CN115847396B - Industrial robot positioning error self-adaptive compensation method based on composite branch neural network - Google Patents
Industrial robot positioning error self-adaptive compensation method based on composite branch neural network Download PDFInfo
- Publication number
- CN115847396B CN115847396B CN202211368818.XA CN202211368818A CN115847396B CN 115847396 B CN115847396 B CN 115847396B CN 202211368818 A CN202211368818 A CN 202211368818A CN 115847396 B CN115847396 B CN 115847396B
- Authority
- CN
- China
- Prior art keywords
- pose
- model
- error
- prediction
- diagnosis
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
Landscapes
- Feedback Control In General (AREA)
- Numerical Control (AREA)
- Manipulator (AREA)
Abstract
The invention discloses an industrial robot positioning error self-adaptive compensation method based on a composite branch neural network, which is characterized in that a neural structure search technology is introduced to design a neural structure search framework of a master-slave controller framework, firstly, under the drive of a master control, a diagnosis slave control automatically creates an error grade diagnosis model according to given pose data with different error grades, then, the prediction slave control automatically creates a compensation pose prediction model according to the pose data with different error grades screened by the diagnosis model, and finally, the master control integrates the diagnosis model and the prediction model to form the composite branch network model. In the application process, the main control firstly activates a diagnosis network branch in the composite branch model to judge the positioning error grade under the current pose, then activates a prediction network branch under the corresponding error grade to generate the compensation pose, and corrects the compensation pose by utilizing the error grade. The method of the invention is applied toAnd UR robot precision compensation cases, the result shows that the performance of the method is superior to that of the similar method.
Description
Technical Field
The invention belongs to the technical field of an absolute positioning accuracy compensation method of an industrial robot, and particularly relates to a self-adaptive compensation method of an industrial robot positioning error based on a composite branch neural network.
Background
The serial industrial robot has the characteristics of universality, operation flexibility, strong space accessibility and the like, and is widely applied under complex working conditions such as welding, polishing and the like. Therefore, the robot can be used for processing some oversized components which cannot be realized by a numerical control machine tool, and a novel efficient and high-precision processing mode is provided for the oversized components, and has attracted importance to academia and industry. However, the application of the robot in the aspect of milling large-scale components is still relatively conservative at present, because the milling precision of the workpiece mainly depends on the static precision and dynamic stability of the robot, and especially the static precision of the robot has higher requirements on the absolute positioning precision of the robot. Therefore, improving and optimizing the absolute positioning accuracy of the robot is an effective way to ensure the milling quality of the workpiece.
Is affected by systematic errors (composed of robot structural errors, namely connecting rod parameter errors, joint rotation angle parameter errors and the like, also called geometric errors) and random errors (mainly caused by factors such as robot joint connection deflection, friction, connection gaps, external temperature and the like, also called non-geometric errors), and the actual pose of the robot end effector is inconsistent with the target pose in actual working, so that the actual running track of the robot has larger deviation from the offline planned target running track. For the two errors, researchers propose a parameter method based on kinematic model compensation and a non-parameter method for directly establishing a mapping relation between the actual pose of the end effector of the robot and the target pose.
Through general researches, the robot positioning error compensation problem has the following two characteristics: in space, the positioning error levels are distributed differently, when the pose of the robot changes greatly, namely the rotation range of each joint is large, the deflection, the gap and the friction of the joints are obviously changed, and the positioning errors of the robot in different sub-working spaces are inevitably caused to be different; the degradation of the working performance of the robot in time may significantly deteriorate the positioning accuracy thereof, because the working performance thereof may gradually degrade as the service time of the robot increases, and the kinematic model parameters may also be continuously changed, resulting in a significant increase in the end positioning error thereof. Whether based on a parameter method or an nonparametric method, the optimization of the positioning accuracy of the robot is mainly realized in an offline mode at present. In addition, in the process of optimizing the positioning accuracy of the robot, the influence of the positioning error grade change and the robot performance degradation on the positioning error compensation is also rarely studied and comprehensively considered, so that the established positioning accuracy optimization model cannot be suitable for the full life cycle of the robot, the model needs to be repeatedly and iteratively trained and optimized, the process is time-consuming, and the error compensation effect possibly cannot meet the use requirement of the robot due to the optimized delay. Therefore, there is a need for an adaptive robot positioning error compensation method to overcome the problems of spatial error level difference distribution and temporal degradation of the robot to exacerbate the positioning error.
Disclosure of Invention
Aiming at the defects of the prior art method, the invention aims to provide an industrial robot positioning error self-adaptive compensation method of a composite branch neural network. The method abstracts two key problems in a robot positioning error compensation space and on time into an error level diagnosis and compensation pose prediction problem, introduces a neural structure search technology to design a neural structure search framework of a master-slave controller architecture, and adaptively and hierarchically compensates the robot positioning error.
In order to achieve the above purpose, the invention provides an industrial robot positioning error self-adaptive compensation method of a composite branch neural network, which specifically comprises the following steps:
s1, setting an error grade by a master control, and automatically creating an error grade diagnosis model by a diagnosis slave control based on pose data;
s2, according to the error level set by the master control, the prediction slave control respectively creates a compensation pose prediction model for pose data of different levels;
s3, integrating the error level diagnosis model and the plurality of compensation pose prediction models into a composite branch network model by the main control, and adjusting the division of the error level according to the performance of the model, guiding the slave control to re-model until the compensation precision requirement is met;
s4, aiming at the robot pose data to be compensated, a diagnosis branch of the active model is controlled to judge the error grade of the current pose;
s5, the main control activates a corresponding prediction branch according to the judging result of the model diagnosis branch, outputs a compensation pose, carries out hierarchical filtering and smoothing on the prediction result by utilizing the diagnosis error level, and finally loads the prediction result into the controller to drive the tail end of the robot to move for work so as to realize self-adaptive compensation of positioning errors;
the main control is responsible for coordinating the training and optimization of the slave control, integrating the composite branch neural network model and activating the corresponding channels of the network to perform pose error grade diagnosis and pose prediction compensation, is essentially a control logic, and is the brain of the method;
the diagnosis slave control is a neural network model, an error grade diagnosis model is created by outputting a decision sequence, and a mapping relation between the pose and the error grade is established, so that the influence of difference distribution of the error grade in a robot working space on high-precision error compensation is overcome;
the prediction slave control is a neural network model, a compensation pose prediction model is created by outputting a decision sequence, a mapping relation between an actual pose and a target pose is established, and errors are compensated into the target pose in advance, so that the influence of the positioning errors on high-precision error compensation caused by the degradation of the machine humanization can be overcome;
in the step S5, the step of performing hierarchical filtering and smoothing on the prediction result by using the diagnosed error level refers to filtering the output of the prediction model by using the diagnosed threshold value of a certain level error, removing abnormal values, and ensuring the validity and stability of the prediction result of the model.
Further, the step S1 specifically includes:
s11, the main control demarcates error levels, and actual and target pose data sampled from a robot working space are classified according to the corresponding error levels;
s12, the diagnosis slave control takes the classified pose data as input, outputs a decision sequence, selects a corresponding network layer from the search space, and automatically creates an error level diagnosis model;
s13, according to the precision performance of the diagnosis model, the diagnosis slave control continuously adjusts the optimization decision sequence until an error grade diagnosis model meeting the precision requirement is automatically created;
the diagnosis slave control in the step S12 is a neural network model mixed in series and parallel, the neural network model is input as a null architecture of a diagnosis candidate model, and a decision sequence corresponding to the null architecture of the candidate model is output through calculation of an intermediate network layer;
the search space in step S12 is a network layer library including 8 layers of networks including a convolution layer, a full connection layer, a long-short-time memory network layer, a bidirectional long-short-time memory network layer, and the like of different super parameter combinations.
Further, the accuracy of the diagnostic model in the step S13 is expressed as follows:
in which A d Measuring finger for representing accuracy performance of diagnostic modelThe scale, N, represents the number of pose points contained in the batch data, K represents the number of error levels set,for the ith component of the diagnostic model for the diagnostic result of the error class in which the jth pose point is located in the batch data +.>The ith component of the error level single-heat coded label corresponding to the jth pose point in the batch data;
the objective function of the diagnostic slave optimization decision sequence in step S13 is as follows:
wherein T is d An objective function of the diagnosis slave control is represented, M represents the number of decision rounds of the diagnosis slave control, T is the length of a decision sequence,for diagnosing the t-th dimensional decision value in the slave control jth round of decision sequence,/for diagnosis>The diagnostic precision of the diagnostic model created by the j-th round of decision, namely the rewarding value of the round of decision, is represented by using cross entropy to describe the distance between the strategy and the return probability distribution, when the two are closer, the cross entropy is smaller, namely when the accumulated rewarding of the decision sequence is larger, the probability of occurrence of the corresponding decision sequence is also larger, the cross entropy loss is smaller, under the given candidate model verification environment, the return corresponding to the same action is the same, and in the interaction process of the controller and the verification environment, the return corresponding to a certain decision sequence is found to be larger, so that the probability of occurrence of the corresponding action is increased by updating the network parameter of the controller through the cross entropy loss function.
Further, the step S2 specifically includes:
s21, the prediction slave control takes pose data under a certain error level as input, outputs a decision sequence, selects a corresponding network layer from a search space, and automatically creates a compensation pose prediction model;
s22, according to the precision performance of the prediction model, the prediction slave control continuously adjusts the optimization decision sequence until a compensation pose prediction model meeting the precision requirement is automatically created;
s23, repeating the process until a compensation pose prediction model meeting the precision requirement is created for pose data under various error grades;
the accuracy of the compensation pose prediction model in the step S22 is expressed as follows:
in the middle ofRepresenting the predicted value of the compensation pose prediction model aiming at the ith dimension coordinate of the jth pose point in the batch data,target value representing the i-th dimensional coordinate, +.>The average value of the target poses is represented, and N represents the number of pose points contained in the batch; the front three-dimensional representation of the coordinates is displaced with an error weight omega 1 The rear three-dimensional representation of the gesture, weighted omega 2 Mu is the weight of the error grade, and A is the weight of the error grade which becomes larger along with the increase of the error grade p The closer to 1, the higher the prediction accuracy of the compensation pose prediction model is, and the worse the prediction reliability is.
Further, the representation of the complex branch network model in the step S3 is expressed as:
wherein A is d For the purpose of diagnosing the accuracy of the model,for the average precision of a plurality of compensation pose prediction models, the coefficient alpha and beta comprehensive evaluation composite branch network model is set in consideration of the influence of the diagnosis precision and the prediction precision, if the coarse granularity divides the error level, the diagnosis difficulty is reduced, but the prediction difficulty is increased; if the error grade is too thin, the diagnosis difficulty can rise, and although the prediction difficulty is reduced, the part of the pose of the diagnosis deviation can call an incorrect prediction model to compensate, and the final prediction precision can be affected.
The beneficial effects of the invention are as follows:
the invention discloses an industrial robot positioning error self-adaptive compensation method based on a composite branch neural network, in the method, influence factors of two aspects of robot positioning error compensation space and time are abstracted into error level diagnosis and compensation pose prediction problems, a neural structure search technology is introduced to design a neural structure search framework of a master-slave controller architecture, firstly, a diagnosis slave control automatically creates an error level diagnosis model according to given pose data with different error levels under the drive of a master control, then the prediction slave control automatically creates a compensation pose prediction model according to pose data with different error levels screened by the diagnosis model, and finally the master control integrates the diagnosis model and the prediction model to form the composite branch network model. In the application process, the main controller firstly activates a diagnosis network branch in the composite branch model to judge the positioning error grade under the current pose, then activates a prediction network branch under the corresponding error grade to generate a compensation pose, corrects the compensation pose by utilizing the error grade, and loads the corrected compensation pose into the robot controller to realize pose compensation. The method of the invention is applied toThe robot and UR robot accuracy compensation cases, the results show that the method will +.>The positioning error of the industrial robot is reduced from 0.884mm to 0.135mm, and the positioning error of the UR robot is reduced from 2.11mm to 0.135mm0.158mm, and the performance is better than that of the similar method.
Drawings
FIG. 1 is a flow chart of an implementation of an adaptive compensation method for positioning errors of an industrial robot based on a composite branch neural network;
FIG. 2 is a slave control structure diagram provided by the invention;
FIG. 3 is a schematic diagram of a long-short-term memory network according to the present invention;
FIG. 4 is a schematic diagram of a verification scheme provided by the present invention;
FIG. 5 is a graph of hierarchical filtering and smoothing effects provided by the present invention;
FIG. 6 is a chart of master-slave performance index records of an automatic modeling process provided by the invention;
FIG. 7 is a diagram showing the output dimension reduction of the error level diagnostic model provided by the invention;
FIG. 8 is a statistical diagram of positioning errors before and after compensation according to the present invention;
FIG. 9 is a three-dimensional representation of the robot trajectory before and after compensation provided by the present invention;
Detailed Description
As shown in fig. 1, the method for adaptively compensating the positioning error of the industrial robot based on the composite branch neural network comprises the following steps:
s1, setting an error grade by a master control, and automatically creating an error grade diagnosis model by a diagnosis slave control based on pose data;
s2, according to the error level set by the master control, the prediction slave control respectively creates a compensation pose prediction model for pose data of different levels;
s3, integrating the error level diagnosis model and the plurality of compensation pose prediction models into a composite branch network model by the main control, and adjusting the division of the error level according to the performance of the model, guiding the slave control to re-model until the compensation precision requirement is met;
s4, aiming at the robot pose data to be compensated, a diagnosis branch of the active model is controlled to judge the error grade of the current pose;
s5, the main control activates a corresponding prediction branch according to the judging result of the model diagnosis branch, outputs a compensation pose, carries out hierarchical filtering and smoothing on the prediction result by utilizing the diagnosis error level, and finally loads the prediction result into the controller to drive the tail end of the robot to move for work so as to realize self-adaptive compensation of positioning errors;
the main control is responsible for coordinating the training and optimization of the slave control, integrating the composite branch neural network model and activating the corresponding channels of the network to perform pose error level diagnosis and pose prediction compensation, is essentially a control logic, and is the brain of the method;
the diagnosis slave control is a neural network model, an error grade diagnosis model is created by outputting a decision sequence, and a mapping relation between the pose and the error grade is established, so that the influence of difference distribution of the error grade in a robot working space on high-precision error compensation is overcome;
the prediction slave control is a neural network model, a compensation pose prediction model is created by outputting a decision sequence, a mapping relation between an actual pose and a target pose is established, and errors are compensated into the target pose in advance, so that the influence of the positioning errors on high-precision error compensation caused by the degradation of the machine humanization can be overcome;
in the step S5, the step of performing hierarchical filtering and smoothing on the prediction result by using the diagnosed error level refers to filtering the output of the prediction model by using the diagnosed threshold value of a certain level error, removing abnormal values, and ensuring the validity and stability of the prediction result of the model.
The step S1 specifically includes:
s11, the main control demarcates error levels, and actual and target pose data sampled from a robot working space are classified according to the corresponding error levels;
s12, the diagnosis slave control takes the classified pose data as input, outputs a decision sequence, selects a corresponding network layer from the search space, and automatically creates an error level diagnosis model;
s13, according to the precision performance of the diagnosis model, the diagnosis slave control continuously adjusts the optimization decision sequence until an error grade diagnosis model meeting the precision requirement is automatically created;
the diagnosis slave control in the step S12 is a neural network model mixed in parallel and series, which is input as a null architecture of the diagnosis candidate model, and outputs a decision sequence corresponding to the null architecture of the candidate model through the calculation of the intermediate network layer. As shown in fig. 2, the input to the diagnostic slave is a zero vector of size (10, 1) representing the layer 10 network for which the candidate diagnostic model is to be determined. Considering that the high-low layer features extracted by the continuous convolution operation cannot be replaced with each other, selecting 5 convolution layers with small sizes for feature extraction, and combining the extracted features to form a mixed feature so as to support the slave control to make better decisions. In addition, a focus mechanism is introduced into the slave control structure, different weights are given to different components in the mixed characteristics, and the characteristic diversity is improved. Specifically, firstly compressing mixed features into row vectors through a global average pooling layer, storing Cheng Yuzhi base, and then converting a threshold base into a threshold coefficient belonging to (0, 1) by utilizing a full-connection layer and a standardization layer; and converting the threshold coefficient into a column vector and multiplying the column vector by a threshold base to obtain a threshold matrix with the same size as the original mixed characteristic, and finally, performing difference between the characteristic matrix and the threshold matrix to obtain a decision matrix (10, 5). On one hand, setting the component smaller than zero in the decision matrix as zero, and eliminating the fine interference in the original characteristics; on the other hand, the values of the different components in the threshold matrix are different and can be dynamically adjusted during training, which is equivalent to applying different "attentions" to the different components of the original feature matrix, which is an embodiment of the attentiveness mechanism. The 10 rows of the decision matrix represent 10-dimensional decisions that determine the candidate model, corresponding to the 10 layers of the candidate model, respectively, and the 5 columns represent 5 options per dimensional decision, namely the empty network layer and the 4 networks provided in the search space. The setting of the empty network layer decision option is convenient for the controller to adjust the depth of the candidate model, and the decision degree of freedom is improved. The structure of the predictive slave is the same as the diagnostic slave.
The search space in the above step S12 is shown in table 1. The search spaces corresponding to the diagnosis slave control and the prediction slave control are the same, and each search space consists of a convolution layer, a full connection layer and a long-short-time memory network layer, as shown in table 1, and each search space comprises four convolution network layers with different super parameters, two full connection layers, one long-short-time memory network layer and one bidirectional long-short-time memory network layer. The upper depth limit of the candidate model is 10 layers, wherein the first 6 layers are constructed by a convolutional neural network and are called a feature extraction module; the latter 4 layers are constructed by fully connected network or long and short memory network layers, and correspond to diagnosis candidate models called diagnosis modules and prediction candidate models called prediction modules. Therefore, the first six rows of decisions in the controller output decision matrix correspond to the convolution layers in the search space, and the last four rows of decisions correspond to the full connection layer and the long-short-term memory network layer in the search space.
Table 1 search space
The activation function of each convolution network layer of the feature extraction module selects the LeakyReLU, so that the gradient hard zero sparsity can be effectively overcome, the feature extraction module is more robust in the optimization process, and the feature extraction module has positive effects on extracting data features which are more representative and are helpful for the diagnosis and prediction module to make correct judgment. And the active function of the network layer in the diagnosis module and the prediction module selects the ReLU to only output components larger than zero, so that the convergence speed is higher, the calculation speed is faster, and the diagnosis and prediction module is supported to output results more efficiently.
LSTM is composed of a plurality of repeating memory cells, each containing both a hidden state and a cell state, as shown in FIG. 3. Providing LSTM and Bi-LSTM for the controller to select in the search space, wherein the LSTM and Bi-LSTM are provided with sequence return attributes, and the memory unit state and the hidden state are output at the same time; the latter takes into account both the front-to-back and back-to-front effects of the memory cell, which are generally more effective for the prediction problem. The slave will automatically create candidate models of different depths, consisting of different types and different hyper-parametric network layers, according to the difficulty of the problem, e.g. LSTM will not typically be finally selected during the creation of the diagnostic model.
An output layer is added after the diagnosis and prediction module to form a complete error level diagnosis and compensation pose prediction module, the output layer is a full-connection layer, the number of the neurons of the full-connection layer is equal to the number of the error levels determined by the master control, and softMax is selected as an activation function; the latter has a neuron number of 6, and a Sigmoid is selected as an activation function according to six-dimensional coordinates of the pose.
The accuracy of the diagnostic model in step S13 is expressed as follows:
in which A d A measurement index indicating the accuracy of the diagnostic model, N indicating the number of pose points included in the batch data, K indicating the number of error levels set,for the ith component of the diagnostic model for the diagnostic result of the error class in which the jth pose point is located in the batch data +.>The ith component of the error level single-heat coded label corresponding to the jth pose point in the batch data;
the objective function of the diagnostic slave optimization decision sequence in step S13 is as follows:
wherein T is d An objective function of the diagnosis slave control is represented, M represents the number of decision rounds of the diagnosis slave control, T is the length of a decision sequence,for diagnosing the t-th dimensional decision value in the slave control jth round of decision sequence,/for diagnosis>Representing the diagnostic accuracy of the diagnostic model created by the j-th round of decisions, i.e., the rewards value of the round of decisions, the distance between the strategy and the return probability distribution is characterized by cross entropy, which is smaller when the two are closer, i.e., when the cumulative rewards of the decision sequenceThe larger the probability of occurrence of the corresponding decision sequence is, the smaller the cross entropy loss is, and under the given candidate model verification environment, the same return corresponding to the same action is the same, and in the interaction process of the controller and the verification environment, the larger the return corresponding to a certain decision sequence is found, so that the probability of occurrence of the corresponding action is increased by updating the network parameters of the controller through the cross entropy loss function.
The step S2 specifically includes:
s21, the prediction slave control takes pose data under a certain error level as input, outputs a decision sequence, selects a corresponding network layer from a search space, and automatically creates a compensation pose prediction model;
s22, according to the precision performance of the prediction model, the prediction slave control continuously adjusts the optimization decision sequence until a compensation pose prediction model meeting the precision requirement is automatically created;
s23, repeating the process until a compensation pose prediction model meeting the precision requirement is created for pose data under various error grades;
the accuracy of the compensation pose prediction model in the step S22 is expressed as follows:
in the middle ofRepresenting the predicted value of the compensation pose prediction model aiming at the ith dimension coordinate of the jth pose point in the batch data,target value representing the i-th dimensional coordinate, +.>The average value of the target poses is represented, and N represents the number of pose points contained in the batch; the front three-dimensional representation of the coordinates is displaced with an error weight omega 1 The rear three-dimensional representation of the gesture, weighted omega 2 Mu is the weight of the error grade, and the error grade is increased continuously，A p The closer to 1, the higher the prediction accuracy of the compensation pose prediction model is, and the worse the prediction reliability is.
The performance of the composite branch network model in the above step S3 is expressed as:
wherein A is d For the purpose of diagnosing the accuracy of the model,for the average precision of a plurality of compensation pose prediction models, the coefficient alpha and beta comprehensive evaluation composite branch network model is set in consideration of the influence of the diagnosis precision and the prediction precision, if the coarse granularity divides the error level, the diagnosis difficulty is reduced, but the prediction difficulty is increased; if the error grade is too thin, the diagnosis difficulty can rise, and although the prediction difficulty is reduced, the part of the pose of the diagnosis deviation can call an incorrect prediction model to compensate, and the final prediction precision can be affected.
In the embodiment of the invention, the experimental process of completing the self-adaptive compensation of the positioning error of the robot by adopting the method is provided:
1. experiment setting:
to verify the method of the invention, the following steps are respectively adoptedThe TX200 robot and the UR10 CB3 robot are used as verification objects, a laser tracker AT960 is matched with a measurement target T-Mac to form a robot tail end pose measurement system based on a developed joint control software system, and an industrial robot actual pose and target pose data acquisition and positioning error compensation verification experiment is designed, as shown in fig. 4.
For the followingTX200 robot, working space for acquiring robot end pose data based on base coordinate system O-XYZ planning, unit moving distance of each axis and rotating angle toAnd the total number of samples collected are shown in table 1. In the data acquisition process, the robot adopts a compound motion mode of combining multi-axis movement and single-axis rotation, namely, the movement of the end effector is compound motion of moving along three coordinate axes, and the rotation of the end is motion of keeping the 90-degree posture unchanged in the directions of the other two coordinate axes when rotating around one coordinate axis.
In order to meet the requirements of data acquisition of a calibration sample library and on-line compensation of the positioning precision of the robot, a comprehensive control system of a laser tracker and an industrial robot is developed, and a system interface is shown as c of fig. 4. The control system can respectively send measurement instructions and motion control commands to the laser tracker and the robot controller through a TCP/IP protocol, and can acquire actual pose and target pose data of the robot.
For the UR10 CB3 robot, a robot positioning error disclosure data set is selected and provided by French research institute of Germany. The experimental setup for this part is shown in fig. 4 d, and the relevant planning of the robot workspace etc. is shown in table 2. In addition, the robot also adopts a compound motion mode combining multi-axis movement and single-axis rotation in the process of acquiring the pose data of the tail end of the robot.
Table 2 experimental planning
2. Method comparison
Before the positioning error is compensated by the application method, the following 5 groups of experiments are conducted to explore the relationship between the classification of error grades and the influence of the error grades on the performance of the diagnosis model and the prediction model, namely, the error grades 1, 5, 10, 15 and 20 are respectively classified, and the time required by the diagnosis slave control and the prediction slave control to automatically create the model meeting the precision requirement under the corresponding setting, the precision of the diagnosis model, the average precision and the minimum precision of the prediction model created for a plurality of error grades and the average compensation error of the model are recorded. In addition, the ideal average prediction accuracy excluding the error diagnosis so that the deviation of the prediction model was erroneously called is recorded to quantitatively demonstrate the interaction between the error level diagnosis and the pose prediction, as shown in table 3.
Table 3 error grading versus recording
As can be seen from the data in the table, as the master control gradually divides the error level in a fine granularity, the difficulty of error level diagnosis is higher and higher for the diagnosis slave control, so that the time required for creating an error level diagnosis model is longer and longer, and the diagnosis precision is lower and lower; for prediction slave control, the magnitude of a certain level error is narrower and the difficulty of compensating pose prediction is relatively reduced, so that the time required for creating a prediction model for each error level is shorter and shorter, and the ideal average prediction precision of the prediction model (namely, the prediction precision of the prediction model per se) is higher and higher. However, the actual average prediction accuracy and the corresponding average compensation error have a trend of rising and then falling, because the master control can call the corresponding pose prediction model according to the diagnosis result of the diagnosis model, when the diagnosis accuracy is lower and lower, the deviation brought to the prediction is larger and larger, so that the average prediction accuracy is gradually lower than the ideal average prediction accuracy. It is worth to put forward that, the prediction slave has also recorded the minimum precision of prediction model below, even if the average precision of prediction model reaches the requirement, but the prediction precision of a certain error level is too low, can't reach the operation requirement again, therefore the training target of prediction slave has considered the requirement of average precision and minimum precision.
The main control and the auxiliary control can not be well guided to carry out coordination training by singly carrying out performance evaluation on the diagnosis model and the prediction model, thereby providing a total performance indexThe final composite branch network model was evaluated, where α was 0.4 and β was 0.6. In the training process of the master control and the slave control, the target requirement with the total precision higher than 0.98 is set, so that the diagnosis precision higher than 0.95 and the prediction precision higher than 0.967 can be ensured.
An important design in the method of the invention is the level of error through diagnosis in step S5The predicted poses are hierarchically filtered and smoothed, respectively. Invoking the corresponding predictive model forThe data set is distributed at pose points between (1.150,1.237) mm for error compensation, the predicted compensation pose is filtered by using error thresholds of 1.150mm and 1.237mm, and finally convolution smoothing is carried out, and positioning errors before, after, and after the compensation, hierarchical filtering and smoothing of the pose point X, Y, Z and the total displacement are shown in fig. 5. According to the graph, although the distribution center of the pose after compensation output by the prediction model is closer to the target pose relative to the pose before compensation, the positioning error is remarkably reduced, a larger fluctuation range still exists, and the positioning error of the individual pose point even exceeds the position before compensation. The method has the advantages that the accurate diagnosis of the error level of the pose point before the compensation pose prediction is carried out is an innovation point and an advantage of the method, and the filtering of the predicted pose by using the error threshold of the diagnosed error level is natural and meaningful. The figure clearly shows that all abnormal points are removed from the filtered predicted pose, the distribution is more reasonable, and the method has practical application significance. And finally, carrying out convolution smoothing on the filtered predicted pose, further reducing the fluctuation range and improving the stability.
3. Method compensation effect verification
The compensation effect of the method is shown in three aspects of master-slave control automatic modeling, error level diagnosis and compensation pose prediction in sequence. For the purpose ofThe robot verifies the object, the master coordinates the diagnosis slave control and the prediction slave control training by adjusting the division of the error level, and the total accuracy of the composite branch network model, the training loss of the diagnosis slave control and the training loss of the prediction slave control in the process are shown in figure 6. Along with the advance of the main control coordination period, the error grade is gradually divided into fine grains, the loss of the diagnosis slave control is higher and higher in the process, and the error of the diagnosis model is larger and larger; the master control precision is reduced after rising, and the loss of the prediction slave control is also reduced and then enters the rising stage, for the reasonThe discussion in the previous section. It can be seen that, approximately, when the robot positioning error is divided into 8 error levels, the total accuracy reaches the target requirement of more than 0.98 at the highest point, so that the error level diagnosis model and the compensation pose prediction model created by the slave control under the setting are selected.
The positioning errors of pose points in the data set are distributed at (0.46,1.32) mm, the positioning errors are equally divided into 8 error grades, the pose points under each error grade are 266, 497, 476, 245, 371, 448, 364 and 105, 8 layers of error grade diagnosis models created by diagnosis slave control exist, the diagnosis precision reaches 98.7%, and the confusion of a small amount of samples exists between the second class, the first class, the third class, the sixth class, the seventh class and the eighth class. FIG. 7 is a visual illustration of a process for classifying pose data for different error levels for a diagnostic model. Fig. 7 a is original pose data, different colors represent different error levels, and pose data of different error levels cannot be divided by pose. After the feature extraction module built by the convolution layer and the classification module formed by the full connection layer are processed, pose data with the same error level begin to gather, and most of pose data are correctly diagnosed after being processed by the output layer softmax activation function, points with the same color are completely contracted together, points with different colors are kept at a long distance in d of fig. 7, and it is worth to propose that the pose points with pink color, namely the pose points with the error level 8, are only 105, and the overlapping degree is higher after gathering, so that the pose points with pink color after gathering occupy only a small area.
For the UR dataset, its positioning error (0.72,2.53) mm is divided into 10 ranks, containing 216,731,1334,1511,1380,1329,1333,1274,1035,356 pose points, respectively. At each error level, 80% of the pose point data were used as training data and 20% were used as verification data. Such an adaptive compensation verification scenario is designed: randomly scrambling pose point data serving as verification data under each error level according to a group of 10 points, sequentially taking out the pose point data, firstly calling a diagnosis model to judge the error level, and then calling a corresponding prediction model according to the error level to carry out compensation pose prediction, hierarchical filtering and smoothing. And finally, merging the pose point data under the same error level according to the diagnosis result, and grouping and counting the positioning errors before and after compensation.
The statistics of the two robots compensating for the front and back positioning errors for each error level are shown in fig. 8. In the figure, the horizontal axis is different sub-workspaces distinguished by pose coordinates, the vertical axis represents positioning errors, orange and purple in a label respectively represent standard deviation and mean of the positioning errors of the robot before compensation, and red and green respectively represent standard deviation and mean of the positioning errors of the robot after compensation. The data in figure 8 (a) shows that after application of the proposed method,the average positioning error of each group of the working space of the robot is reduced to 0.13mm, and preferably, the average positioning error can reach 0.1mm; fig. 8 (b) shows that the average error of the UR robot after compensation is reduced to 0.16mm, preferably to 0.14mm.
In order to more intuitively show the compensation effect of the proposed method, the extreme pose points in the working space are connected into a space ring shape, and the target track, the actual track before compensation and the actual track after compensation of the two robots are shown in comparison, as shown in fig. 9. The result shows that the method improves the motion trail accuracy of the robot by reducing the absolute positioning error of the robot, and has important practical significance for ensuring and improving the processing accuracy of the industrial robot and promoting the application of the industrial robot to actual processing scenes.
The beneficial effects of the invention are as follows:
the invention discloses an industrial robot positioning error self-adaptive compensation method based on a composite branch neural network, in the method of the invention, the influence factors of two aspects of robot positioning error compensation space and time are abstracted into error grade diagnosis and compensation pose prediction problems, a neural structure search technology is introduced to design a neural structure search framework of a master-slave controller architecture, firstly, under the drive of a master controller, a diagnosis slave controller automatically creates an error grade according to given pose data with different error gradesAnd the diagnosis model is used for predicting the pose data of different error grades screened by the slave control according to the diagnosis model, automatically creating a compensation pose prediction model, and finally, integrating the diagnosis model and the prediction model by the master control to form a composite branch network model. In the application process, the main controller firstly activates a diagnosis network branch in the composite branch model to judge the positioning error grade under the current pose, then activates a prediction network branch under the corresponding error grade to generate a compensation pose, corrects the compensation pose by utilizing the error grade, and loads the corrected compensation pose into the robot controller to realize pose compensation. The method of the invention is applied toThe robot and UR robot accuracy compensation cases, the results show that the method will +.>The positioning error of the industrial robot is reduced from 0.884mm to 0.135mm, the positioning error of the UR robot is reduced from 2.11mm to 0.158mm, and the performance is superior to that of the similar method. />
Claims (5)
1. An industrial robot positioning error self-adaptive compensation method based on a composite branch neural network is characterized by comprising the following steps:
s1, setting an error grade by a master control, and automatically creating an error grade diagnosis model by a diagnosis slave control based on pose data;
s2, according to the error level set by the master control, the prediction slave control respectively creates a compensation pose prediction model for pose data of different levels;
s3, integrating the error level diagnosis model and the plurality of compensation pose prediction models into a composite branch network model by the main control, and adjusting the division of the error level according to the performance of the model, guiding the slave control to re-model until the compensation precision requirement is met;
s4, aiming at the robot pose data to be compensated, a diagnosis branch of the active model is controlled to judge the error grade of the current pose;
s5, the main control activates a corresponding prediction branch according to the judging result of the model diagnosis branch, outputs a compensation pose, carries out hierarchical filtering and smoothing on the prediction result by utilizing the diagnosis error level, and finally loads the prediction result into the controller to drive the tail end of the robot to move for work so as to realize self-adaptive compensation of positioning errors;
the main control is responsible for coordinating the training and optimization of the slave control, integrating the composite branch neural network model and activating the corresponding channels of the network to perform pose error level diagnosis and pose prediction compensation, is essentially a control logic, and is the brain of the proposed method;
the diagnosis slave control is a neural network model, an error grade diagnosis model is created by outputting a decision sequence, and a mapping relation between the pose and the error grade is established, so that the influence of difference distribution of the error grade in a robot working space on high-precision error compensation is overcome;
the prediction slave control is a neural network model, a compensation pose prediction model is created by outputting a decision sequence, a mapping relation between an actual pose and a target pose is established, and errors are compensated into the target pose in advance, so that the influence of the positioning errors on high-precision error compensation caused by the degradation of the machine humanization can be overcome;
in the step S5, the step of performing hierarchical filtering and smoothing on the prediction result by using the diagnosed error level refers to filtering the output of the prediction model by using the diagnosed threshold value of a certain level error, removing abnormal values, and ensuring the validity and stability of the prediction result of the model.
2. The adaptive compensation method for positioning errors of an industrial robot based on a composite branch neural network as set forth in claim 1, wherein the step S1 specifically includes:
s11, the main control demarcates error levels, and actual and target pose data sampled from a robot working space are classified according to the corresponding error levels;
s12, the diagnosis slave control takes the classified pose data as input, outputs a decision sequence, selects a corresponding network layer from the search space, and automatically creates an error level diagnosis model;
s13, according to the precision performance of the diagnosis model, the diagnosis slave control continuously adjusts the optimization decision sequence until an error grade diagnosis model meeting the precision requirement is automatically created;
the diagnosis slave control in the step S12 is a neural network model mixed in series and parallel, the neural network model is input as a null architecture of a diagnosis candidate model, and a decision sequence corresponding to the null architecture of the candidate model is output through calculation of an intermediate network layer;
the search space in step S12 is a network layer library including 8 layers of networks including a convolution layer, a full connection layer, a long-short-time memory network layer, a bidirectional long-short-time memory network layer, and the like of different super parameter combinations.
3. The method for adaptively compensating for the positioning error of the industrial robot based on the composite branch neural network according to claim 2, wherein the accuracy of the diagnostic model in the step S13 is represented as follows:
in which A d A measurement index indicating the accuracy of the diagnostic model, N indicating the number of pose points included in the batch data, K indicating the number of error levels set,for the ith component of the diagnostic model for the diagnostic result of the error class in which the jth pose point is located in the batch data +.>The ith component of the error level single-heat coded label corresponding to the jth pose point in the batch data;
the objective function of the diagnostic slave optimization decision sequence in step S13 is as follows:
wherein T is d An objective function of the diagnosis slave control is represented, M represents the number of decision rounds of the diagnosis slave control, T is the length of a decision sequence,for diagnosing the t-th dimensional decision value in the slave control jth round of decision sequence,/for diagnosis>The diagnostic precision of the diagnostic model created by the j-th round of decision, namely the rewarding value of the round of decision, is represented by using cross entropy to describe the distance between the strategy and the return probability distribution, when the two are closer, the cross entropy is smaller, namely when the accumulated rewarding of the decision sequence is larger, the probability of occurrence of the corresponding decision sequence is also larger, the cross entropy loss is smaller, under the given candidate model verification environment, the return corresponding to the same action is the same, and in the interaction process of the controller and the verification environment, the return corresponding to a certain decision sequence is found to be larger, so that the probability of occurrence of the corresponding action is increased by updating the network parameter of the controller through the cross entropy loss function.
4. The adaptive compensation method for positioning errors of an industrial robot based on a composite branch neural network as set forth in claim 1, wherein the step S2 specifically includes:
s21, the prediction slave control takes pose data under a certain error level as input, outputs a decision sequence, selects a corresponding network layer from a search space, and automatically creates a compensation pose prediction model;
s22, according to the precision performance of the prediction model, the prediction slave control continuously adjusts the optimization decision sequence until a compensation pose prediction model meeting the precision requirement is automatically created;
s23, repeating the process until a compensation pose prediction model meeting the precision requirement is created for pose data under various error grades;
the accuracy of the compensation pose prediction model in the step S22 is expressed as follows:
in the middle ofRepresenting predicted values of the compensation pose prediction model for the ith dimension coordinates of the jth pose point in the batch data, < +.>Target value representing the i-th dimensional coordinate, +.>The average value of the target poses is represented, and N represents the number of pose points contained in the batch; the front three-dimensional representation of the coordinates is displaced with an error weight omega 1 The rear three-dimensional representation of the gesture, weighted omega 2 Mu is the weight of the error grade, and A is the weight of the error grade which becomes larger along with the increase of the error grade p The closer to 1, the higher the prediction accuracy of the compensation pose prediction model is, and the worse the prediction reliability is.
5. The adaptive compensation method for positioning errors of an industrial robot based on a complex branched neural network according to claim 1, wherein the representation of the complex branched network model in step S3 is represented as:
wherein A is d For the purpose of diagnosing the accuracy of the model,for the average precision of a plurality of compensation pose prediction models, taking the influence of the diagnosis precision and the prediction precision into consideration, a coefficient alpha and beta comprehensive evaluation composite branch network model is set, if coarseThe granularity is divided into error grades, so that the diagnosis difficulty is reduced, but the prediction difficulty is increased; if the error grade is too thin, the diagnosis difficulty can rise, and although the prediction difficulty is reduced, the part of the pose of the diagnosis deviation can call an incorrect prediction model to compensate, and the final prediction precision can be affected.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211368818.XA CN115847396B (en) | 2022-11-03 | 2022-11-03 | Industrial robot positioning error self-adaptive compensation method based on composite branch neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211368818.XA CN115847396B (en) | 2022-11-03 | 2022-11-03 | Industrial robot positioning error self-adaptive compensation method based on composite branch neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115847396A CN115847396A (en) | 2023-03-28 |
| CN115847396B true CN115847396B (en) | 2023-07-28 |
Family
ID=85662380
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211368818.XA Active CN115847396B (en) | 2022-11-03 | 2022-11-03 | Industrial robot positioning error self-adaptive compensation method based on composite branch neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115847396B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106003034A (en) * | 2016-06-16 | 2016-10-12 | 深圳先进技术研究院 | Master-slave robot control system and control method |
| CN110385720A (en) * | 2019-07-26 | 2019-10-29 | 南京航空航天大学 | A kind of robot localization error compensating method based on deep neural network |
| CN114800529A (en) * | 2022-06-07 | 2022-07-29 | 北京航空航天大学 | Industrial robot positioning error online compensation method based on fixed-length memory window incremental learning and incremental model reconstruction |
| CN114918920A (en) * | 2022-06-01 | 2022-08-19 | 浙江大学 | Industrial robot calibration method based on neural network and distance error model |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11772264B2 (en) * | 2020-11-18 | 2023-10-03 | Dibi (Chongqing) Intelligent Technology Research Institute Co., Ltd. | Neural network adaptive tracking control method for joint robots |
-
2022
- 2022-11-03 CN CN202211368818.XA patent/CN115847396B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106003034A (en) * | 2016-06-16 | 2016-10-12 | 深圳先进技术研究院 | Master-slave robot control system and control method |
| CN110385720A (en) * | 2019-07-26 | 2019-10-29 | 南京航空航天大学 | A kind of robot localization error compensating method based on deep neural network |
| CN114918920A (en) * | 2022-06-01 | 2022-08-19 | 浙江大学 | Industrial robot calibration method based on neural network and distance error model |
| CN114800529A (en) * | 2022-06-07 | 2022-07-29 | 北京航空航天大学 | Industrial robot positioning error online compensation method based on fixed-length memory window incremental learning and incremental model reconstruction |
Non-Patent Citations (1)
| Title |
|---|
| KUKA工业机器人位姿测量与在线误差补偿;史晓佳;张福民;曲兴华;刘柏灵;王俊龙;;机械工程学报(08);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115847396A (en) | 2023-03-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111590581B (en) | Positioning compensation method and device for robot | |
| CN111468989B (en) | Five-axis linkage numerical control manipulator polishing control system and method | |
| CN114237155B (en) | Error prediction and compensation method, system and medium for multi-axis numerical control machining | |
| CN109472057B (en) | Product processing quality prediction device and method based on cross-process implicit parameter memory | |
| CN109146209A (en) | Machine tool spindle thermal error prediction technique based on wavelet neural networks of genetic algorithm | |
| CN112526927B (en) | Quick optimization compensation method for space positioning error of rotating shaft of five-axis numerical control machine tool | |
| CN112558547B (en) | Quick optimization method for geometric error compensation data of translational shaft of five-axis numerical control machine tool | |
| WO2005049284A1 (en) | Method for optimising the performance of a robot | |
| CN105205221B (en) | A kind of heavy digital control machine tool precision reliability analysis method | |
| CN111752151A (en) | Adaptive tracking and compensating method and system for grinding and polishing industrial blade | |
| CN114131611B (en) | Off-line compensation method, system and terminal for joint errors of robot gravity pose decomposition | |
| CN114800529B (en) | Industrial robot error compensation method based on fixed-length memory window increment learning | |
| US11726448B1 (en) | Robotic workspace layout planning | |
| Gorkavyy et al. | Intelligent system for prognostication and optimization of power expenses of technological processes at robotized productions | |
| CN107511823B (en) | The method of robot manipulating task track optimizing analysis | |
| Wegener et al. | Dawn of new machining concepts:: Compensated, intelligent, bioinspired | |
| CN113910001B (en) | Numerical control machine tool space error identification method | |
| CN113910218A (en) | Robot calibration method and device based on kinematics and deep neural network fusion | |
| CN112651080A (en) | Industrial AI technology-based welded structural part process optimization method and system | |
| CN115098978A (en) | RBF neural network-based forward kinematics analysis method for improving Newton iterative algorithm | |
| CN115847396B (en) | Industrial robot positioning error self-adaptive compensation method based on composite branch neural network | |
| Zhou et al. | Adaptive hierarchical positioning error compensation for long-term service of industrial robots based on incremental learning with fixed-length memory window and incremental model reconstruction | |
| CN112507480B (en) | Inertial parameter identification method | |
| CN116383997B (en) | Digital twinning-based large-scale numerical control planer milling machine machining precision prediction method | |
| CN115391936A (en) | Parameter dynamic multi-objective optimization method in robot milling process |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |