A vibration error compensation control method for a tempered glass edge grinding process

Abstract

The application discloses a kind of tempered glass edge grinding process vibration error compensation control methods, it is related to the field of automatic control, the method is by arranging sensor in edge grinding equipment side and collecting multi-source heterogeneous data, vibration signal is carried out by edge computing unit fast fourier transform and time domain statistical feature extraction, error estimate value is calculated based on vibration error model, according to error estimate value and preset compensation control strategy generates feed compensation amount, grinding wheel position compensation amount, speed adjustment amount and pressure adjustment amount, and is issued to corresponding servo driver and frequency converter in real time through industrial field bus, realizes the collaborative adjustment of multiple parameters, the edge grinding state after adjustment is monitored in real time and is re-performed feature extraction and error estimation, forms feedback closed loop and continues iteration optimization compensation control amount until vibration error converges.The application effectively reduces the influence of vibration error on edge grinding precision, and the edge collapse rate is significantly reduced, and the edge grinding quality stability and consistency are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of automation control, and in particular to a vibration error compensation control method for the tempered glass edging process. Background Technology

[0002] Tempered glass is widely used in construction, automobiles, home appliances, and displays. Its edging process is prone to vibration errors due to factors such as grinding wheel vibration, motor vibration, and material inhomogeneity, leading to edge chipping, waviness (i.e., excessive waviness), dimensional deviations, and even breakage. Vibration error compensation control methods achieve high-precision grinding through real-time monitoring and closed-loop adjustment. Edge computing, with its low-latency and highly reliable on-site data processing capabilities, can complete error modeling and compensation calculations at the equipment side, avoiding communication delays caused by cloud transmission and meeting the real-time control requirements for high-frequency dynamic response in edging.

[0003] Several solutions have been proposed in the existing technology for vibration suppression in tempered glass edging equipment. For example, one solution involves setting up a buffer mechanism consisting of rubber pads and buffer springs between the support structure and the grinding mechanism of the edging equipment, using the passive deformation of elastic elements to absorb vibration energy. However, this method is a passive buffer, unable to actively detect or feedforward compensate for vibration errors, and the vibration reduction parameters are fixed in the equipment design stage, making it difficult to adapt to the dynamic suppression requirements under different glass specifications or grinding parameters. Furthermore, it lacks real-time monitoring and closed-loop feedback mechanisms, failing to achieve continuous tracking and adaptive compensation of edging accuracy. Another solution involves setting up a fixing device with a foam layer at the top of the reciprocating device of the linear edging machine to provide vibration buffer protection for the grinding wheel. However, this method is still a typical passive vibration reduction, unable to sense real-time vibration status or perform predictive active compensation. Its design is mainly for reciprocating motion vibration and is not directly applicable to dynamic vibration errors generated by the interaction between the grinding wheel and the glass edge. It also lacks modeling and feedback control of key features such as grinding force fluctuations and contour deviations, making it difficult to form a complete error detection and compensation closed loop.

[0004] To address the above problems, this invention proposes a vibration error compensation and control method for the tempered glass edge grinding process. Summary of the Invention

[0005] The purpose of this invention is to provide a vibration error compensation and control method for the tempered glass edging process, to solve the problems in existing technologies where passive buffer structures cannot actively detect and compensate for vibration errors, vibration reduction parameters cannot be dynamically adjusted according to working conditions, and there is a lack of real-time monitoring and closed-loop feedback mechanisms, making it difficult to achieve adaptive control of edging accuracy. To solve the above technical problems, this invention provides the following technical solution:

[0006] A vibration error compensation and control method for the tempered glass edge grinding process includes:

[0007] Step 1: Collect multi-source heterogeneous data during the tempered glass edging process. Vibration sensors, spindle encoders, displacement sensors, and force sensors are deployed at multiple key measuring points on the edging equipment side to collect multi-source heterogeneous data in real time, including grinding wheel vibration acceleration, spindle speed fluctuations, feed displacement deviations, and grinding force changes.

[0008] Step 2: The edge computing unit deployed on the edge grinding equipment side performs feature extraction and error estimation on the multi-source heterogeneous data. Time-domain and frequency-domain features are extracted from the multi-source heterogeneous data to form a multi-dimensional feature vector. Based on a pre-established vibration error model, the error estimate of the current edge grinding state is calculated.

[0009] Step 3: Generate grinding edge compensation control quantities based on the error estimation results. Based on the error estimation values, calculate the feed compensation quantity, grinding wheel position compensation quantity, speed adjustment quantity, and pressure adjustment quantity according to the preset compensation control strategy.

[0010] Step 4: Adjust the feed, speed, or position parameters of the grinding actuator based on the compensation control values. The feed compensation value, grinding wheel position compensation value, speed adjustment value, and pressure adjustment value are transmitted in real time to the corresponding servo drivers and frequency converters via an industrial fieldbus, enabling coordinated adjustment of multiple parameters including the feed motor speed, spindle motor speed, and grinding wheel position.

[0011] Step 5: Update the compensation control quantity based on the adjusted edge grinding state to achieve closed-loop compensation control of vibration error. The adjusted edge grinding state is monitored in real time, and time-domain and frequency-domain features are extracted and error estimated again to form a feedback closed loop. The compensation control quantity is continuously iterated and optimized until the vibration error converges to within the preset threshold range.

[0012] Preferably, in step 1, the vibration sensor is a triaxial accelerometer, the sampling frequency is set to a preset sampling frequency, the measurement range is a preset acceleration measurement range, the spindle encoder resolution is a preset resolution, the displacement sensor is an eddy current sensor, the measurement range is a preset displacement measurement range, the linearity is a preset linearity range, and the force sensor is a piezoelectric force sensor, the range is a preset force value range, and the bandwidth is a preset signal bandwidth.

[0013] Preferably, the acquisition period of the multi-source heterogeneous data in step 1 is a preset acquisition period. During the data acquisition process, the original signal is subjected to low-pass filtering and power frequency notch processing, and the sampled data is directly transmitted to the buffer area of ​​the edge computing unit using direct memory access technology to eliminate white noise and power frequency interference.

[0014] Preferably, in step 2, the edge computing unit adopts a multi-core heterogeneous processor architecture, is equipped with a dedicated digital signal processing core, has a preset operating frequency, and has a preset on-chip static storage capacity. The edge computing unit is deployed in the control cabinet of the grinding equipment and communicates with the data acquisition module and the actuator controller through a high-speed industrial Ethernet.

[0015] Specifically, in step 2, the extraction of time-domain and frequency-domain features includes:

[0016] For the time-domain vibration signal, spindle speed fluctuation, feed displacement deviation, and grinding force change, time-domain statistical features are extracted, including root mean square value, peak factor, and skewness. In addition, kurtosis features are extracted from the time-domain vibration signal to form the time-domain statistical features of the time-domain vibration signal, and frequency domain features are extracted from the time-domain vibration signal by performing a fast Fourier transform.

[0017] The time-domain statistical features and frequency-domain features of the time-domain vibration signal are combined to form a vibration feature vector; the time-domain statistical features of speed fluctuation, displacement deviation, and force change are combined to form a process feature vector.

[0018] The vibration feature vector and the process feature vector are merged into a multidimensional feature vector;

[0019] The time-domain features include root mean square value, peak factor, kurtosis and skewness, and the frequency-domain features include dominant frequency, spectral centroid frequency and spectral energy percentage. The multidimensional feature vector is composed of time-domain features, frequency-domain features and auxiliary features. The auxiliary features include the maximum value, minimum value, peak-to-peak value, waveform factor and impulse factor within the sampling window.

[0020] Preferably, in step 2, the pre-established vibration error model adopts an autoregressive moving average model based on time series analysis. The model parameters are obtained through offline training using historical data. The training samples include vibration signals from a preset number of consecutive sampling points and their corresponding edge grinding error labels. The model order is set to a preset order, and the number of training samples is not less than the preset number of samples.

[0021] Preferably, in step 3, the preset compensation control strategy adopts a dual closed-loop proportional-integral-derivative control structure, with the outer loop being the position error loop and the inner loop being the vibration error loop;

[0022] The calculation of the feed compensation amount and the speed adjustment amount adopts a combination of proportional, integral and derivative elements. The proportional gain is set to a preset proportional gain, the integral gain is set to a preset integral gain, the derivative gain is set to a preset derivative gain, and the control period is a preset control period.

[0023] The grinding wheel position compensation amount is determined based on the frequency domain characteristics of the vibration error. When the vibration amplitude of a specific frequency component exceeds a preset threshold, the grinding wheel is adjusted to move radially away from the glass surface to reduce the grinding contact stiffness, thereby suppressing the vibration of that frequency component. When the amplitude of that frequency component falls back below the preset threshold, the grinding wheel gradually returns to its original radial position.

[0024] The pressure adjustment amount is calculated based on the grinding force error. The calculation method for the pressure adjustment amount is as follows: the difference between the real-time grinding force collected by the force sensor and the set value is taken as the grinding force error. The pressure adjustment amount is calculated according to the preset control law, and the contact pressure between the grinding wheel and the glass is adjusted by controlling the actuator. When the grinding force error is positive (the measured grinding force is higher than the set value), the contact pressure is reduced to reduce the material removal rate. When the grinding force error is negative (the measured grinding force is lower than the set value), the contact pressure is increased to improve the material removal rate, so that the actual grinding force approaches the set value.

[0025] Preferably, the formula for calculating the feed compensation amount in step 3 is:

[0026]

[0027] in, This represents the positional error in the feed direction. , , These are the proportional, integral, and differential coefficients of the feed compensation, respectively.

[0028] The formula for calculating the speed adjustment ΔN is:

[0029]

[0030] in, Main spindle speed fluctuation value , , These are the proportional, integral, and derivative coefficients for speed adjustment, respectively.

[0031] Preferably, , The value range is 0.1 to 5.0. , The value range is 0.01 to 0.5. , The value range is 0.001 to 0.1.

[0032] Preferably, in step 4, the industrial fieldbus adopts the EtherCAT bus protocol, the communication cycle is a preset communication cycle, and the data frame adopts a double buffering mechanism to ensure real-time performance; the feed motor is driven by a permanent magnet synchronous servo motor, and the servo driver converts the feed compensation amount into a motor torque set value for closed-loop control; the spindle motor is driven by an induction motor in conjunction with a vector frequency converter, and the frequency converter converts the speed adjustment amount into a spindle motor speed set value for closed-loop control; the grinding wheel position adjustment is achieved by a hydraulic cylinder controlled by a proportional valve, wherein the rated speed and rated torque of the feed motor are preset rated speeds, the rated torque is preset rated torque, the rated speed and rated power of the spindle motor are preset rated speeds, and the rated power is preset rated power.

[0033] Preferably, in step 5, the vibration error convergence determination condition is that if the absolute value of the error estimate is less than or equal to a preset error threshold for a consecutive preset number of sampling periods, then convergence is determined; otherwise, an abnormal alarm is triggered and the edge grinding operation is suspended.

[0034] Preferably, the vibration error compensation control method for the tempered glass edging process further includes: establishing a historical database for vibration error compensation, recording the process parameters, multi-source heterogeneous data, error estimates, and compensation control quantities for each edging task; based on the historical database, using a long short-term memory recurrent neural network to predict the edging quality; the input of the prediction model is a multi-dimensional feature vector, which includes vibration signal features, edging process parameters, and environmental parameters; the output of the prediction model is the edge breakage probability and the predicted waviness peak-valley value; the predicted waviness peak-valley value is used to evaluate whether the surface waviness peak-valley value after edging meets the preset peak-valley value threshold requirements.

[0035] As a specific embodiment, in step 2, when the grinding accuracy requirement is higher, triaxial accelerometers can be arranged in the axial and radial directions of the grinding wheel, for a total of 6 acceleration sensing channels; the multi-sensor fusion adopts an extended Kalman filter framework, which inputs the vibration data of each sensor channel into the filter after time alignment and spatial registration, and outputs the fused vibration state estimate.

[0036] As a specific embodiment, in step 2, when the grinding accuracy requirement is higher, an online parameter identification mechanism can be added: during the no-load run phase before the grinding task begins, the actuator is driven to generate test vibration and collect response data. Based on the test data, the inherent frequency, damping ratio and stiffness of the equipment are identified and used to update the parameters of the vibration error model in step 2.

[0037] As a specific embodiment, in step 3, when the edge grinding accuracy requirement is higher, an adaptive parameter adjustment mechanism can be added: when a change in glass thickness or grinding depth is detected, the edge calculation unit automatically switches to the corresponding proportional-integral-derivative control parameter configuration file. In step 3, the proportional gain, integral gain, and derivative gain of the proportional-integral-derivative control switch synchronously with the parameter configuration file, and the parameter switching process gradually transitions within 10 control cycles.

[0038] As a specific embodiment, in step 4, when the edge grinding equipment site does not have the conditions for deploying a complete edge computing unit, or when it needs to be coordinated with the upper system for control, a distributed edge computing architecture can be adopted; under this architecture, in step 4, the compensation control quantity is sent from the edge computing unit to the industrial gateway, and the industrial gateway distributes it to the corresponding servo drive and frequency converter through the Modbus TCP protocol.

[0039] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0040] This invention significantly improves the active compensation control capability. By deploying an edge computing unit on the edge grinding equipment, it enables real-time perception and immediate response to vibration status, effectively avoiding the shortcomings of traditional passive vibration reduction methods that cannot actively detect vibration errors. Through the fusion analysis and feature extraction of multi-source sensor data, it can accurately identify the source and evolution trend of vibration errors in the edge grinding process, providing a reliable data foundation for active compensation control.

[0041] This invention significantly improves control real-time performance and accuracy. By using an edge computing architecture, it eliminates the communication links for data uploading to the cloud and remote processing, reducing the control response time from hundreds of milliseconds in traditional solutions to milliseconds, thus meeting the compensation control requirements for high-frequency dynamic response in the edge grinding process. The dual closed-loop proportional-integral-derivative control structure and multi-parameter collaborative adjustment mechanism achieve comprehensive compensation for feed, speed, and position, effectively reducing control overshoot and settling time.

[0042] This invention enhances adaptive compensation capabilities. Through an online update mechanism for the vibration error model and a prediction model based on a historical database, the compensation control strategy can be dynamically adjusted according to different glass specifications and different edge grinding conditions, solving the problems of fixed vibration reduction effect and poor adaptability of traditional solutions. The introduction of a long short-term memory recurrent neural network enables the system to learn the optimal compensation strategy from historical data, and the compensation effect is continuously optimized over time.

[0043] The edge grinding quality of this invention is significantly improved. Through real-time closed-loop compensation control, the impact of vibration error on edge grinding accuracy is effectively reduced, the edge chipping rate is significantly reduced, the peak and valley values ​​of glass edge waviness are controlled within the preset peak and valley value threshold range, and the contour size deviation is controlled within the preset size deviation range. The stability and consistency of edge grinding quality are greatly improved, the product qualification rate is significantly increased, and the stringent requirements of high-end application scenarios for tempered glass edge grinding accuracy are met.

[0044] The system of this invention has a high degree of integration and strong reliability. It integrates functions such as data acquisition, feature extraction, error modeling, compensation calculation and command issuance into one unit through edge computing unit, which reduces system complexity and dependence on external computing resources. The use of EtherCAT bus and double buffering mechanism ensures the real-time and deterministic nature of control command transmission, and the mean time between failures of the system reaches the preset reliability requirements. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall technical solution architecture of the vibration error compensation and control method for the tempered glass edge grinding process proposed in this invention;

[0046] Figure 2 This is a schematic diagram of the core principle framework of vibration error compensation in the dual closed-loop proportional-integral-derivative control structure of this invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0050] This invention provides a vibration error compensation control method for the tempered glass edging process. The core of its technical approach lies in real-time sensing of multi-source heterogeneous data at the edge, generating vibration error models and compensation control commands, and ultimately achieving closed-loop correction through an actuator. Deploying an edge computing unit on the edging equipment side allows data acquisition, feature extraction, error estimation, and compensation control to be completed locally, effectively avoiding the latency issues caused by cloud transmission and ensuring millisecond-level control response time. Figure 1 As shown, the entire technical solution is implemented sequentially according to the following steps:

[0051] Step 1: Collect multi-source heterogeneous data during the tempered glass edging process. Vibration sensors, spindle encoders, displacement sensors, and force sensors are deployed at multiple key measuring points on the edging equipment side to collect multi-source heterogeneous data in real time, including grinding wheel vibration acceleration, spindle speed fluctuations, feed displacement deviations, and grinding force changes.

[0052] The implementation of multi-source heterogeneous data acquisition involves the coordinated configuration of multiple sensor nodes. The vibration sensor employs a triaxial accelerometer with a sampling frequency set to 5120Hz and a measurement range of ±50g, covering the vibration intensity range that may occur during the edge grinding process. The triaxial accelerometer is arranged along three orthogonal directions—axial, radial, and tangential—along the grinding wheel to ensure complete acquisition of vibration signals. The spindle encoder has a resolution of 2500 lines, which, after quadrupling the frequency, achieves an equivalent resolution of 10000 pulses per revolution, used to accurately acquire spindle rotation angle and speed information. The displacement sensor uses an eddy current sensor with a measurement range of ±5mm and a linearity of ±0.5%, capable of detecting minute displacement changes of the grinding wheel relative to the glass edge. The force sensor uses a piezoelectric force sensor with a range of ±500N and a bandwidth of 0 to 5kHz, used for real-time monitoring of the magnitude and trend of grinding force.

[0053] The acquisition period for multi-source heterogeneous data is set to 0.5ms, or 2000 sampling points per second. The data acquisition module incorporates a digital filtering circuit to perform low-pass filtering on the raw signal with a cutoff frequency of 2000Hz to eliminate high-frequency noise interference. Simultaneously, it implements power frequency notch filtering with a notch frequency of 50Hz to suppress the impact of power frequency interference on the vibration signal. The acquisition module employs direct memory access technology to directly transmit the sampled data to the cache area of ​​the edge computing unit, avoiding latency caused by data transfer by the central processing unit. The acquired data frames include fields such as timestamp, sensor type identifier, raw triaxial acceleration values, coded spindle angle values, displacement values, and force values, all in a uniform 32-bit floating-point format.

[0054] Step 2: The edge computing unit deployed on the edge grinding equipment side performs time-domain and frequency-domain feature extraction on the multi-source heterogeneous data to form a multi-dimensional feature vector, and calculates the error estimate of the current edge grinding state based on the pre-established vibration error model.

[0055] The edge computing unit adopts a multi-core heterogeneous processor architecture, equipped with a dedicated digital signal processing core with a main frequency of 1.2GHz and an on-chip static storage capacity of 2MB. Deployed within the control cabinet of the edge grinding equipment, the edge computing unit communicates with the data acquisition module and actuator controller via a high-speed industrial Ethernet. The industrial Ethernet uses a transmission rate of 100Mbps, with end-to-end communication latency controlled within 100 microseconds. Internally, the edge computing unit is divided into four parallel execution units: a data acquisition thread, a feature extraction thread, an error estimation thread, and a compensation control thread, which interact with each other through shared memory. The shared memory uses a circular buffer structure with a buffer capacity of 1024 frames of data; when the data write speed exceeds the read speed, causing the buffer to fill, newly written data overwrites the oldest unread data frame in the buffer. The data acquisition thread obtains raw data from the data acquisition module in a polling manner, the feature extraction thread preprocesses and calculates features from the raw data, the error estimation thread calls the vibration error model to predict errors, and the compensation control thread generates compensation instructions based on the error prediction results.

[0056] The extraction of time-domain and frequency-domain features specifically involves: extracting time-domain statistical features, including root mean square value, peak factor, and skewness, from the time-domain vibration signal, spindle speed fluctuation, feed displacement deviation, and grinding force change; in addition, kurtosis features are extracted from the time-domain vibration signal to form the time-domain statistical features of the time-domain vibration signal, and frequency-domain features are extracted from the time-domain vibration signal by performing a fast Fourier transform.

[0057] The time-domain statistical features and frequency-domain features of the time-domain vibration signal are combined to form a vibration feature vector; the time-domain statistical features of speed fluctuation, displacement deviation, and force change are combined to form a process feature vector.

[0058] The vibration feature vector and the process feature vector are merged into a multidimensional feature vector.

[0059] Among them, the time-domain features include root mean square value, peak factor, kurtosis and skewness, and the frequency-domain features include dominant frequency, spectral centroid frequency and spectral energy ratio. The multidimensional feature vector is composed of time-domain features, frequency-domain features and auxiliary features.

[0060] The root mean square (RMS) value is calculated as the root mean square value of the vibration acceleration signal within the sampling window, reflecting the average energy level of the vibration signal. The crazing factor is the ratio of the peak value to the RMS value, used to detect the impact component in the vibration signal. Kurtosis is the ratio of the fourth central moment to the square of the variance, reflecting the sharpness of the vibration signal's probability density function. Skewness is the ratio of the third central moment to the standard cube, describing the degree of asymmetry in the vibration signal's probability distribution.

[0061] In this embodiment, the specific implementation of the time-domain and frequency-domain feature extraction is as follows:

[0062] (1) Time window: Feature extraction adopts the sliding time window method. The window length is set to 256 sampling points (corresponding to a time length of 256 / 5120≈50ms), the sliding step is set to 128 sampling points, and the overlap between adjacent windows is 50%.

[0063] (2) Independent processing of three channels: The time domain and frequency domain features of the three channels of the triaxial accelerometer are extracted independently for the axial, radial and tangential directions. The features of each channel are calculated separately and then spliced ​​into a unified feature vector. No arithmetic synthesis between channels is performed to preserve the vibration information in each direction.

[0064] (3) Determination of the dominant frequency: After performing a fast Fourier transform (FFT) on the time-domain vibration signal, the spectrum is first smoothed with an amplitude of 3 frequency points to suppress false spectral peaks caused by random noise. Then, the local maxima in the spectrum are searched, and the frequency corresponding to the local maxima with the largest amplitude is taken as the dominant frequency. If the amplitude ratio between the largest amplitude spectral peak and the second largest amplitude spectral peak is less than 1.2, the frequencies corresponding to the two spectral peaks are retained at the same time, and the vibration characteristics of the frequency band are described by the two features of the dominant frequency and the secondary dominant frequency.

[0065] (4) Centroid frequency of the spectrum: Calculated using the following formula:

[0066] in, This represents the total number of frequency points analyzed in the spectrum. For the first Frequency values ​​at each frequency point For the first The amplitude at each frequency point. The centroid frequency of the spectrum reflects the central position of the vibrational energy distribution in the frequency domain.

[0067] (5) Frequency band division of spectrum energy proportion: The spectrum is divided into the following frequency bands according to frequency range:

[0068] Low frequency band: 0Hz~100Hz, corresponding to equipment foundation vibration and low frequency disturbance;

[0069] Mid-frequency band: 100Hz~400Hz, corresponding to the main chatter frequency range during the edge grinding process;

[0070] High frequency band: 400Hz~1000Hz, corresponding to the high frequency vibration components generated by the interaction between the grinding wheel and the glass.

[0071] The ratio of the sum of squares of the spectral amplitudes in each frequency band to the total energy of the entire frequency band (0Hz to 1000Hz) is calculated as the characteristic of the frequency band energy proportion.

[0072] (6) Calculation of auxiliary features: The maximum value, minimum value, peak-to-peak value, waveform factor and impulse factor within the sampling window are all directly calculated from the sampling points of the time-domain vibration signal within the window, without involving additional algorithms.

[0073] The multidimensional feature vector has 12 dimensions, including 4 dimensions of time-domain features, 3 dimensions of frequency-domain features, and 5 dimensions of auxiliary features. The auxiliary features include the maximum value, minimum value, peak-to-peak value, waveform factor, and impulse factor within the sampling window.

[0074] The pre-established vibration error model adopts an autoregressive moving average model based on time series analysis. The model order is set to 12, with the autoregressive term being of order 8 and the moving average term of order 4. These values ​​are used in this embodiment. The mathematical expression for the model parameters is as follows:

[0075]

[0076] in: This is the estimated vibration error value at the current moment; For the front The vibration error estimate at time 1. ; , where is the autoregressive coefficient, representing the weight of the influence of historical vibration error on the current error; For the front The model residuals at time 1. ; is the moving average coefficient, representing the weighting of historical residuals on the current error correction. Wherein, and Obtained through offline identification or online recursive least squares method.

[0077] The model parameters are obtained through offline training using historically collected data, with a minimum of 5000 training samples; in this embodiment, 10000 training samples are used. Each training sample contains vibration signals from 256 consecutive sampling points (as used in this embodiment) and their corresponding edge-grinding error labels. The gradient descent algorithm is used to optimize the model parameters during training. The specific training process of the gradient descent algorithm is as follows:

[0078] (1) Sample splitting: The 10,000 training samples were divided into training set, validation set and test set in a ratio of 8:1:1. The training set was used for updating model parameters, the validation set was used to monitor the degree of overfitting during the training process, and the test set was used for independent evaluation of the final model accuracy.

[0079] (2) Loss function: The mean squared error is used as the loss function, and its expression is:

[0080]

[0081] in, The number of training set samples (in this embodiment) ), For the first The true edge grinding error value of each training sample For the model to the first The predicted value for each sample.

[0082] (3) Optimization algorithm and hyperparameters: The model parameters are optimized by mini-batch gradient descent. The batch size is set to 256 samples, the initial learning rate is set to 0.001, and the maximum number of iterations is set to 500.

[0083] (4) Termination condition: When the loss function value on the validation set does not decrease in 20 consecutive iterations, training is terminated early to prevent the model from overfitting.

[0084] (5) Model Validation: After training terminates, the model accuracy is independently evaluated using a test set (1000 samples). If the mean squared error on the test set is less than a preset threshold... If the model is deemed valid, it can be deployed to an edge computing unit; otherwise, training data should be collected again or the model order should be adjusted and the model retrained.

[0085] (6) Determination of the order of the autoregressive moving average model: as described above , The order is determined using the Akaike Information Criterion (AIC). Different order combinations are calculated separately. From 1 to 12, The AIC values ​​from 1 to 6 are used to select the order group with the smallest AIC value as the model order. In this embodiment, , The AIC value is at its minimum at that time.

[0086] The objective function is to minimize the root mean square value of the prediction error. After training, the model parameters are stored in the read-only storage area of ​​the edge computing unit, and the stored parameters are directly called for calculation during online inference.

[0087] When the system is first started, all historical terms in the recursive formula of the autoregressive moving average model are initially set to 0. After online inference is started, the estimated value and residual at the current moment are updated to the historical terms after each sampling period. After the first 8 sampling periods, the model enters a fully working state.

[0088] The error estimation process is executed once per sampling period. The edge computing unit first obtains the latest multidimensional feature vector from the feature extraction thread and then inputs it into the autoregressive moving average model. The model output is the vibration error estimate for the current moment, including three components: feed direction error, spindle speed error, and grinding force error. The error estimate is then passed to the compensation control thread for subsequent compensation calculation.

[0089] Step 3: Generate grinding compensation control quantities based on the error estimation results. Based on the error estimation values, calculate the feed compensation quantity, grinding wheel position compensation quantity, speed adjustment quantity, and pressure adjustment quantity according to the preset compensation control strategy.

[0090] like Figure 2 As shown, the compensation control strategy adopts a dual-closed-loop proportional-integral-derivative control structure, with the outer loop being the position error loop and the inner loop being the vibration error loop.

[0091] The proportional gain of the dual closed-loop PID controller Integral gain and differential gain The tuning adopts the Ziegler-Nichols critical proportionality method, and the specific steps are as follows:

[0092] (1) Integral gain and differential gain Set all values ​​to 0, retaining only the proportional gain. ;

[0093] (2) From Initially, gradually increase the step size by 0.1. At the same time, a small step signal is applied to the system for excitation;

[0094] (3) Observe the system response. When the output shows a continuous oscillation with constant amplitude, record the critical proportional gain at this time. and critical oscillation period ;

[0095] (4) Calculate the PID parameters according to the Ziegler-Nichols empirical formula:

[0096] Outer loop (position error loop): , , ;

[0097] Inner ring (vibration error ring): , , .

[0098] Using the above tuning method, the outer loop parameters in this embodiment are determined as follows: , , The inner ring parameters are determined as follows: , , .

[0099] In the actual edge grinding process, online fine-tuning within a range of ±20% is allowed based on the initial parameters. The fine-tuning is based on the system response overshoot and adjustment time: if the overshoot exceeds the set value (20% in this embodiment), it should be appropriately increased. and reduce If the adjustment time is too long, increase it appropriately. and .

[0100] The control cycle is set to 1ms, meaning 1000 control updates per second. The outer loop receives the position error signal from the position sensor and outputs it as the setpoint for the inner loop. The inner loop receives the vibration error signal from the vibration sensor, compares it with the outer loop setpoint, and generates the final control input.

[0101] The formula for calculating the feed compensation is:

[0102]

[0103] Where e is the position error in the feed direction. , , These are the proportional, integral, and differential coefficients of the feed compensation, respectively. Preferably, The value range is 0.1 to 5.0. The value range is 0.01 to 0.5. The value range is 0.001 to 0.1. In this embodiment, The value is 3.2. The value is 1.0. The value is set to 0.4. The formula for calculating the speed adjustment is:

[0104]

[0105] Where f is the spindle speed fluctuation value. , , These are the proportional, integral, and derivative coefficients for speed adjustment, respectively. Preferably... The value range is 0.1 to 5.0. The value range is 0.01 to 0.5. The value range is 0.001 to 0.1. In this embodiment, The value is 2.8. The value is 0.9. The value is 0.35.

[0106] Since the edge computing unit is a digital control system, the above continuous domain formula needs to be converted into a discrete domain recursive form for execution. This embodiment uses an incremental PID algorithm with feed compensation amount... For example, its discretization recursive formula is:

[0107]

[0108] in, This is the sequence number of the current sampling time. The sampling period (in this embodiment) ), The position error at the current moment. This represents the position error at the previous moment. This represents the positional error between the two previous moments.

[0109] In the above formula, the integral term Discretize into That is, a weighted summation of the error at the current time; the differential term Discretize into That is, a second-order approximation using backward difference is adopted.

[0110] To prevent integral saturation, this embodiment employs an integral separation strategy: when the absolute value of the error estimate exceeds the integral separation threshold (0.05 mm in this embodiment), the integral term is forcibly separated. Set to zero, and use only proportional and derivative control; when the absolute value of the error estimate is less than or equal to the integral separation threshold, introduce the integral term normally to eliminate the steady-state error.

[0111] Speed ​​adjustment amount The discretization recursive formula is:

[0112]

[0113] in, This represents the current spindle speed fluctuation value. This represents the spindle speed fluctuation value at the previous moment. The values ​​represent the spindle speed fluctuations at the previous two moments.

[0114] The grinding wheel position compensation amount is determined based on the frequency domain characteristics of the vibration error. When a vibration component of a specific frequency (in this embodiment, it is within the 100Hz to 400Hz frequency band, which is a common flutter frequency range during edge grinding; the specific value can be adjusted according to the actual equipment) is detected, and when the vibration amplitude of the frequency component exceeds the preset threshold, the grinding wheel is adjusted to move radially away from the glass surface to reduce the grinding contact stiffness, thereby suppressing the vibration of the frequency component. When the amplitude of the frequency component falls back to below the preset threshold, the grinding wheel gradually returns to its original radial position at a recovery rate of 0.002mm per control cycle (1ms).

[0115] In this embodiment, the spindle frequency of the edging equipment is 150Hz, the preset frequency is its second harmonic 300Hz, and the preset amplitude threshold is 0.5g. Time-domain vibration signals are acquired in real time and subjected to Fast Fourier Transform. When the amplitude of the 300Hz frequency component exceeds 0.5g, it is determined that flutter has occurred. At this time, based on the deviation of the 300Hz amplitude exceeding 0.5g, the grinding wheel is controlled to move radially away from the glass surface. The radial adjustment range is 0.01mm to 0.1mm, and the adjustment step size is proportional to the amplitude deviation. As the 300Hz amplitude falls below 0.5g, the grinding wheel gradually returns to its original radial position. In this embodiment, the response time from detecting the amplitude deviation to completing the radial position adjustment is no more than 50ms.

[0116] The pressure adjustment is calculated based on the grinding force error. When the grinding force deviates from the set value, a constant material removal rate is maintained by adjusting the contact pressure between the grinding wheel and the glass.

[0117] The pressure adjustment is calculated as follows: the difference between the real-time grinding force collected by the force sensor and the set value is taken as the grinding force error. The pressure adjustment is calculated according to the preset control law, and the contact pressure between the grinding wheel and the glass is adjusted by controlling the actuator. When the grinding force error is positive (the measured grinding force is higher than the set value), the contact pressure is reduced to reduce the material removal rate. When the grinding force error is negative (the measured grinding force is lower than the set value), the contact pressure is increased to improve the material removal rate, so that the actual grinding force approaches the set value.

[0118] In this embodiment, the grinding force is set to 120N, and the control law adopts proportional-integral control with a proportional coefficient of 120N. =2.5, dimensionless, integral coefficient =0.5 The response time of the pressure adjustment actuator is no more than 50ms. When the measured grinding force is 130N (deviation +10N), the calculated pressure adjustment is -3N, that is, the contact pressure is reduced by 3N, so that the grinding force returns to near the set value.

[0119] After the compensation control quantities are generated, the edge computing unit packages each compensation quantity into a control command frame. The control command frame contains fields such as a command type identifier, compensation quantity value, timestamp, and checksum. The command type identifier distinguishes between four types of commands: feed compensation, speed adjustment, position compensation, and pressure compensation. The compensation quantity value is stored in 32-bit floating-point format. The timestamp is a microsecond-level time value at the moment the command was generated. The checksum is calculated using a cyclic redundancy check algorithm to detect errors during data transmission.

[0120] Step 4: Adjust the feed, speed, or position parameters of the grinding actuator based on the compensation control values. The feed compensation value, grinding wheel position compensation value, speed adjustment value, and pressure adjustment value are transmitted in real time to the corresponding servo drivers and frequency converters via the industrial fieldbus, enabling coordinated adjustment of multiple parameters including the feed motor speed, spindle motor speed, and grinding wheel position.

[0121] The industrial fieldbus uses the EtherCAT bus protocol with a communication cycle of 500 microseconds. Data frames employ a double-buffering mechanism to ensure real-time performance. The EtherCAT master station is deployed within the edge computing unit and connects to each slave device via an Ethernet port. The data frames use a ring topology; each slave device reads its own control commands as a data frame passes through, and simultaneously writes its status data into the available space within the data frame. The double-buffering mechanism sets up two buffers at the transmitting end: a transmit buffer and a current buffer. After the master station writes the data to be transmitted into the current buffer, it immediately swaps the buffer pointers to ensure continuous data frame transmission. Similarly, at the receiving end, two buffers are set up: a receive buffer and a processing buffer. Received data is first stored in the receive buffer, and then, during idle periods, the buffer pointers are swapped before data parsing.

[0122] The feed motor has a rated speed of 3000 rpm and a rated torque of 10 Nm, driven by a permanent magnet synchronous servo motor. The servo driver receives the feed compensation from the edge computing unit and converts it into a motor torque setpoint. The torque setpoint is achieved through closed-loop control of the current loop, with a bandwidth of 800 Hz and a response time of less than 0.5 milliseconds. The spindle motor has a rated speed of 6000 rpm and a rated power of 15 kW, driven by an induction motor in conjunction with a vector frequency converter. The frequency converter receives the speed adjustment from the edge computing unit and converts it into a spindle motor speed setpoint. The speed setpoint is achieved through closed-loop control of the speed loop, with a bandwidth of 200 Hz and a response time of less than 2 milliseconds. The grinding wheel position adjustment is achieved through a hydraulic cylinder controlled by a proportional valve, with a proportional valve response time of 10 milliseconds and a hydraulic cylinder stroke of ±10 mm.

[0123] Multi-parameter coordinated adjustment follows priority and constraints. Feed compensation has the highest priority; when there is an error in the feed direction, the feed speed is adjusted first to correct the position deviation. Spindle speed adjustment has the second highest priority and is used to suppress spindle vibration and maintain grinding stability. Position compensation and pressure adjustment have lower priority and are used for fine-tuning and optimizing the grinding edge quality. The adjustment range of each parameter must meet the following physical constraints: feed speed adjustment range is ±20% of the set value, spindle speed adjustment range is ±15% of the set value, grinding wheel position adjustment range is ±5 mm, and grinding pressure adjustment range is ±25% of the set value.

[0124] Step 5: Update the compensation control quantity based on the adjusted edge grinding state to achieve closed-loop compensation control of vibration error. Monitor the adjusted edge grinding state in real time, re-extract features and estimate errors to form a feedback loop, and continuously iterate and optimize the compensation control quantity until the vibration error converges to a preset threshold range. After each iteration, perform a convergence judgment; if the convergence condition is met, exit the compensation adjustment state; otherwise, proceed to the next iteration.

[0125] The vibration error convergence determination condition is as follows: if the absolute value of the error estimate is less than or equal to the preset error threshold (0.01 mm in this embodiment) for a preset number of consecutive sampling periods, it is determined to be converged. The current compensation control quantity is locked, the iterative optimization state is exited, and a status prompt indicating that the vibration error has converged is sent to the human-machine interface. If the error estimate of any one of the 50 sampling periods exceeds the preset error threshold, the convergence count is cleared and accumulation starts again. The previously accumulated number of qualifying periods is invalidated and must start from the beginning. Only when all 50 consecutive sampling periods meet the threshold condition is it determined to be converged. If the convergence condition of 50 consecutive periods is not met within the preset timeout period (3 seconds in this embodiment), it is determined to be a convergence failure, triggering an abnormal alarm and pausing the edge grinding operation.

[0126] The abnormal alarm signal is transmitted to the safety controller of the grinding equipment via hardwiring. The safety controller immediately cuts off the drive signals of the spindle motor and feed motor, causing the equipment to stop. Simultaneously, the abnormal alarm information is sent via industrial Ethernet to the human-machine interface display module of the edge computing unit, alerting the operator to a vibration error non-convergence fault.

[0127] After an abnormal shutdown, the system automatically records the grinding process parameters, vibration data, and error estimates at the time of the shutdown. It then marks the glass currently being processed as an inspection item. Operators use a profilometer to inspect the edge dimensions and waviness of the glass, and based on the inspection results, decide whether to continue processing or scrap it. Once the operator has troubleshooted the problem, they must manually confirm the reset via the human-machine interface. The edge calculation unit then re-executes the initialization process, clears the historical error accumulation values, and restarts the compensation control process from steps 1 to 5 before the grinding operation can resume.

[0128] The iteration cycle of the closed-loop compensation control is 1 millisecond. In each iteration, the edge computing unit first collects vibration data of the current grinding state, then performs feature extraction and error estimation, then calculates the compensation control quantity based on the error estimate, and finally sends the compensation control quantity to the actuator. The total delay of the entire closed-loop control link is controlled within 3 milliseconds, meeting the compensation control requirements for high-frequency dynamic response. The closed-loop control adopts an integral separation strategy. When the error estimate is large, i.e., the absolute value of the error estimate exceeds the integral separation threshold of 0.05mm, the integral term is temporarily shut down to avoid integral saturation. When the error estimate enters within the integral separation threshold (0.05mm in this embodiment), the integral term is gradually introduced to eliminate the steady-state error, pushing the error further down to within the convergence threshold of 0.01mm. The rate at which the integral term is introduced is 5% per cycle.

[0129] In a specific application example, suppose the edge grinding task requires grinding the edge of tempered glass with a thickness of 6 mm, a grinding depth of 1 mm, and a grinding speed of 3 meters per minute. The edge computing unit continuously monitors multi-source heterogeneous data during the edge grinding process. When a vibration error of 0.015 mm is detected in the feed direction, the feed compensation is calculated according to a dual closed-loop proportional-integral-derivative control structure. The proportional term contributes 3.2 multiplied by 0.015, which equals 0.048. The integral term accumulates the error by 0.0005 mm per cycle. The derivative term is calculated based on the error change rate. The overall compensation is 0.052 mm, corresponding to a feed speed reduction of approximately 1.7%. The compensation command is sent to the feed servo driver via the EtherCAT bus. The servo driver adjusts the motor torque output, reducing the feed speed to 2.95 meters per minute. In the next sampling cycle, the edge computing unit collects the adjusted vibration data and re-extracts features and estimates the error. If the estimated error decreases to 0.012 mm, the iterative optimization continues according to the above strategy. If the convergence condition of 50 consecutive cycles is not met within the preset timeout period (3 seconds in this embodiment), an abnormal alarm will be triggered and the edge grinding operation will be suspended.

[0130] The vibration error compensation control method for tempered glass edge grinding also includes establishing a historical database for vibration error compensation, recording the process parameters (including feed rate, spindle speed, grinding depth, and glass thickness) for each edge grinding task, multi-source heterogeneous data (including vibration acceleration, spindle speed fluctuation, feed displacement deviation, and grinding force), error estimates, and compensation control quantities.

[0131] Based on the historical database, a Long Short-Term Memory (LSTM) recurrent neural network is used to predict the edge grinding quality. In this embodiment, the LSTM prediction model adopts a two-layer stacked LSTM structure. The first layer contains 64 LSTM units, the second layer contains 32 LSTM units, and then a fully connected layer with an output dimension of 2 is applied. The model is trained using the Adam optimizer with an initial learning rate of 0.001.

[0132] The output of the LSTM prediction model is a 2-dimensional vector, specifically: (1) edge chipping probability, a dimensionless real number ranging from 0 to 1, representing the predicted probability of edge chipping defects occurring at the glass edge under the current grinding conditions. The closer to 1, the higher the risk of edge chipping; the closer to 0, the lower the risk of edge chipping. (2) Waviness peak-valley prediction result, in mm, representing the predicted value of the waviness peak-valley value on the glass edge surface after grinding, used to assess in advance whether the grinding quality meets the preset threshold (the preset threshold in this embodiment is ≤0.02mm). When the predicted edge chipping probability exceeds 0.7, the system automatically issues a warning. After the operator confirms the warning, they can manually reduce the feed speed or adjust the grinding parameters to reduce the risk of edge chipping. If the system does not confirm the warning three times in a row, the edge calculation unit will automatically reduce the feed speed to 90% of the current value.

[0133] The input feature dimension of the prediction model is 64 dimensions, specifically: (1) Vibration signal features 40 dimensions, including the root mean square value of triaxial acceleration, peak factor, kurtosis, skewness, main frequency, spectral centroid frequency, spectral energy ratio and energy ratio of each frequency band; (2) Grinding process parameters 20 dimensions, including feed speed, spindle speed, grinding depth, glass thickness, grinding wheel particle size, coolant flow rate, grinding wheel wear, feed motor current, spindle motor current, average grinding force, peak grinding force, root mean square value of grinding force and statistics of the above parameters; (3) Environmental parameters 4 dimensions, including ambient temperature, humidity, coolant temperature and machine tool foundation vibration amplitude.

[0134] The edge grinding quality label for each record in the historical database is obtained as follows: After edge grinding, the peak and valley values ​​of the glass edge waviness are detected using an optical profilometer; the percentage (%) of edge chipping defects is calculated using a vision inspection system. The training set contains no fewer than 10,000 records and is divided into training, validation, and test sets in an 8:1:1 ratio.

[0135] The model is updated every 24 hours, using an elastic weighted incremental learning method to fine-tune the model parameters. The regularization coefficient λ=0.5, with at least 200 new training samples added daily, and the learning rate decays to 0.0001. Model training and parameter updates are performed on a cloud server. After training, the model parameters are distributed to edge computing units for deployment. A single incremental training session takes approximately 30 minutes and does not affect normal edge grinding production or real-time prediction services.

[0136] Example 2

[0137] In some application scenarios, edge grinding equipment may not have the conditions to deploy edge computing units, or it may require coordinated control with a higher-level system. In such cases, a distributed edge computing architecture can be adopted, distributing some computing tasks to industrial gateways or field control units in the field. In this distributed edge computing architecture, some computing tasks that were originally handled by the edge computing units are now executed by the field control units. These field control units are functionally equivalent to the edge computing units in the main solution, but their deployment methods differ.

[0138] In a distributed edge computing architecture, the industrial gateway is responsible for data aggregation and protocol conversion. The industrial gateway communicates with each sensor via the Modbus TCP protocol, transmitting the collected raw data to the field control unit. The field control unit uses a programmable logic controller platform, equipped with a dedicated motion control module, responsible for performing feature extraction and error estimation calculations. The calculation of compensation control quantities is uniformly completed by the edge computing unit, and the calculation results are distributed to each actuator through the industrial gateway.

[0139] In the distributed architecture, the data acquisition cycle is set to 1 millisecond, and feature extraction uses a sliding window method with a window length of 256 sampling points and a sliding step size of 128 sampling points. A simplified version of the vibration error model is used, with the model order reduced to 6th order to accommodate the limited computing resources of the programmable logic controller. The compensation control strategy remains unchanged, still employing a dual-closed-loop proportional-integral-derivative control structure. The control response time is extended to approximately 5 milliseconds due to the added protocol conversion overhead of the industrial gateway, but it still meets the compensation control requirements for most edge-grinding operations.

[0140] The industrial gateway is also responsible for uploading critical data to a cloud server for long-term storage and analysis. The cloud server runs machine learning algorithms, periodically updates the vibration error model parameters, and then distributes the updated model parameters to the field control units. This architecture maintains local real-time control capabilities while utilizing cloud resources for model training and optimization, achieving synergy between edge computing and cloud computing.

[0141] Example 3

[0142] In applications requiring extremely high edge grinding precision, multi-sensor fusion and adaptive parameter adjustment mechanisms can be added to the basic technical solution to further improve the accuracy of compensation control.

[0143] The multi-sensor fusion module integrates data from multiple similar sensors for joint estimation. For example, triaxial accelerometers are arranged along the axial and radial directions of the grinding wheel, resulting in a total of six acceleration sensing channels. The multi-sensor fusion algorithm employs an extended Kalman filter framework, aligning and spatially registering the vibration data from each sensor channel before inputting it into the filter for state estimation. The filter output is the fused vibration state estimate, which offers higher estimation accuracy and stronger anti-interference capability compared to single-sensor data.

[0144] The adaptive parameter adjustment mechanism dynamically adjusts the control parameters according to changes in the grinding conditions. Specifically, when a significant change in glass thickness or grinding depth is detected, the edge computing unit automatically switches to the corresponding parameter configuration file. The parameter configuration files are pre-stored in the edge computing unit's storage medium, with each condition corresponding to a set of optimized proportional-integral-derivative (PID) control parameters. The parameter switching process employs a smooth transition strategy, gradually adjusting parameter values ​​over 10 control cycles to avoid the impact of sudden parameter changes on the control system.

[0145] The adaptive parameter adjustment mechanism also supports online parameter identification. During the idle run phase before each edge grinding task, the edge computing unit drives the actuator to generate a series of small-amplitude test vibrations, while simultaneously collecting vibration response data. Based on the test data, the system automatically identifies the dynamic characteristic parameters of the edge grinding equipment, including natural frequency, damping ratio, and stiffness. The identification results are used to update the parameters of the vibration error model, enabling the model to accurately reflect the actual characteristics of the current equipment.

[0146] Comparative Example 1: Passive Buffer Vibration Damping

[0147] A passive damping method was adopted to suppress edge grinding vibration. This involves placing rubber pads and buffer springs between the support structure and the grinding mechanism of the edge grinding equipment, utilizing the passive deformation of the elastic elements to absorb vibration energy. However, in this method, the vibration reduction parameters are fixed during the equipment design phase and cannot be dynamically adjusted according to changes in grinding conditions. Furthermore, it lacks real-time vibration sensing and a closed-loop feedback mechanism, making it impossible to actively detect and compensate for vibration errors. Experimental results show that under the same grinding conditions (glass thickness 6mm, grinding depth 1mm, feed rate 3m / min), the edge chipping rate of this comparative example is 2.8%, and the peak-to-valley value of edge waviness is 0.12mm, which fails to meet the requirements for high-precision edge grinding.

[0148] Comparative Example 2: Single-parameter adjustment control

[0149] This method only adjusts the feed rate to suppress vibration errors, without coordinating the adjustment of spindle speed, grinding wheel position, and grinding pressure. In other words, upon detecting vibration errors, it only reduces the grinding force by decreasing the feed rate, while keeping other process parameters unchanged. This approach does not consider the coupled effects of parameters such as spindle speed, grinding wheel position, and pressure, resulting in slow convergence (approximately three times the adjustment time of this invention) and ineffective suppression under high-frequency chatter conditions (300Hz component amplitude exceeding 0.5g). The final edge chipping rate is 1.5%, significantly higher than the 0.3% of this invention.

[0150] Comparative Example 3: Open-loop one-time compensation

[0151] An open-loop compensation control method is adopted, meaning that after detecting vibration error, the parameters are adjusted once according to the pre-calibrated compensation amount. After adjustment, the edge grinding state is no longer monitored and iteratively optimized in real time, nor is convergence judgment performed. Because the vibration error in the edge grinding process is time-varying and random, one-time open-loop compensation cannot adapt to the dynamic changes of the error. After compensation, there is still a large residual error, with an average residual error of 0.025mm, which is significantly higher than the 0.008mm after closed-loop iterative convergence of the present invention. Moreover, the product quality consistency is poor. The standard deviation of the peak and valley values ​​of edge waviness of the same batch of glass (a statistical measure of product consistency, the smaller the value, the better the consistency) is 4.2 times that of the present invention.

[0152] Comparative Example 4: Cloud Computing Solution

[0153] The entire process of vibration data acquisition, feature extraction, and compensation calculation is uploaded to a cloud server for execution. After the cloud calculation is completed, the compensation command is sent back to the grinding equipment. In this scheme, the cumulative communication delay of the three stages—data upload, cloud processing, and command issuance—is approximately 150ms to 300ms. Because the vibration error changes frequently during the grinding process (the dominant frequency can reach 300Hz), the delay time of the cloud calculation scheme far exceeds the effective compensation window for vibration error (approximately 3ms). This causes the compensation command to lag behind the changes in vibration state, making real-time compensation impossible. In severe cases, the command lag can even trigger control oscillations, resulting in grinding quality that is worse than in the uncompensated state.

[0154] The specific data for Example 1 and the comparative example above are shown in Table 1.

[0155] Table 1. Comparison and analysis of the effects of the comparative example and Example 1

[0156] Example 1 This invention 0.3% 0.015mm 0.008mm 5ms ≤3ms Comparative Example 1 Passive buffering and vibration reduction 2.8% 0.12mm — — — Comparative Example 2 Single parameter adjustment control 1.5% — — 18ms — Comparative Example 3 Open-loop one-time compensation — 0.008mm (standard deviation) 0.025mm — — Comparative Example 4 cloud computing solutions — — — — 150~300ms

[0157] The comparison table shows that:

[0158] Regarding edge chipping rate, the edge chipping rate of Example 1 was 0.3%, compared to 2.8% in Comparative Example 1 and 1.5% in Comparative Example 2. Example 1 showed a reduction of 89.3% compared to Comparative Example 1 and 80% compared to Comparative Example 2. This indicates that the present invention, through multi-source data fusion sensing and multi-parameter collaborative adjustment, can accurately suppress vibration errors and significantly reduce the occurrence rate of edge chipping.

[0159] Regarding edge grinding quality and consistency, the waviness peak-to-valley value of Example 1 was 0.015 mm, while that of Comparative Example 1 was 0.12 mm, a reduction of 87.5%. The residual error of Example 1 was 0.008 mm, while that of Comparative Example 3 was 0.025 mm, a reduction of 68%. The standard deviation of waviness in Example 1 was one-quarter of that in Comparative Example 3. This indicates that the present invention, through closed-loop iterative optimization, effectively reduced residual error and improved product quality consistency.

[0160] Regarding real-time response, the adjustment time in Example 1 is 5ms, compared to 18ms in Comparative Example 2, representing a 72.2% reduction. The end-to-end delay in Example 1 is ≤3ms, compared to 150ms~300ms in Comparative Example 4, a reduction of over 98%. This demonstrates that the present invention, through edge computing localization processing, achieves real-time synchronization between compensation commands and vibration states, meeting the control requirements for high-frequency dynamic response.

[0161] In summary, Example 1 significantly outperforms the comparative examples in all key indicators such as edge chipping rate, edge grinding quality and consistency, and real-time response, fully demonstrating the effectiveness and superiority of the vibration error compensation control method for tempered glass edge grinding process proposed in this invention.

[0162] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A vibration error compensation and control method for the tempered glass edge grinding process, characterized in that, Includes the following steps: Step 1: Vibration sensors, spindle encoders, displacement sensors and force sensors are arranged at multiple key measuring points on the side of the grinding equipment to collect multi-source heterogeneous data in real time, including time-domain vibration signals, spindle speed fluctuations, feed displacement deviations and grinding force changes. Step 2: The edge computing unit deployed on the edge grinding equipment side performs time domain and frequency domain feature extraction on the multi-source heterogeneous data to form a multi-dimensional feature vector, and calculates the error estimate of the current edge grinding state based on the pre-established vibration error model; Step 3: Based on the error estimate, calculate the feed compensation, grinding wheel position compensation, speed adjustment, and pressure adjustment according to the preset compensation control strategy; Step 4: The feed compensation amount, grinding wheel position compensation amount, speed adjustment amount and pressure adjustment amount are sent to the corresponding servo driver and frequency converter in real time through the industrial fieldbus to realize the multi-parameter coordinated adjustment of feed motor speed, spindle motor speed and grinding wheel position; Step 5: Monitor the adjusted edge grinding state in real time, re-extract time and frequency domain features and estimate errors to form a feedback closed loop, and continuously iterate and optimize the compensation control quantity until the vibration error converges to the preset threshold range.

2. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, In step 1, the vibration sensor is a triaxial accelerometer, which is arranged along the three orthogonal directions of the grinding wheel: axial, radial, and tangential. The spindle encoder achieves a preset equivalent resolution after being processed by a quadruple frequency. The displacement sensor is an eddy current sensor. The force sensor is a piezoelectric force sensor.

3. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, In step 1, the acquisition period for multi-source heterogeneous data is a preset acquisition period. During the data acquisition process, the original signal is subjected to low-pass filtering and power frequency notch filtering, and the sampled data is directly transmitted to the cache area of ​​the edge computing unit using direct memory access technology.

4. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, In step 2, the edge computing unit is deployed in the control cabinet of the grinding equipment and communicates with the data acquisition terminal and the actuator control terminal through a high-speed industrial Ethernet. The edge computing unit sequentially performs four processing stages: data acquisition, feature extraction, error estimation, and compensation control. Information is transmitted between each stage through data caching.

5. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, In step 2, the extraction of time-domain and frequency-domain features specifically involves: For the time-domain vibration signal, spindle speed fluctuation, feed displacement deviation, and grinding force change, time-domain statistical features are extracted, including root mean square value, peak factor, and skewness. In addition, kurtosis features are extracted from the time-domain vibration signal to form the time-domain statistical features of the time-domain vibration signal, and frequency domain features are extracted from the time-domain vibration signal by performing a fast Fourier transform. The time-domain statistical features and frequency-domain features of the time-domain vibration signal are combined to form a vibration feature vector; the time-domain statistical features of speed fluctuation, displacement deviation, and force change are combined to form a process feature vector. The vibration feature vector and the process feature vector are merged into a multidimensional feature vector; The time-domain features include root mean square value, peak factor, kurtosis and skewness, and the frequency-domain features include dominant frequency, spectral centroid frequency and spectral energy percentage. The multidimensional feature vector is composed of time-domain features, frequency-domain features and auxiliary features. The auxiliary features include the maximum value, minimum value, peak-to-peak value, waveform factor and impulse factor within the sampling window.

6. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, In step 2, the pre-established vibration error model adopts an autoregressive moving average model based on time series analysis. The model parameters are obtained through offline training using historical data. The training samples include vibration signals from a predetermined number of consecutive sampling points and corresponding edge grinding error labels.

7. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, In step 3, the preset compensation control strategy adopts a dual closed-loop proportional-integral-derivative control structure, with the outer loop being the position error loop and the inner loop being the vibration error loop; the calculation of the feed compensation amount and the speed adjustment amount adopts a combination of proportional, integral and derivative elements; the grinding wheel position compensation amount is determined based on the frequency domain characteristics of the vibration error, and the pressure adjustment amount is calculated based on the grinding force error.

8. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, In step 4, the industrial fieldbus adopts the EtherCAT bus protocol, and the data frame uses a double buffering mechanism to ensure real-time performance; the feed motor is driven by a permanent magnet synchronous servo motor, and the servo driver converts the feed compensation amount into the motor torque set value for closed-loop control; the spindle motor is driven by an induction motor in conjunction with a vector frequency converter, and the frequency converter converts the speed adjustment amount into the spindle motor speed set value for closed-loop control; the grinding wheel position adjustment is achieved by a hydraulic cylinder controlled by a proportional valve.

9. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, The vibration error convergence determination condition in step 5 is as follows: if the absolute value of the error estimate is less than or equal to the preset error threshold for a consecutive preset number of sampling periods, then it is determined to be converged; if the error estimate of any one of the preset number of sampling periods exceeds the preset error threshold, then the convergence count is cleared to zero and re-accumulated. If the convergence condition for a preset number of consecutive cycles is not met within the preset timeout period, an abnormal alarm will be triggered and the edge grinding operation will be suspended.

10. The vibration error compensation and control method for the tempered glass edge grinding process according to claim 1, characterized in that, Also includes: A historical database for vibration error compensation is established, recording the process parameters, multi-source heterogeneous data, error estimates, and compensation control variables for each edge grinding task. Based on this historical database, a long short-term memory recurrent neural network is used to predict the edge grinding quality. The input to the prediction model is a multi-dimensional feature vector, which includes vibration signal features, edge grinding process parameters, and environmental parameters. The output of the prediction model is the edge chipping probability and the predicted waviness peak-valley value. The predicted waviness peak-valley value is used to evaluate whether the surface waviness peak-valley value after edge grinding meets the preset peak-valley value threshold requirements.

Images