Robot grinding prediction method and device based on dual-domain adaptive dual-task learning

Abstract

The application belongs to the technical field of robot grinding, and particularly discloses a robot grinding prediction method and equipment based on dual-domain adaptive double-task learning. The method comprises the following steps: the historical process parameters and the new process parameters of robot grinding are cyclically paired to form a "target-source" sample pair and a "source-source" sample pair; after feature conversion and merging are performed on the "target-source" sample pair and the "source-source" sample pair, "target-source" features and "source-source" features are obtained respectively; an empirical maximum mean difference model is constructed to calculate a domain adaptive loss; an overall loss function of a multi-gate hybrid expert model is constructed for double-task supervised optimization training to obtain a grinding prediction model; and the material removal depth and the average surface roughness of robot grinding are double-predicted based on the grinding prediction model. The application effectively balances the correlation and difference between the material removal depth and the average surface roughness, improves the sample quantity and diversity, and has high prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot grinding, and more specifically, relates to a robot grinding prediction method and device based on dual domain adaptive dual-task learning. Background Art

[0002] Robotic grinding technology has been recognized as an effective technology for finishing and processing difficult-to-machine materials due to its high flexibility, large operating space and high cost-effectiveness. High-precision manufacturing is the goal of robotic grinding technology, which can be achieved by simultaneously improving material removal capability and workpiece surface quality. Over the past few decades, many machine learning methods have been proposed to predict material removal depth and average surface roughness separately. However, these methods usually require training two different models, ignoring the potential correlation between them. Multi-task supervised learning (such as shared underlying method, tensor factorization method and multi-gate expert mixture method) is a more effective technique. Among the multi-task supervised learning methods, the multi-gate expert mixture method is able to balance the correlation and difference between tasks using parameters in a small number of training samples.

[0003] However, traditional multi-task supervised learning methods typically assume that the training and test data are independent and identically distributed. In fact, machining parameters in different intervals exhibit different data distributions, which often leads to prediction bias for unknown machining parameters when using traditional multi-task supervised learning methods. A reasonable approach is to utilize deep adaptation-based methods to alleviate the data distribution difference between existing and new machining parameters in robotic grinding, thereby achieving accurate joint prediction of material removal depth and average surface roughness when processing new machining parameters. However, the predictive ability of deep adaptation-based methods depends heavily on the number of samples available in the source and target domains. However, the cost of collecting and annotating data in the new machining parameter space for robotic grinding is too high, resulting in only a small number of labeled samples available for deep adaptation-based model training. When adopting deep adaptation-based methods, the limited test samples in the target domain may lead to inefficient training and inconsistent data distribution between training and test data.

[0004] Therefore, it is necessary to propose a new adaptive strategy to eliminate the distribution difference between training data and test data when a small number of test samples are provided. Summary of the Invention

[0005] In view of the above defects or improvement needs of the prior art, the present application provides a robot grinding prediction method and device based on dual-domain adaptive dual-task learning, wherein a dual-domain adaptive dual-task learning method is designed in correspondence with the characteristics of robot grinding itself and the distribution characteristics of new process parameters and historical process parameter data, a domain adaptation strategy is used to alleviate the problem of data distribution difference between the source domain and the target domain, and a dual-task prediction algorithm MMoE (Multi-Gated Mixture-of-Experts Model) is used to enhance the domain adaptation capability when only a small amount of labeled target samples are available for task supervised learning, so as to realize joint prediction of material removal depth and average surface roughness.

[0006] To achieve the above-mentioned purpose, according to one aspect of the present application, a robot grinding prediction method based on dual-domain adaptive dual-task learning is provided, comprising the following steps:

[0007] Step one, the historical process parameters and the corresponding new process parameters of robot grinding are cyclically paired to form "target-source" sample pairs and "source-source" sample pairs, and after feature conversion and merging of the "target-source" sample pairs and the "source-source" sample pairs, "target-source" features and "source-source" features are obtained respectively;

[0008] Step two, an experience maximum mean difference model between the "target-source" features and the "source-source" features is constructed, and a domain adaptation loss between the "target-source" features and the "source-source" features is calculated;

[0009] Step three, based on the domain adaptation loss, an overall loss function of a multi-gated mixture-of-experts model is constructed, the "target-source" features and the "source-source" features are input into the multi-gated mixture-of-experts model for dual-task supervised optimization training, so as to obtain a grinding prediction model;

[0010] Step four, based on the grinding prediction model, the material removal depth and the average surface roughness of robot grinding are double-predicted.

[0011] As a further preferred, in step one, a domain adaptation strategy is used to align the data between the source domain and the target domain, and after the "target-source" sample pairs and the "source-source" sample pairs are input into a feature converter for conversion and merging, "target-source" features and "source-source" features are obtained respectively.

[0012] As a further preferred, in step one, in order to learn the shared knowledge between MRD and Ra, two fully connected layers FC1 and FC2 are used for feature transformation;

[0013] The merging layer of the feature transformer is used to merge the transformed features of the paired data to reduce the domain adaptation loss between the “target-source” sample pairs and the “source-source” sample pairs.

[0014] As a further preferred embodiment, in step 2, the domain adaptive loss calculation model includes:

[0015]

[0016] Among them, L d represents the domain adaptation loss between the merged features of the target-source sample pair and the source-source sample pair, [·,·] represents the concatenation of the two vectors, h(·) is the feature conversion network that maps the input data to the shared feature space, and X s is the source feature, X t is the target feature, x s is the source domain sample, x t is the target domain sample.

[0017] As a further preference, the Gaussian kernel is used as the kernel function in the adaptive loss calculation model, which is defined as follows:

[0018] φ(x i ,x j )=exp(-||x i -x j || 2 / (2σ 2 ))

[0019] Here, σ is a hyperparameter representing the width of the Gaussian kernel.

[0020] As a further preference, the output of the kth (k=1, 2, . . . , 1) task in the multi-gated hybrid expert model is calculated as follows:

[0021]

[0022] Among them, x represents the input data or its transformed features, s is the number of expert networks, and f i (·) represents the mapping function of the i-th expert network, is the output of the k-th gating network, is the weight matrix, d is the dimension of x, represents the mapping function of the k-th tower network, y k corresponds to the output of the tower network.

[0023] As a further preference, the Adam optimization algorithm is used to minimize the overall loss function of the multi-gated hybrid expert model and update the weights.

[0024] As a further preference, the overall loss function is expressed as follows:

[0025] L=L r +λL d ,

[0026] Among them, L r is the regression loss of labeled source and target samples, L d represents the domain adaptation loss between the merged features of the “target-source” sample pair and the “source-source” sample pair, and the parameter λ serves as a penalty term to balance the regression loss and the domain adaptation loss.

[0027] According to another aspect of the present invention, a robot grinding prediction system based on dual-domain adaptive dual-task learning is provided, comprising:

[0028] A first main control module is used to cyclically pair historical process parameters of robot grinding with corresponding new process parameters to form "target-source" sample pairs and "source-source" sample pairs, and perform feature conversion and merging on the "target-source" sample pairs and the "source-source" sample pairs to obtain "target-source" features and "source-source" features respectively;

[0029] The second main control module is used to build the empirical maximum mean difference model between the "target-source" features and the "source-source" features, and calculate the domain adaptation loss between the "target-source" features and the "source-source" features;

[0030] A third main control module is used to construct an overall loss function of a multi-gated hybrid expert model based on the domain adaptive loss, and input the "target-source" feature and the "source-source" feature into the multi-gated hybrid expert model for dual-task supervised optimization training to obtain a grinding prediction model;

[0031] The fourth main control module is used to perform dual prediction of the material removal depth and average surface roughness of robot grinding according to the grinding prediction model.

[0032] According to another aspect of the present invention, there is also provided an electronic device, comprising:

[0033] At least one processor, at least one memory and a communication interface; wherein,

[0034] The processor, memory and communication interface communicate with each other;

[0035] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute a robot grinding prediction method based on dual-domain adaptive dual-task learning according to any of the above embodiments or a combination of multiple embodiments.

[0036] According to another aspect of the present invention, a non-transitory computer-readable storage medium is also provided, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute a robot grinding prediction method based on dual domain adaptive dual-task learning according to any of the above embodiments or a combination of multiple embodiments.

[0037] In general, the above technical solutions conceived by the present invention have the following technical advantages compared with the existing technology:

[0038] 1. This paper proposes a dual-domain adaptive dual-task learning method for the joint prediction of material removal depth and average surface roughness in robotic grinding. The training process of this method is divided into two main stages. First, the historical process parameters and new processing parameters in robotic grinding are cyclically paired to form "target-source" sample pairs and "source-source" sample pairs. Subsequently, these two types of sample pairs are input into a feature converter for domain adaptation and supervised dual-task model training, thereby achieving joint prediction of material removal depth and average surface roughness.

[0039] 2. The present invention proposes a domain adaptation strategy to make the data distribution between training data and test data correspond, so as to eliminate the negative impact of the data distribution difference between historical process parameters (training data) and new process parameters (test data) during training.

[0040] 3. The dual strategy of the present invention can be regarded as a data augmentation method that expands a limited number of target samples by cyclically grouping target samples and source samples. Compared with traditional data augmentation methods (such as generative adversarial networks (GANs), synthetic minority oversampling techniques (SMOTE), and adaptive synthesis (ADASYN) algorithms), the proposed dual strategy can increase the sample size using a small number of samples (even as few as one sample). Compared with the sample replication method, the proposed pairing strategy can significantly improve the diversity of enhanced target samples by combining a limited number of target samples with a sufficient number of source samples, thereby enhancing the ability of domain adaptation when there are fewer target samples.

[0041] 4. This paper utilizes the Multi-MoE (MMoE) to perform dual-task predictions for material removal capability and workpiece surface quality. The MMoE models the relationship between tasks in a weighted manner. If the provided tasks are less relevant, the shared expert network is penalized. As a result, the gated networks for these tasks learn to select different expert networks, effectively balancing the correlation and disparity between material removal depth and average surface roughness. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of a robot grinding prediction method based on dual-domain adaptive dual-task learning provided by an embodiment of the present invention;

[0043] Figure 2 1 is a network diagram of a robot grinding prediction method based on dual-domain adaptive dual-task learning according to the present invention;

[0044] Figure 3 Schematic diagram of the MMoE method provided by an embodiment of the present invention;

[0045] Figure 4 Schematic diagram of the joint prediction results of MRD and Ra using the PW-DA-DTL method under different normal forces provided by an embodiment of the present invention;

[0046] Figure 5 3 is a schematic diagram of the joint prediction results of MRD and Ra of the PW-DA-DTL method considering different robot feed speeds provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0048] Example 1

[0049] like Figure 1 and Figure 2 As shown, the dual-task learning method based on dual domain adaptation of the present invention mainly includes the following steps:

[0050] (1) First, the historical process parameters and new machining parameters in robotic grinding are cyclically paired to form "target-source" sample pairs and "source-source" sample pairs. A domain adaptation (DA) strategy is used to alleviate the problem of data distribution misalignment between the source domain and the target domain. The two sample pairs (represented as "target-source" sample pairs and "source-source" sample pairs) are simultaneously input into the feature converter, and the converted features are merged to obtain "target-source" and "source-source" features respectively. In this embodiment, the new machining parameters refer to the machining parameters obtained under the new normal force and new robot feed rate.

[0051] (2) The empirical maximum mean difference (MMD) is used to characterize the distribution difference between the source domain and the target domain. The domain adaptation loss is the empirical maximum mean difference between the combined features of the "target-source" sample pair and the "source-source" sample pair. The calculation of the domain adaptation loss is converted into the following form:

[0052]

[0053] Where [·,·] represents the concatenation of two vectors, and h(·) is a feature transformation network that maps the input data to a shared feature space. The Gaussian kernel is used as the kernel function, which is defined as follows:

[0054] φ(x i ,x j )=exp(-||x i -x j || 2 / (2σ 2 ))

[0055] Here, σ is a hyperparameter representing the width of the Gaussian kernel.

[0056] (3) Further, Figure 3 As shown, the above features are input into MMoE for supervised training. The output of the kth (k=1, 2, ···, l) task in the MMoE method is calculated as follows:

[0057]

[0058] Where x represents the input data or its transformed features, s is the number of expert networks, and f i (·) represents the mapping function of the i-th expert network. is the output of the k-th gating network, where is the weight matrix, and d represents the dimension of x. represents the mapping function of the k-th tower network, y k corresponds to the output of the tower network.

[0059] (4) Furthermore, the Adam optimization algorithm is used to minimize the overall loss function and update the weight θ. The overall loss function is expressed as follows:

[0060] L=L r +λL d ,

[0061] Among them L r L is the regression loss (calculated by mean square error (MSE)) of labeled source and target samples. d represents the domain adaptation loss between the merged features of the target-source sample pair and the source-source sample pair. The parameter λ acts as a penalty term to balance the regression loss and the domain adaptation loss.

[0062] In order to test the health status prediction effect of the regression meta-learning framework of the present invention under small sample conditions, an industrial robot and a belt grinder equipped with a six-dimensional force sensor were tested. The test workpiece was a unidirectional carbon fiber reinforced composite material. In this experiment, four key control parameters (i.e., normal force F n , robot feed speed V w, belt linear speed V s and the equivalent radius of the abrasive grain ) to study the MRD and R a The controlled processing parameters and their selected values ​​are recorded in the table below. The controlled parameters include: normal force F n , feed speed V w , feed speed V w , belt linear speed V s , equivalent radius of abrasive particles

[0063] After randomly combining the four controlled machining parameters, orthogonal grinding experiments with the specified group of parameters were conducted to evaluate the performance of RBG, and the corresponding MRD and Ra were measured as labels. n and V w In order to verify the capability of the proposed PW-DA-DTL method for unknown processing parameters, two different verification experiments were conducted based on the F n and V w The value of is used to divide the training data set and the test data set in a cross-validation manner. n For example, in each cross-validation iteration, a specific F n The samples with the same value are used as the test dataset, and the rest of the samples are used as the training dataset. In this experiment, only a labeled target sample is randomly selected from the target dataset to train the PW-DA-DTL method, hoping to reduce the dependence on the target sample.

[0064] First, the historical and new machining parameters from robotic grinding are cyclically paired to form "target-source" and "source-source" sample pairs. A domain adaptation (DA) strategy is used to mitigate the data distribution misalignment between the source and target domains. The two sample pairs (represented as "target-source" and "source-source") are simultaneously input into a feature converter. The converted features are merged to generate "target-source" and "source-source" features, respectively.

[0065] Furthermore, the empirical maximum mean difference (MMD) is used to characterize the distribution difference between the source domain and the target domain. The domain adaptation loss is the empirical maximum mean difference between the combined features of the "target-source" sample pair and the "source-source" sample pair. The calculation of the domain adaptation loss is converted into the following form:

[0066]

[0067] Where [·,·] represents the concatenation of two vectors, and h(·) is a feature transformation network that maps the input data to a shared feature space. The Gaussian kernel is used as the kernel function, which is defined as follows:

[0068] φ(xi ,x j )=exp(-||x i -x j || 2 / (2σ 2 ))

[0069] Here, σ is a hyperparameter representing the width of the Gaussian kernel.

[0070] Furthermore, the above features are input into MMoE for supervised training. The output of the kth (k=1, 2, ···, l) task in the MMoE method is calculated as follows:

[0071]

[0072] Where x represents the input data or its transformed features, s is the number of expert networks, and f i (·) represents the mapping function of the i-th expert network. is the output of the k-th gating network, where is the weight matrix, and d represents the dimension of x. represents the mapping function of the k-th tower network, y k corresponds to the output of the tower network.

[0073] Furthermore, the Adam optimization algorithm is used to minimize the overall loss function and update the weight θ. The overall loss function is expressed as follows:

[0074] L=L r +λL d ,

[0075] Among them L r L is the regression loss (calculated by mean square error (MSE)) of labeled source and target samples. d represents the domain adaptation loss between the merged features of the target-source sample pair and the source-source sample pair. The parameter λ acts as a penalty term to balance the regression loss and the domain adaptation loss.

[0076] In the experiments, the network inputs are target-source or source-source pairs that control processing parameters. To learn the shared knowledge between MRD and Ra, two fully connected layers (FC1 and FC2) are used to transform features. A merging layer is used to combine the transformed features of the paired data to reduce the domain adaptation loss between target-source and source-source pairs.

[0077] Due to the weights transferred from the standard MMoE method, the training cycle is set to 100. The test process is to input the target data of the four machining parameters into the trained PW-DA-DTL model, and the corresponding output is the joint predicted MRD and Ra. The supervised MSE loss between the observed label and the predicted label is minimized using the Adam optimizer and trained from scratch. When using the new normal force to jointly predict MRD and Ra, the PW-DA-DTL method performs better than the other two methods at different test F n The MRD and Ra co-prediction results under showed consistency between the predicted and observed values, such as Figure 4 As shown; the results of joint prediction using robot feed speed are shown as Figure 5 The evaluation metrics for MRD and Ra co-prediction using the proposed PW-DA-DTL method, shared bottom method, and MMoE method are listed in the table below.

[0078]

[0079] Compared with the shared bottom method and MMoE method, the prediction accuracy of the PW-DA-DTL method is further improved. The above MRD and Ra joint prediction results prove that the PW-DA-DTL method is effective for various F n This highlights the flexibility and effectiveness of the proposed method in predicting new processing parameters of RBG.

[0080] In the joint prediction of MRD and Ra with robot feed rate. PW-DA-DTL method has different V w The combined MRD and Ra prediction results under the same conditions also showed significant consistency between the predicted and observed values. The following table lists the evaluation metrics for combined MRD and Ra prediction using the proposed PW-DA-DTL method, the shared floor method, and the MMoE method.

[0081]

[0082] The above joint prediction results of MRD and Ra are consistent with the joint prediction results of normal force, which also reflects the superiority of the PW-DA-DTL method.

[0083] That is, the method proposed in the present invention is used to jointly predict the material removal depth and average surface roughness under new process parameters of robotic grinding. This method uses multiple hybrid experts to balance the correlation and difference between the material removal depth and the average surface roughness to achieve the purpose of high-precision dual-task prediction. In order to eliminate the negative impact of the data distribution difference between historical process parameters (training data) and new process parameters (test data) during training, a domain adaptation strategy is proposed to make the data distribution between training data and test data correspond. In order to solve the problem that the data distribution cannot be fully aligned due to insufficient sample size of new process parameters, a dual strategy is developed to significantly improve the number and diversity of samples. The present invention can simultaneously and accurately predict the material removal depth and average surface roughness. The present invention has been tested and verified on an independently developed robotic grinding system. Experimental results show that compared with traditional methods, this invention shows obvious superiority in joint prediction.

[0084] Example 2

[0085] Based on any of the above embodiments or a combination of multiple embodiments, the implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functionality. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention can be encapsulated into various modules. Based on this reality, on the basis of the above embodiments, an embodiment of the present invention provides a robot grinding prediction system with dual domain adaptive dual-task learning, including:

[0086] The first main control module is used to cyclically pair the historical process parameters of robot grinding with the corresponding new process parameters to form "target-source" sample pairs and "source-source" sample pairs, and after feature conversion and merging of the "target-source" sample pairs and "source-source" sample pairs, obtain "target-source" features and "source-source" features respectively; the second main control module is used to construct an empirical maximum mean difference model between the "target-source" features and the "source-source" features, and calculate the domain adaptive loss between the "target-source" features and the "source-source" features; the third main control module is used to construct the overall loss function of the multi-gated hybrid expert model based on the domain adaptive loss, and input the "target-source" features and the "source-source" features into the multi-gated hybrid expert model for dual-task supervised optimization training to obtain a grinding prediction model; the fourth main control module is used to perform dual prediction of the material removal depth and average surface roughness of robot grinding based on the grinding prediction model.

[0087] Example 3

[0088] Based on any of the above embodiments or a combination of multiple embodiments, the method of this embodiment is implemented by relying on an electronic device, so it is necessary to introduce the relevant electronic device. For this purpose, an embodiment of the present invention provides an electronic device, which includes: at least one processor (processor), a communication interface (CommunicationsInterface), at least one memory (memory) and a communication bus, wherein at least one processor, a communication interface, and at least one memory communicate with each other through the communication bus. At least one processor can call the logic instructions in at least one memory to execute all or part of the steps of the method provided by the aforementioned various method embodiments.

[0089] In addition, the logic instructions in the at least one memory mentioned above can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each method embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0090] 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, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0091] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0092] The flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. Based on this understanding, each box in the flowchart or block diagram can represent a module, program segment or part of the code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, or sometimes in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0093] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A robot grinding prediction method based on dual-domain adaptive dual-task learning, characterized in that: The following steps are involved: Step 1: cyclically pair the historical process parameters of robot grinding with the corresponding new process parameters to form "target-source" sample pairs and "source-source" sample pairs. After feature conversion and merging of the "target-source" sample pairs and "source-source" sample pairs, the "target-source" feature and "source-source" feature are obtained respectively. In step 1, a domain adaptation strategy is used to align the data between the source domain and the target domain. The target-source sample pairs and source-source samples are input into the feature converter for conversion and merging to obtain target-source features and source-source features respectively. Step 2: Build an empirical maximum mean difference model between the target-source feature and the source-source feature, and calculate the domain adaptation loss between the target-source feature and the source-source feature. Step 3: constructing an overall loss function of a multi-gated hybrid expert model based on the domain adaptive loss, inputting the "target-source" feature and the "source-source" feature into the multi-gated hybrid expert model for dual-task supervised optimization training to obtain a grinding prediction model; Step 4: Dual prediction of the material removal depth and average surface roughness of robot grinding is performed based on the grinding prediction model.

2. The robot grinding prediction method based on dual domain adaptive dual-task learning according to claim 1 is characterized in that: In step 1, in order to learn the shared knowledge between MRD and Ra, two fully connected layers FC1 and FC2 are used for feature transformation; The merging layer of the feature converter is used to merge the transformed features of the paired data to reduce the domain adaptation loss between the "target-source" sample pairs and the "source-source" sample pairs.

3. The robot grinding prediction method based on dual domain adaptive dual-task learning according to claim 1 is characterized in that: In step 2, the domain adaptive loss calculation model includes: Among them, L d represents the domain adaptation loss between the merged features of the target-source sample pair and the source-source sample pair, [·,·] represents the concatenation of the two vectors, h(·) is the feature conversion network that maps the input data to the shared feature space, and X s is the source feature, X t is the target feature, x s is the source domain sample, x t is the target domain sample.

4. The robot grinding prediction method based on dual domain adaptive dual-task learning according to claim 3 is characterized in that: The Gaussian kernel is used as the kernel function in the adaptive loss calculation model, which is defined as follows: φ(x i ,x j )=exp(-||x i -x j || 2 / (2σ 2 )) Here, σ is a hyperparameter representing the width of the Gaussian kernel.

5. The robot grinding prediction method based on dual domain adaptive dual-task learning according to claim 3 is characterized in that: The output of the kth task in the multi-gated hybrid expert model is calculated as follows: Among them, x represents the input data or its transformed features, s is the number of expert networks, and f i (·) represents the mapping function of the i-th expert network, is the output of the k-th gating network, is the weight matrix, d is the dimension of x, represents the mapping function of the k-th tower network, y k Corresponding to the output of the tower network, k = 1, 2,…, n.

6. The robot grinding prediction method based on dual domain adaptive dual-task learning according to claim 1 is characterized in that: The Adam optimization algorithm is used to minimize the overall loss function of the multi-gated hybrid expert model and update the weights.

7. The robot grinding prediction method based on dual domain adaptive dual-task learning according to claim 6 is characterized in that: The overall loss function is expressed as follows: L=L r +λL d Among them, L r is the regression loss of the labeled source sample target sample, L d represents the domain adaptation loss between the merged features of the "target-source" sample pair and the "source-source" sample pair, and the parameter λ serves as a penalty term to balance the regression loss and the domain adaptation loss.

8. A robot grinding prediction system based on dual domain adaptive dual-task learning, characterized by: include: a first main control module, configured to cyclically pair historical process parameters of robot grinding with corresponding new process parameters to form "target-source" sample pairs and "source-source" sample pairs, and perform feature conversion and merging on the "target-source" sample pairs and the "source-source" sample pairs to obtain "target-source" features and "source-source" features, respectively; wherein a domain adaptation strategy is adopted to align data between the source domain and the target domain, and the "target-source" sample pairs and the "source-source" samples are input into a feature converter for conversion and merging to obtain "target-source" features and "source-source" features, respectively; The second main control module is used to build the empirical maximum mean difference model between the "target-source" features and the "source-source" features, and calculate the domain adaptation loss between the "target-source" features and the "source-source" features; A third main control module is configured to construct an overall loss function of a multi-gated hybrid expert model based on the domain adaptive loss, input the target-source feature and the source-source feature into the multi-gated hybrid expert model for dual-task supervised optimization training to obtain a grinding prediction model; The fourth main control module is used to perform dual prediction of the material removal depth and average surface roughness of robot grinding according to the grinding prediction model.

9. An electronic device, characterized in that: include: At least one processor, at least one memory and a communication interface; wherein, The processor, memory and communication interface communicate with each other; The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the robot grinding prediction method based on dual domain adaptive dual-task learning according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the robot grinding prediction method based on dual domain adaptive dual-task learning according to any one of claims 1 to 7.

Images