material removal threshold and thermal conductivity, directly using empirical values ​​or lookup table values ​​may lead to large prediction errors. An adaptive correction method for model parameters based on online data assimilation can be adopted. The aim is to dynamically correct key parameters of the physical model using real-time monitoring data, thereby improving the model's prediction accuracy. Specifically, a parameter estimation framework is established, treating the key parameters of the physical model as state variables to be estimated. Data assimilation methods such as Kalman filtering or particle filtering are used, with real-time monitoring data as the observation input and the physical model as the state transition model. Through Bayesian inference, the optimal values ​​and uncertainties of the model parameters are estimated online. For example, for the material removal threshold, an empirical value is initially used. As processing progresses, the true removal threshold is inferred based on the actual measured removal depth and corresponding laser energy density. Through the accumulation of multiple measurement data, the value gradually converges to the true parameter value. This online correction mechanism enables the physical model to adaptively adjust to the current specific grinding wheel and processing conditions, improving prediction accuracy.

[0159] Step 5: Based on the comprehensive evaluation results and comprehensive prediction results of the current processing status, perform multi-objective compensation optimization on the basic path planning results to obtain the comprehensive path compensation amount;

[0160] Step 5.1: Establish a multi-objective compensation optimization problem;

[0161] Based on the comprehensive evaluation results of the current processing status output in step 3, the comprehensive prediction results output in step 4, and the basic path planning results output in step 1.2, the path compensation problem is formulated as a multi-objective optimization problem. The optimization objectives include: minimizing shape deviation, making the shape of the dressed grinding wheel as close as possible to the target profile; maximizing removal uniformity, avoiding local over-removal or under-removal; controlling the temperature level, avoiding material damage and thermal deformation caused by heat accumulation; and maintaining execution smoothness, avoiding equipment vibration and processing shock caused by abrupt changes in compensation commands.

[0162] Establish mathematical expressions for each objective. The objective function for shape deviation is the norm of the compensated shape deviation field, such as the maximum deviation or root mean square deviation. The objective function for removing uniformity is the variance or standard deviation of the velocity spatial distribution. The objective function for temperature control is the difference between the peak temperature and the upper limit of the safe temperature or the area of ​​the over-temperature region. The objective function for smoothness is the rate of change of curvature or the rate of change of acceleration of the compensated path.

[0163] Establish compensation variables. Compensation variables include the position adjustment of path points, that is, the displacement of the basic path planning result points along the normal direction to adjust the distance or action position between the laser and the grinding wheel surface; the scanning speed adjustment, that is, using different scanning speeds in different path segments to control the energy input per unit area; and the laser power adjustment, that is, using different laser powers in different positions to adapt to the spatial changes in material properties.

[0164] Establish constraints. Constraints include: the magnitude of the position adjustment cannot exceed the working range of the actuator; the scanning speed cannot exceed the maximum speed of the equipment and cannot be lower than the minimum speed to ensure processing quality; the laser power cannot exceed the rated power of the laser and cannot be lower than the minimum power to maintain effective removal; the compensated path must meet kinematic constraints, such as acceleration and jerk limits; and the temperature cannot exceed the material's tolerance temperature.

[0165] Step 5.2: Solve the collaborative optimization problem using a hierarchical iterative strategy;

[0166] Due to the complexity of multi-objective optimization problems and the coupling between variables, a hierarchical iterative strategy is adopted for solving them. The optimization problem is decomposed into multiple subproblems, each corresponding to a different optimization objective. By iteratively coordinating the solutions of each subproblem, a comprehensive optimal solution is finally obtained.

[0167] The first layer is the geometric compensation layer, with the goal of minimizing shape deviation. Based on the current shape deviation assessment output from step 3 and the future shape deviation prediction output from step 4, the required positional compensation amount to eliminate the deviation is calculated. An inverse compensation strategy is employed: reducing the laser depth or removal amount in positive deviation regions and increasing the removal amount in negative deviation regions. Specifically, for each location in the deviation field, based on the deviation value and the material removal model, the path position or energy input that needs adjustment is calculated in reverse, generating a preliminary geometric compensation scheme.

[0168] The second layer is the process compensation layer, which aims to maximize removal uniformity. Based on the removal uniformity evaluation index output from step 3 and the geometric compensation scheme of the first layer, the expected removal rate at each location after performing geometric compensation is analyzed. If the removal rate at some locations deviates from the average level, compensation is made by adjusting the scanning speed or laser power. In areas with excessively fast removal rates, the laser power is reduced or the scanning speed is increased to reduce energy input; in areas with excessively slow removal rates, the laser power is increased or the scanning speed is reduced to increase energy input. The process parameter adjustment amount is calculated to make the removal rate at each location tend to be consistent.

[0169] The third layer is the temperature compensation layer, aiming to control the temperature within a safe range. Based on the temperature field evolution prediction results output in step 4 and the compensation schemes of the first two layers, the temperature distribution after compensation is evaluated. If the predicted temperature exceeds the safe threshold, cooling measures are taken. Cooling measures include: further reducing laser power or increasing scanning speed in high-temperature regions to reduce heat input; adjusting the scanning path sequence and inserting cooling waiting time after processing in high-temperature regions to allow heat to dissipate; and adopting a segmented processing strategy to disperse continuous high-energy processing into multiple time periods to avoid concentrated heat accumulation. The process parameters required for temperature control are further adjusted.

[0170] The fourth layer is the coordination iteration layer, which checks the coordination between the compensation amounts of each layer. Because there is coupling between position compensation, velocity adjustment, and power adjustment, an adjustment in one layer may affect the performance of other layers. For example, reducing the power of the temperature compensation layer will affect the removal rate, thus affecting the performance of the geometry compensation layer. An iterative coordination method is used, taking the comprehensive compensation scheme output from the third layer as new input, re-executing the calculations from the first to the third layer, and evaluating the effect after iteration.

[0171] Establish clear convergence criteria: the relative change of each objective function value is less than a set threshold, such as shape deviation improvement less than 0.1%, temperature reduction less than 1%, and uniformity removal improvement less than 0.5%; the change in compensation amount between two consecutive iterations is less than a set threshold, such as position adjustment change less than 0.01mm and power adjustment change less than 1%; and the conflict index between compensation strategies at each layer is less than a set threshold. If the convergence conditions are met and the number of iterations has not reached the upper limit, the iteration ends, and the current compensation scheme is output; if convergence is not achieved and the number of iterations has not reached the upper limit, the weight coefficients of each layer are adjusted, and the iteration continues.

[0172] If convergence is not achieved after reaching the maximum number of iterations, the alternative solution processing logic is activated: a sub-objective optimization strategy is adopted, prioritizing the highest priority objective such as shape accuracy, while appropriately relaxing the requirements of secondary objectives; a regional processing strategy is adopted, dividing the grinding wheel surface into multiple independent regions, with each region employing a locally optimal compensation scheme; a conservative compensation strategy is adopted, reducing the compensation magnitude, replacing a single large compensation with multiple small compensations to ensure system stability, and outputting a suboptimal but executable compensation scheme.

[0173] Step 5.3: Conduct a feasibility check and constraint handling for the compensation plan;

[0174] The obtained comprehensive compensation scheme is subjected to a feasibility check to verify whether it meets all constraints. The position compensation amount is checked to see if it exceeds the working stroke of the actuator; if so, amplitude limiting is applied or the compensation amount is distributed across multiple path segments. The adjusted scanning speed and laser power are checked to see if they are within allowable ranges; if they exceed limits, amplitude limiting is applied. The compensated path is checked to see if it meets kinematic constraints; the path velocity, acceleration, and jerk are calculated. If they exceed equipment performance limits, the constraints are met by increasing the path point density, extending the execution time, or modifying the compensation amount distribution. The compensated path is checked to see if there is a risk of collision with the workpiece or fixture; if so, the compensation direction is adjusted or the local path is replanned.

[0175] For cases where constraints cannot be satisfied through simple adjustments, compensation scheme optimization is initiated. Constraint optimization algorithms, such as sequential quadratic programming or interior-point methods, are employed to search for the optimal compensation scheme within the feasible region that satisfies the constraints. During optimization, the requirements of some secondary objectives can be appropriately relaxed, prioritizing key objectives such as shape accuracy and safety.

[0176] Step 5.4: Calculate the priority and execution strategy for compensation;

[0177] Different types of compensation have varying degrees of impact and urgency on the processing results. Prioritizing compensation guides the execution strategy. Shape deviation compensation has the highest priority because shape accuracy is the primary goal of finishing. Temperature control compensation has the next highest priority because exceeding temperature limits can lead to irreversible material damage. Uniformity removal compensation has the next lowest priority because local non-uniformities can be improved through subsequent processing within a certain range. Smoothness optimization has the lowest priority and mainly plays a supporting role.

[0178] Implementation strategies are formulated based on priorities. When multiple compensation needs conflict, the higher-priority compensation is prioritized. For example, if shape compensation requires increasing the removal amount at a certain location, but temperature control requires reducing the energy input at that location, a trade-off decision is made based on the current temperature level and the magnitude of the shape deviation: if the temperature is close to the limit, temperature control is prioritized, sacrificing some shape compensation effect; if the temperature is within the safe range, shape compensation is prioritized.

[0179] Develop a phased implementation strategy. For large-scale compensations, implement them in stages, with each stage achieving a partial compensation amount to avoid the risks associated with a one-time adjustment. For emergency compensations, such as those due to temperature exceeding limits, execute immediately; for gradual compensations, such as those aimed at progressively improving shape accuracy, execute continuously over multiple update cycles.

[0180] Output comprehensive path compensation, including the position adjustment vector of each path point, the scanning speed adjustment value of each path segment, the laser power adjustment value of each path segment, as well as the priority identifier and execution strategy description of the compensation.

[0181] In some embodiments, due to the possibility that the optimization algorithm may get stuck in local optima under certain extreme conditions and fail to find a globally optimal compensation scheme, a multi-starting-point parallel optimization and solution space exploration method can be adopted. The aim is to increase the probability of finding a globally optimal solution by starting optimization from multiple different initial points and exploring different regions of the solution space. Specifically, based on the current state and empirical rules, multiple different initial compensation schemes are generated. These schemes focus on different optimization objectives or adopt different compensation strategies. Optimization calculations are performed in parallel on each initial scheme. Since the starting points of each optimization thread are different, they may converge to different local optima. The optimal solutions obtained by each thread are compared, and their objective function values, constraint satisfaction, implementation complexity, etc., are evaluated. The solution with the best overall performance is selected as the final compensation scheme. This multi-starting-point parallel strategy makes full use of multi-core computing resources and improves the optimization quality of the compensation scheme under the premise of real-time performance. For scenarios with limited computing resources, heuristic methods can be used to generate a few high-quality initial schemes instead of traversing all possibilities, achieving a balance between optimization quality and computation time.

[0182] Step 6: Based on the comprehensive path compensation amount and the basic path planning results, perform path correction and smoothing to obtain the compensation path instructions;

[0183] Step 6.1: Generate the compensated path point coordinate sequence;

[0184] Based on the comprehensive path compensation amount output in step 5 and the basic path planning result output in step 1.2, the basic path planning result is corrected to generate a compensated path. For each path point in the path, its position adjustment is superimposed on the original coordinates to obtain the new compensated coordinates. The position adjustment is usually performed along the normal direction of the grinding wheel surface. The three-dimensional position of the new coordinates is calculated based on the local surface normal vector and the adjustment amount.

[0185] The connection method between path points needs to be processed. Path points in the basic path planning results are typically connected by straight line segments or circular arc segments. However, the positions of the compensated path points change, requiring the regeneration of connection curves. Interpolation methods are used to generate smooth connection curves between the compensated path points. Commonly used interpolation methods include linear interpolation, circular interpolation, and spline interpolation. An appropriate interpolation method is selected based on the spatial distribution and curvature requirements of the path points.

[0186] For adjustments to scanning speed and laser power, the adjusted values ​​are associated with the corresponding path segments. Process parameter information, including the desired scanning speed, laser power, and defocusing amount, is attached to each path segment or path point. These parameters are read and set by the control system during execution, outputting a compensated sequence of path point coordinates.

[0187] Step 6.2: Smooth the compensation path;

[0188] Based on the compensated path point coordinate sequence output in step 6.1, the compensated path is smoothed. The compensated path may contain local discontinuities or abrupt curvature changes, affecting the smoothness of execution and processing quality. A path smoothing algorithm is used to optimize the path. Commonly used smoothing methods include B-spline fitting, Bézier curve fitting, and low-pass filtering.

[0189] Establish constraints for smoothing: Geometric accuracy constraints require that the maximum deviation between the smoothed path and the original compensation path does not exceed 10% of the compensation accuracy requirement, and the root mean square deviation does not exceed 5%; curvature continuity constraints require that the rate of curvature change at each point on the path be within the set range to avoid impact during execution; key point preservation constraints require that the smoothed path must pass through or approach key compensation points, such as the position with the largest shape deviation, with the deviation not exceeding 0.01mm; boundary condition constraints require that the position and tangential of the path's start and end points remain unchanged.

[0190] A B-spline fitting method is used, with compensated path points serving as control points or data points to generate a B-spline curve. B-spline curves offer good smoothness and local controllability; by adjusting the spline order and node distribution, the smoothness of the curve and the accuracy of its approximation to the original points can be controlled. During the fitting process, the constrained spline curve maintains a certain accuracy through or near key path points, while allowing appropriate deviations in non-critical regions to achieve smoothness.

[0191] The smoothed path is evaluated for geometric error, and the maximum deviation and root mean square deviation between the smoothed path and the original compensated path are calculated. If the geometric error is within the constraints, the smoothing result is accepted; if the geometric error exceeds the limits, the smoothing parameters such as spline order and number of nodes are adjusted, and the fitting is performed again until the accuracy requirements are met, and the smoothed path is output.

[0192] Step 6.3: Perform path dynamics constraint checks and optimizations;

[0193] Perform dynamic analysis on the smoothed path output in step 6.2, calculating the velocity, acceleration, and jerk of each axis during path execution. Based on the path geometry and the desired scanning speed, solve for the motion parameters of each axis at each moment through kinematic calculations.

[0194] Check whether the motion parameters of each axis exceed the performance limits of the actuator. If the speed of a certain axis exceeds the maximum speed, or the acceleration exceeds the maximum acceleration, or the jerk exceeds the maximum jerk, it is determined that the constraint has been violated and adjustment is required.

[0195] The adjustment methods include: reducing the scanning speed by extending the execution time to reduce the dynamic requirements of each axis; increasing the path point density to make the changes in curvature and acceleration smoother through more refined path description; and modifying the local path shape by appropriately increasing the radius of curvature or decreasing the turning angle to reduce the dynamic requirements.

[0196] Iterative optimization is performed, with the dynamic parameters recalculated after each adjustment to check if the constraints are met. If they are met, the optimization ends; otherwise, adjustments continue. During the adjustment process, efforts are made to maintain the satisfaction of the compensation objective, avoiding excessive sacrifice of compensation effect to meet dynamic constraints. The dynamically optimized path is then output.

[0197] Step 6.4 employs segmented trajectory generation and look-ahead buffering;

[0198] To improve the real-time performance and continuity of compensation, a segmented trajectory generation strategy is adopted. The entire adjustment path is divided into multiple path segments, each corresponding to a certain length or execution time. For the currently executing or about-to-be-executed path segment, detailed compensation calculations and trajectory generation are performed; for subsequent path segments, preliminary compensation calculations are performed first, or the compensation strategy of the previous segment is reused.

[0199] A look-ahead buffer mechanism is established so that while the current path segment is being executed, the background simultaneously calculates the compensation trajectories for several subsequent segments and stores the generated trajectories in a buffer. The buffer adopts a first-in, first-out queue structure, with a buffer depth of 3 to 5 path segments, dynamically adjusted according to the calculation and execution speed. When the current segment's execution progress reaches 80%, the next trajectory segment is retrieved from the buffer and preloaded. If the current segment's execution is complete, the system immediately switches to the next trajectory segment, achieving seamless transition.

[0200] Establish a buffer management mechanism: monitor the buffer's fill status, and when the remaining buffer capacity is lower than a set threshold, increase the priority of background calculation to accelerate trajectory generation; when an abnormal status is detected or the compensation strategy needs to be adjusted, clear outdated trajectories in the buffer and recalculate subsequent trajectories; when the buffer overflows, retain the latest calculated trajectory and discard the earlier trajectory to ensure execution continuity.

[0201] Segmented generation also facilitates dynamic adjustment. If a state change or anomaly is detected during the execution of the current segment, the compensation strategy for subsequent segments can be updated in a timely manner without waiting for the entire path calculation to complete, thus improving the system's response speed and adaptability.

[0202] Step 6.5: Output the executable compensation path instructions;

[0203] The optimized compensation path is converted into an instruction format recognizable by the actuator. The instructions include the coordinates of the path points, the motion speed of each axis, acceleration parameters, and process parameters such as laser power and auxiliary gas flow rate. The instruction format conforms to the standards of CNC systems or motion controllers, such as G-code format or dedicated control instruction formats.

[0204] The instruction sequence is sorted and time-calibrated to determine the execution time and order of each instruction. An execution timetable is generated, specifying the start time, duration, and end time of each path segment, providing a time reference for real-time execution and monitoring.

[0205] The output compensation path instruction, including the complete path coordinate sequence, process parameter sequence, execution timetable, compensation identification information, etc., is transmitted to the execution control module to drive the actuator to perform the trimming operation.

[0206] In some embodiments, path smoothing may alter the shape of the compensated path, leading to a deviation between the actual compensation effect and the calculated compensation target. An iterative smoothing method based on error compensation can be employed. The aim is to evaluate the error introduced by smoothing after the smoothing process and compensate for it in subsequent path segments, keeping the accumulated error within a controllable range. Specifically, after smoothing a path segment, the deviation between the smoothed path and the original compensated path is calculated; this deviation is called the smoothing error. The smoothing error is recorded. When generating the compensation amount for the next path segment, in addition to the compensation amount calculated based on the current state, the smoothing error of the previous segment is also added. This ensures that the compensation of the next segment achieves the current target while also correcting the deviation of the previous segment. Through this iterative mechanism of error accumulation and compensation, the gradual accumulation of smoothing error is avoided, maintaining overall compensation accuracy. This method is similar to integral control in a control system, capable of eliminating steady-state errors. For long-path or high-precision trimming tasks, this method can effectively improve the long-term stability and consistency of compensation.

[0207] Step 7: Execute the compensation path command, monitor the execution process in real time and evaluate the execution effect, perform adaptive feedback, and obtain a complete data record and trimming report of the processing process;

[0208] Step 7.1: Issue the compensation path command and start execution;

[0209] The executable compensation path command output from step 6 is sent to the motion control system and the laser control system. The motion control system parses the path command, plans the motion trajectory of each axis, generates an interpolation point sequence, and drives the multi-axis servo motors to move along the planned trajectory. The laser control system sets parameters such as the laser's output power, pulse frequency, and pulse width according to the command. Auxiliary systems such as the cooling system and dust removal system are also configured accordingly based on the command.

[0210] Once the dressing process is initiated, the laser head, driven by a multi-axis mechanism, moves relative to the rotating grinding wheel along a compensation path. The laser beam is focused on the surface of the grinding wheel, interacting with the grinding wheel material to remove surface material and achieve dressing. During the process, various sensors continuously collect and monitor data, and the monitoring system displays the processing status in real time. Operators can observe the dressing progress and status indicators through a human-machine interface.

[0211] Step 7.2: Continuously collect monitoring data and evaluate the implementation effect;

[0212] During the execution of the compensation path, based on the multi-sensor monitoring system established in step 1, the multi-sensor system continues to synchronously collect monitoring data. The collected data is preprocessed and spatiotemporally aligned to form a new monitoring dataset. Based on the new monitoring data, feature extraction in step 2 and state assessment in step 3 are performed to obtain the latest processing status information.

[0213] The current actual state is compared with the expected state to evaluate the effectiveness of the compensation. The expected state is the state evolution prediction output in step 4 or the target state during the compensation calculation in step 5. The actual shape deviation is compared with the expected deviation, the actual temperature with the expected temperature, and the actual removal state with the expected removal state to calculate the state tracking error.

[0214] If the state tracking error is within an acceptable range, it indicates that the compensation is effective and the system continues to execute according to the current strategy. If the state tracking error exceeds the acceptable range, it indicates that there is an unforeseen disturbance or that the compensation strategy is inappropriate, requiring adjustment.

[0215] Step 7.3: Implement adaptive feedback control to adjust the compensation strategy;

[0216] Based on the state tracking error and corresponding evaluation confidence level output from step 7.2, adaptive feedback control is implemented. The feedback control gain is set, and its magnitude is dynamically adjusted according to the error magnitude and confidence level. When the error is small and the confidence level is high, a smaller feedback gain is used, mainly relying on feedforward prediction compensation to maintain smooth execution. When the error is large or the confidence level decreases, the feedback gain is increased to strengthen the correction of actual deviations and improve robustness.

[0217] The adjustment methods of feedback control include: incrementally adjusting the compensation amount by adding a feedback correction amount to the original calculated compensation amount, with the correction amount being proportional to the state error; fine-tuning the process parameters by adjusting the laser power or scanning speed according to the difference between the actual and expected states to make the actual removal effect closer to the expectation; and updating the prediction model parameters. If prediction errors occur multiple times, it indicates that the model parameters may deviate from reality. The model parameters are updated through online parameter identification or incremental learning to improve the accuracy of subsequent predictions.

[0218] Establish specific judgment criteria for anomaly detection: the temperature anomaly threshold is set to exceed 90% of the material's tolerance temperature or the temperature rise rate exceeds 3 times the set value; the shape deviation anomaly threshold is set to the deviation value exceed 2 times the target accuracy requirement or the deviation change rate exceeds 5 times the normal range; the processing stability anomaly is judged by the sudden change in the acoustic emission signal amplitude exceeding 10 times the average value or the spectrum characteristics shifting to a certain extent; the actuator anomaly is identified by the position tracking error exceeding the set threshold or the speed fluctuation exceeding the normal range.

[0219] When an abnormal state is detected, a tiered anomaly handling mechanism is activated: Level 1 anomalies are minor deviations, handled through parameter fine-tuning and enhanced monitoring; Level 2 anomalies are moderate deviations, immediately reducing laser power to a safe level, adjusting the scanning strategy, and increasing cooling time; Level 3 anomalies are severe deviations, immediately suspending processing, initiating safety protection procedures, and performing system self-checks and fault diagnosis. The recovery mechanism after anomaly handling includes: verifying whether the anomaly has been eliminated, reassessing the current state, adjusting subsequent compensation strategies, restarting with lower process parameters, and gradually restoring to normal processing status.

[0220] Step 7.4: Achieve closed-loop data flow and system self-learning;

[0221] The entire repair process forms a complete closed-loop data flow: monitoring data acquisition → feature extraction → state assessment → state prediction → compensation calculation → trajectory generation → execution control → monitoring feedback, with the output of each stage serving as the input for the next stage.

[0222] During closed-loop operation, data is continuously accumulated, including raw data from each sensor, extracted feature parameters, state assessment results, compensation decision-making basis, and performance evaluation. This data is stored in a processing database, forming a traceable data chain.

[0223] Utilize accumulated data to implement system self-learning and performance optimization. Regularly analyze historical data to uncover processing patterns and identify key factors affecting processing quality. Update the state assessment model, prediction model, and compensation strategies to enable the system to learn from past experience and continuously improve its intelligence and adaptability.

[0224] For new grinding wheel types, new combinations of process parameters, and new machining tasks, applicable models and strategies can be quickly established through transfer learning or incremental learning, reducing debugging time and improving the versatility of the system.

[0225] Step 7.5: Complete the trimming and output the results;

[0226] The dressing process ends when the dressing task is completed, meaning all planned path segments have been executed and the status assessment indicates that the dressing objective has been achieved. The laser is turned off, the motion mechanism returns to its initial position, and the dressed grinding wheel is unloaded.

[0227] The dressed grinding wheel undergoes quality inspection using high-precision measuring equipment such as a coordinate measuring machine, profilometer, and surface roughness tester to measure its shape accuracy, surface roughness, and boundary integrity. The measurement results are then compared with the target requirements to assess whether the dressing quality is up to standard.

[0228] Generate a dressing report, which includes: the initial state of the grinding wheel, target requirements, dressing process parameters, implemented compensation strategies, actual quality indicators after dressing, shape accuracy test results, surface quality evaluation, and records of abnormal events during the dressing process. The dressing report is archived to provide a basis for subsequent process improvements and quality traceability.

[0229] The system outputs a complete record of the processing data, including time-series monitoring data, status assessment sequences, compensation decision sequences, and execution trajectory records. This data can be used for process analysis, model validation, system optimization, and other purposes.

[0230] The dressed and qualified grinding wheels are delivered to the subsequent grinding process for precision grinding of aero-engine blades to ensure the quality of blade processing.

[0231] A real-time compensation system for laser trimming paths that integrates online monitoring, such as Figure 3 As shown, a real-time compensation method for laser trimming path based on integrated online monitoring, as described above, includes:

[0232] The multi-sensor monitoring module collects monitoring data from multiple heterogeneous sensors during laser dressing of grinding wheels, plans the laser dressing path, and obtains the basic path planning results; it also performs data preprocessing and spatiotemporal alignment to obtain standard multi-source monitoring data.

[0233] The feature extraction module, based on the standard multi-source monitoring dataset, performs multimodal feature extraction and feature selection to obtain a multimodal key feature set;

[0234] The status assessment module, based on the multimodal key feature set, performs multimodal feature fusion and assesses the current processing status of the grinding wheel laser dressing, obtaining a comprehensive assessment result of the current processing status;

[0235] The status prediction module predicts the processing status of laser dressing of grinding wheels based on the comprehensive evaluation results of the current processing status and outputs the comprehensive prediction results.

[0236] The compensation optimization module performs multi-objective compensation optimization on the basic path planning results based on the comprehensive evaluation results and comprehensive prediction results of the current processing status, and obtains the comprehensive path compensation amount.

[0237] The path generation module performs path correction and smoothing based on the comprehensive path compensation amount and the basic path planning results to obtain the compensated path instructions.

[0238] The execution control module executes compensation path instructions, monitors the execution process in real time and evaluates the execution effect, provides adaptive feedback, and obtains complete data records and trimming reports of the processing process.

[0239] In one embodiment of the present invention, a specific example is provided:

[0240] This invention utilizes a typical microcrystalline corundum CBN grinding wheel with an outer diameter of 300 mm, a width of 40 mm, a grit size of 180 mesh, and a ceramic binder. This grinding wheel is used for profile grinding of titanium alloy blades. The blade profile is a complex, twisted curved surface with a shape accuracy requirement of 5 micrometers and a surface roughness requirement of Ra 0.2 micrometers. Due to the difficulty in machining the blade material, the grinding wheel wears quickly and requires frequent dressing to maintain shape accuracy and cutting performance.

[0241] System Configuration: The laser trimming system is equipped with a fiber laser with a wavelength of 1064 nm, a maximum output power of 500 W, and a beam quality factor of less than 1.5. The multi-sensor monitoring system includes: a laser triangulation displacement sensor with a measurement range of 50 mm, a resolution of 5 μm, and a sampling frequency of 2000 Hz; an industrial camera with a resolution of 2048 pixels by 2048 pixels, a frame rate of 30 frames per second, and a telecentric lens; an acoustic emission sensor with a frequency response range of 100 kHz to 1 MHz and a sampling frequency of 2 MHz; and an infrared temperature sensor with a temperature measurement range of 100°C to 1500°C and a response time of 10 milliseconds. The actuator is a five-axis CNC platform with a travel of at least 500 mm on each axis, a positioning accuracy of 3 μm, and a maximum speed of 10 m / min.

[0242] Dressing task: Dress grinding wheels with significant shape deviations after use, aiming to restore the outer diameter of the grinding wheel to the designed profile, with shape error controlled within 3 micrometers. The offline planned dressing path contains 200 path points, and the estimated dressing time is 15 minutes.

[0243] Data Acquisition and Processing: During the trimming process, each sensor synchronously collects data according to a set frequency. Table 1 shows examples of multi-source monitoring data collected at five typical moments during the trimming process.

[0244] Table 1: Examples of multi-source monitoring data acquisition;

[0245]

[0246] As can be seen from Table 1, as the trimming process progresses, the distance measured by the laser displacement sensor gradually approaches the target value, the boundary sharpness index evaluated by the vision system gradually increases, the acoustic emission signal energy fluctuates within a reasonable range, and the peak temperature shows a trend of first rising and then stabilizing.

[0247] State assessment and compensation decision: Based on the collected monitoring data, the system extracts multimodal features in real time to perform state assessment. The state assessment results and corresponding compensation decisions at five typical time points are shown in Table 2.

[0248] Table 2: Examples of Condition Assessment and Compensation Decisions;

[0249]

[0250] As can be seen from Table 2, the system calculates the corresponding position compensation amount based on the assessed shape deviation; dynamically adjusts the laser power according to the change in heat accumulation level, and reduces the power to control heat accumulation when the temperature is high; as the trimming progresses, the shape deviation gradually decreases, the compensation amount decreases accordingly, and the uniformity score gradually improves.

[0251] Dressing Results: After dressing, the grinding wheel was precisely measured. The measurement results showed that the shape error of the outer diameter of the grinding wheel was 2.3 micrometers, meeting the accuracy requirement of 3 micrometers. The grinding wheel surface showed no obvious burn marks, had a uniform texture, and intact boundaries. The dressed grinding wheel was applied to blade grinding; the grinding process was stable, the workpiece surface quality met the requirements, and the effective service life of the grinding wheel was guaranteed.

[0252] This application example verifies the effectiveness of the method of the present invention in the actual scenario of dressing grinding wheels for grinding aero-engine blades. By integrating multi-source online monitoring information, evaluating the processing status in real time, predicting the status evolution, and calculating adaptive compensation, the system can overcome the influence of multiple factors such as grinding wheel material inhomogeneity, dynamic disturbances during processing, and thermal accumulation effects, achieving high-precision real-time path compensation and ensuring that the shape accuracy and surface quality of the dressed grinding wheel meet the requirements of subsequent high-precision grinding.

[0253] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A real-time compensation method for laser trimming paths integrating online monitoring, characterized in that, Includes the following steps: Data from multi-source heterogeneous sensors were collected for laser dressing of grinding wheels. The laser dressing path was planned to obtain the basic path planning results. Data preprocessing and spatiotemporal alignment were performed to obtain standard multi-source monitoring data. Based on the standard multi-source monitoring dataset, multimodal feature extraction and feature selection are performed to obtain a multimodal key feature set; Based on the multimodal key feature set, multimodal feature fusion is performed, and the current processing status of the grinding wheel laser dressing is evaluated to obtain a comprehensive evaluation result of the current processing status; Based on the comprehensive evaluation results of the current processing status, the processing status of laser dressing of grinding wheels is predicted, and the comprehensive prediction results are output. Based on the comprehensive evaluation and prediction results of the current processing status, the basic path planning results are optimized by multi-objective compensation to obtain the comprehensive path compensation amount. Based on the comprehensive path compensation amount and the basic path planning results, path correction and smoothing are performed to obtain the compensation path instructions. The system executes compensation path instructions, monitors the execution process in real time, evaluates the execution effect, provides adaptive feedback, and obtains complete data records and trimming reports of the processing process.

2. The real-time compensation method for laser trimming path integrating online monitoring according to claim 1, characterized in that, The monitoring data collected by the multi-source heterogeneous sensor for laser dressing of grinding wheels includes: Configure a multi-sensor monitoring system, including a laser displacement sensor, a visual imaging system, an acoustic emission sensor, and an infrared temperature sensor; For the laser displacement sensor, a calibration curve between the distance measurement value and the actual distance is established using standard gauge blocks. For the visual imaging system, a calibration board is used to calibrate the camera intrinsic parameters, and a rotation matrix and translation vector between the camera coordinate system and the machining coordinate system are established through hand-eye calibration. For the acoustic emission sensor, its frequency response characteristics and sensitivity parameters are tested using a known sound source. For the infrared temperature sensor, temperature calibration is performed based on the emissivity characteristics of the material being tested. Thus, the calibration parameters of the multi-sensor monitoring system are obtained. Multi-source heterogeneous sensor monitoring data are collected synchronously according to their respective sampling frequencies and timestamped during data collection, outputting multi-source raw monitoring data.

3. The real-time compensation method for laser trimming path integrating online monitoring according to claim 1, characterized in that, The data preprocessing and spatiotemporal alignment include: The monitoring data from multi-source heterogeneous sensors are filtered, denoised, and preliminarily feature-extracted to output preprocessed monitoring data. Based on the calibration parameters of the multi-sensor monitoring system and the preprocessed monitoring data, the spatiotemporal alignment of the multi-source monitoring data is performed, a unified time grid is established, the timestamps of each sensor are compensated for time delay, the spatial positions measured by different sensors are mapped to a unified processing coordinate system, and a standard multi-source monitoring dataset is generated.

4. The real-time compensation method for laser trimming path integrating online monitoring according to claim 1, characterized in that, The multimodal feature extraction includes: Geometric deviation features are extracted from topographic point cloud data. The point cloud data is registered with the target trimming surface. The normal distance from each measurement point in the point cloud to the target surface is calculated to generate the shape deviation field distribution. The global features of the deviation are extracted, and the spatial distribution features of the deviation are extracted. Surface quality features are extracted from visual image data, texture feature parameters are extracted using the gray-level co-occurrence matrix method, the contour boundary of the grinding wheel is extracted by edge detection to assess boundary integrity, and burn areas are identified by color space analysis. Material removal process features are extracted from acoustic emission signals, the main frequency component and its time domain variation are extracted, the cumulative energy and energy change rate of the signal within each time window are calculated, and acoustic emission events are identified and event features are statistically analyzed by setting amplitude thresholds. Extract heat accumulation effect features from temperature data, extract the standard deviation of spatial statistical features, extract the temporal evolution features of temperature, and extract the features of the heat-affected zone.

5. The real-time compensation method for laser trimming path integrating online monitoring according to claim 1, characterized in that, The multimodal feature fusion includes: Establish a feature fusion rule base based on expert knowledge; A lightweight model is used for feature fusion and state assessment. The state assessment model adopts gradient boosting decision tree or support vector machine to output shape deviation score, uniformity removal score, thermal accumulation level, processing stability score, and surface quality level.

6. The real-time compensation method for laser trimming path integrating online monitoring according to claim 1, characterized in that, The prediction of the processing state of laser dressing of grinding wheels includes: A simplified physical model for material removal is established, and an energy threshold criterion is adopted. When the energy density received per unit area of ​​the laser-acted region exceeds the material removal threshold, the material is removed. The removal depth is proportional to the energy density, and the amount of material removed is predicted. A simplified physical model of heat conduction is established. The surface of the grinding wheel is divided into multiple lumped elements using a lumped parameter model. The heat balance equations of each element are established, and the temperature field evolution is predicted by numerically solving the heat balance equations of each element. A simplified physical model of thermal deformation is established, the thermal strain at each location is calculated based on the temperature field distribution, the stress-strain relationship is established based on the material's elastic modulus and Poisson's ratio to solve for the thermal stress distribution, and the thermal stress and thermal deformation of the grinding wheel are predicted. Using data-driven models to predict state evolution trends; By integrating the prediction results from the physical model and the data-driven model, a comprehensive prediction result is obtained.

7. The real-time compensation method for laser trimming path integrating online monitoring according to claim 1, characterized in that, The multi-objective compensation optimization of the basic path planning results includes: A multi-objective compensation optimization problem is established, with optimization objectives including minimizing shape deviation, maximizing removal uniformity, controlling temperature level, and maintaining execution smoothness. Compensation variables include path point position adjustment, scanning speed adjustment, and laser power adjustment. A hierarchical iterative strategy is employed to solve the collaborative optimization problem; Conduct a feasibility check and implement constraint measures for the compensation plan; Calculate the priority and execution strategy for compensation.

8. The real-time compensation method for laser trimming path integrating online monitoring according to claim 1, characterized in that, The path correction and smoothing process includes: Generate a sequence of compensated path point coordinates. For each path point in the path, add its position adjustment amount to the original coordinates to obtain the new compensated coordinates. The position adjustment is performed along the normal direction of the grinding wheel surface. The compensation path is smoothed, and the B-spline fitting method is used to fit the compensated path points as control points or data points to generate a B-spline curve. Perform dynamic constraint checks and optimizations on the path, and calculate the velocity, acceleration, and jerk of each axis when executing the path; The trimming path is segmented by using segmented trajectory generation and look-ahead buffering.

9. The real-time compensation method for laser trimming path integrating online monitoring according to claim 1, characterized in that, The adaptive feedback includes: Continuously collect and monitor data and evaluate the execution effect, compare the current actual state with the expected state, and calculate the state tracking error; An adaptive feedback control adjustment and compensation strategy is implemented, the feedback control gain is set, and incremental adjustments are made based on the state tracking error. The adjustment methods include incremental adjustment of the compensation amount, fine-tuning of process parameters, and updating of prediction model parameters. Achieve closed-loop data flow and system self-learning, forming a complete closed loop from monitoring data acquisition to execution control; Establish an anomaly detection and graded anomaly handling mechanism. When an anomaly is detected, the graded anomaly handling mechanism is activated, including: Level 1 anomalies are handled by parameter fine-tuning and enhanced monitoring; Level 2 anomalies are handled by immediately reducing the laser power to a safe level; and Level 3 anomalies are handled by immediately suspending processing and activating the safety protection program.

10. A real-time compensation system for laser trimming paths integrating online monitoring, characterized in that, A method for real-time compensation of laser trimming path based on integrated online monitoring as described in any one of claims 1-9 includes: The multi-sensor monitoring module collects monitoring data from multiple heterogeneous sensors during laser dressing of grinding wheels, plans the laser dressing path, and obtains the basic path planning results; it also performs data preprocessing and spatiotemporal alignment to obtain standard multi-source monitoring data. The feature extraction module, based on the standard multi-source monitoring dataset, performs multimodal feature extraction and feature selection to obtain a multimodal key feature set; The status assessment module, based on the multimodal key feature set, performs multimodal feature fusion and assesses the current processing status of the grinding wheel laser dressing, obtaining a comprehensive assessment result of the current processing status; The status prediction module predicts the processing status of the grinding wheel laser dressing based on the comprehensive evaluation results of the current processing status and outputs the comprehensive prediction results. The compensation optimization module performs multi-objective compensation optimization on the basic path planning results based on the comprehensive evaluation results and comprehensive prediction results of the current processing status, and obtains the comprehensive path compensation amount. The path generation module performs path correction and smoothing based on the comprehensive path compensation amount and the basic path planning results to obtain the compensated path instructions. The execution control module executes compensation path instructions, monitors the execution process in real time and evaluates the execution effect, provides adaptive feedback, and obtains complete data records and trimming reports of the processing process.

Images