A robot positioning error compensation method and system driven by mechanism data fusion

Abstract

The application belongs to the technical field of error compensation, and discloses a robot positioning error compensation method and system driven by mechanism data fusion, which comprises the following steps: S1, calculating robot body joint torque and load joint torque, wherein the robot body joint torque and the load joint torque constitute joint torque; at the same time, the difference between the end theoretical position and the end actual position is obtained to obtain the positioning error; S2, using a data set to train the positioning error prediction model of the robot, and using the trained positioning error prediction model to predict the positioning error, and then realizing the positioning error compensation of the robot based on the positioning error; wherein the data set comprises an input space and an output space, the command joint position and the joint torque constitute the input space, and the positioning error constitutes the output space. The application can adapt to different workloads.

Description

Technical Field

[0001] This invention belongs to the technical field of error compensation, and more specifically, relates to a robot positioning error compensation method and system driven by mechanism data fusion. Background Technology

[0002] With the continuous development of science and technology, the digitalization, networking, and intelligentization of manufacturing have become core technologies of the new round of industrial revolution. Against this backdrop, industrial robots, as multi-degree-of-freedom mechanical devices, are widely used in fields such as machinery manufacturing and aerospace due to their advantages of high efficiency, long working hours, and large workspace. The actuators installed at the end effector of robots can perform various tasks such as assembly, handling, and welding, and their widespread application has greatly liberated labor and developed productivity.

[0003] Positioning accuracy is a quantitative indicator of the positioning error of industrial robots. It reflects the accuracy with which the robot achieves the end-effector pose according to instructions. Too low positioning accuracy may lead to failure due to low operation accuracy, which directly affects the quality and efficiency of a process or even the entire production line. It restricts the large-scale application of industrial robots in high-load and high-precision scenarios in modern manufacturing.

[0004] Positioning accuracy can be divided into absolute positioning accuracy and repeatability. Absolute positioning accuracy refers to the absolute accuracy with which a robot's end effector reaches a designated position according to instructions, while repeatability refers to the repeatability of the robot reaching the same position multiple times according to the same instructions. Typically, industrial robots have high repeatability, reaching below 0.1mm; while absolute positioning accuracy is relatively low, often above 1mm. Furthermore, considering the advantages of offline operation, such as low cost, simple equipment, and convenient operation, most work scenarios currently use offline methods to operate robots to complete tasks. This method requires the robot's absolute positioning accuracy to meet the accuracy requirements of the work task. The positioning error addressed in this technology specifically refers to absolute positioning error, and the corresponding indicator is absolute positioning accuracy.

[0005] Based on the different inducing factors, positioning errors can be divided into two categories: geometric errors and non-geometric errors. Geometric errors, also known as kinematic parameter errors, refer to the small deviations between the actual and theoretical values ​​of kinematic parameters. These errors can be effectively compensated for through kinematic parameter calibration. Non-geometric errors are more complex in composition, mainly including joint flexibility errors, link flexibility errors, and errors caused by thermal effects. These error factors are often complex and interdependent, making it difficult to model them individually based on their mechanisms. In recent years, deep learning technology has developed rapidly. By leveraging the powerful nonlinear mapping capabilities of models such as deep neural networks, it is possible to accurately predict non-geometric errors, thereby improving positioning accuracy through error compensation.

[0006] The positioning error prediction and compensation technology can be roughly divided into three steps. The first step is to use a laser tracker to measure and acquire the robot's positioning error dataset. The second step is to train a neural network model based on the dataset to obtain the error prediction value of the target point. The third step is to compensate the target point with the error prediction value, thereby reducing the positioning error of the target point. Current technologies first measure the positioning error of the end effector corresponding to the joint point in joint space, or collect stiffness and deformation data as error data in Cartesian space. Then, a neural network model is used to predict the positioning error of the target point, and the command position of the point is modified according to the error prediction value to achieve error compensation. However, the error prediction model only consists of the command joint position as the input space and lacks prior knowledge of the dynamic mechanism. This means that the model is only applicable to situations where the workload is determined, and the workload size needs to be the same as the workload size when the model was trained and the data was collected, which has certain limitations. Summary of the Invention

[0007] To address the above-mentioned deficiencies or improvement needs of existing technologies, this invention provides a mechanism-data fusion-driven robot positioning error compensation method and system, which aims to solve the problem that existing error compensation methods are only applicable to situations where the workload is determined.

[0008] To achieve the above objectives, according to one aspect of the present invention, a robot localization error compensation method driven by mechanism data fusion is provided, the method comprising the following steps:

[0009] S1, calculate the joint torque of the robot body and the joint torque of the load, which together constitute the joint torque; at the same time, the difference between the theoretical position and the actual position of the end effector is calculated to obtain the positioning error;

[0010] S2, the robot's positioning error prediction model is trained using a dataset, and the trained positioning error prediction model is used to predict the positioning error. Then, the positioning error of the robot is compensated based on the positioning error. The dataset includes an input space and an output space. The input space consists of the command joint position and joint torque, and the output space consists of the positioning error.

[0011] Furthermore, before training, a load index is introduced, and the dataset is then expanded based on the load index and an interpolation method; wherein, the load index is:

[0012]

[0013] In the formula, This represents the terminal load value under full load conditions. This indicates the end load value under no-load conditions.

[0014] Furthermore, assuming n is the number of data sets, the input space... Represented as {q c ,τ}∈R n×12 , where the command joint position {q c The joint torques {τ} are respectively expressed as {q} c1 ,q c2 ,…,q c6} and {τ1,τ2,…,τ6} can be represented in matrix form as follows:

[0015]

[0016] Output space It consists of positioning error data, represented as {ΔP}={ΔX,ΔY,ΔZ}∈R n×3 Represented in matrix form as follows:

[0017]

[0018] The input space and the output space together constitute the dataset.

[0019] Furthermore, the joint torques of the robot body are calculated based on the robot's dynamic model, and the end effector load is converted to each joint to obtain the corresponding load joint torques; the joint torques borne by the robot under any load are expressed as:

[0020]

[0021] Furthermore, when obtaining the theoretical position of the end flange, it is first necessary to calibrate the base coordinate system {b}, the end flange center coordinate system {e}, and the tool coordinate system {t}. After calibration, each coordinate system is represented as {b′}, {e′}, and {t′} respectively. The homogeneous transformation matrix is ​​then calculated using forward kinematics:

[0022]

[0023] in, Let be the homogeneous transformation matrix from the laser tracker coordinate system to the base coordinate system. Let be the transformation matrix from the base coordinate system to the center coordinate system of the end flange. Let n be the transformation matrix from the tool coordinate system to the end flange center coordinate system. x ,n y ,n z ,o x ,o y ,o z ,a x ,a y ,a zLet p represent the rotation components in the homogeneous transformation matrix, respectively. x ,p y ,p z Representing the position components respectively, after substituting the command joint position, in the homogeneous transformation matrix T cmd The theoretical position of the end is obtained from the middle.

[0024] Furthermore, the robot was first measured under full load. and no-load state The positioning error value ΔP max and ΔP min Then, the positioning error value under any load is calculated using interpolation, specifically expressed as follows:

[0025] ΔP=ΔP min +η·(ΔP max -ΔP min ).

[0026] Furthermore, the error prediction model is a deep neural network, which is divided into two parts: a feature extractor and a regression predictor. Each part consists of several layers of hidden layers with different numbers of neurons. Each hidden layer consists of a linear layer, a batch normalization layer, and an activation layer. The hidden layers are connected by a Dropout mechanism.

[0027] Furthermore, the positioning error prediction result is as follows: Represented in component form as The robot's command position is updated based on the compensation amount. After the new command position is entered into the system, positioning error compensation can be achieved.

[0028] The present invention provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the mechanism-data fusion-driven robot positioning error compensation method as described above.

[0029] The present invention also provides a robot positioning error compensation system driven by mechanism data fusion, the system comprising four parts: a torque calculation module, an error measurement module, a network prediction module, and an error compensation module;

[0030] The torque calculation module is used to calculate the joint torque of the robot body based on the robot's dynamic model, and to convert the robot's end effector load to each joint to obtain the corresponding load joint torque. The joint torque of the robot body and the load joint torque together constitute the joint torque.

[0031] The error measurement module is used to calculate the difference between the theoretical end position obtained by calculation and the actual end position obtained by experiment to obtain the positioning error;

[0032] The network prediction module is trained using a dataset and is used to predict positioning errors; the command joint positions and joint torques form the input space, the positioning errors form the output space, and the input space and output space together constitute the dataset.

[0033] The error compensation module is used to compensate for the robot's positioning error based on the positioning error.

[0034] In summary, compared with the prior art, the mechanism data fusion-driven robot positioning error compensation method and system provided by the present invention have the following advantages:

[0035] 1. This invention improves the composition of the input space of the positioning error prediction model by introducing joint torque as a component of the input space. The corresponding compensation method can adapt to different workloads, improving applicability and possessing certain engineering application value.

[0036] 2. By introducing a load index and using joint torque as prior knowledge of the mechanism, the dataset used for model training is expanded based on existing measurement data, which enhances the effectiveness of the data-driven method and improves the engineering practicality of the positioning error prediction model.

[0037] 3. This invention uses the theoretical position of the robot's end effector as the compensation target. The error factors to be compensated are not limited to joint flexibility errors, and are more comprehensive than existing technologies. After using this invention for positioning error prediction and compensation, the robot has higher absolute positioning accuracy. Attached Figure Description

[0038] Figure 1 This is a flowchart of a mechanism-based data fusion-driven robot positioning error compensation method provided by the present invention;

[0039] Figure 2 It is a diagram showing the relationship between the coordinate system of the data acquisition experiment;

[0040] Figure 3 This is a schematic diagram of a neural network structure;

[0041] Figure 4 This is a schematic diagram of the hidden layer structure;

[0042] Figure 5 This is a diagram illustrating the Dropout mechanism. Detailed Implementation

[0043] 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.

[0044] Please see Figure 1 This invention provides a mechanism data fusion-driven robot positioning error compensation system, which includes four parts: a torque calculation module, an error measurement module, a network prediction module, and an error compensation module.

[0045] The torque calculation module is used to calculate the joint torques of the robot body based on the robot's dynamic model, and to convert the robot's end effector load to each joint to obtain the corresponding load joint torques. The robot body joint torques and load joint torques together constitute the joint torques. Specifically, the torque consists of two parts: one part is the robot body joint torque calculated by the dynamic model, and the other part is the load joint torque corresponding to each joint converted from the robot's end effector load.

[0046] The dynamic model can be expressed as:

[0047]

[0048] In the formula, M(q) c ) represents the inertial force term. G(q) represents the Coriolis force and centrifugal force terms. c ) represents the gravity term, q c , These represent the command joint position, velocity, and acceleration, τ. D The body joint torques are calculated from the dynamic model. Since positioning error refers to the absolute error between the robot's end effector reaching the specified position according to the command and the theoretical position, error measurement is based on a quasi-static condition. In the quasi-static condition, both the robot's joint velocity and acceleration are 0, i.e.:

[0049]

[0050] At this point, the dynamic model can be simplified to:

[0051] τ D =G(q) c (3)

[0052] This allows us to obtain the body joint torque of the robot, the magnitude of which is related to the position of the command joint, i.e., the robot configuration.

[0053] When a load is applied to the robot's end effector, each joint generates additional torque to balance the external force applied by the load. The load-bearing joint torque can be expressed as:

[0054] τ L =J T (q c )·M Load (4)

[0055] Wherein J(q) c M is the Jacobian matrix of the robot, which is related to the robot's configuration. Load The external force τ applied to the end load. L This refers to the corresponding load joint torque.

[0056] Therefore, the total torque borne by the robot joints can be expressed as:

[0057]

[0058] The error measurement module is used to calculate the positioning error by subtracting the calculated theoretical end position from the experimentally obtained actual end position. Specifically, the error is obtained by subtracting the calculated theoretical end position from the experimentally obtained actual end position. When obtaining the theoretical end position, the base coordinate system {b}, the end flange center coordinate system {e}, and the tool coordinate system {t} first need to be calibrated. For example... Figure 2 As shown, after calibration, each coordinate system can be represented as {b′}, {e′}, {t′}, respectively. The homogeneous transformation matrix is ​​calculated using forward kinematics as follows:

[0059]

[0060] in, Let be the homogeneous transformation matrix from the laser tracker coordinate system to the base coordinate system. The transformation matrix from the base coordinate system to the end flange center coordinate system is related to the robot joint positions calculated using the given commands. Let n be the transformation matrix from the tool coordinate system to the end flange center coordinate system. x ,n y ,n z ,o x ,o y ,o z ,a x ,a y ,a z Let p represent the rotation components in the homogeneous transformation matrix, respectively. x ,p y ,p z These represent the position components. Substituting the command joint positions, we can find the homogeneous transformation matrix T. cmd The theoretical position of the end is obtained from the middle.

[0061] Similarly, the actual position of the end effector can be obtained through error acquisition experiments:

[0062]

[0063] Therefore, the positioning error can be obtained by subtraction:

[0064]

[0065] The absolute positioning accuracy before compensation can then be calculated:

[0066]

[0067] The network prediction module is trained using a dataset and is used to predict positioning errors; the command joint positions and joint torques form the input space, the positioning errors form the output space, and the input space and output space together constitute the dataset.

[0068] Specifically, the network used to predict the positioning error is a deep neural network. Its function is to acquire data features through a training dataset, and then, given the input space of the target data, to predict the output space of the target data. The dataset can be divided into an input space and an output space; where n is the number of data sets, the input space... It can be represented as {q c ,τ}∈R n×12 , where the command joint position {q c The joint torques {τ} can be expressed as {q}. c1 ,q c2 ,…,q c6} and {τ1,τ2,…,τ6} can be represented in matrix form as follows:

[0069]

[0070] Output space The data consists of positioning error data and can be represented as {ΔP}={ΔX,ΔY,ΔZ}∈R n×3 Represented in matrix form as follows:

[0071]

[0072] The input space and the output space together constitute the dataset. In the input space, the joint torque {τ} consists of the body joint torque and the load joint torque. The body joint torque is only related to the robot's joint configuration, while the load joint torque depends on the weight of the load applied to the end effector. Considering the significant time cost required to measure the error data corresponding to any load through measurement methods, which is generally impractical in engineering, we consider introducing the concept of a load exponent to expand the dataset used for model training based on existing measurement data. The load exponent is defined as:

[0073]

[0074] in, This represents the terminal load value under full load conditions. This represents the end effector load value under no-load conditions. Therefore, the joint torque experienced by the robot under any load can be expressed as:

[0075]

[0076] During the experiment, the robot was first measured under full load. and no-load state The positioning error value ΔP max and ΔP min Then, the positioning error value under any load is calculated using interpolation, which can be specifically expressed as:

[0077] ΔP=ΔP min +η·(ΔP max -ΔP min (14)

[0078] Therefore, the augmented dataset used for neural network training is represented as Where N represents the number of expanded data sets. The purpose of the network prediction module is to find the mapping relationship between the input space and the output space, so as to achieve accurate prediction of the output space given the input space of the target data. This mapping relationship can be expressed as:

[0079]

[0080] The neural network structure used for training is as follows: Figure 3 As shown, this neural network can be divided into two parts: a feature extractor and a regression predictor. Each part consists of several layers of hidden layers with varying numbers of neurons. Figure 4 and Figure 5 As shown, each hidden layer consists of a linear layer, a batch normalization layer, and an activation layer. The Dropout mechanism is used between the hidden layers to reduce overfitting during training.

[0081] The training process can be divided into four steps: feature extraction, regression prediction, loss calculation, and parameter update.

[0082] The first step is feature extraction. Input spatial data x i (q c ;τ) after passing through feature extractor G f After feature extraction, it is transformed into space-invariant features F. i (q c ,τ;θ f Specifically, it can be expressed as:

[0083] F i (q c ,τ;θ f ) = G f (x i (q c ;τ);θ f (16)

[0084] Where, θ f These are the variable parameters that are updated during the training process in the feature extractor.

[0085] The second step is regression prediction. The invariant features are processed by the regression predictor G. r Therefore, the output space is predicted, and the predicted output space value is obtained:

[0086]

[0087] Where, θ r These are the variable parameters that are updated during the training process in the regression predictor.

[0088] The third step is loss calculation. In regression prediction problems, it is necessary to calculate the error between the predicted value in the output space and the data in the output space. The mean squared error (MSE) is usually set as the loss function, specifically expressed as:

[0089]

[0090] Among them, y i The output space data is N, where N is the size of the training set data.

[0091] The fourth step is parameter updating. Based on the calculated loss value, the variable parameters of the feature extractor and regression predictor are updated using the backpropagation (BP) algorithm. The update process is as follows:

[0092]

[0093] After multiple iterations following the steps described above, the deep neural network model used for localization error prediction is trained.

[0094] The error compensation module is used to compensate for the robot's positioning error. Specifically, the positioning error prediction module makes a prediction based on the input spatial data of the target point, and outputs the compensation amount, i.e., the positioning error prediction result. Represented in component form as The error compensation module updates the robot's command position based on the compensation amount. After inputting the new command position into the system, positioning error compensation can be achieved. The specific method is as follows:

[0095] Predict the compensation amount based on the input spatial data of the target location:

[0096]

[0097] The location of the update instruction based on the compensation amount is:

[0098]

[0099] After the system inputs a new command position, the end-effector position measurement result is P′. measure Therefore, the compensated positioning error can be obtained as follows:

[0100]

[0101] Then, the compensated absolute positioning accuracy is calculated:

[0102]

[0103] At this point, the robot's positioning error compensation is complete.

[0104] This invention also provides a robot localization error compensation method driven by mechanism data fusion, the error compensation method mainly includes the following steps:

[0105] S1. The robot body joint torque is calculated based on the robot's dynamic model, and the robot end-effector load is converted to each joint to obtain the corresponding load joint torque. The robot body joint torque and the load joint torque constitute the joint torque. At the same time, the difference between the theoretical position and the actual position of the end-effector is calculated to obtain the positioning error.

[0106] In this embodiment, the dynamic model can be expressed as:

[0107]

[0108] In the formula, M(q) c ) represents the inertial force term. G(q) represents the Coriolis force and centrifugal force terms. c) represents the gravity term, q c , These represent the command joint position, velocity, and acceleration, τ. D The body joint torques are calculated from the dynamic model. Since positioning error refers to the absolute error between the robot's end effector reaching the specified position according to the command and the theoretical position, error measurement is based on a quasi-static condition. In the quasi-static condition, both the robot's joint velocity and acceleration are 0, i.e.:

[0109]

[0110] At this point, the dynamic model can be simplified to:

[0111] τ D =G(q) c )

[0112] This allows us to obtain the body joint torque of the robot, the magnitude of which is related to the position of the command joint, i.e., the robot configuration.

[0113] After a load is applied to the robot's end effector, each joint generates additional torque to balance the external force applied by the load. The load-bearing joint torque can be expressed as:

[0114] τ L =J T (q c )·M Load

[0115] Wherein J(q) c M is the Jacobian matrix of the robot, which is related to the robot's configuration. Load The external force τ applied to the end load. L This refers to the corresponding load joint torque.

[0116] Therefore, the total torque borne by the robot joints can be expressed as:

[0117]

[0118] When obtaining the theoretical position of the end flange, it is first necessary to calibrate the base coordinate system {b}, the end flange center coordinate system {e}, and the tool coordinate system {t}. For example... Figure 2 As shown, after calibration, each coordinate system can be represented as {b′}, {e′}, and {t′} respectively. The homogeneous transformation matrix is ​​calculated using forward kinematics as follows:

[0119]

[0120] In the formula, Let be the homogeneous transformation matrix from the laser tracker coordinate system to the base coordinate system. The transformation matrix from the base coordinate system to the end flange center coordinate system is related to the robot joint positions calculated using the given commands. Let n be the transformation matrix from the tool coordinate system to the end flange center coordinate system. x ,n y ,n z ,o x ,o y ,o z ,a x ,a y ,a z Let p represent the rotation components in the homogeneous transformation matrix, respectively. x ,p y ,p z These represent the position components. Substituting the command joint positions, we can find the homogeneous transformation matrix T. cmd The theoretical position of the end is obtained from the middle.

[0121] Similarly, after the error acquisition experiment, the actual position of the end effector can be obtained, expressed as:

[0122]

[0123] Therefore, the positioning error can be obtained by subtraction:

[0124]

[0125] The absolute positioning accuracy before compensation can be calculated:

[0126]

[0127] S2, the commanded joint positions and joint torques form the input space, the positioning error forms the output space, and the input and output spaces form a dataset. A load index is introduced, and the dataset is then expanded based on the load index and an interpolation method; wherein, the load index is:

[0128]

[0129] In the formula, This represents the terminal load value under full load conditions. This indicates the end load value under no-load conditions.

[0130] The robot's localization error prediction model is a deep neural network. Its function is to acquire data features from a training dataset, and then, given the input space of the target data, predict the output space of the target data. The dataset can be divided into an input space and an output space. Here, we assume n is the number of data sets, and the input space... It can be represented as {q c ,τ}∈R n×12, where the command joint position {q c The joint torques {τ} can be expressed as {q}. c1 ,q c2 ,…,q c6} and {τ1,τ2,…,τ6} can be represented in matrix form as follows:

[0131]

[0132] Output space The data consists of positioning error data and can be represented as {ΔP}={ΔX,ΔY,ΔZ}∈R n×3 Represented in matrix form as follows:

[0133]

[0134] The input space and the output space together constitute the dataset. In the input space, the joint torque {τ} consists of the body joint torque and the load joint torque. The body joint torque is only related to the robot's joint configuration, while the load joint torque depends on the weight of the load applied to the end effector. Considering the significant time cost required to measure the error data corresponding to any load through measurement methods, which is generally impractical in engineering, we consider introducing the concept of a load exponent to expand the dataset used for model training based on existing measurement data. The load exponent is defined as:

[0135]

[0136] In the formula, This represents the terminal load value under full load conditions. This represents the end effector load value under no-load conditions. Therefore, the joint torque experienced by the robot under any load can be expressed as:

[0137]

[0138] In practice, the augmented dataset used for neural network training is represented as follows: Where N represents the number of expanded data sets.

[0139] S3, the robot's positioning error prediction model is trained using the dataset, and the trained positioning error prediction model is used to predict the positioning error, and then the robot's positioning error compensation is realized based on the positioning error.

[0140] The positioning error prediction model determines the mapping relationship between the input space and the output space, thereby enabling accurate prediction of the output space given the input space of the target data. This mapping relationship can be expressed as:

[0141]

[0142] The neural network structure used for training is as follows: Figure 3 As shown, the network can be divided into two parts: a feature extractor and a regression predictor. Each part consists of several layers of hidden layers with varying numbers of neurons. Figure 4 and Figure 5 As shown, each hidden layer consists of a linear layer, a batch normalization layer, and an activation layer. The Dropout mechanism is used between the hidden layers to reduce overfitting during training.

[0143] The training process can be divided into four steps: feature extraction, regression prediction, loss calculation, and parameter update.

[0144] The first step is feature extraction. Input spatial data x i (q c ;τ) after passing through feature extractor G f After feature extraction, it is transformed into space-invariant features F. i (q c ,τ;θ f Specifically, it can be expressed as:

[0145] F i (q c ,τ;θ f ) = G f (x i (q c ;τ);θ f );

[0146] In the formula, θ f These are the variable parameters that are updated during the training process in the feature extractor.

[0147] The second step is regression prediction. The invariant features are processed by the regression predictor G. r Therefore, the output space is predicted, and the predicted output space value is obtained:

[0148]

[0149] In the formula, θ r These are the variable parameters that are updated during the training process in the regression predictor.

[0150] The third step is loss calculation. In regression prediction problems, it is necessary to calculate the error between the predicted value in the output space and the data in the output space. The mean squared error (MSE) is usually set as the loss function, specifically expressed as:

[0151]

[0152] In the formula, y iThe output space data is N, where N is the size of the training set data.

[0153] The fourth step is parameter updating. Based on the calculated loss value, the variable parameters of the feature extractor and regression predictor are updated using the backpropagation (BP) algorithm. The update process is as follows:

[0154]

[0155] After multiple iterations following the steps described above, the deep neural network used for localization error prediction is trained.

[0156] After training, the positioning error prediction model is first used to predict the input spatial data of the target point location, and the output compensation amount, i.e., the positioning error prediction result, is... Represented in component form as Next, the robot's command position is updated based on the compensation amount. After inputting the new command position into the system, positioning error compensation can be achieved. The specific method is as follows:

[0157] Predict the compensation amount based on the input spatial data of the target location:

[0158]

[0159] The location of the update instruction based on the compensation amount is:

[0160]

[0161] After the system inputs a new command position, the end-effector position measurement result is P. m ′ easure Therefore, the compensated positioning error can be obtained as follows:

[0162]

[0163] Then, the compensated absolute positioning accuracy is calculated:

[0164]

[0165] At this point, the robot's positioning error compensation is complete.

[0166] 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 mechanism-data fusion-driven robot positioning error compensation method as described above.

[0167] 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 robot localization error compensation driven by mechanism data fusion, characterized in that, The method includes the following steps: S1, calculate the joint torque of the robot body and the joint torque of the load, which together constitute the joint torque; at the same time, the difference between the theoretical position and the actual position of the end effector is calculated to obtain the positioning error; S2, the robot's localization error prediction model is trained using a dataset, and the trained localization error prediction model is used to predict localization errors. Then, the robot's localization error is compensated based on the predicted localization error. The dataset includes an input space and an output space. The input space consists of the commanded joint positions and joint torques, and the output space consists of the localization error. Before training, the dataset is augmented using a load index and interpolation; the load index is: In the formula, This represents the terminal load value under full load conditions. This represents the end load value under no-load conditions; The end-effector load borne by the robot; First, measure the robot under full load. and no-load state Positioning error value below and Then, the positioning error value under any load is calculated using interpolation, specifically expressed as follows: 。 2. The robot localization error compensation method driven by mechanism data fusion as described in claim 1, characterized in that: Assumption Number of data sets, input space Represented as The command joint position Joint torque They are respectively represented as and Represented in matrix form as follows: Output space It consists of positioning error data, represented as Represented in matrix form as follows: The input space and the output space together constitute the dataset. .

3. The robot localization error compensation method driven by mechanism data fusion as described in claim 2, characterized in that: The joint moments of the robot body are calculated based on the robot's dynamic model, and the end effector load is converted to each joint to obtain the corresponding load joint moments; the joint moments borne by the robot under any load are expressed as: The body joint torques calculated from the dynamic model; For load joint torque; This is the term related to gravity. Let be the Jacobian matrix of the robot.

4. The robot positioning error compensation method driven by mechanism data fusion as described in claim 3, characterized in that: When obtaining the theoretical position of the end point, it is first necessary to establish the base coordinate system. End flange center coordinate system Tool coordinate system After calibration, the coordinate systems are represented as follows: , , The homogeneous transformation matrix is ​​calculated using forward kinematics as follows: in, Let be the homogeneous transformation matrix from the laser tracker coordinate system to the base coordinate system. Let be the transformation matrix from the base coordinate system to the center coordinate system of the end flange. Let be the transformation matrix from the tool coordinate system to the end flange center coordinate system. Let them represent the rotation components in the homogeneous transformation matrix, Representing the position components respectively, after substituting the command joint position, in the homogeneous transformation matrix... The theoretical position of the end is obtained from the middle. .

5. The robot localization error compensation method driven by mechanism data fusion as described in claim 1, characterized in that: The error prediction model is a deep neural network, which is divided into two parts: a feature extractor and a regression predictor. Each part consists of several layers of hidden layers with different numbers of neurons. Each hidden layer consists of a linear layer, a batch normalization layer, and an activation layer. The hidden layers are connected by a Dropout mechanism.

6. The mechanism-based data fusion-driven robot positioning error compensation method as described in any one of claims 1-5, characterized in that: The positioning error prediction result is Represented in component form as The compensation amount is calculated based on the positioning error prediction result, and then the robot command position is updated based on the compensation amount. After the new command position is input into the system, the positioning error compensation can be realized.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the mechanism-data fusion-driven robot positioning error compensation method according to any one of claims 1-6.

8. A robot positioning error compensation system driven by mechanism data fusion, characterized in that: The system is used to implement the mechanism data fusion driven robot positioning error compensation method according to any one of claims 1-6, and includes four parts: torque calculation module, error measurement module, network prediction module and error compensation module; The torque calculation module is used to calculate the joint torque of the robot body based on the robot's dynamic model, and to convert the robot's end effector load to each joint to obtain the corresponding load joint torque. The joint torque of the robot body and the load joint torque together constitute the joint torque. The error measurement module is used to calculate the difference between the theoretical end position obtained by calculation and the actual end position obtained by experiment to obtain the positioning error; The network prediction module is trained using a dataset and is used to predict localization errors. The commanded joint position and joint torque constitute the input space, the positioning error constitutes the output space, and the input space and output space constitute the dataset. The error compensation module is used to compensate for the robot's positioning error based on the positioning error.

Images