Self-adaptive control method for micro-nano high-precision motion platform

Abstract

The invention relates to the technical field of micro-nano motion control, and discloses a self-adaptive control method of a micro-nano high-precision motion platform. The method comprises the following steps: acquiring real-time pose feedback data and target trajectory data of a motion platform, and extracting dynamic response features; calling the trained motion feature analysis network to perform multi-modal feature separation, and generating a platform pose feature set; based on the set, performing space-time coupling analysis on the working environment parameters through an environment disturbance perception model to obtain a fusion result containing mechanical deformation characteristics and environment disturbance characteristics; inputting a fusion result into a dynamic compensation model to calculate a track correction amount, and outputting a driving compensation instruction; and calibrating the target trajectory data in real time according to the compensation instruction, and generating an actual control signal. According to the method, through multi-modal feature analysis, space-time coupling perception and dynamic compensation, the control precision and stability of the motion platform in a complex environment are improved, and the method is suitable for a micro-nano high-precision motion control scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of micro-nano motion control, in particular to an adaptive control method of a micro-nano high-precision motion platform. BACKGROUND

[0002] In the fields of high-end manufacturing and scientific research such as semiconductor manufacturing, precision measurement, and biomedical treatment, a micro-nano high-precision motion platform is a core device for realizing precision operation. Its running precision directly affects product quality and experimental results, and therefore puts forward very high requirements on motion control technology.

[0003] Traditional motion platform control methods mostly adopt PID control or linear control strategies based on models. Such methods rely on accurate mathematical models. However, the mechanical structure of a micro-nano motion platform is complex, and there are nonlinear factors such as friction, clearance, and flexible deformation, which are difficult to completely describe through accurate modeling. At the same time, external factors such as temperature fluctuations, vibrations, and air flow disturbances in the working environment will further exacerbate the pose deviation of the platform.

[0004] With the development of technology, the precision requirement for the motion platform has been improved from the micron level to the nanometer level, and the traditional control method is difficult to cope with complex dynamic characteristics and environmental disturbances. For example, in the control of a lithography machine workbench, a slight temperature change or mechanical vibration will cause trajectory deviation, affecting the lithography precision. In addition, the dynamic response of the motion platform under different loads and running speeds is different, and the control method with fixed parameters cannot realize adaptive adjustment, resulting in a decrease in control precision and insufficient stability.

[0005] Existing adaptive control methods mostly adjust based on a single feature or model, lack comprehensive analysis of multi-modal features, and are difficult to fully reflect the dynamic characteristics of the motion platform. At the same time, in the aspect of environmental disturbance perception and compensation, the spatio-temporal coupling relationship is not effectively combined, resulting in compensation lag or inaccuracy, which limits the further improvement of control precision. SUMMARY

[0006] The purpose of the present application is to provide an adaptive control method for a micro-nano high-precision motion platform to solve the problems raised in the background.

[0007] To achieve the above purpose, the present application provides an adaptive control method for a micro-nano high-precision motion platform, which comprises:

[0008] acquiring real-time pose feedback data and target trajectory data of the motion platform, and extracting dynamic response features in the pose feedback data;

[0009] calling a trained motion feature analysis network to perform multi-modal feature separation processing on the real-time pose feedback data to generate a platform pose feature set;

[0010] Based on the platform pose feature set, an environment disturbance perception model is called to perform spatio-temporal coupling analysis on the working environment parameters, and a feature fusion result containing mechanical deformation features and environment disturbance features is generated;

[0011] The feature fusion result is input into a dynamic compensation model to calculate a trajectory correction amount, and a platform driving compensation instruction is output;

[0012] According to the platform driving compensation instruction, real-time calibration processing is performed on the target trajectory data, and actual control signals of the motion platform are generated.

[0013] Preferably, the trained motion feature analysis network is called to perform multi-modal feature separation processing on the real-time pose feedback data, and a platform pose feature set is generated, including:

[0014] The real-time pose feedback data is processed in time-frequency domain to generate a plurality of frequency band feature components and corresponding phase response parameters;

[0015] The energy distribution graph of each frequency band feature component is extracted, and the energy distribution graph is processed by spatial attention weighting to generate a weighted motion feature vector;

[0016] The motion feature vector is matched with a preset dynamic response threshold to filter out a feature candidate set that meets a confidence condition;

[0017] The feature candidate set is processed by non-uniform sampling to generate the platform pose feature set containing position encoding and attitude encoding.

[0018] Preferably, based on the platform pose feature set, an environment disturbance perception model is called to perform spatio-temporal coupling analysis on the working environment parameters, and a feature fusion result containing mechanical deformation features and environment disturbance features is generated, including:

[0019] According to the position encoding in the platform pose feature set, the working environment parameters are processed by region mapping to generate an environment parameter distribution matrix;

[0020] The environment parameter distribution matrix is processed by gradient field calculation to generate an environment disturbance intensity distribution graph;

[0021] A spatio-temporal convolution subnetwork in the environment disturbance perception model is called to extract local features from the environment disturbance intensity distribution graph to generate an environment disturbance feature tensor;

[0022] A mechanical deformation prediction model is called in parallel to match the platform pose feature set with a stress distribution to generate a mechanical deformation probability distribution;

[0023] The environmental disturbance feature tensor is cross-fused with the mechanical deformation probability distribution to generate the feature fusion result.

[0024] Preferably, the feature fusion result is input into a dynamic compensation model to calculate a trajectory correction amount, and a platform driving compensation instruction is output, comprising:

[0025] The trajectory constraint conditions of the current motion task are analyzed to generate a compensation priority weight associated with the feature fusion result;

[0026] An error propagation matrix is generated based on the compensation priority weight, and a trajectory point sequence in the target trajectory data is sorted;

[0027] An optimal compensation sequence is selected from the error propagation matrix using a reinforcement learning strategy to generate a basic compensation instruction set containing a position compensation amount, an attitude compensation amount, and an acceleration compensation amount;

[0028] The basic compensation instruction set is integrated with platform dynamic response parameters through instruction encoding to generate the platform driving compensation instruction.

[0029] Preferably, the method further comprises:

[0030] Platform execution state data is collected in real time during the motion control process to generate a feedback log containing a trajectory offset amount and a vibration spectrum feature;

[0031] Dynamic error features in the feedback log are extracted, and the dynamic error features are matched with a historical control case library to generate adaptive compensation parameters;

[0032] The dynamic response threshold of the motion feature analysis network is updated online based on the adaptive compensation parameters;

[0033] The updated dynamic response threshold is injected into the dynamic compensation model to recalculate the compensation priority weight in the error propagation matrix.

[0034] Preferably, the extraction of the dynamic error features in the feedback log and the matching of the dynamic error features with the historical control case library to generate adaptive compensation parameters comprises:

[0035] The dynamic error features are subjected to time series segmentation processing to generate a plurality of error segments and their corresponding platform motion state sequences;

[0036] A completed error classification model is called to perform root cause analysis on each error segment to generate a classification label containing mechanical resonance, environmental interference, and control system lag;

[0037] retrieving a compensation strategy template matching the classification label from the historical control case library, to generate a candidate compensation parameter set;

[0038] sorting the platform motion state sequence and the candidate compensation parameter set based on the matching degree, selecting the parameter with the highest confidence to generate the adaptive compensation parameter.

[0039] Preferably, the method further comprises:

[0040] injecting a virtual trajectory disturbance parameter before outputting the control signal, the virtual trajectory disturbance parameter being used to simulate a platform sudden motion deviation scenario;

[0041] monitoring the correction processing result of the dynamic compensation model on the disturbed trajectory, to generate a control robustness evaluation index;

[0042] when the control robustness evaluation index is lower than a preset threshold, triggering the parameter learning mode of the motion feature analysis network;

[0043] updating the convolution kernel parameters of the motion feature analysis network based on the gradient backpropagation of the trajectory difference data before and after the disturbance.

[0044] Preferably, the method further comprises:

[0045] spatial-temporal alignment processing the trajectory difference data before and after the disturbance, to generate a matching sequence containing a set of trajectory points before the disturbance and a set of trajectory points after the disturbance;

[0046] extracting the trajectory deviation amount in the matching sequence, to generate a position deviation vector and a posture deviation vector of each trajectory point;

[0047] constructing a motion regression loss function according to the position deviation vector and the posture deviation vector;

[0048] differentiable transformation processing the motion regression loss function, to generate a loss value tensor required for gradient backpropagation;

[0049] iterating the convolution kernel parameters of the motion feature analysis network, to calculate the partial derivative of the loss value tensor to each convolution kernel, and to generate a convolution kernel gradient matrix;

[0050] iterative incremental adjustment of the convolution kernel parameters of the motion feature analysis network based on the convolution kernel gradient matrix and a preset learning rate parameter.

[0051] Preferably, the method further comprises:

[0052] A cross-platform control instruction adapter is constructed, and the instruction protocol differences of different driving devices are resolved through the cross-platform control instruction adapter.

[0053] The platform driving compensation instruction is converted into underlying control signals supported by the target driving device.

[0054] The timing constraint relationship and dynamic compensation context of the actual control signal are preserved during the conversion process.

[0055] The precision optimization parameters matched by the target driving device are injected to generate a motion control instruction stream that meets the cross-platform execution conditions.

[0056] Preferably, the cross-platform control instruction adapter is constructed, and the instruction protocol differences of different driving devices are resolved through the cross-platform control instruction adapter, including:

[0057] A driving device protocol rule library is established to store the control instruction mapping table and parameter transmission path of each platform.

[0058] The platform driving compensation instruction is subjected to instruction structure analysis to generate intermediate control representation layer data.

[0059] Based on the intermediate control representation layer data, the control instruction mapping table is queried to generate a device-compatible instruction conversion scheme.

[0060] The parameter transmission paths with conflicts are subjected to semantic adaptation processing to generate a control instruction conversion result without ambiguity.

[0061] Compared with the prior art, the present application has the following advantages:

[0062] By obtaining real-time pose feedback data and target trajectory data, and extracting dynamic response features, the running state changes of the motion platform can be captured in a timely manner. This real-time data acquisition and feature extraction enables the control process to closely follow the actual motion of the platform, avoiding control deviations caused by information lag.

[0063] The trained motion feature analysis network is called to perform multi-modal feature separation processing to generate a platform pose feature set, which can separate multiple related features from complex pose data. These features cover different aspects of platform motion, such as position, attitude, speed, etc. Through comprehensive analysis of these multi-modal features, the motion characteristics of the platform can be more fully understood, providing rich basis for subsequent control adjustment.

[0064] Based on the platform pose feature set, the environmental disturbance perception model is called to perform spatio-temporal coupling analysis on the working environment parameters, to generate a feature fusion result containing mechanical deformation features and environmental disturbance features, realizing comprehensive consideration of environmental factors and mechanical changes. Environmental disturbance and mechanical deformation have mutual influence in space and time. Through this coupling analysis, the comprehensive effect of the two on platform motion can be accurately identified, avoiding one-sidedness that may occur when considering a single factor.

[0065] The feature fusion result is input into the dynamic compensation model to calculate the trajectory correction amount, and the platform drive compensation instruction is output, so that the compensation instruction can accurately correspond to the actual trajectory deviation reason. The dynamic compensation model calculates the correction amount according to the fused features, which can give corresponding compensation strategies for different disturbances and deformations, ensuring the pertinence and effectiveness of compensation.

[0066] According to the platform drive compensation instruction, real-time calibration processing is performed on the target trajectory data to generate the actual control signal of the motion platform, realizing a closed-loop process from data acquisition, feature analysis to compensation control. This closed-loop control can continuously calibrate the trajectory, so that the actual motion trajectory of the platform always keeps consistent with the target trajectory, reducing the trajectory deviation caused by various factors, and improving the running stability and trajectory precision of the motion platform in complex environments and dynamic change conditions. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 The working principle diagram of the adaptive control method of the micro-nano level high-precision motion platform is described.

[0068] Figure 2 The flowchart of environmental disturbance perception and feature fusion is described.

[0069] Figure 3 The flowchart of dynamic compensation and instruction generation is described.

[0070] Figure 4 The flowchart of dynamic error feature matching is described.

[0071] Figure 5 The flowchart of cross-platform control instruction adaptation is described. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0073] Please refer to Figure 1The application provides an adaptive control method for a micro-nano high-precision motion platform, and the method comprises the following steps:

[0074] By integrating data processing and model calling, dynamic calibration of the platform pose is realized. Specifically, real-time pose feedback data and target trajectory data of the motion platform are obtained, six-degree-of-freedom position and attitude information of the platform is collected by a sensor system, and dynamic response characteristics such as acceleration spectrum and displacement change rate in the pose feedback data are extracted. Subsequently, a pre-trained motion feature analysis network is called to perform multi-modal feature separation processing on the real-time pose feedback data. The network is based on a deep convolutional structure, separates mechanical vibration features and motion offset components through multiple levels, and generates a platform pose feature set containing position encoding and attitude encoding vectors. Based on the platform pose feature set, an environment disturbance perception model is called to perform spatio-temporal coupling analysis on the working environment parameters. The environment parameters include real-time monitoring variables such as temperature gradient, air pressure fluctuation and vibration interference, and the model fuses nonlinear filtering and regression algorithms to generate a feature fusion result containing mechanical deformation features and environmental disturbance features. The feature fusion result is transmitted as input to a dynamic compensation model to calculate the trajectory correction amount. The dynamic compensation model adopts a reinforcement learning framework, outputs platform driving compensation instructions including position correction and attitude adjustment values. According to the platform driving compensation instructions, real-time calibration processing is performed on the target trajectory data. The calibration process involves interpolation algorithms and constraint optimization, and the trajectory point sequence is updated to generate actual control signals of the motion platform. The signals directly drive the actuator system of the platform to ensure that the trajectory tracking accuracy is at the micro-nano level.

[0075] Embodiment 1: refer to Figure 2 The real-time pose feedback data is obtained by high-precision encoders and laser interferometers, and the sampling frequency is not less than 20 kHz. The original data contains six-degree-of-freedom coordinates and attitude quaternions. Time-frequency domain decomposition processing is performed on the real-time pose feedback data, and a discrete wavelet transform algorithm is used to decompose the original signal into 8 frequency band feature components. Each frequency band covers a specific mechanical response frequency band, such as a 0-100Hz base vibration band and a 1-5kHz actuator resonance band. Each frequency band feature component is processed by Hilbert transform to extract phase response parameters, including instantaneous amplitude envelope and phase shift angle. The above decomposition process uses Daubechies wavelet basis function, and the decomposition level is dynamically configured by the platform dynamics characteristics.

[0076] For each frequency band component, its power spectral density is calculated and a two-dimensional energy distribution map is generated. The energy distribution map is mapped to the three-dimensional workspace of the motion platform, and each spatial unit records the energy amplitude. Spatial attention weighting processing is performed on the energy distribution map: a normalized exponential function is used to calculate the spatial unit weight factor, and the weight coefficient is positively correlated with the energy density in the unit. High-energy aggregation areas are assigned a weight gain coefficient, and low-energy areas are attenuated. The weighted result is converted into a 512-dimensional motion feature vector, and the vector dimension is decoupled from the platform degree of freedom, independently encoding the dynamic characteristics of different motion directions.

[0077] The motion feature vector is introduced into the feature screening module and matched with the preset dynamic response threshold library. The threshold library stores the baseline feature vector under typical working conditions, and the matching process calculates the cosine similarity. Feature vectors with a similarity less than 0.85 are excluded. The screened feature candidate set performs non-uniform sampling: according to the energy amplitude gradient distribution, sampling points are densified in areas where the gradient change rate exceeds 5%. Each sampling point generates a 32-bit position encoding and a 16-bit pose encoding. The position encoding is based on the Cartesian coordinate system to quantify displacement deviation, and the pose encoding uses a quaternion rotation angle compression algorithm. Finally, the platform pose feature set is output, with a data structure of a time series tensor containing a feature sequence of 2000 sampling points per second.

[0078] Based on the platform pose feature set, an environmental disturbance perception model is called to perform analysis. The model first analyzes the position encoding sequence and maps it to the physical coordinate space of the temperature, pressure, and vibration sensors. An environmental parameter distribution matrix is generated through cubic spline interpolation, with the matrix dimension aligned with the motion platform workspace. Each matrix element integrates multi-sensor data, such as temperature gradient parameters generated by fusing infrared thermal imager data and thermocouple data. Gradient field calculation is performed on the environmental parameter distribution matrix: three-dimensional gradient detection is performed using the Sobel operator, and an environmental disturbance intensity distribution map is output. High-intensity disturbance areas (gradient value ≥ 15 units) and low-intensity areas (gradient value ≤ 3 units) are marked in the map, and the spatial direction of the disturbance vector is recorded.

[0079] The spatio-temporal convolution subnet built into the environmental disturbance perception model adopts a three-layer convolution structure: the first layer is a 3x3x3 convolution kernel that extracts local disturbance patterns; the second layer applies a dilated convolution to expand the receptive field; and the third layer compresses the features through max pooling. The subnet outputs a 256-channel environmental disturbance feature tensor, with tensor elements encoding disturbance amplitude, frequency, and spatial correlation. At the same time, a mechanical deformation prediction model is started in parallel: the model loads the platform pose feature set and calls a finite element analysis engine to perform stress simulation. The finite element mesh division precision reaches 0.1 mm, and the boundary conditions are dynamically adjusted according to the real-time pose encoding. The output is a mechanical deformation probability distribution map, with the probability value converted into a deformation risk index by the von Mises stress criterion.

[0080] The feature fusion stage performs a tensor dot product operation: align the environmental disturbance feature tensor (256 dimensions) with the mechanical deformation probability distribution (64 dimensions) in the feature dimension, and expand the latter dimension by transposed convolution. The fusion process uses Hadamard product to calculate the cross features, generating a 320-dimensional feature fusion result. This result preserves the spatio-temporal properties of the original tensor, such as the coupling features of high-frequency vibration bands and local deformation regions encoded in high-order tensors. The final output multi-dimensional data array contains timestamp markers and can be directly input into the downstream compensation model.

[0081] Embodiment 2: refer to Figure 3 The input interface of the dynamic compensation model receives the feature fusion result from the upstream processing flow, and the data structure is a 320-dimensional tensor sequence with a time resolution of 1 ms. The model initialization stage loads a preset trajectory constraint condition library, which contains performance boundary parameters of the motion platform under various working conditions, including maximum allowed acceleration, velocity smoothness coefficient, and position tracking error threshold. The constraint condition parser associates the feature fusion result with the current task number, and retrieves the corresponding constraint parameter group through hash indexing. Each parameter group contains 12 weight factors, corresponding to the compensation priorities of position, attitude, and acceleration in three dimensions. The assignment logic of the weight factors is based on the platform dynamics characteristics, such as increasing the acceleration compensation weight in high-speed motion segments and enhancing the position compensation weight in precision positioning segments.

[0082] The construction process of the error transfer matrix adopts a hierarchical calculation architecture. The first layer processes the position compensation quantity, multiplies the position-related dimensions (first 128 dimensions) in the feature fusion result with the constraint weight, generating a position error vector. The second layer analyzes the attitude compensation demand, extracts 64-dimensional attitude feature data from the feature tensor, and converts it into Euler angle deviation through quaternion interpolation algorithm. The third layer calculates the acceleration compensation component, and the remaining 128-dimensional feature data is reconstructed into time-domain acceleration spectrum through inverse Fourier transform. The three-layer outputs are combined through Kronecker product to form a 24x24 error transfer matrix, with diagonal elements representing compensation quantities of each degree of freedom, and non-diagonal elements encoding the coupling effects between degrees of freedom.

[0083] The traversal ordering of the trajectory point sequence uses an improved A* search algorithm. The algorithm heuristic function considers both the compensation quantity and the motion smoothness requirement, generating an optimal access sequence in a search space of 500 trajectory points. Each trajectory point is associated with 6 key elements of the error transfer matrix (X / Y / Z position deviation and Rx / Ry / Rz attitude deviation), and the priority score of the point is updated in real time during the ordering process. The reinforcement learning strategy deploys a double deep Q network architecture, with the main network evaluating the Q value of the current state-action pair and the target network providing a stable learning target. The action space is defined as 7 basic compensation operations, including linear displacement compensation, angular displacement compensation, and composite compensation mode. The reward function design focuses on three aspects: position error convergence speed, energy consumption efficiency, and vibration suppression effect.

[0084] The generation of the base compensation instruction set undergoes three processing stages. The first stage performs coarse compensation, selecting a preset compensation template according to the action type output by the Q network. The second stage performs fine adjustment, using a PID control algorithm to fine-tune the compensation amount amplitude. The third stage implements safety verification, comparing the compensation instructions with platform physical limit parameters, and eliminating instruction items that exceed the execution capability. The final generated instruction set contains three types of data: position compensation amount represented by nanometer-level floating point numbers, attitude compensation amount stored as normalized quaternions, and acceleration compensation amount recorded as differential pulse numbers.

[0085] The instruction encoding integration module realizes the standardized conversion of control signals. The platform dynamic response parameter database stores the response delay, gain characteristics and nonlinear correction table of each axis motor. The integration process first performs time alignment processing on the base compensation instructions to compensate for the response speed differences of each axis. Then, the inverse dynamics model is applied for conversion, converting the target compensation amount into motor drive signals. The encoding output uses the industrial Ethernet protocol, and each data packet contains 32 bytes of control instructions and 8 bytes of check code. The real-time guarantee mechanism uses hardware timestamp synchronization to ensure that the control period is strictly controlled within 1 ms.

[0086] The feedback data acquisition system is composed of a multi-source sensor network. Optical encoders record the actual position of the platform at a sampling rate of 50 kHz, and a vibration sensor array monitors the mechanical resonance in six degrees of freedom. The data preprocessing unit performs sensor fusion and eliminates measurement noise through Kalman filtering. The feedback log generator converts the raw data into structured records, with each log containing a timestamp, trajectory offset spectrum, and vibration feature fingerprint. The dynamic error feature extractor uses sliding window analysis, with the window width adaptively adjusted according to the motion speed, and the typical value is 20 ms. The feature vector contains 15 key indicators, such as the root mean square value of the position error and the energy proportion of the main frequency band of the vibration.

[0087] The historical control case library uses a time series database architecture to store control records from the past 30 days. The pattern matching engine uses the dynamic time warping algorithm to find similar patterns between the current error features and historical cases. The matching process considers three dimensions: working condition similarity, error pattern similarity, and control effect, returning the top K best matching cases. The adaptive compensation parameter generator synthesizes the compensation strategies of these cases and generates new control parameters through a weighted voting mechanism. The parameter update system uses a hot switching mechanism, and the new parameters take effect immediately after verification, without the need to restart the control system.

[0088] The threshold update of the motion feature analysis network adopts an incremental learning strategy. The dynamic response threshold is stored in a distributed key-value database, and the update request is processed asynchronously through a message queue. The learning rate scheduler automatically adjusts the update amplitude according to the error convergence, and the initial learning rate is set to 0.001. The threshold verification module performs a quick test after each update, using the data of the last 100 control cycles for verification testing. The anomaly detector monitors the update process and automatically rolls back to the last stable version when the threshold drift is found to be outside the safe range.

[0089] Embodiment 3: refer to Figure 4 The processing flow of dynamic error features starts with the time series segmentation of the feedback log. The sensor system continuously records the platform motion data at a sampling rate of 20 kHz, generating an original log containing trajectory offsets, vibration spectra, and environmental parameters. The time series segmenter uses an adaptive window width algorithm, with a window length Decided by the dynamic characteristics of the current motion state:

[0090]

[0091] Where: : preset maximum window width, : preset minimum window width, : second norm of velocity change in window, : window width adjustment coefficient, : small constant to prevent division by zero.

[0092] The length of the segmented error segment is dynamically adjusted between 5ms and 50ms, and each segment is associated with a corresponding platform motion state sequence, including a six-degree-of-freedom velocity profile and an acceleration distribution.

[0093] The error classification model uses a multi-scale convolutional neural network architecture. The input layer receives the standardized error segment, with a data dimension of time steps x sensor channels. The first convolutional layer uses a set of parallel convolutional kernels with variable width, with narrow kernels (length ≤ 5) capturing high-frequency vibration features and wide kernels (length ≥ 15) extracting low-speed drift patterns. The feature fusion layer dynamically weights each scale feature through a gating mechanism, outputting a 128-dimensional mixed feature vector. The classification head contains three fully connected branches, corresponding to the mechanical resonance, environmental disturbance, and control system lag three types of root cause labels. The label probability is adjusted by the branch-specific temperature coefficient to avoid overconfidence in the classification result.

[0094] The retrieval process of the historical control case library adopts a hierarchical indexing strategy. The first-level index is hash-sharded according to motion task type, and the second-level index is constructed based on error spectrum features using a KD-tree. The query phase first locates the task type partition, and then performs an approximate nearest neighbor search in the feature space. The generation of the candidate compensation parameter set involves two screening steps: initial screening retains the top 100 similar cases, and fine screening calculates the temporal matching degree using a dynamic time warping algorithm, ultimately retaining cases with a matching degree higher than 0.7. The parameter optimization module uses Pareto front analysis to find non-dominated solutions across three objective dimensions: error convergence speed, energy consumption, and vibration suppression.

[0095] The injection of virtual trajectory disturbance parameters is triggered 5ms before the control signal output. The disturbance generator includes three modes: white noise disturbance, which adds a Gaussian random variable in the time domain, with the amplitude limited by the platform's maximum permissible deviation; harmonic disturbance, which simulates mechanical resonance at a specific frequency, with the frequency value randomly selected from the platform's inherent frequency library; and step disturbance, which generates sudden positional transitions to test the system's transient response capability. The three disturbance modes are mixed in a preset ratio, and the mixing weight is dynamically adjusted according to the current motion phase.

[0096] The robustness assessment module monitors the difference in trajectory correction before and after the disturbance. The assessment metrics include both time-domain and frequency-domain metrics: time-domain metrics calculate the point-by-point deviation between the actual trajectory and the target trajectory, while frequency-domain metrics analyze the spectral distribution of residual vibration energy. Robustness score. Weighted synthesis from multidimensional indicators:

[0097]

[0098] in: : No. The deviation value of each evaluation dimension, The tolerance threshold for the corresponding dimension. Preset weighting coefficients : Total number of evaluation dimensions. When When the value falls below the warning threshold of 0.6, the system automatically activates the parameter learning mode. Before starting the learning mode, a data snapshot is taken to save a complete record of the state before and after the disturbance, including the controller's internal state variables and the raw sensor data.

[0099] The online update of the motion feature analysis network adopts an elastic weight solidification strategy. The adjustment range of the network parameters is determined by two factors: the importance weight of the current parameters and the gradient direction of the new data. The importance weight is approximated by the diagonal elements of the Fisher information matrix, recording the sensitivity of the parameters in historical tasks. The gradient calculation uses the comparison data before and after the perturbation, and the loss function includes two components: trajectory position error and feature reconstruction error. The parameter update process applies an elastic constraint to prevent drastic changes in high-importance parameters. The proportion of parameters updated each time is controlled within 15%, and the bias terms of the convolution kernels and the parameters of the normalization layer are adjusted first.

[0100] The stability monitoring of the control system runs throughout the entire implementation process. The monitor tracks the moving average of key performance indicators, including command delay, computational load, and communication error rate. The anomaly detection algorithm is based on the Isolation Forest model, which identifies abnormal deviations in the indicator data. When persistent performance degradation is detected, the system automatically reverts to the parameter configuration of the last stable version and generates a diagnostic report for subsequent analysis. The state information of all subsystems is synchronized by a high-precision clock, with a timestamp accuracy of 100 nanoseconds, ensuring the timing accuracy of fault analysis.

[0101] Cross-protocol data conversion is performed at the end of the command transmission link. The protocol adapter maintains a library of device description files, each defining the command syntax, data encoding, and timing requirements of the target device. The conversion process uses an intermediate representation layer architecture: first, the platform driver compensates the command by parsing it into a sequence of device-independent atomic operations, and then re-encodes it according to the description file of the target device.

[0102] Example 4: The parameter update process of the motion feature analysis network adopts a backpropagation mechanism based on trajectory difference data. This process starts with spatiotemporal alignment processing, where the system collects two sets of trajectory data before and after perturbation, each containing the position coordinates of the platform in the X / Y / Z three-axis directions and the attitude data of the three rotational degrees of freedom Rx / Ry / Rz. The alignment algorithm first establishes a time axis mapping relationship, unifying the time base of the two sets of data through dynamic time warping technology, eliminating the deviation caused by different sampling times. The aligned data forms a matching sequence, each sequence point containing the corresponding state values before and after perturbation.

[0103] The trajectory offset calculation module processes the aligned matching sequence. For each trajectory point, the system calculates the position deviation vector and the attitude deviation vector. The position deviation vector records the displacement difference in the three-axis directions, and the attitude deviation vector calculates the rotation difference using the quaternion difference method. These deviation values are converted into a unified error metric, forming the training samples for network update. Part of the trajectory deviation data collected during a certain update process is shown in Table 1.

[0104] Table 1: Part of the trajectory deviation data collected during a certain update process.

[0105] Timestamp (ms) X deviation (nm) Y deviation (nm) Z deviation (nm) Rx deviation (μrad) Ry deviation (μrad) Rz deviation (μrad) 1024 12.5 8.7 15.2 2.1 1.8 3.5 1048 14.3 9.2 16.8 2.5 2.1 4.2 1072 11.8 7.9 14.5 1.9 1.6 3.1 1096 13.6 8.4 15.9 2.3 1.9 3.8

[0106] The construction of the motion regression loss function considers the comprehensive influence of multi-dimensional errors. The position deviation adopts a weighted Euclidean distance metric, and the weight coefficients of the three axes are dynamically configured according to the platform structure characteristics. The attitude deviation calculation adopts a quaternion logarithmic mapping method to convert the rotation difference into an equivalent axis angle representation. The loss function integrates the deviation data of all trajectory points to generate a scalar value reflecting the overall control error. This value is used as the objective function for network parameter optimization, driving the subsequent backpropagation process.

[0107] The generation of the loss value tensor is realized using automatic differentiation technology. The system constructs a computation graph to record the complete operation process from input data to loss value. The nodes of the computation graph include various tensor operation operations such as convolution, pooling, matrix multiplication, etc. In the backpropagation phase, the system calculates the gradient value of each parameter in reverse along the computation graph. The gradient calculation process uses the chain rule to derive layer by layer from the loss function, ultimately obtaining the partial derivative of each trainable parameter.

[0108] The calculation process of the convolution kernel gradient matrix has the characteristics of selective updating. The system maintains a parameter importance score table, which records the sensitivity index of each convolution kernel in historical tasks. When calculating the gradient matrix, the gradients of important parameters are amplified, while the gradients of secondary parameters are correspondingly attenuated. This selective updating mechanism ensures that the network does not destroy important feature representations learned when adapting to new data. Each iteration update, the system only adjusts about 15% of the network parameters, prioritizing the convolution kernels with the highest relevance to the current error pattern.

[0109] Parameter updating adopts a phased and progressive strategy. In the initial stage, a larger learning rate is used to quickly approach the optimal solution, and the learning rate is gradually attenuated as the number of iterations increases. The update process includes three verification links: first, check the rationality of the gradient value, exclude abnormal values; second, evaluate the change of network output after parameter update, ensure that the error is indeed reduced; finally, test the performance of the network on the validation set to prevent overfitting. Only updates that pass all verification links will be finally applied.

[0110] The performance evaluation of the updated network uses the cross-validation method. The system reserves part of the trajectory data as the test set, which does not participate in the training process. After updating, the test set data is used to evaluate the generalization ability of the network. Evaluation indicators include position control accuracy, attitude stability, and response speed, etc. If the test results do not meet the requirements, the system will automatically roll back to the last stable version and adjust the update strategy for retraining.

[0111] The entire parameter updating process has a real-time guarantee mechanism. The system uses a pipeline processing architecture to parallelize data acquisition, preprocessing, training, and verification. The dynamic scheduling algorithm is used to allocate computing resources, ensuring that network updates do not affect normal control task execution. The end-to-end delay of each update is controlled within 50ms, meeting the real-time requirements of high-precision motion control.

[0112] The updated network parameters are securely stored in a multi-version management system. Each version is accompanied by complete metadata, including training time, data source, and performance indicators. The system maintains a parameter version library, supporting quick rollback and version comparison functions. When performance degradation is detected, the appropriate parameter configuration can be selected from the version library for recovery.

[0113] The persistent storage of network parameters uses differential encoding technology. Each update only stores the parameter differences relative to the previous version, significantly reducing storage space occupancy. The instant recovery mechanism is used during parameter loading, which can complete the network state switching within milliseconds. This design not only ensures the security of parameters, but also meets the strict real-time requirements of control systems.

[0114] The system also implements a parameter health monitoring mechanism. Regularly scan the network parameter value range to detect possible value overflow or gradient disappearance problems. When abnormal parameters are found, automatically trigger the repair process, including parameter reinitialization, gradient clipping, and other recovery methods. The monitoring results will generate detailed reports to help analysts understand the evolution trend of network parameters.

[0115] Embodiment 5: refer to Figure 5 The construction of the cross-platform control instruction adapter is based on a protocol abstraction layer architecture. The architecture includes a device description file library, an instruction conversion engine, and a protocol verification module. The device description file is stored in XML format, defining the data encoding rules, control instruction syntax, and timing constraints of the target driven device. Each description file contains an instruction set mapping table, which clearly shows the correspondence between the source instruction opcode and the target device opcode, and the parameter transfer path details the assembly order and byte sequence rules of multi-byte parameters. The conversion engine loads all device description files during initialization and establishes an in-memory index table to speed up the retrieval process.

[0116] After the platform driver compensation instruction input is converted into the engine, the instruction structure is first parsed. The parser identifies the atomic operation type of the instruction, such as position setting, speed curve loading or emergency braking command. Each atomic operation is decomposed into an operation code field and a parameter data segment to generate intermediate control representation layer data. The representation layer data is organized in a tree structure, with the root node identifying the operation category and the child nodes storing parameter key-value pairs. The engine iteratively queries the control instruction mapping table in memory to locate the equivalent instruction of the target device through operation code hashing. The mapping process takes into account the differences in device capabilities, and when the target device does not support a specific operation, the engine automatically decomposes it into multiple basic instruction combinations.

[0117] The semantic adaptation process of the parameter transmission path solves the protocol conflict problem. The conflict detector analyzes the bit distribution difference between the source parameters and the target parameters, and identifies three conflict types: precision mismatch (such as 32-bit floating point to 16-bit fixed point), unit inconsistency (nanometer to micrometer) or parameter missing. The adapter adopts different processing strategies for different types: precision conversion performs linear scaling and adds rounding compensation; unit conversion adjusts through a proportionality coefficient; missing parameters are filled by the default value generator based on context inference. The preservation of timing constraints uses a double buffering mechanism: the instruction sequence is temporarily stored in a circular queue with timestamps, and a waiting period is inserted according to the device response delay before output, maintaining the original time interval of adjacent instructions.

[0118] The reservation of dynamic compensation context realizes the integrity of instruction semantics. The conversion engine encapsulates a special data structure to record the metadata of the current compensation operation: including the compensation trigger reason, the previous compensation effect and the expected control target. This data structure is embedded as a hidden field in the output instruction, which can be used by the device driver to optimize the local control strategy. The injection of precision optimization parameters is implemented in the final encoding stage: the resolution coefficient defined in the device description file acts on the parameter encoding process, for example, re-quantizing a position quantity of 1.25 nanometers to 2 nanometer resolution, and the quantization error is accumulated to subsequent instruction compensation.

[0119] The control instruction stream generation adopts a frame transmission strategy. A single frame structure contains a frame header, an instruction body and a check code: the frame header marks the data stream type and the frame sequence number; the instruction body encapsulates the operation code and parameters according to the device requirements; the check code uses the CRC-32 algorithm to guarantee the integrity of the transmission. The flow control module monitors the device state feedback and automatically adjusts the frame sending rate when it detects the risk of buffer overflow. The device registration module realizes the plug-and-play function: when a new device is connected, it reports the description information, and the engine dynamically updates the memory index table and mapping rules, without the need to restart the system to complete the protocol adaptation. The entire conversion process meets the strict time constraints: the delay of batch conversion of 1000 instructions does not exceed 5 milliseconds.

[0120] The security verification mechanism runs through the whole conversion process. Syntax checker detects the compliance of output instructions, excluding out-of-bound parameters and illegal operation codes; semantic analyzer verifies the logical consistency of instruction sequence, preventing conflict operations; timing checker calibrates the interval deviation between instructions through hardware clock. Instructions that fail verification trigger rollback mechanism: original compensation instructions are stored in the safe zone, while a diagnostic report is generated to locate protocol adaptation defects. This implementation supports seamless access of ten types of industrial driving devices, including linear motor platforms, piezoelectric ceramic actuators, and magnetic levitation tables, etc. heterogeneous systems.

[0121] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0122] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences, modifications, substitutions and variations of the embodiments and can be modified in various ways by those skilled in the art without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An adaptive control method for a micro-nano high-precision motion platform, characterized in that, include: Acquire real-time pose feedback data and target trajectory data from the motion platform, and extract dynamic response features from the pose feedback data; The trained motion feature parsing network is invoked to perform multimodal feature separation processing on the real-time pose feedback data, generating a platform pose feature set. Based on the platform pose feature set, the environmental disturbance perception model is invoked to perform spatiotemporal coupling analysis on the working environment parameters, generating a feature fusion result that includes mechanical deformation features and environmental disturbance features. The feature fusion result is input into the dynamic compensation model to calculate the trajectory correction amount, and the platform-driven compensation command is output. The target trajectory data is calibrated in real time according to the platform drive compensation command to generate the actual control signal of the motion platform.

2. The adaptive control method of the micro-nano high-precision motion platform according to claim 1, wherein, The call to the trained motion feature parsing network performs multimodal feature separation processing on the real-time pose feedback data to generate a platform pose feature set, including: The real-time pose feedback data is decomposed in the time-frequency domain to generate multiple frequency band feature components and their corresponding phase response parameters. Extract the energy distribution map of each frequency band feature component, and perform spatial attention weighting processing on the energy distribution map to generate a weighted motion feature vector; The motion feature vectors are matched with a preset dynamic response threshold for similarity to select a set of feature candidates that meet the confidence criteria. The candidate feature set is subjected to non-uniform sampling to generate the platform pose feature set containing position encoding and pose encoding.

3. The adaptive control method of the micro-nano high-precision motion platform according to claim 2, wherein, Based on the platform pose feature set, the environmental disturbance perception model is invoked to perform spatiotemporal coupling analysis on the working environment parameters, generating a feature fusion result containing mechanical deformation features and environmental disturbance features, including: Based on the position encoding in the platform pose feature set, the working environment parameters are processed by region mapping to generate an environment parameter distribution matrix; The gradient field is calculated on the environmental parameter distribution matrix to generate an environmental disturbance intensity distribution map. The spatiotemporal convolutional subnet in the environmental disturbance perception model is invoked to extract local features from the environmental disturbance intensity distribution map, generating an environmental disturbance feature tensor. The mechanical deformation prediction model is invoked in parallel to match the stress distribution of the platform pose feature set and generate a mechanical deformation probability distribution. The environmental disturbance feature tensor and the mechanical deformation probability distribution are cross-fused to generate the feature fusion result.

4. The adaptive control method of the micro-nano high-precision motion platform according to claim 3, wherein, The step of inputting the feature fusion result into the dynamic compensation model to calculate the trajectory correction amount and outputting platform-driven compensation instructions includes: The trajectory constraints of the current motion task are analyzed, and compensation priority weights associated with the feature fusion results are generated. An error propagation matrix is ​​generated based on the compensation priority weights, and the trajectory point sequence in the target trajectory data is traversed and sorted. The optimal compensation sequence is selected from the error propagation matrix using a reinforcement learning strategy to generate a basic compensation instruction set containing position compensation, attitude compensation, and acceleration compensation. The basic compensation instruction set and the platform dynamic response parameters are integrated by instruction encoding to generate the platform-driven compensation instruction.

5. The adaptive control method of the micro-nano high-precision motion platform according to claim 4, wherein, The method further includes: Real-time acquisition of platform execution state data during motion control, generating a feedback log containing trajectory offset and vibration spectrum features; Extracting dynamic error features from the feedback log and performing pattern matching with the historical control case library to generate adaptive compensation parameters; Online updating of the dynamic response threshold of the motion feature analysis network based on the adaptive compensation parameters; Injecting the updated dynamic response threshold into the dynamic compensation model to recalculate the compensation priority weight in the error transfer matrix.

6. The adaptive control method of the micro-nano high-precision motion platform according to claim 5, wherein, The method further comprises: Time series segmentation processing of the dynamic error features to generate multiple error segments and their corresponding platform motion state sequences; Calling a trained error classification model to perform root cause analysis on each error segment and generating classification labels containing mechanical resonance, environmental interference, and control system lag; Retrieving compensation strategy templates matching the classification labels from the historical control case library to generate a candidate compensation parameter set; Based on the matching degree of the platform motion state sequence and the candidate compensation parameter set, the parameter with the highest confidence is selected to generate the adaptive compensation parameters.

7. The adaptive control method of the micro-nano high-precision motion platform according to claim 1, wherein, The method further comprises: Injecting a virtual trajectory disturbance parameter before outputting the control signal, which is used to simulate sudden platform motion offset scenarios; Monitoring the correction processing results of the dynamic compensation model on the disturbed trajectory to generate a control robustness evaluation index; When the control robustness evaluation index is lower than the preset threshold, triggering the parameter learning mode of the motion feature analysis network; Gradient backpropagation updating of the convolution kernel parameters of the motion feature analysis network based on the trajectory difference data before and after disturbance.

8. The adaptive control method of the micro-nano high-precision motion platform according to claim 7, wherein, The method further comprises: Temporal and spatial alignment processing of the trajectory difference data before and after disturbance to generate a matching sequence containing a set of trajectory points before disturbance and a set of trajectory points after disturbance; Extracting trajectory offset from the matching sequence to generate position deviation vectors and attitude deviation vectors for each trajectory point; Constructing a motion regression loss function based on the position deviation vectors and attitude deviation vectors; Differentiable transformation of the motion regression loss function to generate a loss value tensor required for gradient backpropagation; Iterative incremental adjustment of the convolution kernel parameters of the motion feature analysis network based on the convolution kernel gradient matrix and the preset learning rate parameter. The method further comprises:

9. The adaptive control method of the micro-nano high-precision motion platform according to claim 1, wherein, Constructing a cross-platform control instruction adapter and analyzing the instruction protocol differences of different driving devices through the cross-platform control instruction adapter; Converting the platform driving compensation instructions into underlying control signals supported by the target driving device; Preserving the timing constraint relationship and dynamic compensation context of the actual control signal during conversion; ​ Injecting the precision optimization parameters matched with the target driving device to generate motion control instruction stream meeting cross-platform execution conditions.

10. The adaptive control method of the micro-nano high-precision motion platform according to claim 9, wherein, The cross-platform control instruction adapter is constructed, and instruction protocol differences of different driving devices are resolved through the cross-platform control instruction adapter, including: A driving device protocol rule library is established to store control instruction mapping tables and parameter transmission paths of each platform; The platform driving compensation instructions are subjected to instruction structure analysis to generate intermediate control representation layer data; Based on the intermediate control representation layer data, the control instruction mapping table is queried to generate a device-compatible instruction conversion scheme; The parameter transmission paths with conflicts are subjected to semantic adaptation processing to generate an unambiguous control instruction conversion result.

Images