Adaptive control method for micro-nano high-precision motion platform

Abstract

The application relates to the technical field of micro-nano motion control, and discloses an adaptive control method for a micro-nano high-precision motion platform. The method acquires real-time pose feedback data and target trajectory data of the motion platform, and extracts dynamic response characteristics; a trained motion characteristic analysis network is called to separate multi-modal characteristics, and a platform pose characteristic set is generated; based on the set, time-space coupling analysis is performed on working environment parameters through an environment disturbance perception model, and a fusion result containing mechanical deformation characteristics and environment disturbance characteristics is obtained; subsequently, the fusion result is input into a dynamic compensation model to calculate a trajectory correction amount, and a driving compensation instruction is output; the target trajectory data is calibrated in real time according to the compensation instruction, and an actual control signal is generated. Through multi-modal characteristic analysis, time-space 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 micro-nano high-precision motion control scenes.

Description

Technical Field

[0001] This invention relates to the field of micro-nano motion control technology, specifically to an adaptive control method for a micro-nano-level high-precision motion platform. Background Technology

[0002] In high-end manufacturing and scientific research fields such as semiconductor manufacturing, precision measurement, and biomedicine, micro-nano-level high-precision motion platforms are core equipment for achieving precision operations. Their operational accuracy directly affects product quality and experimental results, thus placing extremely high demands on motion control technology.

[0003] Traditional motion platform control methods often employ model-based PID control or linear control strategies, which rely on precise mathematical models. However, the mechanical structures of micro / nano-scale motion platforms are complex and contain nonlinear factors such as friction, clearance, and flexible deformation, which are difficult to fully describe through precise modeling. Furthermore, external factors in the working environment, such as temperature fluctuations, vibrations, and airflow disturbances, can further exacerbate the platform's pose deviations.

[0004] With technological advancements, the precision requirements for motion platforms have increased from the micrometer level to the nanometer level, making it difficult for traditional control methods to handle complex dynamic characteristics and environmental disturbances. For example, in the control of a lithography machine stage, even minute temperature changes or mechanical vibrations can cause trajectory deviations, affecting lithography accuracy. Furthermore, the dynamic response of the motion platform varies under different loads and operating speeds, and fixed-parameter control methods cannot achieve adaptive adjustment, leading to decreased control accuracy and insufficient stability.

[0005] Existing adaptive control methods are mostly based on single features or models for adjustment, lacking comprehensive analysis of multimodal characteristics and failing to fully reflect the dynamic characteristics of the motion platform. Furthermore, in terms of environmental disturbance perception and compensation, they fail to effectively integrate spatiotemporal coupling, leading to compensation lag or inaccuracy, thus limiting further improvements in control accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive control method for a micro-nano-scale high-precision motion platform to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an adaptive control method for a micro / nano-scale high-precision motion platform, the method comprising:

[0008] Acquire real-time pose feedback data and target trajectory data from the motion platform, and extract dynamic response features from the pose feedback data;

[0009] 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.

[0010] 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.

[0011] 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.

[0012] 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.

[0013] Preferably, the step of calling the trained motion feature parsing network to perform multimodal feature separation processing on the real-time pose feedback data to generate a platform pose feature set includes:

[0014] 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.

[0015] 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;

[0016] 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.

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

[0018] Preferably, the step of calling the environmental disturbance perception model to perform spatiotemporal coupling analysis on the working environment parameters based on the platform pose feature set, and generating a feature fusion result containing mechanical deformation features and environmental disturbance features, includes:

[0019] 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;

[0020] The gradient field is calculated on the environmental parameter distribution matrix to generate an environmental disturbance intensity distribution map.

[0021] 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.

[0022] 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.

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

[0024] Preferably, the step of inputting the feature fusion result into the dynamic compensation model to calculate the trajectory correction amount and outputting the platform-driven compensation command includes:

[0025] The trajectory constraints of the current motion task are analyzed, and compensation priority weights associated with the feature fusion results are generated.

[0026] 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.

[0027] 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.

[0028] The basic compensation instruction set and the platform dynamic response parameters are integrated by instruction encoding to generate the platform-driven compensation instruction.

[0029] Preferably, the method further includes:

[0030] During motion control, the platform's execution status data is collected in real time to generate feedback logs containing trajectory offset and vibration spectrum characteristics.

[0031] Extract the dynamic error features from the feedback log, and perform pattern matching between the dynamic error features and the historical control case library to generate adaptive compensation parameters;

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

[0033] The updated dynamic response threshold is injected into the dynamic compensation model, and the compensation priority weights in the error propagation matrix are recalculated.

[0034] Preferably, the step of extracting dynamic error features from the feedback log and performing pattern matching between the dynamic error features and the historical control case library to generate adaptive compensation parameters includes:

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

[0036] The trained error classification model is invoked to perform root cause analysis on each error segment, generating classification labels that include mechanical resonance, environmental interference, and control system hysteresis.

[0037] Retrieve compensation strategy templates that match the classification labels from the historical control case library to generate a set of candidate compensation parameters;

[0038] Based on the matching degree ranking between the platform motion state sequence and the candidate compensation parameter set, the parameter with the highest confidence is selected to generate the adaptive compensation parameter.

[0039] Preferably, the method further includes:

[0040] Virtual trajectory disturbance parameters are injected before the control signal output, and the virtual trajectory disturbance parameters are used to simulate the sudden movement deviation scenario of the platform;

[0041] Monitor the results of the dynamic compensation model's correction of the trajectory after disturbance, and generate control robustness evaluation index;

[0042] When the control robustness evaluation index is lower than a preset threshold, the parameter learning mode of the motion feature parsing network is triggered.

[0043] The convolution kernel parameters of the motion feature parsing network are updated by gradient backpropagation based on the trajectory difference data before and after the perturbation.

[0044] Preferably, the step of updating the convolutional kernel parameters of the motion feature parsing network based on the trajectory difference data before and after the perturbation includes:

[0045] The trajectory difference data before and after the disturbance are spatiotemporally aligned to generate a matching sequence containing the set of trajectory points before the disturbance and the set of trajectory points after the disturbance.

[0046] Extract the trajectory offset from the matching sequence to generate the position deviation vector and attitude deviation vector for each trajectory point;

[0047] A motion regression loss function is constructed based on the position deviation vector and the attitude deviation vector;

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

[0049] Iterate through the convolutional kernel parameters of the motion feature parsing network, calculate the partial derivative of the loss tensor with respect to each convolutional kernel, and generate the convolutional kernel gradient matrix;

[0050] Based on the convolution kernel gradient matrix and the preset learning rate parameters, the convolution kernel parameters of the motion feature parsing network are iteratively and incrementally adjusted.

[0051] Preferably, the method further includes:

[0052] A cross-platform control instruction adapter is constructed, and the differences in instruction protocols between different driving devices are resolved through the cross-platform control instruction adapter;

[0053] The platform drive compensation command is converted into the underlying control signal supported by the target drive device;

[0054] The timing constraints and dynamic compensation context of the actual control signals are preserved during the conversion process;

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

[0056] Preferably, the step of constructing a cross-platform control instruction adapter and resolving the instruction protocol differences between different driver devices through the cross-platform control instruction adapter includes:

[0057] Establish a driver device protocol rule base and store the control command mapping table and parameter transmission path for each platform;

[0058] The platform-driven compensation instructions are parsed to generate intermediate control representation layer data;

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

[0060] Semantic adaptation is performed on conflicting parameter passing paths to generate unambiguous control command conversion results.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

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

[0063] The trained motion feature parsing network is invoked to perform multimodal feature separation processing, generating a platform pose feature set. This allows for the separation of various relevant features from complex pose data. These features cover different aspects of platform motion, such as position, attitude, and velocity. Through comprehensive analysis of these multimodal features, a more complete understanding of the platform's motion characteristics can be achieved, providing rich data for subsequent control and adjustment.

[0064] Based on the platform's pose feature set, an environmental disturbance perception model is invoked to perform spatiotemporal coupling analysis on the working environment parameters, generating feature fusion results that include mechanical deformation features and environmental disturbance features. This achieves a comprehensive consideration of environmental factors and the changes in the machine itself. Environmental disturbances and mechanical deformations interact in space and time. Through this coupling analysis, the combined effect of the two on the platform's motion can be accurately identified, avoiding the one-sidedness that may occur when considering only one factor.

[0065] The feature fusion results are input into the dynamic compensation model to calculate the trajectory correction amount, and the output platform-driven compensation command ensures that the compensation command accurately corresponds to the actual cause of the trajectory deviation. The dynamic compensation model calculates the correction amount based on the fused features and can provide corresponding compensation strategies for different disturbances and deformations, ensuring the pertinence and effectiveness of the compensation.

[0066] The platform-driven compensation commands are used to perform real-time calibration of the target trajectory data, generating the actual control signals for the motion platform. This achieves a closed-loop process from data acquisition and feature analysis to compensation control. This closed-loop control continuously calibrates the trajectory, ensuring that the platform's actual motion trajectory always remains consistent with the target trajectory. This reduces trajectory deviations caused by various factors and improves the stability and trajectory accuracy of the motion platform under complex environments and dynamic conditions. Attached Figure Description

[0067] Figure 1 This is a schematic diagram illustrating the working principle of the adaptive control method for the micro / nano-level high-precision motion platform described in this invention.

[0068] Figure 2 A flowchart for environmental disturbance perception and feature fusion;

[0069] Figure 3 A flowchart for dynamic compensation and instruction generation;

[0070] Figure 4 A flowchart for dynamic error feature matching;

[0071] Figure 5 A flowchart for adapting control commands to cross-platform use. Detailed Implementation

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

[0073] Please see Figure 1This invention provides an adaptive control method for a micro / nano-scale high-precision motion platform, the method comprising:

[0074] Dynamic calibration of the platform's pose is achieved through integrated data processing and model invocation. Specifically, this includes: acquiring real-time pose feedback data and target trajectory data of the motion platform; using a sensor system to collect the platform's six-degree-of-freedom position and attitude information; and extracting dynamic response features from the pose feedback data, such as acceleration spectra and displacement rates of change. Subsequently, a pre-trained motion feature parsing network is invoked to perform multimodal feature separation processing on the real-time pose feedback data. This network, based on a deep convolutional structure, separates mechanical vibration features and motion offset components through multiple layers to generate a platform pose feature set, which includes position and attitude encoding vectors. Based on this platform pose feature set, an environmental disturbance perception model is invoked to perform spatiotemporal coupling analysis of the working environment parameters. These environmental parameters include real-time monitored variables such as temperature gradients, air pressure fluctuations, and vibration disturbances. The model integrates nonlinear filtering and regression algorithms to generate a feature fusion result containing mechanical deformation features and environmental disturbance features. This feature fusion result is used as input to a dynamic compensation model for trajectory correction calculation. The dynamic compensation model employs a reinforcement learning framework and outputs platform-driven compensation commands, including position correction and attitude adjustment values. According to the platform-driven compensation instructions, the target trajectory data is calibrated in real time. The calibration process involves interpolation algorithms and constraint optimization, updating the trajectory point sequence to generate the actual control signal for the motion platform. This signal directly drives the platform's actuator system, ensuring trajectory tracking accuracy at the micro-nano level.

[0075] Example 1: See Figure 2 Real-time pose feedback data is acquired via a high-precision encoder and laser interferometer, with a sampling frequency of no less than 20kHz. The raw data includes six-degree-of-freedom coordinates and attitude quaternions. Time-frequency domain decomposition processing is performed on the real-time pose feedback data using a discrete wavelet transform algorithm to decompose the raw signal into eight frequency band feature components. Each frequency band covers a specific mechanical response frequency band, such as the 0-100Hz base vibration band and the 1-5kHz actuator resonance band. Each frequency band feature component undergoes Hilbert transform processing to extract phase response parameters, including the instantaneous amplitude envelope and phase shift angle. The above decomposition process uses Daubechies wavelet basis functions, and the decomposition level is dynamically configured based on the platform's dynamic characteristics.

[0076] For each frequency band characteristic component, its power spectral density is calculated, and a two-dimensional energy distribution map is generated. This energy distribution map is mapped to the three-dimensional workspace of the motion platform, with each spatial cell recording the energy amplitude. Spatial attention weighting is performed on the energy distribution map: a normalized exponential function is used to calculate the spatial cell weight factor, and the weight coefficient is positively correlated with the energy density within the cell. High-energy-concentration regions are assigned weight gain coefficients, while low-energy regions are attenuated. The weighted result is converted into a 512-dimensional motion feature vector. The vector dimension is decoupled from the platform's degrees of freedom, independently encoding dynamic features in different motion directions.

[0077] The motion feature vectors are imported into the feature filtering module and matched against a preset dynamic response threshold library. The threshold library stores baseline feature vectors under typical operating conditions. The matching process calculates cosine similarity, and feature vectors with a similarity lower than 0.85 are discarded. The filtered candidate feature set undergoes non-uniform sampling: based on the energy amplitude gradient distribution, sampling points are densified in regions where the gradient change rate exceeds 5%. Each sampling point generates a 32-bit position code and a 16-bit attitude code. The position code quantizes displacement deviation based on Cartesian coordinates, and the attitude code uses a quaternion rotation angle compression algorithm. The final output is a platform pose feature set, with a temporal tensor data structure containing a feature sequence of 2000 sampling points / second.

[0078] Based on the platform's pose feature set, an environmental disturbance perception model is invoked for analysis. The model first parses the position encoding sequence and maps it to the physical coordinate space of temperature, air pressure, and vibration sensors. An environmental parameter distribution matrix is ​​generated through cubic spline interpolation, with the matrix dimensions aligned with the motion platform's workspace. Each matrix unit integrates multi-sensor data; for example, the temperature gradient parameter is generated by fusing infrared thermal imager data and thermocouple data. Gradient field calculations are performed on the environmental parameter distribution matrix: a three-dimensional gradient detection is performed using the Sobel operator, outputting an environmental disturbance intensity distribution map. The map marks high-intensity disturbance regions (gradient value ≥ 15 units) and low-intensity regions (gradient value ≤ 3 units), and records the spatial orientation of the disturbance vector.

[0079] The spatiotemporal convolutional subnet built into the environmental disturbance perception model adopts a three-layer convolutional structure: the first layer uses a 3×3×3 convolutional kernel to extract local disturbance patterns; the second layer applies dilated convolution to expand the receptive field; and the third layer compresses features through max pooling. The subnet outputs a 256-channel environmental disturbance feature tensor, with tensor elements encoding disturbance amplitude, frequency, and spatial correlation. Simultaneously, the mechanical deformation prediction model is launched in parallel: this model loads the platform pose feature set and calls the finite element analysis engine to perform stress simulation. The finite element mesh generation accuracy reaches 0.1mm, and the boundary conditions are dynamically adjusted based on real-time attitude encoding. It outputs a mechanical deformation probability distribution map, with probability values ​​converted into a deformation risk index using the von Mises stress criterion.

[0080] The feature fusion stage performs tensor dot product operations: aligning the environmental perturbation feature tensor (256 dimensions) with the mechanical deformation probability distribution (64 dimensions) in the feature dimensions, and expanding the latter's dimension through transposed convolution. The fusion process uses Hadamard product to calculate cross-features, generating a 320-dimensional feature fusion result. This result preserves the spatiotemporal properties of the original tensors; for example, the coupling features between high-frequency vibration zones and local deformation regions are encoded in high-order tensor form. The final output multidimensional data array includes timestamps and can be directly input into downstream compensation models.

[0081] Example 2: See Figure 3 The dynamic compensation model's input interface receives feature fusion results from the upstream processing flow. This data structure is a 320-dimensional tensor sequence with a time resolution of 1 ms. During model initialization, a pre-defined trajectory constraint library is loaded. This library contains performance boundary parameters of the motion platform under various operating conditions, including maximum permissible acceleration, velocity smoothness coefficient, and position tracking error threshold. The constraint parser associates the feature fusion results with the current task number and retrieves the corresponding constraint parameter group using a hash index. Each parameter group contains 12 weight factors, corresponding to the compensation priority in the three dimensions of position, attitude, and acceleration. The assignment logic of the weight factors is based on the platform's dynamic characteristics; for example, increasing the acceleration compensation weight in high-speed motion segments and enhancing the position compensation weight in precise positioning segments.

[0082] The construction of the error propagation matrix adopts a hierarchical computational architecture. The first layer processes the position compensation amount, multiplying the position-related dimensions (first 128 dimensions) in the feature fusion result with the constraint weights to generate a position error vector. The second layer analyzes the attitude compensation requirements, extracting 64-dimensional attitude feature data from the feature tensor and converting it into Euler angle deviations using a quaternion interpolation algorithm. The third layer calculates the acceleration compensation components, and the remaining 128-dimensional feature data is reconstructed into a time-domain acceleration spectrum through inverse Fourier transform. The outputs of the three layers are combined into a 24×24 error propagation matrix through Kronecker product. The diagonal elements of the matrix represent the compensation amount for each degree of freedom, and the off-diagonal elements encode the coupling effect between degrees of freedom.

[0083] The traversal and sorting of the trajectory point sequence employs an improved A* search algorithm. The algorithm's heuristic function comprehensively considers the magnitude of compensation and the requirements for motion smoothness, generating the optimal access sequence within a search space of 500 trajectory points. Each trajectory point is associated with six key elements of the error propagation matrix (X / Y / Z positional deviation and Rx / Ry / Rz pose deviation), and the priority score of each point is updated in real time during the sorting process. A reinforcement learning strategy deploys a dual-deep Q-network architecture; the main network evaluates the Q-value of the current state-action pair, while the target network provides a stable learning objective. The action space is defined as seven basic compensation operations, including linear displacement compensation, angular displacement compensation, and a composite compensation mode. The reward function design focuses on three aspects: positional error convergence speed, energy consumption efficiency, and vibration suppression effect.

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

[0085] The instruction encoding and integration module standardizes the conversion of control signals. The platform's dynamic response parameter database stores the response delay, gain characteristics, and nonlinearity correction tables for each axis motor. The integration process first performs time alignment on the basic compensation instructions to compensate for differences in response speed between axes. Then, an inverse dynamics model is applied to convert the target compensation amount into a motor drive signal. The encoded output uses the industrial Ethernet protocol, with each data packet containing 32 bytes of control instructions and 8 bytes of checksum. A real-time guarantee mechanism employs hardware timestamp synchronization to ensure the control cycle is strictly controlled within 1ms.

[0086] The feedback data acquisition system consists of a multi-source sensor network. An optical encoder records the platform's actual position at a 50kHz sampling rate, while a vibration sensor array monitors mechanical resonances across six degrees of freedom. A data preprocessing unit performs sensor fusion and eliminates measurement noise using Kalman filtering. A feedback log generator converts the raw data into structured records, each containing a timestamp, trajectory offset spectrum, and vibration feature fingerprint. A dynamic error feature extractor employs sliding window analysis, with the window width adaptively adjusted based on motion speed, typically 20ms. The feature vector contains 15 key indicators, such as the root mean square value of the position error and the energy proportion of the dominant vibration frequency band.

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

[0088] The threshold update for the motion feature parsing network employs an incremental learning strategy. Dynamic response thresholds are stored in a distributed key-value database, and update requests are processed asynchronously via a message queue. The learning rate scheduler automatically adjusts the update magnitude based on error convergence, with an initial learning rate set to 0.001. The threshold validation module performs a rapid test after each update, using data from the most recent 100 control cycles for validation. An anomaly detector monitors the update process and automatically rolls back to the previous stable version when threshold drift exceeds a safe range.

[0089] Example 3: See Figure 4 The processing flow for dynamic error characteristics begins with time-series segmentation of the feedback log. The sensor system continuously records platform motion data at a sampling rate of 20kHz, generating raw logs containing trajectory offsets, vibration spectra, and environmental parameters. The time-series segmenter employs an adaptive window width algorithm, with a window length... Determined by the dynamic characteristics of the current motion state:

[0090]

[0091] in: : Preset maximum window width : Preset minimum window width The second norm of the velocity change within the window. Window width adjustment coefficient Small constants to prevent division by zero.

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

[0093] The error classification model employs a multi-scale convolutional neural network architecture. The input layer receives standardized error segments, with the data dimension being time step × number of sensor channels. The first convolutional layer uses a group of parallel convolutional kernels with variable widths; narrow kernels (length ≤ 5) capture high-frequency vibration features, while wide kernels (length ≥ 15) extract low-speed drift patterns. The feature fusion layer dynamically weights features at each scale through a gating mechanism, outputting a 128-dimensional hybrid feature vector. The classification head contains three fully connected branches, corresponding to three root cause labels: mechanical resonance, environmental interference, and control system hysteresis. Label probabilities are adjusted using a branch-specific temperature coefficient to avoid overconfidence in the classification results.

[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 parsing network employs an elastic weight fixation strategy. The adjustment range of network parameters is determined by two factors: the importance weights of the current parameters and the gradient direction of the new data. The importance weights are approximated using the diagonal elements of the Fisher information matrix, recording the sensitivity of the parameters in historical tasks. Gradient calculation uses comparative data before and after perturbation, and the loss function includes two components: trajectory position error and feature reconstruction error. Elastic constraints are applied during the parameter update process to prevent drastic changes in highly important parameters. The proportion of parameters updated in each step is controlled within 15%, prioritizing adjustments to the bias terms of the convolutional kernels and the parameters of the normalization layers.

[0100] Stability monitoring of the control system is conducted throughout the entire implementation process. Monitors track moving averages of key performance indicators, including command latency, computational load, and communication error rate. Anomaly detection algorithms, based on the isolated forest model, identify abnormal deviations in indicator data. When persistent performance degradation is detected, the system automatically reverts to the parameter configuration of the previous stable version and generates a diagnostic report for subsequent analysis. The status information of all subsystems is synchronized via a high-precision clock with timestamp accuracy down to the 100 nanosecond level, ensuring the timing accuracy of fault analysis.

[0101] Cross-protocol data conversion is performed at the very 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 employs an intermediate presentation layer architecture: first, platform-driven compensation commands are parsed into device-independent sequences of atomic operations, and then re-encoded according to the target device's description file.

[0102] Example 4: The parameter update process of the motion feature parsing network adopts a backpropagation mechanism based on trajectory difference data. This process begins with spatiotemporal alignment. The system collects two sets of trajectory data before and after the disturbance. Each set contains the platform's position coordinates along the X / Y / Z axes and attitude data for the three rotational degrees of freedom (Rx / Ry / Rz). The alignment algorithm first establishes a time axis mapping relationship and uses dynamic time warping to unify the time reference of the two sets of data, eliminating deviations caused by asynchronous sampling times. The aligned data forms a matching sequence, with each sequence point containing the corresponding state values ​​before and after the disturbance.

[0103] The trajectory offset calculation module processes the aligned matching sequence. For each trajectory point, the system calculates a position deviation vector and an attitude deviation vector. The position deviation vector records the displacement difference along three axes, and the attitude deviation vector uses quaternion interpolation to calculate the rotational difference. These deviation values ​​are converted into a unified error metric to form training samples for network updates. Partial trajectory deviation data collected during a particular update is shown in Table 1.

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

[0105] Timestamp (ms) X-bias (nm) Y-bias (nm) Z-bias (nm) Rx deviation (μrad) Ry bias (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 motion regression loss function is constructed considering the combined effects of multi-dimensional errors. Position deviation is measured using a weighted Euclidean distance metric, with the weight coefficients for the three axes dynamically configured based on the platform's structural characteristics. Attitude deviation is calculated using a quaternion logarithmic mapping method, converting rotational differences into equivalent axis angle representations. The loss function integrates deviation data from all trajectory points to generate a scalar value reflecting the overall control error. This value serves as the objective function for network parameter optimization, driving the subsequent backpropagation process.

[0107] The loss tensor is generated using automatic differentiation. The system constructs a computation graph to record the complete computation process from input data to the loss value. Nodes in the computation graph include various tensor operations, such as convolution, pooling, and matrix multiplication. During backpropagation, the system calculates the gradient value of each parameter backward along the computation graph. The gradient calculation process uses the chain rule, taking derivatives layer by layer starting from the loss function, ultimately obtaining the partial derivatives of each trainable parameter.

[0108] The calculation of the convolutional kernel gradient matrix features selective updating. The system maintains a parameter importance score table, recording the sensitivity index of each convolutional kernel in historical tasks. When calculating the gradient matrix, the gradients corresponding to important parameters are amplified, while the gradients of less important parameters are correspondingly decayed. This selective updating mechanism ensures that the network does not destroy the important feature representations it has already learned when adapting to new data. In each iteration, the system only adjusts about 15% of the network parameters, prioritizing the convolutional kernels most correlated with the current error pattern.

[0109] The parameter updates employ a phased, gradual strategy. Initially, a large learning rate is used to quickly approach the optimal solution, and this rate is gradually decreased with each iteration. The update process includes three validation steps: first, the reasonableness of the gradient values ​​is checked to eliminate outliers; second, the changes in the network output after the parameter updates are evaluated to ensure that the error has indeed decreased; and finally, the network's performance on the validation set is tested to prevent overfitting. Only updates that pass all validation steps are ultimately applied.

[0110] The performance evaluation of the updated network employs cross-validation. A portion of the trajectory data is retained as a test set and not used in the training process. After the update, the network's generalization ability is evaluated using the test set data. Evaluation metrics include position control accuracy, attitude stability, and response speed. If the test results do not meet the requirements, the system automatically rolls back to the previous stable version and retrains with adjusted update strategies.

[0111] The entire parameter update process has a real-time guarantee mechanism. The system adopts a pipelined processing architecture, parallelizing data acquisition, preprocessing, training, and verification. Computational resource allocation uses a dynamic scheduling algorithm to ensure that network updates do not affect the execution of normal control tasks. The end-to-end latency for each update is controlled within 50ms, meeting the real-time requirements of high-precision motion control.

[0112] Updated network parameters are securely stored in a multi-version management system. Each version comes with complete metadata, including training time, data source, and performance metrics. The system maintains a parameter version repository, supporting fast rollback and version comparison. When a performance degradation is detected, appropriate parameter configurations can be selected from the repository for recovery.

[0113] The persistent storage of network parameters employs differential encoding technology. Each update only stores the parameter differences relative to the previous version, significantly reducing storage space usage. An instant recovery mechanism is used during parameter loading, enabling network state switching to be completed within milliseconds. This design ensures both parameter security and meets the stringent real-time requirements of the control system.

[0114] The system also implements a parameter health monitoring mechanism. It periodically scans the range of network parameter values ​​to detect potential overflow or vanishing gradient issues. When abnormal parameters are detected, a repair process is automatically triggered, including parameter reinitialization and gradient pruning. The monitoring results generate detailed reports to help analysts understand the evolution trends of the network parameters.

[0115] Example 5: See Figure 5 The cross-platform control command adapter is built on a protocol abstraction layer architecture. This architecture includes a device description file library, a command conversion engine, and a protocol verification module. Device description files are stored in XML format, defining the data encoding rules, control command syntax, and timing constraints of the target driver device. Each description file contains an instruction set mapping table, which clearly defines the correspondence between the source instruction opcode and the target device opcode. The parameter passing path details the assembly order and byte order rules of multi-byte parameters. During engine initialization, all device description files are loaded, and an in-memory index table is established to accelerate the retrieval process.

[0116] After the platform-driven compensation command is input into the conversion engine, it first performs command structure parsing. The parser identifies the atomic operation type of the command, such as position setting, speed curve loading, or emergency braking command. Each atomic operation is broken down into an opcode field and parameter data segment, generating intermediate control presentation layer data. The presentation layer data is organized in a tree structure, with the root node identifying the operation category and child nodes storing parameter key-value pairs. The engine traverses and queries the control command mapping table in memory, using opcode hash matching to locate the equivalent command for the target device. The mapping process considers differences in device capabilities; when the target device does not support a specific operation, the engine automatically breaks it down into a combination of multiple basic commands.

[0117] Semantic adaptation of parameter passing paths resolves protocol conflicts. A conflict detector analyzes the bit distribution differences between source and target parameters, identifying three conflict types: precision mismatch (e.g., 32-bit floating-point to 16-bit fixed-point conversion), unit inconsistency (nanometer to micrometer conversion), or missing parameters. The adapter employs different handling strategies for each type: precision conversion uses linear scaling with rounding compensation; unit conversion is adjusted using a scaling factor; and missing parameters are filled by a default value generator based on context inference. A double-buffering mechanism is used to maintain timing constraints: the instruction sequence is temporarily stored in a timestamped circular queue, and a wait period is inserted before output based on the device response delay to maintain the original time interval between adjacent instructions.

[0118] Preserving the dynamic compensation context ensures the semantic integrity of the instructions. The conversion engine encapsulates a dedicated data structure to record the metadata of the current compensation operation, including the reason for the compensation trigger, the effect of previous compensation, and the expected control target. This data structure is embedded as a hidden field in the output instructions, allowing the device-side driver to optimize its local control strategy. The injection of precision optimization parameters is implemented in the final encoding stage: the resolution coefficients defined in the device description file are applied to the parameter encoding process, for example, requantizing a 1.25-nanometer position quantity at a 2-nanometer resolution, with the quantization error accumulating in subsequent instruction compensation.

[0119] The control command stream generation employs a framed transmission strategy. A single frame structure includes a frame header, command body, and checksum: the frame header identifies the data stream type and frame sequence number; the command body encapsulates opcodes and parameters according to device requirements; and the checksum uses the CRC-32 algorithm to ensure transmission integrity. The flow control module monitors device status feedback and automatically adjusts the frame transmission rate when a buffer overflow risk is detected. The device registration module implements plug-and-play functionality: when a new device connects, it reports descriptive information, and the engine dynamically updates the memory index table and mapping rules, completing protocol adaptation without requiring a system restart. The entire conversion process meets strict time constraints: the batch conversion latency for 1000 commands does not exceed 5 milliseconds.

[0120] A security verification mechanism is implemented throughout the entire conversion process. The syntax validator checks the compliance of output instructions, excluding out-of-bounds parameters and illegal opcodes; the semantic analyzer verifies the logical consistency of the instruction sequence to prevent conflicting operations; and the timing checker calibrates instruction interval deviations using a hardware clock. Instructions that fail verification trigger a rollback mechanism: the original compensation instruction is saved to the safe zone, and a diagnostic report is generated to locate protocol adaptation defects. This implementation supports seamless integration with ten types of industrial drive equipment, including heterogeneous systems such as linear motor platforms, piezoelectric ceramic actuators, and magnetic levitation platforms.

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive control method for a micro / nano-scale 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. 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; 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.

2. The adaptive control method for a micro / nano-level high-precision motion platform as described in claim 1, characterized in that, 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.

3. The adaptive control method for a micro / nano-level high-precision motion platform as described in claim 2, characterized in that, The method further includes: During motion control, the platform's execution status data is collected in real time to generate feedback logs containing trajectory offset and vibration spectrum characteristics. Extract the dynamic error features from the feedback log, and perform pattern matching between the dynamic error features and the historical control case library to generate adaptive compensation parameters; The dynamic response threshold of the motion feature parsing network is updated online based on the adaptive compensation parameters. The updated dynamic response threshold is injected into the dynamic compensation model, and the compensation priority weights in the error propagation matrix are recalculated.

4. The adaptive control method for a micro / nano-level high-precision motion platform as described in claim 3, characterized in that, The step of extracting dynamic error features from the feedback log and performing pattern matching between the dynamic error features and the historical control case library to generate adaptive compensation parameters includes: The dynamic error features are subjected to time series segmentation processing to generate multiple error segments and their corresponding platform motion state sequences; The trained error classification model is invoked to perform root cause analysis on each error segment, generating classification labels that include mechanical resonance, environmental interference, and control system hysteresis. Retrieve compensation strategy templates that match the classification labels from the historical control case library to generate a set of candidate compensation parameters; Based on the matching degree ranking between the platform motion state sequence and the candidate compensation parameter set, the parameter with the highest confidence is selected to generate the adaptive compensation parameter.

5. The adaptive control method for a micro / nano-level high-precision motion platform as described in claim 1, characterized in that, The method further includes: Virtual trajectory disturbance parameters are injected before the control signal output, and the virtual trajectory disturbance parameters are used to simulate the sudden movement deviation scenario of the platform; Monitor the results of the dynamic compensation model's correction of the trajectory after disturbance, and generate control robustness evaluation index; When the control robustness evaluation index is lower than a preset threshold, the parameter learning mode of the motion feature parsing network is triggered. The convolution kernel parameters of the motion feature parsing network are updated by gradient backpropagation based on the trajectory difference data before and after the perturbation.

6. The adaptive control method for a micro / nano-level high-precision motion platform as described in claim 5, characterized in that, The step of updating the convolution kernel parameters of the motion feature parsing network by gradient backpropagation based on trajectory difference data before and after the perturbation includes: The trajectory difference data before and after the disturbance are spatiotemporally aligned to generate a matching sequence containing the set of trajectory points before the disturbance and the set of trajectory points after the disturbance. Extract the trajectory offset from the matching sequence to generate the position deviation vector and attitude deviation vector for each trajectory point; A motion regression loss function is constructed based on the position deviation vector and the attitude deviation vector; The motion regression loss function is subjected to a differentiable transformation to generate the loss value tensor required for gradient backpropagation; Iterate through the convolutional kernel parameters of the motion feature parsing network, calculate the partial derivative of the loss tensor with respect to each convolutional kernel, and generate the convolutional kernel gradient matrix; Based on the convolution kernel gradient matrix and the preset learning rate parameters, the convolution kernel parameters of the motion feature parsing network are iteratively and incrementally adjusted.

7. The adaptive control method for a micro / nano-level high-precision motion platform as described in claim 1, characterized in that, The method further includes: A cross-platform control instruction adapter is constructed, and the differences in instruction protocols between different driving devices are resolved through the cross-platform control instruction adapter; The platform drive compensation command is converted into the underlying control signal supported by the target drive device; The timing constraints and dynamic compensation context of the actual control signals are preserved during the conversion process; The precision optimization parameters matched by the target driving device are injected to generate a motion control command stream that meets the cross-platform execution conditions.

8. The adaptive control method for a micro / nano-level high-precision motion platform as described in claim 7, characterized in that, The construction of a cross-platform control instruction adapter, and the parsing of instruction protocol differences between different driver devices through the cross-platform control instruction adapter, includes: Establish a driver device protocol rule base and store the control command mapping table and parameter transmission path for each platform; The platform-driven compensation instructions are parsed to generate intermediate control representation layer data; Based on the intermediate control representation layer data, the control instruction mapping table is traversed and queried to generate a device-compatible instruction conversion scheme; Semantic adaptation is performed on conflicting parameter passing paths to generate unambiguous control command conversion results.

Images