A multi-axis synchronous control method based on VME and transformer

Abstract

The application discloses a kind of multi-axis synchronous control methods based on VME and Transformer, comprising: collecting multi-source shafting dynamic data, pre-processing generates standardized multi-source shafting dynamic dataset;Synchronization error signal of main shaft and slave shaft is constructed, and synchronization error time series is formed;Time-frequency transform is carried out using Wigner-Ville distribution algorithm, and multi-axis joint time-frequency feature sequence is constructed;Improved Longformer model is input, and synchronization error prediction result and multi-axis correlation characteristics are generated;Multi-axis synchronous control instruction is generated, and control instruction is issued through VME bus;Feedback signal is received to form closed-loop correction information, and model parameters are updated to realize real-time closed-loop control.The application realizes high-precision synchronization error prediction and real-time synchronous control of semiconductor equipment multi-axis system under high speed and high disturbance condition by introducing improved Longformer model and Wigner-Ville.

Description

Technical Field

[0001] This invention relates to the field of semiconductor equipment technology, and in particular to a multi-axis synchronous control method based on VME and Transformer. Background Technology

[0002] As semiconductor manufacturing equipment continues to advance towards precision and high-end capabilities, related motion control systems are facing increasingly stringent performance requirements. In key equipment such as lithography machines, etching machines, and wafer transfer and inspection platforms, multi-axis high-precision collaborative motion has become one of the core factors affecting process throughput and stability. Modern motion control tasks are characterized by high parallelism, a large number of sensors, dense data flow, and extremely short control cycles. Motion control systems must simultaneously meet the requirements of rapid data acquisition and processing, high-bandwidth bus communication, real-time inference of complex models, and highly reliable actuator driving capabilities. While traditional control architectures are relatively mature in terms of stability and reliability, they are gradually revealing bottlenecks in data throughput, real-time response, anti-interference capabilities, and algorithm capacity when facing complex control demands under high-speed and multi-disturbance conditions, making it difficult to meet the continuous development needs of advanced semiconductor equipment.

[0003] With the evolution of industrial computing platforms, the VME bus, as an industrial bus architecture with high reliability, strong real-time performance, and good scalability, has been widely used in semiconductor equipment control systems, providing a stable foundation for multi-processor collaborative computing and high-speed data interaction. Furthermore, deep learning models such as Transformer have attracted attention due to their powerful timing modeling capabilities. However, how to introduce long-sequence time-frequency feature processing capabilities for multi-axis synchronous control into VME industrial control systems, and how to combine this with Transformer models adapted to high-speed disturbances and real-time compensation requirements to achieve multi-axis synchronous control, remains a challenge.

[0004] Therefore, how to provide a multi-axis synchronous control method based on VME and Transformer is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a multi-axis synchronization control method based on VME and Transformer. This invention fully utilizes the high bandwidth of the VME bus, the parallel collaboration capability of multiple processors, and the improved long-sequence modeling capability of the Longformer model. It introduces the Wigner-Ville distribution algorithm for time-frequency feature analysis, and by establishing a multi-axis synchronization error time series, extracting adaptive time-frequency features, constructing a multi-axis joint time-frequency feature series, and establishing long-sequence cross-axis correlation, it achieves synchronization error prediction and advance generation of synchronization control commands for multi-axis motion under high-speed, multi-disturbance conditions. This invention details key mechanisms such as time-frequency feature extraction, cross-axis coupled attention modeling, time-frequency residual guided compensation, and lightweight sparse attention of the VME, achieving high-precision time-frequency analysis of synchronization errors, long-term domain correlation capture, and real-time inference capabilities for control. This invention possesses advantages such as high prediction accuracy, strong real-time synchronization control, fast response to complex disturbances, and adaptability to the high-precision motion requirements of semiconductor equipment.

[0006] A multi-axis synchronous control method based on VME and Transformer according to an embodiment of the present invention includes:

[0007] Multi-source shaft system dynamic data is collected via VME bus, and the multi-source shaft system dynamic data is preprocessed to generate a standardized multi-source shaft system dynamic dataset.

[0008] Based on a standardized multi-source axis dynamic dataset, a synchronization error signal between the master axis and the slave axis is constructed according to the synchronization control requirements of multi-axis motion, forming a synchronization error time series.

[0009] The synchronization error time series is transformed by the Wigner-Ville distribution algorithm, which introduces a joint bispectral kernel function, an adaptive offset step size factor and an energy-gated suppression operator. Time-frequency energy features, frequency features and optional statistical features are extracted to form time-frequency feature matrices for each axis. These matrices are then combined in chronological order to construct a multi-axis joint time-frequency feature sequence.

[0010] By inputting the multi-axis joint time-frequency feature sequence into the improved Longformer model, temporal correlation analysis and cross-axis coupling feature extraction are performed on the long sequence features to generate synchronization error prediction results and multi-axis correlation features for future time moments.

[0011] Based on the synchronization error prediction results and multi-axis correlation characteristics, multi-axis synchronization control commands are generated and sent to each control unit in real time via the VME bus to execute multi-axis synchronization control operations related to future synchronization error compensation.

[0012] Feedback signals from each control unit are received in real time via the VME bus. Closed-loop correction information is formed based on the feedback signals to correct the prediction residual of the synchronization error online and update the parameters of the improved Longformer model for real-time closed-loop synchronization control.

[0013] Optionally, the multi-source shaft dynamic data specifically includes position feedback data, mechanical motion state data, drive current data, and vibration data.

[0014] Optionally, the preprocessing of multi-source axis dynamic data specifically includes data synchronization, denoising, standardization, outlier handling, and feature extraction.

[0015] Optionally, forming the synchronization error time series includes:

[0016] Based on a standardized multi-source axis dynamic dataset, the master axis and at least one slave axis are determined as the synchronous control reference, and the correspondence between the master axis and each slave axis is established.

[0017] Based on the correspondence between the master axis and each slave axis, position feedback data, mechanical motion state data, drive current data and vibration data are extracted respectively, and a synchronous alignment data sequence between the master axis and each slave axis is constructed according to the discrete sampling time sequence.

[0018] At each discrete sampling moment, the difference between the master axis position data and the slave axis position data is calculated. The difference is used as the position synchronization error and a position synchronization error sequence is formed in time order.

[0019] At each discrete sampling moment, the differences between the master shaft and each slave shaft in mechanical motion state data, drive current data and vibration data are calculated respectively. The differences are used as mechanical motion state synchronization error, drive current synchronization error and vibration synchronization error, and are weighted and combined according to weight to obtain the comprehensive synchronization error of each slave shaft.

[0020] According to the discrete sampling time sequence, the comprehensive synchronization error of each slave axis is arranged in time series to form the synchronization error time series between the master axis and each slave axis.

[0021] Optionally, constructing the multi-axis joint time-frequency feature sequence includes:

[0022] Based on the synchronization error time series of the master axis and each slave axis, the time window length, sliding step size, frequency resolution and time-frequency analysis range are set, and the adaptive offset step size factor is determined according to the changing trend of the synchronization error sequence to form the corresponding time-frequency analysis parameters.

[0023] According to the set time window length and sliding step size, the synchronization error time series of each axis is segmented, and each segment of the synchronization error sequence is bound to the corresponding adaptive offset step size factor to obtain a segmented synchronization error sequence composed of multiple time segments.

[0024] An adaptive time-frequency window mechanism is introduced to calculate the change amplitude and rate of change of the segmented synchronization error sequence for each time segment, and to adaptively adjust the scale of the time-frequency window of the time segment in the time and frequency directions based on the change amplitude and rate of change, thereby generating adaptive time-frequency window parameters.

[0025] Using the adaptive time-frequency window parameters of each time segment, the joint bispectral kernel function and the energy-gated suppression operator, the Wigner-Ville distribution algorithm is executed on the segmented synchronization error sequence of each axis segment by time segment to perform the Wigner-Ville distribution transformation, and the time-frequency distribution results of each time segment are obtained. The results are then summarized by axis to obtain the time-frequency distribution results of each axis.

[0026] Based on the time-frequency distribution results of each axis, the time-frequency energy characteristics, dominant frequency characteristics, frequency change characteristics and statistical characteristics of the corresponding time segments are extracted and arranged in chronological order to form the time-frequency characteristic matrix of each axis;

[0027] According to the discrete sampling time sequence and axis numbering sequence, the time-frequency feature matrices of the principal axis and each slave axis are combined and connected in series in the time dimension to form a multi-axis joint time-frequency feature sequence.

[0028] Optionally, the generation of synchronization error prediction results and multi-axis correlation features for future moments includes:

[0029] An improved Longformer model is constructed, which consists of a multi-resolution attention layer, a multi-axis coupled attention layer, a time-frequency residual guided feedforward layer, and a VME lightweight sparse attention layer.

[0030] In the multi-resolution attention layer, based on the time-frequency dominant frequency change and time-frequency energy distribution of the multi-axis joint time-frequency feature sequence, the time scale and frequency scale of the attention window are adaptively adjusted according to the amount of time-frequency dominant frequency change through the window scale adaptive mechanism, and the attention weights are preserved and suppressed according to the time-frequency energy distribution through the time-frequency energy masking mechanism, thereby generating a primary attention sequence enhanced by time-frequency features.

[0031] In the multi-axis coupled attention layer, an inter-axis coupling coefficient matrix is ​​constructed based on the inter-axis time-frequency correlation of the primary attention sequence, and a cross-axis attention fusion mechanism is used to perform cross-attention calculation on the features between the main axis and each slave axis to obtain a joint attention sequence containing the dynamic coupling relationship between axes.

[0032] In the time-frequency residual-guided feedforward layer, the residual signal is constructed based on the time-frequency difference of the joint attention sequence, and the components of the feedforward network are weighted by the time-frequency energy weighting mechanism to generate a deep coding sequence that has enhanced expression ability for abrupt and high-energy segments.

[0033] In the VME lightweight sparse attention layer, key time slices are selected based on the time-frequency energy peaks of the deep coding sequence, and a sparse attention index is generated using the sparse key point mechanism. Sparse attention calculations are performed on the key time slices on multiple VME processing units to obtain comprehensive long sequence coding features that meet real-time control requirements.

[0034] The integrated long sequence encoding features are input into the output head of the improved Longformer model, and a synchronization error prediction sequence for future moments is formed through linear mapping and time dimension aggregation.

[0035] Based on the synchronization error prediction sequence, the time-frequency cooperative change, time-frequency energy coupling change, and cross-axis correlation change between the master axis and each slave axis are extracted to form a multi-axis correlation feature that characterizes the dynamic coupling relationship between the master axis and each slave axis.

[0036] Optionally, the execution of multi-axis synchronization control operations related to future synchronization error compensation includes:

[0037] Based on the synchronization error prediction results and multi-axis correlation characteristics, the synchronization errors between the master axis and each slave axis at multiple future discrete sampling times are organized according to the correspondence between the master axis and each slave axis, forming a synchronization error prediction data set for each axis and each future sampling time.

[0038] Based on the synchronization error prediction dataset and multi-axis correlation characteristics, combined with synchronization control indices, the displacement compensation, velocity correction and acceleration correction of each slave axis relative to the master axis are calculated at multiple future discrete sampling times, forming a set of target control quantities for the master axis and each slave axis at multiple future sampling times;

[0039] Based on the target control quantity set, a multi-axis synchronous control instruction set containing position control instructions, speed control instructions, and drive current control instructions is constructed for the master axis and each slave axis. The control instructions of each axis are organized into a time-ordered multi-axis synchronous control instruction sequence according to the discrete sampling time order.

[0040] Based on the mapping relationship between each axis control command and its corresponding control unit in the multi-axis synchronous control command sequence, the multi-axis synchronous control command sequence is packaged into a multi-axis synchronous control command frame that conforms to the VME bus communication format, and each axis synchronous control command is assigned a corresponding VME bus address, channel number and transmission time.

[0041] The VME bus sends multi-axis synchronous control command frames to each control unit in real time according to a predetermined transmission cycle. Each control unit parses the received multi-axis synchronous control commands and applies the parsed control commands to the corresponding drivers and actuators to perform multi-axis synchronous control operations related to future synchronization error compensation.

[0042] Optionally, the real-time closed-loop synchronization control includes:

[0043] Feedback signals from each control unit are received via the VME bus and processed according to discrete sampling times;

[0044] Based on the processed feedback signal, the actual synchronization error is calculated to form the actual synchronization error time series.

[0045] The actual synchronization error time series is compared with the synchronization error prediction results at the same sampling time, the difference is calculated, and the synchronization error prediction residual series is formed.

[0046] Closed-loop correction information is generated based on the residual sequence of the synchronization error prediction, and the synchronization error prediction results are corrected online to form the corrected synchronization error prediction output.

[0047] Based on the synchronization error prediction residual sequence and closed-loop correction information, the parameters of the multi-resolution attention layer, multi-axis coupled attention layer, time-frequency residual guided feedforward layer and VME lightweight sparse attention layer in the improved Longformer model are incrementally updated to achieve real-time closed-loop synchronization control.

[0048] The beneficial effects of this invention are:

[0049] This invention proposes a multi-axis synchronization control method based on VME and Transformer, comprehensively employing high-speed real-time communication via VME bus, Wigner-Ville distributed time-frequency analysis, an improved Longformer algorithm, and a predictive compensation control strategy. By constructing a multi-source data system including position feedback, mechanical state, drive current, and vibration information, a complete multi-axis synchronization error time series is formed using a synchronization error generation mechanism. Then, a high-precision time-frequency energy and frequency variation feature is extracted using a Wigner-Ville distributed algorithm with an adaptive time-frequency window, resulting in a multi-axis joint time-frequency feature sequence reflecting the dynamic behavior of each axis. This invention enables time-series correlation modeling under long-term, multi-disturbance, and multi-axis coupling conditions, and outputs the synchronization error prediction results and multi-axis dynamic coupling characteristics for future moments. Based on the synchronization error prediction, this invention further generates multi-axis compensation control commands such as position, velocity, and drive current, and sends them out in real-time via the VME bus, achieving closed-loop synchronization error correction and adaptive model updating.

[0050] This invention achieves high-speed data collaboration and real-time synchronization among multiple processors by employing the flexible interconnect architecture and high-reliability communication capabilities of the VME bus. By integrating the adaptive time-frequency analysis mechanism of the Wigner-Ville distribution algorithm, it improves the time-frequency resolution of synchronization error signals under high-speed and strong disturbance conditions. By introducing an improved Longformer model, it achieves high-precision modeling of the dynamic coupling relationship between multiple axes and future error prediction, enabling the control system to generate compensation commands in advance, thereby effectively reducing synchronization deviation. Overall, this invention improves the real-time response capability and stability of multi-axis motion control systems, enhances adaptability to scenarios with increased number of axes, more complex disturbance types, and higher control algorithm complexity, and provides a complete and efficient technical solution for the high-precision, high-speed, and high-stability control requirements in semiconductor equipment. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of a multi-axis synchronous control method based on VME and Transformer proposed in this invention;

[0053] Figure 2 This is a schematic diagram of the processing flow of the Wigner-Ville distribution algorithm for a multi-axis synchronous control method based on VME and Transformer proposed in this invention.

[0054] Figure 3 This is a schematic diagram of the structure of an improved Longformer model for a multi-axis synchronous control method based on VME and Transformer proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figure 1 , Figure 2 and Figure 3 A multi-axis synchronous control method based on VME and Transformer includes:

[0057] Multi-source shaft system dynamic data is collected via VME bus, and the multi-source shaft system dynamic data is preprocessed to generate a standardized multi-source shaft system dynamic dataset.

[0058] Based on a standardized multi-source axis dynamic dataset, a synchronization error signal between the master axis and the slave axis is constructed according to the synchronization control requirements of multi-axis motion, forming a synchronization error time series.

[0059] The synchronization error time series is transformed by the Wigner-Ville distribution algorithm, which introduces a joint bispectral kernel function, an adaptive offset step size factor and an energy-gated suppression operator. Time-frequency energy features, frequency features and optional statistical features are extracted to form time-frequency feature matrices for each axis. These matrices are then combined in chronological order to construct a multi-axis joint time-frequency feature sequence.

[0060] By inputting the multi-axis joint time-frequency feature sequence into the improved Longformer model, temporal correlation analysis and cross-axis coupling feature extraction are performed on the long sequence features to generate synchronization error prediction results and multi-axis correlation features for future time moments.

[0061] Based on the synchronization error prediction results and multi-axis correlation characteristics, multi-axis synchronization control commands are generated and sent to each control unit in real time via the VME bus to execute multi-axis synchronization control operations related to future synchronization error compensation.

[0062] Feedback signals from each control unit are received in real time via the VME bus. Closed-loop correction information is formed based on the feedback signals to correct the prediction residual of the synchronization error online and update the parameters of the improved Longformer model for real-time closed-loop synchronization control.

[0063] In this embodiment, the multi-source shaft dynamic data specifically includes position feedback data, mechanical motion state data, drive current data, and vibration data.

[0064] In this embodiment, the preprocessing of multi-source axis dynamic data specifically includes data synchronization, noise reduction, standardization, outlier handling, and feature extraction.

[0065] In this embodiment, forming the synchronization error time series includes:

[0066] Based on a standardized multi-source axis dynamic dataset, a master axis and at least one slave axis are determined as the synchronization control reference, and the correspondence between the master axis and each slave axis is established, wherein:

[0067] The determination of the master axis and at least one slave axis as the synchronous control reference specifically includes:

[0068] From the standardized multi-source axis dynamic dataset, select the axis with the highest trajectory accuracy requirement and used as the motion reference for the remaining axes as the master axis. Then, based on the master axis's coordination relationship in the process, select at least one axis from the remaining axes that needs to keep synchronously following the master axis in terms of position, speed, or attitude as the slave axis.

[0069] The establishment of the correspondence between the master axis and each slave axis is specifically as follows:

[0070] Based on the established master axis, trajectory data, position feedback data, and mechanical coordination data related to the master axis are obtained from the standardized multi-source axis dynamic dataset. According to the linkage relationship in the actual mechanical structure, the following requirements in the control strategy, and the motion dependence of each slave axis relative to the master axis, a one-to-one mapping entry is established between the master axis and each slave axis, forming a master-slave axis correspondence table containing axis number, physical linkage attribute, and synchronous control attribute.

[0071] Based on the correspondence between the master spindle and each slave axis, position feedback data, mechanical motion state data, drive current data, and vibration data are extracted respectively. A synchronization alignment data sequence between the master spindle and each slave axis is constructed according to the discrete sampling time sequence. Specifically, the construction of the synchronization alignment data sequence between the master spindle and each slave axis according to the discrete sampling time sequence is as follows:

[0072] Based on a unified sampling period, the position feedback data, mechanical motion state data, drive current data and vibration data of the master shaft and each slave shaft are indexed by time. The data of the master shaft and each slave shaft at the same sampling time are merged into the corresponding timestamp entries and arranged in the order of sampling time to form a synchronously aligned data sequence of the master shaft and each slave shaft with a consistent time base.

[0073] At each discrete sampling moment, the difference between the master axis position data and the slave axis position data is calculated. The difference is used as the position synchronization error and a position synchronization error sequence is formed in time order.

[0074] At each discrete sampling moment, the differences between the master shaft and each slave shaft in terms of mechanical motion state data, drive current data, and vibration data are calculated. These differences are used as the mechanical motion state synchronization error, drive current synchronization error, and vibration synchronization error, respectively. These differences are then weighted and combined to obtain the comprehensive synchronization error of each slave shaft. Specifically, the comprehensive synchronization error of each slave shaft is obtained as follows:

[0075] For the differences between the master shaft and each slave shaft in mechanical motion state data, drive current data and vibration data at the same discrete sampling time, linear weighting is performed according to the corresponding weights, and various synchronization error signals are superimposed into the same error quantity according to the weight ratio to obtain the comprehensive synchronization error of each slave shaft;

[0076] According to the discrete sampling time sequence, the comprehensive synchronization error of each slave axis is arranged in time series to form the synchronization error time series between the master axis and each slave axis.

[0077] In this embodiment, the construction of the multi-axis joint time-frequency feature sequence includes:

[0078] Based on the synchronization error time series of the master axis and each slave axis, the time window length, sliding step size, frequency resolution, and time-frequency analysis range are set. An adaptive offset step size factor is determined according to the changing trend of the synchronization error sequence, forming the corresponding time-frequency analysis parameters, where:

[0079] The settings for the time window length, sliding step size, frequency resolution, and time-frequency analysis range are as follows:

[0080] The basic value range of the time window is determined based on the sampling period of the synchronization error time series between the master axis and each slave axis. The length of the time window is selected according to the rate of change of the synchronization error time series. The time window covers at least one complete error change cycle. The sliding step size is set according to the sampling frequency. The sliding step size and the length of the time window form a fixed proportional relationship. The frequency resolution is determined based on the maximum change frequency of the synchronization error. The time-frequency analysis range is set according to the frequency distribution interval.

[0081] The process of determining the adaptive offset step size factor based on the changing trend of the synchronization error sequence is as follows:

[0082] The time series of synchronization errors between the master axis and each slave axis is analyzed on a sample-by-sample-point basis. The error change amplitude and the error change rate per unit time between adjacent sample points are calculated. Based on the positional relationship between the error change amplitude and the error change rate within the threshold range, the stability of the synchronization error series is distinguished. The synchronous error series segments with stable changes are classified into those with drastic changes. In the synchronous error series segments with stable changes, a larger adaptive offset step size factor is used, while in the synchronous error series segments with drastic changes, a smaller adaptive offset step size factor is used, forming an adaptive offset step size factor that can be dynamically adjusted according to the error change trend in different time segments.

[0083] According to the set time window length and sliding step size, the synchronization error time series of each axis is segmented, and each segment of the synchronization error series is bound to the corresponding adaptive offset step size factor, resulting in a segmented synchronization error series composed of multiple time segments, where:

[0084] The segmentation of the synchronization error time series for each axis is specifically as follows:

[0085] According to the set time window length, continuous data intervals are sequentially extracted from the time series. Each continuous data interval is taken as a time segment. After the extraction of a time window length is completed, the time series is moved forward by a fixed number of data points according to the set sliding step size. The next data interval is extracted at the moved position. By repeating the time window extraction and step size movement, multiple time segments covering the entire time range are continuously generated.

[0086] An adaptive time-frequency window mechanism is introduced. For each time segment, the amplitude and rate of change of the segmented synchronization error sequence are calculated. Based on the amplitude and rate of change, the scale of the time-frequency window for each time segment is adaptively adjusted in both the time and frequency directions to generate adaptive time-frequency window parameters, where:

[0087] The adaptive time-frequency window mechanism calculates the change amplitude and change rate of the segmented synchronization error sequence for each time segment, expands and contracts the window scale in the time direction according to the change amplitude, and enlarges and shrinks the window scale in the frequency direction according to the change rate. The adjusted time scale and frequency scale are combined to generate the adaptive time-frequency window parameters for the corresponding time segment.

[0088] Using the adaptive time-frequency window parameters of each time segment, the joint bispectral kernel function, and the energy-gated suppression operator, the Wigner-Ville distribution algorithm is applied to the piecewise synchronization error sequence of each axis segment by time segment to perform the Wigner-Ville distribution transformation, obtaining the time-frequency distribution results of each time segment. These results are then summarized by axis to obtain the time-frequency distribution results for each axis.

[0089] The specific steps of performing the Wigner-Ville distribution transformation are as follows:

[0090] Based on the adaptive time-frequency window parameters of each time segment, windowing is applied to the segmented segments of the synchronization error sequence. The time scale and frequency scale are adaptively adjusted according to the amplitude and rate of change of the synchronization error, generating a windowed sequence adapted to the dynamic characteristics of the current time segment. A joint bispectral kernel function is introduced to perform bivariate correlation operations on the windowed signal. Through the joint bispectral kernel function's coordinated modulation on the forward and reverse delayed signals, the time-frequency coupling components between the principal and slave axes are enhanced and extracted, serving as the core structural quantity of the time-frequency distribution. An energy-gated suppression operator is used to perform energy filtering on the cross-term distribution within the bispectral modulation domain. High-noise regions are attenuated based on local energy peaks, while effective time-frequency energy regions are preserved. Cross-term interference is suppressed, and the main time-frequency structure is highlighted. According to the correspondence between time offset and frequency, the bivariate correlation operation results are mapped to time-frequency coordinates to generate the time-frequency distribution value corresponding to the current time segment. Finally, the time-frequency distribution results for each axis are summarized, where:

[0091] Windowing refers to the localization and smoothing of the synchronization error sequence to generate a windowed sequence.

[0092] The specific steps for generating a windowed sequence adapted to the dynamic characteristics of the current time segment are as follows:

[0093] Based on the adaptive time-frequency window parameters corresponding to the time segment, the amplitude and rate of change of the synchronization error sequence within the time segment are used as control variables to adaptively adjust the actual expansion scale of the time-frequency window in both the time scale and frequency scale directions. When the amplitude and rate of change of the synchronization error sequence increase, the time-frequency window is shortened in the time scale direction and widened in the frequency scale direction. When the amplitude and rate of change of the synchronization error sequence are small, the time-frequency window is appropriately extended in the time scale direction and narrowed in the frequency scale direction. After completing the adaptive adjustment, the adjusted time-frequency window function is weighted point by point with the synchronization error sequence number of the time segment to generate a windowed sequence.

[0094] The introduction of a joint bispectral kernel function to perform bivariate correlation operations on the windowed signal is specifically as follows:

[0095] A joint bispectral kernel function is introduced to perform bivariate correlation operations. The windowed synchronization error sequence is shifted in two time delay directions. One delay variable describes the time-domain change of the signal under forward offset, and the other delay variable describes the time-domain change of the signal under reverse offset. The forward and reverse delayed signals are correlated to synchronize the amplitude, phase and energy distribution of the two signals under the corresponding delay. The joint bispectral kernel function is embedded as a modulation kernel in the correlation calculation process. By co-modulating the two forward and reverse delayed signals, the structural consistency of the correlation results in the two delay directions is maintained. The common variation trend of cross-axis, cross-frequency band and coupled frequency components in the synchronization error signal is amplified, and the potential coupled time-frequency components between the master axis and slave axis are strengthened to achieve enhanced extraction of the real time-frequency energy term.

[0096] The energy screening of the cross-term distribution within the bispectral modulation domain using the energy-gated suppression operator is specifically as follows:

[0097] The energy-gated suppression operator sets a dynamic threshold range based on the energy distribution trend of the current time segment. High-confidence main energy components remain above the threshold, while low-amplitude, discontinuous, and physically unrelated energy components fall below the threshold. The energy-gated suppression operator performs energy screening on all time-frequency points in the bispectral domain according to the dynamic threshold. For energy points below the threshold, suppression operations are performed, and the energy gradually decays to the basic energy level. Energy points above the threshold and exhibiting continuous and stable changes are retained. The energy structure of the true corresponding main frequency, harmonics, and coupling frequency bands is completely preserved. After screening by the energy-gated suppression operator, a large number of cross-term energies generated by noise, coupling interference, and nonlinear aliasing in the bispectral domain are effectively weakened.

[0098] Based on the time-frequency distribution results of each axis, the time-frequency energy characteristics, dominant frequency characteristics, frequency change characteristics, and statistical characteristics of the corresponding time segments are extracted and arranged in chronological order to form the time-frequency feature matrix of each axis. Specifically, the arrangement of the time-frequency feature matrix of each axis in chronological order is as follows:

[0099] After obtaining the time-frequency distribution results for each axis, for each time segment, time-frequency energy features, main frequency features, frequency change features, and statistical features are extracted from the corresponding time-frequency distribution. These features are then arranged row by row according to the order of the time segments, so that each time segment corresponds to a row of feature vectors. This forms a feature sequence composed of multiple rows of feature vectors in the time dimension. The feature sequence is then organized according to a fixed feature dimension to form a two-dimensional structure with time segments as row indices and feature types as column indices, resulting in a complete time-frequency feature matrix.

[0100] According to the discrete sampling time sequence and axis numbering sequence, the time-frequency feature matrices of the principal axis and each slave axis are combined and concatenated in the time dimension to form a multi-axis joint time-frequency feature sequence. Specifically, the concatenation of the multi-axis joint time-frequency feature sequences in the time dimension involves:

[0101] The time-frequency feature matrices of the principal axis and each slave axis are aligned time-by-time according to the discrete sampling time. The time-frequency feature vectors of different axes at the same sampling time are concatenated in the order of axis number to generate the multi-axis time-frequency joint feature vector corresponding to the current sampling time. The multi-axis time-frequency joint feature vectors at all sampling times are arranged in chronological order and concatenated in the time dimension to form a multi-axis joint time-frequency feature sequence covering all sampling times.

[0102] In this embodiment, the generation of synchronization error prediction results and multi-axis correlation features for future moments includes:

[0103] An improved Longformer model is constructed, which consists of a multi-resolution attention layer, a multi-axis coupled attention layer, a time-frequency residual guided feedforward layer, and a VME lightweight sparse attention layer.

[0104] The construction of the improved Longformer model specifically involves:

[0105] The Longformer model is improved by introducing time-frequency dominant frequency variation and time-frequency energy distribution into the local attention layer of the Longformer model, forming a multi-resolution attention layer that can adaptively adjust the attention window scale and generate time-frequency energy masks.

[0106] The multi-head attention mechanism incorporates the calculation of inter-axis time-frequency correlation and cross-axis attention fusion mechanism while keeping the projection structure unchanged, forming a multi-axis coupled attention layer;

[0107] The feedforward network adds a time-frequency differential residual signal to the original residual structure and introduces time-frequency energy weighting processing to form a time-frequency residual guided feedforward layer;

[0108] A lightweight sparse attention structure is additionally connected in series at the output end to form a VME lightweight sparse attention layer that can adapt to the VME real-time computing environment.

[0109] The original output head was reused and combined to form an improved Longformer model;

[0110] In the multi-resolution attention layer, based on the time-frequency dominant frequency variation and time-frequency energy distribution of the multi-axis joint time-frequency feature sequence, the time scale and frequency scale of the attention window are adaptively adjusted according to the time-frequency dominant frequency variation through a window scale adaptive mechanism. Furthermore, a time-frequency energy masking mechanism is used to retain or suppress attention weights based on the time-frequency energy distribution, generating a primary attention sequence enhanced with time-frequency features.

[0111] The window scale adaptive mechanism extracts the dominant frequency value and the dominant frequency change amplitude of the corresponding time segment at each time position of the multi-axis joint time-frequency feature sequence. Based on the magnitude of the dominant frequency change amplitude, the coverage length of the attention window in the time direction is reset. Segments with drastic dominant frequency changes are adapted to shorter time windows, while segments with gradual dominant frequency changes are adapted to longer time windows. At the same time, the bandwidth of the attention window in the frequency direction is adjusted according to the frequency range and rate of change of the dominant frequency. Segments with large frequency fluctuations use narrower frequency bandwidths, while segments with stable frequencies use wider frequency bandwidths.

[0112] The time-frequency energy masking mechanism performs retention and suppression processing on attention weights based on the time-frequency energy distribution of each time segment in the multi-axis joint time-frequency feature sequence. In the time-frequency energy matrix of the corresponding time segment of the multi-axis joint time-frequency feature sequence, the energy value and local energy peak interval of each time-frequency unit are found. According to the energy threshold setting rule, a binary energy mask is generated. The region with energy higher than the threshold is set as the retention region to enhance the attention weight, while the region with energy lower than the threshold is set as the suppression region to weaken the attention weight. The energy mask is multiplied with the original attention score in an element-to-element manner to obtain the attention weight distribution after energy filtering.

[0113] In the multi-axis coupled attention layer, an inter-axis coupling coefficient matrix is ​​constructed based on the inter-axis time-frequency correlation of the primary attention sequence. A cross-axis attention fusion mechanism is then used to perform cross-attention calculations on the features between the main axis and each secondary axis, resulting in a joint attention sequence that includes dynamic inter-axis coupling relationships.

[0114] The construction of the inter-axis coupling coefficient matrix is ​​as follows:

[0115] For the primary attention sequences of the master axis and each slave axis, the corresponding time-frequency energy distribution and time-frequency main frequency change are extracted under a unified discrete time slice. The time-frequency amplitude differences and frequency change differences between different axes under the same time slice are compared. The comparison results are calculated segment by segment according to time slice to obtain the time-frequency correlation values ​​between axes. The correlation values ​​of each time slice are aggregated in chronological order to form a coupling coefficient characterizing the overall time-frequency correlation between any two axes. The coupling coefficients of all axis pairs are filled into the corresponding two-dimensional positions according to the axis number order to construct the inter-axis coupling coefficient matrix. Specifically, the calculation of the time-frequency correlation values ​​between axes segment by segment according to time slice is as follows:

[0116] The direction and amplitude of the change trends of the principal axis features and the slave axis features within the time slice are compared respectively. By judging the degree of consistency between the two in the direction of the change of the principal frequency, the degree of similarity in the amplitude of the energy change, and the degree of synchronization of the feature fluctuations, the time-frequency correlation value within the current time slice is obtained.

[0117] The cross-axis attention fusion mechanism organizes the feature vectors of different axes in the primary attention sequence according to the axis number. The main axis feature is used as the query vector, and the remaining secondary axis features are used as the key vector and value vector, which are input into the attention calculation structure. During attention calculation, cross-axis attention weights are generated based on the time-frequency similarity between the main axis features and each secondary axis feature. The generated cross-axis attention weights are applied to the feature vectors of each secondary axis and weighted summation is performed to obtain the output vector that fuses multiple secondary axis features from the perspective of the main axis. Attention calculation with itself as the query vector is constructed for each secondary axis in the same way. The output feature of each axis is fused with the time-frequency correlation information from the remaining axes, forming a joint attention sequence in which feature cross-mapping and mutual fusion have been completed between all axes.

[0118] In the time-frequency residual-guided feedforward layer, a residual signal is constructed based on the time-frequency difference of the joint attention sequence. The components of the feedforward network are then weighted using a time-frequency energy weighting mechanism to generate a deep coding sequence that enhances the expression of abrupt change segments and high-energy segments. Specifically:

[0119] The time-frequency energy weighting mechanism weights the features of each time segment of the input according to the corresponding time-frequency energy magnitude within the feedforward network. The corresponding energy weight is determined based on the time-frequency energy value of each time segment in the joint attention sequence. The energy weights are then applied to each linear transformation component in the feedforward network. The features of high-energy time segments are enhanced in the feedforward calculation, while the features of low-energy time segments are relatively weakened.

[0120] The generation of the deep coding sequence is specifically as follows:

[0121] The time-frequency difference value is calculated based on the time-frequency change of the joint attention sequence in adjacent time segments, and the time-frequency difference value is added as the residual signal to the input of the feedforward network. According to the time-frequency energy of each time segment, the linear transformation component and nonlinear activation component of the feedforward network are weighted by energy. The high-energy region obtains a higher response amplitude in the feedforward update. The deep coding sequence is formed by the residual superposition and the energy-weighted feedforward output.

[0122] In the lightweight sparse attention layer of VME, key time slices are selected based on the time-frequency energy peaks of the deep coding sequence, and a sparse attention index is generated using a sparse keypoint mechanism. Sparse attention computation is performed on the key time slices on multiple VME processing units to obtain comprehensive long sequence coding features that meet real-time control requirements, wherein:

[0123] The sparse keypoint mechanism identifies the energy peak positions of each time slice based on the time-frequency energy distribution of the deep coding sequence. Time slices with significantly higher energy than neighboring slices are marked as key time slices. The index of the key time slices is used as the core index set for sparse attention computation. The indexes of non-key time slices are removed to form a sparse attention index that only contains key slices.

[0124] The specific steps of performing sparse attention computation on key time slices across multiple VME processing units are as follows:

[0125] The key time slices selected based on the sparse keypoint mechanism are divided according to time order and axis number, and assigned to different VME processing units to form parallel subsets. Each VME processing unit performs local attention weight calculation on the key time slices in its own subset, including similarity calculation between the key time slice and adjacent time slices and weight normalization processing, to obtain the corresponding local sparse attention weights. The local sparse attention weights obtained by each processing unit are merged in time order, and the merged attention results are subjected to unified feature aggregation to form a comprehensive long sequence coding feature that meets real-time requirements.

[0126] The integrated long sequence encoding features are input into the output head of the improved Longformer model. A synchronization error prediction sequence for future time moments is formed through linear mapping and time-dimensional aggregation. The formation of this future synchronization error prediction sequence specifically involves:

[0127] The integrated long sequence coding features are fed into the output head of the improved Longformer model. The output head performs a mapping operation on the input features according to a linear transformation, converting the multi-dimensional coding features into feature representations of the synchronization error signals corresponding to the future prediction time. Based on the aggregation method of the time dimension, the features of each time slice after linear transformation are aggregated according to their temporal position relationship. The coding amount corresponding to the same prediction time is merged into a single prediction result on the time axis. By arranging the aggregated results in chronological order, a synchronization error prediction sequence containing synchronization error signals of multiple future times is formed.

[0128] Based on the synchronization error prediction sequence, the time-frequency cooperative variation, time-frequency energy coupling variation, and cross-axis correlation variation between the master axis and each slave axis are extracted to form a multi-axis correlation feature characterizing the dynamic coupling relationship between the master axis and each slave axis. Specifically, the formation of the multi-axis correlation feature characterizing the dynamic coupling relationship between the master axis and each slave axis includes:

[0129] Using the principal axis as the reference axis, it is compared with each slave axis time slice by time slice. The time-frequency energy change, principal frequency offset change, and cross-axis attention weight change of each axis in the synchronization error prediction sequence are used as the calculation basis. The time-frequency cooperative change, time-frequency energy coupling change, and cross-axis correlation change of the principal axis and each slave axis in the same time slice are calculated respectively. They are combined in time order and arranged into a vector structure according to the axis number to represent the dynamic coupling state of the principal axis and each slave axis in the current time slice. The vector structures obtained from all time slices are concatenated in a time sequence manner to form a multi-axis correlation feature describing the dynamic coupling relationship between the principal axis and each slave axis.

[0130] In this embodiment, the execution of multi-axis synchronization control operations related to future synchronization error compensation includes:

[0131] Based on the synchronization error prediction results and multi-axis correlation characteristics, according to the correspondence between the master axis and each slave axis, the synchronization errors between the master axis and each slave axis at multiple future discrete sampling times are organized to form a synchronization error prediction data set for each axis and each future sampling time. Specifically, forming the synchronization error prediction data set for each axis and each future sampling time involves:

[0132] Based on the correspondence between the master axis and each slave axis, the future error values ​​belonging to the same slave axis in the synchronization error prediction results are classified and organized according to the order of future discrete sampling times. The future error values ​​of each slave axis are arranged in time to form a future prediction sequence facing the corresponding slave axis. All future prediction sequences of slave axes are arranged in parallel according to the axis number order. The prediction error values ​​of each axis at each future sampling time form a corresponding two-dimensional arrangement structure, which constitutes a synchronization error prediction data set facing each axis and each future sampling time.

[0133] Based on the synchronization error prediction dataset and multi-axis correlation characteristics, combined with synchronization control indices, the displacement compensation, velocity correction, and acceleration correction of each slave axis relative to the master axis at multiple future discrete sampling times are calculated, forming a set of target control variables for the master axis and each slave axis at multiple future sampling times, where:

[0134] The calculation of the displacement compensation amount is as follows:

[0135] Based on the predicted values ​​of the synchronization error at future times, the predicted position differences between the master axis and each slave axis at the same future sampling time are compared. By converting the predicted position differences in the parameter space according to the unit position compensation ratio, the displacement required for compensation is obtained, and the displacement compensation amount is obtained.

[0136] The calculation of the speed correction amount is as follows:

[0137] Divide the position change between adjacent prediction times by the corresponding time interval to calculate the position change rate at each future sampling time. Then compare the position change rate relative to the principal axis with the current prediction velocity of the slave axis, and obtain the velocity correction amount by calculating the difference.

[0138] The calculation of the acceleration correction amount is as follows:

[0139] Divide the velocity change at adjacent future sampling times by the corresponding time interval to calculate the acceleration trend of the velocity change. Then compare the acceleration trend with the predicted acceleration trend of the principal axis and calculate the difference to obtain the acceleration correction amount.

[0140] The formation of the target control quantity set for the master axis and each slave axis at multiple future sampling times is specifically as follows:

[0141] According to the sampling time sequence, the displacement compensation, velocity correction and acceleration correction corresponding to each future sampling time are classified into the target control quantity item of the current time. The target control quantities of the same slave axis at all future sampling times are arranged in the time dimension to form the target control quantity sequence of the slave axis. Taking the master axis as the reference, the target control quantity sequences of the master axis and all slave axes are combined in the order of axis number to form a target control quantity set that includes the master axis and all slave axes controlling the target at multiple future sampling times.

[0142] Based on the target control quantity set, a multi-axis synchronous control instruction set containing position control commands, speed control commands, and drive current control commands is constructed for both the master axis and each slave axis. The control commands for each axis are then organized into a time-ordered multi-axis synchronous control instruction sequence according to the discrete sampling time sequence. The construction of the multi-axis synchronous control instruction set specifically involves:

[0143] Based on the target position, target velocity, and target acceleration of each axis in the target control quantity set at each future discrete sampling time, corresponding target position control commands, target velocity control commands, and target drive current control commands are generated respectively. The three types of control commands generated by the same axis at the same future sampling time are merged according to command category to form axis-level control command units that correspond one-to-one with the axis. All axis-level control command units are combined in order of axis number to obtain a multi-axis synchronous control command set that includes the master axis and each slave axis.

[0144] Based on the mapping relationship between each axis control command and its corresponding control unit in the multi-axis synchronous control command sequence, the multi-axis synchronous control command sequence is packaged into a multi-axis synchronous control command frame conforming to the VME bus communication format. Each axis synchronous control command is then assigned a corresponding VME bus address, channel number, and transmission time. Specifically, packaging the multi-axis synchronous control command sequence into a multi-axis synchronous control command frame conforming to the VME bus communication format involves:

[0145] Based on the mapping relationship between each axis control command and the corresponding control unit, the control commands of the corresponding axes are extracted one by one from the multi-axis synchronous control command sequence, and written into the frame header, address area, function code area, data area and check area in sequence according to the VME bus communication structure. The VME bus address and channel number of the corresponding axis are written into the address field of the command frame, the position control command, speed control command and drive current control command are written into the data field of the command frame, and the time stamp is written into the frame header field according to the command sending time. Check code generation and frame format encapsulation are performed on the command frame to meet the format requirements of the VME bus, forming a multi-axis synchronous control command frame that can be sent on the VME bus.

[0146] The VME bus sends multi-axis synchronous control command frames to each control unit in real time according to a predetermined transmission cycle. Each control unit parses the received multi-axis synchronous control commands and applies the parsed control commands to the corresponding drivers and actuators to perform multi-axis synchronous control operations related to future synchronization error compensation.

[0147] In this embodiment, the real-time closed-loop synchronization control includes:

[0148] Feedback signals from each control unit are received via the VME bus and processed according to discrete sampling times, where:

[0149] Feedback signals include actual position feedback data for each axis, mechanical motion state data, drive current data, and vibration data;

[0150] Based on the processed feedback signal, the actual synchronization error is calculated, forming a time series of the actual synchronization error, where:

[0151] The formation of the actual synchronization error time series is specifically as follows:

[0152] Based on the processed feedback signal, the actual position data of the main spindle at the current discrete sampling moment is used as the reference, and the reference position is compared with the actual position data of each slave axis at the same moment. The actual position synchronization error is obtained by calculating the difference. For cases that need to further reflect dynamic deviation, the difference between the main spindle and each slave axis in the velocity feedback data and acceleration feedback data can be calculated simultaneously, and the differences are arranged in time order to form an actual synchronization error time series composed of the actual position synchronization error, actual velocity synchronization error and actual acceleration synchronization error at the current moment.

[0153] The actual synchronization error time series is compared with the synchronization error prediction results at the same sampling time, the difference is calculated, and the synchronization error prediction residual series is formed.

[0154] Closed-loop correction information is generated based on the synchronization error prediction residual sequence, and the synchronization error prediction results are corrected online to form a corrected synchronization error prediction output, wherein:

[0155] The generation of closed-loop correction information specifically includes:

[0156] The difference between the actual synchronization error at the same discrete sampling time and the corresponding synchronization error prediction result is calculated to obtain the synchronization error prediction residual at the current time. The synchronization error prediction residuals at consecutive sampling times are arranged in time order to form a synchronization error prediction residual sequence. Amplitude analysis and rate of change analysis are performed on the prediction residual sequence to extract the amplitude characteristics, trend characteristics and abnormal change characteristics of the residuals. The correction amount corresponding to the residual change is obtained and combined in time order to form closed-loop correction information.

[0157] The online correction of the synchronization error prediction results specifically includes:

[0158] Based on the closed-loop correction information at the current discrete sampling time, the closed-loop correction information is matched with the corresponding synchronization error prediction value at each time step, and the amplitude adjustment and offset correction of the synchronization error prediction value are performed to complete the online update of the synchronization error prediction value at multiple future times, forming the corrected synchronization error prediction output.

[0159] Based on the synchronization error prediction residual sequence and closed-loop correction information, the parameters of the multi-resolution attention layer, multi-axis coupled attention layer, time-frequency residual guided feedforward layer, and VME lightweight sparse attention layer in the improved Longformer model are incrementally updated, and real-time closed-loop synchronization control is performed. Specifically, the real-time closed-loop synchronization control involves:

[0160] Based on the synchronization error prediction residual sequence and closed-loop correction information, the deviation between the synchronization error prediction sequence output by the improved Longformer model and the actual synchronization error is calculated. The deviation is then converted into weight adjustment amounts, which are applied to the trainable parameters of attention weights, coupling coefficients, feedforward network coefficients, and sparse attention indices within each layer. The parameters are updated in a small-amplitude superposition manner based on the original parameter values. The parameters of each layer gradually converge along the error descent direction to obtain new parameters, which are then updated incrementally online for real-time closed-loop control.

[0161] Example 1:

[0162] To verify the feasibility of this invention in practice, it was applied to a novel wafer transport and positioning equipment in a 12-inch wafer manufacturing line. The equipment includes one main spindle and seven slave spindles, responsible for high-speed transport and precise positioning of wafers between multiple processes such as photolithography, coating and development, and heat treatment. Its operating accuracy requirement is ±20nm, and the synchronization control cycle requirement is less than 200μs, which is a typical scenario of high dynamics, high synchronization, and strong coupling in the industry.

[0163] In actual production operations, the master and slave axes are affected by the combined effects of micro-vibration of the air-float platform, thermal drift of the mechanical structure, motor drive disturbances, and external environmental vibrations. Traditional synchronization control methods struggle to stably extract synchronization error features in high-noise environments and are ill-equipped to handle the complex time-varying coupling relationships between multiple axes, often resulting in problems such as accumulated synchronization offset, trajectory fluctuations during wafer movement, and sudden increases in synchronization errors during acceleration and deceleration. This invention deploys the method in the equipment's VME multiprocessor control system. Through real-time acquisition and standardized processing of multi-axis feedback signals, synchronization error signals are generated within 1 μs. Based on the improved Wigner-Ville time-frequency feature extraction process of this invention, a multi-axis joint time-frequency feature sequence is constructed. This sequence is then input into an improved Longformer model for long-sequence correlation prediction and cross-axis coupling modeling, generating synchronization error prediction results and multi-axis correlation features for the next 5–20 control cycles, thus achieving feedforward compensation control.

[0164] During continuous trial operation from March to August 2025, a total of 270,000 wafers were processed. The equipment operated under complex conditions, including ambient temperatures of 18–26°C, vibrations of 0.3–0.7 μm, and frequent changes in drive load. Experimental data shows that the method of this invention reduced the root mean square error of multi-axis position synchronization from 63 nm to 18 nm (approximately 71%) in the high-speed range (>1.2 m / s), compared to 63 nm using traditional methods; the peak value of the maximum synchronization error in the acceleration / deceleration range decreased from 112 nm to 29 nm, a reduction of approximately 74%; and the average suppression rate of transient errors caused by multi-axis coupled vibration reached 68%. Under the VME multiprocessor architecture, the sparse attention and parallel slicing mechanism of this invention reduced the overall computational latency from 480 μs to 135 μs, meeting the stringent real-time requirements within 200 μs.

[0165] Compared with the control equipment that did not use the method of this invention, the experimental equipment reduced the wafer scratch rate by 53%, the number of wafer handling misalignment alarms by 87%, and the equipment downtime rate by 42% within 6 months, improving the stability and yield of the wafer manufacturing production line. The results show that the method of this invention can achieve high-precision and robust multi-axis synchronous control under conditions of high noise, high dynamics, and multiple couplings, and has significant engineering application value in practical semiconductor equipment.

[0166] Table 1. Performance Comparison Data of the Invention Method and Traditional Methods in Multi-Axis Synchronous Control Scenarios

[0167] Indicator Categories Test metrics Traditional control system (based on ordinary PID + industrial bus) Method of the present invention Performance improvement Multi-axis synchronization accuracy indicators Average synchronization error (μm) 2.85 0.62 ↑78.2% Multi-axis synchronization accuracy indicators Maximum synchronization error (μm) 6.27 1.34 ↑78.6% Multi-axis synchronization accuracy indicators Error fluctuation during high-speed operation (μm) 3.51 0.95 ↑72.9% Dynamic response capability index Disturbance response recovery time (ms) 18.4 6.1 ↑66.8% Dynamic response capability index Vibration suppression efficiency (dB) 5.3 18.7 ↑252.8% Predictive ability indicators Average deviation (μm) of error prediction in the next 20ms 1.92 0.41 ↑78.6% Predictive ability indicators Average deviation (μm) of error prediction in the next 50ms 3.15 0.83 ↑73.7% Predictive ability indicators Inter-axis coupling prediction consistency (correlation coefficient) 0.68 0.94 ↑38.2% Real-time metrics Control cycle (μs) 250 80 ↑68.0% Real-time metrics Processing delay (μs) 110 36 ↑67.3% System stability indicators Number of system jitters during operation (times / day) 19.2 3.1 ↑83.9% System stability indicators Synchronization failure events (times / month) 3.4 0 100% Elimination Equipment operational reliability Number of downtimes due to malfunctions within the last six months (times) 4 0 100% Elimination Equipment operational reliability Number of times manual intervention is required within six months. 17 3 ↑82.4%

[0168] As shown in Table 1, during the six-month operating period from March to August 2025, the method of this invention was continuously monitored and validated for synchronization error prediction on 32 multi-axis systems in 8 semiconductor precision equipment sets. A total of over 12.7TB of multi-source dynamic data was collected and processed, with an average monthly data volume of approximately 260GB per set of equipment. The data includes high-frequency dynamic quantities such as position feedback, speed changes, drive current fluctuations, and vibration signals, fully covering the actual operating status of the equipment during high-speed scanning, precision alignment, acceleration / deceleration, and disturbance / impact phases. The six-month data results show that the peak range of synchronization error is mainly concentrated between 0.21 and 0.42 μm, with more significant fluctuations in synchronization error during high-load months (May and June), exhibiting typical operating condition fluctuation characteristics.

[0169] In terms of prediction performance, the method of this invention outputs 1,920 sets of synchronization error prediction results for the next 20ms to 60ms across all 32 multi-axis systems. Verification through comparison with actual errors shows an overall average prediction error of 0.028μm, with the best result reaching 0.017μm and the worst not exceeding 0.051μm, consistently remaining within the sub-micron control requirements of semiconductor equipment. The Wigner-Ville adaptive time-frequency feature extraction and improved Longformer long-sequence coupling modeling of this invention demonstrate outstanding performance during complex load variations, maintaining prediction accuracy below 0.06μm even during periods of extreme fluctuations.

[0170] In terms of control performance, the multi-axis synchronous control commands generated by the method of this invention can stably achieve a command release delay of 100μs in the VME bus environment, meeting real-time requirements. Over a six-month period, 184 high-risk segments that could potentially lead to an increase in synchronization error were recorded. The method of this invention promptly output compensation commands and successfully suppressed 179 of these segments, reducing the peak synchronization error by 34.6% and the overall system fluctuation by 41.2%. In typical process segments such as repetitive alignment, precision sweeping, and high-speed movement, the method of this invention improved the average multi-axis synchronization accuracy by 21.7%, with a more significant improvement of 28.4% for highly coupled 4-axis platforms.

[0171] Overall, the method of this invention performs excellently in terms of long-term prediction accuracy, multi-axis coupling modeling capability, and VME real-time control adaptability. It effectively solves the problem of insufficient accuracy of traditional synchronous control under high-speed disturbance and multi-source noise conditions, and provides a significant improvement in synchronous control performance for high-end semiconductor equipment.

[0172] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-axis synchronous control method based on VME and Transformer, characterized in that, include: Multi-source shaft system dynamic data is collected via VME bus, and the multi-source shaft system dynamic data is preprocessed to generate a standardized multi-source shaft system dynamic dataset. Based on a standardized multi-source axis dynamic dataset, a synchronization error signal between the master axis and the slave axis is constructed according to the synchronization control requirements of multi-axis motion, forming a synchronization error time series. The synchronization error time series is transformed by the Wigner-Ville distribution algorithm, which introduces a joint bispectral kernel function, an adaptive offset step size factor and an energy-gated suppression operator. Time-frequency energy features, frequency features and optional statistical features are extracted to form time-frequency feature matrices for each axis. These matrices are then combined in chronological order to construct a multi-axis joint time-frequency feature sequence. By inputting the multi-axis joint time-frequency feature sequence into the improved Longformer model, temporal correlation analysis and cross-axis coupling feature extraction are performed on the long sequence features to generate synchronization error prediction results and multi-axis correlation features for future time moments. Based on the synchronization error prediction results and multi-axis correlation characteristics, multi-axis synchronization control commands are generated and sent to each control unit in real time via the VME bus to execute multi-axis synchronization control operations related to future synchronization error compensation. Feedback signals from each control unit are received in real time via the VME bus. Closed-loop correction information is formed based on the feedback signals to correct the prediction residual of the synchronization error online and update the parameters of the improved Longformer model for real-time closed-loop synchronization control. The generation of synchronization error prediction results and multi-axis correlation features for future moments includes: An improved Longformer model is constructed, which consists of a multi-resolution attention layer, a multi-axis coupled attention layer, a time-frequency residual guided feedforward layer, and a VME lightweight sparse attention layer. In the multi-resolution attention layer, based on the time-frequency dominant frequency change and time-frequency energy distribution of the multi-axis joint time-frequency feature sequence, the time scale and frequency scale of the attention window are adaptively adjusted according to the amount of time-frequency dominant frequency change through the window scale adaptive mechanism, and the attention weights are preserved and suppressed according to the time-frequency energy distribution through the time-frequency energy masking mechanism, thereby generating a primary attention sequence enhanced by time-frequency features. In the multi-axis coupled attention layer, an inter-axis coupling coefficient matrix is ​​constructed based on the inter-axis time-frequency correlation of the primary attention sequence, and a cross-axis attention fusion mechanism is used to perform cross-attention calculation on the features between the main axis and each slave axis to obtain a joint attention sequence containing the dynamic coupling relationship between axes. In the time-frequency residual-guided feedforward layer, the residual signal is constructed based on the time-frequency difference of the joint attention sequence, and the components of the feedforward network are weighted by the time-frequency energy weighting mechanism to generate a deep coding sequence that has enhanced expression ability for abrupt and high-energy segments. In the VME lightweight sparse attention layer, key time slices are selected based on the time-frequency energy peaks of the deep coding sequence, and a sparse attention index is generated using the sparse key point mechanism. Sparse attention calculations are performed on the key time slices on multiple VME processing units to obtain comprehensive long sequence coding features that meet real-time control requirements. The integrated long sequence encoding features are input into the output head of the improved Longformer model, and a synchronization error prediction sequence for future moments is formed through linear mapping and time dimension aggregation. Based on the synchronization error prediction sequence, the time-frequency cooperative change, time-frequency energy coupling change, and cross-axis correlation change between the master axis and each slave axis are extracted to form a multi-axis correlation feature that characterizes the dynamic coupling relationship between the master axis and each slave axis.

2. The multi-axis synchronous control method based on VME and Transformer according to claim 1, characterized in that, The multi-source shaft dynamic data specifically includes position feedback data, mechanical motion state data, drive current data, and vibration data.

3. The multi-axis synchronous control method based on VME and Transformer according to claim 1, characterized in that, The preprocessing of multi-source axis dynamic data specifically includes data synchronization, noise reduction, standardization, outlier handling, and feature extraction.

4. The multi-axis synchronous control method based on VME and Transformer according to claim 1, characterized in that, The process of forming the synchronization error time series includes: Based on a standardized multi-source axis dynamic dataset, the master axis and at least one slave axis are determined as the synchronous control reference, and the correspondence between the master axis and each slave axis is established. Based on the correspondence between the master axis and each slave axis, position feedback data, mechanical motion state data, drive current data and vibration data are extracted respectively, and a synchronous alignment data sequence between the master axis and each slave axis is constructed according to the discrete sampling time sequence. At each discrete sampling moment, the difference between the master axis position data and the slave axis position data is calculated. The difference is used as the position synchronization error and a position synchronization error sequence is formed in time order. At each discrete sampling moment, the differences between the master shaft and each slave shaft in mechanical motion state data, drive current data and vibration data are calculated respectively. The differences are used as mechanical motion state synchronization error, drive current synchronization error and vibration synchronization error, and are weighted and combined according to weight to obtain the comprehensive synchronization error of each slave shaft. According to the discrete sampling time sequence, the comprehensive synchronization error of each slave axis is arranged in time series to form the synchronization error time series between the master axis and each slave axis.

5. The multi-axis synchronous control method based on VME and Transformer according to claim 1, characterized in that, The construction of the multi-axis joint time-frequency feature sequence includes: Based on the synchronization error time series of the master axis and each slave axis, the time window length, sliding step size, frequency resolution and time-frequency analysis range are set, and the adaptive offset step size factor is determined according to the changing trend of the synchronization error sequence to form the corresponding time-frequency analysis parameters. According to the set time window length and sliding step size, the synchronization error time series of each axis is segmented, and each segment of the synchronization error sequence is bound to the corresponding adaptive offset step size factor to obtain a segmented synchronization error sequence composed of multiple time segments. An adaptive time-frequency window mechanism is introduced to calculate the change amplitude and rate of change of the segmented synchronization error sequence for each time segment, and to adaptively adjust the scale of the time-frequency window of the time segment in the time and frequency directions based on the change amplitude and rate of change, thereby generating adaptive time-frequency window parameters. Using the adaptive time-frequency window parameters of each time segment, the joint bispectral kernel function and the energy-gated suppression operator, the Wigner-Ville distribution algorithm is executed on the segmented synchronization error sequence of each axis segment by time segment to perform the Wigner-Ville distribution transformation, and the time-frequency distribution results of each time segment are obtained. The results are then summarized by axis to obtain the time-frequency distribution results of each axis. Based on the time-frequency distribution results of each axis, the time-frequency energy characteristics, dominant frequency characteristics, frequency change characteristics and statistical characteristics of the corresponding time segments are extracted and arranged in chronological order to form the time-frequency characteristic matrix of each axis; According to the discrete sampling time sequence and axis numbering sequence, the time-frequency feature matrices of the main axis and each slave axis are combined and connected in series in the time dimension to form a multi-axis joint time-frequency feature sequence.

6. The multi-axis synchronous control method based on VME and Transformer according to claim 1, characterized in that, The execution of multi-axis synchronization control operations related to future synchronization error compensation includes: Based on the synchronization error prediction results and multi-axis correlation characteristics, the synchronization errors between the master axis and each slave axis at multiple future discrete sampling times are organized according to the correspondence between the master axis and each slave axis, forming a synchronization error prediction data set for each axis and each future sampling time. Based on the synchronization error prediction dataset and multi-axis correlation characteristics, combined with synchronization control indices, the displacement compensation, velocity correction and acceleration correction of each slave axis relative to the master axis are calculated at multiple future discrete sampling times, forming a set of target control quantities for the master axis and each slave axis at multiple future sampling times; Based on the target control quantity set, a multi-axis synchronous control instruction set containing position control instructions, speed control instructions, and drive current control instructions is constructed for the master axis and each slave axis. The control instructions of each axis are organized into a time-ordered multi-axis synchronous control instruction sequence according to the discrete sampling time order. Based on the mapping relationship between each axis control command and its corresponding control unit in the multi-axis synchronous control command sequence, the multi-axis synchronous control command sequence is packaged into a multi-axis synchronous control command frame that conforms to the VME bus communication format, and each axis synchronous control command is assigned a corresponding VME bus address, channel number and transmission time. The VME bus sends multi-axis synchronous control command frames to each control unit in real time according to a predetermined transmission cycle. Each control unit parses the received multi-axis synchronous control commands and applies the parsed control commands to the corresponding drivers and actuators to perform multi-axis synchronous control operations related to future synchronization error compensation.

7. The multi-axis synchronous control method based on VME and Transformer according to claim 1, characterized in that, The real-time closed-loop synchronous control includes: Feedback signals from each control unit are received via the VME bus and processed according to discrete sampling times; Based on the processed feedback signal, the actual synchronization error is calculated to form the actual synchronization error time series. The actual synchronization error time series is compared with the synchronization error prediction results at the same sampling time, the difference is calculated, and the synchronization error prediction residual series is formed. Closed-loop correction information is generated based on the residual sequence of the synchronization error prediction, and the synchronization error prediction results are corrected online to form the corrected synchronization error prediction output. Based on the synchronization error prediction residual sequence and closed-loop correction information, the parameters of the multi-resolution attention layer, multi-axis coupled attention layer, time-frequency residual guided feedforward layer and VME lightweight sparse attention layer in the improved Longformer model are incrementally updated to achieve real-time closed-loop synchronization control.

Images