Method for improving gear machining precision

Abstract

The invention discloses a method for improving gear machining precision. The method comprises the steps that a tool base body is subjected to cleaning, surface modification and stress relief treatment; a dynamic coupling error field model is constructed, and dynamic compensation is carried out through a feedforward and feedback composite strategy; performing tool dynamic balance grading calibration and cutting vibration active suppression; a digital twinborn model of the physical processing system is constructed, and closed-loop control is formed through virtual and real data linkage and iterative optimization; according to the method, fusion prediction is carried out on the multi-source error through the dynamic coupling error field model, adaptive fusion of physical constraint and data driving is realized, and the technical problem that an existing error compensation model is limited to a single physical mechanism and cannot accurately pre-judge multi-source error coupling is solved; through deep cooperation of error compensation, process management and control and a digital twinning technology, dynamic adaptation and full-period closed-loop regulation and control of all processes are realized, and the technical problems that in the prior art, all links are poor in collaboration, and parameters are fixed and cannot adapt to working condition changes are solved.

Description

Technical Field

[0001] This invention relates to the field of gear machining technology, and in particular to a method for improving gear machining accuracy. Background Technology

[0002] As a core component of mechanical transmission systems, the machining accuracy of gears directly determines transmission efficiency, noise levels, and equipment lifespan. In the high-end equipment sector, the precision requirements for gears have been raised to ISO 3-4 level. Existing methods for controlling gear machining accuracy mostly employ segmented control strategies, which have significant technical bottlenecks and struggle to meet the demands of high-precision, high-volume processing.

[0003] On the one hand, existing error compensation models are mostly limited to a single physical mechanism, such as only considering thermal deformation or geometric errors or pure data-driven modeling, lacking the ability to adapt and integrate physical constraints with data.

[0004] On the other hand, error compensation has poor coordination with machining process control and digital twin technology. Vibration suppression and tool dynamic balance calibration mostly adopt fixed parameter control, which is disconnected from the error compensation process and cannot dynamically adapt to changes in working conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method for improving gear machining accuracy, thereby solving the problems mentioned in the background section. To achieve the above objective, the present invention employs the following technical solution: The method for improving gear machining accuracy includes the following steps: S1. Clean, modify and stress-relieve the tool substrate to avoid machining accuracy deviations caused by deformation and defects at the substrate level. S2. Construct and apply a dynamic coupled error field model based on physical information neural network to perform fusion prediction of machine tool geometric error, cutting force, temperature and tool wear, and perform dynamic compensation through a feedforward feedback composite strategy. S3. Implement tool dynamic balancing calibration and active suppression of cutting vibration to reduce the impact of dynamic interference on tooth surface accuracy; S4. Construct a digital twin model of the physical processing system, and form a closed-loop control through the linkage and iterative optimization of virtual and real data.

[0006] Optionally, step S2 includes the following steps: S21. Establish a set of partial differential equations describing the thermal deformation and force-induced deformation during the system processing, as the physical kernel of the model; S22. Construct an attention mechanism neural network, with time-series data including cutting path curvature and instantaneous material removal rate as input; S23. Through the adaptive fusion module, the output weights of the physical kernel and the attention mechanism neural network are dynamically adjusted according to the processing state, and the error prediction values ​​output by the physical kernel and the error prediction values ​​output by the attention mechanism neural network are fused.

[0007] S24. Based on the fusion results, output a dynamic error field distribution map covering the tooth surface in the future time domain.

[0008] Optionally, the dynamic compensation is calculated using the following formula: σ(t)=σ0(t)-k(T,F)·[d·ΔG(t)+e·ΔF(t)+f·ΔT(t)] Where σ(t) is the compensated error, σ0(t) is the original error, d, e, and f are the weighting coefficients of geometric error, cutting force error, and temperature error, respectively, and ΔG(t), ΔF(t), and ΔT(t) are the real-time error quantities of geometry, cutting force, and temperature, respectively.

[0009] Optionally, the method further includes the following steps before step S2: The model is preprocessed and outliers are removed from the multi-error source data. The model is then trained and optimized. Overfitting is suppressed through iterative algorithms. The accuracy of the model compensation is verified under various processing conditions.

[0010] Optionally, the dynamic weight adjustment includes: the dynamic adjustment of the output weights of the physical kernel and the attention mechanism neural network includes: when the cutting force and vibration signal fluctuations are below a preset threshold in the stable cutting stage, the physical kernel output is given a weight value higher than the first weight value to ensure modeling accuracy; when the sensor detects that the cutting force or vibration signal changes exceed the threshold, the output weight of the attention mechanism neural network is increased to a weight value higher than the second weight value within a preset time.

[0011] Optionally, the active suppression of cutting vibration includes the following steps: S31. Real-time acquisition of vibration signals during the processing; S32. Perform spectral analysis on the vibration signal to identify the key vibration frequency bands that affect the accuracy of the tooth surface; S33. Dynamically adjust damping control parameters based on analysis results to suppress key vibration frequency bands in a targeted manner; S34. Adjust the processing parameters in conjunction with the control and verify the suppression effect to ensure that the vibration amplitude is within the allowable accuracy range.

[0012] Optionally, the spectral analysis of the vibration signal includes the following steps: S321. Filter the original vibration signal to remove environmental interference noise; S322. The frequency amplitude distribution of the vibration signal is obtained by using the Fast Fourier Transform algorithm; S323. Set the amplitude threshold according to the gear machining accuracy requirements and lock the high-frequency vibration band that exceeds the threshold.

[0013] Optionally, the dynamic adjustment of damping control parameters based on the analysis results includes the following steps: S331. Match the initial damping parameters according to the locked vibration frequency band, and initiate targeted suppression control; S332. Calculate the vibration suppression rate and determine whether it meets the preset requirements; S333. If the suppression rate does not meet the standard, fine-tune the damping parameters through an iterative algorithm until the vibration amplitude drops to the allowable range. S334. Record the optimal parameters and store them in the database.

[0014] Optionally, step S4 includes the following steps: S41. Construct a digital twin model that includes machine tools, workpieces, fixtures, and environmental elements; S42. Predict future trends in machining accuracy based on a long short-term memory network model; S43. Feed back the online detected tooth surface error data to the twin model and iteratively optimize the cutting parameters and error compensation coefficients.

[0015] Optionally, after step S43, a dynamic calibration of the digital twin model is further included, comprising the following steps: S44. Real-time calculation of the residual between the measured values ​​and the twin simulation values ​​of key physical quantities; S45. When the residual continuously exceeds the preset residual threshold and exceeds the preset number of times, it is determined that the twin is inaccurate and the calibration process is triggered. S46. The calibration process uses a reverse identification algorithm to prioritize adjusting the machine tool dynamic parameters and tool wear state parameters in the twin, so that the matching degree between the simulated value and the measured value is restored to within the preset residual threshold.

[0016] Compared to existing technologies, the advantages of this invention are that, by adopting the above-mentioned scheme, the present invention achieves adaptive fusion of physical constraints and data-driven approaches by using a dynamically coupled error field model to fuse and predict multi-source errors. This solves the technical problem that existing error compensation models are limited to a single physical mechanism and are difficult to accurately predict the coupling of multi-source errors. Through the deep collaboration of error compensation, process control and digital twin technology, the invention achieves dynamic adaptation and full-cycle closed-loop control of each process, solving the technical problems of poor coordination between links and fixed parameters that cannot adapt to changes in operating conditions in existing technologies. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the method for improving gear machining accuracy according to the present invention; Figure 2This is a schematic diagram illustrating the creation process of the dynamic coupling error field model of the present invention; Figure 3 This is a schematic diagram of the active suppression of cutting vibration according to the present invention; Figure 4 This is a schematic diagram of the spectrum analysis process for vibration signals according to the present invention; Figure 5 This is a schematic diagram of the process for dynamically adjusting damping control parameters based on analysis results according to the present invention. Figure 6 This is a schematic diagram illustrating the process of constructing a digital twin model of the physical processing system according to the present invention; Figure 7 This is a schematic diagram of the dynamic calibration process for the digital twin model according to the present invention. Detailed Implementation

[0018] To facilitate understanding of this application, a more detailed description of the application is provided below with reference to the accompanying drawings and specific embodiments; preferred embodiments of the application are shown in the drawings; however, the application may be implemented in many different forms and is not limited to the embodiments described in this specification; rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of this application.

[0019] It should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. In the embodiments shown in the accompanying drawings, directional indications (such as up, down, left, right, front, and back) are used to explain the structure and movement of various components and are not absolute but relative. These descriptions are appropriate when these components are in the positions shown in the drawings. If the descriptions of the positions of these components change, these directional indications also change accordingly.

[0020] It should also be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence; it should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in a sequence other than those illustrated or described herein.

[0021] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit this application.

[0022] like Figure 1 As shown, one embodiment of the present invention is as follows: Figure 1 As shown, one embodiment of the present invention is a method for improving gear machining accuracy, comprising the following steps: S1. Select a cemented carbide tool substrate and first use an ultrasonic cleaning device to clean the tool to remove oil, oxide layer and impurities; then perform surface modification through physical vapor deposition process to prepare a TiAlN coating to improve the surface hardness and wear resistance of the tool; finally, perform low temperature aging treatment to eliminate residual stress inside the tool substrate and avoid machining accuracy deviation caused by deformation and defects from the substrate level.

[0023] S2. Construct a dynamic coupled error field model based on a physical information neural network (PINN), simultaneously collect machine tool geometric errors, cutting force signals, cutting zone temperature and tool wear, input the above data into the model for fusion prediction, and then use a feedforward feedback composite strategy to dynamically correct the machine tool motion commands based on the prediction results. At the same time, fine-tune the tool holder tilt angle according to real-time tooth surface detection data to achieve accurate error compensation.

[0024] Among these methods, machine tool geometric errors can be detected by a laser interferometer, cutting force signals can be collected by a piezoelectric force sensor, the temperature of the cutting zone can be collected by an infrared thermometer, and the tool wear can be collected by a vision inspection system.

[0025] S3. Perform coarse calibration using a dynamic balancing instrument to quickly eliminate imbalances greater than 5 g·mm; then perform fine calibration to achieve a balance accuracy of G1 level, meeting the stability requirements of high-speed cutting; at the same time, activate the active cutting vibration suppression module to collect vibration signals in real time and dynamically adjust damping parameters to reduce the impact of dynamic interference on tooth surface accuracy.

[0026] S4. Construct a full-element digital twin model covering machine tools, workpieces, fixtures, and machining environment. Real-time linkage between virtual and real data is achieved through industrial Ethernet. Based on the model, the trend of machining accuracy changes is predicted. The actual detected tooth surface error data is fed back to the model. The cutting parameters and error compensation coefficients are iteratively optimized to form a complete closed-loop control.

[0027] This application achieves adaptive fusion of physical constraints and data-driven approaches by using a dynamically coupled error field model to predict multi-source errors. This solves the technical problem that existing error compensation models are limited to a single physical mechanism and are difficult to accurately predict the coupling of multi-source errors. Through the deep collaboration of error compensation, process control and digital twin technology, dynamic adaptation and full-cycle closed-loop control of each process are achieved, solving the technical problems of poor coordination between links and fixed parameters that cannot adapt to changes in operating conditions in existing technologies.

[0028] In one embodiment, such as Figure 2 As shown, step S2 includes the following steps: S21. Establish a set of partial differential equations describing the thermal deformation and force-induced deformation of the machine tool, cutting tool, and workpiece system. The thermal deformation equation adopts the Fourier heat conduction equation, and the force-induced deformation equation adopts the elasticity equilibrium equation. The thermal deformation equation adopts the Fourier heat conduction equation, which is adapted to the machine tool, cutting tool, and workpiece system. The thermal conductivity matches the material of the cutting tool and workpiece. The boundary conditions are set as the heat source input in the cutting zone and the heat dissipation constraint of the machine tool components. The force-induced deformation equation adopts the elasticity equilibrium equation, with the cutting force as the external load. The constraint conditions match the fixed state of the machine tool fixture and the force characteristics of the tool cantilever. The above set of equations is used as the physical kernel of the model to constrain the prediction direction of the model and avoid prediction deviations caused by pure data-driven approaches.

[0029] S22. Construct an attention mechanism neural network, select a self-attention layer, and adapt it to the extraction of associated features of time-series data. The network includes an input layer, an attention layer, a hidden layer, and an output layer. The input data is the time-series data of cutting path curvature and instantaneous material removal rate. The time step is set to 10ms to ensure the timeliness of the data. The cutting path curvature is obtained through the machine tool CNC system, and the instantaneous material removal rate is calculated based on the cutting speed, feed rate, and cutting depth.

[0030] S23. Design an adaptive fusion module, which includes a working condition identification unit, a weight calculation unit, and a fusion execution unit. The working condition identification unit is linked with sensors that collect cutting force and vibration signals in real time to monitor the fluctuation status of cutting force and vibration signals. The weight calculation unit presets weight thresholds for stable working conditions and abrupt working conditions, and dynamically calculates the output weights of the physical kernel and the attention mechanism neural network based on the detection results of the working condition identification unit. The fusion execution unit receives the weight calculation results and performs weighted fusion on the error prediction values ​​output by the physical kernel and the error prediction values ​​output by the attention mechanism neural network to achieve accurate adaptation and fusion of the prediction results of the two.

[0031] S24. Based on the fusion results, output a dynamic error field distribution map covering the entire tooth surface of the gear within the time domain of the next 500ms. The resolution of the distribution map is set to 0.01mm, and the error values ​​at each position are accurately marked to provide an intuitive basis for subsequent error compensation.

[0032] This embodiment ensures the rationality of model predictions through physical kernel constraints, strengthens the influence of key time-series data through attention mechanism neural networks, and achieves accurate adaptation under different processing conditions through adaptive fusion modules. The final output of the full tooth surface dynamic error field map provides an accurate and comprehensive predictive basis for error compensation and improves the pertinence of the compensation strategy.

[0033] In one embodiment, the dynamic compensation formula is σ(t)=σ0(t)-k(T,F)·[d·ΔG(t)+e·ΔF(t)+f·ΔT(t)], where the parameter determination and application process includes: the weighting coefficients d, e, and f are calibrated through orthogonal experiments, and based on the degree of error influence under different processing conditions, d=0.35, e=0.4, and f=0.25 are set to ensure that the coefficients are adapted to the multi-source error coupling characteristics; the dynamic correction coefficient k(T,F) is obtained by fitting the coupling relationship between cutting temperature T and cutting force F, and the fitting formula is k(T,F)=0.002T+0.001F+0.5 (T is in °C, F is in N), which is adjusted in real time according to the changes in the working conditions; ΔG(t), ΔF(t), and ΔT(t) are collected in real time by a laser interferometer, a piezoelectric force sensor, and an infrared thermometer, respectively, with the sampling frequency set to 1kHz to ensure data real-time performance. During compensation, the original error σ0(t) is first calculated, then substituted into the formula to calculate the compensation amount. The motion command is then corrected through the machine tool CNC system to achieve dynamic compensation.

[0034] This embodiment is based on real-time acquisition data of multi-source errors. It distinguishes the degree of influence of each error source by weighting coefficients, dynamically corrects coefficients to adapt to changes in working conditions, calculates accurate compensation amount by formula, and reverses the machine tool motion parameters to achieve dynamic cancellation of multi-source errors.

[0035] In one embodiment, before model training, the multi-error source data is preprocessed. Outliers are removed using the 3σ criterion to eliminate extreme data caused by environmental interference. Missing data is supplemented using linear interpolation to ensure data integrity. The min-max normalization algorithm is used to normalize the data, mapping the values ​​to the [0,1] interval to improve model training convergence speed. During training, the Adam optimizer is selected, with a learning rate of 0.001 and 1000 iterations. A gradient descent iterative algorithm combined with an early stopping strategy is used to suppress overfitting. The early stopping strategy includes: dividing the preprocessed error source data into training and validation sets in a 7:3 ratio; calculating the validation set error every 10 iterations during training; and immediately stopping training and saving the current optimal model parameters when the validation set error shows no decrease for 20 consecutive iterations and the training set error continues to decrease. The gradient descent algorithm step size is set to 0.0005 to adapt to the model training convergence requirements. During the verification phase, five typical machining conditions were selected, such as different combinations of cutting speed and feed rate. Three sets of gear samples were machined for each condition. The tooth surface error was detected by the gear measurement center to verify the accuracy of the model compensation and ensure that the tooth surface error after compensation is controlled within the standard accuracy range.

[0036] This embodiment improves the model's generalization ability and stability through data preprocessing and targeted training strategies, avoiding insufficient adaptation to working conditions caused by overfitting; multi-working-condition verification ensures that the model can accurately compensate in different processing scenarios.

[0037] In one embodiment, dynamically adjusting the output weights of the physical kernel and the attention mechanism neural network includes: setting the cutting force fluctuation threshold to ±5%, the vibration signal fluctuation threshold to ±3μm, setting the first weight value to 70%, the preset time to 50ms, and the second weight value to 60%. When the cutting force and vibration signal fluctuations are both below the corresponding thresholds and the cutting process is stable, the adaptive fusion module assigns the physical kernel output a weight of more than 70% (default 75%), relying on physical constraints to ensure modeling accuracy. When the sensor detects a sudden change in cutting force or vibration signal exceeding the threshold, such as a sudden increase in cutting force due to uneven workpiece material, the module increases the output weight of the attention mechanism neural network to more than 60% (default 65%) within 50ms, enhancing the data-driven real-time adaptation capability and quickly responding to changes in working conditions.

[0038] This embodiment identifies the processing state based on cutting force and vibration signals. In a stable state, it relies on the physical kernel to ensure prediction accuracy. When the working condition changes abruptly, it enhances the real-time response capability of the attention network. By dynamically allocating weights to adapt to different processing scenarios, it balances prediction accuracy and real-time performance, solving the problem that a single weight cannot take both accuracy and real-time performance into account. This ensures that the model can maintain excellent prediction performance under different processing states.

[0039] In one embodiment, such as Figure 3 As shown, the active suppression process for cutting vibration includes the following steps: S31. Install an acceleration sensor at the end of the machine tool spindle, set the acquisition frequency to 2kHz, and collect vibration signals during the machining process in real time, and transmit them synchronously to the signal processing unit.

[0040] S32. The signal processing unit performs spectrum analysis on the collected vibration signal to identify the key vibration frequency bands that affect the accuracy of the tooth surface, which are usually the high-frequency band of 2000-5000Hz.

[0041] S33. Based on the identified key frequency bands, dynamically adjust the damping coefficient of the machine tool damper, with an adjustment range of 0.1-0.8, to specifically suppress key vibration frequency bands and reduce vibration amplitude.

[0042] S34. After adjustment, the vibration amplitude is monitored in real time by the sensor to verify the suppression effect and ensure that the vibration amplitude is controlled within 2μm. At the same time, the cutting speed and feed rate are adjusted in conjunction with the adjustment range of ±10% to further optimize the vibration suppression effect. This forms a dynamic synergy with the error compensation in step S2 to avoid vibration interference from offsetting the compensation effect.

[0043] This embodiment captures vibration signals in real time using sensors, identifies key interference frequency bands through spectrum analysis, achieves targeted suppression through dynamic damping adjustment, optimizes processing parameters in conjunction with error compensation, forms a vibration suppression closed loop, weakens the impact of dynamic interference on accuracy, improves the stability of the processing, and avoids defects such as tooth surface ripples and scratches caused by vibration.

[0044] In one embodiment, such as Figure 4 As shown, the spectral analysis of vibration signals includes the following steps: S321. The raw vibration signal collected by the accelerometer is filtered using a low-pass filter with a cutoff frequency of 10kHz to remove environmental interference noise, such as workshop environmental vibration and equipment electromagnetic interference, and retain the effective vibration signal.

[0045] S322. The Fast Fourier Transform (FFT) algorithm is used to process the filtered signal. Combined with the 2kHz sampling frequency of the vibration signal mentioned above, the number of FFT sampling points is set to 4096. The Hanning window is selected to suppress spectral leakage, and the time domain signal is converted into a frequency domain signal to obtain the frequency-amplitude distribution curve of the vibration signal and clarify the vibration amplitude corresponding to each frequency band.

[0046] Understandably, spectral leakage refers to the fact that the cutting vibration signals acquired in this step, such as machine tool spindle vibration, are unlikely to be exactly integer periods. Directly processing them through FFT will cause the vibration energy of a single frequency to spread to surrounding frequency bands, resulting in frequency ambiguity and affecting the accurate identification of key vibration frequency bands. The Hanning window, as a weighting method for signal preprocessing, does not require additional hardware and can make the acquired vibration signal gradually change to 0 at both ends, simulating a signal shape close to an integer period, thereby effectively suppressing the above-mentioned spectral leakage phenomenon and ensuring that the frequency amplitude distribution curve of the FFT output is more accurate.

[0047] S323. Based on the gear machining accuracy requirements, the vibration amplitude threshold is set to 2μm. In the frequency amplitude distribution curve, the high-frequency vibration band with amplitude exceeding 2μm is locked as the target frequency band for subsequent damping control.

[0048] This embodiment first filters and purifies the signal, then uses Fourier transform to convert the time domain to the frequency domain. Based on the accuracy requirements, a threshold is set to lock the high-frequency vibration band that affects the accuracy, providing a precise target for damping parameter control and avoiding resource waste and poor results caused by blind suppression.

[0049] In one embodiment, such as Figure 5 As shown, dynamically adjusting the damping control parameters based on the analysis results includes the following steps: S331. Retrieve the initial damping parameters for the corresponding frequency band from the parameter database, and start the damper for targeted suppression and control by preset historical optimization data.

[0050] S332. After 100ms of suppression and control, calculate the vibration suppression rate. The suppression rate formula is (amplitude before control - amplitude after control) / amplitude before control × 100%. The preset requirement for the suppression rate is greater than or equal to 80%.

[0051] S333. If the suppression rate does not meet the standard, such as being less than 80%, the damping coefficient is fine-tuned using the gradient descent algorithm with a learning rate of 0.01 and a fine-tuning step size of 0.05. After each adjustment, the vibration amplitude is checked for 50ms until the vibration amplitude drops to within 2μm, at which point the suppression rate meets the standard.

[0052] S334. The optimal damping parameters, corresponding vibration frequency bands, and processing condition information after reaching the standard are associated and stored in the database to provide parameter references for similar processing scenarios and shorten the subsequent adjustment time.

[0053] This embodiment uses historical data to quickly match initial parameters, verifies the control effect through the inhibition rate, fine-tunes the parameters through an iterative algorithm until the target is met, and finally stores the optimal parameters to form a closed loop, achieving precise and efficient control of damping parameters, reducing the control cost under similar working conditions, and improving the consistency of batch processing.

[0054] In one embodiment, such as Figure 6 As shown, step S4 includes the following steps: S41. A digital twin model is constructed using a Python simulation platform. The construction process mainly includes three core steps to ensure accurate matching between the model and the physical processing system: First, full-element modeling: corresponding sub-models are established for the machine tool, workpiece, fixture, and processing environment, and each sub-model fits the gear processing scenario of this patent. Second, parameter calibration: based on the rated parameters in the machine tool manual, such as the spindle speed range and tool holder stroke, combined with the machine tool operation data and workpiece size data collected in the actual processing scenario, the parameters of each sub-model are calibrated to ensure that the characteristics of the sub-model are consistent with the physical components. Third, consistency verification: the no-load operation data of the physical system and the standard part processing data are input into the model, the model output results are compared with the physical detection results, and the model parameters are adjusted until the error between the two is controlled within 1%, finally forming a digital twin model covering all elements, laying the foundation for subsequent virtual-physical linkage. Among them, the machine tool includes the motion characteristics of the spindle and tool holder, the workpiece includes the gear blank and the geometric shape changes during the processing, the fixture includes the clamping stiffness and positioning error, and the processing environment includes temperature and humidity.

[0055] S42. A time-series prediction unit is constructed based on a Long Short-Term Memory (LSTM) network model. This unit is adapted to the gear machining accuracy prediction scenario of this patent. The specific adaptation details are as follows: The LSTM model is set with 3 hidden layers, each with 64 neurons, and the number of iterations is set to 500. The Adam optimizer is selected with a learning rate of 0.001. The input data uses the historical machining accuracy data, real-time process parameters and error compensation data mentioned above. The data is first normalized and mapped to the [0,1] interval to improve the model convergence speed. The model is trained based on the actual measurement data of the gear machining scenario to ensure that the prediction accuracy error is less than or equal to 2%. After training, it is used to predict the machining accuracy change trend in the next 1000ms. The prediction step size is set to 50ms to predict the accuracy fluctuation risk in advance.

[0056] S43. Iterative optimization: The gear surface error data is detected online through the gear measurement center and fed back to the digital twin model in real time. The model compares the predicted value with the actual detected value. When the deviation between the two is greater than 0.01mm, iterative optimization is started. The cutting speed and feed rate are adjusted by ±5%, and the error compensation coefficient is adjusted by ±0.05, until the deviation is less than or equal to 0.01mm and the machining accuracy is stable within the standard range, and the iteration stops.

[0057] In one embodiment, such as Figure 7 As shown, after step S43, dynamic calibration of the digital twin model is also included, comprising the following steps: S44. Select the core key physical quantities corresponding to the digital twin model and the physical machining system as the residual calculation benchmark, including machine tool spindle speed, cutting force, and tool wear. The measured values ​​of the machine tool spindle speed are acquired in real time through the machine tool CNC system, the measured values ​​of the cutting force are acquired through a piezoelectric force sensor, and the measured values ​​of the tool wear are acquired through a vision inspection system. The acquisition frequency is consistent with the data transmission frequency of the digital twin model, such as 1kHz, to ensure data synchronization. The residual calculation uses the absolute difference between the measured value and the simulated value, calculated as: δ = |X_measured - X_simulated|, where δ is the real-time residual of a single key physical quantity, X_measured is the real-time measured data of that physical quantity, and X_simulated is the synchronous simulation data of the corresponding physical quantity in the digital twin model. The calculation results are transmitted in real time to the control unit of the digital twin model for subsequent misalignment determination.

[0058] S45. In accordance with the high-precision gear machining requirements, preset residual thresholds for each type of key physical quantity are established to ensure that the judgment accuracy closely matches the actual machining scenario. Specific threshold settings are as follows: machine tool spindle speed residual threshold ≤ 5 r / min, cutting force residual threshold ≤ 5 N, and tool wear residual threshold ≤ 0.005 mm. Simultaneously, the number of consecutive exceedances of the residual threshold is set to 3 to avoid false calibration triggers due to instantaneous interference, such as environmental vibration or data fluctuations, ensuring the reliability of the judgment results. The digital twin model control unit monitors the residual changes of each type of key physical quantity in real time. When the residual of any key physical quantity exceeds the corresponding preset residual threshold 3 times consecutively, the digital twin model is deemed inaccurate, triggering a dynamic calibration process. If it only exceeds the preset residual threshold once or twice, it is judged as instantaneous interference, and calibration is not triggered; only fluctuation data is recorded to avoid frequent calibration affecting machining efficiency. After triggering the calibration process, the iterative optimization function of the digital twin model is paused, but the physical machining process is not stopped, achieving parallel calibration and machining to ensure machining continuity.

[0059] S46. The calibration process employs a reverse identification algorithm, specifically the gradient descent reverse identification algorithm, with the following parameters set: learning rate of 0.005, number of iterations of 200, and iteration step size of 0.001, ensuring smooth parameter adjustment and preventing over-calibration. During calibration, priority is given to adjusting the machine tool dynamics parameters and tool wear state parameters in the digital twin model, while other non-critical parameters are not adjusted, balancing calibration efficiency and model stability. Specifically, the machine tool dynamics parameters are adjusted primarily for spindle stiffness and guideway damping coefficient, with adjustments controlled within ±3% to ±8% to adapt to the actual operating characteristics of the machine tool. The tool wear state parameters are adjusted primarily for tool wear rate and wear threshold, with the adjustment process linked to the measured tool wear data collected by the vision inspection system to ensure that the adjusted parameters closely match the actual tool wear state. After each parameter adjustment, the residuals of the corresponding key physical quantities are calculated in real time to determine whether they have fallen within the preset residual threshold. If they have not fallen within the preset residual threshold, iterative adjustments continue until the residuals of all key physical quantities have fallen within the corresponding preset residual thresholds and the matching degree between the simulation value and the measured value has recovered to more than 99%. At this point, the calibration process is stopped, the iterative optimization function of the digital twin model is restored, and a dynamic calibration is completed.

[0060] In addition, to improve the efficiency of subsequent calibration and realize parameter reuse, the dynamic calibration process also includes a calibration recording step. The time of each calibration, the type of inaccuracy (i.e., which type of key physical quantity residual exceeds the threshold), the adjusted machine tool dynamic parameters and tool wear state parameters, and the matching degree data between the calibrated simulation values ​​and measured values ​​are associated and stored in the parameter database mentioned above. This provides data support for model maintenance and parameter optimization in subsequent batch processing, reduces the number of repeated calibrations, and improves the accuracy consistency of batch processing.

[0061] This embodiment achieves accurate identification and efficient calibration of digital twin model inaccuracies through a dynamic calibration process, avoiding problems such as accuracy prediction deviations and iterative optimization failures caused by parameter drift and disconnection from the physical machining system due to long-term model operation. It prioritizes the adjustment of machine tool dynamic parameters and tool wear parameters, without needing to comprehensively adjust all model parameters, thus balancing calibration efficiency and accuracy and reducing calibration costs. The calibration process runs in parallel with machining, without affecting machining continuity and ensuring machining efficiency. The recording and reuse of calibration data further improves the stability and consistency of the model in batch machining.

[0062] It should be noted that the above-mentioned technical features can be combined with each other to form various embodiments not listed above, all of which are considered to be within the scope of this invention specification; and, for those skilled in the art, improvements or modifications can be made based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for improving gear machining accuracy, characterized in that, Includes the following steps: S1. Clean, modify and stress-relieve the tool substrate to avoid machining accuracy deviations caused by deformation and defects at the substrate level. S2. Construct and apply a dynamic coupled error field model based on physical information neural network to perform fusion prediction of machine tool geometric error, cutting force, temperature and tool wear, and perform dynamic compensation through a feedforward feedback composite strategy. S3. Implement tool dynamic balancing calibration and active suppression of cutting vibration to reduce the impact of dynamic interference on tooth surface accuracy; S4. Construct a digital twin model of the physical processing system, and form a closed-loop control through the linkage and iterative optimization of virtual and real data.

2. The method for improving gear machining accuracy according to claim 1, characterized in that, Step S2 includes the following steps: S21. Establish a set of partial differential equations describing thermal deformation and force-induced deformation during the system processing, as the physical kernel of the model; S22. Construct an attention mechanism neural network, with time-series data including cutting path curvature and instantaneous material removal rate as input; S23. Through the adaptive fusion module, the output weights of the physical kernel and the attention mechanism neural network are dynamically adjusted according to the processing state, and the error prediction values ​​output by the physical kernel and the error prediction values ​​output by the attention mechanism neural network are fused. S24. Based on the fusion results, output a dynamic error field distribution map covering the tooth surface in the future time domain.

3. The method for improving gear machining accuracy according to claim 1, characterized in that, The dynamic compensation is calculated using the following formula: σ(t)=σ0(t)-k(T,F)·[d·ΔG(t)+e·ΔF(t)+f·ΔT(t)] Where σ(t) is the compensated error, σ0(t) is the original error, d, e, and f are the weighting coefficients of geometric error, cutting force error, and temperature error, respectively, and ΔG(t), ΔF(t), and ΔT(t) are the real-time error quantities of geometry, cutting force, and temperature, respectively.

4. The method for improving gear machining accuracy according to claim 2, characterized in that, The procedure preceding step S2 also includes: The model is preprocessed and outliers are removed from the multi-error source data. The model is then trained and optimized. Overfitting is suppressed through iterative algorithms. The accuracy of the model compensation is verified under various processing conditions.

5. The method for improving gear machining accuracy according to claim 2, characterized in that, The dynamic weight adjustment includes: the dynamic adjustment of the output weights of the physical kernel and the attention mechanism neural network includes: when the cutting force and vibration signal fluctuations are below a preset threshold in the stable cutting stage, the physical kernel output is given a weight value higher than the first weight value to ensure modeling accuracy; when the sensor detects that the cutting force or vibration signal changes exceed the threshold, the output weight of the attention mechanism neural network is increased to a weight value higher than the second weight value within a preset time.

6. The method for improving gear machining accuracy according to claim 1, characterized in that, The active suppression of cutting vibration includes the following steps: S31. Real-time acquisition of vibration signals during the processing; S32. Perform spectral analysis on the vibration signal to identify the key vibration frequency bands that affect the accuracy of the tooth surface; S33. Dynamically adjust damping control parameters based on analysis results to suppress key vibration frequency bands in a targeted manner; S34. Adjust the processing parameters in conjunction with the control and verify the suppression effect to ensure that the vibration amplitude is within the allowable accuracy range.

7. The method for improving gear machining accuracy according to claim 6, characterized in that, The spectral analysis of the vibration signal includes the following steps: S321. Filter the original vibration signal to remove environmental interference noise; S322. The frequency amplitude distribution of the vibration signal is obtained by using the Fast Fourier Transform algorithm; S323. Set the amplitude threshold according to the gear machining accuracy requirements and lock the high-frequency vibration band that exceeds the threshold.

8. The method for improving gear machining accuracy according to claim 6, characterized in that, The dynamic adjustment of damping control parameters based on analysis results includes the following steps: S331. Match the initial damping parameters according to the locked vibration frequency band, and initiate targeted suppression control; S332. Calculate the vibration suppression rate and determine whether it meets the preset requirements; S333. If the suppression rate does not meet the standard, fine-tune the damping parameters through an iterative algorithm until the vibration amplitude drops to the allowable range. S334. Record the optimal parameters and store them in the database.

9. The method for improving gear machining accuracy according to claim 1, characterized in that, Step S4 includes the following steps: S41. Construct a digital twin model that includes machine tools, workpieces, fixtures, and environmental elements; S42. Predict future trends in machining accuracy based on a long short-term memory network model; S43. Feed back the online detected tooth surface error data to the twin model and iteratively optimize the cutting parameters and error compensation coefficients.

10. The method for improving gear machining accuracy according to claim 9, characterized in that, Following step S43, dynamic calibration of the digital twin model is also included, comprising the following steps: S44. Real-time calculation of the residual between the measured values ​​and the twin simulation values ​​of key physical quantities; S45. When the residual continuously exceeds the preset residual threshold and exceeds the preset number of times, it is determined that the twin is inaccurate and the calibration process is triggered. S46. The calibration process uses a reverse identification algorithm to prioritize adjusting the machine tool dynamic parameters and tool wear state parameters in the twin, so that the matching degree between the simulated value and the measured value is restored to within the preset residual threshold.

Images