Industrial robot absolute positioning precision compensation method and device

Abstract

The invention provides an industrial robot absolute positioning precision compensation method and device. The method provided by the invention comprises the steps that multiple targets trained by a precision compensation model are determined, and the multiple targets at least comprise a sub-target of a training dimension and a sub-target of a compensation precision dimension; determining a to-be-trained original parameter set of the deep belief network; calculating the influence degree of each parameter in the original parameter set on multiple targets; screening parameters of which the influence degrees are higher than a preset threshold value as candidate hyper-parameters; acquiring current pose data of a plurality of sampling points of the robot based on a laser tracker; generating a position error based on predetermined expected pose data and current pose data of the robot, training a deep belief network to obtain a trained target deep belief network, and calculating a contribution weight of each feature vector based on an evidence theory; and generating a compensation result. The industrial robot absolute positioning precision compensation method provided by the invention is used for improving the positioning precision of the robot during operation.

Description

Technical Field

[0001] This application relates to the field of robot manufacturing technology, and in particular to a method and apparatus for compensating the absolute positioning accuracy of an industrial robot. Background Technology

[0002] With the rapid development of the manufacturing industry, industrial robots are playing an increasingly important role in automated production and assembly due to their high efficiency, stability, and flexibility. However, due to the complexity of industrial robot systems, their absolute positioning accuracy is often affected by various factors, making it difficult to meet the requirements of high-precision tasks in practical applications.

[0003] To improve the absolute positioning accuracy of industrial robots, a common existing method is traditional manual calibration, which relies on the experience and skills of professional technicians to manually adjust the robot to correct its position and posture errors. However, the accuracy and consistency of manual calibration are difficult to guarantee, affecting the robot's stability and reliability. Another method based on neural networks uses experimental semivariograms to fit the robot's joint angles and corresponding position errors, and then uses the semivariogram parameters and RBF neural networks to modify the target position. However, this method has low accuracy and large errors in model processing. Therefore, there is an urgent need for a more efficient and accurate absolute positioning accuracy compensation technology for industrial robots to overcome the shortcomings of existing technologies. Summary of the Invention

[0004] In view of this, this application provides an absolute positioning accuracy compensation method and apparatus for industrial robots to improve the positioning accuracy of robots during operation.

[0005] Specifically, this application is implemented through the following technical solution:

[0006] The first aspect of this application provides a method for compensating the absolute positioning accuracy of an industrial robot, the method comprising:

[0007] Determine the multiple objectives for training the accuracy compensation model, wherein the multiple objectives include at least sub-objectives of the training dimension and sub-objectives of the accuracy compensation dimension;

[0008] Determine the original set of parameters to be trained in the deep belief network;

[0009] Calculate the degree of influence of each parameter in the original parameter set on the multi-objective;

[0010] Parameters whose influence exceeds a preset threshold are selected as candidate hyperparameters;

[0011] The current pose data of the robot at multiple sampling points is obtained using a laser tracker;

[0012] Based on the predetermined desired pose data and the current pose data of the robot, a position error is generated. The desired pose data is input into the deep belief network, and the position error is used as the output of the deep belief network to train the deep belief network.

[0013] While training the deep belief network, the candidate hyperparameters of the deep belief network are optimized based on a genetic algorithm to obtain the trained target deep belief network.

[0014] Acquire the data to be compensated during robot operation;

[0015] The target deep belief network is used to extract feature vectors of the data to be compensated, and the contribution weight of each feature vector is calculated based on evidence theory.

[0016] The feature vector is weighted based on the contribution weights to generate a compensation result.

[0017] A second aspect of this application provides an absolute positioning accuracy compensation device for an industrial robot, the device comprising an acquisition module, a training module, a calculation module, and a processing module; wherein...

[0018] The processing module is used to determine the multiple objectives of training the accuracy compensation model, wherein the multiple objectives include at least the sub-objectives of the training dimension and the sub-objectives of the accuracy compensation dimension.

[0019] The processing module is also used to determine the original parameter set of the deep belief network to be trained;

[0020] The calculation module is used to calculate the degree of influence of each parameter in the original parameter set on the multi-objective;

[0021] The processing module is also used to filter parameters whose influence is higher than a preset threshold as candidate hyperparameters;

[0022] The acquisition module is used to acquire the current pose data of multiple sampling points of the robot based on the laser tracker;

[0023] The training module is used to generate a position error based on the predetermined desired pose data and the current pose data of the robot, input the desired pose data into the depth belief network, and use the position error as the output of the depth belief network to train the depth belief network.

[0024] The training module is also used to optimize the candidate hyperparameters of the deep belief network based on a genetic algorithm while training the deep belief network, so as to obtain the trained target deep belief network.

[0025] The acquisition module is also used to acquire the data to be compensated during robot operation;

[0026] The calculation module is also used to extract feature vectors of the data to be compensated using the target deep belief network, and to calculate the contribution weight of each feature vector based on evidence theory.

[0027] The processing module is further configured to weight the feature vector based on the contribution weight to generate a compensation result.

[0028] The industrial robot absolute positioning accuracy compensation method and device provided in this application start from the robot's control results, that is, from the coordinate position after movement, ignoring the fluctuations of internal forces generated in the intermediate process. This effectively reduces the computational complexity and ensures the accuracy of the calculation. Firstly, by combining deep belief networks (DBNs) with genetic algorithms and applying evidence theory, the DBN uses a multi-layer restricted Boltzmann machine to extract deep features from the input data. The genetic algorithm further optimizes the hyperparameters of the DBN, ensuring optimal network model performance. Combined with evidence theory to assess prediction uncertainty, the organic integration of these three elements in different directions improves the accuracy and stability of the network output. Secondly, when optimizing the DBN, the genetic algorithm not only globally searches for the optimal hyperparameter combination but also uses a fitness function to screen individuals, ensuring that only high-performing individuals enter the next generation. By screening out the key factors that have the greatest impact on model performance, the genetic algorithm effectively avoids local optima, enabling the DBN to better adapt to complex industrial environments and improve the accuracy of prediction and compensation. This optimization strategy significantly reduces the complexity of model training while improving the efficiency and accuracy of compensation control. Thirdly, the application of evidence theory in deep belief networks goes beyond simply correcting the final prediction result; it penetrates the intermediate processes of the network. By precisely calculating the contribution weights of each feature vector, the network can more accurately capture key information and ignore irrelevant noise during feature extraction and pattern recognition. This intervention method actually improves the accuracy and robustness of the network itself, enabling deep belief networks to maintain stable prediction performance even in complex and ever-changing task environments. Thus, by training deep belief networks with expected pose data and position errors, the deep belief network learns the complex mapping relationship between robot pose and position error from the data, enabling accurate prediction. The robot's position error under arbitrary poses provides a reliable predictive model for subsequent compensation. By optimizing the hyperparameters of the deep belief network using a genetic algorithm, the mechanisms of natural selection and genetics are simulated. The optimal parameter combination is searched globally, effectively avoiding the problem of local optima and enabling more accurate prediction of the robot's position error under different poses. Based on evidence theory, the contribution weight of each feature vector is calculated. Evidence theory can handle uncertain information and assign reasonable weights to each feature vector, reflecting their importance in the compensation process. This ensures that each feature vector can reasonably influence the compensation result, guaranteeing the accuracy of positioning. In this way, through the organic combination of deep belief network, genetic algorithm, and evidence theory, accurate compensation results are obtained, achieving an improvement in the absolute positioning accuracy of industrial robots. Attached Figure Description

[0029] Figure 1A flowchart of Embodiment 1 of the industrial robot absolute positioning accuracy compensation method provided in this application;

[0030] Figure 2 This is a schematic diagram illustrating an exemplary embodiment of the present application of the absolute positioning accuracy compensation process for an industrial robot.

[0031] Figure 3 A schematic diagram illustrating a communication control mode as shown in an exemplary embodiment of this application;

[0032] Figure 4 This is a schematic diagram illustrating the partitioning of a dataset as shown in an exemplary embodiment of this application;

[0033] Figure 5 This is a schematic diagram illustrating dataset partitioning during individual training of a deep belief network, as shown in an exemplary embodiment of this application.

[0034] Figure 6 A schematic diagram illustrating a fitness function curve for an exemplary embodiment of this application;

[0035] Figure 7 An error prediction result diagram shown as an exemplary embodiment of this application;

[0036] Figure 8 An error prediction result diagram shown as another exemplary embodiment of this application;

[0037] Figure 9 An error prediction result diagram is shown as another exemplary embodiment of this application;

[0038] Figure 10 This is a schematic diagram illustrating the x-direction comparison in an exemplary embodiment of this application;

[0039] Figure 11 A schematic diagram illustrating a comparison in the y-direction for another exemplary embodiment of this application;

[0040] Figure 12 A schematic diagram illustrating a comparison in the z-direction, as shown in yet another exemplary embodiment of this application;

[0041] Figure 13 This is a schematic diagram illustrating the compensation effect of an exemplary embodiment of this application;

[0042] Figure 14 A schematic diagram of an industrial robot absolute positioning accuracy compensation device provided for an exemplary embodiment of this application. Detailed Implementation

[0043] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0044] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0045] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0046] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0047] Figure 1 This is a flowchart of an embodiment of the industrial robot absolute positioning accuracy compensation method provided in this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0048] S101. Determine the multiple objectives for training the accuracy compensation model, wherein the multiple objectives include at least the sub-objectives of the training dimension and the sub-objectives of the accuracy compensation dimension.

[0049] Specifically, the accuracy compensation model is a mathematical model used to improve the positioning accuracy of industrial robots. By analyzing and compensating for the actual positioning error of the robot through the accuracy compensation model, the positioning error can be reduced and the positioning accuracy can be improved.

[0050] Furthermore, the sub-objectives of the training dimension focus on accuracy compensation for the model training process itself. Through the sub-objectives of the training dimension, the improvement of the model's learning ability and the efficiency of the optimization process can be characterized.

[0051] Furthermore, the compensation accuracy dimension sub-targets are directly correlated with the accuracy compensation model to compensate for the positioning error of industrial robots.

[0052] In practical implementation, the specific application scenarios and processing accuracy requirements of the industrial robot are clearly defined, and a first set of candidate sub-objectives for the compensation accuracy dimension is determined accordingly. For example, in one embodiment, if the robot is used for precision assembly, it may be necessary to include "maximum positioning error not exceeding 0.1mm" as a sub-objective. Then, considering factors such as model training time and computational resource consumption, a second set of candidate sub-objectives for the training dimension is determined. For example, in one embodiment, sub-objectives such as "model training time not exceeding 24 hours" or "using no more than 10GB of GPU memory" can be set. Then, for each sub-objective in the first and second objective configurations, potential conflicts between them are analyzed, and based on the degree of repulsion between sub-objectives, multi-objective optimization methods (such as weighted summation, Pareto front, etc.) are used to select the optimal combination of sub-objectives, so that the model can reach an acceptable level in multiple dimensions such as compensation accuracy and training efficiency. Furthermore, based on actual needs and technical feasibility, the final multi-objective set is further expanded. Finally, based on the selected optimal combination of sub-objectives, the accuracy compensation model is trained and optimized to achieve the predetermined multi-objective requirements.

[0053] Furthermore, the accuracy compensation model can be defined as optimizing the training dimensions of 6 dimensions and the compensation accuracy dimensions of 9 parameters in these 6 dimensions, to ensure the optimization effect.

[0054] S102. Determine the original parameter set to be trained in the deep belief network.

[0055] Specifically, deep belief networks are composed of multiple restricted Boltzmann machines stacked together to extract deep features from data.

[0056] In a specific implementation, for example, in one embodiment, each layer of the deep belief network has its corresponding parameters, so the original parameter set to be trained may include 20 dimensions, and each dimension has 8 corresponding features.

[0057] S103. Calculate the degree of influence of each parameter in the original parameter set on the multi-objective.

[0058] In practical implementation, parameters in the original parameter set can be clustered based on their network layer, function type (such as weights or biases), and other attributes. For example, in one embodiment, all weights belonging to the connection between the first hidden layer and the input layer can be clustered into one category, all biases belonging to the first hidden layer into another, and so on. Then, for each dimension of parameters, one or more representative parameters can be selected to evaluate the impact of that category of parameters on multi-objectives. For example, in one embodiment, other parameters can be kept constant, and only the size of a certain category of weights can be adjusted, then the changes in compensation accuracy and training efficiency of the model can be evaluated. Then, it can be determined which parameter categories have a higher degree of influence on multi-objectives, and these parameters in the subsequent optimization process should be given more attention. Then, for parameter categories with a high degree of influence, the specific impact of each parameter in that category on multi-objectives can be analyzed in detail. For example, in one embodiment, the value of each parameter can be adjusted one by one and the changes in model performance can be observed. Finally, the parameters are sorted according to their degree of influence on multi-objectives, and the parameters with the greatest influence are selected for optimization. For example, in one embodiment, during the optimization process, optimization methods such as genetic algorithms and gradient descent can be used to continuously adjust the values ​​of these parameters to improve the overall performance of the model.

[0059] S104. Select parameters whose influence exceeds a preset threshold as candidate hyperparameters.

[0060] Specifically, the specific value of the preset threshold is set according to actual needs, and this embodiment does not limit it.

[0061] In practice, the collected parameter influence data can be compared with a preset threshold. If the influence of a parameter is higher than the threshold, it is considered a candidate hyperparameter; otherwise, it is excluded. The final candidate hyperparameters are shown in Table 1, which is a candidate hyperparameter table provided by an exemplary embodiment of this application.

[0062] Table 1

[0063] S105. Obtain the current pose data of multiple sampling points of the robot based on the laser tracker.

[0064] Specifically, the exact location of the sampling point is determined according to actual needs, and this embodiment does not impose any limitations on it.

[0065] In practice, sampling points can be evenly distributed within the robot's workspace to fully reflect the robot's performance under different positions and postures. For example, in one embodiment, the robot's robotic arm and joints can be designated as sampling points.

[0066] Furthermore, the laser tracker can be the Leica AT901-B, with an error of ±15μm+6μm / m.

[0067] It should be noted that the error of the laser tracker increases with the distance. For example, in one embodiment, combined with the above embodiment, the AT901-B uses an angle encoder to measure the angle and an absolute interferometer to measure the distance. The absolute interferometer in the AT901 integrates a helium-neon laser interferometer and an absolute rangefinder. These two lasers can work independently. The laser beam emitted by the laser is directed towards the target through a gimbaled mirror. The interferometer laser beam also serves as the line of sight of the tracker. The reflected laser is measured by the dual-axis position detector built into the tracker. The pulse generated by the position detector is processed by the tracker processor and fed back to the tracker's servo motor, driving the motor to track the target mirror in real time, ultimately achieving tracking and distance measurement, and obtaining the actual pose of the robot's end effector.

[0068] As an optional embodiment, before step S101, the method further includes: determining the plurality of sampling points, which includes: establishing a spatial environment model of the robot; determining the spatial working trajectory of the robot in the spatial environment model; determining the change curve of the change rate of the spatial working trajectory; and determining a plurality of sampling points based on the amplitude change of the change curve, wherein the number of sampling points at positions with large amplitude is less than the number of sampling points at positions with small amplitude. Specifically, during robot processing, if the processing parameters change frequently, the frequency of robot position compensation calculation is relatively high. If the robot processing parameters remain unchanged, the position compensation can use the original calculated value to remain constant. By adopting a variable-period sampling method, unnecessary resampling can be reduced in response to the actual changes in robot processing, thus improving sampling efficiency. At the same time, the number of sampling points is increased in areas of frequent change, thereby improving sampling accuracy. Specifically, within the range where the amplitude change is less than a preset threshold, the target position in the change curve is determined with a fixed period, and the target sampling position on the spatial working trajectory is determined based on the value of the change rate of the target position. Within the range where the amplitude change is greater than a preset threshold, the product of the amplitude change rate and the fixed period is used as the corrected adaptive variable period value, and the target sampling position on the spatial working trajectory is determined according to the adaptive variable period value.

[0069] S106. Based on the predetermined desired pose data of the robot and the current pose data, generate a position error, input the desired pose data into the depth belief network, use the position error as the output of the depth belief network, and train the depth belief network.

[0070] Specifically, the desired pose data is preset by the industrial robot's control system or programming environment, representing the ideal position and orientation that the robot should reach at each sampling point.

[0071] In practice, the expected pose data can include the same data type as the current pose data, and can include the robot's X, Y, and Z coordinates in the workspace as well as rotation angles (such as pitch, yaw, and roll) around these three axes.

[0072] Furthermore, the input layer of the deep belief network has nine channels, which are the theoretical position coordinates of the robot and the corresponding angles of its six joints. The output layer of the deep belief network has three channels, representing the robot's position error. .

[0073] Furthermore, the calculation determined that the maximum number of hidden layers in the deep belief network (DBN) is 4, with the number of nodes in each hidden layer ranging from 10 to 101. The initial learning rate is 0.01, the momentum factor is 0.8, the activation function is Sigmoid, and the mean squared error (MSE) is used as the loss function. The hyperparameters that need to be optimized are the number of hidden layers, the number of hidden nodes, the learning rate, the momentum factor, the number of RBM iterations, and the number of DBN fine-tuning iterations.

[0074] Furthermore, the desired pose data is compared with the current pose data to calculate the robot's position error at each sampling point.

[0075] In a specific implementation, for example, in one embodiment, the position error can be characterized by the vector difference between the desired position and the actual position: that is, the position error includes three translation errors (ΔX, ΔY, ΔZ) and three rotation errors (ΔRoll, ΔPitch, ΔYaw).

[0076] Furthermore, the calculated position error is used as the output target of the deep belief network, while the expected pose data is used as the input feature of the network. The deep belief network is trained using these two data.

[0077] The following is a specific example to illustrate the training process of a deep belief network:

[0078] (1) Adjust the parameters of the deep belief network based on the layer-by-layer greedy algorithm to obtain the initialized model parameters; wherein the training process of the layer-by-layer greedy algorithm is unsupervised pre-training.

[0079] Specifically, a greedy algorithm is a training method that builds a model step by step by obtaining the current optimal parameters at each step, thus gradually reaching the global optimum through local optimum selection.

[0080] In practice, the input data can be used as the input to the visible layer, and the weights and biases can be adjusted by unsupervised learning algorithms such as contrastive divergence, so that the deep belief network can learn an effective representation of the data.

[0081] The following is a specific example to illustrate the process of initializing model parameters:

[0082] Step 1: Process the desired pose data based on the visible layer of the first restricted Boltzmann machine, and output the target corresponding data of the desired pose data through the hidden layer of the restricted Boltzmann machine.

[0083] Specifically, in a deep belief network, the visible layer of a restricted Boltzmann machine (RBM) is used to receive external input data, and there can be multiple RBMs. In practice, the desired pose data is input into the visible layer of the first RBM.

[0084] Furthermore, the Boltzmann machine is restricted to the visible layer. and hidden layers This consists of two layers. The visible layer receives training data, specifically the desired end effector position and current robot joint angles in the robot coordinate system. The hidden layer takes the output of the visible layer as its input and is used to extract features. There is no intra-layer connection but full inter-layer connection between the neurons in the two layers.

[0085] Step 2: Adjust the weights and biases of the visible layer based on the data feedback corresponding to the target to obtain the trained first layer of the restricted Boltzmann machine.

[0086] Specifically, each restricted Boltzmann machine contains a visible layer and a hidden layer. The input data is processed through the visible layer, and the output is completed through the hidden layer.

[0087] In practice, the desired pose data is processed by a Boltzmann machine, and the reconstructed data is output based on the activation state of the hidden layer. The reconstruction error is calculated by comparing the difference between the reconstructed data and the original input data.

[0088] Furthermore, based on the reconstruction error, algorithms such as contrastive divergence are used to update the weights and biases between the visible and hidden layers.

[0089] In practice, the target data is passed to the hidden layer. The state of the hidden layer neurons is sampled according to their activation probabilities. Then, the sampled hidden layer states are used to reconstruct the state of the visible layer using weights and biases. For each visible layer neuron, its reconstruction probability is calculated. The original visible layer input and the reconstructed visible layer output are compared, and the difference between them is calculated. Gradient descent or other optimization algorithms are used to update the weights and biases based on the reconstruction error.

[0090] Understandably, by minimizing the reconstruction error, the restricted Boltzmann machine can better capture the intrinsic structure of the data and determine accurate weights and biases.

[0091] Furthermore, step two can be repeated multiple times until the Boltzmann machine training converges or a predetermined number of iterations is reached, in order to ensure that the weights and biases are gradually adjusted to the optimal state.

[0092] Step 3: Process the desired pose data based on the trained first-layer restricted Boltzmann machine, and use the output data as the input of the second-layer restricted Boltzmann machine, repeating the steps of acquiring the target corresponding data.

[0093] In practice, the hidden layer activation states of the trained first-layer Restricted Boltzmann Machine are used as the visible layer inputs of the second-layer Restricted Boltzmann Machine to achieve data transfer layer by layer.

[0094] Furthermore, the training process in steps one and two is repeated for the second-layer Restricted Boltzmann Machine. The hidden layer output of the second-layer Restricted Boltzmann Machine is used as the visible layer input of the next-layer Restricted Boltzmann Machine, and the training process continues.

[0095] Step 4: After the number of layers of the trained restricted Boltzmann machine reaches the preset number, confirm that the model parameter initialization is complete.

[0096] Specifically, the exact number of preset layers is set according to actual needs, and this embodiment does not limit it. In practice, the number of layers can be determined based on the complexity of robot positioning.

[0097] The industrial robot absolute positioning accuracy compensation method provided in this embodiment uses algorithms such as divergence to iteratively adjust the weights and biases between the visible and hidden layers based on the reconstruction error from the hidden layer to the visible layer. This unsupervised training process can minimize the reconstruction error, enabling the restricted Boltzmann machine to better capture the intrinsic structure of the data. At the same time, by stacking restricted Boltzmann machines layer by layer to construct a deep neural network, it is able to handle more complex data relationships and patterns, ensuring the reliability and accuracy of the deep belief network.

[0098] (2) Fine-tuning the parameters of the deep belief network based on the BP algorithm to complete the parameter adjustment of the deep belief network; wherein, the fine-tuning process of the BP algorithm is supervised fine-tuning.

[0099] Specifically, the BP algorithm is the Backpropagation Algorithm, which is used to calculate the difference between the network output and the expected output, and then backpropagate the difference back to the network to update the weights in the network, so that the network output gets closer and closer to the expected output.

[0100] Furthermore, before fine-tuning the BP algorithm, a labeled dataset needs to be prepared. The labeled dataset can contain the expected pose data of the industrial robot and the corresponding actual position error (where the actual position error is used as the label).

[0101] The following is a specific example to illustrate the fine-tuning process of the BP algorithm:

[0102] Step 1: Input the desired pose data into the deep belief network to obtain the output prediction result.

[0103] In practice, the desired pose data is input into a deep belief network, which processes the data layer by layer to obtain the robot's possible pose data as the prediction result.

[0104] Step 2: Calculate the loss function based on the prediction result and the position error, and obtain the gradient parameters of each gradient of the loss function; each gradient corresponds to one layer of the restricted Boltzmann machine.

[0105] In practice, the position error is determined based on the label corresponding to the desired pose data. The position error is then compared with the prediction result to calculate the loss function. For example, in one embodiment, the average of the sum of squares of the differences between the position error and the prediction result can be used as the loss function.

[0106] Furthermore, each layer of the Restricted Boltzmann Machine corresponds to a layer of gradients. The chain rule can be used to calculate the gradient of the loss function with respect to the weights of each Restricted Boltzmann Machine layer, starting from the output layer. These gradient parameters represent the direction and magnitude of the weight adjustment of each Restricted Boltzmann Machine layer, thereby updating the network parameters.

[0107] It is understandable that since each layer of the Restricted Boltzmann Machine corresponds to a gradient, each gradient corresponds to a layer of the Restricted Boltzmann Machine.

[0108] Step 3: Backpropagate the gradient parameters in the deep belief network to fine-tune the parameters of the deep belief network.

[0109] In practice, the calculated gradient parameters are backpropagated back through the network layer by layer, starting from the output layer. During backpropagation, each Restricted Boltzmann Machine adjusts its weights based on the received gradient parameters to reduce prediction errors.

[0110] The industrial robot absolute positioning accuracy compensation method provided in this embodiment compares the model's prediction results with the actual position error, and then uses the chain rule to differentiate the loss function. Since each gradient corresponds to a restricted Boltzmann machine layer in the deep belief network, the restricted Boltzmann machine that generates the error can be accurately located, providing precise guidance for subsequent parameter fine-tuning and ensuring the accuracy of model parameter adjustment.

[0111] S107. Optimize the hyperparameters of the deep belief network based on the genetic algorithm to obtain the trained target deep belief network.

[0112] Specifically, genetic algorithms search for the optimal solution to a problem by simulating phenomena such as selection, inheritance, mutation, and crossover.

[0113] Furthermore, the performance of the deep belief network is measured using a genetic algorithm to obtain the corresponding fitness value, and the hyperparameters of the deep belief network are adjusted using the fitness value.

[0114] The following is a specific embodiment to illustrate in detail the process of optimizing the hyperparameters of the deep belief network based on a genetic algorithm:

[0115] (1) Generate optimized cataloging and initialization configuration for the hyperparameters of the deep belief network.

[0116] In practice, an optimized catalog is generated based on the range of hyperparameters, and the optimized catalog forms the basis of the genetic algorithm's search space.

[0117] Furthermore, the initialization configuration comprises multiple individuals, each representing a combination of hyperparameters. In practice, optimized cataloging and initialization configurations can be generated randomly.

[0118] (2) Calculate the mean square error between the output bias of the deep belief network and the actual bias to obtain the fitness function of the genetic algorithm.

[0119] Specifically, the deep belief network makes predictions based on optimized cataloging and initial configuration to obtain the predicted robot position. It then calculates the mean square error of the deviation between the predicted position and the actual position, and uses the mean square error as the fitness function.

[0120] The following is a specific example to illustrate the process of obtaining the mean squared error:

[0121] Step 1: Input the desired pose data into the deep belief network to obtain the prediction output.

[0122] In practice, the desired pose data is input into the deep belief network, and the deep belief network outputs the corresponding prediction results.

[0123] Step 2: Calculate the square of the difference between the predicted output and the position error to obtain the target median value.

[0124] Furthermore, the difference between the predicted output and the position error is calculated, and the square of the difference is used to obtain the target median value.

[0125] In practice, the target intermediate value can be calculated based on the following formula:

[0126] E= ;

[0127] Where E is the target median value. To predict the output results, This represents the positional error.

[0128] Step 3: Calculate the mean square error based on the sum of squares of the target intermediate values, the output dimension of the deep belief network, and the number of data contained in the expected pose data.

[0129] In practice, the mean square error can be calculated based on the following formula.

[0130] ;

[0131] Where N is the number of data points, and m is the output dimension of the deep belief network. It is the predicted output result. It is a positional error. It is the mean squared error.

[0132] (3) Sort the initial configurations based on the fitness function and select a portion as the first target configuration.

[0133] Specifically, the initial configurations are sorted according to the fitness function values, and individuals with higher fitness are selected as the first target configuration.

[0134] It is understandable that individuals corresponding to the first target configuration perform better in the genetic process and have higher genetic value.

[0135] (4) Perform cross-operation and compilation operation on the first target configuration to obtain a set of second target configurations, and update the initial configuration based on the second target configurations to obtain the updated configuration.

[0136] Specifically, crossover operation involves randomly selecting two primary target configurations and generating new individuals by swapping some of their hyperparameter values.

[0137] Furthermore, mutation operations are performed on the offspring individuals obtained after crossover, that is, one or more hyperparameter values ​​of the offspring individuals are randomly changed with a certain probability.

[0138] Furthermore, the second target configuration obtained after crossover and mutation operations replaces some individuals in the initial configuration to form an updated configuration, which is the new hyperparameter value.

[0139] (5) Replace the initialization configuration with the updated configuration, re-execute the step of obtaining the updated configuration a preset number of times, and optimize the hyperparameters of the deep belief network based on the obtained updated configuration.

[0140] In practice, the hyperparameter configuration of the deep belief network is updated to the new configuration. The process of updating the hyperparameters of the deep belief network using the new configuration is repeated multiple times until the hyperparameter values ​​are stable.

[0141] The industrial robot absolute positioning accuracy compensation method provided in this embodiment continuously updates the individuals in the population, so that the algorithm can maintain the excellent genes of the population and gradually eliminate individuals with poor performance, thereby continuously improving the overall performance of the population during the iteration process and ensuring the robustness and accuracy of the deep belief network.

[0142] S108. Obtain the data to be compensated during robot operation.

[0143] In practice, the data to be compensated can be obtained from the training data used during the training of the deep belief network.

[0144] Furthermore, the specific data type of the data to be compensated is determined according to actual needs, and this embodiment does not limit it. For example, in one embodiment, the data to be compensated may include the position coordinates of each robotic arm of the robot and the joint angles of each joint of the robot, etc.

[0145] S109. Use the target deep belief network to extract features of the data to be compensated to obtain feature vectors, and calculate the contribution weight of each feature vector based on evidence theory.

[0146] In practice, the data to be compensated is input into the target deep belief network, and feature vectors are extracted based on the restricted Boltzmann machine.

[0147] Furthermore, evidence theory is a mathematical tool for dealing with uncertainty, which can effectively integrate information from multiple sources and assess its uncertainty.

[0148] It should be noted that the evidence theory involves obtaining the eigenvectors calculated from each layer of the restricted Boltzmann machine and calculating their contribution weights.

[0149] The following is a specific example to illustrate the process of generating contribution weights:

[0150] (1) Based on the evidence theory, the feature vector is processed to obtain multiple basic probability assignment functions corresponding to each feature vector.

[0151] Specifically, the basic probability assignment function reflects the degree to which the feature vector supports the prediction result.

[0152] In practice, evidence theory is used to process feature vectors, treating each feature vector as a source of evidence. Based on the value of the feature vector and its performance in the deep belief network, one or more basic probability assignment functions are assigned to it.

[0153] (2) For a given feature vector, fuse all the corresponding basic probability assignment functions.

[0154] In practice, the Dempster synthesis rule can be applied to consider the conflict and consistency between different basic probability assignment functions to obtain the fusion result corresponding to the feature vector.

[0155] (3) Generate the contribution weight based on the trust interval and the fused basic probability assignment function.

[0156] Specifically, the confidence interval is used to quantify the range of proposition confidence. The confidence interval represents the minimum and maximum confidence estimates for a correctly calculated proposition.

[0157] Furthermore, the trust interval can be a closed interval composed of the trust function (Bel) and the likelihood function (Pl), used to represent the likelihood of an event. level of support Bel(A) is the sum of the basic probability assignments of all subsets of A, representing the degree of trust in A, and satisfies the formula:

[0158] ;

[0159] Where Pl(A) is the sum of the basic probability assignments of all subsets intersecting with A, representing the degree of non-denial of A, and satisfies the formula:

[0160] ;

[0161] Where Pl(A) is the sum of the basic probability assignments of all subsets intersecting with A, representing the degree to which A is not denied.

[0162] Furthermore, let there be a finite nonempty set. To identify the frame, the function To identify the frame The basic probability assignment function.

[0163] It should be noted that Bel and Pl are used to represent the upper and lower bounds of the reliability of the positioning accuracy of industrial robots.

[0164] In practice, after fusing the basic probability assignment functions of all feature vectors, the contribution weight of each feature vector is generated based on the trust interval and the fused basic probability assignment functions.

[0165] Furthermore, the contribution weight can be calculated based on the distribution of the fused basic probability assignment function within the trust interval. The larger the weight value, the greater the contribution of the feature vector to the prediction result.

[0166] Furthermore, assuming These are the n feature vectors of the input sample, within the recognition framework. Above, each feature It can be viewed as evidence for a subset within the corresponding identification framework, for each category. This can be considered evidence. Either support Either support The specific support depends on the evidence. Weighting coefficients:

[0167] ;

[0168] in, and There are two weight parameters.

[0169] Furthermore, let's assume and The weights of evidence are respectively The positive and negative parts, then for each and each category Both have two basic probability assignment functions:

[0170] ;

[0171] ;

[0172] in, express right The level of support express right The level of support.

[0173] The industrial robot absolute positioning accuracy compensation method provided in this embodiment introduces evidence theory to process feature vectors and generate contribution weights. The entire error compensation system has higher robustness and adaptability when facing complex and ever-changing environments. At the same time, the error compensation method based on contribution weights can concentrate more computational resources and attention on feature vectors that have a greater impact on the final prediction results, thereby improving the overall efficiency of the system and ensuring the accuracy of the deep belief network.

[0174] S110. The feature vector is weighted based on the contribution weight, and a compensation result is generated by combining the data to be compensated.

[0175] In practice, the contribution weight can be directly multiplied by the feature vector to obtain the location information. Based on the difference between this location information and the data to be compensated, the location compensation method can be determined to obtain the compensation result.

[0176] Figure 2 This is a schematic diagram illustrating an exemplary embodiment of the absolute positioning accuracy compensation process for an industrial robot. Please refer to... Figure 2 First, the multi-objective training of the accuracy compensation model and the original parameter set are determined. Then, candidate hyperparameters with an influence level higher than a preset threshold are obtained by screening. The deep belief network is trained in combination with the candidate hyperparameters to obtain the trained target deep belief network. Furthermore, the contribution weight of each feature vector is calculated by combining evidence theory, and the compensation result is generated based on the contribution weight.

[0177] The following is a specific example to illustrate the calculation process of the compensation result:

[0178] (1) Obtain the position coordinates and joint angles in the data to be compensated, input the position coordinates and joint angles into the visible layer of the target depth belief network, and extract features through the hidden layer of the target depth belief network to obtain the feature vector.

[0179] In practice, the robot's position coordinates and the angles of each joint are determined from the compensation data obtained from the industrial robot. This data is then input into an optimized target depth confidence network. Deep feature extraction is performed through the visible and hidden layers of the target depth confidence network to obtain feature vectors.

[0180] (2) Multiply the feature vector with the corresponding contribution weight to obtain a weighted vector, calculate the compensation results for the robot in three directions, and output the compensation results through the target depth belief network.

[0181] In practice, after extracting the feature vectors, the contribution weights for each feature vector are calculated according to the evidence theory. These weights are then multiplied by the corresponding feature vectors to obtain a weighted vector.

[0182] It is understandable that by multiplying the feature vector with the corresponding contribution weight, we can emphasize those features that have a greater impact on positioning accuracy, while weakening those features that have a smaller impact.

[0183] Furthermore, after obtaining the weighted vectors, these weighted vectors are used to calculate the compensation results of the robot in three directions (X-axis, Y-axis, and Z-axis).

[0184] In practice, the deep belief network calculates the prediction error of the robot in each direction based on the weighted vector and the pre-trained model parameters, and uses this error value as the amount that needs to be compensated.

[0185] The industrial robot absolute positioning accuracy compensation method provided in this embodiment has stronger robustness because the deep learning model can automatically learn and adapt to data features under different environments, and the uncertainty handling method can reduce the uncertainty in the prediction.

[0186] The following is a specific embodiment to illustrate the robot position correction process in detail:

[0187] (1) Determine the robot joint angle adjustment vector and the robot position coordinate adjustment vector based on the compensation results.

[0188] Specifically, the joint angle adjustment vector indicates the angle and direction that each joint needs to be adjusted, while the position coordinate adjustment vector directly indicates the distance and direction that the robot's position in three-dimensional space needs to be adjusted.

[0189] In practice, the compensation results are analyzed in depth to determine the robot joint angle adjustment vectors and robot position coordinate adjustment vectors in the X, Y, and Z directions.

[0190] (2) Correct the joint angle of the robot based on the robot joint angle adjustment vector, and correct the position of the end effector of the robot in the Cartesian coordinate system based on the robot joint angle adjustment vector.

[0191] In practice, adjustment commands can be sent to the corresponding joint actuators to adjust the joint angles according to the requirements of the adjustment vector.

[0192] Furthermore, the position of the robot's end effector can be corrected in the Cartesian coordinate system. By precisely controlling the position of the robot's end effector, it can be ensured that the robot can accurately reach the predetermined working point.

[0193] The industrial robot absolute positioning accuracy compensation method provided in this embodiment reduces robot malfunctions and downtime caused by positioning errors through high-precision positioning, thereby reducing the additional costs and time losses caused by fault repair. At the same time, the combination of robot joint angle adjustment vector and robot position coordinate adjustment vector can ensure the accuracy of robot adjustment.

[0194] Corresponding to the aforementioned embodiment of the industrial robot absolute positioning accuracy compensation method, this application also provides an experimental verification process for the industrial robot absolute positioning accuracy compensation method:

[0195] The experimental platform for absolute positioning accuracy compensation of industrial robots uses a KUKA KR6_R700 sixx_CR industrial robot, which has a 6kg load capacity, a working radius of 700mm, a working space volume of 1.36m³, a position repeatability of ±0.03mm, and an absolute position accuracy of ±0.6mm. The laser tracker used to measure the position error is a Leica AT901-B, with an error of ±15μm +6μm / m. The error of the laser tracker increases with increasing distance.

[0196] The AT901-B uses an angle encoder to measure angles and an absolute interferometer to measure distances. The absolute interferometer in the AT901 integrates a helium-neon laser interferometer and an absolute rangefinder. These two lasers can operate independently. The laser beam emitted by the laser is directed towards the target via a gimbaled mirror. The interferometer laser beam also serves as the line of sight of the tracker. The reflected laser light is measured by the tracker's built-in dual-axis position detector. The pulses generated by the position detector are processed by the tracker's processor and fed back to the tracker's servo motor, driving the motor to track the target mirror in real time. This ultimately achieves tracking and ranging, used to measure the actual pose of the end effector of an industrial robot.

[0197] The control and communication diagram of the error compensation platform is as follows: Figure 3 As shown. Figure 3 This is a schematic diagram illustrating a communication control mode in an exemplary embodiment of this application. A computer is used as the TwinCAT host, i.e., the main controller of the control system; the TwinCAT host communicates with the industrial robot using industrial Ethernet (EtherCAT), with a control cycle of 12ms; the laser tracker communicates with the TwinCAT host via Ethernet (TCP / IP protocol), and the sampling cycle of the laser tracker can reach 1ms.

[0198] A laser tracker can be used to obtain the actual position of the robotic arm, calculate the error, and finally use a deep belief network based on a genetic algorithm to achieve the error compensation function.

[0199] Under the premise of meeting the measurement requirements of the laser tracker, a 240mm×240mm×200mm workspace was selected within the target ball of the fixed fixture on the industrial robot to measure approximately 8000 sets of data. To ensure the universality and randomness of the experimental data, the random number module drand was used in TwinCAT3 to randomly generate specific sampling data within a pre-defined sampling space. To obtain the true pose of the robot and laser tracker in a stable state, each sampling step consisted of three steps. First, the robot paused for 2 seconds after reaching the sampling point; then, the laser tracker recorded the data for 1 second; finally, a 1-second delay was applied for reset. The theoretical position coordinates and joint angles of the robot served as the inputs to the model, while the absolute position error of the robot's end effector constituted the model's output.

[0200] The dataset was divided into training and test sets at a ratio of 0.3. The collected 8000 data sets were divided into 5600 training sets and 2400 test sets. For example... Figure 4 As shown, Figure 4 This is a schematic diagram illustrating the partitioning of a dataset as shown in an exemplary embodiment of this application. Blue dots represent the training set, and red dots represent the test set.

[0201] Furthermore, the input layer of the deep belief network has nine channels, which are the robot's theoretical position coordinates and joint angles, respectively. The output layer of the deep belief network has three channels, representing the robot's position error. , Figure 5 This diagram illustrates the dataset partitioning during individual training of a deep belief network, as shown in an exemplary embodiment of this application. A genetic algorithm is used to determine the number of hidden layers, hidden nodes, learning rate, momentum factor, number of RBM iterations, and number of deep belief network fine-tuning iterations for the deep belief network. Figure 6 This is a schematic diagram of the fitness function curve shown in an exemplary embodiment of this application. The fitness decrease curve of the DE-optimized deep belief network shows that the fitness value decreases with the increase of the number of iterations, from 0.387 to 0.152, a reduction of 60.7%; after 150 iterations of training, the optimal number of fitness iterations can be found at the end of the 92nd iteration, and the optimal parameters of the deep belief network are output.

[0202] The genetic algorithm (GA) is used to determine the number of hidden layers, hidden layer nodes, learning rate, momentum factor, number of RBM iterations, and number of deep belief network fine-tuning iterations. The fitness of each particle is calculated according to the fitness calculation conditions. The genetic algorithm terminates when the training error reaches the allowable value or the number of iterations reaches the maximum value.

[0203] Finally, the hyperparameters of the deep belief network determined by the genetic algorithm are shown in Table 2.

[0204] Table 2

[0205] Supervised learning in machine learning essentially involves establishing a mapping relationship based on a set of training samples, so that the fitted result is as close as possible to the true output. The loss function is a crucial metric for analyzing the quality of training results. There are various ways to define the loss function; this paper uses five metrics for analysis in deep belief networks: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Absolute Error (Mean Absolute Error), and R-Score. The definitions and explanations of these five metrics are given below.

[0206] ;

[0207] ;

[0208] ;

[0209] ;

[0210] ;

[0211] In the above formula, Represents the true values ​​of the dataset. This represents the predicted values ​​of the dataset. This represents the average of the predicted values ​​in the dataset. Indicates the number of datasets. This represents the variance of the data.

[0212] The mean square error is the average of the sum of squares of the errors between the predicted data and the corresponding points in the original data; the root mean square error is the square root of the mean square error, also called the standard deviation of the regression system; the mean absolute percentage error is often used to measure the accuracy of prediction. The mean absolute error refers to the average of the absolute values ​​of the deviations of each measurement, which can accurately reflect the magnitude of the actual prediction error; the closer these four indicators are to 0, the closer the predicted value is to the true value, and the better the prediction effect.

[0213] R² represents the coefficient of determination of the model. The best score is 1.0, indicating that the model perfectly predicts the true value; it can also be negative because the model can vary arbitrarily, meaning there is no mapping relationship between the predicted data and the true data. Table 3 shows the values ​​of MSE, RMSE, MAPE, MAR, and R² for the deep belief network used to fit the robot's end effector error.

[0214] Table 3

[0215] Using the proposed industrial robot position error prediction model, the MSE, RMSE, MAPE, and MAE of the predicted position error of the robot end effector are all close to 0. The coefficient of determination R² of the predicted position error is as follows: Figure 7-11 As shown, Figure 7 The diagram illustrates the error prediction results for an exemplary embodiment of this application. Figure 8 The diagram illustrates the error prediction results for another exemplary embodiment of this application. Figure 9 The diagram illustrates the error prediction results for another exemplary embodiment of this application. It can be seen that the predicted values ​​(blue dots) are closely distributed around the true values ​​(red lines), and the R² values ​​are all close to 1, indicating a high correlation between the predicted and actual values ​​and high fitting accuracy. Therefore, the proposed machine learning model demonstrates good adaptability and robustness in predicting the positional errors of industrial robots.

[0216] The R² of the robot's end-effector accuracy compensation error is around 0.87 to 0.95, indicating a good overall performance. However, because the deep belief network is trained and iterated simultaneously across all three dimensions, the features of these three dimensions influence and couple with each other. Furthermore, it is affected by nonlinear factors such as data acquisition accuracy and environmental conditions, leading to variations in the robot's end-effector accuracy error compensation performance.

[0217] To verify the effectiveness and improvement of the proposed genetic algorithm-based deep belief network, 150 random verification points were selected within the robot's motion space of 240mm×240mm×200mm. Figure 10 This is a schematic diagram illustrating the x-direction comparison as an exemplary embodiment of this application. Figure 11 This is a schematic diagram illustrating a comparison in the y-direction for another exemplary embodiment of this application. Figure 12 This is a schematic diagram illustrating the comparison in the z-direction, as shown in another exemplary embodiment of this application. The compensation results show that before compensation, the x-direction error was basically uniformly distributed around 0; the y-direction error was also distributed around 0, but more negative; and the z-direction error was almost entirely negative. Using the error compensation technique proposed in this paper, which utilizes a genetic algorithm to optimize a deep belief network, the errors in all three directions are basically distributed around 0, fluctuating around approximately ±0.2mm, ±0.1mm, and ±0.05mm. The fluctuation range is very small, indicating that the accuracy after compensation has high stability and can improve the accuracy of robot operation.

[0218] The static statistical analysis results before and after robot end-effector position error compensation are shown in Table 4, which is a table of compensation effects. The x, y, and z directions improved by 65.56%, 55.22%, and 49.12%, respectively.

[0219] Table 4

[0220] The experimental platform for data acquisition and verification of the robot is the KR6_R700 sixx_CR light industrial robot. Its error range is much smaller than that of traditional heavy industrial robots, making feature extraction, model training, and optimization using deep belief networks more challenging. This paper proposes using a genetic algorithm to optimize the deep belief network and, combined with evidence theory, establishes a position error mapping model for the industrial robot. Figure 13 This is a schematic diagram illustrating the compensation effect of an exemplary embodiment of this application. Please refer to... Figure 13 This involves analyzing the compensation effect on the robot's end effector accuracy from three perspectives combined.

[0221] To verify the experimental results, the method presented in this paper was compared with previous methods, and the results are shown in Table 5. Table 5 shows the comparison results of different methods. After offline compensation, the minimum value increased from 0.097 mm to 0.006 mm, and the average value increased from 0.110 mm to 0.083 mm. Therefore, the GA-depth belief network method proposed in this paper improves both the minimum and average values ​​after robot end-effector error compensation.

[0222] Table 5

[0223] Corresponding to the aforementioned embodiment of an industrial robot absolute positioning accuracy compensation method, this application also provides an embodiment of an industrial robot absolute positioning accuracy compensation device.

[0224] An embodiment of the absolute positioning accuracy compensation device for industrial robots disclosed in this application can be applied to absolute positioning accuracy compensation equipment for industrial robots. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the industrial robot absolute positioning accuracy compensation equipment loading the corresponding computer program instructions from the non-volatile memory into memory and executing them. From a hardware perspective, the industrial robot absolute positioning accuracy compensation equipment in which the device is located in the embodiment may also include other hardware depending on the actual function of the industrial robot absolute positioning accuracy compensation equipment, which will not be elaborated further.

[0225] Figure 14This is a schematic diagram of an industrial robot absolute positioning accuracy compensation device provided for an exemplary embodiment of this application. Please refer to... Figure 14 The apparatus provided in this embodiment includes an acquisition module 1410, a training module 1420, a calculation module 1430, and a processing module 1440; wherein,

[0226] The acquisition module 1410 is used to acquire the current pose data of multiple sampling points of the robot based on the laser tracker;

[0227] The training module 1420 is used to generate a position error based on the predetermined desired pose data and the current pose data of the robot, input the desired pose data into the depth belief network, and use the position error as the output of the depth belief network to train the depth belief network.

[0228] The training module 1420 is also used to optimize the hyperparameters of the deep belief network based on a genetic algorithm to obtain a trained target deep belief network.

[0229] The acquisition module 1410 is also used to acquire the data to be compensated during robot operation;

[0230] The calculation module 1430 is used to extract features from the data to be compensated using the target deep belief network, obtain feature vectors, and calculate the contribution weight of each feature vector based on evidence theory.

[0231] The processing module 1440 is used to weight the feature vector based on the contribution weight to generate a compensation result.

[0232] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0233] This application also provides an absolute positioning accuracy compensation device for an industrial robot, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the methods provided in the first aspect of this application.

[0234] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods provided in this application.

[0235] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0236] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0237] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for compensating the absolute positioning accuracy of an industrial robot, characterized in that, The method includes: Determine the multiple objectives for training the accuracy compensation model, wherein the multiple objectives include at least sub-objectives of the training dimension and sub-objectives of the accuracy compensation dimension; Determine the original set of parameters to be trained in the deep belief network; Calculate the degree of influence of each parameter in the original parameter set on the multi-objective; Parameters whose influence exceeds a preset threshold are selected as candidate hyperparameters; The current pose data of the robot at multiple sampling points is obtained using a laser tracker; Based on the predetermined desired pose data and the current pose data of the robot, a position error is generated. The desired pose data is input into the deep belief network, and the position error is used as the output of the deep belief network to train the deep belief network. While training the deep belief network, the candidate hyperparameters of the deep belief network are optimized based on a genetic algorithm to obtain the trained target deep belief network. Acquire the data to be compensated during robot operation; The target deep belief network is used to extract features from the data to be compensated to obtain feature vectors, and the contribution weight of each feature vector is calculated based on evidence theory. The feature vector is weighted based on the contribution weights, and a compensation result is generated by combining the data to be compensated.

2. The method according to claim 1, characterized in that, The step of inputting the desired pose data into a depth belief network and using the position error as the output of the depth belief network to train the depth belief network includes: The parameters of the deep belief network are adjusted based on a layer-by-layer greedy algorithm to obtain the initialized model parameters; wherein, the training process of the layer-by-layer greedy algorithm is unsupervised pre-training. The parameters of the deep belief network are fine-tuned based on the BP algorithm to complete the parameter adjustment of the deep belief network; wherein, the fine-tuning process of the BP algorithm is supervised fine-tuning.

3. The method according to claim 2, characterized in that, The process of adjusting the parameters of the deep belief network based on a layer-by-layer greedy algorithm to obtain the initialized model parameters includes: The desired pose data is processed based on the visible layer of the first restricted Boltzmann machine, and the target corresponding data of the desired pose data is output through the hidden layer of the restricted Boltzmann machine. Based on the data feedback corresponding to the target, the weights and biases of the visible layer are adjusted to obtain the trained first layer of the restricted Boltzmann machine; The desired pose data is processed based on the trained first-layer restricted Boltzmann machine, and the output data is used as the input of the second-layer restricted Boltzmann machine. The steps of obtaining the target corresponding data are repeated. Once the trained restricted Boltzmann machine has reached the preset number of layers, the model parameter initialization is confirmed to be complete.

4. The method according to claim 2, characterized in that, The fine-tuning of the parameters of the deep belief network based on the BP algorithm, to complete the parameter adjustment of the deep belief network, includes: The desired pose data is input into the deep belief network to obtain the output prediction result; Based on the prediction results and the position error, a loss function is calculated, and the gradient parameters of each gradient of the loss function are obtained; each gradient corresponds to a restricted Boltzmann machine layer. The gradient parameters are backpropagated in the deep belief network to fine-tune the parameters of the deep belief network.

5. The method according to claim 1, characterized in that, The optimization of the candidate hyperparameters of the deep belief network based on the genetic algorithm includes: For the candidate hyperparameters of the deep belief network, generate an optimized catalog and initialization configuration; The mean square error between the output bias of the deep belief network and the actual bias is calculated to obtain the fitness function of the genetic algorithm. The initial configurations are sorted based on the fitness function, and the portion of the fitness function value that is higher than a preset threshold is selected as the first target configuration. Perform cross-operation and compilation operations on the first target configuration to obtain a set of second target configurations, and update the initial configuration based on the second target configurations to obtain the updated configuration; Replace the initialization configuration with the updated configuration, re-execute the step of obtaining the updated configuration a preset number of times, and optimize the candidate hyperparameters of the deep belief network based on the obtained updated configuration.

6. The method according to claim 1, characterized in that, The calculation of the contribution weight of each feature vector based on evidence theory includes: Based on the evidence theory, the performance of the feature vector is calculated to obtain multiple basic probability assignment functions corresponding to each feature vector; For a given feature vector, fuse all the corresponding basic probability assignment functions; The contribution weights are generated based on the trust interval and the fused basic probability assignment function.

7. The method according to claim 1, characterized in that, The step of weighting the feature vector based on the contribution weights to generate a compensation result includes: The position coordinates and joint angles in the data to be compensated are obtained, and the position coordinates and joint angles are input into the visible layer of the target depth belief network. Feature extraction is performed through the hidden layer of the target depth belief network to obtain the feature vector. The feature vector is multiplied by the corresponding contribution weight to obtain a weighted vector. The weighted vector is then input into the restricted Boltzmann machine of the target depth belief network to obtain the compensation results for the robot in three directions. The compensation results are then output through the target depth belief network.

8. The method according to claim 5, characterized in that, The calculation of the mean square error between the output bias of the deep belief network and the actual bias includes: The desired pose data is input into the deep belief network to obtain the prediction output. The target median value is obtained by calculating the square of the difference between the predicted output and the position error. The mean square error is calculated based on the sum of squares of the target intermediate values, the output dimension of the deep belief network, and the number of data points contained in the expected pose data.

9. The method according to claim 1, characterized in that, After the method generates a compensation result by weighting the feature vector based on the contribution weight, it further includes: Based on the compensation results, determine the robot joint angle adjustment vector and the robot position coordinate adjustment vector; The robot's joint angles are corrected based on the robot joint angle adjustment vector, and the robot's end effector position is corrected in Cartesian coordinates based on the robot joint angle adjustment vector.

10. An offline compensation device for absolute positioning accuracy of an industrial robot, characterized in that, The device includes an acquisition module, a training module, a calculation module, and a processing module; wherein, The processing module is used to determine the multiple objectives of training the accuracy compensation model, wherein the multiple objectives include at least the sub-objectives of the training dimension and the sub-objectives of the accuracy compensation dimension. The processing module is also used to determine the original parameter set of the deep belief network to be trained; The calculation module is used to calculate the degree of influence of each parameter in the original parameter set on the multi-objective; The processing module is also used to filter parameters whose influence is higher than a preset threshold as candidate hyperparameters; The acquisition module is used to acquire the current pose data of multiple sampling points of the robot based on the laser tracker; The training module is used to generate a position error based on the predetermined desired pose data and the current pose data of the robot, input the desired pose data into the depth belief network, and use the position error as the output of the depth belief network to train the depth belief network. The training module is also used to optimize the candidate hyperparameters of the deep belief network based on a genetic algorithm while training the deep belief network, so as to obtain the trained target deep belief network. The acquisition module is also used to acquire the data to be compensated during robot operation; The calculation module is also used to extract feature vectors of the data to be compensated using the target deep belief network, and to calculate the contribution weight of each feature vector based on evidence theory. The processing module is further configured to weight the feature vector based on the contribution weight to generate a compensation result.

Images