Dynamic error real-time compensation method and system for heavy-load vertical machining center

Abstract

The invention discloses a dynamic error real-time compensation method and system for a heavy-load vertical machining center, and belongs to the technical field of high-end numerical control equipment, precision manufacturing and intelligent control. Inputting the state vector into a dynamic error model to solve a three-dimensional space dynamic error vector; processing the error vector to generate a real-time compensation instruction; a compensation instruction is injected into the numerical control system to correct the machining track online; and obtaining a real error to update the dynamic error model on line. According to the method, the technology of combining multi-physics field data fusion and a neural network agent model is adopted, frequency decoupling and dual-channel compensation injection are performed on errors, and an online model self-optimization feedback closed loop is established, so that real-time and high-precision compensation on multi-source coupling dynamic errors such as thermal-induced and force-induced multi-source coupling dynamic errors can be realized; and the limit machining precision and stability under the heavy-load machining condition are remarkably improved.

Description

[0001] The application relates to the fields of high-end numerical control equipment, precision manufacturing and intelligent control technology, in particular to a dynamic error real-time compensation method and system for a heavy-duty vertical machining center. BACKGROUND

[0002] The heavy-duty vertical machining center is a core machining equipment in national key fields such as aerospace, energy equipment and large molds, and is used for completing precision machining of large, heavy and complex structure workpieces. The final contour accuracy and surface quality of the workpiece are extremely high for such machining tasks, and the accuracy and stability of the machining process are the basis for guaranteeing the product quality of high-end manufacturing industry.

[0003] However, in the actual machining process, the machining center will produce complex dynamic errors due to the coupling of multiple physical field factors such as spindle high-speed rotation heating, cutting force change and heavy workpiece load. The traditional static geometric error compensation method is difficult to cope with the real-time changes of such errors with time and working conditions, and some existing dynamic compensation schemes still have limitations in real-time and accurate decoupling and comprehensive compensation of multiple error sources, which restricts the improvement of the limit machining precision of the machine tool. SUMMARY

[0004] To solve the above problems, the application provides a dynamic error real-time compensation method and system for a heavy-duty vertical machining center, which adopts a technology combining multi-physical field data fusion and a neural network proxy model, performs frequency decoupling and double-channel compensation injection on the error, and establishes an online model self-optimization feedback closed loop, so as to realize real-time and high-precision compensation of multi-source coupled dynamic errors such as thermal and force-induced errors, and significantly improve the limit machining precision and stability under heavy-duty machining conditions.

[0005] The above object can be achieved by the following scheme: The dynamic error real-time compensation method for the heavy-duty vertical machining center comprises the following steps: collecting multi-source data in real time through a multi-physical field sensor and fusing the data to form a state vector; inputting the state vector into a preset multi-physical field dynamic error model for real-time calculation, calculating and generating a three-dimensional space dynamic error vector; performing reverse conversion and format processing on the three-dimensional space dynamic error vector to generate a real-time compensation instruction; injecting the real-time compensation instruction into a numerical control system of the machining center to perform online correction of a path and generate an actual machining trajectory; acquiring an actual error measurement value, comparing the actual error measurement value with the three-dimensional space dynamic error vector, calculating a model prediction residual, and performing online update on the dynamic error model by using the model prediction residual.

[0006] Optionally, the fusion to form a state vector includes: acquiring the temperature data of the machining center spindle and bed in real time, and performing time domain feature extraction to generate a thermodynamic state feature set; synchronously acquiring the force and vibration data of the machining center column and worktable, and performing frequency domain feature analysis to generate a mechanical dynamic feature set; normalizing and vectorizing the thermodynamic state feature set and the mechanical dynamic feature set to generate a state vector.

[0007] Optionally, the preset multi-physics field dynamic error model includes: collecting historical state vectors and historical spatial error values ​​during historical operation to obtain a historical training data set; using the state vector as input and the spatial error value as output, and using the historical training data set to establish and train a neural network model to obtain a multi-physics field dynamic error model.

[0008] Optionally, the calculation and generation of the three-dimensional spatial dynamic error vector includes: inputting the thermodynamic state feature set into the thermal error prediction subnetwork within the dynamic error model, calculating and generating the thermally induced spatial error component; inputting the mechanical dynamic feature set into the force-induced error prediction subnetwork within the dynamic error model, calculating and generating the force-induced spatial error component; calculating and generating dynamic fusion weights based on the thermodynamic state feature set and the mechanical dynamic feature set, performing weighted summation according to the dynamic fusion weights, and solving the three-dimensional spatial dynamic error vector.

[0009] Optionally, the generating of real-time compensation instructions includes: performing multi-band decomposition on the three-dimensional spatial dynamic error vector to separate and generate low-frequency error components and high-frequency error components; performing path inverse processing on the low-frequency error component to generate a reference path compensation instruction; performing active vibration suppression processing on the high-frequency error component to generate a dynamic damping compensation instruction; synthesizing the reference path compensation instruction and the dynamic damping compensation instruction in real time, and encapsulating them into a CNC system protocol format to generate a real-time compensation instruction.

[0010] Optionally, the online correction path includes: performing spectral analysis on the time series of the high-frequency error component to identify and extract the main vibration frequency; performing calculations through a digital filter based on the main vibration frequency to construct and generate an adaptive notch filter; filtering the high-frequency error component through the adaptive notch filter to generate a dynamic damping compensation instruction.

[0011] Optionally, the online path correction includes: converting the reference path compensation instruction into a coordinate system bias signal, and applying the coordinate system bias signal to the external workpiece coordinate system of the CNC system to achieve quasi-static trajectory correction; converting the dynamic damping compensation instruction into a pulse correction signal, and injecting the pulse correction signal into the servo drive loop of the CNC system to achieve real-time active suppression.

[0012] Optionally, the online correction path includes: periodically measuring the actual spatial position of the motion execution component of the machining center to obtain an actual error measurement value; performing vector calculation on the actual error measurement value and the three-dimensional spatial dynamic error vector predicted by the dynamic error model at the same time to obtain a model prediction residual; and based on the model prediction residual, using an algorithm to iteratively correct the network weights of the dynamic error model.

[0013] Optionally, the method also includes: based on the model prediction residual, performing time series correlation analysis on the thermal-induced spatial error component and the force-induced spatial error component to calculate the correlation coefficient; based on the correlation coefficient, performing decoupling calculation to generate error source attribution weights to update the dynamic error model online.

[0014] Based on the same inventive concept, the present invention also provides a real-time dynamic error compensation system for a heavy-load vertical machining center, the system comprising: a multi-physics field state perception module, for collecting multi-source data in real time through multi-physics field sensors, and fusing them to form a state vector; a dynamic error real-time solution module, for inputting the state vector into a preset multi-physics field dynamic error model for real-time solution, calculating and generating a three-dimensional space dynamic error vector; a compensation instruction generation module, for performing inverse conversion and formatting processing on the three-dimensional space dynamic error vector, and generating a real-time compensation instruction; an instruction injection and execution module, for injecting the real-time compensation instruction into the numerical control system of the machining center, performing online path correction, and generating an actual machining trajectory; a model online self-optimization module, for obtaining actual error measurement values, comparing the actual error measurement values ​​with the three-dimensional space dynamic error vectors, calculating the model prediction residuals, and using the model prediction residuals to perform online update of the dynamic error model.

[0015] Compared with the prior art, the present invention has the following advantages: 1. By integrating real-time data from multiple sources, such as temperature, force, and vibration, and solving it using a multi-physics field dynamic error model, this method breaks through the limitation of traditional static compensation that can only handle geometric errors. It can accurately predict the real-time dynamic errors caused by the coupling of multiple physical factors such as thermal deformation and cutting force, greatly improving the comprehensiveness and accuracy of compensation. 2. By frequency decoupling the errors, slow thermal drift and high-speed vibration errors are separated and compensated using coordinate system bias and servo loop injection. This frequency-divided, dual-channel compensation strategy achieves efficient, feedforward suppression of errors with different characteristics, significantly enhancing the dynamic stability and surface quality of the machining process. 3. The dynamic error model is updated online by periodically acquiring actual errors and calculating model prediction residuals. This mechanism enables the compensation system to autonomously learn and automatically compensate for precision drift caused by machine wear or environmental changes, ensuring long-term precision retention and enabling the self-evolution of the compensation model.

[0016] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 It is a flow chart of a method for real-time compensation of dynamic errors of a heavy-load vertical machining center according to an embodiment of the present invention.

[0019] Figure 2 2 is a diagram showing the composition and evolution of dynamic error components according to an embodiment of the present invention.

[0020] Figure 3 4 is a spectrum comparison and analysis diagram of the active vibration suppression process according to an embodiment of the present invention.

[0021] Figure 4 It is a three-dimensional convergence trajectory diagram of the model prediction residual in an embodiment of the present invention.

[0022] Figure 5 It is a structural diagram of a real-time dynamic error compensation system for a heavy-load vertical machining center according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0024] Reference Figure 1One embodiment of the present invention proposes a real-time compensation method for dynamic errors of a heavy-load vertical machining center. It adopts a technology that combines multi-physics field data fusion with a neural network proxy model. By performing frequency decoupling and dual-channel compensation injection on the errors, and establishing an online model self-optimization feedback closed loop, it can achieve real-time, high-precision compensation for multi-source coupled dynamic errors such as thermal and force-induced errors, significantly improving the ultimate machining accuracy and stability under heavy-load machining conditions.

[0025] The method of this embodiment specifically includes: Multi-source data is collected in real time through multi-physics field sensors and fused to form a state vector; Inputting the state vector into a preset multi-physics field dynamic error model for real-time solution, calculating and generating a three-dimensional space dynamic error vector; Performing inverse conversion and formatting processing on the three-dimensional spatial dynamic error vector to generate a real-time compensation instruction; Injecting the real-time compensation instruction into the numerical control system of the machining center to perform online path correction and generate the actual machining trajectory; An actual error measurement value is obtained, the actual error measurement value is compared with the three-dimensional space dynamic error vector, a model prediction residual is calculated, and the dynamic error model is updated online using the model prediction residual.

[0026] By adopting a technology that combines multi-physics field data fusion with a neural network proxy model, frequency decoupling and dual-channel compensation injection are performed on the errors, and an online model self-optimization feedback loop is established. This enables real-time, high-precision compensation for multi-source coupled dynamic errors such as thermal and force-induced errors, significantly improving the ultimate machining accuracy and stability under heavy-load machining conditions.

[0027] Optionally, the combining and fusing to form a state vector includes: Acquire the temperature data of the machining center spindle and bed in real time, extract time-domain features, and generate a thermodynamic state feature set; Specifically, in the step of acquiring temperature data in real time and performing time-domain feature extraction to generate a thermodynamic state feature set, in order to fully grasp the thermodynamic state of the machine tool, multiple temperature sensors, such as Pt100 platinum resistance thermometers or K-type thermocouples, are deployed at key heat sources and heat-sensitive parts of the machining center, such as spindle bearings, motor housings, and bed guide rails. A data acquisition unit records the temperature data stream of these measuring points in real time at a high frequency. In order to reflect more than just the instantaneous temperature, a feature extraction process also performs time-domain analysis on the temperature data stream, calculating multiple time-domain features including the current temperature value, the temperature change rate within the last minute, and the temperature sliding average. These multiple characteristic values ​​that can fully describe the current state and change trend of the thermal field together constitute the thermodynamic state feature set.

[0028] Synchronously acquire force and vibration data of the machining center column and worktable, perform frequency domain feature analysis, and generate a mechanical dynamic feature set; Specifically, during the steps of simultaneously acquiring force and vibration data and performing frequency domain feature analysis to generate a mechanical dynamic feature set, strain gauges are installed on key structural components such as the column and worktable to indirectly measure cutting forces to capture the force and vibration characteristics of the machining process. A triaxial accelerometer is also installed on the spindle end. The collected high-frequency vibration signal is converted from the time domain to the frequency domain by performing a Fast Fourier Transform (FFT). Key frequency domain features are then extracted from the transformed spectrum, such as the amplitude of one or more dominant frequencies and the energy of characteristic frequencies corresponding to machining chatter. These frequency domain features, along with the quasi-static and dynamic cutting force components calculated from the strain gauge data, constitute the mechanical dynamic feature set.

[0029] The thermodynamic state feature set and the mechanical dynamic feature set are normalized and vectorized to generate a state vector.

[0030] Specifically, in the step of normalizing and vectorizing the thermodynamic state feature set and the mechanical dynamic feature set to generate the state vector, in order to enable the features of different physical dimensions and numerical ranges to be effectively processed by the subsequent dynamic error model, this step first normalizes all elements of the thermodynamic state feature set and the mechanical dynamic feature set generated in the first two steps. A commonly used normalization method is minimum-maximum scaling, as shown in the formula: , in, is the normalized eigenvalue, whose range is [0, 1]; is the original eigenvalue; and The minimum and maximum values ​​of this feature appearing in the historical data. After all eigenvalues ​​are normalized, they are concatenated in a preset order into a single, high-dimensional vector. This vector is the final state vector, which comprehensively and digitally represents the complete multi-physics field working condition of the machine tool at the current moment.

[0031] Exemplarily, in a heavy load cutting process, a data acquisition process obtains at a certain time: the real-time temperature of the main shaft front bearing is 65.2℃, and the temperature change rate is 0.2℃ / min; at the same time, the frequency domain analysis of the accelerometer shows that there is a significant vibration peak at 350Hz, and the amplitude is 0.5g, and the current cutting force calculated by the strain gauge is 1200N. A feature processing process will normalize these physical quantities (65.2, 0.2, 350, 0.5, 1200,...) according to the formula, and obtain a set of dimensionless values such as [0.85, 0.32,..., 0.91, 0.75]. These values are sequentially spliced to form a state vector that can represent the "signs" of the machine tool at this moment, and are immediately transmitted to the next step of the dynamic error model for calculation.

[0032] Optionally, the preset multi-physical field dynamic error model comprises: In the historical running process, the historical state vector and the historical space error value are collected to obtain the historical training data set; Specifically, in the step of collecting the historical state vector and the historical space error value to obtain the historical training data set, this step is usually completed in a controlled offline calibration stage. In this stage, the machining center is driven to execute a series of specially designed motion trajectories that can fully stimulate its dynamic error, such as circular interpolation or S-shaped curve motion at high speed. During the entire running process, on the one hand, a plurality of state vectors containing thermodynamic and mechanical dynamic characteristics are synchronously collected by multi-physical field sensors. On the other hand, an external high-precision measuring device, such as a laser tracker or a ball bar, is used to measure the actual three-dimensional space position of the machining center motion execution component in real time, which is strictly clock-synchronized with the sensor data acquisition. By comparing the actual measured position with the command position of the numerical control system, the actual space error value corresponding to each state vector can be obtained. The collection of all these working condition-error pairing data constitutes the historical training data set for model training.

[0033] A neural network model is established and trained using the historical training data set, taking the state vector as the input and the space error value as the output, to obtain the multi-physical field dynamic error model.

[0034] Specifically, in the step of establishing and training a neural network model to obtain a multi-physics field dynamic error model, a neural network architecture that can handle temporal dependencies, such as a long short-term memory (LSTM) network or a gated recurrent unit (GRU) network, is used to construct the dynamic error model. The training process uses the state vector sequence in the historical training data set as the input of the model and the corresponding actual spatial error value as the expected output or label of the model. Through the backpropagation algorithm, the weights and bias parameters of the network are iteratively optimized with the goal of minimizing the mean square error loss function between the model prediction error and the actual error. A mean square error loss function can be expressed by the formula: , in, is the mean square error loss value; is the total number of training samples; is the error vector predicted by the model for the i-th state vector; is the actual spatial error value recorded in the historical training dataset corresponding to the i-th state vector. When this loss value converges to a sufficiently small preset value, training is complete, and this set of optimized network parameters is solidified into the final multiphysics dynamic error model, which can be used in the real-time solution phase.

[0035] For example, an engineer performed offline calibration on a heavy-load vertical machining center. He mounted a laser tracker target on the machine's spindle and then wrote a G-code program that instructed the machine to execute complex helical motions within the worktable, continuously varying the feed rate and spindle speed. During the 10-minute program run, a data acquisition system simultaneously recorded data from sensors located throughout the machine tool and the real-time spatial errors measured by the laser tracker at a frequency of 100 Hz. This data was compiled into a historical training dataset. A neural network consisting of two layers of LSTM units was then trained using this dataset. After several hours of training, training ceased when the model's mean squared error loss fell below 1 square micron. The resulting set of trained network weights was saved and used as a multiphysics dynamic error model for real-time compensation.

[0036] Optionally, calculating and generating a three-dimensional space dynamic error vector includes: Inputting the thermodynamic state feature set into the thermal error prediction subnetwork within the dynamic error model to calculate and generate a thermally induced spatial error component; Specifically, in the step of inputting the thermodynamic state feature set into the thermal error prediction subnetwork to calculate and generate the thermally induced spatial error component, this step is designed to specifically deal with the relatively slowly changing geometric drift caused by machine tool heating. The thermal error prediction subnetwork is a specialized component of the neural network model. This subnetwork receives as input the thermodynamic state feature set that represents the temperature and its changing trend, and outputs a three-dimensional thermally induced spatial error component through its internal trained nonlinear mapping relationship. This component mainly reflects the slow, quasi-static position deviation of the machining center's motion actuators in the X, Y, and Z directions due to the thermal expansion of components such as the spindle, motor, and ball screw.

[0037] Inputting the mechanical dynamic feature set into the force-induced error prediction subnetwork within the dynamic error model to calculate and generate a force-induced spatial error component; Specifically, the step of inputting the mechanical dynamic feature set into the force-induced error prediction subnetwork to calculate and generate the force-induced spatial error component is designed to specifically address the rapidly changing dynamic errors caused by cutting forces, gravity changes, and vibrations. The force-induced error prediction subnetwork is another specialized component of the neural network model. It receives as input a mechanical dynamic feature set that characterizes the force and vibration spectrum characteristics. After training, this subnetwork can highly dynamically predict structural elastic deformation and high-frequency position errors caused by factors such as tool forces, workpiece and worktable gravity changes, and machining chatter based on real-time force and vibration signals, and output a three-dimensional force-induced spatial error component.

[0038] Based on the thermodynamic state feature set and the mechanical dynamic feature set, dynamic fusion weights are calculated and generated, and weighted summation is performed according to the dynamic fusion weights to solve the three-dimensional space dynamic error vector.

[0039] Specifically, in the step of calculating the dynamic fusion weight based on the thermodynamic and mechanical dynamic feature sets, and performing weighted summation to solve the three-dimensional space dynamic error vector, this step embodies the innovative idea of ​​intelligently fusing errors from different sources. A working condition discrimination process analyzes the significance of the thermodynamic state feature set and the mechanical dynamic feature set in real time. For example, when the force or vibration amplitude in the mechanical dynamic feature set exceeds the preset threshold, it is determined that the current working condition is "force-dominated" such as heavy cutting; otherwise, it is determined to be a "heat-dominated" working condition such as finishing or idling. Based on this judgment result, a function is used to generate a set of dynamically changing fusion weights. For example, the Softmax function is used to calculate the weights, as shown in the formula: , in, and are the dynamic fusion weights of the thermal error and force error components, and their sum is 1; and are the working condition significance scores calculated based on the thermodynamic and mechanical dynamic feature sets, respectively. Finally, the total three-dimensional dynamic error vector is calculated by weighted summing the two error components, as shown in the following formula: , in, is the final solved three-dimensional dynamic error vector; and are the thermally and mechanically induced spatial error components generated in the first two steps, respectively. Figure 2 As shown in the figure, in the form of a stacked area chart, it clearly shows how the total dynamic error in a simulated machining process is composed of the dynamic superposition of the slowly growing thermal error component and the force-induced error component that is only significant in the cutting stage. The area of ​​each region in the figure intuitively reflects the contribution of different error sources at any time.

[0040] For example, in a machining task, the machine tool first performs rough machining with large cutting depth and high load. At this time, a working condition discrimination process detects that the cutting force signal in the mechanical dynamic feature set is very strong, and the calculated force working condition significance score is Much larger than the thermal condition significance score , according to the generated dynamic fusion weights may be =0.9, =0.1. When solving the total error, the contribution of the force-induced error component is dominant. Subsequently, the machine tool enters the high-speed, small-depth finishing stage, where the cutting force is greatly reduced, but the spindle temperature continues to rise due to long-term operation. At this time, the working condition discrimination process will dynamically adjust the fusion weight to, for example, =0.2, = 0.8, which makes the contribution of the thermally induced error component the main part of the total error prediction.

[0041] Optionally, generating a real-time compensation instruction includes: Performing multi-band decomposition on the three-dimensional spatial dynamic error vector to separate and generate a low-frequency error component and a high-frequency error component; Specifically, in the step of performing multi-band decomposition on the three-dimensional dynamic error vector, this step aims to separate error sources with different physical characteristics. A signal processing process uses a digital filter bank to process the received three-dimensional dynamic error vector in the form of a time series. The filter bank contains at least a low-pass filter (LPF) and a high-pass filter (HPF). The low-pass filter is used to extract the slowly changing trend portion of the error signal, which mainly corresponds to the machine tool structure drift caused by thermal effects. Its output is the low-frequency error component. The high-pass filter is used to extract the rapidly changing portion of the signal, which mainly corresponds to the machining chatter caused by sudden changes in cutting force or structural resonance. Its output is the high-frequency error component.

[0042] Performing path inverse processing on the low-frequency error component to generate a reference path compensation instruction; Specifically, in the step of performing path inversion processing on the low-frequency error component to generate the reference path compensation instruction, the processing logic of this step is relatively straightforward. Since the low-frequency error component represents the overall, quasi-static geometric position deviation of the machine tool, the method for compensating it is to simply perform inverse processing on the error vector. The calculation process is shown in the formula: , in, is the generated reference path compensation instruction; This is the low-frequency error component separated in the previous step. This command will be used to make an overall and smooth correction to the nominal tool path.

[0043] performing active vibration suppression processing on the high-frequency error component to generate a dynamic damping compensation instruction; Specifically, the step of actively suppressing vibrations of high-frequency error components to generate dynamic damping compensation instructions employs an adaptive control strategy based on real-time spectrum analysis. First, spectrum analysis, such as a fast Fourier transform, is performed on the time series of the high-frequency error components to identify and extract the dominant vibration frequency under the current machining state in real time. Subsequently, based on this dominant vibration frequency, a digital filter design algorithm is used to construct and generate an adaptive notch filter capable of accurately suppressing this dominant vibration frequency. Finally, the high-frequency error components are filtered through this custom-designed adaptive notch filter, the output of which is a dynamic damping compensation instruction capable of actively canceling or damping machining chatter.

[0044] The reference path compensation instruction and the dynamic damping compensation instruction are synthesized in real time and packaged into a numerical control system protocol format to generate a real-time compensation instruction.

[0045] Specifically, during the real-time synthesis and packaging of the baseline path compensation and dynamic damping compensation instructions, a command synthesis process is responsible for fusing the two compensation instructions generated by the previous two steps. By performing vector addition on the baseline path compensation instruction representing slow drift compensation and the dynamic damping compensation instruction representing high-speed vibration suppression, a final composite compensation vector containing full-band error correction information is generated. This composite compensation vector is then encapsulated into a standard data packet according to the communication protocol requirements of the specific CNC system. This encapsulated data packet is the final real-time compensation instruction that can be directly parsed and executed by the CNC system.

[0046] For example, the total error vector predicted by a dynamic error model at time t appears as a slowly decreasing ramp superimposed with a 250Hz sine wave. A digital filter bank successfully decomposes this signal into a low-frequency error component representing the ramp and a high-frequency error component representing the sine wave. A path processing process inverts the low-frequency component to generate a slowly increasing reference path compensation command. Simultaneously, a vibration suppression process identifies the main vibration frequency of 250Hz and generates a dynamic damping compensation command that specifically filters out the 250Hz frequency band. Finally, these two commands are combined into a single compensation data stream and packaged into a format recognizable by the CNC system, ready for injection into the servo control loop.

[0047] Optionally, the online correction path includes: Performing spectrum analysis on the time series of the high-frequency error component to identify and extract the main vibration frequency; Specifically, the step of performing spectral analysis on the time series of high-frequency error components to identify and extract the dominant vibration frequency aims to diagnose the dominant and most damaging flutter frequency in real time from a complex vibration signal. By applying a frequency domain transformation algorithm to the separated time series data of the high-frequency error components, the signal is converted from a time domain signal into a spectrum that displays the energy or amplitude of each frequency component. By searching for peak points on this spectrum, the frequency with the highest energy at the current moment is identified and extracted, and this frequency is determined to be the dominant vibration frequency that needs to be actively suppressed.

[0048] According to the master oscillation frequency, a digital filter is used to perform calculations to construct and generate an adaptive notch filter; Specifically, in the step of constructing and generating an adaptive notch filter based on the master oscillation frequency, this step aims to design a digital filter specifically for eliminating vibrations of a specific frequency. A digital filter design process uses the master oscillation frequency identified in the previous step as the core design parameter. The goal of this process is to generate a transfer function for a second-order infinite impulse response (IIR) notch filter whose zero point in the digital domain is precisely set at the master oscillation frequency, thereby achieving maximum attenuation of signals at that frequency. This transfer function can be expressed by the formula: , in, is the transfer function of the generated adaptive notch filter; is the gain factor, which is used to ensure that signals of other frequencies can pass through without loss; It is the digital angular frequency normalized according to the master oscillation frequency; is a pole radius close to 1, which controls the width of the notch. It is dynamically set according to the real-time master frequency, so the generated notch filter is adaptive.

[0049] The high-frequency error component is filtered through the adaptive notch filter to generate a dynamic damping compensation instruction.

[0050] Specifically, in the step of filtering the high-frequency error component through an adaptive notch filter to generate a dynamic damping compensation instruction, this step is the actual execution of the vibration "filtering" operation. The original time series data of the high-frequency error component is used as the input signal and convolved through the adaptive notch filter generated in the previous step. Since the filter is precisely designed to attenuate only the main vibration frequency, its output signal will be a new time series that retains all other frequency components but significantly suppresses the main vibration frequency component. This "targeted" and smoother signal is the final generated dynamic damping compensation instruction, which will be used in subsequent steps to offset the real-time vibration of the machine tool, such as Figure 4 As shown in the figure, by comparing the confidence interval and mean of the signal spectrum before and after filtering, it is intuitively demonstrated that the adaptive filtering method of the present invention can not only accurately eliminate the peak of the main vibration frequency, but also significantly improve the overall stability of the signal.

[0051] Exemplarily, in a high-precision side milling process, a high-frequency error component is obtained from the vibration signal collected by an accelerometer. A spectrum analysis process is performed on the signal to find a sharp peak at 420 Hz, and the frequency of 420 Hz is determined as the main vibration frequency. A filter design process is immediately performed to construct an adaptive notch filter that specifically suppresses the vibration at 420 Hz, by calculating a set of specific filter coefficients. The original high-frequency error component signal is greatly weakened in the vibration component at 420 Hz after passing through the filter. The "purified" signal is used as a dynamic damping compensation instruction, which is finally injected into the servo loop to effectively suppress the machining chatter and ensure the surface quality of the workpiece.

[0052] Optionally, the online correction path includes: The reference path compensation instruction is converted into a coordinate system offset signal, and the coordinate system offset signal is applied to an external workpiece coordinate system of the numerical control system to realize quasi-static trajectory correction. Specifically, in the step of converting the reference path compensation instruction into a coordinate system offset signal and applying it to the external workpiece coordinate system, this step aims to compensate for the quasi-static error caused by factors such as thermal drift and slow changes. In this embodiment, the external workpiece coordinate system offset function commonly provided by numerical control systems is used. The generated reference path compensation instruction is a slowly changing three-dimensional vector, which is sent to the PLC (Programmable Logic Controller) or CNC kernel of the machining center in real time through a high-speed input / output interface. After the instruction is parsed by the PLC program, its components on the X, Y, and Z axes are written into the corresponding external coordinate system offset registers. When the numerical control system executes the G code program, it automatically adds the offset values in the registers to the program coordinate values, thereby realizing overall and smooth correction of the nominal tool path. This process can be represented by the formula: , wherein, is the final tool instruction coordinate after quasi-static trajectory correction; is the instruction coordinate in the original G code program; is the coordinate system offset signal converted from the reference path compensation instruction.

[0053] The dynamic damping compensation instruction is converted into a pulse correction signal, and the pulse correction signal is injected into the servo drive loop of the numerical control system to realize real-time active suppression.

[0054] Specifically, in the step of converting the dynamic damping compensation instruction into a pulse correction signal and injecting it into the servo drive loop, this step aims to achieve real-time active suppression of high-frequency dynamic errors such as machining chatter. This process adopts a strategy of directly intervening in the servo drive feedback loop. The generated dynamic damping compensation instruction is a high-frequency, rapidly changing correction signal. This signal is converted into a pulse correction signal stream compatible with the servo motor encoder pulse format through a high-speed data output card. This pulse correction signal stream is superimposed in real time with the original position feedback pulse signal returned from the servo motor encoder in a hardware adder. The synthesis principle can be expressed by the formula: , in, It is the corrected feedback signal that finally enters the servo controller; It is the original feedback pulse signal output by the encoder; This is a pulse correction signal converted from the dynamic damping compensation command. In this way, the servo controller "mistakenly" believes that an error has occurred in the opposite direction of the actual vibration and immediately outputs a reverse driving torque to offset it, thus achieving active, real-time "damping" of high-frequency vibrations.

[0055] For example, during one machining operation, a compensation system predicted a -10 micron downward Z-axis drift due to thermal effects, coupled with 200Hz high-frequency vibrations from cutting chatter. Two instructions were generated: a baseline path compensation instruction with the value (0, 0, +10μm); and a dynamic damping compensation instruction containing a 200Hz inverted waveform. The first instruction was sent to the CNC via the I / O port, setting the Z-axis offset register value of the external workpiece coordinate system to +0.010mm. All subsequent Z-axis instructions were automatically raised by 10 microns. Simultaneously, the second instruction was converted into a high-frequency pulse stream and superimposed with the feedback pulses from the Z-axis motor encoder. Upon sensing this feedback signal, the servo drive instantly adjusted the motor torque to actively suppress the 200Hz physical vibrations, thereby ensuring a high-quality machined surface finish.

[0056] Optionally, the online correction path includes: Periodically measure the actual spatial position of the machining center's motion execution components to obtain actual error measurement values; Specifically, in the step of periodically measuring the actual spatial position of the moving actuator of the machining center to obtain the actual error measurement value, this step is intended to provide a high-precision "real-world" benchmark for the self-correction of the model. This measurement is usually performed during non-cutting machining intervals, such as the interval between turning the workpiece or changing the tool. A laser tracker or laser interferometer with high sampling frequency and micron-level accuracy installed outside the working area of ​​the machine tool is used to perform fast, non-contact three-dimensional spatial position measurement of a fixed reference point of the moving actuator of the machining center. By comparing this measured position with the theoretical command position in the CNC system, an actual error measurement value representing the comprehensive error of the machine tool under the current thermal and mechanical state can be obtained.

[0057] Performing vector calculation on the actual error measurement value and the three-dimensional dynamic error vector predicted by the dynamic error model at the same time to obtain a model prediction residual; Specifically, in the step of performing vector calculations between the actual error measurements and the three-dimensional dynamic error vector to obtain the model prediction residual, the core of this step is to accurately quantify the deviation between the model prediction and physical reality. A data synchronization process ensures that the actual error measurements obtained from the external measurement device and the predicted error vector output by the dynamic error model have exactly the same timestamp. Subsequently, a vector subtraction operation is performed to calculate the model prediction residual, as shown in the formula: , in, is the calculated model prediction residual vector; It is the actual error measurement value measured by a high-precision measuring device; is the three-dimensional dynamic error vector predicted by the dynamic error model at the same time. The magnitude and direction of this residual vector accurately indicate the deficiencies in the current model prediction.

[0058] Based on the model prediction residuals, an algorithm is used to iteratively correct the network weights of the dynamic error model.

[0059] Specifically, in the step of iteratively correcting the network weights of the dynamic error model based on the model prediction residual using an algorithm, this step is the execution link for realizing the self-evolution of the model. The calculated model prediction residual is used as the new error signal, and the gradient of the residual relative to the weights of each layer of the neural network is calculated through the backpropagation algorithm. Then, an online optimization algorithm, such as Stochastic Gradient Descent with Momentum (SGDM), is used to perform one or several small iterative corrections on the network weights of the trained dynamic error model. The core idea of ​​its weight update can be expressed by the formula: , in, is the updated network weight matrix; is the current network weight matrix; is the online learning rate, whose size can be dynamically adjusted according to the residual; is the gradient of the loss function with respect to the current weight, which is directly related to the model prediction residual. Through this step, the model can learn from its own prediction errors and continuously adapt to the characteristic drift of the machine tool due to factors such as wear and aging, such as Figure 4 As shown in the figure, the three-dimensional space trajectory line vividly shows how the model's prediction residual vector gradually spirals down in the three-dimensional space as the number of online updates increases, and finally converges and stabilizes within the target sphere representing the high-precision interval, proving the effectiveness and convergence of the online update method.

[0060] For example, after a machining center has been running continuously for two hours, the spindle temperature has increased significantly. During a tool change, a laser tracker measured the actual Z-axis error of the spindle end to be -35μm. However, at the same time, the dynamic error model predicted a Z-axis error of -30μm based on real-time sensor data. A calculation process concluded that the model prediction residual in the Z-axis direction was -5μm. This residual was then used to drive an online correction. A stochastic gradient descent algorithm calculated the gradient based on the -5μm residual and fine-tuned the heat-related weight parameters in the neural network. After this correction, the model "learned" that under the current thermal state, its prediction of Z-axis thermal extension needs to be appropriately increased. In subsequent processing, the model's prediction will be closer to the actual -35μm, thereby achieving an adaptive improvement in compensation accuracy.

[0061] Optionally, the method further includes: Based on the model prediction residual, a time series correlation analysis is performed on the thermal-induced spatial error component and the force-induced spatial error component to calculate a correlation coefficient; Specifically, the step of performing time series correlation analysis to calculate the correlation coefficient aims to intelligently determine which physical error source the total model prediction residual is more similar to in terms of dynamic characteristics. A time series analysis process extracts the model prediction residual sequence within the most recent time window and performs cross-correlation calculations with the thermal-induced spatial error component sequence and the force-induced spatial error component sequence generated within the same time window. The cross-correlation calculation quantifies the degree of similarity between the two time series in terms of morphology and phase. The results are two correlation coefficients: one for the correlation between the residual and the thermal error, and the other for the correlation between the residual and the force error.

[0062] Based on the correlation coefficient, a decoupling calculation is performed to generate an error source attribution weight to update the dynamic error model online.

[0063] Specifically, in the step of performing decoupling calculations to generate error source attribution weights, the purpose of this step is to convert the qualitative correlations obtained in the previous step into quantitative weights that can be used to guide model updates. A normalized exponential function, such as the Softmax function, is used to convert the two correlation coefficients into a set of normalized weights that sum to 1. The calculation process is shown in the formula: , in, and are the final generated error source attribution weights corresponding to thermal error and force-induced error respectively; and is the correlation coefficient between the calculated residual and the thermal and force-induced errors. This set of attribution weights is subsequently used to weight the model prediction residuals during the online update step described in claim 1. For example, when modifying the neural network weights, the majority of the modification will be attributed to or applied to the error prediction subnetwork corresponding to the higher weight, thereby achieving targeted, differentiated online updates of the error model.

[0064] For example, during a machining task, the feedback loop calculates a model prediction residual. A time series analysis process reveals that this residual signal exhibits rapid, burr-like fluctuations synchronized with the cutting force. Therefore, its correlation coefficient with the time series of the force-induced spatial error component is calculated to be a high value of 0.9, while its correlation coefficient with the slowly varying series of the thermal-induced spatial error component is only 0.1. Subsequently, a decoupling calculation process converts these two correlation coefficients into a set of error source attribution weights based on a formula. The result may be: a force-induced error attribution weight of 0.95 and a thermal-induced error attribution weight of 0.05. This means that 95% of the "responsibility" for the deviation in this model prediction lies with the force-induced error prediction subnetwork. Therefore, in subsequent online updates, the vast majority of the corrections will be used to fine-tune the weights of the force-induced error prediction subnetwork, while essentially maintaining the stability of the thermal error prediction subnetwork.

[0065] Based on the same inventive concept, the present invention also provides a real-time dynamic error compensation system for a heavy-load vertical machining center, such as Figure 5 As shown, the system includes: Multi-physics state perception module, which is used to collect multi-source data in real time through multi-physics sensors and fuse them into a state vector; A dynamic error real-time solution module is used to input the state vector into a preset multi-physics field dynamic error model for real-time solution, calculate and generate a three-dimensional dynamic error vector; a compensation instruction generation module, configured to perform inverse conversion and formatting processing on the three-dimensional spatial dynamic error vector to generate a real-time compensation instruction; An instruction injection and execution module is used to inject the real-time compensation instruction into the numerical control system of the machining center, perform online path correction, and generate an actual machining trajectory; The model online self-optimization module is used to obtain actual error measurement values, compare the actual error measurement values ​​with the three-dimensional space dynamic error vector, calculate the model prediction residual, and use the model prediction residual to update the dynamic error model online.

[0066] It should be noted that the functional division and information interaction between the aforementioned modules are logical. Physically, they can be integrated into the same software platform or deployed in a distributed manner. The connections between them represent data and control flows, designed to collaboratively achieve the dynamic optimization of building energy consumption of the present invention. The foregoing description is merely an exemplary embodiment of the present invention and is not intended to limit its scope.

Claims

1. Collect multi-source data in real time through multi-physics field sensors and fuse them to form a state vector; Inputting the state vector into a preset multi-physics field dynamic error model for real-time solution, calculating and generating a three-dimensional space dynamic error vector; Performing inverse conversion and formatting processing on the three-dimensional spatial dynamic error vector to generate a real-time compensation instruction; Injecting the real-time compensation instruction into the numerical control system of the machining center to perform online path correction and generate the actual machining trajectory; An actual error measurement value is obtained, the actual error measurement value is compared with the three-dimensional space dynamic error vector, a model prediction residual is calculated, and the dynamic error model is updated online using the model prediction residual.

2. The real-time dynamic error compensation method for a heavy-load vertical machining center according to claim 1 is characterized in that: The fusion to form a state vector includes: Acquire the temperature data of the machining center spindle and bed in real time, extract time-domain features, and generate a thermodynamic state feature set; Synchronously acquire force and vibration data of the machining center column and worktable, perform frequency domain feature analysis, and generate a mechanical dynamic feature set; The thermodynamic state feature set and the mechanical dynamic feature set are normalized and vectorized to generate a state vector.

3. The real-time dynamic error compensation method for a heavy-load vertical machining center according to claim 1 is characterized in that: The preset multi-physics field dynamic error model includes: During the historical operation process, historical state vectors and historical spatial error values ​​are collected to obtain historical training data sets; With the state vector as input and the spatial error value as output, a neural network model is established and trained using the historical training data set to obtain a multi-physics field dynamic error model.

4. The real-time dynamic error compensation method for a heavy-load vertical machining center according to claim 2 is characterized in that: The calculating and generating of the three-dimensional space dynamic error vector comprises: Inputting the thermodynamic state feature set into the thermal error prediction subnetwork within the dynamic error model to calculate and generate a thermally induced spatial error component; Inputting the mechanical dynamic feature set into the force-induced error prediction subnetwork within the dynamic error model to calculate and generate a force-induced spatial error component; Based on the thermodynamic state feature set and the mechanical dynamic feature set, dynamic fusion weights are calculated and generated, and weighted summation is performed according to the dynamic fusion weights to solve the three-dimensional space dynamic error vector.

5. The real-time dynamic error compensation method for a heavy-load vertical machining center according to claim 1 is characterized in that: Generating a real-time compensation instruction comprises: Performing multi-band decomposition on the three-dimensional spatial dynamic error vector to separate and generate a low-frequency error component and a high-frequency error component; Performing path inverse processing on the low-frequency error component to generate a reference path compensation instruction; performing active vibration suppression processing on the high-frequency error component to generate a dynamic damping compensation instruction; The reference path compensation instruction and the dynamic damping compensation instruction are synthesized in real time and packaged into a numerical control system protocol format to generate a real-time compensation instruction.

6. The real-time dynamic error compensation method for a heavy-load vertical machining center according to claim 5, characterized in that: The online correction path includes: Performing spectrum analysis on the time series of the high-frequency error component to identify and extract the main vibration frequency; According to the master oscillation frequency, a digital filter is used to perform calculations to construct and generate an adaptive notch filter; The high-frequency error component is filtered through the adaptive notch filter to generate a dynamic damping compensation instruction.

7. The real-time dynamic error compensation method for a heavy-load vertical machining center according to claim 5, characterized in that: The online correction path includes: Converting the reference path compensation instruction into a coordinate system offset signal, and applying the coordinate system offset signal to the external workpiece coordinate system of the numerical control system to achieve quasi-static trajectory correction; The dynamic damping compensation instruction is converted into a pulse correction signal, and the pulse correction signal is injected into the servo drive loop of the numerical control system to achieve real-time active suppression.

8. The real-time dynamic error compensation method for a heavy-load vertical machining center according to claim 1 is characterized in that: The online correction path includes: Periodically measure the actual spatial position of the machining center's motion execution components to obtain actual error measurement values; Performing vector calculation on the actual error measurement value and the three-dimensional dynamic error vector predicted by the dynamic error model at the same time to obtain a model prediction residual; Based on the model prediction residuals, an algorithm is used to iteratively correct the network weights of the dynamic error model.

9. The real-time dynamic error compensation method for a heavy-load vertical machining center according to claim 4, characterized in that: The method further comprises: Based on the model prediction residual, a time series correlation analysis is performed on the thermal-induced spatial error component and the force-induced spatial error component to calculate a correlation coefficient; Based on the correlation coefficient, a decoupling calculation is performed to generate an error source attribution weight to update the dynamic error model online.

10. A real-time dynamic error compensation system for a heavy-load vertical machining center, applied to a real-time dynamic error compensation method for a heavy-load vertical machining center according to any one of claims 1 to 9, characterized in that: The system comprises: Multi-physics state perception module, which is used to collect multi-source data in real time through multi-physics sensors and fuse them into a state vector; A dynamic error real-time solution module is used to input the state vector into a preset multi-physics field dynamic error model for real-time solution, calculate and generate a three-dimensional space dynamic error vector; a compensation instruction generation module, configured to perform inverse conversion and formatting processing on the three-dimensional spatial dynamic error vector to generate a real-time compensation instruction; An instruction injection and execution module is used to inject the real-time compensation instruction into the numerical control system of the machining center, perform online path correction, and generate an actual machining trajectory; The model online self-optimization module is used to obtain actual error measurement values, compare the actual error measurement values ​​with the three-dimensional space dynamic error vector, calculate the model prediction residual, and use the model prediction residual to update the dynamic error model online.

Images