A multi-source error dynamic compensation method for numerical control machine tools under complex working conditions

Abstract

This invention discloses a dynamic compensation method for multi-source errors in CNC machine tools under complex working conditions, belonging to the field of precision control technology for CNC machine tools. This method constructs a distributed sensor network to collect multi-physics field data such as temperature field, cutting force, vibration, and environmental parameters in real time, establishes a geometric-thermal-mechanical coupling error propagation model, and uses an improved variational mode decomposition algorithm to achieve high-precision decoupling of multi-source errors. Based on a bidirectional LSTM deep learning network, dynamic compensation amounts are predicted, and combined with a feedforward-feedback composite controller and adaptive coupling matrix adjustment, multi-axis collaborative compensation commands are generated. The compensation effect is evaluated online using a laser interferometer, triggering dynamic updates of model parameters and forming a closed-loop optimization mechanism. This invention overcomes the limitations of traditional single error compensation, solves the problem of real-time analysis and dynamic suppression of multi-physics field coupled errors under complex working conditions, significantly improves the machining accuracy and adaptive capability of CNC machine tools, and is suitable for high-precision machining scenarios such as aerospace precision parts and optical molds.

Description

Technical Field

[0001] This invention belongs to the field of precision control technology for CNC machine tools, specifically relating to a dynamic compensation method for multi-source errors of CNC machine tools under complex working conditions. It is particularly aimed at the multi-physics field coupling problem of geometric errors, thermal errors, and dynamic errors, and achieves dynamic improvement of machine tool accuracy through multi-source data fusion, error decoupling modeling, and intelligent compensation control. Background Technology

[0002] With the increasing demand for high-precision machining, error compensation technology for CNC machine tools has become a key means to improve machining accuracy. Traditional error compensation methods are mostly based on modeling a single error source (such as thermal or geometric error) and adopt offline calibration and static compensation strategies, without considering the coupling effects of dynamic factors such as cutting force and vibration. Although such methods are effective under stable working conditions, in complex machining scenarios with varying working conditions and multiple disturbances, the compensation accuracy decreases significantly due to the neglect of multi-physics coupling effects.

[0003] Studies have shown that machine tool errors are a nonlinear system involving multiple sources of coupled errors, including geometric errors (guide rail wear, assembly deviations), thermal errors (spindle heating, environmental temperature changes), and dynamic errors (cutting force disturbances, vibration excitation). Existing technologies, such as multi-sensor data fusion, generate compensation quantities by weighted superposition of error components, but they do not address the nonlinearity of error coupling and propagation. Especially during long-term machining, the interaction between thermal accumulation and dynamic loads causes the error propagation path to vary over time. Traditional linear superposition models cannot accurately describe the error evolution, resulting in compensation lag or even overcompensation.

[0004] The core challenge currently facing dynamic error compensation technology for CNC machine tools lies in the insufficient real-time analysis and adaptive compensation capabilities for multi-physics coupling effects. Traditional methods generally employ fixed-parameter models calibrated offline, which are ill-suited to the time-varying error characteristics caused by the dynamic coupling of multiple factors such as sudden changes in cutting force, thermal accumulation effects, and mechanical vibrations during machining. Especially in high-precision machining scenarios, the nonlinear superposition of geometric errors, thermally induced errors, and dynamic errors leads to significant residuals in traditional linear compensation models. Furthermore, decoupling algorithms based on single physical quantities (such as frequency domain filtering or temperature-displacement regression) often cause modal aliasing and misjudgment of error components because they ignore cross-scale coupling mechanisms. More critically, existing compensation systems lack the ability to online perceive and collaboratively respond to changes in the dynamic characteristics of the machine tool structure (such as spindle stiffness decay and guideway wear) and environmental disturbances (coolant temperature changes and workshop air pressure fluctuations). This results in compensation accuracy under complex working conditions generally being limited to above 15μm, failing to meet the stringent requirements for sub-micron precision in fields such as aerospace precision component machining and ultra-precision optical mold machining. Against this backdrop, there is an urgent need to construct a closed-loop control system that integrates real-time perception of multi-source heterogeneous data, dynamic decoupling of nonlinear coupling errors, and multi-axis collaborative intelligent compensation, in order to overcome the triple technical barriers of traditional compensation methods in terms of model adaptability, decomposition accuracy, and control real-time performance. Summary of the Invention

[0005] To address the shortcomings of existing CNC machine tool error compensation methods, such as insufficient analysis of multi-physics coupling effects, low error decoupling accuracy, and poor real-time performance, this invention proposes a multi-source dynamic error compensation method under complex working conditions. This method constructs a real-time sensing system for multi-source heterogeneous data, designs an improved variational mode decomposition (VMD) algorithm to achieve dynamic decoupling of nonlinear errors, and innovatively develops a feedforward-feedback-self-learning composite control architecture to form a multi-axis collaborative closed-loop dynamic compensation system. The core of this invention lies in integrating geometric-thermal-mechanical multi-physics coupling modeling, adaptive error component extraction, and intelligent dynamic compensation technology. This overcomes the limitations of traditional linear superposition models, enabling online identification of machine tool error propagation paths and dynamic optimization of compensation strategies. This significantly improves compensation accuracy and system robustness under complex and variable working conditions, meeting the stringent requirements of aerospace, precision optics, and other fields for sub-micron level machining accuracy.

[0006] The core of this invention lies in constructing a dynamic compensation method for multi-source errors in CNC machine tools under complex working conditions. The specific technical solution includes the following key steps: (1) Multi-physics data sensing and fusion, which involves real-time acquisition of machine tool operating status data through a distributed sensor network, including: Spindle temperature field distribution: Fiber grating temperature sensors are arranged at key points of the spindle bearing, lead screw nut and bed, with a sampling frequency ≥100Hz;

[0007] in These are the thermal mode basis functions; Three-dimensional cutting force components: A piezoelectric force sensor installed at the tool holder is used, with a measurement bandwidth ≥5 kHz; ; Vibration acceleration spectrum of the guide rail: Vibration signals in the 0.1-2000Hz frequency band are detected by an array of MEMS accelerometers arranged on the slide. ; Environmental parameters: including barometric pressure sensor, humidity sensor, and coolant temperature sensor; This includes air pressure, humidity, and coolant temperature.

[0008] Preferably, the temperature field reconstruction uses the Kriging interpolation algorithm, which determines the optimal weight coefficients by solving the covariance matrix equation, with an interpolation error of < ±0.5℃.

[0009] Weighting coefficient By solving the covariance matrix equation Obtain, among which .

[0010] (2) Multi-source error coupling modeling: Establish an error propagation model that includes geometric-thermal-mechanical coupling:

[0011] In the formula: The structural transfer matrix was identified by fusing hammer impact experiments with the finite element model. The temperature, force, and vibration sensitivity matrices are obtained through least squares system identification. This represents the machine tool reference error.

[0012] Preferably, the sensitivity matrix identification employs a recursive least squares method:

[0013] Among them, forgetting factor λ ∈[0.95,0.99], the initial value of the covariance matrix P is set to 104 .

[0014] (3) Variational mode cooperative decomposition: An improved adaptive VMD algorithm is used to analyze the synthesis error. Ψ(t) Perform multi-scale decomposition:

[0015] ; ; in ∈[1000,5000], number of modes K Based on the Bayesian information criterion, the geometric error is adaptively selected and decomposed. Thermal error Dynamic error Quantity.

[0016] Preferably, the modal component selection criterion of the improved VMD algorithm is: Frequency band energy percentage: ; Temporal kurtosis index: ; Modal components that simultaneously meet the above conditions are retained.

[0017] (4) Prediction of compensation amount using deep learning: Construct a bidirectional LSTM network incorporating an attention mechanism, and establish a nonlinear mapping from error components to compensation quantities:

[0018] In the formula: This is the hidden state of a bidirectional LSTM; Attention weight Wa ∈Rd×2d For the hidden layer dimension.

[0019] Network training uses a weighted loss function:

[0020] in For dynamic weights.

[0021] Preferably, the bidirectional LSTM network adopts a distributed training architecture, comprising: Feature extraction layer: 3-layer convolutional neural network (kernel size 3×3, stride 1); Temporal processing layer: 64 bidirectional LSTM units, with the initial bias value of the forget gate set to 1.0; Attention mechanism layer: computes the context vector; ,

[0022] Output layer: Fully connected network with Dropout layer (dropout rate 0.2).

[0023] (5) Multi-axis collaborative dynamic compensation: Design a feedforward-feedback composite controller to generate compensation instructions: Feedforward channel: ; Feedback channel: ; Synthesized output: ; In the formula The adaptive coupling matrix satisfies: .

[0024] Preferably, the adjustment process of the adaptive coupling matrix incorporates a dynamic limiting strategy:

[0025] in , The update period is an integer multiple of the control period (1ms).

[0026] (6) Online evaluation and updating of compensation effect: The actual position deviation was measured using a laser interferometer. When the following conditions are met: ; Trigger online updates of model parameters, where Tw ∈[5,60]s is a sliding window. η =3 is the update threshold.

[0027] Furthermore, it also includes a compensation priority decision module, which allocates compensation resources according to the following rules when multiple axis errors exceed limits simultaneously:

[0028] Among them, the weighting coefficient , Axial axes with a priority > 0.5 are processed first.

[0029] Furthermore, the model parameters are updated online using the momentum gradient descent method:

[0030] Momentum coefficient β =0.9, learning rate adjusted according to cosine annealing strategy: ; in , K=1000 is the number of training steps.

[0031] Furthermore, the online update strategy for the structural transfer matrix is ​​as follows:

[0032] Adaptive learning rate γ=1 / (n+1)0.6 The update trigger condition is an ambient temperature change ΔT > 2℃ or continuous running time > 4 hours.

[0033] Furthermore, the spatial mapping of the compensation amount includes a backlash compensation term, the calculation of which satisfies:

[0034] In the formula The backlash calibration values ​​for each axis. For speed threshold, The instantaneous velocity of each axis.

[0035] The beneficial effects of this invention are mainly reflected in the following aspects: Improving machining accuracy: By sensing and fusing multi-physics data, real-time machine tool operating status data is collected, including spindle temperature field, three-dimensional cutting force, guideway vibration acceleration, and environmental parameters. This allows for a comprehensive and accurate understanding of the machine tool's actual operation under complex conditions. Based on this, a geometry-thermal-mechanical coupled error propagation model is established, which can more accurately describe the generation and propagation of machine tool errors. Compared to traditional single-error-source modeling methods, this method can more effectively consider the coupled effects of dynamic factors such as cutting force and vibration, thereby achieving more precise error compensation and significantly improving machining accuracy. The compensation accuracy is increased from the traditional 15μm or higher to the sub-micron level, meeting the high-precision machining needs of aerospace, precision optics, and other fields.

[0036] Enhancing System Adaptability: An improved adaptive VMD algorithm is employed to decompose the synthetic error into multiple scales. This algorithm adaptively selects the number of modes based on the error characteristics under different operating conditions, accurately decomposing geometric, thermal, and dynamic error components. Compared to traditional fixed-parameter models or single-physical-quantity decoupling algorithms, this method better addresses the time-varying characteristics of error propagation paths under complex operating conditions, avoiding modal aliasing and misjudgment of error components. This results in a more adaptable and robust system, maintaining stable compensation performance in processing scenarios with varying operating conditions and multiple disturbances.

[0037] Enhancing Real-Time Compensation: The constructed bidirectional LSTM network, combined with an attention mechanism, can efficiently learn and rapidly predict the nonlinear mapping relationship between error components and compensation quantities. This network structure can fully exploit the temporal features and inherent patterns in the error data, while the attention mechanism can highlight key features, improving prediction accuracy and speed. Compared with traditional compensation methods, this deep learning prediction method can generate compensation commands faster, achieving real-time dynamic compensation for machine tool errors, effectively reducing compensation lag, and improving the efficiency and quality of the machining process.

[0038] Optimized compensation strategy: The designed feedforward-feedback composite controller and adaptive coupling matrix can dynamically adjust the compensation amount and rationally allocate compensation resources among multiple axes based on the magnitude and trend of the error. The compensation priority decision module ensures that when multiple axes simultaneously exceed their limits, the axis with the larger error is processed first, avoiding resource waste and overcompensation. Simultaneously, backlash compensation further improves the accuracy and reliability of the compensation, enabling the entire compensation system to operate more intelligently and efficiently, optimizing the compensation strategy and enhancing the overall system performance.

[0039] Online evaluation and continuous optimization are achieved: The compensation effect is evaluated online using measurement tools such as laser interferometers. When the actual position deviation exceeds a set threshold, the model parameters are updated online in a timely manner. Momentum gradient descent and cosine annealing strategies are employed to adaptively adjust the model parameters, while the structural transfer matrix is ​​updated in real time to adapt to changes in the dynamic characteristics of the machine tool structure and the influence of environmental disturbances. This online evaluation and update mechanism ensures that the accuracy and effectiveness of the compensation model are always at their optimal state, enabling the system to continuously self-optimize and improve, continuously enhancing compensation accuracy and better coping with complex and ever-changing machining environments.

[0040] Reduced production costs and increased production efficiency: Although this invention requires some investment in hardware and software, it significantly improves machining accuracy, reduces scrap rates and rework frequency, thereby lowering production costs. Simultaneously, the real-time dynamic compensation function can shorten machining cycles and improve machine tool production efficiency. Attached Figure Description

[0041] Figure 1 This is a flowchart of a method for dynamic compensation of multi-source errors in CNC machine tools under complex working conditions, according to an embodiment of the present invention.

[0042] Figure 2 This is a system architecture diagram of a multi-source error dynamic compensation method for CNC machine tools under complex working conditions, according to an embodiment of the present invention.

[0043] Figure 3This is an error coupling diagram of a multi-source error dynamic compensation method for CNC machine tools under complex working conditions, according to an embodiment of the present invention. Detailed Implementation

[0044] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0045] Figure 1 This is a flowchart illustrating a method for dynamic compensation of multi-source errors in CNC machine tools under complex working conditions, according to an embodiment of the present invention. Figure 2 This is a system architecture diagram of a multi-source error dynamic compensation method for CNC machine tools under complex working conditions according to an embodiment of the present invention. The following is in conjunction with... Figure 1 and Figure 2 The implementation process of this invention will be described in detail.

[0046] Step 1: Multiphysics Data Sensing and Fusion (1) Sensor network topology design: Temperature field monitoring: such as Figure 3 As shown in the thermal error source section, three sets of FISO FOT-M fiber optic temperature sensors (range -50~300℃, resolution 0.1℃) are arranged at the front and rear bearings of the spindle, two sets are installed at the ball screw nuts of the X / Y / Z axes, and four sets are arranged at the thermally symmetrical points of the bed, for a total of 18 temperature measurement points (exceeding the 12 sets required in claim 2). The sensor spacing is optimized according to the thermal conduction model to ensure the accuracy of temperature gradient capture.

[0047] Cutting force sensing: A Kistler 9257B triaxial piezoelectric force gauge is integrated at the tool holder interface, with a range of ±8kN (X / Y direction) and ±12kN (Z direction), and a natural frequency >5kHz (meeting the ≥5kHz requirement of claim 1). The signal is conditioned by a charge amplifier (Kistler 5070A) with an AD sampling rate of 10kHz.

[0048] Vibration monitoring: Three sets of PCB 352C33 MEMS accelerometers (range ±50g, frequency response 0.5-3000Hz) are installed on each of the X / Y / Z axis guide sliders, and are arranged in an equilateral triangle to capture pitch, yaw, and roll vibration components. The signals are synchronously acquired at a sampling rate of 5kHz after being filtered by anti-aliasing (cutoff frequency 2500Hz).

[0049] Environmental parameter acquisition: One BME688 environmental sensor (air pressure ±0.12kPa, humidity ±3%RH, temperature ±0.3℃) is deployed inside / outside the machine tool protective cover. A Pt100 platinum resistance thermometer (±0.1℃) is installed on the coolant pipeline. The data is transmitted to the industrial control computer via CAN bus.

[0050] (2) Data synchronization and preprocessing: Xilinx Zynq-7000 FPGA is used to achieve hard synchronization of multi-source data. The trigger signal is triggered by the Z pulse of the spindle encoder, and the timing jitter is <100ns.

[0051] Temperature field reconstruction: Construct the spatial covariance matrix according to the Kriging algorithm in claim 2: ,

[0052] Solving for weight coefficients ωi After that, the maximum interpolation error of the reconstructed temperature field was 0.4℃.

[0053] Step 2: Multi-source error coupling modeling (1) Identification of structural dynamic characteristics: Hammer impact test: A wideband excitation (0.1-2000Hz) was applied to the X / Y / Z axis ends of the machine tool using a PCB 086D20 pulse hammer (force sensitivity 10mV / N). The frequency response function (FRF) at 24 points was acquired using LMS SCADAS Mobile. Figure 3 The source of geometric error is shown.

[0054] Finite element model fusion: A parametric model of the machine tool was established based on ANSYS Workbench, and the stiffness parameters of the joint were corrected using the recursive least squares method of claim 3. After optimization, the frequency band correlation coefficient between the simulation and the measured FRF reached 0.98, which meets the requirement of >0.95 of claim 3.

[0055] (2) Online identification of sensitivity matrix: Under typical operating conditions of a spindle speed of 6000 rpm and a feed rate of 10 m / min, the formula is updated according to claim 3:

[0056] Initial covariance matrixP=104I When the cutting force suddenly increases from 2000N to 5000N, the system completes the operation within 20ms. K F The update process ensures that the sensitivity coefficient change rate is less than 5% / s, guaranteeing the model's dynamic tracking capability.

[0057] Step 3: Variational Mode Co-decomposition (1) Improved VMD algorithm implementation: Parameter initialization: Set the initial penalty factor α 0=2500, energy decay factor σ =0.8, and the number of modes K=6 is dynamically selected through the Bayesian Information Criterion (BIC).

[0058] Modal screening: according to the dual criteria of claim 4: Frequency band energy distribution: Calculate the power spectral density (PSD) of each mode and retain... ρ k ≥0.85 components (actual measurement) ρ 1=0.91, ρ2=0.87 ); Temporal kurtosis: Removed κk<3 Low impact component (measured) κ3=2.7 Removed); Decomposition effect: Synthesis error Ψ(t) = 15.2 μm Decomposed into ε g =6.3μm (geometric error) εt =7.1μm (thermal error) ε d =1.5μm (dynamic error), residual ε r =0.3μm.

[0059] (2) Real-time optimization: The VMD parallel computing is achieved by using GPU acceleration (NVIDIA Jetson AGX Xavier), with a single decomposition time of <8ms, which meets the real-time requirements of five-axis linkage control.

[0060] Step 4: Deep Learning Compensation Quantity Prediction (1) Network architecture and training: Input layer: Receive timing length is 50. ε g,εt,εd Data (sampling interval 10ms).

[0061] Feature extraction: 3-layer 1D CNN (3×1 kernel, stride 1, number of channels 64 / 128 / 256), ReLU activation, batch normalization.

[0062] Temporal modeling: bidirectional LSTM (64 hidden units, initial value of forget gate bias 1.0), dropout rate 0.2.

[0063] Attention mechanism: Calculating the context vector: ,

[0064] Training strategy: The Adam optimizer is used, with an initial learning rate of 0.001 and dynamic weights. ,when hour Increased to 3 times.

[0065] (2) Online inference performance: Deploying the TensorRT accelerated model on the NX platform, the single prediction time is 3.2ms (<5ms control cycle), and the mean square error (MSE) between the compensated output and the true value is 0.8μm. 2 .

[0066] Step 5: Multi-axis collaborative dynamic compensation (1) Composite controller parameter tuning: Feedforward channel: Online adjustment of the Lyapunov adaptive law according to claim 6: ,

[0067] initial value K p=0.8,Kd=0.15 It automatically converges to the cutting force under the disturbance. K p=1.2,Kd=0.18 .

[0068] Feedback Channel: Integral Coefficient Ki =0.25, anti-saturation function To prevent the actuator from exceeding its limits.

[0069] (2) Adaptive coupling control: When X-axis error hour, λx =1 (Full Feedback); Y-axis Error At that time, calculated according to claim 6:

[0070] Dynamic limiting: As per claim 7, set λmax =1.2, λmin =0.3, when λy When the calculated value is 0.8, the actual output remains at 0.75 (without triggering the amplitude limit).

[0071] (3) Backlash compensation: According to claim 10, when the Z-axis motion speed v z=0.12mm / s>vth=0.1mm / s At that time, apply compensation amount:

[0072] Verification using grating ruler feedback showed that the reverse clearance decreased from 3.2 μm before compensation to 0.7 μm.

[0073] Step 6: Online evaluation and updating of compensation effect (1) Residual monitoring and threshold determination: Position deviation was measured using a Renishaw XL-80 laser interferometer (resolution 0.001 μm) with a sliding window. Tw =The root mean square (RMS) residual within 30 seconds is 0.9 μm. According to the formula in claim 6:

[0074] The current model is determined not to need updating.

[0075] (2) Dynamic parameter update: Triggering condition: When the coolant temperature suddenly drops by 4°C (ΔT>2°C in claim 9), the structure transfer matrix is ​​updated.

[0076] Update process: according to the formula in claim 9:

[0077] The update took 1.8 seconds, and the model prediction residuals decreased by 22% after the update.

[0078] This invention significantly enhances the dynamic compensation capability of CNC machine tools under complex working conditions through technological innovations such as multi-source heterogeneous data fusion, adaptive error decoupling, and intelligent collaborative control. The system effectively solves the problem of strong coupling of errors from multiple physical fields (geometric, thermal, and mechanical), accurately separating error components using improved variational mode decomposition, overcoming the limitations of traditional linear models. Through a feedforward-feedback composite control architecture and an online parameter update mechanism, it achieves dynamic optimization of the compensation strategy, greatly enhancing adaptability and real-time response to time-varying conditions. Combining deep learning prediction and multi-axis collaborative compensation technology, it stably achieves sub-micron level accuracy in precision machining scenarios such as aerospace blades and optical molds, meeting the stringent requirements of high-end manufacturing for the accuracy of complex curved surface contours. The system exhibits excellent robustness, optimizing control resource allocation through intelligent priority decision-making to ensure balanced convergence of multi-axis errors. Simultaneously, relying on a closed-loop evaluation mechanism, it significantly reduces maintenance needs, providing reliable end-to-end accuracy assurance for the machining of high-value-added parts, and promoting the development of intelligent and precise CNC machine tools.

[0079] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this invention should be included within the protection scope of this invention. Therefore, the protection scope of this invention should be determined by the scope of the claims.

Claims

1. A method for dynamic compensation of multi-source errors in CNC machine tools under complex working conditions, characterized in that: Includes the following steps: (1) Multiphysics data sensing and fusion Real-time acquisition of machine tool operating status data through a distributed sensor network, including: Spindle temperature field distribution: Fiber grating temperature sensors are arranged at key points of the spindle bearing, lead screw nut and bed, with a sampling frequency ≥100Hz; ； in These are the thermal mode basis functions; Three-dimensional cutting force components: A piezoelectric force sensor installed at the tool holder is used, with a measurement bandwidth ≥5 kHz; ; Vibration acceleration spectrum of the guide rail: Vibration signals in the 0.1-2000Hz frequency band are detected by an array of MEMS accelerometers arranged on the slide. ; Environmental parameters: including barometric pressure sensor, humidity sensor, and coolant temperature sensor; This includes air pressure, humidity, and coolant temperature; (2) Multi-source error coupling modeling: Establish an error propagation model that includes geometric-thermal-mechanical coupling: ； In the formula: The structural transfer matrix was identified by fusing hammer impact experiments with the finite element model. The temperature, force, and vibration sensitivity matrices are obtained through least squares system identification. This refers to the machine tool reference error. (3) Variational mode cooperative decomposition: An improved adaptive VMD algorithm is used to analyze the synthesis error. Perform multi-scale decomposition: ; ; ; in ∈[1000,5000], number of modes K Based on the Bayesian information criterion, the geometric error is adaptively selected and decomposed. Thermal error Dynamic error Quantity; (4) Deep learning compensation amount prediction: Construct a bidirectional LSTM network incorporating an attention mechanism, and establish a nonlinear mapping from error components to compensation quantities: ； In formula 1: This is the hidden state of a bidirectional LSTM; Attention weight Wa ∈Rd×2d For the hidden layer dimension; Network training uses a weighted loss function: ； in Dynamic weights; (5) Multi-axis collaborative dynamic compensation: Design a feedforward-feedback composite controller to generate compensation instructions: Feedforward channel: ; Feedback channel: ; Synthesized output: ; In the formula The adaptive coupling matrix satisfies: ； (6) Online evaluation and updating of compensation effect: The actual position deviation was measured using a laser interferometer. When the following conditions are met: ； Trigger online updates of model parameters, where Tw ∈[5,60]s is a sliding window. η =3 is the update threshold.

2. The method according to claim 1, characterized in that: The temperature field reconstruction in step (1) uses the Kriging interpolation algorithm. The optimal weight coefficients are determined by solving the covariance matrix equation, and the interpolation error is < ±0.5℃. ； Weighting coefficient By solving the covariance matrix equation Obtain, among which 。 3. The method according to claim 1, characterized in that: In step (2), the sensitivity matrix identification uses the recursive least squares method: ； Among them, forgetting factor λ ∈[0.95,0.99], the initial value of the covariance matrix P is set to 104 .

4. The method according to claim 1, characterized in that: The modal component selection criteria for the improved VMD algorithm in step (3) are as follows: Frequency band energy percentage: ; Temporal kurtosis index: ； Modal components that simultaneously meet the above conditions are retained.

5. The method according to claim 1, characterized in that: The bidirectional LSTM network described in step (4) adopts a distributed training architecture, including: Feature extraction layer: 3-layer convolutional neural network (kernel size 3×3, stride 1); Temporal processing layer: 64 bidirectional LSTM units, with the initial bias value of the forget gate set to 1.0; Attention mechanism layer: computes the context vector; , ； Output layer: Fully connected network with Dropout layer (dropout rate 0.2).

6. The method according to claim 1, characterized in that: The adjustment process of the adaptive coupling matrix introduces a dynamic limiting strategy: ； in , The update period is an integer multiple of the control period (1ms).

7. The method according to claim 1, characterized in that: It also includes a compensation priority decision module, which allocates compensation resources according to the following rules when multiple axis errors exceed the limit simultaneously (prioritizing axes with priority > 0.5): ； Furthermore, the spatial mapping of the compensation amount includes a backlash compensation term, the calculation of which is based on the calibration value. Triggered by instantaneous motion velocity (velocity threshold 0.1 mm / s): When hour, Other situations, .

8. The method according to claim 1, characterized in that: Online model parameter updates employ the momentum gradient descent method (momentum coefficients) The learning rate is adjusted according to the cosine annealing strategy: ； Furthermore, the online update strategy for the structural transfer matrix is ​​based on an adaptive learning rate. Perform incremental updates to the predicted residuals and output terms. The trigger condition is a change in ambient temperature. Or running time > 4 hours:

Images