Rapid and accurate prediction positioning method for linear motion module of industrial robot

Abstract

The invention relates to the crossing field of mechanical engineering and automatic control, discloses a rapid and accurate prediction positioning method and system for a linear motion module of an industrial robot, and aims to solve the problem of dynamic positioning deviation caused by factors such as elastic deformation, a mechanical gap and a heat effect under high-speed motion. The method comprises the following steps: collecting vibration, strain, temperature and motor load data in real time through a multi-mode sensor array; constructing a space-time dynamic state encoder based on module physical topology, and fusing time sequence convolution and a graph neural network to generate system state representation; predicting a future positioning error of the end effector by using an attention mechanism in combination with a future motion instruction; and a feedforward compensation signal is generated through an inverse dynamic model, and is superposed with an original instruction to form a composite driving instruction. According to the method, control normal form transition from passive feedback to active prediction is realized, and the positioning precision and the movement efficiency under the high-speed working condition are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of mechanical engineering and automatic control, and specifically relates to a rapid and accurate predictive positioning method for linear motion modules of industrial robots. Background Technology

[0002] Industrial robots play a central role in modern intelligent manufacturing systems, and their motion control precision and response speed directly determine production efficiency and product quality. Linear motion modules, as the basic execution unit for high-precision positioning in industrial robots, are widely used in key processes such as assembly, handling, welding, and inspection. These modules typically consist of servo motors, ball screws, guide rails, and position feedback devices, relying on a closed-loop control system to track the target position. However, under high-speed operation or frequent start-stop conditions, the system is inevitably affected by a combination of factors such as mechanical inertia, transmission backlash, frictional nonlinearity, and external disturbances, causing the actual trajectory to deviate from the ideal path, resulting in overshoot, oscillation, or steady-state errors, making it difficult to simultaneously achieve both speed and positioning accuracy.

[0003] In existing technologies, some solutions attempt to introduce Kalman filtering or state observers for position estimation to overcome the aforementioned technical defects. However, their model parameters are mostly based on static calibration, making it difficult to adapt to load changes, stiffness decay caused by temperature rise, or transmission characteristic drift caused by long-term wear.

[0004] Other existing technologies use neural networks for end-to-end learning. Although they have a certain nonlinear fitting ability, they suffer from inaccurate predictions in unseen working conditions due to insufficient training data coverage and weak generalization ability. Furthermore, they lack interpretability and real-time performance guarantees.

[0005] Furthermore, most prediction strategies do not fully integrate multi-source sensing information such as current, vibration, and high-frequency encoder sampling, nor do they consider the computational resource limitations under control cycle constraints, making them difficult to deploy in practical embedded controllers.

[0006] Therefore, under the premise of ensuring the lightweight and real-time performance of the algorithm, how to construct a predictive positioning mechanism that can adapt to environmental disturbances, accurately predict future positions, and support high-frequency updates has become an urgent technical problem that restricts the performance leap of linear motion modules of industrial robots. Summary of the Invention

[0007] The technical problem this invention aims to solve is that, when existing industrial robot linear motion modules perform high-speed motion tasks, the coupling effect of multiple complex physical factors such as elastic deformation of the drive chain, mechanical backlash, frictional nonlinearity, and thermal effects leads to significant dynamic deviations and vibration hysteresis between the theoretical position commands output by the servo controller and the actual physical position of the end effector of the motion module. This deviation is particularly severe during high-speed start-stop and direction-changing processes, thus failing to simultaneously meet the dual requirements of high speed and high precision. Traditional closed-loop control methods based on position feedback are essentially a form of hysteresis compensation, which can only correct errors after they occur and cannot anticipate or actively suppress their generation.

[0008] To address the aforementioned technical problems, this invention provides a rapid and accurate predictive positioning method for linear motion modules of industrial robots. The core idea of ​​this invention is to move away from relying on a single position loop hysteresis feedback, and instead deploy a multimodal heterogeneous sensor array on key physical nodes of the linear motion module to capture high-dimensional physical data streams in real time that characterize the complete dynamic characteristics of the system, including multi-dimensional information such as vibration, strain, temperature, and motor load. Next, a deep fusion model of graph neural networks and temporal convolutional networks is constructed to accurately map the physical topology of the module and the dynamic mechanical transmission relationship. This model encodes the discrete, multi-source sensor data streams in real time into a unified low-dimensional spatiotemporal state feature vector that characterizes the current and historical dynamic behavior of the entire system. Then, this spatiotemporal state feature vector is combined with... The sequence of future motion commands issued by the layer motion planner is input into a sequence prediction decoder based on an attention mechanism to predict with high accuracy the specific positioning error trajectory that the module's end effector will experience within a very short time window in the future. Finally, based on the predicted error trajectory, a feedforward compensation control signal sequence is generated in real time through an inverse dynamic compensation model. This sequence is vector-superimposed with the original motion command sequence to form the final composite drive command that has been pre-distorted. This drives the motor to move in a way that can compensate for the expected physical error in advance, ultimately achieving nanosecond-level accurate prediction and micrometer-level active positioning control of the end effector position under high-speed motion.

[0009] A method for rapid and accurate predictive positioning of a linear motion module in an industrial robot, comprising the following steps: Real-time acquisition of the operating data of the linear motion module during the motion process. The operating data includes at least: triaxial vibration acceleration signals collected by multiple acceleration sensors arranged along the transmission chain of the linear motion module; mechanical stress strain signals collected by strain sensors arranged at the ball screw nut seat or linear guide slider of the linear motion module; temperature distribution data collected by temperature sensors arranged on the ball screw, guide rail and drive motor housing; and real-time torque output current signal and rotor position encoder signal of the motor obtained from the drive motor servo driver. A physical structure spatial topology diagram of the linear motion module is constructed, and the real-time acquired running data is mapped to the corresponding nodes of the spatial topology diagram. The mapped data is processed by a spatiotemporal dynamic state encoder to generate a running state representation vector that represents the overall real-time dynamic characteristics of the linear motion module. The sequence of future motion commands to be executed is obtained from the main motion controller of the industrial robot. The sequence of future motion commands includes the target position, target velocity and target acceleration commands for a series of future time steps. The running state representation vector and the future motion command sequence are simultaneously fed into a future positioning error prediction module. Based on the current physical state information of the system contained in the input running state representation vector and the future input excitation defined by the future motion command sequence, the future positioning error prediction module calculates and outputs a future positioning error sequence synchronized with the time axis of the future motion command sequence. The future positioning error sequence is input into a feedforward compensation control generation module. The feedforward compensation control generation module calculates and generates a feedforward compensation control sequence to offset the expected error based on the future positioning error sequence. The feedforward compensation control sequence and the original future motion command sequence are algebraically summed at the control clock cycle level and fused into a final drive command sequence. The final drive command sequence is then sent to the servo driver of the linear motion module to control the drive motor to perform the compensated and refined motion.

[0010] Furthermore, in the step of acquiring operational data, the accelerometer is a microelectromechanical system triaxial accelerometer with a sampling frequency of not less than 10 kHz, and its placement includes the drive motor mounting flange, the center of the motion slide, and both ends of the linear motion module base; the strain sensor is a metal foil strain gauge, which is attached to the outer wall of the ball screw nut seat in the form of a full-bridge circuit, and is used to measure the axial and torsional combined stress during the transmission process; the temperature sensor is a platinum resistance temperature sensor, which is closely attached to the middle of the ball screw surface, the bearing seats at both ends of the screw, and near the stator winding of the drive motor; the real-time torque output current signal of the motor and the rotor position encoder signal are read from the servo driver at a frequency of not less than 4 kHz via a high-speed industrial Ethernet bus.

[0011] Furthermore, the steps for generating the runtime state representation vector specifically include: First, the raw data streams acquired from each sensor are timestamped and resampled to form a multi-channel time series dataset with a unified time reference. Secondly, a multi-layer one-dimensional temporal convolutional network is used to process the temporal data of each channel. The temporal convolutional network adopts causal convolution and dilated convolution structure to efficiently extract the dynamic fluctuation characteristics and long-term dependencies of each physical quantity at different time scales without the leakage of future information, and generate the preliminary temporal features of each measurement point. Furthermore, based on the mechanical structure drawings of the linear motion module, an undirected graph is constructed with each sensor deployment point as a node and the reciprocal of the physical connection stiffness or the shortest physical distance between nodes as the edge weight, serving as a spatial topology graph. Finally, the preliminary temporal features are used as the initial attribute features of each node. An iterative graph convolution operation is performed on the spatial topology graph through a graph neural network. In each iteration, the feature vector of each node will aggregate the feature information of its neighboring nodes to simulate the transmission and coupling effects of vibration waves, heat and stress in the mechanical structure. After multi-layer graph convolution operation, the feature vectors of all nodes are globally pooled to generate the final, unique, fixed-dimensional running state representation vector.

[0012] Furthermore, the future positioning error prediction module is a sequence-to-sequence prediction model based on an attention mechanism. Its specific structure includes an encoder and a decoder. The encoder receives the historical sequence of the running state representation vector and captures the inherent laws of the system's dynamic state evolution through a self-attention mechanism. The decoder combines the encoder's output, the future motion command sequence, and its own predicted output from the previous time step. Through the synergistic effect of cross-attention and self-attention mechanisms, it generates predicted values ​​for the future positioning error step by step until the entire future positioning error sequence is output. The positioning error is defined as the difference between the actual position and the commanded position of the moving slide. The actual position of the moving slide is measured in real time by an independent high-precision linear grating ruler and serves as the ground truth label during the prediction model training phase.

[0013] Furthermore, the feedforward compensation control generation module is specifically an inverse dynamics neural network model. During offline training, this model learns from a large amount of actual motion data to establish a nonlinear mapping relationship between a desired compensatory acceleration (i.e., acceleration opposite to the prediction error) and the additional motor drive torque required to generate that acceleration. During real-time operation, after receiving the future positioning error sequence, the module first performs a quadratic difference to obtain the desired compensatory acceleration sequence. This compensatory acceleration sequence is then input into the inverse dynamics neural network model to directly calculate the required additional motor drive torque sequence. This torque sequence is then converted into a feedforward compensation position or speed command sequence executable by the servo controller.

[0014] In summary, this application includes at least one of the following beneficial technical effects: I. Achieving a paradigm shift from passive compensation to active prediction in control. This invention constructs a real-time representation of the system's deep physical state and combines it with future motion commands for forward-looking prediction. It can calculate the shape of dynamic errors and generate compensation commands before they actually occur, fundamentally solving the inherent delay problem of traditional feedback control and greatly improving trajectory tracking accuracy under high-speed motion.

[0015] Second, a more comprehensive and profound dynamic model of the linear motion module has been constructed. This invention employs multimodal sensor fusion and spatiotemporal graph neural network technology, which can capture panoramic dynamic characteristics including mechanical vibration, structural stress transmission, thermal expansion and contraction effects, and complex nonlinear coupling relationships between various physical quantities. Its model fidelity far exceeds that of traditional control theory models that simplify the system to a linear time-invariant system.

[0016] Third, it effectively solves the core contradiction between speed and accuracy in industrial applications. By accurately feeding forward to compensate for expected errors, this invention allows linear motion modules to operate with higher acceleration and speed, while eliminating the need for additional time spent on vibration attenuation and tuning near the positioning point. This significantly improves overall motion cycle time and production efficiency without sacrificing positioning accuracy.

[0017] Fourth, it possesses high adaptability and stability. Data-driven neural network models can automatically adapt to system characteristic drift caused by mechanical wear, load changes, or ambient temperature fluctuations through online learning or periodic retraining, maintaining prediction accuracy at all times. Compared to physical models that rely on fixed parameters, it has stronger environmental adaptability and long-term operational stability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention. Detailed Implementation

[0019] This invention provides a rapid and accurate predictive positioning method for linear motion modules of industrial robots. Its purpose is to solve the problem in the prior art that the theoretical position command output by the servo controller and the actual physical position of the end effector are significantly deviated and vibratory hysteresis due to the coupling effect of multiple physical factors such as elastic deformation of the drive chain, mechanical clearance, frictional nonlinearity and thermal effect, during high-speed motion.

[0020] This invention provides a rapid and accurate predictive positioning method for linear motion modules of industrial robots. Its purpose is to solve the problem in the prior art that the theoretical position command output by the servo controller and the actual physical position of the end effector are significantly deviated and vibratory hysteresis due to the coupling effect of multiple physical factors such as elastic deformation of the drive chain, mechanical clearance, frictional nonlinearity and thermal effect, during high-speed motion.

[0021] Existing traditional closed-loop control methods can only perform lag corrections for errors that have already occurred, and cannot achieve active error suppression. This embodiment constructs a multimodal heterogeneous sensor array, a spatiotemporal graph neural network encoding model, an attention mechanism sequence prediction module, and an inverse dynamics feedforward compensation mechanism to form a complete predictive-compensation control closed loop, thereby achieving rapid response capability under high-speed motion conditions.

[0022] The rapid and accurate prediction and positioning method of this application includes the following steps: S1: Real-time acquisition of the linear motion module's operating data during the motion process; S2: Construct a physical structure space topology diagram of the linear motion module, map the running data to the corresponding nodes of the diagram, and generate a running state representation vector; S3: Obtain the future motion command sequence from the main motion controller; S4: Input the running state representation vector and the future motion command sequence into the future positioning error prediction module, and output the future positioning error sequence; S5: Input the error sequence into the feedforward compensation control generation module to calculate and generate the feedforward compensation control sequence; algebraically sum the feedforward compensation control sequence with the original future motion command sequence to form the final drive command sequence and send it to the servo driver.

[0023] The steps described above are explained below: S1: Real-time acquisition of operational data of the linear motion module during its motion. The collected data includes at least four types of physical quantities: the first type is triaxial vibration acceleration signals collected by multiple acceleration sensors deployed along the transmission chain; the second type is mechanical stress and strain signals collected by strain sensors deployed at the ball screw nut seat or linear guide slider; the third type is temperature distribution data collected by temperature sensors deployed on the ball screw, guide rail, and drive motor housing; and the fourth type is the real-time torque output current signal and rotor position encoder signal obtained from the drive motor servo driver.

[0024] Specifically, the following steps are included: S1.1: Acquire vibration acceleration signals. A microelectromechanical system (MEMS) triaxial accelerometer is used as the acceleration sensor, with a sampling frequency of not less than 10 kHz. The sensor is deployed at locations including the drive motor mounting flange, the center of the motion slide, and both ends of the linear motion module base, to comprehensively cover the vibration mode distribution of the module under high-frequency excitation.

[0025] S1.2: Acquisition of mechanical stress and strain signals. A metal foil strain gauge is used as the strain sensor, attached to the circumferential direction of the outer wall of the ball screw nut seat in a full-bridge circuit configuration. This is used to simultaneously measure axial tensile and compressive stresses and torsional shear stresses during transmission. Its signal conditioning circuit has a temperature self-compensation function to eliminate the influence of environmental temperature drift on the measurement results.

[0026] S1.3: Acquire temperature distribution data. A platinum resistance temperature sensor is used, with a temperature measurement accuracy better than 0.5 degrees Celsius. The sensor is closely attached to the center of the ball screw surface, the bearing housings at both ends of the screw, and near the stator windings of the drive motor to capture the heat source distribution and heat conduction path.

[0027] S1.4: Acquire motor operating data. The real-time torque output current signal and rotor position encoder signal of the motor are read from the servo driver at a frequency of no less than 4kHz via a high-speed industrial Ethernet bus, ensuring time synchronization with external sensor data.

[0028] S2: Construct the physical structure space topology diagram of the linear motion module, map the running data to the corresponding nodes of the diagram, and generate the running state representation vector. This includes the following steps: S2.1: Multi-sensor data synchronization and alignment. All raw sensor data streams from S1 undergo timestamp alignment and resampling. Due to differences in sampling frequencies among different sensors, the system employs a hardware clock synchronization-based triggering mechanism. At the beginning of each servo control cycle, all sensors are uniformly triggered to sample. Interpolation algorithms are used to downsample high-frequency data (such as acceleration signals) to a unified reference frequency, or upsample low-frequency data (such as temperature signals) to that frequency, forming a multi-channel time-series dataset with a unified time reference.

[0029] Each time step of a multi-channel time-series dataset contains a feature vector whose dimension is equal to the total number of sensors multiplied by the feature vector of its output dimension.

[0030] S2.2: Temporal Feature Extraction. The temporal data for each channel obtained in S2.1 is processed using a multi-layer one-dimensional temporal convolutional network. This network employs a causal convolution and dilated convolution structure, consisting of several alternating stacked causal and dilated convolutional layers. Each convolutional kernel has a size of 3, and the dilation factor increases in powers of 2. This allows for efficient extraction of the dynamic fluctuation characteristics and long-term dependencies of various physical quantities at different time scales without leakage of future information, and exponentially expands the receptive field.

[0031] Causal convolution ensures that computation at any given time step does not depend on future inputs, maintaining the causality of the prediction system; dilated convolution effectively captures long-range temporal dependencies, such as slow positional drift caused by thermal expansion or the exponential envelope characteristics in vibration decay processes. After processing by this network, each sensor measurement point outputs a preliminary temporal feature vector of fixed length, the dimension of which is determined by the number of channels in the last layer of the network.

[0032] S2.3: Construct a spatial topology graph. Based on the existing mechanical structure drawings of the linear motion module, construct a physical structure spatial topology graph. This graph is an undirected graph, and its node set corresponds to all sensor deployment points.

[0033] The edge set is determined by the physical connections between nodes: if two nodes are directly connected in a mechanical structure through a rigid component, an edge is created in the graph; the weight of the edge is defined as the reciprocal of the shortest physical distance between the two nodes, or determined based on the elements of the local stiffness matrix obtained from finite element analysis. This graph structure is permanently stored in the system memory as the static adjacency matrix of the physical structure spatial topological graph neural network.

[0034] S2.4: Graph Convolution Fusion and State Vector Generation. The preliminary temporal features obtained in step S2.2 are used as the initial attribute features of each node in the spatial topological relationship graph in S2.3, and input into a graph neural network for iterative graph convolution operations.

[0035] Graph neural networks contain several graph convolutional layers, each performing the following operations to update node features: This graph neural network contains several graph convolutional layers, each performing the following operation to update node features: For any node v in the graph, its position at the 1st... New feature vectors of the layer Generate in the following way: Neighborhood aggregation: Calculate its own position in the [missing information] direction. eigenvectors of the layer Features of all its neighboring nodes The weighted average, where the weights are determined by the edge weights in the adjacency matrix of the spatial topological graph, forms the aggregated feature vector.

[0036] Linear Transformation and Projection: The aggregated feature vector is subjected to a learnable linear transformation matrix. Multiply (i.e., perform a linear projection) and add a learnable bias vector. .

[0037] Nonlinear activation: The result of the above linear projection is transformed by a nonlinear activation function ρ (such as the ReLU function).

[0038] The mathematical expression of this process can be simplified as follows: ; This process simulates the propagation and coupling effects of vibrational energy, heat flow, and stress waves in a mechanical structure. The learnable linear transformation matrix... With bias vector , which are the model parameters obtained by optimizing the graph neural network through the backpropagation algorithm during the offline training phase in a graph neural network such as S6. The optimization objective is to minimize the prediction loss of future positioning errors.

[0039] After multi-layer graph convolution, the system performs global max pooling on the final feature vectors of all nodes to generate a fixed-dimensional running state representation vector, which fully encodes the current overall dynamic state of the linear motion module.

[0040] S3: Obtain the future motion command sequence from the main motion controller, specifically including the following steps: S3.1: Establish a motion command interface. The system receives motion trajectory planning data to be executed in real time from the main motion controller of the industrial robot through the motion command interface module.

[0041] S3.2: Parse the instruction data format. The motion trajectory planning data is given in discrete time series form. The data packet for each time step contains three components: target position, target velocity, and target acceleration.

[0042] S3.3: Synchronization control cycle. The time step of the future motion command sequence is consistent with the servo control cycle, with a typical value of 250 microseconds.

[0043] S3.4: Set the prediction window. The sequence covers several future servo control cycles. The length of the prediction window can be dynamically adjusted according to the system's computing power. A typical value is 20 control cycles, or 5 milliseconds.

[0044] S4: Input the running state representation vector from S2 and the future motion command sequence from S3 into the future positioning error prediction module, and output the future positioning error sequence. Specifically, this includes the following steps: S4.1: Input the historical sequence of the aforementioned running state representation vector and the future motion command sequence into a sequence-to-sequence prediction model based on an attention mechanism.

[0045] S4.2: The encoder of the sequence-to-sequence prediction model receives the sequence of operating state representation vectors over several past control cycles, calculates the correlation between states at each time step through a multi-head self-attention mechanism, and generates a context-aware state evolution representation.

[0046] It should be noted that multi-head self-attention is a standard and mature component in deep learning, especially in the Transformer architecture. Its successful applications in natural language processing, time series prediction, and other fields have been ongoing for several years. Its mathematical principles and computational processes are generally known and understood by those skilled in the field (i.e., engineers engaged in the development of intelligent control and machine learning algorithms). Therefore, further details will not be elaborated upon here.

[0047] S4.3: The decoder of the sequence-to-sequence prediction model operates in an autoregressive manner, generating a sequence of future positioning errors step-by-step. When predicting the positioning error at the t-th future time step, the decoder's input includes: All outputs of the encoder; The instructions for the first t time steps in the future motion instruction sequence; The error value predicted by the decoder itself in the first t-1 time steps.

[0048] For the prediction at the initial time step (t=1), the "predicted error value" in the input is replaced by a specific initial token (such as a zero vector).

[0049] The decoder contains both a self-attention layer and a cross-attention layer. The former is used to model the temporal dependencies of the error sequence itself, while the latter is used to establish a dynamic mapping relationship between the error and the system state and motion commands.

[0050] S4.4: Output Future Positioning Error Sequence. The decoder ultimately outputs a future positioning error sequence that is strictly synchronized with the timeline of the future motion command sequence. The output at each time step is a scalar representing the difference between the actual position and the commanded position of the moving slide at that moment.

[0051] S4.5: The true value of the positioning error is provided by an independently deployed high-precision linear grating ruler with a resolution better than 0.1 micrometers. This grating ruler is mounted between the base and the slide of the linear motion module, directly measuring the absolute displacement of the slide relative to the base. The signal from the grating ruler does not participate in the real-time control loop; it is used only as a supervision label during the offline model training phase.

[0052] S5: Input the error sequence into the feedforward compensation control generation module to calculate and generate the feedforward compensation control sequence; algebraically sum the feedforward compensation control sequence with the original future motion command sequence to form the final drive command sequence and send it to the servo driver. This includes the following steps: S5.0: Inverse dynamics model training.

[0053] During the offline training phase, the inverse dynamics neural network model learns the nonlinear mapping relationship from the desired compensation acceleration to the required additional motor drive torque through a large amount of experimental data. The network structure is a multi-layer fully connected network, with both the input and output layers having dimensions equal to the prediction window length, and the hidden layers employing a modified linear unit activation function.

[0054] During training, the loss function is defined as the mean square error between the actual compensated acceleration and the expected input acceleration after the actual applied torque of the model output.

[0055] S5.1: Calculate the desired compensation acceleration. The future positioning error sequence e(t) predicted in S4 is input to the feedforward compensation control generation module. This module first performs a quadratic difference operation on e(t) to obtain the desired compensation acceleration sequence a_c(t), with the relationship a_c(t) = -d²e(t) / dt².

[0056] S5.2: Inverse dynamics mapping generates compensating torque. The compensating acceleration sequence a_c(t) calculated in S5.1 is input into the inverse dynamics neural network model at the core of the module. This model directly outputs the corresponding additional motor drive torque sequence τ_c(t) used to generate the compensating acceleration.

[0057] S5.3: Convert the compensation command format. During the real-time operation phase, the additional motor drive torque sequence τ_c(t) generated in S5.2 is converted into feedforward compensation commands that can be executed by the servo controller.

[0058] Since servo drives typically accept position or speed commands, the system converts the torque sequence τ_c(t) into an equivalent speed increment sequence through integration, and then further integrates to obtain the position increment sequence Δp_c(t), which serves as the final feedforward compensation control sequence.

[0059] The integral operation is implemented using the forward Euler method in the discrete-time system, i.e.: velocity increment Δv(t) = τ_c(t) * Δt / J, position increment Δp_c(t) = ΣΔv(t) * Δt, where Δt is the servo control period and J is the equivalent rotational inertia of the system.

[0060] S5.4: Fusion and execution of the final drive instruction. The feedforward compensation control sequence Δp_c(t) obtained in S5.3 and the original future motion instruction sequence p_cmd(t) are algebraically summed at the control clock cycle level and fused into the final drive instruction sequence p_final(t), with the relationship p_final(t) = p_cmd(t) + Δp_c(t).

[0061] This fusion operation is performed by a dedicated adder circuit in the drive command fusion and execution module, ensuring that the calculation latency is less than one servo control cycle. The fused final drive command sequence is sent to the servo driver in real time via a high-speed communication interface to control the drive motor to perform compensated, refined motion.

[0062] Finally, the model training method S6 in this application is as follows: The spatiotemporal dynamic state encoder (corresponding to S2.2 and S2.4), the future positioning error prediction module (corresponding to S4), and the inverse dynamic neural network model in the feedforward compensation control generation module (corresponding to S5) are all supervised training in the offline training phase through the following unified process.

[0063] S6.1: Training data acquisition.

[0064] The linear motion module is controlled to execute a series of excitation trajectories covering its entire workspace, including but not limited to step signals, ramp signals, sinusoidal sweep signals, and composite motion trajectories simulating actual tasks with different amplitudes and frequencies. During this process, the following operations are performed simultaneously: All operational data are collected through the S1 multimodal sensor array.

[0065] The actual position of the moving slide is measured in real time using a high-precision linear grating ruler (corresponding to S4), and the true value of the positioning error e_true(*t*) at each moment is calculated accordingly.

[0066] Record all historical motion commands from the main motion controller.

[0067] S6.2: Construction of training dataset.

[0068] The raw data collected by S6.1 is processed to construct the data pairs required for model training: Input data: The synchronized sensor data stream, historical motion command sequence, and historical operating state representation vector (for sequence prediction models) are used as input features.

[0069] Target data: For the future positioning error prediction module, its training objective is to make the prediction error sequence e_pred(*t*) output by the module as close as possible to the error true value sequence e_true(*t*) measured by the grating ruler.

[0070] For the inverse dynamics neural network model, the training objective is to make the predicted torque τ_pred(*t*) output by the model after inputting the compensated acceleration a_c(*t*) as close as possible to the actual torque value that can produce that acceleration in actual motion.

[0071] S6.3: Model Training and Optimization.

[0072] The overall system is optimized by combining end-to-end and phased training strategies.

[0073] Loss function definition: The prediction module loss L_pred is calculated using the mean squared error loss function, which calculates the difference between the prediction error sequence and the true value sequence, i.e., L_pred=Σ(e_true(*t*)-e_pred(*t*))².

[0074] Inverse dynamics model loss L_inv: The mean square error loss function is used to calculate the difference between the predicted torque and the true torque (or the equivalent derived torque), i.e., L_inv=Σ(τ_true(*t*)-τ_pred(*t*))².

[0075] Training process: Phase 1 (Joint Training of Spatiotemporal Encoder and Prediction Module): With the inverse dynamics model parameters fixed, the input data constructed in S6.2 is fed into the spatiotemporal dynamic state encoder, and its output, along with future motion commands, is input into the future positioning error prediction module. By minimizing the prediction loss L_pred, the parameters of the spatiotemporal dynamic state encoder (including all parameters of the temporal convolutional network and graph neural network, such as convolutional kernel weights, and the learnable linear transformation matrix W^(*l*) and bias vector b^(*l*) in S2.4) and the future positioning error prediction module are jointly optimized using the backpropagation algorithm.

[0076] The second stage (training of the inverse dynamics model): With the pre-trained spatiotemporal encoder and prediction module fixed, the inverse dynamics neural network model is trained separately using the dataset and L_inv loss.

[0077] The third stage (end-to-end fine-tuning, optional): Connect the entire system (encoder, prediction module, inverse dynamics model) and perform end-to-end fine-tuning with the final motion positioning accuracy as the optimization target.

[0078] Optimization algorithm: The training process uses an adaptive moment estimation optimizer and employs gradient pruning techniques to ensure the stability of the training process.

[0079] S6.4: Model Deployment.

[0080] After the above model training is completed and the validation set loss converges, the optimized model parameters (including all weight matrices and bias vectors) are quantized and solidified, and burned into the memory of the edge computing hardware unit to complete the system deployment and enter the real-time prediction and compensation control process from S1 to S5.

[0081] Through the above method, this embodiment achieves deep perception, accurate prediction, and active compensation of the dynamic behavior of the linear motion module of an industrial robot. In actual testing, when the module performs reciprocating motion at a speed of two meters per second, the peak-to-peak value of the steady-state positioning error under traditional closed-loop control is eighteen micrometers. However, after adopting this method, this error is suppressed to within three and a half micrometers, and the vibration decay time is reduced by 76%. This result verifies the significant technical advantages of this invention in high-speed and high-precision scenarios.

[0082] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0083] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A rapid and accurate predictive positioning method for a linear motion module of an industrial robot, characterized in that, include: Real-time acquisition of operational data of the linear motion module during its motion process; A spatial topology diagram of the linear motion module is constructed, and the real-time acquired operation data is mapped to the corresponding nodes of the spatial topology diagram. The mapped data is processed by a spatiotemporal dynamic state encoder to generate an operation state representation vector that represents the overall real-time dynamic characteristics of the linear motion module. The sequence of future motion commands to be executed is obtained from the main motion controller of the industrial robot. The sequence of future motion commands includes the target position, target velocity and target acceleration commands for a series of future time steps. The running state representation vector and the future motion command sequence are simultaneously fed into the future positioning error prediction module. Based on the current physical state information of the system contained in the input running state representation vector and the future input excitation defined by the future motion command sequence, the future positioning error prediction module calculates and outputs a future positioning error sequence that is synchronized with the time axis of the future motion command sequence. The future positioning error sequence is input into the feedforward compensation control generation module, which calculates and generates a feedforward compensation control sequence to offset the expected error based on the future positioning error sequence. The feedforward compensation control sequence and the original future motion command sequence are algebraically summed at the control clock cycle level to form the final drive command sequence. The final drive command sequence is then sent to the servo driver of the linear motion module to control the drive motor to perform the compensated and refined motion.

2. The rapid and accurate prediction and positioning method for the linear motion module of an industrial robot according to claim 1, characterized in that, Real-time acquisition of operational data of the linear motion module during its motion process, including: A triaxial accelerometer with microelectromechanical system is used as the acceleration sensor, and its placement locations include the drive motor mounting flange, the center of the motion slide, and both ends of the linear motion module base; Metal foil strain gauges are used as strain sensors and are attached to the outer wall of the ball screw nut seat in the form of a full-bridge circuit. A platinum resistance temperature sensor is used as the temperature sensor and is closely attached to the middle of the ball screw surface, the support bearing seats at both ends of the screw, and the vicinity of the stator winding of the drive motor. The real-time torque output current signal and rotor position encoder signal of the motor are read from the servo drive at a frequency of not less than 4 kHz via a high-speed industrial Ethernet bus.

3. The rapid and accurate prediction and positioning method for the linear motion module of an industrial robot according to claim 1, characterized in that, Generate a runtime state representation vector, including: The raw data streams acquired from each sensor are timestamped and resampled to form a multi-channel time series dataset with a unified time reference. The temporal data of each channel is processed by a multi-layer one-dimensional temporal convolutional network. The temporal convolutional network adopts causal convolution and dilated convolution structure to extract the dynamic fluctuation characteristics and long-term dependencies of each physical quantity at different time scales, and generate the preliminary temporal features of each measurement point. Based on the mechanical structure drawings of the linear motion module, an undirected graph is constructed with each sensor deployment point as a node and the reciprocal of the physical connection stiffness or the shortest physical distance between nodes as the edge weight, which serves as a spatial topology graph. The initial temporal features are used as the initial attribute features of each node. An iterative graph convolution operation is performed on the spatial topology graph through a graph neural network. In each iteration, the feature vector of each node is aggregated with the feature information of its neighboring nodes. After multiple layers of graph convolution operation, the feature vectors of all nodes are globally pooled to generate the running state representation vector.

4. The rapid and accurate prediction and positioning method for the linear motion module of an industrial robot according to claim 1, characterized in that, The running state representation vector and the future motion command sequence are simultaneously fed into a future positioning error prediction module, including: The historical sequence of the running state representation vector is input into an encoder of a sequence-to-sequence prediction model based on an attention mechanism, and the inherent laws of the dynamic state evolution of the system are captured through the self-attention mechanism. The future motion command sequence, the encoder output, and the decoder's predicted output from the previous time step are all input into the decoder of the sequence-to-sequence prediction model. Through the synergistic effect of cross-attention and self-attention mechanisms, the predicted values ​​for the future positioning error are generated step by step until the entire future positioning error sequence is output.

5. The rapid and accurate predictive positioning method for the linear motion module of an industrial robot according to claim 1, characterized in that, The positioning error is defined as the difference between the actual position and the commanded position of the moving slide. The actual position of the moving slide is measured in real time by an independent high-precision linear grating ruler and serves as the ground truth label during the training phase of the prediction model.

6. The rapid and accurate predictive positioning method for the linear motion module of an industrial robot according to claim 1, characterized in that, The future positioning error sequence is input into the feedforward compensation control generation module, including: The future positioning error sequence is differentially divided twice to obtain the desired compensating acceleration sequence; The compensating acceleration sequence is input into the inverse dynamics neural network model to directly calculate the required additional motor drive torque sequence; The motor drive torque sequence is converted into a feedforward compensation position or speed command sequence that can be executed by the servo controller, and used as a feedforward compensation control sequence.

7. The rapid and accurate predictive positioning method for the linear motion module of an industrial robot according to claim 1, characterized in that, In the offline training phase, the inverse dynamics neural network model learns from a large amount of actual motion data to establish a nonlinear mapping relationship between the desired compensatory acceleration and the additional motor drive torque required to produce that acceleration.

8. The rapid and accurate predictive positioning method for the linear motion module of an industrial robot according to claim 1, characterized in that, The feedforward compensation control sequence and the original future motion command sequence are algebraically summed at the control clock cycle level, including: If the original instruction is the position instruction p_cmd(t) and the compensation sequence is Δp_c(t), then the final driving instruction is p_final(t) = p_cmd(t) + Δp_c(t); The algebraic summation operation is performed through a dedicated adder circuit, and the final drive command sequence is sent to the servo driver in real time through a high-speed communication interface.

Images