Robot hybrid manufacturing system error self-adaptive distribution and distributed compensation method and equipment

Abstract

The invention belongs to the technical field related to robot machining precision control, and discloses a robot hybrid manufacturing system error self-adaptive distribution and distributed compensation method and device, and the method comprises the steps: (1) building a robot body pose error prediction model based on working condition data in a robot machining process, the robot body pose error prediction model is adopted to predict a robot body six-dimensional pose error; meanwhile, physical information constraints are constructed for external motion units; (2) constructing an unsupervised error distribution model based on the predicted six-dimensional pose error, and mapping the robot body six-dimensional pose error of the robot to be compensated into a robot body error compensation amount and an external motion unit error compensation amount by adopting the unsupervised error distribution model; and (3) carrying out cooperative compensation on the machining error of the robot machining hybrid manufacturing system based on the obtained error compensation amount of the robot body and the error compensation amount of the external motion unit. According to the invention, the error compensation precision is improved.

Description

Technical Field

[0001] This invention belongs to the technical field of robot machining accuracy control, and more specifically, relates to a method and device for adaptive allocation and distributed compensation of errors in a robot hybrid manufacturing system. Background Technology

[0002] Industrial robots are increasingly widely used in manufacturing due to their advantages such as large workspace, high flexibility, and relatively low cost. However, due to factors such as weak rigidity caused by their serial configuration and kinematic parameter errors, the absolute positioning accuracy and trajectory accuracy of robots are poor, making them difficult to apply directly to high-precision machining scenarios such as milling.

[0003] To improve the machining accuracy of robots, existing technologies are mainly divided into two categories: offline compensation and online compensation. Offline compensation methods (such as kinematic parameter calibration, error modeling and prediction) correct instructions in advance before machining, but they have poor adaptability to time-varying and complex machining processes. Online compensation methods compensate through real-time measurement and feedback, offering high accuracy, but are often limited by the openness of the robot system, safety strategies, and the high cost of measurement equipment.

[0004] In recent years, hybrid manufacturing systems have emerged, integrating robots with external motion units (such as micro-motion platforms and additional linear axes) to compensate for the insufficient precision of the robot itself. However, in such multi-source compensation systems, how to rationally and efficiently distribute the total machining error among the robot and external motion units has become a critical issue. Inappropriate distribution can lead to: 1) The robot's sensitive joints are over-adjusted, introducing unnecessary "regeneration errors"; 2) The compensation capability of the external motion unit is not fully utilized; 3) The overall compensation effect failed to reach the optimal level.

[0005] Existing allocation methods are mostly based on experience or simple rules (such as allocating all position errors to external axes and all posture errors to the robot), which fail to fully consider the spatiotemporal characteristics of robot posture errors, the differences in the sensitivity of joint movements, and other physical laws, resulting in limited compensation accuracy. Summary of the Invention

[0006] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and device for adaptive allocation and distributed compensation of errors in a robot hybrid manufacturing system, which aims to solve the problem of low accuracy of existing robot error compensation.

[0007] To achieve the above objectives, according to one aspect of the present invention, a method for adaptive error allocation and distributed compensation in a robot hybrid manufacturing system is provided, comprising the following steps: (1) Based on the working condition data during the robot processing, a robot body pose error prediction model is constructed, and the robot body pose error prediction model is used to predict the six-dimensional pose error of the robot body; at the same time, physical information constraints are constructed for the external motion unit; wherein, the physical information constraints include error allocation equivalence, robot body sensitivity and maximum controllability of external unit; (2) Based on the predicted six-dimensional pose error, an unsupervised error allocation model is constructed. The unsupervised error allocation model is used to map the six-dimensional pose error of the robot body to be compensated into the robot body error compensation amount and the external motion unit error compensation amount. (3) Based on the obtained robot body error compensation amount and external motion unit error compensation amount, the processing error of the robot processing hybrid manufacturing system is compensated in a coordinated manner.

[0008] Furthermore, the robot body pose error prediction model is a spatiotemporal attention embedded neural network model, with a temporal convolutional network as the backbone network and embedded spatial attention mechanism and channel attention mechanism.

[0009] Furthermore, error allocation equivalence is used to ensure that the tool pose after distributed compensation is equivalent to the pose after direct robot compensation; robot body sensitivity is used to limit the compensation amount of highly sensitive joints based on the sensitivity analysis of joint space and pose error space; and external unit maximum controllability is used to maximize the use of the compensation capability of external motion units under the premise of satisfying equivalence and sensitivity constraints.

[0010] Furthermore, the equivalence of the error allocation is quantified by the following formula:

[0011] in For the differential motion vector matrix conversion operator, It is its inverse operation. This describes the tool coordinate system within the robot's base coordinate system. This is the original amount of compensation required on the robot; The compensation amount allocated to external components is in 3 columns; The compensation amount allocated to the robot is in the form of 6 columns.

[0012] Furthermore, define the sensitivity constraint measurement function. Specifically, it is expressed as:

[0013] in, ; For a single sensitivity indicator, it means S One of the elements; For the entire set of sensitivity sets; Let be the i-th joint angle of the robot, with indices from 1 to 6; For the recipient Sensitivity constraints on joint effects, i.e., single sensitivity index It is determined by both the overall set S and the current joint state; The identifier for the x-direction component; All sensitivities calculated for the i-th joint; express The influence of the component in the x-direction; This is the identifier for the y-direction component; Identifier for the z-direction component; The identifier for the Rx direction component; This is the identifier for the Ry direction component; This is the identifier for the Rz direction component.

[0014] Furthermore, the maximum controllability constraint of the external motion unit is quantified by the following formula:

[0015] in Represents the modulus operator. Specifically refers to The first three dimensions.

[0016] Furthermore, the unsupervised error allocation model is a neural network model, and its overall loss function is a weighted sum of the losses from the three physical information constraints:

[0017] in , and These are harmonic parameters used to balance the relationship between the three physical information constraint losses; and Redistribution functions The parameter set and the regularization coefficients of the training process; Constraints on robot body sensitivity.

[0018] Furthermore, , and It is obtained through an external optimization algorithm, with the optimization objective being to minimize the overall loss function.

[0019] The present invention also provides an adaptive allocation and distributed compensation system for errors in a robot hybrid manufacturing system. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the adaptive allocation and distributed compensation method for errors in a robot hybrid manufacturing system as described above.

[0020] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the adaptive allocation and distributed compensation method for errors in a robotic hybrid manufacturing system as described above.

[0021] In summary, compared with the prior art, the robot hybrid manufacturing system error adaptive allocation and distributed compensation method and equipment provided by the present invention have the following beneficial effects: 1. By introducing physical information constraints, this invention makes the error allocation process not only consider mathematical equivalence, but also the dynamic characteristics (sensitivity) of the robot body and the performance advantages of external units, thus achieving a more reasonable allocation at the physical level, suppressing the risk of regeneration error from the source, and thereby improving the accuracy of error compensation.

[0022] 2. By accurately predicting errors using a spatiotemporal attention model and combining it with optimized distributed compensation, globally optimal compensation results are achieved. Experiments show that compared to single-source compensation or simple splitting strategies, this invention significantly reduces processing errors and improves compensation accuracy.

[0023] 3. It adopts unsupervised learning for allocation, eliminating the need for manually formulating complex allocation rules. It can automatically determine the optimal compensation allocation scheme based on the current processing status and error characteristics, making it highly adaptable.

[0024] 4. This invention provides a complete technical solution from error perception, physical constraint modeling, intelligent allocation to collaborative compensation, and constructs a corresponding hybrid manufacturing system that is easy to implement and promote. Attached Figure Description

[0025] Figure 1 This is a flowchart of an adaptive error allocation and distributed compensation method for a robot hybrid manufacturing system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the data preprocessing process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the results of the pose error prediction model with spatiotemporal attention embedding in an embodiment of the present invention; Figure 4This is a schematic diagram of the physical constraints of the unsupervised error allocation model in an embodiment of the present invention; Figure 5 (a) and (b) in the figure are schematic diagrams of the harmonic weight optimization process and the final result during the execution of the embodiments of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0027] This invention provides an adaptive allocation and distributed compensation method for errors in a robot hybrid manufacturing system. The compensation method solves the problem of lack of theoretical basis for the allocation of compensation sources in robot hybrid manufacturing systems, realizes the collaborative compensation control of the robot body and external motion units for hybrid manufacturing systems, and improves the accuracy performance of robot hybrid manufacturing systems.

[0028] The compensation method guides unsupervised learning through physical information constraints, intelligently and reasonably allocating the predicted robot pose error to the robot body and external motion units to achieve optimal collaborative compensation.

[0029] The compensation method mainly includes the following steps: S1. A robot body pose error prediction model is constructed based on the working condition data during the robot's machining process. This model is then used to predict the six-dimensional pose error of the robot body. The robot body pose error prediction model is a spatiotemporal attention-embedded neural network model.

[0030] By collecting working condition data during the robot's machining process and using a spatiotemporal attention-embedded neural network model, the pose error of the robot's end effector can be predicted. By performing robot machining on a standard workpiece, working condition data during the machining process can be collected, and a six-dimensional pose error of the robot's end effector can be predicted using a spatiotemporal attention-embedded neural network model.

[0031] The operating data includes the robot's joint angles during processing. Calculated joint velocities Joint torque All three are uniformly described as Furthermore, the generalized forces at the robot's end effector are related to machining. Including cutting force F and torque M Output torque on the electric spindle It is also collected to reflect the load status. The actual pose data of the robot's end effector is measured by an external laser tracker. .

[0032] The spatiotemporal attention embedding neural network model uses a temporal convolutional network as its backbone and embeds spatial and channel attention mechanisms to achieve the fusion of spatiotemporal features. The model training process can be described as utilizing... , and As input data, a spatiotemporal attention-embedded robot body pose error prediction model is trained. The trained model can then predict the six-dimensional pose error of the robot body during machining. The prediction result is expressed as follows: The prediction process can be described as follows: ,in After one-dimensional dilated convolution, the elements in the sequence s The output can be represented as:

[0033] in Represents the convolution operator. d This is an expansion factor used to increase the receptive field of the model. k This represents the size of the filter. The first in the filter i Element-wise operation functions. For all input sequences The first in Each element, It is also related to the direction of historical data. Overall, the effective length of historical time series data is... In this real-time reversal method, d As the number of layers increases, according to 2 n ( n =0,1,2) The exponent increases, and the filter size is set to... k =3.

[0034] However, a larger receptive field requires a deeper network structure. In this case, to prevent gradient vanishing during backpropagation, residual blocks

[37] similar to those in ResNet and weight normalization are added to the backbone network. The network output after residuals can be expressed as:

[0035] in ReLU is used as the activation function. Train the parameter matrix for the model. This is the mapping function corresponding to the first layer model, with the upper right corner indicating the layer number. This is the mapping function corresponding to the second-layer model. This is the mapping function corresponding to the third layer model. For input features.

[0036] Subsequently, to achieve attention extraction and focusing of temporal and spatial features in the mapping process of processing errors, spatial attention and channel attention mechanisms are employed to extract attention from the original high-dimensional features and effectively embed them into downstream tasks, ensuring accurate prediction of processing errors. The weight mask under the spatial attention mechanism can be calculated as follows:

[0037] in and These represent the max pooling and average pooling operators, respectively. After concatenating the results, a 1D convolutional network is used to transform the original features into single-channel feature maps. Finally, the maps are processed... The activation function yields a spatial attention weight mask. In this embodiment, It is set to the sigmoid function. Based on this, the output features are weighted by a spatial attention mechanism. It can be represented as:

[0038] Similarly, the weight mask under the channel attention mechanism can be calculated as follows:

[0039] in, and These represent the max pooling and average pooling operators, respectively. Both are processed by encoders that first compress and then expand to obtain their corresponding outputs. The results are then summed and passed through... The activation function yields the corresponding weight mask. Based on this, the output features are weighted using a channel attention mechanism. It can be represented as:

[0040] After constructing new features using spatial and channel attention mechanism mask weighting, targeting and The dimension, design the corresponding 1 d A convolutional neural network is used to concatenate the two data parts using a feature augmentation method. A subsequent linear sequence (1) is designed to map the fused high-dimensional data onto the measured part machining error. The mapping relationship can be expressed as:

[0041] in This represents the direct prediction result of the robot's pose error. and Each represents a specific design 1 d Convolutional neural networks are primarily used to unify the dimensions of data; the concatenated data is then fed into the design. This is used to predict robot pose errors. The loss function during model training can be expressed as:

[0042] in The L2 regularization coefficient is used to reduce the risk of overfitting and improve the model's generalization performance to some extent. For the loss function described above, the Adam optimizer is used to optimize gradient descent.

[0043] S2, construct physical information constraints for external motion units; among which, physical information constraints include error allocation equivalence, robot body sensitivity, and maximum controllability of external units.

[0044] After constructing the robot body pose error prediction model, the six-dimensional pose error of the robot body under the data acquisition of the machining process can be predicted for any robot machining task. However, for hybrid manufacturing systems, the control unit involves external units in addition to the robot body. The existence of external units necessitates consideration of how to reasonably, accurately, and reliably allocate the compensation control quantities of different motion units. Therefore, this invention constructs three physical information constraints from three different perspectives: S2.1 error allocation equivalence, S2.2 robot body sensitivity constraint, and S2.3 maximum controllability of external units.

[0045] Specifically, the S2.1 physical information constraint is to ensure that the tool pose after distributed compensation is equivalent to the pose after direct robot compensation; the S2.2 physical information constraint is to limit the compensation amount of highly sensitive joints based on the sensitivity analysis of joint space and pose error space; the S2.3 information physical constraint is to maximize the use of the compensation capability of external motion units under the premise of satisfying equivalence and sensitivity constraints.

[0046] The error allocation equivalence is quantified by the following formula:

[0047] in For the differential motion vector matrix conversion operator, It is its inverse operation. This describes the tool coordinate system within the robot's base coordinate system. The above formula describes the original compensation amount required on the robot. With the process The effect deviation in the tool coordinate system after distribution is quantified.

[0048] The purpose of robot body sensitivity constraints is to account for the differentiated sensitivity between the robot joint space and the end effector pose error space under different external loads during robot machining, depending on the joint configuration. This sensitivity essentially characterizes which joints' small movements have the greatest impact on the end effector error. In compensation, units with smaller impacts are expected to be adjusted, while units with larger impacts are expected to remain in their original state. A sensitivity constraint metric function is defined. Specifically, it is expressed as:

[0049] in, Conventional compensation strategies in Under the compensation amount, the corresponding change in robot joint space can be expressed as:

[0050] in Let P be the inverse kinematics function of the robot, and P be the theoretical pose of the tool during the machining process. After redistribution, the new joint space changes can be expressed as:

[0051] To ensure that the redistributed robot joint space compensation meets the expectations of the sensitivity analysis, the robot motion sensitivity constraint is defined as follows: .

[0052] The maximum controllability constraint of the external motion unit is quantified by the following formula:

[0053] in Represents the modulus operator. Specifically refers to The first three dimensions.

[0054] S3. Based on the predicted six-dimensional pose error, an unsupervised error allocation model is constructed. The unsupervised error allocation model is used to map the six-dimensional pose error of the robot body to be compensated into the robot body error compensation amount and the external motion unit error compensation amount.

[0055] Based on operating condition data The robot pose error predicted by S2 As input, through the error allocation model The robot's pose error can be mapped to... ,in This is the compensation amount for the robot's motion unit. The compensation amount for the external motion unit is represented by the overall mapping process as follows: .

[0056] The unsupervised error allocation model is a neural network model, and its overall loss function is a weighted sum of the losses from the three physical information constraints:

[0057] in , and The harmonic parameter is used to balance the relationship between the losses of the above-mentioned constraints. and Redistribution functions The parameter set and the regularization coefficients of the training process.

[0058] During model training, the three harmonic parameters are obtained through an external optimization algorithm. However, due to the redistribution mapping model... The training process is essentially unsupervised, and the merits of the chosen harmonic parameters cannot be directly evaluated based on the model's predictive performance. Therefore, with the goal of minimizing the numerical value of the loss function after model training, the overall optimization problem can be defined as:

[0059] Where j represents the optimization round. The current objective function definition eliminates the influence of different harmonic parameters on the optimization objective value. The PSO algorithm is used during optimization, with the initial sample size set to 100. A linearly decreasing weight update strategy is employed during optimization. This adjusted optimization algorithm significantly improves its local optimization capability and mitigates the premature convergence defect of the particle swarm optimization algorithm to some extent.

[0060] S4. Based on the obtained robot body error compensation amount and external motion unit error compensation amount, the processing error of the robot processing hybrid manufacturing system is compensated collaboratively.

[0061] The robot body error to be compensated and the external motion unit error to be compensated obtained in step S3 are sent to the robot controller and the external motion unit controller respectively to perform collaborative compensation for robot processing errors.

[0062] This invention also provides a robot hybrid manufacturing system employing the error compensation method described above. The system includes: an industrial robot, an external motion unit installed at the end effector of the industrial robot, a data acquisition system for collecting operational data from the robot's processing. The system is equipped with two types of controllers: a central controller, which deploys a pose error perception model, a physical information constraint model, and an unsupervised error allocation model, for executing steps S1 to S3 of the method described above; and a robot controller, which receives the robot body error to be compensated from the central controller and performs compensation. In specific implementation, the external unit controller receives the external motion unit error to be compensated from the central controller and performs compensation.

[0063] The present invention will be further described in detail below with reference to specific embodiments.

[0064] This invention provides a method for adaptive error allocation and distributed compensation in a robot hybrid manufacturing system, such as... Figure 1 As shown, it includes: 1 Robot body pose error prediction model 1.1 System Construction The robotic machining hybrid manufacturing system includes a Staubli TX2-90L industrial robot and a three-degree-of-freedom high-precision linear module (as an end effector) mounted at the robot's end effector, which is equipped with an electric spindle and cutting tools. For data acquisition, it includes internal robot sensors (collecting joint angles and torque), a six-dimensional force sensor (collecting end-effector forces), a spindle status sensor (collecting torque), and a laser tracker (used to collect actual pose errors during model training; not required during compensation and verification). For control, it includes a central industrial control computer (running a Windows system and deploying the algorithms used in this embodiment), a robot body controller, and an external linear axis controller (using EtherCAT-based servo drives). The central industrial control computer communicates with the robot controller and the external linear axis controller via EtherCAT.

[0065] 1.2 Data Acquisition and Preprocessing during Processing When the robot is machining a representative test workpiece (such as a complex curved surface), joint angles, joint velocities, and joint torques are acquired at 1000Hz; the end effector's six-dimensional force / torque and spindle torque are acquired at 250Hz; simultaneously, a laser tracker acquires the actual end effector pose at 1000Hz, and compares it with the theoretical pose to calculate the true pose error for model training. All acquired data are synchronized and time-aligned.

[0066] The collected data is segmented into time series using an asynchronous look-ahead sliding window strategy. During the sliding window process, the sequence length of the label space samples and the length of the leading time window remain consistent. The specific process is as follows: Figure 2 As shown.

[0067] 1.3 Model Construction and Training Build as Figure 3 The spatiotemporal attention-embedded robot pose error prediction model shown has the following basic network: a three-layer TCN (Temporal Convolutional Network) with an exponentially increasing dilation factor of 2 and a filter size k=3. The TCN effectively captures the temporal dependence of the error. The spatiotemporal attention mechanism involves parallel incorporating a spatial attention module (SAM) and a channel attention module (CAM) onto the features output by the TCN, enabling the model to focus on the most critical feature dimensions and time points for the current error prediction. The output layer fuses the output features of SAM and CAM and maps them to the six-dimensional pose error through a fully connected layer.

[0068] During training, the mean squared error loss function with a regularization term is used. The model is trained under supervision using pre-acquired labeled data (working condition data + pose error ground truth).

[0069] 2 Physical Information Constraint Modeling Please refer to section 4. For this step, it is necessary to perform actual modeling of the three types of physical constraints involved in the robotic processing hybrid manufacturing system, specifically: 2.1 Error Allocation Equivalence Error allocation equivalence is quantified by the following formula:

[0070] in For the differential motion vector matrix conversion operator, It is its inverse operation. This describes the tool coordinate system within the robot's base coordinate system. The above formula describes the original compensation amount required on the robot. With the process The effect deviation in the tool coordinate system after distribution is quantified.

[0071] 2.2 Robot Body Sensitivity Constraints Define the robot's ontological sensitivity constraint metric function Specifically, it is expressed as:

[0072] in, Conventional compensation strategies in Under the compensation amount, the corresponding change in robot joint space can be expressed as:

[0073] in Let P be the inverse kinematics function of the robot, and P be the theoretical pose of the tool during the machining process. After redistribution, the new joint space changes can be expressed as:

[0074] To ensure that the redistributed robot joint space compensation meets the expectations of the sensitivity analysis, the robot motion sensitivity constraint is defined as follows:

[0075] 2.3 Maximum controllability of external motion units The maximum controllability constraint of the external unit is quantified by the following formula:

[0076] in Represents the modulus operator. Specifically refers to The first three dimensions. In the constructed hybrid manufacturing system, the external axes are three orthogonal linear axes, therefore the physical constraint here is described as maximizing the adjustment of the external linear units.

[0077] Different hybrid manufacturing systems have different configurations. The model structure in the above embodiments is only for illustrative purposes and does not represent the only option.

[0078] 3. Construction of Unsupervised Error Assignment Model After completing the physical information constraint modeling at different levels, a simple multilayer perceptron (MLP) is established. The input is the concatenation of the robot body pose error (6-dimensional) predicted by the S1 model and all working condition data. The output is a 9-dimensional vector: the first 6 dimensions are the compensation amounts allocated to the robot, and the last 3 dimensions are the compensation amounts allocated to the external axes.

[0079] The training of this unsupervised error assignment model does not require "correct assignments" as labels; its training objective is to minimize the overall loss function.

[0080] Harmonic parameter optimization: Using the Particle Swarm Optimization (PSO) algorithm to minimize To find the optimal value on the training set. , and Combine them. After optimization, fix these parameters. Figure 5 The diagram shows the harmonic weight optimization process and the final result during the execution of the embodiment.

[0081] 4 Distributed compensation control Online Operation: During actual machining, models S1 and S3 are deployed in a central industrial control computer. Real-time data collection of the machining process is input into model S1 to obtain predicted values ​​for the robot's pose error. The prediction results, along with the operating condition data, Simultaneously inputting the model from S3 yields the compensation values ​​to be allocated to the robot body. and the compensation value assigned to the external motion unit .

[0082] Based on this, The path correction is sent to the robot controller via the robot's Alter online path correction interface and superimposed onto the theoretical trajectory in real time. Simultaneously, The FIFO instruction stack of the external linear axis controller is written to control the linear axis motion in real time. In the same minimum motion control cycle, the robot body and the external linear axis move synchronously according to their respective compensation amounts to jointly offset machining errors.

[0083] In a milling experiment of a complex curved surface part made of aerospace aluminum alloy, five strategies were compared: W1: no compensation (blank control experiment); W2: all errors are compensated on the robot; W3: position errors are compensated only by external axes; W4: simple splitting (position errors are distributed to external axes, and attitude errors are distributed to the robot); W5: the method provided by this invention. The results are detailed in Table 1, showing that the method provided by this invention (W5) achieved the best machining quality, reducing the average machining error to 0.029 mm and the peak error to only 0.139 mm, a reduction of nearly 80% compared to the uncompensated state.

[0084] Table 1. Processing errors under different compensation strategies

[0085] The present invention also provides an adaptive allocation and distributed compensation system for errors in a robot hybrid manufacturing system. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the adaptive allocation and distributed compensation method for errors in a robot hybrid manufacturing system as described above.

[0086] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the adaptive allocation and distributed compensation method for errors in a robotic hybrid manufacturing system as described above.

[0087] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive error allocation and distributed compensation in a robot hybrid manufacturing system, characterized in that, The steps are: (1) constructing a robot body pose error prediction model based on working condition data in a robot machining process, predicting a robot body six-dimensional pose error using the robot body pose error prediction model, and constructing physical information constraints for an external motion unit; the physical information constraints include error distribution equivalence, robot body sensitivity, and maximum controllability of the external unit; (2) constructing an unsupervised error distribution model based on the predicted six-dimensional pose error, and using the unsupervised error distribution model to map the robot body six-dimensional pose error of the robot to be compensated into robot body error compensation and external motion unit error compensation; (3) based on the obtained robot body error compensation and external motion unit error compensation, the machining error of the robot hybrid manufacturing system is compensated.

2. The robotic hybrid manufacturing system error self-adaptive allocation and distributed compensation method of claim 1, wherein: The robot body pose error prediction model is a neural network model with space-time attention embedding, and the time convolution network is a basic backbone network, and the space attention mechanism and channel attention mechanism are embedded.

3. The robotic hybrid manufacturing system error self-adaptive allocation and distributed compensation method of claim 1, wherein: The error distribution equivalence is used to ensure that the tool pose after distributed compensation is equivalent to the pose after direct robot compensation; the robot body sensitivity is used to limit the compensation amount of high sensitivity joints according to the sensitivity analysis of the joint space and the pose error space; and the maximum controllability of the external unit is used to maximize the compensation ability of the external motion unit under the premise of meeting the equivalence and sensitivity constraints.

4. The robotic hybrid manufacturing system error self-adaptive allocation and distributed compensation method of claim 3, wherein: The error distribution equivalence is quantified by the following formula: wherein is the differential motion vector matrix operator, is its inverse operation, represents the description of the tool coordinate system in the robot base coordinate system; is the original compensation amount required on the robot; is the compensation amount assigned to the external elements, which has a size of 3 columns; is the compensation amount assigned to the robot, which has a size of 6 columns.

5. The robotic hybrid manufacturing system error self-adaptive allocation and distributed compensation method of claim 4, wherein: Defining a sensitivity constraint metric function , which is expressed as: wherein, ; is a single sensitivity index, representing S one element in the set of sensitivities; is the universal set of sensitivities; is the i-th joint angle of the robot, with index from 1 to 6; is the sensitivity constraint affected by the i-th joint, i.e. a single sensitivity index is determined according to the overall set S and the current joint state; is the x-direction component identifier; is the overall sensitivity calculated for the i-th joint; represents the component influence in the x-direction; is the y-direction component identifier; is the z-direction component identifier; is the Rx-direction component identifier; is the Ry-direction component identifier; is the Rz-direction component identifier.

6. The robotic hybrid manufacturing system error self-adaptive allocation and distributed compensation method of claim 5, wherein: The maximum controllability constraint of the external motion unit is quantified by the following formula: wherein denotes the length of the module, specifically refers to the first three dimensions.

7. The robotic hybrid manufacturing system error self-adaptive allocation and distributed compensation method of claim 6, wherein: The unsupervised error distribution model is a neural network model, and the overall loss function is the weighted sum of the three physical information constraint losses: where , and are harmonic parameters used to balance the relationship between the three physical information constraint losses; and are respectively a set of parameters of the redistribution function and a regularization coefficient of the training process; is a constraint on the robot's proprioceptive sensitivity.

8. The robotic hybrid manufacturing system error self-adaptive allocation and distributed compensation method of claim 7, wherein: , and are obtained by an external optimization algorithm, and the optimization is performed with the numerical value of the total loss function minimized as the optimization objective.

9. A robotic hybrid manufacturing system error self-adaptive allocation and distributed compensation system, characterized in that: The system includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to execute the robot hybrid manufacturing system error adaptive distribution and distributed compensation method of any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: The computer readable storage medium stores machine executable instructions, and when the machine executable instructions are called and executed by the processor, the machine executable instructions cause the processor to implement the robot hybrid manufacturing system error adaptive distribution and distributed compensation method of any one of claims 1-8.

Images