A robot milling cutting force monitoring method considering pose-dependent dynamic characteristics

Abstract

This application belongs to the field of robot machining state monitoring technology, specifically disclosing a robot milling cutting force monitoring method considering pose-dependent dynamic characteristics. The method includes: acquiring robot monitoring data; and obtaining a predicted cutting force value based on the monitoring data and a depth regression model corresponding to the current pose. The depth regression model is used to predict the cutting force based on joint angle monitoring values ​​and acceleration signal monitoring values. The robot has multiple poses, including a source pose and at least one target pose. The depth regression model corresponding to the source pose is obtained through supervised training based on samples and corresponding labels at the source pose. The depth regression model corresponding to the target pose is obtained through transfer learning based on the depth regression model corresponding to the source pose. By considering pose-dependent dynamic characteristics and training corresponding depth regression models for each pose, accurate monitoring of cutting force under different poses can be achieved.

Description

Technical Field

[0001] This application belongs to the field of robot machining status monitoring technology, and more specifically, relates to a robot milling cutting force monitoring method that considers pose-dependent dynamic characteristics. Background Technology

[0002] Modern manufacturing industries place higher demands on product processing quality, efficiency, and economy. Processing condition monitoring is particularly important for high-precision, high-efficiency machining and intelligent manufacturing. Metal cutting is widely used in aerospace, automotive, shipbuilding, and electronics industries. Cutting force is generated by the interaction between the tool and the material during the cutting process. Compared to vibration and sound signals, it can more sensitively and quickly reflect the processing status. Therefore, cutting force is used in processing condition monitoring, tool wear monitoring, vibration characteristic analysis, and process parameter optimization.

[0003] Due to their low rigidity, robots are prone to chattering under milling forces, leading to severe tool wear and damage to the surface finish of the workpiece. Therefore, accurate monitoring of cutting forces during robotic milling is crucial for precisely calculating robot deformation, optimizing milling process parameters, and improving part machining accuracy. Currently, the mainstream methods for monitoring cutting forces in robotic milling include direct measurement using a force gauge and indirect measurement using integrated spindle sensors. However, benchtop force gauges are unsuitable for monitoring cutting forces in large-sized parts, are expensive, and struggle to maintain optimal performance under harsh environments with chips, cutting fluid, and vibration. Integrated spindle sensors offer greater flexibility, as vibration signals have a high bandwidth in the frequency domain, including displacement and acceleration signals. Mounting accelerometers on the spindle offers advantages such as not affecting robot machining operations, convenient signal acquisition and installation, and high economic efficiency. Indirect measurement of cutting forces using acceleration signals has been extensively studied in CNC milling. However, unlike traditional CNC machine tools, the dynamic characteristics of robotic milling systems are heavily dependent on pose, resulting in significant differences in their distribution within the workspace. This results in low stability in robotic milling, with significant variations in milling performance across different machining zones, and a narrow window of available process parameter options. Indirect cutting force identification methods used on machine tools cannot be directly applied to robotic milling.

[0004] Data-driven approaches can learn end-to-end relationships using acquired data, establishing models that map the relationship between acceleration signals and cutting forces. Traditional machine learning algorithms, however, require data to be uniformly distributed. Robots exhibit different dynamic characteristics in different poses, and predictive force models built for a single pose have poor generalization ability for other poses. Therefore, accurately monitoring cutting forces under different poses is a pressing technical problem in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application aims to achieve accurate monitoring of cutting forces under different poses. Through robotic milling experiments in various poses, joint angles of each robot joint, acceleration signals of the robot spindle, and cutting force signals are collected during the cutting process. A deep transfer regression algorithm (using deep learning technology for regression analysis, constructing a deep regression model, and using transfer learning to transfer the deep regression model from the source pose to the target pose) is used to implement a regression model for predicting cutting forces under different poses using robot joint angles and acceleration signals. During actual machining, after changing the robot pose, combined with a small amount of data from the target pose, the monitored joint angles and acceleration signals are input into the established prediction model for iterative prediction, thus obtaining the real-time cutting force under the new pose. This reduces the number of cutting force measurement experiments under new poses, significantly lowering the complexity and cost of modeling and experimentation.

[0006] To achieve the above objectives, in a first aspect, this application provides a method for monitoring robot milling cutting forces considering pose-dependent dynamic characteristics, comprising:

[0007] Acquire monitoring data for the robot, including joint angle monitoring values ​​and acceleration signal monitoring values;

[0008] Based on the monitoring data and the depth regression model corresponding to the current pose, the predicted value of the cutting force is obtained;

[0009] Among them, the deep regression model is used to predict the cutting force based on the joint angle monitoring value and the acceleration signal monitoring value. The robot has multiple poses, including a source pose and at least one target pose. The deep regression model corresponding to the source pose is obtained by supervised training based on the samples and corresponding labels at the source pose. The deep regression model corresponding to the target pose is obtained by transfer learning based on the deep regression model corresponding to the source pose.

[0010] The robot milling cutting force monitoring method provided in this application considers the pose-dependent dynamic characteristics. For multiple robot poses (including a source pose and at least one target pose), during the model training phase, supervised training can be performed using samples and corresponding labels at the source pose to obtain a depth regression model corresponding to the source pose. Then, through transfer learning, the depth regression model is transferred from the source pose to the target pose to obtain a depth regression model corresponding to the target pose. In the model application phase, monitoring data of the robot can be acquired in real time, including joint angle monitoring values ​​and acceleration signal monitoring values. The joint angle monitoring values ​​and acceleration signal monitoring values ​​are input into the depth regression model corresponding to the robot's current pose (which can be either the source pose or the target pose). The depth regression model is used to predict the cutting force at the current moment. The depth regression model corresponding to the source pose can accurately predict the cutting force at the source pose, and the depth regression model corresponding to the target pose can accurately predict the cutting force at the target pose, thus achieving accurate monitoring of cutting force under different poses.

[0011] Optionally, the robot can be configured with multiple sets of different machining process parameters in each pose, and cutting experiments can be performed for each set of machining process parameters to obtain the dataset corresponding to that set of machining process parameters.

[0012] For any set of machining parameters at the source pose (for ease of description, this set of machining parameters is named the Sm-th set of machining parameters), the initial depth regression model can be trained in a supervised manner using the labeled dataset corresponding to the Sm-th set of machining parameters at the source pose. This yields a pre-trained depth regression model corresponding to the Sm-th set of machining parameters at the source pose. Then, when the robot is in the source pose and performing cutting operations using the Sm-th set of machining parameters, the pre-trained depth regression model can be used to predict the cutting force. In essence, for the robot's source pose, supervised training can be performed using the labeled dataset corresponding to each set of machining parameters to obtain a pre-trained depth regression model for each set of machining parameters.

[0013] For multiple sets of machining process parameters under the target pose, a labeled dataset can be constructed for one set of machining process parameters (for ease of description, this set of machining process parameters is named the Tn1 set of machining process parameters), and an unlabeled dataset can be constructed for the other sets of machining process parameters.

[0014] If the dataset corresponding to a set of machining process parameters at the target pose (let's call this set of machining process parameters the Tn2th set for ease of description) is an unlabeled dataset, then based on the Tn2th set of machining process parameters, we can find a set of machining process parameters (named the Smth set of machining process parameters) that is similar to the Tn2th set of machining process parameters among multiple sets of machining process parameters at the source pose. It's understandable that finding a similar set of machining process parameters can reduce the difficulty of transfer learning and improve its efficiency. For example, based on the spindle speed in the Tn2th set of machining process parameters, we can find the Smth set of machining process parameters among multiple sets of machining process parameters at the source pose, where the spindle speed in the Tn2th set of machining process parameters is the same as the spindle speed in the Smth set of machining process parameters. Then, the pre-trained deep regression model corresponding to the Sm group of processing parameters can be used as the initial model for transfer learning. Then, the labeled dataset corresponding to the Sm group of processing parameters, the labeled dataset corresponding to the Tn1 group of processing parameters, and the unlabeled dataset corresponding to the Tn2 group of processing parameters can be used for transfer learning to obtain the pre-trained deep regression model corresponding to the Tn2 group of processing parameters.

[0015] Accordingly, the above-mentioned "obtaining the predicted value of cutting force based on the monitoring data and the depth regression model corresponding to the current pose" specifically includes: obtaining the predicted value of cutting force based on the monitoring data and the depth regression model corresponding to the target machining process parameters under the current pose, wherein the target machining process parameters are the machining process parameters currently used by the robot.

[0016] In one possible implementation, the data input to the deep regression model includes monitoring data from the current and previous time points, as well as the cutting force signal value from the previous time point (during model training, the cutting force signal value can be the cutting force signal label value; during model application, the cutting force signal value can be the cutting force signal prediction value). The data output by the deep regression model can be the cutting force prediction value at the current time point. Accordingly, the above-mentioned "obtaining the cutting force prediction value based on the monitoring data and the deep regression model corresponding to the current pose" specifically includes:

[0017] Input the monitoring data of the current time and the previous time, as well as the cutting force signal value of the previous time, into the depth regression model corresponding to the current pose, and obtain the cutting force prediction value of the current time output by the depth regression model corresponding to the current pose.

[0018] One possible implementation also includes obtaining the depth regression model corresponding to the source pose through the following steps:

[0019] A source pose dataset is constructed based on the monitoring signal samples (including joint angle samples and acceleration signal samples) and the corresponding cutting force signal labels of the source pose.

[0020] Based on the source pose dataset, an initial deep regression model is trained to obtain the deep regression model corresponding to the source pose.

[0021] One possible implementation also includes obtaining the depth regression model corresponding to the target pose through the following steps:

[0022] Based on the labeled monitoring signal samples of the target pose and the corresponding cutting force signal labels, a labeled dataset of the target pose is constructed. Based on the unlabeled monitoring signal samples of the target pose and the corresponding pseudo-labels, an unlabeled dataset of the target pose is constructed. The pseudo-labels are obtained by performing principal component analysis on the unlabeled monitoring signal samples.

[0023] Based on the labeled dataset of the target pose, the unlabeled dataset of the target pose, and the deep regression model corresponding to the source pose, transfer learning is performed iteratively to obtain the deep regression model corresponding to the target pose.

[0024] In one possible implementation, the aforementioned "transfer learning iterative training" includes updating the model parameters by minimizing the computed value of the following weighted ensemble loss function:

[0025]

[0026] in, The regression loss is for the labeled monitoring signal samples at the target pose. This represents the difference in conditional embedding distribution between the monitored signal samples of the source pose and the unlabeled monitored signal samples of the target pose. The conditional embedding distribution difference between labeled and unlabeled monitoring signal samples represents the target pose, with α and β being trade-off parameters.

[0027] The following section elaborates on the transfer learning method adopted in this application and the aforementioned weighted ensemble loss function used in the transfer learning process.

[0028] This application, through experiments, reveals that when the difference between the source and target domains is small (e.g., the difference lies in the spindle speed of the source domain versus the spindle speed of the target domain), the weight parameters of one or more modules in the trained model of the source domain are highly similar to the weight parameters of the corresponding modules in the trained model of the target domain. Therefore, in this case, transfer learning can be performed by sharing some weights between the source and target domain models (one or more modules in the source domain model and the corresponding modules in the target domain model use the same weight parameters, for example, the feature extractor in the source domain model and the feature extractor in the target domain model use the same weight parameters), thereby reducing the difficulty of transfer learning and improving the training effect.

[0029] This application also found through experiments that when there are significant differences between the source domain and the target domain (for example, differences between the source domain and the target domain may include different poses in the source domain and different machining process parameters (including spindle speed) in the source domain and different machining process parameters (including spindle speed) in the target domain), the weight parameters of each module in the trained model of the source domain are significantly different from the weight parameters of the corresponding modules in the trained model of the target domain. Therefore, in this case, if transfer learning is still performed by sharing some weights between the model of the source domain and the model of the target domain, the training effect will be reduced.

[0030] The application scenario targeted by model transfer in this application is the transfer from a source pose to a target pose. The difference between the source and target domains is significant, and the approach of "sharing some weights between the source and target domain models" is not suitable for this application scenario. Therefore, a new transfer learning method needs to be designed to reduce the transfer difficulty and improve training performance. Specifically, the source pose model is used as the initial model for the target pose. Transfer learning is performed based on this initial model, and the source and target pose models do not share weights during the transfer learning process. Labeled and unlabeled data are constructed for the target pose, and two conditions embedding distribution differences are designed accordingly (one condition embedding distribution difference represents the difference in conditional distribution metrics between the source pose data and the unlabeled target pose data, and the other represents the difference in conditional distribution metrics between the labeled and unlabeled target pose data). The target pose regression loss is combined with the two conditions embedding distribution differences. The network is trained by minimizing the regression loss and the conditions embedding distribution differences. The model obtained through transfer learning can adapt to the target pose, ensuring training effectiveness.

[0031] It is also worth noting that, for the transfer learning method provided in this application, the sample data under the target pose includes labeled sample data and unlabeled sample data. It is not necessary to assign real labels to all samples under the target pose, which can effectively reduce the workload required to prepare transfer-related data, reduce the difficulty of transfer, and improve the efficiency of transfer.

[0032] In one possible implementation, the deep regression model includes a cascaded convolutional neural network feature extractor and a regressor. The regressor includes multiple fully connected layers, with the number of fully connected layers being greater than or equal to three. These fully connected layers are connected in a cascaded manner. During transfer learning iterative training, both unlabeled and labeled monitoring signal samples of the target pose are processed by the same convolutional neural network feature extractor. During transfer learning iterative training, the unlabeled and labeled monitoring signal samples of the target pose are processed by different regressors. The model also includes:

[0033] The first dataset is constructed by processing the monitoring signal samples based on the source pose through the second fully connected layer of the regressor in the deep regression model corresponding to the source pose, and combining the features formed by these features with the labels corresponding to the monitoring signal samples based on the source pose.

[0034] The second dataset is constructed from the features formed by processing the unlabeled monitoring signal samples of the target pose through the second fully connected layer in the corresponding regressor, and the predicted cutting force signal values ​​obtained by processing the unlabeled monitoring signal samples of the target pose through the regressor.

[0035] A third dataset is constructed based on the features generated from the labeled monitoring signal samples at the target pose through the second fully connected layer in the corresponding regressor, and the labels corresponding to the labeled monitoring signal samples at the target pose.

[0036] Analyze the differences between the conditional distribution features of the first dataset and the conditional distribution features of the second dataset in the RKHS, and obtain the first conditional embedding distribution differences.

[0037] Analyze the differences in RKHS between the conditional distribution features corresponding to the second dataset and the conditional distribution features corresponding to the third dataset to obtain the differences in the second conditional embedding distribution.

[0038] In one possible implementation, the convolutional neural network feature extractor is determined by cascading a first convolutional layer, a first normalization layer, a second convolutional layer, a pooling layer, and a second normalization layer.

[0039] Secondly, this application provides a robot milling cutting force monitoring device that considers pose-dependent dynamic characteristics, comprising:

[0040] The monitoring data acquisition module is used to acquire monitoring data of the robot, including joint angle monitoring values ​​and acceleration signal monitoring values.

[0041] The cutting force prediction module is used to obtain the predicted cutting force value based on the monitoring data and the depth regression model corresponding to the current pose;

[0042] The deep regression model is used to predict cutting force based on joint angle monitoring values ​​and acceleration signal monitoring values. The robot has multiple poses, including a source pose and at least one target pose. The deep regression model corresponding to the source pose is obtained through supervised training based on samples and corresponding labels at the source pose. The deep regression model corresponding to the target pose is obtained through transfer learning based on the deep regression model corresponding to the source pose.

[0043] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0044] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0045] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0046] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art:

[0047] (1) The cutting force is monitored by using the vibration signal of the robot spindle, which realizes the monitoring of cutting force under different robot postures. The sensor installation does not affect the robot's processing operation. It is simple and convenient to use and has the potential to be promoted to practical industrial applications.

[0048] (2) The mapping relationship between vibration signal and cutting force is established by using data-driven model, without the need for complex parameter identification process or compensation strategy.

[0049] (3) The deep transfer regression algorithm was used to predict the cutting force of the robot in other target poses by using labeled cutting data in finite poses, which greatly reduced the number of experiments and thus reduced the cost of workpiece material, cutting tools and time. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the robot milling cutting force monitoring method that considers pose-dependent dynamic characteristics provided in the embodiments of this application.

[0051] Figure 2 This is a flowchart illustrating the model training method provided in the embodiments of this application;

[0052] Figure 3 This is a schematic diagram of the cutting force depth migration regression prediction model under the target pose, which considers pose-dependent dynamic characteristics, provided in an embodiment of this application.

[0053] Figure 4 This is a schematic diagram of robot milling in different poses provided in the embodiments of this application;

[0054] Figure 5This is one of the schematic diagrams showing the comparison between measured and predicted cutting force values ​​in linear trajectory milling provided in the embodiments of this application;

[0055] Figure 6 This is the second schematic diagram showing the comparison between the measured and predicted cutting force values ​​in linear trajectory milling provided in the embodiments of this application;

[0056] Figure 7 This is one of the schematic diagrams showing the comparison between the measured and predicted cutting force values ​​during arc trajectory milling provided in the embodiments of this application;

[0057] Figure 8 This is the second schematic diagram comparing the measured and predicted cutting force values ​​during arc trajectory milling provided in the embodiments of this application;

[0058] Figure 9 This is one of the schematic diagrams comparing different cutting force prediction methods under linear trajectory milling provided in the embodiments of this application;

[0059] Figure 10 This is the second schematic diagram comparing different cutting force prediction methods under linear trajectory milling provided in the embodiments of this application;

[0060] Figure 11 This is one of the schematic diagrams comparing different cutting force prediction methods under circular arc trajectory milling provided in the embodiments of this application;

[0061] Figure 12 This is the second schematic diagram comparing different cutting force prediction methods under circular arc trajectory milling provided in the embodiments of this application;

[0062] Figure 13 This is a schematic diagram of the structure of the robot milling cutting force monitoring device that considers pose-dependent dynamic characteristics provided in the embodiments of this application;

[0063] Figure 14 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0065] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0066] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0067] The embodiments of this application are described below with reference to the accompanying drawings.

[0068] Figure 1 This is a flowchart illustrating the robot milling cutting force monitoring method considering pose-dependent dynamic characteristics provided in the embodiments of this application, as shown below. Figure 1 As shown, the subject executing this method can be an electronic device, such as a server. The method includes the following steps S101 and S102.

[0069] Step S101: Obtain the robot's monitoring data, which includes joint angle monitoring values ​​and acceleration signal monitoring values.

[0070] Step S102: Based on the monitoring data and the depth regression model corresponding to the current pose, obtain the predicted value of the cutting force.

[0071] Among them, the deep regression model is used to predict the cutting force based on the joint angle monitoring value and the acceleration signal monitoring value. The robot has multiple poses, including a source pose and at least one target pose. The deep regression model corresponding to the source pose is obtained by supervised training based on the samples and corresponding labels at the source pose. The deep regression model corresponding to the target pose is obtained by transfer learning based on the deep regression model corresponding to the source pose.

[0072] By considering the pose-dependent dynamic characteristics, for multiple robot poses (including a source pose and at least one target pose), during the model training phase, supervised training can be performed using samples and corresponding labels at the source pose to obtain a deep regression model corresponding to the source pose. Then, through transfer learning, the deep regression model can be transferred from the source pose to the target pose to obtain a deep regression model corresponding to the target pose. In the model application phase, real-time monitoring data of the robot can be acquired, including joint angle monitoring values ​​and acceleration signal monitoring values. The joint angle monitoring values ​​and acceleration signal monitoring values ​​are input into the deep regression model corresponding to the robot's current pose (which can be either the source pose or the target pose). The deep regression model is used to predict the cutting force at the current moment. The deep regression model corresponding to the source pose can accurately predict the cutting force at the source pose, and the deep regression model corresponding to the target pose can accurately predict the cutting force at the target pose, thus achieving accurate monitoring of the cutting force under different poses.

[0073] Figure 2This is a flowchart illustrating the model training method provided in the embodiments of this application, as shown below. Figure 2 As shown, milling experiments can be conducted on robots in different poses. By monitoring acceleration signals, joint angles, and cutting force signals, datasets can be constructed, such as labeled datasets for the source pose (or source domain), labeled data for the target pose (or target domain), and unlabeled data for the target pose. Pseudo-labels are assigned to the unlabeled data for the target pose. After constructing the datasets, an initial deep regression model (including a feature extractor and a regressor) can be trained based on the labeled dataset for the source pose to obtain the deep regression model corresponding to the source pose (or source domain prediction model). Then, transfer learning can be used to transfer the model parameters to obtain the deep regression model corresponding to the target pose (or target domain prediction model). Through the above training, cutting force prediction models for different poses can be obtained.

[0074] In one possible implementation, the cutting force prediction model under different poses can be obtained through the following steps S201 to S204.

[0075] Step S201: Using the planned machining process parameters, conduct robot milling experiments under different poses, and collect joint angles, acceleration signals and cutting force signals to construct the source pose experiment original dataset and the target pose experiment original dataset.

[0076] In step S202, the original signals collected in step S201 are processed to obtain the model input and output signals under the corresponding process parameters and to construct a dataset that can be used for network training and testing.

[0077] Step S203: For the prediction of the target pose, it is necessary to assign pseudo-labels to the unlabeled target pose data using the principal component analysis (PCA) method. The pseudo-labels and the target pose monitoring input data are processed together to form the multidimensional input data of the target pose to meet the input data dimensionality requirements of the network.

[0078] Step S204: Based on the dataset obtained in step S203, establish a deep transfer regression prediction model.

[0079] Specifically, a deep learning prediction model for the source pose is established and trained based on the source pose data. This source pose deep learning prediction model serves as the initial model for transfer learning during the transfer of model parameters to the target pose model. Then, labeled and unlabeled target pose data are input into the target pose model with initialized parameters. First, feature extraction is performed, followed by regression prediction on the extracted features. The regression prediction values ​​are calculated using the features from the unlabeled target pose data, the differences in conditional embedding distributions between the source pose data features and their labels, and the differences in conditional embedding distributions between the labeled target pose data and their labels. Finally, the regression loss is calculated using the predicted values ​​from the labeled target pose data and their labels. The network is trained by minimizing the target pose regression loss and the differences in the two conditional embedding distributions to obtain a cutting force prediction model for the target pose, thus achieving model knowledge transfer.

[0080] In one possible implementation, in step S201, an acceleration sensor is attached to the spindle, acceleration signals during milling are acquired through a data acquisition system, joint angles of various robot joints during machining are obtained through the robot's open data acquisition system, and cutting force signals during milling are measured and acquired using a force measuring system.

[0081] In one possible implementation, in step S202, the first and last non-cutting segment data of the original signals under different poses are first removed. For cutting force and acceleration signals with high sampling frequency, they need to be synchronized with the cutting start point of the robot joint angle data, and then downsampling is performed to make them have the same time data length as the joint angle data.

[0082] It is understandable that the acceleration-force frequency response function is expressed in the Laplace domain as:

[0083]

[0084] In the formula, m is the number of modes, and R k For the model residuals, ξ k and ω nk These are the k-th order damping ratio and the natural frequency, respectively.

[0085] Converting it into a differential equation, the cutting force can be calculated using the inverse estimate of the displacement:

[0086]

[0087] Where A j (j=0,...,n), B j (j=1,...,l) are the coefficients of the differential equation, a j (j=0,...,n) and Let be the j-th derivative with respect to time, and n and l be a and F, respectively. a The order of.

[0088] Data-driven methods can be used to establish a dynamic nonlinear relationship between monitored acceleration and joint angle signals and cutting force. In the nonlinear dynamic prediction model, the relationship between robot joint angles, acceleration signals, and cutting force is represented as a discrete-time dynamic system model:

[0089]

[0090] Where k represents the current time, F a (k) represents the cutting force at the current moment, q1(kj), q2(kj), q3(kj), q4(kj), q5(kj), q6(kj), a(kj), (j=0,1,...,n) represent the joint angles and acceleration signals at previous and current moments. F a (kj), (j = 1, 2, ..., l) represent the predicted cutting force values ​​at previous time steps (using the corresponding labels or pseudo-labels of the samples during the training phase), and n and l represent the time delay orders related to the input and output, for example, n = 6 and l = 5. f represents the nonlinear relationship between the input and output, which is constructed using a data-driven model (i.e., constructing f by training a deep regression model using the dataset). During model training, the time-aligned data is converted into multidimensional data according to the above formula.

[0091] In one possible implementation, in step S203, after obtaining the source pose dataset, for the unlabeled data under the new target pose, in order to meet the size of the model input dimension, pseudo-labels need to be assigned to the unlabeled monitoring input data before training. The PCA method is used to reduce the dimensionality of the multidimensional time-series data, and principal components are selected as pseudo-label data. PCA preserves the main direction of data change, thus reflecting the deep feature representation and structure of the input data. Based on the minimum projective distance of the entire sample set, the following formula can be obtained:

[0092]

[0093] Tr(·) represents the trace of the matrix. When the above expression takes its minimum value, the corresponding W is determined by the covariance matrix XX. T The eigenvectors corresponding to the largest eigenvalue can be transformed into the following formula using the Lagrange function:

[0094] J(W) = -Tr(W) T XX T W+λ(W T WI));

[0095] Differentiate W with respect to W and summarize as follows:

[0096] XX T W = λW;

[0097] W is XX T A matrix consisting of n eigenvectors, where λ is XX. T The matrix is ​​composed of the eigenvalues. When reducing the dataset from m dimensions to n dimensions, it is necessary to find the n largest eigenvalues ​​to form an eigenvector, which serves as the pseudo-label data.

[0098] In one possible implementation, in step S204, Figure 3 This is a schematic diagram of the cutting force depth migration regression prediction model under the target pose, considering pose-dependent dynamic characteristics, provided in an embodiment of this application. Figure 3 As shown, a deep network model was established, including a feature extraction network and a fully connected network, to extract features from time-series data and perform regression prediction on these features. First, a CNN feature extraction network and a fully connected regression prediction module were trained using the source pose data. Then, the pre-trained model parameters were transferred to the target pose model with the same structure. The feature extractor is defined as follows:

[0099]

[0100] Where, the subscript S represents the source pose, and F ES This is to represent the depth features contained in the input temporal signal extracted from the source pose. This represents a feature extractor for the source pose data. s It is the input source pose data sample. This represents the feature extractor parameters corresponding to the source pose.

[0101] A deep fully connected network is built to map the extracted features to the corresponding cutting force values, with an output dimension of 1. The regression loss of the regressor is as follows:

[0102]

[0103] in, This represents the regressor corresponding to the source pose, y s n represents the label corresponding to the source pose data sample. s The number of source pose samples, The calculation of mean squared error (MSE) is as follows: This represents the regressor parameters.

[0104] Based on the kernel embedding theory of conditional distributions, a conditional distribution embedding computation operator is proposed to measure the difference in conditional probability distributions. For O in the original space X (the space formed by the samples) and O YThe variables X and Y (in the space formed by the labels corresponding to the samples) have their corresponding reproducing kernel Hilbert spaces (RKHS) as follows: and By embedding p(Y|X) into RKHS, the conditional mean embedding can be defined as:

[0105]

[0106] in Let ψ(x) represent the expectation. And φ(Y): and yes The operator, which characterizes the features of the conditional distribution, yields a conditional mean embedding value μ for each fixed x. Y|x .

[0107] According to the formula for calculating the conditional probability distribution, it is equal to the ratio of the joint probability distribution to the marginal probability distribution of the input. The conditional embedding operator can be defined as a combination of the cross-covariate operator and the autocovariate operator:

[0108]

[0109] in, Represents the tensor product. Given a dataset D = {(x1,y1),(x2,y2),...,(x...}... n ,y n When x in dataset D is in the range )}, n Let y represent a sample in dataset D. n This indicates the label corresponding to the sample. The estimated value can be defined as follows:

[0110]

[0111] in, In this context, n, λ, and I represent the number of samples in the dataset, the preset parameters, and the identity matrix, respectively. Φ = (φ(y1), φ(y2), ..., φ(y...). n )), Υ=(ψ(x1),ψ(x2),...,ψ(x n )),and It is the Gram matrix of the variable samples, with an added regularization term (nλI) to ensure that it is well-posed.

[0112] Given two datasets D with different distributions s ={(x1,y1),(x2,y2),...,(x m ,y m )} and D t={(x1,y1),(x2,y2),...,(x n ,y n The marginal probability distribution of X is generally calculated using the following formula. This nonparametric computation (i.e., nonparametric estimation) method measures the mean embedding. and Differences in RKHS.

[0113]

[0114] and denoted by m and n, respectively, represent the samples in the source and target domains, where m represents the number of samples in the source domain and n represents the number of samples in the target domain. φ represents the feature map. This represents an RKHS with a characteristic kernel κ.

[0115] A conditional distribution embedding computation operation is designed to estimate the differences between conditional probability distributions. Definition and The following formula is used as the loss function for updating network parameters to minimize domain differences as much as possible.

[0116]

[0117] in, The first dataset is constructed by processing the features formed by the monitoring signal samples based on the source pose through the second fully connected layer in the regressor of the deep regression model corresponding to the source pose, and the labels corresponding to the monitoring signal samples based on the source pose. The second dataset is constructed by processing the unlabeled monitoring signal samples based on the target pose through the second fully connected layer in the corresponding regressor, and the predicted value of the cutting force signal obtained by processing the unlabeled monitoring signal samples based on the target pose through the regressor. The third dataset is constructed by processing the labeled monitoring signal samples based on the target pose through the second fully connected layer in the corresponding regressor to form features and labels corresponding to the labeled monitoring signal samples based on the target pose. Represents the first dataset Corresponding conditional distribution characteristics; Represents the second dataset Corresponding conditional distribution characteristics; Represents the third dataset Corresponding conditional distribution characteristics; express and Differences in RKHS; express and Differences in RKHS. Matrix K XX'Calculations were performed using the Gaussian kernel κ:

[0118] K XX' (i,j)=κ(X i ,X' j )= <f(X i ),f(X' j )>;

[0119] Where <·,·> denotes the vector dot product, the subscript tl represents labeled samples of the target pose, the subscript tu represents unlabeled samples of the target pose, and n s n represents the number of samples in the source pose dataset. tu n represents the number of unlabeled samples in the target pose dataset. tl This indicates the number of labeled samples in the target pose dataset.

[0120] Therefore, the conditional embedding distribution difference can be calculated by using the new features formed by the target pose unlabeled input data features after passing through the second fully connected layer, the regression network's predicted values, the new features formed by the source pose data features after passing through the second fully connected layer, and their label values. The new features formed by passing unlabeled input data features through a second fully connected layer can be used to regress the network's predicted values. The new features formed by passing labeled data features through a second fully connected layer and their label values ​​can be calculated based on the conditional embedding distribution differences.

[0121] The regression loss is calculated using the target pose predicted values ​​and their corresponding label values. It can be calculated as follows:

[0122]

[0123] Here, the subscript t represents the target pose.

[0124] For training the feature extractor and task-specific regressor for the target data, the following weighted ensemble loss function is used for parameter updates:

[0125]

[0126] in, The regression loss is used for labeled data at the target pose. This represents the difference in conditional embedding distribution between the source pose data (or the source pose monitoring signal samples) and the target pose unlabeled data (or the target pose unlabeled monitoring signal samples). The conditional embedding distribution difference between labeled and unlabeled data (or unlabeled monitoring signal samples) represents the difference in target pose between labeled and unlabeled data, where α and β are trade-off parameters.

[0127] The network is trained by minimizing the target pose regression loss and the difference between the two conditional embedding distributions. The number of iterations is set to 100, and the initial learning rate of the regressor is set to 0.0001. The network parameters are updated using the Adam optimizer. First, the gradient of the weighted loss function equivalent to the network parameters at time step t is calculated:

[0128]

[0129] in, and In the subscripts f, f represents the feature extractor; R represents the regressor; S represents the source pose; T represents the target pose; TL represents the labeled case at the target pose; and TU represents the unlabeled case at the target pose. These represent the gradients of the loss functions of the feature extraction network and the regressor network with respect to the network parameters at time step t, respectively.

[0130] Then update the first and second moment estimates in momentum form:

[0131]

[0132] Where the subscript t represents the current time step, and β1 and β2 represent the exponential decay rate of the moment estimation, which can effectively smooth the fluctuations in gradient calculation and make the optimization results more accurate and stable.

[0133] Next, the bias-corrected first-order moment estimates and second-order moment estimates are calculated:

[0134]

[0135] in and It is the t power of β1 and β2.

[0136] Finally, update the network parameters:

[0137]

[0138] In the formula, Θ t Let η be the network parameter vector, and η be the learning rate. An initial learning rate can be set, and then the learning rate can be automatically updated according to the program. This learning rate optimization strategy is beneficial for the rapid convergence of the training model and the improvement of model performance. ε represents a small constant equal to 10e-8 to avoid the denominator being equal to 0.

[0139] After iteration, the cutting force prediction model for the target pose is obtained:

[0140]

[0141] In this context, f represents the feature extractor, R represents the regressor, T represents the target pose, and TU represents the unlabeled case under the target pose.

[0142] Therefore, by minimizing the target pose regression loss and the difference between the two conditional embedding distributions, the network is trained to obtain a cutting force prediction model for the target pose, thus achieving model knowledge transfer. During actual machining, after changing the robot pose, the monitored robot joint angles and acceleration signals are input into the established prediction model along with a small amount of labeled data for training and iterative prediction, thereby obtaining the real-time cutting force under the new pose.

[0143] The following is an optional example of this application, but it is not intended to limit this application.

[0144] Milling experiments were conducted using a robot according to the planned machining parameters. Cutting force signals were measured by a high-precision benchtop force gauge mounted on the robot's worktable, with the workpiece mounted on the gauge. Vibration signals during the milling process were collected using an accelerometer magnetically attached to the robot's spindle. An open data acquisition system was used to collect the joint angles of each axis during the robot's machining process. The sampling time interval for joint angles was 1 ms, and the acquisition frequency for acceleration and cutting force signals could reach tens of kHz, with higher frequencies being preferable; the sampling frequency was set to 25.6 kHz.

[0145] The cutting process was conducted using linear and circular trajectory milling experiments. The workpiece was aluminum alloy 7075, and the cutting tool was a solid carbide end mill with a diameter of 12mm. The machining process parameters are shown in Table 1. Figure 4 This is a schematic diagram of robot milling in different poses provided in the embodiments of this application, such as... Figure 4 As shown, cutting experiments with different machining parameters were conducted in multiple robot poses to construct a dataset.

[0146] Each set of machining process parameters corresponds to a dataset. For example, for "Straight trajectory source pose 1" in Table 1, there are 6 sets of machining process parameters. Through cutting experiments, the datasets corresponding to the 6 sets of machining process parameters can be obtained (the dataset under the source pose can be a labeled dataset). For example, for "Straight trajectory target pose 2" in Table 1, there are 4 sets of machining process parameters. Through cutting experiments, the datasets corresponding to the 4 sets of machining process parameters can be obtained (the dataset under the target pose can be a labeled dataset or an unlabeled dataset).

[0147] For any set of machining parameters at the source pose (for ease of description, this set of machining parameters is named the Sm-th set of machining parameters), the initial depth regression model can be trained in a supervised manner using the labeled dataset corresponding to the Sm-th set of machining parameters at the source pose. This yields a pre-trained depth regression model corresponding to the Sm-th set of machining parameters at the source pose. Then, when the robot is in the source pose and performing cutting operations using the Sm-th set of machining parameters, the pre-trained depth regression model can be used to predict the cutting force. In essence, for the robot's source pose, supervised training can be performed using the labeled dataset corresponding to each set of machining parameters to obtain a pre-trained depth regression model for each set of machining parameters.

[0148] For multiple sets of machining process parameters under the target pose, a labeled dataset can be constructed for one set of machining process parameters (for ease of description, this set of machining process parameters is named the Tn1 set of machining process parameters), and an unlabeled dataset can be constructed for the other sets of machining process parameters.

[0149] If the dataset corresponding to a set of machining process parameters at the target pose (let's call this set of machining process parameters the Tn2th set for ease of description) is an unlabeled dataset, then based on the Tn2th set of machining process parameters, we can find a set of machining process parameters (named the Smth set of machining process parameters) that is similar to the Tn2th set of machining process parameters among multiple sets of machining process parameters at the source pose. It's understandable that finding a similar set of machining process parameters can reduce the difficulty of transfer learning and improve its efficiency. For example, based on the spindle speed in the Tn2th set of machining process parameters, we can find the Smth set of machining process parameters among multiple sets of machining process parameters at the source pose, where the spindle speed in the Tn2th set of machining process parameters is the same as the spindle speed in the Smth set of machining process parameters. Then, the pre-trained deep regression model corresponding to the Sm group of processing parameters can be used as the initial model for transfer learning. Then, the labeled dataset corresponding to the Sm group of processing parameters, the labeled dataset corresponding to the Tn1 group of processing parameters, and the unlabeled dataset corresponding to the Tn2 group of processing parameters can be used for transfer learning to obtain the pre-trained deep regression model corresponding to the Tn2 group of processing parameters.

[0150] Table 1 Milling Machining Process Parameters

[0151]

[0152]

[0153] Taking the acquired Y-axis feed direction signal as an example, the data processing flow of the acquired joint angle, acceleration signal, and cutting force signal is described as follows:

[0154] First, the non-cutting segments of the data at the beginning and end are removed, along with outliers. A Butterworth low-pass filter with a cutoff frequency of 1000Hz is then used to preprocess the obtained acceleration and cutting force signals to eliminate noise frequencies, followed by downsampling. Finally, the time-series data is synchronized and processed according to discrete time relationships into a format conforming to the proposed network architecture.

[0155] Table 1 lists the milling experimental process parameters performed at the extreme positions of the four corners of the robot table. Cutting data in pose 1 is selected as the source pose data, and cutting data in poses 2, 3, and 4 are used as the target pose data. To demonstrate the effectiveness of the method provided in this application, the target pose 3, located at the diagonal of the table with the largest pose difference, is used as an example. In the target pose, a set of machining parameters identical to that in the source pose and a set of machining parameters different from that in the source pose are selected as test data.

[0156] Before model training, each sample point in the input monitoring signal of the source pose data is transformed from 7 dimensions to 7×7+5=54 dimensions according to the time lag order (i.e., the time series lag order). This transforms the input signal into the monitoring signals of the current and previous times, as well as the cutting force signal from the previous time. The output signal remains unchanged, still representing the cutting force value to be predicted at the current time. For the test data under the target pose, pseudo-labels are assigned using the PCA method to meet the network input dimension requirements.

[0157] The obtained source pose data and target pose data are input into, for example... Figure 3 The deep transfer regression model shown first establishes a deep network model, including a feature extraction network and a fully connected network, to extract time-series data features and perform regression prediction on these features. For example... Figure 3 As shown, a CNN feature extraction network and an FC regression prediction module are first trained using source pose data. The CNN feature extraction network has two convolutional layers with kernel sizes of 3 and 2, respectively. Batch normalization layers are inserted between each convolutional layer, and a max pooling layer is inserted after the second convolutional layer. The FC regression prediction module consists of four fully connected layers with 128, 64, 16, and 1 neurons, respectively. Then, the parameters of the pre-trained model are transferred to the target pose model with the same structure.

[0158] Based on the training process described above, the depth transfer regression prediction model established using target pose data selected in the linear and circular trajectory cutting experiments is evaluated. The processing method of the test set samples differs from that of the training data. The input signal of the test set is also 54-dimensional, but unlike the training data, it consists of the input signals at the current and previous times, as well as given initial random label data. After iteration, the input signal consists of the input signals at the current and previous times, as well as the predicted cutting force value at the previous time. All force prediction values ​​are obtained after all input signals are iterated.

[0159] Figure 5 This is one of the schematic diagrams showing the comparison between measured and predicted cutting force values ​​in linear trajectory milling provided in the embodiments of this application. Figure 5 The diagram shows the comparison between the measured and predicted cutting force values ​​when the source domain is pose 1 and the target domain is pose 3 in linear trajectory milling, with the first set of process parameters.

[0160] Figure 6 This is the second schematic diagram showing the comparison between measured and predicted cutting force values ​​in linear trajectory milling provided in the embodiments of this application. Figure 6 The diagram shows the comparison between the measured and predicted cutting force values ​​when the source domain is pose 1, the target domain is pose 3, and the third set of process parameters are used in linear trajectory milling.

[0161] Figure 7 This is one of the schematic diagrams showing the comparison between measured and predicted cutting force values ​​during arc trajectory milling provided in the embodiments of this application. Figure 7 The diagram shows the comparison between the measured and predicted cutting force values ​​when the source domain is pose 1 and the target domain is pose 3 in the third set of process parameters for arc trajectory milling.

[0162] Figure 8 This is the second schematic diagram comparing the measured and predicted cutting force values ​​during arc trajectory milling provided in the embodiments of this application. Figure 8 The comparison between the measured and predicted cutting force values ​​is shown when the source domain is pose 1, the target domain is pose 3, and the fourth set of process parameters is used in circular arc trajectory milling.

[0163] like Figure 5-8 As shown, the cutting force curve predicted by the method proposed in this application fits well within the tool's revolution cycle time, and the overall envelope matches well with the measured value envelope profile, that is, the peak value of the cutting force and the phase match well, which is of great significance for the monitoring of cutting force under different poses.

[0164] To verify the superiority of the proposed model, two methods were selected to compare the cutting force prediction performance. (1) Compared with a neural network trained directly using source pose data, it is called benchmark method-1; (2) Compared with a neural network trained directly using labeled data under target pose, it is called benchmark method-2.

[0165] Figure 9 This is one of the schematic diagrams comparing different cutting force prediction methods under linear trajectory milling provided in the embodiments of this application. Figure 9 The paper shows the cutting force measurement values ​​under linear trajectory milling with the source domain as pose 1, the target domain as pose 3, and the fourth set of process parameters. The comparison is between the predicted values ​​of the method proposed in this application (hereinafter referred to as the proposed method) and the predicted values ​​of the network directly trained using the source pose.

[0166] Figure 10 This is the second schematic diagram comparing different cutting force prediction methods under linear trajectory milling provided in the embodiments of this application. Figure 10 The paper presents the cutting force measurement values ​​under the fourth set of process parameters in linear trajectory milling with source domain as pose 1, target domain as pose 3, and the comparison between the predicted values ​​of the proposed method and the predicted values ​​of the network trained with labeled data under the target pose.

[0167] For linear trajectory milling, the source domain is selected as pose 1 (i.e., pose 1 is the source pose) and the target domain is pose 3 (i.e., pose 3 is the target pose). For circular trajectory milling, the source domain is selected as pose 1 and the target domain is pose 3. The results are compared with benchmark-1. Figure 9 and Figure 10 As shown in the figure. The results demonstrate that this method effectively improves prediction accuracy compared to networks trained directly using the source pose. The root mean square error of the predicted straight line and circular trajectory is reduced by 80.5% and 63.2%, respectively.

[0168] Figure 11 This is one of the schematic diagrams comparing different cutting force prediction methods for circular arc trajectory milling provided in the embodiments of this application. Figure 11 The paper presents the cutting force measurement values ​​when the source domain is pose 1, the target domain is pose 3, and the third set of process parameters are used in circular arc trajectory milling. It compares the predicted values ​​of the proposed method with the predicted values ​​of the network directly trained using the source pose.

[0169] Figure 12 This is the second schematic diagram comparing different cutting force prediction methods for circular arc trajectory milling provided in the embodiments of this application. Figure 12The paper presents the cutting force measurement values ​​when the source domain is pose 1, the target domain is pose 3, and the third set of process parameters are used in circular arc trajectory milling. It compares the predicted values ​​of the proposed method with the predicted values ​​of the network trained with labeled data under the target pose.

[0170] We selected linear trajectory milling with the source domain as pose 1 and the target domain as pose 3, and circular trajectory milling with the source domain as pose 1 and the target domain as pose 3, and compared them with benchmark-2. The results are as follows. Figure 11 and Figure 12 As shown, compared with neural networks trained directly using labeled data under the target pose, the method proposed in this application has higher prediction accuracy, with the root mean square error values ​​of the predicted straight line and circular arc trajectories reduced by 59.0% and 52.0%, respectively.

[0171] The comparison results with the two methods show that the proposed method has higher prediction accuracy than both networks trained directly using the source pose and neural networks trained directly using labeled data under the target pose, verifying the effectiveness and advantages of the method proposed in this application.

[0172] The robot milling cutting force monitoring device considering pose-dependent dynamic characteristics provided in this application is described below. The robot milling cutting force monitoring device considering pose-dependent dynamic characteristics described below can be referred to in correspondence with the robot milling cutting force monitoring method considering pose-dependent dynamic characteristics described above.

[0173] Figure 13 This is a schematic diagram of the structure of the robot milling cutting force monitoring device considering pose-dependent dynamic characteristics provided in the embodiments of this application, as shown below. Figure 13 As shown, the device includes a monitoring data acquisition module 10 and a cutting force prediction module 20.

[0174] The monitoring data acquisition module 10 is used to acquire the robot's monitoring data, which includes joint angle monitoring values ​​and acceleration signal monitoring values.

[0175] The cutting force prediction module 20 is used to obtain the predicted value of cutting force based on the monitoring data and the depth regression model corresponding to the current pose;

[0176] Among them, the deep regression model is used to predict the cutting force based on the joint angle monitoring value and the acceleration signal monitoring value. The robot has multiple poses, including a source pose and at least one target pose. The deep regression model corresponding to the source pose is obtained by supervised training based on the samples and corresponding labels at the source pose. The deep regression model corresponding to the target pose is obtained by transfer learning based on the deep regression model corresponding to the source pose.

[0177] It is understood that the detailed functional implementation of each of the above units / modules can be found in the description in the aforementioned method embodiments, and will not be repeated here.

[0178] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0179] Based on the methods in the above embodiments, this application provides an electronic device. Figure 14 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 14 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the methods in the above embodiments.

[0180] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0181] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0182] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0183] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0184] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0185] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0186] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0187] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements 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 monitoring cutting force in robot milling considering pose-dependent dynamic characteristics, characterized in that, include: Acquire monitoring data for the robot, including joint angle monitoring values ​​and acceleration signal monitoring values; Based on the monitoring data and the depth regression model corresponding to the current pose, the predicted value of the cutting force is obtained; The deep regression model is used to predict cutting force based on joint angle monitoring values ​​and acceleration signal monitoring values. The robot has multiple poses, including a source pose and at least one target pose. The deep regression model corresponding to the source pose is obtained through supervised training based on samples and corresponding labels at the source pose. The deep regression model corresponding to the target pose is obtained through transfer learning based on the deep regression model corresponding to the source pose. It also includes obtaining the depth regression model corresponding to the target pose through the following steps: Based on the labeled monitoring signal samples of the target pose and the corresponding cutting force signal labels, a labeled dataset of the target pose is constructed, and based on the unlabeled monitoring signal samples of the target pose and the corresponding pseudo-labels, an unlabeled dataset of the target pose is constructed. The pseudo-labels are obtained by performing principal component analysis on the unlabeled monitoring signal samples. Based on the labeled dataset of the target pose, the unlabeled dataset of the target pose, and the deep regression model corresponding to the source pose, transfer learning is performed iteratively to obtain the deep regression model corresponding to the target pose. The transfer learning iterative training includes updating model parameters by minimizing the calculated value of the following weighted ensemble loss function: ； in, The regression loss is the sum of the labeled monitoring signal samples for the target pose. This represents the difference in conditional embedding distribution between the monitoring signal samples of the source pose and the unlabeled monitoring signal samples of the target pose. This represents the difference in conditional embedding distribution between labeled and unlabeled monitoring signal samples representing the target pose. and It is a trade-off parameter.

2. The robot milling cutting force monitoring method considering pose-dependent dynamic characteristics according to claim 1, characterized in that, The deep regression model is obtained through a cascaded convolutional neural network feature extractor and regressor. The regressor includes multiple fully connected layers, with the number of fully connected layers being greater than or equal to three. These fully connected layers are connected in a cascaded manner. During transfer learning iterative training, the unlabeled and labeled monitoring signal samples of the target pose are processed by the same convolutional neural network feature extractor. Conversely, during transfer learning iterative training, the unlabeled and labeled monitoring signal samples of the target pose are processed by different regressors. The first conditional embedding distribution differs. Difference between the second conditional embedding distribution It was obtained through the following steps: The first dataset is constructed based on the features formed by the monitoring signal samples of the source pose through the second fully connected layer in the regressor of the deep regression model corresponding to the source pose and the labels corresponding to the monitoring signal samples of the source pose. The second dataset is constructed based on the features formed by processing the unlabeled monitoring signal samples of the target pose through the second fully connected layer in the corresponding regressor, and the predicted value of the cutting force signal obtained by processing the unlabeled monitoring signal samples of the target pose through the corresponding regressor. A third dataset is constructed based on the features formed by processing the labeled monitoring signal samples of the target pose through the second fully connected layer in the corresponding regressor and the labels corresponding to the labeled monitoring signal samples of the target pose. Analyze the differences between the conditional distribution features corresponding to the first dataset and the conditional distribution features corresponding to the second dataset in the reproducing kernel Hilbert space (RKHS) to obtain the first conditional embedding distribution difference; The difference between the conditional distribution features corresponding to the second dataset and the conditional distribution features corresponding to the third dataset in the reproducing kernel Hilbert space RKHS is obtained to obtain the difference in the second conditional embedding distribution.

3. The robot milling cutting force monitoring method considering pose-dependent dynamic characteristics according to claim 2, characterized in that, The convolutional neural network feature extractor is determined by cascading a first convolutional layer, a first normalization layer, a second convolutional layer, a pooling layer, and a second normalization layer.

4. The robot milling cutting force monitoring method considering pose-dependent dynamic characteristics according to any one of claims 1-3, characterized in that, The data input to the depth regression model includes monitoring data from the current and previous times, as well as the cutting force signal value from the previous time. The data output by the depth regression model is the predicted cutting force value at the current time. Obtaining the predicted cutting force value based on the monitoring data and the depth regression model corresponding to the current pose includes: Input the monitoring data of the current time and the previous time, as well as the cutting force signal value of the previous time, into the depth regression model corresponding to the current pose, and obtain the cutting force prediction value of the current time output by the depth regression model corresponding to the current pose.

5. The robot milling cutting force monitoring method considering pose-dependent dynamic characteristics according to any one of claims 1-3, characterized in that, It also includes obtaining the depth regression model corresponding to the source pose through the following steps: Based on the monitoring signal samples of the source pose and the corresponding cutting force signal labels, a source pose dataset is constructed; Based on the source pose dataset, an initial depth regression model is trained to obtain the depth regression model corresponding to the source pose.

6. A robot milling cutting force monitoring device considering pose-dependent dynamic characteristics, characterized in that, The robot milling cutting force monitoring method considering pose-dependent dynamic characteristics as described in claim 1 includes: The monitoring data acquisition module is used to acquire monitoring data of the robot, including joint angle monitoring values ​​and acceleration signal monitoring values. The cutting force prediction module is used to obtain the predicted cutting force value based on the monitoring data and the depth regression model corresponding to the current pose; The deep regression model is used to predict cutting force based on joint angle monitoring values ​​and acceleration signal monitoring values. The robot has multiple poses, including a source pose and at least one target pose. The deep regression model corresponding to the source pose is obtained through supervised training based on samples and corresponding labels at the source pose. The deep regression model corresponding to the target pose is obtained through transfer learning based on the deep regression model corresponding to the source pose.

7. An electronic device, characterized in that, include: At least one memory for storing computer programs; At least one processor is configured to execute a program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to perform the method as described in any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is run on the processor, it causes the processor to perform the method as described in any one of claims 1-5.

Images