Machine learning based intelligent compensation method and system for machine tool hydraulic control

Abstract

This invention relates to the field of hydraulic control compensation, and more particularly to a machine tool hydraulic control intelligent compensation method and system based on machine learning. The method includes: fitting a global wear decay rate to obtain a coefficient of determination; calculating the wear equivalent time of the grinding process to be compensated, and constructing a process feature vector for the grinding process to be compensated; training a regression prediction model based on the process feature vectors of historical grinding processes, inputting the process feature vector of the grinding process to be compensated into the trained regression prediction model to output a predicted value of the removal rate of the grinding process to be compensated; correcting the predicted value of the removal rate based on the coefficient of determination to obtain a corrected predicted rate of the grinding process to be compensated; calculating the hydraulic pressure compensation amount of the grinding process to be compensated; and generating control commands based on the hydraulic pressure compensation amount to complete the intelligent hydraulic compensation. The technical solution of this invention can improve the accuracy of hydraulic compensation results.

Description

Technical Field

[0001] This invention relates to the field of hydraulic control compensation. In particular, it relates to a machine tool hydraulic control intelligent compensation method and system based on machine learning. Background Technology

[0002] In precision machining equipment, the hydraulic system, as a key actuator in regulating the contact pressure between the grinding disc and the workpiece, directly determines the removal rate and the consistency of the final thickness. However, as the cumulative usage time of the grinding disc increases, abrasive passivation and bond wear cause nonlinear drift in the process state. Relying solely on fixed process parameter settings makes it difficult to maintain the stability of the machining process. Existing machine learning methods, which directly use cumulative time as feature input, are often based on the assumption of stationarity and cannot distinguish between the rapid degradation in the early stages of wear and the non-stationary characteristics that tend to level off later, thus producing systematic prediction biases.

[0003] Existing technologies typically use cumulative grinding time as a feature input to the prediction model, or employ a purely data-driven approach to establish a mapping relationship between process parameters and removal rates. These methods are largely based on the assumption of stationarity, assuming that the amount of process state change caused by the same time interval is constant. However, in double-end grinding processes, the wear of the grinding disc exhibits significant non-stationary degradation characteristics. Influenced by the coupling effect of abrasive passivation and binder wear, the initial wear rate is relatively fast, while the rate tends to level off later. Directly inputting the original time into the model will cause the model to be unable to distinguish the physical differences between different wear stages, incorrectly treating the initial rapid degradation and the later slow wear as equivalent states, thus generating systematic prediction bias and leading to inaccurate hydraulic compensation results. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention provides solutions in the following aspects.

[0005] In a first aspect, the present invention provides a machine tool hydraulic control intelligent compensation method based on machine learning, comprising: acquiring the hydraulic pressure, cumulative grinding time, and initial workpiece thickness of any historical grinding process; fitting a global wear attenuation rate based on a historical database, and simultaneously obtaining the determination coefficients in the fitting process; taking the previous historical grinding process adjacent to the grinding process to be compensated as a reference process; calculating the wear equivalent time of the grinding process to be compensated based on the global wear attenuation rate and the cumulative grinding time of the reference process; and constructing the wear equivalent time, the hydraulic pressure of the grinding process to be compensated, and the initial workpiece thickness of the grinding process to be compensated as the working time of the grinding process to be compensated. Similarly, the process feature vector for any historical grinding process is obtained; a regression prediction model is trained based on the process feature vector of the historical grinding process; the process feature vector of the grinding process to be compensated is input into the trained regression prediction model, and the predicted value of the removal rate of the grinding process to be compensated is output; the predicted value of the removal rate is corrected based on the coefficient of determination to obtain the corrected predicted rate of the grinding process to be compensated; the hydraulic pressure compensation amount of the grinding process to be compensated is calculated based on the corrected predicted rate, the hydraulic pressure of the grinding process to be compensated, and the obtained standard removal rate; control commands are generated according to the hydraulic pressure compensation amount to complete the intelligent hydraulic compensation.

[0006] Preferably, the step of fitting the global wear attenuation rate based on the historical database and obtaining the coefficient of determination in the fitting process includes: taking several historical grinding processes completed under standard pressure in the historical database as experimental processes, obtaining data pairs for any experimental process, wherein the data pairs include the true values ​​of the cumulative grinding time and removal rate; using a regression fitting algorithm on all data pairs of experimental processes to obtain the global wear attenuation rate; and recording the coefficient of determination of the fitting curve in the fitting process.

[0007] Preferably, the calculation of the wear equivalent time of the grinding process to be compensated includes: calculating the product of the global wear decay rate and the cumulative grinding time of the reference process as a first product, calculating the negative exponent value of the first product, and taking the difference between the constant 1 and the negative exponent value as the wear equivalent time of the grinding process to be compensated.

[0008] Preferably, the regression prediction model adopts a Gaussian process regression model, a neural network, a support vector machine regression model, or a random forest regression model.

[0009] Preferably, correcting the predicted value of the removal rate based on the coefficient of determination to obtain the corrected predicted rate of the grinding process to be compensated includes: obtaining the predicted value of the removal rate of the reference process and the corrected predicted rate; taking the difference between the predicted value of the removal rate of the reference process and the corrected predicted rate as a first difference; taking the ratio of the first difference to the predicted value of the removal rate of the reference process as the prediction error residual of the reference process; taking the difference between the constant 1 and the coefficient of determination as a second difference; taking the product of the second difference and the prediction error residual as a second product; taking the sum of the constant 1 and the second product as a correction factor; and taking the product of the correction factor and the predicted value of the removal rate of the grinding process to be compensated as the corrected predicted rate of the grinding process to be compensated.

[0010] Preferably, the calculation of the hydraulic pressure compensation amount for the grinding process to be compensated includes: obtaining a standard removal rate, taking the difference between the standard removal rate and the corrected predicted rate of the grinding process to be compensated as a third difference, and taking the ratio of the third difference to the corrected predicted rate of the grinding process to be compensated as a rate ratio; and taking the product of the hydraulic pressure of the grinding process to be compensated and the rate ratio as the hydraulic pressure compensation amount for the grinding process to be compensated.

[0011] Preferably, the step of generating control commands based on the hydraulic pressure compensation amount to complete hydraulic intelligent compensation includes: obtaining the minimum and maximum working pressures allowed in the hydraulic system; and taking the sum of the hydraulic pressure during the grinding process to be compensated and the hydraulic pressure compensation amount as the comprehensive pressure.

[0012] The combined pressure is limited between the minimum and maximum working pressure by a limiting function, thereby generating control commands.

[0013] Preferably, the method further includes: measuring the final thickness of the workpiece after the compensation grinding process is completed using a grating thickness gauge, calculating the true value of the removal rate of the compensation grinding process, and updating the true value of the removal rate of the compensation grinding process to the historical database.

[0014] Secondly, the present invention also provides a machine tool hydraulic control intelligent compensation system based on machine learning, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned machine tool hydraulic control intelligent compensation method based on machine learning is implemented.

[0015] The present invention has the following effects:

[0016] This invention maps the non-stationary, nonlinear wear degradation process of the grinding disc into a stationary equivalent time domain by constructing an equivalent wear time. This effectively solves the systematic prediction bias caused by the rapid degradation in the early stage of wear and the gradual degradation in the later stage, significantly improving the accuracy of removal rate prediction. Secondly, this invention innovatively introduces a dynamic correction mechanism based on the coefficient of determination. It adaptively adjusts the error feedback weights using the model's goodness of fit, reducing interference when the model is reliable and enhancing compensation when the model is uncertain. This significantly improves the robustness and adaptability of the hydraulic system at different wear stages.

[0017] This invention combines the advantages of physical attenuation models and machine learning regression models, utilizing both the interpretability of physical laws and the nonlinear fitting capabilities of data-driven approaches to achieve a precise mapping between process parameters, wear state, and removal rate. Through a real-time feedback update mechanism and pressure limiting protection, the safety and continuity of the processing are ensured, effectively maintaining the consistency of workpiece thickness and improving the accuracy of hydraulic compensation and production efficiency. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent compensation method for machine tool hydraulic control based on machine learning, according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0020] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] Reference Figure 1 The machine tool hydraulic control intelligent compensation method based on machine learning includes steps S1-S4, as follows:

[0022] S1: Obtain the hydraulic pressure, cumulative grinding time, and initial workpiece thickness for any historical grinding process, fit the global wear attenuation rate based on the historical database, and obtain the determination coefficients in the fitting process.

[0023] In one embodiment, key process parameters are acquired synchronously through a multi-source sensor fusion strategy. Specifically: a pressure sensor integrated into the hydraulic closed-loop control loop collects hydraulic pressure, which is the normal pressure actually applied to the workpiece surface by the upper grinding disc, directly characterizing the contact stress level during the grinding process; a high-precision timer in the equipment control system collects the cumulative grinding time, which is the cumulative running time from the moment the grinding disc is replaced or repaired until the end of any historical grinding process, serving as a core feature characterizing the wear evolution state of the grinding disc; a high-resolution laser displacement sensor installed at the loading station collects the initial thickness of the workpiece in a single instance triggered by workpiece loading, which is the measured thickness value of the workpiece before entering the grinding process, used to characterize individual differences in the workpiece.

[0024] Several historical grinding processes completed under standard pressure were selected as experimental procedures to eliminate the interference of pressure compensation on the removal rate observation in historical processing and to ensure that the data truly reflects the natural decay trend of the grinding disc's cutting ability.

[0025] For any experimental process, data pairs containing the true values ​​of cumulative grinding time and removal rate are extracted. The cumulative grinding time is defined as the total running time from the replacement of the grinding disc to the end of the experimental process, and the true value of the removal rate is calculated based on the actual thickness measured by the grating thickness gauge. Subsequently, a regression fitting algorithm is used to fit the trend of the data pairs of all experimental processes. The exponential decay model is selected to characterize the non-stationary degradation law of abrasive passivation and bond wear, thereby solving for the global wear decay rate that characterizes the rate of decay of cutting capability per unit time. At the same time, the coefficient of determination of the fitting curve is recorded during the fitting process. The coefficient of determination quantifies the degree of fit between the physical model and historical data.

[0026] For example, in the process of fitting the global wear decay rate based on historical database in this invention, the regression fitting algorithm preferably adopts the linearized least squares method to achieve efficient parameter calculation. Specifically, for the exponential decay model that characterizes the nonlinear decay of the grinding disc's cutting ability over time, the true value of the removal rate obtained during the experiment under standard pressure is first logarithmically transformed to convert the originally complex exponential relationship into a linear relationship, thereby constructing a linear mapping between the cumulative grinding time and the transformed removal rate.

[0027] Subsequently, the transformed data is fitted with a straight line using the least squares method. The slope of the fitted line directly corresponds to the global wear decay rate. The coefficients of determination during the fitting process are recorded to quantify the model's reliability. Compared to direct nonlinear iterative optimization methods, this linearization approach significantly reduces computational complexity, allowing for rapid solution of key parameters without a cumbersome iterative convergence process. This makes it particularly suitable for resource-constrained embedded controllers, ensuring efficient real-time solutions and guaranteeing the timely generation of hydraulic pressure compensation commands to meet real-time control requirements in continuous processing scenarios.

[0028] In another embodiment, the regression fitting algorithm employs a nonlinear least squares method to further improve parameter identification accuracy. Specifically, based directly on the true values ​​of cumulative grinding time and removal rate obtained during the experiment, an error sum-of-squares function is constructed to characterize the deviation between the model's predicted values ​​and the measured values. This aims to quantify the overall fitting error between the exponential decay model curve and all historical data points.

[0029] Subsequently, through iterative optimization algorithms (exemplarily, the Gauss-Newton method), the global wear decay rate is directly calculated by continuously seeking the optimal value within the parameter space, thus minimizing the sum of squared errors. Compared to linearization methods, this nonlinear fitting strategy eliminates the need for logarithmic transformation of the removal rate data, thereby avoiding distribution distortion and uneven error weighting caused by data transformation. It more realistically reflects the nonlinear physical laws of grinding disc wear, significantly improving the fitting accuracy and reliability of the global wear decay rate, and providing a more accurate physical parameter basis for subsequent mapping calculations of wear equivalent time.

[0030] In another embodiment, the regression fitting algorithm employs gradient descent to improve the generality of parameter solving and the model's scalability. Specifically, an initial estimate of the global wear decay rate is set. Then, a loss function characterizing the deviation between the model's predicted values ​​and historical measured data is constructed, and the gradient direction of this loss function with respect to the global wear decay rate is calculated. Next, the value of the global wear decay rate is iteratively updated along the opposite direction of the gradient. By continuously decreasing the loss function value, the parameter solving process is driven until convergence, thereby obtaining the optimal global wear decay rate.

[0031] Compared to specific analytical solutions, iterative optimization methods are more versatile and do not depend on the specific mathematical form of the model. This makes it easier to extend the wear model to more complex nonlinear structures or multi-parameter coupled scenarios, providing flexible algorithmic support for the accurate identification of process conditions. It ensures that a reliable global wear decay rate can be stably obtained under different wear model complexities, thereby providing a more accurate physical parameter basis for the subsequent mapping calculation of wear equivalent time.

[0032] S2: Take the previous historical grinding process adjacent to the grinding process to be compensated as the reference process. Calculate the wear equivalent time of the grinding process to be compensated based on the global wear decay rate and the cumulative grinding time of the reference process. Construct the process feature vector of the grinding process to be compensated by the wear equivalent time, the hydraulic pressure of the grinding process to be compensated, and the initial thickness of the workpiece in the grinding process to be compensated. Similarly, obtain the process feature vector of any historical grinding process.

[0033] It should be noted that during continuous processing in a double-end grinding machine, the regression prediction model requires a stable correspondence between the time difference and the change in removal rate in the input feature space. However, physically, the wear of the grinding disc exhibits significant non-stationary degradation characteristics. In the initial stage, abrasive passivation is rapid, and the decrease in removal rate per unit cumulative grinding time is substantial. In the later stage, wear tends to level off, and the change in removal rate caused by the same time increment is weak. If the original cumulative grinding time is directly used as the feature input model, the model will be unable to distinguish the physical differences between different wear stages, incorrectly treating the rapid degradation in the initial stage and the slow wear in the later stage as equivalent states, thus producing a systematic prediction bias. To resolve this structural contradiction, this invention constructs a wear equivalent time feature that can characterize the cumulative wear state. By using the fitted global wear decay rate and the current cumulative grinding time for nonlinear transformation, the non-stationary actual time domain is mapped to a stationary equivalent time domain. This mapping effectively balances the impact of rapid initial wear and slow later wear on the model by nonlinearly compressing the later time weights. This ensures that the wear effect corresponding to any identical time difference within the equivalent time space remains essentially constant. Thus, the physical laws of wear are embedded into the model input features, making the input space satisfy the stable change assumption of the regression prediction model. This provides a feature basis that conforms to mathematical assumptions for accurate prediction of the removal rate.

[0034] In one embodiment, in order to accurately characterize the wear state of the grinding process to be compensated, the previous historical grinding process adjacent to the grinding process to be compensated is selected as a reference process, and the cumulative grinding time at the end of the reference process is used as the benchmark for the current wear progress.

[0035] When specifically calculating the wear equivalent time of the grinding process to be compensated, the product of the global wear decay rate and the cumulative grinding time of the reference process is first calculated as the first product, which represents the physical dimension of the cumulative wear effect. Then, the negative exponent value of the first product is calculated, and the nonlinear law of the wear rate decreasing with time is simulated by the exponential decay characteristic. Finally, the difference between the constant 1 and the negative exponent value is taken as the wear equivalent time of the grinding process to be compensated.

[0036] The calculation of the wear equivalent time essentially involves constructing a nonlinear mapping function to transform the non-stationary actual cumulative time domain into a stationary equivalent time domain. This ensures that the equivalent time value converges to the interval between 0 and 1, where 0 represents the initial state of a new grinding disc and 1 represents the wear saturation state. This mapping effectively balances the weighted impact of rapid initial wear and slow subsequent wear on the model input, ensuring that the wear effect corresponding to any identical time difference within this equivalent time space remains essentially constant. This satisfies the regression prediction model's assumption of stable change in the input feature space, providing a physically sound feature basis for accurate prediction of the subsequent removal rate, and thus ensuring the accuracy of the hydraulic pressure compensation calculation.

[0037] To ensure the consistency and standardization of the input feature space of the regression prediction model, the wear equivalent time, the hydraulic pressure of the grinding process to be compensated, and the initial thickness of the workpiece in the grinding process to be compensated are constructed as the process feature vector of the grinding process to be compensated. Similarly, the process feature vector of any historical grinding process is obtained.

[0038] Specifically, for the grinding process to be compensated, the wear equivalent time after nonlinear mapping transformation is used to replace the original cumulative grinding time to eliminate the influence of wear non-stationarity. This time, the input is combined with the real-time collected hydraulic set pressure and workpiece initial thickness data in a fixed order to form a standardized input. For historical grinding processes used for model training, the same data structure and processing logic are used to convert the historical cumulative time into historical wear equivalent time. The historical hydraulic pressure and historical workpiece initial thickness are then combined to construct the process feature vector of the historical grinding process, ensuring that the training samples and the prediction input are in the same feature space.

[0039] This construction method not only realizes the transformation from the actual time domain to the equivalent time domain, but also ensures the consistency of data distribution between the model training and inference stages through a unified vector structure. It eliminates prediction bias caused by differences in feature definitions, enabling the regression prediction model to accurately learn the nonlinear mapping relationship between process parameters, wear state and removal rate, and providing a reliable and standardized data input basis for the accurate calculation of subsequent hydraulic pressure compensation.

[0040] S3: Train a regression prediction model based on the process feature vector of the historical grinding process. Input the process feature vector of the grinding process to be compensated into the trained regression prediction model and output the predicted value of the removal rate of the grinding process to be compensated. Correct the predicted value of the removal rate based on the coefficient of determination to obtain the corrected predicted rate of the grinding process to be compensated.

[0041] In one embodiment, the process feature vectors of several historical grinding processes completed in the equipment's historical database are used as training samples to train the regression prediction model, enabling the model to learn and establish a nonlinear mapping relationship between process parameters, wear state, and removal rate. The historical process feature vectors are also processed through equivalent time mapping to ensure consistency between the training space and the prediction space.

[0042] The process feature vector of the grinding process to be compensated is input into a trained regression prediction model. The model performs real-time inference based on the current wear equivalent time, hydraulic pressure, and initial workpiece thickness, outputting a predicted value for the removal rate of the grinding process. This predicted removal rate quantifies the expected processing efficiency under the current grinding disc cutting capability, eliminating the interference of wear non-stationarity on the prediction results. It provides core data support for subsequent calculation of hydraulic pressure compensation based on prediction deviation, ensuring the accuracy and adaptability of the compensation command, thereby achieving machine learning-based intelligent hydraulic compensation.

[0043] Among them, the regression prediction model uses a machine learning algorithm with strong nonlinear fitting ability, preferably a Gaussian process regression model, or a neural network, support vector machine regression or random forest regression model can also be used to adapt to the complex nonlinear mapping relationship between process parameters and removal rate.

[0044] The training data for the regression prediction model comes from completed grinding batch records in the equipment's historical database. Each training sample's input is a process feature vector, and the output is the predicted removal rate for the corresponding historical grinding process. The label is the actual removal rate for the corresponding historical grinding process, ensuring high confidence in the training labels. The training process employs existing optimization algorithms such as gradient descent or maximum likelihood estimation to adjust the model's internal parameters, aiming to minimize the error between the model's predicted values ​​and the actual removal rate, thereby accurately establishing the nonlinear mapping relationship between process parameters, wear state, and removal rate.

[0045] The trained regression prediction model can infer the predicted value of the removal rate in real time based on the process feature vector of the grinding process to be compensated, and quantify the expected processing efficiency under the current cutting capability of the grinding disc.

[0046] In the process of dynamically correcting the removal rate prediction based on the coefficient of determination, the predicted removal rate of the reference process and its corrected prediction rate after error feedback are first obtained. The difference between the two is calculated as the first difference, and the ratio of the first difference to the predicted removal rate of the reference process is defined as the prediction error residual of the reference process. This is used to quantify the degree of actual prediction deviation of the model in the previous batch.

[0047] Subsequently, the difference between constant 1 and the determination coefficient obtained when fitting the global wear decay rate is taken as the second difference. This second difference represents the complement of the physical model fitting credibility, i.e. the weight of the model uncertainty. The product of the second difference and the prediction error residual is taken as the second product. Then, the sum of constant 1 and the second product is used to construct the correction factor, thereby realizing the adaptive adjustment of the error feedback strength according to the model fitting quality.

[0048] Finally, the product of the correction factor and the predicted value of the removal rate of the grinding process to be compensated is used as the corrected predicted rate of the grinding process to be compensated.

[0049] The error experience from the previous historical grinding process is weighted and transferred to the grinding process to be compensated. When the coefficient of determination is high, the correction factor approaches 1, mainly relying on the physical model. When the coefficient of determination is low, the correction factor deviates significantly from 1 to enhance error feedback compensation. This eliminates short-term random disturbances while adapting to wear non-stationarity, and improves the robustness of the prediction model throughout the entire life cycle and the accuracy of hydraulic pressure compensation.

[0050] S4: Calculate the hydraulic pressure compensation amount for the grinding process to be compensated based on the corrected prediction rate, the hydraulic pressure of the grinding process to be compensated, and the obtained standard removal rate; generate control commands based on the hydraulic pressure compensation amount to complete the intelligent hydraulic compensation.

[0051] In one embodiment, a standard removal rate is obtained, which is defined as the true value of the removal rate of the first historical grinding process completed under standard pressure after the grinding disc is replaced, representing the ideal processing efficiency benchmark in the initial state of the new disc.

[0052] The difference between the standard removal rate and the corrected predicted rate of the grinding process to be compensated is used as the third difference. The third difference quantifies the degree of absolute deviation between the expected removal capacity and the ideal benchmark under the current wear state. The ratio of the third difference to the corrected predicted rate of the grinding process to be compensated is used as the rate ratio. The rate ratio characterizes the relative proportion of the removal capacity deviation to the current predicted value, thus achieving the normalization of the deviation.

[0053] The product of the hydraulic pressure and the rate ratio in the grinding process to be compensated is used as the hydraulic pressure compensation amount. The calculation is based on the physical law that hydraulic pressure and removal rate are linearly proportional, directly mapping the relative deviation of the removal rate to the adjustment ratio of the hydraulic pressure. Specifically, when the corrected predicted rate is lower than the standard removal rate, the third difference is positive, the rate ratio is positive, and the hydraulic pressure compensation amount is positive, instructing the hydraulic system to increase the contact pressure to compensate for the reduction in removal capacity caused by wear. When the corrected predicted rate is higher than the standard removal rate, the third difference is negative, the rate ratio is negative, and the hydraulic pressure compensation amount is negative, instructing the hydraulic system to appropriately reduce the pressure to avoid the risk of over-grinding.

[0054] The calculation process of hydraulic pressure compensation accurately transforms the removal rate deviation predicted by the regression prediction model into an executable hydraulic pressure adjustment command, realizing feedforward compensation of wear effect, ensuring that the workpiece thickness can be maintained in different wear stages, while avoiding reliance on high-frequency intermediate thickness measurement feedback, effectively improving continuous processing efficiency and dimensional control accuracy.

[0055] When the corrected predicted rate is lower than the standard removal rate, it indicates that the wear of the grinding disc has led to insufficient cutting capacity, making it impossible to maintain the initial processing efficiency. In this case, the calculated hydraulic pressure compensation is positive, indicating that the hydraulic system increases the contact pressure of the upper grinding disc on the workpiece to improve the removal efficiency and compensate for the decrease in capacity caused by wear. Conversely, when the corrected predicted rate is higher than the standard removal rate, it indicates that the removal capacity under the current process condition is excessive, and there is a risk of over-grinding. In this case, the hydraulic pressure compensation is negative, indicating that the hydraulic system appropriately reduces the contact pressure to avoid insufficient workpiece thickness or surface damage, and to ensure processing safety.

[0056] Based on the working principle of the double-end grinding machine, the hydraulic system adopts an independent closed-loop control architecture, physically decoupled from the motion systems such as the grinding wheel rotation and carrier revolution. This allows for rapid response to pressure adjustment commands without interfering with the stable operation of other motion axes, ensuring real-time and accurate pressure regulation. Therefore, the hydraulic pressure compensation is a key variable in converting the model-predicted removal rate deviation into physically executable control commands. Based on the physical law that hydraulic pressure and removal rate are linearly proportional, the relative deviation of the removal rate is directly mapped to the adjustment ratio of the hydraulic pressure. Through a feedforward adjustment mechanism, the wear effect is actively offset, ensuring that the workpiece thickness remains consistent at different wear stages. This eliminates the need for high-frequency intermediate thickness measurement feedback, enabling dimensional accuracy control during continuous processing.

[0057] Obtain the minimum and maximum working pressures allowed by the hydraulic system. The minimum and maximum working pressures are set by the equipment manufacturer based on the rated load of the hydraulic cylinder, the pressure resistance level of the seals, and the process safety margin, which constitute the physical boundary constraints for pressure regulation.

[0058] The sum of the hydraulic pressure during the grinding process to be compensated and the hydraulic pressure compensation amount is used as the comprehensive pressure, which characterizes the target contact pressure setpoint after wear compensation correction. Then, a limiting function is used to restrict the comprehensive pressure between the minimum and maximum working pressures, generating the final control command. This limiting process effectively avoids the risk of pressure exceeding limits due to model prediction deviations, instantaneous disturbances, or extreme wear conditions, ensuring that the hydraulic system always operates within a safe range and preventing impact damage to the workpiece surface quality caused by sudden pressure changes. After the control command is generated, it is sent to the proportional valve of the hydraulic system via the communication bus, driving the hydraulic cylinder piston rod to adjust the vertical pressure of the upper grinding disc in real time, achieving feedforward compensation for wear effects.

[0059] After the grinding process to be compensated is completed, the final thickness of the workpiece is measured using a grating thickness gauge equipped with the equipment. Based on the difference between the initial thickness and the final thickness of the workpiece, the true value of the removal rate of the grinding process to be compensated is calculated. The true value of the removal rate serves as a high-confidence process feedback label and is automatically updated to the equipment's historical database. This is used to correct the global wear attenuation rate fitting and regression prediction model training for the next grinding process. The update mechanism does not require additional intermediate thickness measurement frequency and can fully utilize conventional process data to complete the model self-evolution. This ensures that the grinding disc can maintain the workpiece thickness tolerance within a very small range throughout its entire life cycle, significantly improving continuous processing efficiency and dimensional consistency, and realizing machine tool hydraulic intelligent compensation based on machine learning.

[0060] The machine tool hydraulic control intelligent compensation system based on machine learning includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the machine tool hydraulic control intelligent compensation method based on machine learning according to the first aspect of the present invention is implemented.

[0061] The machine tool hydraulic control intelligent compensation system based on machine learning also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0062] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A machine learning based intelligent compensation method for machine tool hydraulic control, characterized in that, include: The hydraulic pressure, cumulative grinding time, and initial workpiece thickness of any historical grinding process are obtained. The cumulative grinding time is the cumulative running time from the moment the grinding disc is replaced or repaired until the end of any historical grinding process. Based on the historical database, a regression fitting algorithm is used to perform trend fitting, and the global wear attenuation rate, which characterizes the rate of attenuation of cutting capability per unit time, is calculated. At the same time, the determination coefficient in the fitting process is obtained. Using the previous historical grinding process adjacent to the grinding process to be compensated as a reference process, a wear equivalent time feature that can characterize the cumulative wear state is constructed. The fitted global wear decay rate and the current cumulative grinding time are nonlinearly transformed to map the non-stationary actual time domain to a stationary equivalent time domain. The wear equivalent time of the grinding process to be compensated is calculated based on the global wear decay rate and the cumulative grinding time of the reference process. The wear equivalent time, the hydraulic pressure of the grinding process to be compensated, and the initial thickness of the workpiece in the grinding process to be compensated are used to construct the process feature vector of the grinding process to be compensated. Similarly, the process feature vector of any historical grinding process is obtained. The regression prediction model is trained based on the process feature vector of the historical grinding process. The process feature vector of the grinding process to be compensated is input into the trained regression prediction model, and the predicted value of the removal rate of the grinding process to be compensated is output. The predicted value of the removal rate is corrected based on the determination coefficient to obtain the corrected predicted rate of the grinding process to be compensated. The hydraulic pressure compensation amount for the grinding process to be compensated is calculated based on the corrected prediction rate, the hydraulic pressure of the grinding process to be compensated, and the obtained standard removal rate; control commands are generated according to the hydraulic pressure compensation amount to complete the intelligent hydraulic compensation.

2. The machine learning based machine tool hydraulic control intelligent compensation method according to claim 1, characterized in that, The process of fitting the global wear rate based on historical databases and obtaining the coefficients of determination during the fitting process includes: The experiment process is to take several historical grinding processes completed under standard pressure in the historical database as experimental processes and obtain data pairs for any experimental process. The data pairs include the true values ​​of the cumulative grinding time and removal rate. A regression fitting algorithm was used on all experimental data pairs to obtain the global wear attenuation rate. Record the coefficients of determination of the fitted curve during the fitting process.

3. The machine learning based machine tool hydraulic control intelligent compensation method according to claim 1, characterized in that, The calculation of the wear equivalent time of the grinding process to be compensated includes: The product of the global wear decay rate and the cumulative grinding time of the reference process is calculated as the first product. The negative exponent of the first product is calculated, and the difference between the constant 1 and the negative exponent is taken as the wear equivalent time of the grinding process to be compensated.

4. The intelligent compensation method for machine tool hydraulic control based on machine learning according to claim 1, characterized in that, The regression prediction model employs Gaussian process regression, neural networks, support vector machine regression, or random forest regression.

5. The intelligent compensation method for machine tool hydraulic control based on machine learning according to claim 1, characterized in that, The predicted value of the removal rate is corrected based on the determination coefficient to obtain the corrected predicted rate of the grinding process to be compensated, including: Obtain the predicted value of the removal rate of the reference process and the corrected predicted rate, take the difference between the predicted value of the removal rate of the reference process and the corrected predicted rate as the first difference, and take the ratio of the first difference to the predicted value of the removal rate of the reference process as the prediction error residual of the reference process. The difference between constant 1 and the coefficient of determination is taken as the second difference, the product of the second difference and the prediction error residual is taken as the second product, and the sum of constant 1 and the second product is taken as the correction factor. The product of the correction factor and the predicted value of the removal rate of the grinding process to be compensated is used as the corrected predicted rate of the grinding process to be compensated.

6. The intelligent compensation method for machine tool hydraulic control based on machine learning according to claim 1, characterized in that, The calculation of the hydraulic pressure compensation amount for the grinding process to be compensated includes: Obtain the standard removal rate, take the difference between the standard removal rate and the corrected predicted rate of the grinding process to be compensated as the third difference, and take the ratio of the third difference to the corrected predicted rate of the grinding process to be compensated as the rate ratio. The product of the hydraulic pressure and the rate ratio in the grinding process to be compensated is used as the hydraulic pressure compensation amount for the grinding process to be compensated.

7. The intelligent compensation method for machine tool hydraulic control based on machine learning according to claim 1, characterized in that, The step of generating control commands based on the hydraulic pressure compensation amount to complete intelligent hydraulic compensation includes: Obtain the minimum and maximum permissible working pressure in the hydraulic system; The sum of the hydraulic pressure during the grinding process to be compensated and the hydraulic pressure compensation amount is taken as the comprehensive pressure; The combined pressure is limited between the minimum and maximum working pressure by a limiting function, thereby generating control commands.

8. The intelligent compensation method for machine tool hydraulic control based on machine learning according to claim 1, characterized in that, Also includes: The final thickness of the workpiece after the compensation grinding process is measured by a grating thickness gauge, and the true value of the removal rate of the compensation grinding process is calculated and updated to the historical database.

9. A machine tool hydraulic control intelligent compensation system based on machine learning, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the machine tool hydraulic control intelligent compensation method based on machine learning according to any one of claims 1-8.

Images