Welding industrial robot parameter planning method and device

Abstract

The embodiment of the invention provides a welding industrial robot parameter planning method and device, and the method comprises the steps: constructing two task exclusive classification output layers corresponding to an industrial robot classification model and a machining object classification model, and a shared bottom feature extraction network, a task loss function of the task exclusive classification output layer is optimized at the same time through a joint learning framework of multi-task learning, a shared bottom-layer feature extraction network is obtained, and the shared bottom-layer feature extraction network is fused to a welding parameter configuration model based on a neural network; a welding request sent by a user is received, the industrial robot brand and machining object information in the welding request is extracted and input into the welding parameter configuration model, corresponding optimal welding parameters are obtained through the welding parameter configuration model, and a welding machine control program of the industrial robot is generated according to the optimal welding parameters; the automation and accuracy of the industrial robot capable of executing the welding machine control program can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of industrial robots, and in particular to a method and device for parameter planning of a welding industrial robot. Background Art

[0002] Welding is an indispensable process in modern manufacturing, and the application of industrial robots has greatly improved the efficiency and quality of welding. However, since the welding process involves complex physical and chemical reactions and is affected by many factors, how to formulate the optimal welding parameters for different brands of industrial robots and different processing objects has always been a major challenge facing the industry.

[0003] Traditional welding parameter settings mainly rely on manual adjustments by experienced engineers, which is not only time-consuming and labor-intensive, but also difficult to cope with rapidly changing production needs. In addition, industrial robots of different brands have different kinematic characteristics and control systems, which further increases the difficulty of parameter setting.

[0004] In recent years, with the development of artificial intelligence technology, some researchers have begun to try to apply machine learning methods to welding parameter optimization. Among them, parameter prediction models based on historical data have become an important research direction. This method attempts to establish a relationship model between welding parameters and welding quality by analyzing a large amount of historical welding data. However, how to effectively process and utilize heterogeneous data from different brands of robots and different processing objects remains a challenge.

[0005] Convolutional neural networks (CNNs) have been introduced into the field of industrial robot control due to their excellent performance in image recognition and pattern classification. However, how to apply CNNs to welding parameter optimization, especially how to effectively extract and utilize robot kinematic features and welding machine control features, still needs further research.

[0006] As a classic machine learning algorithm, support vector machine (SVM) has advantages in processing small samples and high-dimensional data. In the field of welding, SVM is used for material classification and quality prediction. However, how to combine SVM with deep learning methods to better handle complex welding parameter optimization problems is a direction worth exploring.

[0007] Multi-task learning is a hot research direction in the field of machine learning in recent years. In welding parameter optimization, how to simultaneously consider two related but different tasks, industrial robot classification and processing object classification, and improve the generalization ability and prediction accuracy of the model through a multi-task learning framework is a challenging problem.

[0008] Feature extraction and representation learning play a key role in machine learning. For the problem of welding parameter optimization, how to design a shared underlying feature extraction network that can effectively capture the key features of the welding process is the key to improving model performance.

[0009] In addition, existing welding parameter optimization methods are often targeted at specific robot brands or processing objects, lacking versatility and scalability. How to design a universal framework that can adapt to different brands of robots and various processing objects is an important direction of current research.

[0010] Finally, how to effectively convert the optimized welding parameters into a welding machine control program executable by the industrial robot to ensure that the parameter optimization results can be accurately implemented in actual production is also a key issue that needs to be solved.

[0011] In summary, although welding parameter optimization technology has made certain progress, there are still many challenges in robot feature extraction, processing object classification, multi-task learning, feature representation, and conversion of parameters to programs, which require further research and innovation. Summary of the invention

[0012] In view of the problems in the prior art, the present application provides a welding industrial robot parameter planning method and device, which can effectively improve the automation and accuracy of the welding machine control program executable by the industrial robot.

[0013] In order to solve at least one of the above problems, the present application provides the following technical solutions:

[0014] In a first aspect, the present application provides a welding industrial robot parameter planning method, comprising:

[0015] Obtain historical welding control data of industrial robots of different brands, extract robot kinematic features and welding machine control features from the historical welding control data, and use them as first training data to obtain an industrial robot classification model through training with a preset convolutional neural network model;

[0016] Acquire historical welding processing data of different processing objects, extract processing object quality features and processing object material features from the historical welding processing data, and use them as second training data to obtain a processing object classification model through training with a preset support vector machine model;

[0017] Constructing two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, optimizing the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model simultaneously through a joint learning framework of multi-task learning, obtaining a shared underlying feature extraction network after model optimization, and fusing the shared underlying feature extraction network into a welding parameter configuration model based on a neural network;

[0018] A welding request sent by a user is received, the industrial robot brand and processing object information in the welding request are extracted and input into the welding parameter configuration model, the corresponding optimal welding parameters are obtained through the welding parameter configuration model, and a welding machine control program of the industrial robot is generated according to the optimal welding parameters.

[0019] Furthermore, the historical welding control data of industrial robots of different brands are obtained, robot kinematic features and welding machine control features are extracted from the historical welding control data, and the industrial robot classification model is obtained by training a preset convolutional neural network model as the first training data, including:

[0020] Acquire historical welding control data of industrial robots of different brands, wherein the historical welding control data includes at least one of current, voltage, joint angle, and end tool posture parameters, and extract robot kinematic characteristics and welding machine control characteristics from the historical welding control data through video tracking algorithm and machine vision detection, and use them as first training data;

[0021] By presetting a convolutional neural network model as an industrial robot classifier, the first training data is used as the input of the convolutional neural network model for model training to obtain an industrial robot classification model.

[0022] Furthermore, the historical welding processing data of different processing objects are obtained, and the processing object quality characteristics and processing object material characteristics are extracted from the historical welding processing data, and used as the second training data to train a preset support vector machine model to obtain a processing object classification model, including:

[0023] Acquire historical welding processing data of different processing objects, wherein the historical welding processing data includes at least one of material properties, size, shape, and surface quality parameters, and automatically extract potential quality features and material features of the processing objects from the historical welding processing data by an unsupervised learning autoencoder, and use the data as second training data;

[0024] A support vector machine model is preset as a processing object classifier, and the second training data is used as an input of the support vector machine model for model training to obtain a processing object classification model.

[0025] Furthermore, the construction of two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model includes:

[0026] Constructing two task-specific classification output layers corresponding to the industrial robot classification task and the processing object classification task respectively, wherein the task-specific classification output layers are responsible for the classification prediction of the respective tasks;

[0027] A shared underlying feature extraction network is constructed as a shared component of multi-task learning, wherein the shared underlying feature extraction network is responsible for extracting a common feature representation from input data to provide support for two tasks.

[0028] Furthermore, the joint learning framework through multi-task learning simultaneously optimizes the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model to obtain a shared underlying feature extraction network after model optimization, including:

[0029] The prediction errors of the industrial robot classification task and the processing object classification task are calculated respectively, and the loss functions specific to the two tasks are obtained to update the shared underlying feature extraction network.

[0030] The loss functions of the two tasks are jointly optimized through the back-propagation algorithm, and a shared underlying feature extraction network is obtained that can learn common feature representations that are effective for both tasks.

[0031] Furthermore, the step of fusing the shared underlying feature extraction network into a neural network-based welding parameter configuration model comprises:

[0032] The optimized shared underlying feature extraction network is used as the underlying input feature extraction module of the welding parameter configuration model;

[0033] The welding parameter configuration model is end-to-end optimized and trained to obtain an optimized and trained welding parameter configuration model.

[0034] Further, extracting the industrial robot brand and processing object information in the welding request and inputting them into the welding parameter configuration model, and obtaining the corresponding optimal welding parameters through the welding parameter configuration model, includes:

[0035] Extracting the industrial robot brand and processing object information in the welding request and integrating them into a complete feature vector;

[0036] The feature vector is received as input by a welding parameter configuration model integrated with a shared feature extraction network to obtain optimal welding parameters output by the welding parameter configuration model, wherein the optimal welding parameters include at least one of current, voltage, and welding speed.

[0037] In a second aspect, the present application provides a welding industrial robot parameter planning device, comprising:

[0038] An industrial robot classification model building module is used to obtain historical welding control data of industrial robots of different brands, extract robot kinematic features and welding machine control features from the historical welding control data, and use them as first training data to obtain an industrial robot classification model through a preset convolutional neural network model training;

[0039] A processing object classification model building module is used to obtain historical welding processing data of different processing objects, extract processing object quality characteristics and processing object material characteristics from the historical welding processing data, and use them as second training data to obtain a processing object classification model through training with a preset support vector machine model;

[0040] A welding parameter configuration model construction module is used to construct two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimize the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network after model optimization, and fuse the shared underlying feature extraction network into a welding parameter configuration model based on a neural network;

[0041] The automatic welding control module is used to receive a welding request sent by a user, extract the industrial robot brand and processing object information in the welding request and input them into the welding parameter configuration model, obtain the corresponding optimal welding parameters through the welding parameter configuration model, and generate a welding machine control program for the industrial robot according to the optimal welding parameters.

[0042] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the welding industrial robot parameter planning method when executing the program.

[0043] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the welding industrial robot parameter planning method.

[0044] In a fifth aspect, the present application provides a computer program product, including a computer program / instruction, which implements the steps of the welding industrial robot parameter planning method when executed by a processor.

[0045] It can be seen from the above technical solution that the present application provides a welding industrial robot parameter planning method and device, which constructs two task-specific classification output layers corresponding to the industrial robot classification model and the processing object classification model and a shared underlying feature extraction network, and simultaneously optimizes the task loss function of the task-specific classification output layer through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network, and integrates the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; receives a welding request sent by a user, extracts the industrial robot brand and processing object information in the welding request and inputs them into the welding parameter configuration model, obtains the corresponding optimal welding parameters through the welding parameter configuration model, and generates the welding machine control program of the industrial robot according to the optimal welding parameters, thereby effectively improving the automation and accuracy of the welding machine control program executable by the industrial robot. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 This is one of the flow charts of the parameter planning method for a welding industrial robot in an embodiment of the present application;

[0048] Figure 2 This is a second flow chart of a parameter planning method for a welding industrial robot in an embodiment of the present application;

[0049] Figure 3 This is a third flow chart of the parameter planning method for a welding industrial robot in an embodiment of the present application;

[0050] Figure 4 This is a fourth flow chart of a parameter planning method for a welding industrial robot in an embodiment of the present application;

[0051] Figure 5 This is a fifth flow chart of the parameter planning method for a welding industrial robot in an embodiment of the present application;

[0052] Figure 6 This is a sixth flow chart of the parameter planning method for a welding industrial robot in an embodiment of the present application;

[0053] Figure 7 FIG7 is a flow chart of a parameter planning method for a welding industrial robot in an embodiment of the present application;

[0054] Figure 8 This is a structural diagram of a parameter planning device for a welding industrial robot in an embodiment of the present application;

[0055] Fig. 9 It is a schematic diagram of the structure of an electronic device in an embodiment of the present application.

[0056] Reference numerals:

[0057] Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver program storage unit 9144, antenna 9111, speaker 9131, microphone 9132. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0059] The acquisition, storage, use, and processing of data in the technical solution of this application comply with the relevant provisions of national laws and regulations.

[0060] In view of the problems existing in the prior art, the present application provides a welding industrial robot parameter planning method and device, which constructs two task-specific classification output layers corresponding to the industrial robot classification model and the processing object classification model and a shared underlying feature extraction network, and simultaneously optimizes the task loss function of the task-specific classification output layer through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network, and integrates the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; receives a welding request sent by a user, extracts the industrial robot brand and processing object information in the welding request and inputs them into the welding parameter configuration model, obtains the corresponding optimal welding parameters through the welding parameter configuration model, and generates the welding machine control program of the industrial robot according to the optimal welding parameters, thereby effectively improving the automation and accuracy of the welding machine control program executable by the industrial robot.

[0061] In order to effectively improve the automation and accuracy of the welding machine control program executable by the industrial robot, the present application provides an embodiment of a welding industrial robot parameter planning method, see Figure 1 The welding industrial robot parameter planning method specifically includes the following contents:

[0062] Step S101: acquiring historical welding control data of industrial robots of different brands, extracting robot kinematic features and welding machine control features from the historical welding control data, and using the data as first training data to obtain an industrial robot classification model through training with a preset convolutional neural network model;

[0063] Optionally, in this embodiment, step S101 is a key data processing and machine learning model training process, which aims to establish an intelligent system that can accurately classify industrial robots of different brands. The core of this step is to use historical welding control data to train a convolutional neural network model to achieve automatic classification of industrial robots.

[0064] First, the system needs to collect and organize historical welding control data of industrial robots of different brands. These data usually come from the factory's production logs, records of the robot control system, and measurement results of various sensors. During the data collection process, it is necessary to ensure the integrity, accuracy, and representativeness of the data so that subsequent feature extraction and model training can be built on a reliable foundation.

[0065] Next, the system extracts two types of key features from these historical data: robot kinematic features and welding machine control features. Robot kinematic features may include parameters such as joint angles, end effector position, speed, acceleration, etc. These features reflect the robot's motion ability and accuracy. Welding machine control features may include parameters such as current, voltage, wire feed speed, shielding gas flow, etc. These features reflect the control ability and stability of the welding process. The feature extraction process may involve a variety of technical methods such as signal processing, statistical analysis, and Fourier transform.

[0066] The extracted features are organized into a format suitable for convolutional neural network processing, usually in the form of multidimensional arrays or tensors. These processed data are called first training data, which contain the key information required for robot brand classification.

[0067] The preset convolutional neural network model is the core component of this step. Convolutional neural networks are particularly suitable for processing data with spatial or temporal correlation, such as images or time series data. In this scenario, the robot's motion trajectory and the changes in welding parameters can be regarded as a special kind of time series data. The convolution layer can automatically learn the local features in the data, the pooling layer can extract the most significant information from these features, and the fully connected layer is responsible for mapping these features to the final classification results.

[0068] The model training process usually uses optimization methods such as back propagation algorithm and stochastic gradient descent. During the training process, the model will continuously adjust its internal parameters to minimize the gap between the predicted results and the true labels. This process may require multiple iterations until the performance of the model reaches the expected level or no longer improves significantly.

[0069] This step solves several key problems faced by traditional industrial robot classification methods. First, it realizes automatic recognition of robot brands without human intervention or complex rule design. This greatly improves the efficiency and accuracy of classification, especially when faced with a large number of robots of different brands and models.

[0070] Second, this method can adapt to the subtle differences between different robots. Traditional methods may have difficulty capturing the small but critical differences in kinematics and control characteristics of robots from different brands, while deep learning models can automatically learn these detailed features, thereby improving classification accuracy.

[0071] Furthermore, this approach has strong generalization capabilities. Once the model is trained, it can handle unseen robot data, which is very useful for quickly identifying newly introduced robot models.

[0072] In terms of technical effects, this deep learning-based industrial robot classification method has brought significant improvements. First, it greatly improves the accuracy of classification. In practical applications, this method may achieve a classification accuracy of more than 95%, far exceeding traditional rule-based or simple statistical methods.

[0073] Secondly, this approach greatly improves the speed of classification. Once the model is trained, the classification process can be completed in milliseconds, which is critical in industrial environments that require real-time response.

[0074] Finally, this method lays the foundation for subsequent intelligent applications. Accurate robot classification is the prerequisite for realizing advanced functions such as personalized parameter optimization and predictive maintenance.

[0075] Step S102: acquiring historical welding processing data of different processing objects, extracting processing object quality features and processing object material features from the historical welding processing data, and using the data as second training data to obtain a processing object classification model through training with a preset support vector machine model;

[0076] Optionally, in this embodiment, step S102 is an important data processing and machine learning model training process, which aims to establish an intelligent system that can accurately classify different processing objects. The core of this step is to use historical welding processing data to train a support vector machine model to achieve automatic classification of processing objects.

[0077] First, the system needs to collect and organize historical welding processing data of different processing objects. This data may come from various sensors, quality inspection equipment, material analysis instruments, etc. on the production line. The data collection process needs to ensure the diversity and representativeness of the data to cover a variety of possible processing objects and processing conditions. This may include different types of metal materials, workpieces of different shapes and sizes, and processing results under different welding process parameters.

[0078] Next, the system extracts two types of key features from these historical data: processing object quality features and processing object material features. Quality features may include parameters such as weld width, depth, strength, and surface roughness, which reflect various aspects of welding quality. Material features may include chemical composition, crystal structure, hardness, thermal conductivity and other properties of the material, which determine the behavior of the material during the welding process. The feature extraction process may involve technologies from multiple fields such as image processing, spectral analysis, and materials science.

[0079] The extracted features are organized into a format suitable for support vector machine processing, usually in the form of high-dimensional feature vectors. These processed data are called second training data, which contain the key information required for the classification of the processed objects.

[0080] The preset support vector machine model is the core component of this step. Support vector machine is a powerful classification algorithm, which is particularly suitable for processing high-dimensional data and complex classification problems. Its core idea is to find an optimal hyperplane in high-dimensional space to separate data points of different categories. Support vector machine can handle nonlinear classification problems through kernel techniques, which makes it able to adapt to the complex nonlinear relationships that may exist in the classification of processing objects.

[0081] The model training process usually uses algorithms such as quadratic programming or sequential minimum optimization. During the training process, the model will continuously adjust the position and parameters of the hyperplane to maximize the interval between data points of different categories. For multi-category classification problems, a one-to-many or one-to-one strategy is usually used to convert the problem into multiple binary classification problems.

[0082] This step solves several key problems faced by traditional object classification methods. First, it realizes the automatic recognition and classification of processing objects without manual inspection or complex rule design. This greatly improves the efficiency and accuracy of classification, especially when faced with a large number of processing objects of different types and materials.

[0083] Secondly, this method can handle subtle differences between processing objects. Traditional methods may have difficulty distinguishing certain processing objects of similar materials or similar quality levels, while the support vector machine model can accurately classify based on small differences in high-dimensional feature space.

[0084] Furthermore, this method has strong generalization and robustness. The maximum margin characteristic of the support vector machine makes the model have good anti-interference ability to noise and outliers, which is very important for dealing with data fluctuations and anomalies that may exist in actual production environments.

[0085] In terms of technical effects, this support vector machine-based object classification method has brought significant improvements. First, it greatly improves the classification accuracy. In practical applications, this method may achieve a classification accuracy of more than 98%, especially when processing objects of similar materials or complex shapes, its advantages are more obvious.

[0086] Secondly, this method improves the consistency and reliability of classification. Compared with manual classification, the judgment criteria of machine learning models are consistent and will not be affected by human factors such as fatigue and emotions. This is crucial to ensure the consistency of product quality.

[0087] Finally, this approach provides a basis for intelligent manufacturing and quality control. Accurate classification of processing objects is a prerequisite for realizing advanced functions such as automated production line adjustment and personalized processing parameter optimization.

[0088] Step S103: constructing two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimizing the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network after model optimization, and integrating the shared underlying feature extraction network into a welding parameter configuration model based on a neural network;

[0089] Optionally, in this embodiment, step S103 is a complex and innovative machine learning model construction and optimization process, which aims to simultaneously improve the performance of industrial robot classification and processing object classification through a multi-task learning method, and apply the optimized feature extraction capability to welding parameter configuration. The core of this step is to use the joint learning framework of multi-task learning to achieve knowledge transfer and model optimization.

[0090] First, the system needs to build a neural network architecture consisting of three main parts: two task-specific classification output layers and a shared underlying feature extraction network. The shared underlying feature extraction network is the foundation of this architecture, which is responsible for extracting common, high-level feature representations from the raw input data. This network may consist of multiple layers of convolutional neural networks (CNNs) or recurrent neural networks (RNNs), depending on the nature of the input data. For example, if the input data contains time series information, a long short-term memory network (LSTM) or a gated recurrent unit (GRU) may be used.

[0091] The two task-specific classification output layers correspond to the industrial robot classification and processing object classification tasks respectively. These output layers are usually fully connected layers that receive features extracted by the shared underlying network and output the classification results of the corresponding tasks. Each output layer has its own loss function to measure the performance of the model on a specific task.

[0092] The core idea of ​​multi-task learning is to improve the generalization ability and performance of the model by learning multiple related tasks at the same time. In this scenario, although industrial robot classification and processing object classification are two different tasks, they are closely related in the welding process. By sharing the underlying feature extraction network, the model can learn more robust and general feature representations that are helpful for both tasks.

[0093] The optimization process of the joint learning framework usually adopts the strategy of alternating training or joint optimization. In each iteration, the system calculates the loss function of the two tasks and then updates the model parameters through the back-propagation algorithm. In this process, the parameters of the shared underlying network will be affected by both tasks, while the parameters of the task-specific output layer will be mainly affected by each task.

[0094] A key challenge in the optimization process is to balance the importance of different tasks. This is usually achieved by adjusting the weights of the loss function for each task. For example, if industrial robot classification is considered more important or challenging, it can be given a higher weight. Another approach is to use a dynamic weight adjustment strategy to automatically adjust the weights according to the learning progress of each task.

[0095] This step solves several key technical problems. First, it overcomes the limitations of single-task learning. Traditional single-task learning models may overfit the features of a specific task and ignore more general and essential features. Multi-task learning forces the model to learn more general and robust feature representations by learning multiple related tasks simultaneously.

[0096] Secondly, this method effectively utilizes limited training data. In actual industrial environments, it is often difficult and expensive to obtain a large amount of high-quality labeled data. Multi-task learning allows the model to learn from multiple related tasks, which is equivalent to increasing the number of effective training samples and helps improve the generalization ability of the model.

[0097] Furthermore, this approach enables knowledge transfer. By sharing the underlying network, the knowledge learned in one task can help improve the performance of another task. For example, the knowledge about the machine motion characteristics learned from the industrial robot classification task may be helpful for understanding the processing characteristics of different processing objects.

[0098] In terms of technical results, this multi-task learning-based approach brings significant improvements. First, it improves the overall performance of the model. Compared with training two classification models separately, the multi-task learning model is usually able to achieve better classification accuracy on both tasks. In some cases, the performance improvement can reach 5-10%.

[0099] Second, this approach improves the generalization ability of the model. Since more general feature representations are learned, the model performs better when faced with new, unseen data. This is particularly important for dealing with new types of robots or new materials that may appear in industrial environments.

[0100] Finally, this approach lays a solid foundation for the subsequent welding parameter configuration task. The optimized shared underlying feature extraction network contains rich, welding-related feature information, which is crucial for accurately configuring welding parameters.

[0101] The key step to achieve knowledge transfer is to integrate the optimized shared underlying feature extraction network into the neural network-based welding parameter configuration model. This process can be carried out through transfer learning, that is, the pre-trained shared network is used as part of the welding parameter configuration model, and then fine-tuned using data related to the welding parameter configuration. This method not only speeds up the training process of the welding parameter configuration model, but also improves the performance and stability of the model.

[0102] Step S104: receiving a welding request sent by a user, extracting the industrial robot brand and processing object information in the welding request and inputting them into the welding parameter configuration model, obtaining corresponding optimal welding parameters through the welding parameter configuration model, and generating a welding machine control program for the industrial robot according to the optimal welding parameters.

[0103] Optionally, in this embodiment, step S104 is the key execution stage of the entire welding parameter configuration system, which puts the model trained and optimized in the previous steps into practice, and realizes the automatic generation process from user request to specific welding control program. This step involves multiple technical links, including information extraction, model reasoning, parameter optimization and program generation, which reflects the core value of the intelligent manufacturing system.

[0104] First, the system needs to receive and parse the welding request sent by the user. This request may come from multiple channels, such as a production management system, an operator workstation, or a mobile device. The format of the request may be structured data (such as JSON or XML) or a natural language description. The system needs to be robust enough to handle a variety of possible input formats.

[0105] Information extraction is the first key technical point in this step. The system needs to accurately extract two core pieces of information from the welding request: the industrial robot brand and the processing object information. This may involve natural language processing (NLP) technology, especially when processing unstructured text input. For example, the system may use named entity recognition (NER) algorithms to identify brand names, and use keyword extraction and semantic analysis techniques to understand the description of the processing object.

[0106] The extracted information is then fed into the welding parameter configuration model optimized in the previous step. This model is the core of the entire system, integrating the functions of industrial robot classification, processing object classification, and parameter optimization. The input layer of the model receives the feature representation of the industrial robot brand and the processing object, which may include the robot's technical parameters (such as load capacity, accuracy, etc.) and the material properties of the processing object (such as composition, thickness, shape, etc.).

[0107] The welding parameter configuration model converts input information into optimal welding parameters through a complex neural network structure. This process may involve multiple hidden layers, each of which extracts and combines higher-level features. The output layer of the model gives a series of welding parameters, which may include current, voltage, welding speed, swing amplitude, shielding gas flow, etc.

[0108] The model's reasoning process is real-time, usually only milliseconds are needed to give results. This fast response capability is essential to meet the high efficiency requirements in modern manufacturing environments. At the same time, the model may also output the confidence level of the parameters or the possible parameter range, providing space for subsequent fine-tuning.

[0109] After obtaining the optimal welding parameters, the next challenge facing the system is how to convert these parameters into a specific industrial robot welding machine control program. This process involves complex industrial automation knowledge and needs to consider the programming language and control protocol of a specific robot brand.

[0110] The control program is usually generated using a templated approach. The system maintains a library of program templates for different brands and models of industrial robots. Based on the identified robot brand, the system selects the corresponding template and then inserts the optimized welding parameters into the specific position of the template. This process may also need to consider factors such as the robot's motion planning, safety restrictions, and process requirements.

[0111] The generated control program may include multiple parts: the initialization part sets the basic parameters of the robot; the main program part defines the welding path and specific welding operations; the subroutine part may contain some commonly used functional modules, such as weld tracking or automatic obstacle avoidance. The program may also include error handling and countermeasures for abnormal situations to improve the robustness of the system.

[0112] This step solves several key problems in the traditional welding parameter configuration and program writing process. First, it realizes the automation and intelligence of parameter configuration. The traditional method usually relies on experienced engineers to manually set parameters, which is not only time-consuming but also difficult to cope with rapidly changing production needs. The intelligent parameter configuration system can quickly respond to various new welding requirements, greatly improving the flexibility of production.

[0113] Secondly, this method improves the accuracy and consistency of parameter configuration. The machine learning model is trained based on a large amount of historical data and can capture subtle relationships that human experts may overlook, thereby providing the optimal parameter configuration under various conditions. This consistency is crucial to ensuring product quality.

[0114] Furthermore, the automated program generation process greatly reduces the risk of human error. Manually writing welding control programs is not only time-consuming but also prone to errors, especially when dealing with complex welding tasks. Automatically generated programs not only ensure syntactically correct but also ensure the accuracy of parameter settings.

[0115] In terms of technical effects, this step has brought significant improvements in many aspects. First, it has greatly improved production efficiency. The entire process from receiving a request to generating a control program may only take a few seconds, while traditional methods may take hours or even days. This high efficiency enables companies to respond quickly to customer needs and reduce production preparation time.

[0116] Secondly, this approach improves the consistency of product quality. Since parameter configuration is based on an optimized model rather than human subjective judgment, it can maintain a high degree of consistency between different batches and operators. This is particularly important for meeting strict quality standards, especially in high-demand industries such as aerospace and automotive.

[0117] Finally, this approach builds valuable knowledge assets for the company. Every successful welding operation can be recorded and used to further optimize the model. Over time, the performance of the system will continue to improve, forming a virtuous cycle.

[0118] From the above description, it can be seen that the welding industrial robot parameter planning method provided in the embodiment of the present application can construct two task-specific classification output layers corresponding to the industrial robot classification model and the processing object classification model and a shared underlying feature extraction network, and simultaneously optimize the task loss function of the task-specific classification output layer through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network, and integrate the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; receive a welding request sent by a user, extract the industrial robot brand and processing object information in the welding request and input it into the welding parameter configuration model, obtain the corresponding optimal welding parameters through the welding parameter configuration model, and generate the welding machine control program of the industrial robot according to the optimal welding parameters, thereby effectively improving the automation and accuracy of the welding machine control program executable by the industrial robot.

[0119] In one embodiment of the welding industrial robot parameter planning method of the present application, see Figure 2 , and can also include the following:

[0120] Step S201: acquiring historical welding control data of industrial robots of different brands, wherein the historical welding control data includes at least one of current, voltage, joint angle, and end tool posture parameters, and extracting robot kinematic characteristics and welding machine control characteristics from the historical welding control data through video tracking algorithm and machine vision detection, and using them as first training data;

[0121] Step S202: Preset a convolutional neural network model as an industrial robot classifier, use the first training data as input of the convolutional neural network model to perform model training, and obtain an industrial robot classification model.

[0122] Optionally, in this embodiment, steps S201 and S202 constitute a complete industrial robot classification model training process, involving key links such as data collection, feature extraction and model training. This process aims to establish an intelligent system that can accurately identify industrial robots of different brands, laying the foundation for subsequent welding parameter optimization.

[0123] First, the core of step S201 is to obtain and process historical welding control data. These data are the cornerstone of model training, and their quality and quantity directly affect the performance of the final model. Historical welding control data usually comes from the factory's production record system, the log file of the robot controller, or a specially set data acquisition device. These data contain rich information, such as current, voltage, joint angle, and end tool posture parameters.

[0124] Current and voltage data reflect the energy input during welding and are essential for identifying the welding characteristics of different robots. Robots of different brands may have different current control strategies and voltage regulation mechanisms, and these differences can be an important basis for classification.

[0125] Joint angle data provides direct information about the robot's kinematics. Each brand of robot may have unique mechanical structures and kinematic parameters, which are reflected in the change pattern of joint angles. For example, certain brands of robots may have characteristic joint angle combinations when performing specific actions.

[0126] The end tool posture parameters describe the position and posture of the welding gun or other end effector in three-dimensional space. These data reflect the robot's workspace characteristics and motion planning strategy. Robots of different brands may have different posture control algorithms.

[0127] The application of video tracking algorithms and machine vision detection is an innovation in this step. The robot's motion process is captured by a high-speed camera, and then computer vision technology is used to extract kinematic features. For example, the motion trajectory of each part of the robot can be tracked using an optical flow algorithm, or the key points of the robot can be identified and tracked using a target detection algorithm. This method can capture some subtle features that are difficult to obtain from traditional sensor data, such as the robot's vibration characteristics or dynamic response patterns.

[0128] Machine vision technology can also be used to analyze visual features of the welding process, such as weld morphology and molten pool dynamics. Robots of different brands may have different welding control strategies, and these differences may be reflected in the visual performance of the welding process. For example, by analyzing the shape and dynamic changes of the molten pool, the robot's welding control characteristics can be inferred.

[0129] Feature extraction is the key technical point in this step. Extracting meaningful features from raw data requires a combination of domain knowledge and advanced signal processing techniques. For example, Fourier transform can be used to analyze the spectral characteristics of current and voltage signals, principal component analysis (PCA) can be used to reduce the dimension of joint angle data, or wavelet transform can be used to capture the multi-scale features of time series data.

[0130] Step S202 focuses on the construction and training of the model. It is a wise decision to choose a convolutional neural network (CNN) as a classifier. CNN performs well in processing data with spatial or temporal local correlations and is very suitable for processing complex multidimensional data from industrial robots.

[0131] The architectural design of CNN requires careful consideration. A typical architecture may include multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layers are responsible for extracting local features, the pooling layers are used for dimensionality reduction and extracting main features, and the fully connected layers combine these features to make the final classification decision. It may also be necessary to add batch normalization layers to speed up training and improve the generalization ability of the model, as well as dropout layers to prevent overfitting.

[0132] During model training, we need to pay attention to the following key points:

[0133] Data preprocessing: Normalize the input data to ensure that different types of features (such as current and angle) are on the same scale.

[0134] Data enhancement: Increase the diversity of training data and improve the robustness of the model by adding noise, time warping, etc.

[0135] Learning rate scheduling: Use dynamic learning rate strategies such as learning rate decay or periodic learning rate adjustment to achieve better convergence.

[0136] Cross-validation: Use techniques such as k-fold cross-validation to evaluate the generalization ability of the model and avoid overfitting.

[0137] Model ensemble: Train multiple models and perform ensembles, such as bagging or boosting, to further improve classification performance.

[0138] This process solves several key technical problems. First, it overcomes the limitations of traditional robotic recognition methods that rely on manual labeling or simple rules. Through deep learning, the system is able to automatically learn complex feature combinations and capture subtle differences that human experts may overlook.

[0139] Secondly, this method improves the accuracy and robustness of recognition. Traditional methods may be susceptible to noise and abnormal data, while deep learning models are better able to handle various variations and interferences through training with large amounts of data.

[0140] Furthermore, this method realizes the automation and real-time nature of the recognition process. Once the model training is completed, it can be quickly deployed in the production environment to achieve real-time recognition of industrial robots.

[0141] In terms of technical effects, this deep learning-based method has brought significant improvements. First, it greatly improves the accuracy of recognition. In an ideal situation, this method may achieve a classification accuracy of more than 95%, far exceeding traditional methods.

[0142] Secondly, this approach improves the adaptability of the system. When a new model of robot is introduced, it only needs to collect a certain amount of new data and retrain the model to quickly adapt to new classification tasks. This flexibility is especially important in a rapidly changing manufacturing environment.

[0143] Finally, this method provides a reliable basis for subsequent parameter optimization and control. Accurate robot brand identification can help the system select the most suitable welding parameters and control strategies, thereby improving the overall welding quality and efficiency.

[0144] In one embodiment of the welding industrial robot parameter planning method of the present application, see Figure 3 , and can also include the following:

[0145] Step S301: acquiring historical welding processing data of different processing objects, wherein the historical welding processing data includes at least one of material properties, size, shape, and surface quality parameters, and automatically extracting potential quality features and material features of the processing objects from the historical welding processing data through an unsupervised learning autoencoder, and using the data as second training data;

[0146] Step S302: Preset a support vector machine model as a processing object classifier, use the second training data as input of the support vector machine model to perform model training, and obtain a processing object classification model.

[0147] Optionally, in this embodiment, steps S301 and S302 are intended to establish an efficient and accurate classification model for processing objects, which is of great significance for subsequent welding parameter optimization and quality control. This process combines unsupervised learning and supervised learning techniques, makes full use of historical data, automatically extracts key features, and finally builds a robust classification model.

[0148] First, we need to obtain a large amount of historical welding processing data. This data usually comes from multiple sources, including sensor records on the production line, inspection reports from the quality control department, specifications provided by material suppliers, and outputs from measuring equipment such as 3D scanners. The data content covers many aspects of the processing object, such as material properties (such as chemical composition, hardness, thermal conductivity, resistivity, etc.), size (geometric parameters such as length, width, thickness), shape (may be a simple flat plate, or a complex curved surface or special-shaped structure), and surface quality (such as roughness, coating type, degree of oxidation, etc.). These multi-dimensional data provide a rich information basis for subsequent feature extraction and classification.

[0149] Next, we use an unsupervised learning autoencoder to automatically extract the potential quality features and material characteristics of the processed object from these raw data. An autoencoder is a special neural network structure that attempts to learn a compressed representation (encoding) of the input data and then reconstruct the original input from this representation (decoding). In this embodiment, the structure of the autoencoder may include an input layer, multiple encoding hidden layers, a potential representation layer, multiple decoding hidden layers, and an output layer. For example, for an input containing 100 original features, the encoder may compress it to a 10-dimensional potential representation through a layer structure such as 50-25-10, and then the decoder attempts to reconstruct the original input through a 10-25-50-100 structure.

[0150] The training process of the autoencoder is carried out by minimizing the reconstruction error. Specifically, we can use the mean square error (MSE) as the loss function and adjust the network parameters through the back-propagation algorithm. During the training process, the encoder part gradually learns how to extract the key features of the processed object, while the decoder learns how to reconstruct the original data from these features. This process does not require manually annotated data, thus greatly reducing the workload of data preparation, and can also discover potential features that human experts may overlook.

[0151] After training, we use the encoder part to extract features. For each processed object, we input its original data into the encoder to obtain a low-dimensional potential representation. This potential representation becomes our second training data, which contains the key quality characteristics and material characteristics of the processed object. For example, for a processed object that originally has 100 features, we may compress it into a 10-dimensional vector, and each element in this vector represents a complex pattern or feature combination in the original data.

[0152] In step S302, we use a support vector machine (SVM) model as a processing object classifier. SVM is a powerful classification algorithm that is particularly suitable for processing high-dimensional data and complex classification boundaries. In this embodiment, we use features extracted from an autoencoder as input to the SVM. The core idea of ​​SVM is to find an optimal hyperplane in the feature space so that the sample intervals of different categories are maximized. For situations where linear separation may not be possible, SVM maps data to a higher-dimensional space through the kernel technique, where a linear separation hyperplane is found.

[0153] The training process of the SVM model includes selecting an appropriate kernel function (such as the RBF kernel), adjusting hyperparameters (such as the C parameter and the gamma parameter), and using cross-validation to evaluate model performance. During the training process, we may need to deal with class imbalance problems, such as using weight adjustment or oversampling techniques. After training, we evaluate the model performance on an independent test set. Evaluation indicators may include accuracy, precision, recall, and F1 score.

[0154] This process solves several key technical problems. First, it automates feature engineering. In traditional methods, feature engineering often relies on the experience of domain experts, which is time-consuming and may overlook some potentially important features. Through autoencoders, the system can automatically learn complex feature combinations, which not only reduces manual workload, but also may discover patterns that are difficult for human experts to detect. Second, it effectively handles the problem of high-dimensional data. The characteristics of the processed object may involve a large number of parameters, and directly using these original features may lead to dimensionality disasters. Autoencoders can compress these high-dimensional data into a low-dimensional latent space, which not only reduces the complexity of subsequent classification tasks, but also may remove noise in the data and improve the generalization ability of the model.

[0155] In addition, this method also handles the problem of nonlinear relationships very well. There may be complex nonlinear relationships between material properties and processing quality, which are difficult to capture with simple linear models. The nonlinear activation function of the autoencoder and the kernel trick of the SVM can handle this complexity very well. Finally, this method has good adaptability and scalability. When new materials or processing objects are introduced, the system can quickly adapt by retraining the autoencoder and SVM models without redesigning the entire system.

[0156] In terms of technical effects, this method has brought significant improvements. The first is a significant improvement in classification accuracy. Compared with traditional rule-based methods, this machine learning-based method can greatly improve the classification accuracy of processing objects. In a complex manufacturing environment, the accuracy may be increased from the original 80% to more than 95%. This high accuracy is crucial for subsequent welding parameter optimization and quality control. Secondly, this method provides better feature understanding. The potential features learned by the autoencoder can help engineers better understand which factors are most important for processing quality. This insight can guide subsequent process optimization and may even lead to new material design ideas.

[0157] In one embodiment of the welding industrial robot parameter planning method of the present application, see Figure 4 , and can also include the following:

[0158] Step S401: constructing two task-specific classification output layers corresponding to the industrial robot classification task and the processing object classification task respectively, wherein the task-specific classification output layers are responsible for classification prediction of the respective tasks;

[0159] Step S402: construct a shared underlying feature extraction network as a shared component of multi-task learning, wherein the shared underlying feature extraction network is responsible for extracting common feature representations from input data to provide support for two tasks.

[0160] Optionally, in this embodiment, steps S401 and S402 are intended to construct a multi-task learning model that can simultaneously handle two related but independent tasks, namely, industrial robot classification and processing object classification. This method can not only improve the efficiency and generalization ability of the model, but also utilize the correlation between tasks to improve the overall performance.

[0161] First, let's discuss step S401 in detail. In this step, we build dedicated classification output layers for the industrial robot classification task and the processing object classification task respectively. These two output layers are the last layers of the model and are directly responsible for the classification prediction of their respective tasks. For the industrial robot classification task, the output layer may be a fully connected layer with the number of neurons equal to the number of possible robot categories. For example, if we have 5 different types of industrial robots, then this output layer will have 5 neurons, one for each robot type. Similarly, for the processing object classification task, if we need to classify 10 types of processing objects, then the dedicated output layer for this task will have 10 neurons.

[0162] These output layers typically use a softmax activation function to convert the input into a probability distribution. For example, for an industrial robot classification task, the model might output [0.1, 0.05, 0.7, 0.1, 0.05], indicating that the input data most likely corresponds to the third class of robots with a probability of 70%. This design allows the model to make predictions for each task independently without interfering with each other.

[0163] Step S402 is the core of the multi-task learning model, which is to build a shared underlying feature extraction network. As a shared component of multi-task learning, this network is responsible for extracting common feature representations from the input data to support the two classification tasks. The shared network usually consists of multiple convolutional layers and pooling layers if the input data is an image. For non-image data, a fully connected layer or other suitable network structure may be used.

[0164] For example, we can design a shared network with three convolutional layers. The first layer may have 32 3x3 convolution kernels, the second layer 64 3x3 convolution kernels, and the third layer 128 3x3 convolution kernels. Each convolution layer is followed by a ReLU activation function and a max pooling layer. The role of this network is to gradually extract more and more abstract features from the raw input data. The bottom layer may capture simple edge and texture information, the middle layer may recognize more complex shapes and structures, and the upper layer may extract high-level features that are most useful for classification tasks.

[0165] The technical principles of this design are based on the following points: First, multi-task learning can exploit the correlation between tasks. Although industrial robot classification and processing object classification are two different tasks, they may share some basic features in some aspects. For example, both tasks may need to identify the basic shape, size or material of the object. By sharing the underlying network, the model can learn these common features more effectively. Secondly, sharing the network can reduce the risk of overfitting. Because the underlying network needs to meet the needs of two tasks at the same time, it is forced to learn a more general and robust feature representation instead of overfitting to a single task. Finally, this design can also improve computational efficiency. Compared with training a complete model for each task separately, sharing most of the parameters can significantly reduce the demand for computing resources.

[0166] This process solves several key technical problems. First, it solves how to effectively use limited training data. In industrial settings, it can be difficult or expensive to obtain large amounts of labeled data. Through multi-task learning, we can let the model learn from multiple related tasks, even if the amount of data for each task is not large, but together they can provide enough information to train a powerful model. Second, it solves the problem of how to balance task specificity and generality. The shared network extracts common features, while the task-specific output layer handles the needs of specific tasks. This structure ensures the generality of the model without losing flexibility.

[0167] In terms of technical results, this approach brings significant improvements. The first is the improvement in classification accuracy. Compared with single-task learning, multi-task learning models can usually achieve higher accuracy on both tasks. This is because the model can exploit the correlation between tasks and learn from more data. Secondly, this approach improves the generalization ability of the model. Because the underlying features are jointly optimized for multiple tasks, they tend to be more robust and can better cope with new, unseen data.

[0168] In addition, this approach can also speed up the convergence of the model. Because the underlying network receives gradient updates from multiple tasks at the same time, it can learn useful feature representations faster. In practical applications, this may mean shorter training time and less computing resource requirements. Finally, this approach provides better feature understanding. By analyzing the features learned by the shared network, we can better understand which features are important for multiple tasks, which may bring new insights and guide subsequent system optimization.

[0169] In one embodiment of the welding industrial robot parameter planning method of the present application, see Figure 5 , and can also include the following:

[0170] Step S501: respectively calculating the prediction errors of the industrial robot classification task and the processing object classification task, and obtaining loss functions specific to the two tasks to update the shared underlying feature extraction network;

[0171] Step S502: jointly optimize the loss functions of the two tasks through a back-propagation algorithm to obtain a shared underlying feature extraction network that can learn a common feature representation that is effective for both tasks.

[0172] Optionally, in this embodiment, steps S501 and S502 are the core links of model training and optimization. These two steps are intended to obtain a shared underlying feature extraction network that can effectively handle industrial robot classification and processing object classification by jointly optimizing the loss functions of the two classification tasks. This method can not only improve the overall performance of the model, but also make full use of the correlation between tasks and enhance the generalization ability of the model.

[0173] First, let's discuss step S501 in detail. In this step, we need to calculate the prediction errors of the industrial robot classification task and the processing object classification task respectively, and obtain the loss functions specific to the two tasks based on this. For classification tasks, we usually use the cross entropy loss function. Specifically, for the industrial robot classification task, assuming that we have m categories, its loss function can be expressed as the negative log-likelihood: L1 = -∑(y_i*log(p_i)), where y_i is the one-hot encoding of the true label and p_i is the probability distribution predicted by the model. Similarly, for the processing object classification task, if there are n categories, its loss function can be expressed as: L2 = -∑(z_j*log(q_j)), where z_j is the one-hot encoding of the true label and q_j is the probability distribution predicted by the model.

[0174] These two loss functions reflect the performance of the model on their respective tasks. A smaller loss value means that the model's prediction is closer to the true label. By calculating these loss functions, we can quantify the performance of the model on each task and provide guidance for subsequent optimization.

[0175] Next, in step S502, we use the back propagation algorithm to jointly optimize the loss functions of the two tasks. Back propagation is one of the most commonly used optimization algorithms in deep learning. It calculates the gradient of the loss function with respect to each parameter and then updates the parameters in the opposite direction of the gradient to minimize the loss function. In the case of multi-task learning, we need to consider the losses of the two tasks at the same time.

[0176] A common approach is to weight the sum of the losses of the two tasks: L_total = α*L1+β*L2, where α and β are weight coefficients used to balance the importance of the two tasks. These weights can be set based on the relative importance or difficulty of the tasks, or they can be automatically adjusted through hyperparameter optimization. For example, if we believe that the industrial robot classification task is more important or more difficult, we may give it a larger weight.

[0177] In actual implementation, we may use mini-batch stochastic gradient descent (mini-batch SGD) or its variants (such as Adam, RMSprop, etc.) for optimization. In each iteration, we randomly extract a small batch of samples from the training data, containing data from both tasks. Then, we feed this data into the model, calculate the loss of the two tasks, and calculate their weighted sum. Next, we use the back-propagation algorithm to calculate the gradient and update the model parameters.

[0178] A key point of this process is that the parameters of the shared underlying feature extraction network are jointly affected by the two tasks. For example, if a particular convolution kernel is important for both tasks, it will get a stronger update signal; on the contrary, if a feature is only useful for one task and not helpful or even harmful for the other task, it may be weakened or eliminated. This mechanism ensures that the shared network can learn a common feature representation that is effective for both tasks.

[0179] This approach solves several key technical problems. First, it solves the problem of how to effectively allocate model capacity among multiple related tasks. Traditional single-task learning may waste model capacity on task-specific features, while multi-task learning can more effectively utilize model capacity by sharing parameters. Second, it solves the problem of how to balance the needs of different tasks. By adjusting the weights of the loss function, we can flexibly control the model's attention to each task.

[0180] In addition, this method can effectively deal with the problem of data imbalance. In practical applications, the amount of training data for different tasks may vary greatly. For example, we may have a large amount of processing object data, but only a small amount of industrial robot data. Through multi-task learning, even if a task has less data, it can learn useful features from other tasks, thereby improving performance.

[0181] In terms of technical results, this approach has brought significant improvements. The first is the overall improvement in model performance. Compared with training two models separately, multi-task learning can usually achieve better performance on both tasks. This is because the model is able to learn from more data and take advantage of the correlation between tasks. For example, in practical applications, we may see the accuracy of industrial robot classification increase from 85% to 90%, and the accuracy of processing object classification increase from 88% to 93%.

[0182] Secondly, this approach improves the generalization ability of the model. Since the shared network needs to meet the needs of two tasks at the same time, it is forced to learn more general and robust feature representations. This means that the model will perform better when faced with new, unseen data. For example, when introducing new types of industrial robots or new processing objects, the multi-task model may perform better than the single-task model because it has learned more general features.

[0183] In addition, this approach can improve computational efficiency and resource utilization. Compared with training a complete model for each task separately, sharing most of the parameters can significantly reduce the demand for computing resources. In practical applications, this may mean faster training speed, lower hardware requirements, and easier model deployment and maintenance.

[0184] In one embodiment of the welding industrial robot parameter planning method of the present application, see Figure 6 , and can also include the following:

[0185] Step S601: using the optimized shared underlying feature extraction network as the underlying input feature extraction module of the welding parameter configuration model;

[0186] Step S602: performing end-to-end optimization training on the welding parameter configuration model to obtain an optimized and trained welding parameter configuration model.

[0187] Optionally, in this embodiment, steps S601 and S602 are key optimization links in the intelligent welding system, which are intended to further optimize the welding parameter configuration model by using the shared underlying feature extraction network obtained by multi-task learning. This process can not only improve the performance of the model, but also make full use of the previously learned common features to achieve more accurate and efficient welding parameter configuration.

[0188] First, let's discuss step S601 in detail. In this step, we use the optimized shared underlying feature extraction network as the underlying input feature extraction module of the welding parameter configuration model. This process is actually an application of transfer learning. The core idea of ​​transfer learning is to apply the knowledge and skills learned in one task or field to another related but different task or field. In this scenario, we transfer the feature extraction capabilities learned from the industrial robot classification and processing object classification tasks to the welding parameter configuration task.

[0189] Specifically, we can copy the weights of the shared underlying feature extraction network directly into the corresponding layers of the welding parameter configuration model. For example, if the shared network contains three convolutional layers and two fully connected layers, we can copy these layers and their weights directly into the welding parameter configuration model. The benefit of this is that we do not need to train the feature extraction part of the welding parameter configuration model from scratch, but can leverage the useful features that have already been learned.

[0190] The technical principle of this method is based on the following assumption: although the three tasks of industrial robot classification, processing object classification, and welding parameter configuration have different goals, they may have some commonalities in underlying features. For example, they may all need to recognize features such as the shape, size, and material of an object. By using a shared network optimized by multi-task learning, we can provide a good starting point for the welding parameter configuration task, allowing the model to learn task-related features faster and better.

[0191] Next, in step S602, we perform end-to-end optimization training on the welding parameter configuration model. This process involves several key technical points:

[0192] First, we need to prepare training data for the welding parameter configuration task. These data usually include input features (such as workpiece type, welding position, environmental conditions, etc.) and corresponding optimal welding parameters (such as current, voltage, welding speed, etc.). These data may come from historical records, expert experience, or experimental results.

[0193] Secondly, we need to design the output layer of the welding parameter configuration model. Unlike classification tasks, welding parameter configuration is usually a regression problem, where we need to predict continuous parameter values. Therefore, the output layer may be a fully connected layer with the number of neurons equal to the number of parameters that need to be configured. For example, if we need to configure the three parameters of current, voltage, and welding speed at the same time, then the output layer will have three neurons.

[0194] Then, we need to choose a suitable loss function. For regression problems, commonly used loss functions include mean squared error (MSE) or mean absolute error (MAE). For example, we can use MSE to measure the gap between the predicted parameters and the actual optimal parameters.

[0195] During the training process, we use the back-propagation algorithm to optimize the model parameters. It is worth noting that although we inherit the underlying feature extraction modules from the shared network, at this stage, we usually allow the parameters of these layers to be fine-tuned as well. This is because the welding parameter configuration task may require some specific features that may not be very important in the previous task. By allowing parameter updates of the entire network, we can make the model better adapt to the needs of the new task.

[0196] This process solves several key technical problems. First, it solves the problem of how to effectively utilize limited welding parameter configuration data. In practical applications, it can be very difficult and expensive to obtain a large amount of optimal welding parameter data. By using a pre-trained feature extraction network, we can train a model with good performance on less data.

[0197] Secondly, it solves the problem of how to balance the generality and specificity of the model. The shared network provides a good starting point and ensures that the model has a certain degree of generality. And through end-to-end fine-tuning, we can make the model better adapted to the specific task of welding parameter configuration.

[0198] In terms of technical results, this approach has brought significant improvements. The first is the improvement in model performance. Compared with training a welding parameter configuration model from scratch, using a pre-trained feature extraction network can usually achieve better performance. This is reflected in more accurate parameter predictions and better generalization capabilities. For example, in practical applications, we may see a 20% reduction in the average error of welding parameter predictions, or a 30% improvement in the performance of the model on new types of welding tasks over traditional methods.

[0199] Secondly, this approach can significantly speed up the convergence of the model. Since the underlying network has learned to extract useful features, the model only needs to learn how to use these features to predict welding parameters instead of learning from raw data. In practice, this may mean that the training time is reduced from days to hours, greatly improving the efficiency of model development and iteration.

[0200] In addition, this approach also improves the robustness of the model. Because the underlying features are learned from multiple related tasks, they are generally more general and stable. This means that the model will perform better when faced with new, unseen welding situations. For example, when a new alloy material needs to be welded, the model may give relatively reasonable parameter recommendations even without dedicated training data.

[0201] In one embodiment of the welding industrial robot parameter planning method of the present application, see Figure 7 , and can also include the following:

[0202] Step S701: extracting the industrial robot brand and processing object information in the welding request and integrating them into a complete feature vector;

[0203] Step S702: receiving the feature vector as input through a welding parameter configuration model integrated with a shared feature extraction network, and obtaining optimal welding parameters output by the welding parameter configuration model, wherein the optimal welding parameters include at least one of current, voltage, and welding speed.

[0204] Optionally, in this embodiment, steps S701 and S702 are key links in the practical application of the intelligent welding system, aiming to apply the previously trained model to the actual welding parameter configuration task. This process can not only realize automatic welding parameter configuration, but also make full use of previously learned knowledge to achieve more accurate and efficient welding operations.

[0205] First, let's discuss step S701 in detail. In this step, we need to extract key information from the welding request and integrate it into a complete feature vector. This process involves two important technical links: data preprocessing and feature engineering.

[0206] Specifically, welding requests usually contain a variety of information, such as the brand and model of the industrial robot, the material, size, and shape of the processing object, etc. We need to extract features that are meaningful for welding parameter configuration from this raw information. For example, for the brand of the industrial robot, we may need to convert it into a numerical code or one-hot code. For the processing object, we may need to extract its material (such as steel, aluminum, copper, etc.), thickness, welding position (such as flat welding, vertical welding, horizontal welding, etc.), and other information.

[0207] During the feature extraction process, we may also need to perform some data cleaning and standardization operations. For example, we may need to handle missing values, convert values ​​in different units to a unified standard, or convert text information to numerical representation. The purpose of these operations is to ensure that the data input to the model is clean, consistent, and standardized, thereby improving the prediction accuracy of the model.

[0208] Integrating all these features into a complete feature vector is a key step. This vector should contain all the key information that may affect the welding parameters while maintaining a certain degree of compactness and interpretability. For example, a possible feature vector may include: [robot brand code, robot power, processing object material code, processing object thickness, welding position code, ambient temperature, ambient humidity].

[0209] This process solves the technical problem of how to transform complex and diverse welding request information into standardized inputs that the model can process. Through a carefully designed feature extraction and integration process, we can ensure that the model receives the most relevant and informative data, thereby improving the accuracy and reliability of predictions.

[0210] Next, in step S702, we use the welding parameter configuration model integrated with the shared feature extraction network to process the feature vector and output the optimal welding parameters. This process involves two key links: model reasoning and parameter interpretation.

[0211] First, the feature vector is input into the shared feature extraction network of the model. This network has been previously optimized through multi-task learning and can effectively extract common and meaningful features. For example, it may be able to identify the physical properties of the material, the geometric characteristics of the welding location, etc. These extracted features are then passed to the subsequent layers of the welding parameter configuration model.

[0212] Subsequent layers of the welding parameter configuration model may include multiple fully connected layers, which are responsible for mapping the extracted features to specific welding parameters. In this process, the model considers the complex interactions between all input features and their combined impact on the final welding parameters.

[0213] Finally, the model outputs the predicted optimal welding parameters, which typically include current, voltage, and welding speed. The specific values ​​of these parameters vary depending on the characteristics of the input. For example, for thicker materials, the model may recommend higher current and slower welding speeds, while for thinner materials, it may recommend lower current and faster speeds.

[0214] This process solves the technical problem of how to automatically configure the optimal welding parameters based on complex input information. By utilizing the pre-trained model, we can give parameter recommendations within milliseconds, greatly improving the efficiency and flexibility of welding operations.

[0215] In terms of technical effects, this approach has brought significant improvements. The first is the accuracy and consistency of parameter configuration. Compared with manual configuration, model-based automatic configuration can better consider the complex interaction of various factors, thereby giving better parameter recommendations. For example, in practical applications, we may see the pass rate of welding quality increase from 90% to 98%, or the welding defect rate decrease by 50%.

[0216] Secondly, this approach greatly improves the efficiency of welding operations. Traditional parameter configuration may require experienced engineers to spend a lot of time on trial and adjustment. Using this automated method, we can get parameter suggestions in seconds, greatly reducing production preparation time. In large-scale production, this may mean producing dozens or even hundreds of more products per day.

[0217] In addition, this method also improves the adaptability and flexibility of the system. When faced with new welding tasks or new materials, traditional methods may require a lot of re-testing and adjustments. The machine learning-based method can quickly give reasonable initial parameters based on existing knowledge, and then achieve the best results with a small amount of fine-tuning. This greatly reduces the cost and time of switching production lines to new products or new processes.

[0218] In order to effectively improve the automation and accuracy of the welding machine control program executable by the industrial robot, the present application provides an embodiment of a welding industrial robot parameter planning device for implementing all or part of the contents of the welding industrial robot parameter planning method, see Figure 8 The welding industrial robot parameter planning device specifically includes the following contents:

[0219] An industrial robot classification model building module 10 is used to obtain historical welding control data of industrial robots of different brands, extract robot kinematic features and welding machine control features from the historical welding control data, and use them as first training data to obtain an industrial robot classification model through a preset convolutional neural network model training;

[0220] A processing object classification model building module 20 is used to obtain historical welding processing data of different processing objects, extract processing object quality features and processing object material features from the historical welding processing data, and use them as second training data to obtain a processing object classification model through training with a preset support vector machine model;

[0221] A welding parameter configuration model construction module 30 is used to construct two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimize the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network after model optimization, and fuse the shared underlying feature extraction network into a welding parameter configuration model based on a neural network;

[0222] The automatic welding control module 40 is used to receive a welding request sent by a user, extract the industrial robot brand and processing object information in the welding request and input the welding parameter configuration model, obtain the corresponding optimal welding parameters through the welding parameter configuration model, and generate a welding machine control program for the industrial robot according to the optimal welding parameters.

[0223] From the above description, it can be seen that the welding industrial robot parameter planning device provided in the embodiment of the present application can construct two task-specific classification output layers corresponding to the industrial robot classification model and the processing object classification model and a shared underlying feature extraction network, and simultaneously optimize the task loss function of the task-specific classification output layer through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network, and integrate the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; receive a welding request sent by a user, extract the industrial robot brand and processing object information in the welding request and input it into the welding parameter configuration model, obtain the corresponding optimal welding parameters through the welding parameter configuration model, and generate the welding machine control program of the industrial robot according to the optimal welding parameters, thereby effectively improving the automation and accuracy of the welding machine control program executable by the industrial robot.

[0224] From the hardware level, in order to effectively improve the automation and accuracy of the welding machine control program executable by the industrial robot, the present application provides an embodiment of an electronic device for implementing all or part of the contents of the welding industrial robot parameter planning method, and the electronic device specifically includes the following contents:

[0225] Processor, memory, communication interface and bus; wherein the processor, memory and communication interface communicate with each other through the bus; the communication interface is used to realize information transmission between the parameter planning device of the welding industrial robot and related equipment such as the core business system, user terminal and related database; the logic controller can be a desktop computer, a tablet computer and a mobile terminal, etc., but the present embodiment is not limited thereto. In the present embodiment, the logic controller can be implemented with reference to the embodiment of the parameter planning method of the welding industrial robot and the embodiment of the parameter planning device of the welding industrial robot in the embodiment, and the contents thereof are incorporated herein, and the repeated parts are not repeated.

[0226] It is understandable that the user terminal may include a smart phone, a tablet electronic device, a network set-top box, a portable computer, a desktop computer, a personal digital assistant (PDA), a vehicle-mounted device, a smart wearable device, etc. Among them, the smart wearable device may include smart glasses, a smart watch, a smart bracelet, etc.

[0227] In practical applications, part of the parameter planning method for the welding industrial robot can be executed on the electronic device side as described above, or all operations can be completed in the client device. The selection can be made based on the processing capability of the client device and the limitations of the user's usage scenario. This application does not limit this. If all operations are completed in the client device, the client device may also include a processor.

[0228] The client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side, and other implementation scenarios may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, or a server cluster consisting of multiple servers, or a server structure of a distributed device.

[0229] Fig. 9 FIG. 9 is a schematic block diagram of the system structure of the electronic device 9600 according to an embodiment of the present application. Fig. 9 As shown, the electronic device 9600 may include a central processor 9100 and a memory 9140; the memory 9140 is coupled to the central processor 9100. It is worth noting that Fig. 9 is exemplary; other types of structures may also be used to supplement or replace this structure to implement telecommunication functions or other functions.

[0230] In one embodiment, the parameter planning method function of the welding industrial robot can be integrated into the central processor 9100. The central processor 9100 can be configured to perform the following control:

[0231] Step S101: acquiring historical welding control data of industrial robots of different brands, extracting robot kinematic features and welding machine control features from the historical welding control data, and using the data as first training data to obtain an industrial robot classification model through training with a preset convolutional neural network model;

[0232] Step S102: acquiring historical welding processing data of different processing objects, extracting processing object quality features and processing object material features from the historical welding processing data, and using the data as second training data to obtain a processing object classification model through training with a preset support vector machine model;

[0233] Step S103: constructing two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimizing the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network after model optimization, and integrating the shared underlying feature extraction network into a welding parameter configuration model based on a neural network;

[0234] Step S104: receiving a welding request sent by a user, extracting the industrial robot brand and processing object information in the welding request and inputting them into the welding parameter configuration model, obtaining corresponding optimal welding parameters through the welding parameter configuration model, and generating a welding machine control program for the industrial robot according to the optimal welding parameters.

[0235] From the above description, it can be seen that the electronic device provided in the embodiment of the present application constructs two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimizes the task loss function of the task-specific classification output layer through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network, and integrates the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; receives a welding request sent by a user, extracts the industrial robot brand and processing object information in the welding request and inputs it into the welding parameter configuration model, obtains the corresponding optimal welding parameters through the welding parameter configuration model, and generates a welding machine control program for the industrial robot according to the optimal welding parameters, thereby effectively improving the automation and accuracy of the welding machine control program executable by the industrial robot.

[0236] In another embodiment, the welding industrial robot parameter planning device can be configured separately from the central processing unit 9100. For example, the welding industrial robot parameter planning device can be configured as a chip connected to the central processing unit 9100, and the welding industrial robot parameter planning method function can be implemented through the control of the central processing unit.

[0237] like Fig. 9 As shown, the electronic device 9600 may also include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily have to include Fig. 9 In addition, the electronic device 9600 may also include Fig. 9 For components not shown, reference may be made to the prior art.

[0238] like Fig. 9 As shown, the central processing unit 9100 is sometimes also referred to as a controller or an operation control, and may include a microprocessor or other processor device and / or logic device. The central processing unit 9100 receives input and controls the operation of various components of the electronic device 9600.

[0239] The memory 9140 may be, for example, one or more of a cache, a flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory or other suitable devices. The above-mentioned information related to the failure may be stored, and a program for executing the relevant information may also be stored. The CPU 9100 may execute the program stored in the memory 9140 to implement information storage or processing, etc.

[0240] The input unit 9120 provides input to the central processing unit 9100. The input unit 9120 is, for example, a key or a touch input device. The power supply 9170 is used to provide power to the electronic device 9600. The display 9160 is used to display display objects such as images and texts. The display may be, for example, an LCD display, but is not limited thereto.

[0241] The memory 9140 may be a solid-state memory, such as a read-only memory (ROM), a random access memory (RAM), a SIM card, etc. It may also be a memory that saves information even when the power is off, can be selectively erased, and is provided with more data, examples of which are sometimes referred to as EPROMs, etc. The memory 9140 may also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142, which is used to store application programs and function programs or processes for executing the operation of the electronic device 9600 through the central processor 9100.

[0242] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for communication functions of the electronic device and / or for executing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0243] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processor 9100 to provide input signals and receive output signals, which may be the same as the case of a conventional mobile communication terminal.

[0244] Based on different communication technologies, multiple communication modules 9110 may be provided in the same electronic device, such as a cellular network module, a Bluetooth module and / or a wireless LAN module, etc. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide an audio output via the speaker 9131 and receive an audio input from the microphone 9132, thereby realizing a common telecommunication function. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. In addition, the audio processor 9130 is also coupled to the central processor 9100, so that recording can be performed on the local machine through the microphone 9132, and the sound stored on the local machine can be played through the speaker 9131.

[0245] The embodiments of the present application also provide a computer-readable storage medium capable of implementing all the steps in the parameter planning method for a welding industrial robot in the above-mentioned embodiment, where the execution subject is a server or a client. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, all the steps in the parameter planning method for a welding industrial robot in the above-mentioned embodiment are implemented, where the execution subject is a server or a client. For example, when the processor executes the computer program, the following steps are implemented:

[0246] Step S101: acquiring historical welding control data of industrial robots of different brands, extracting robot kinematic features and welding machine control features from the historical welding control data, and using the data as first training data to obtain an industrial robot classification model through training with a preset convolutional neural network model;

[0247] Step S102: acquiring historical welding processing data of different processing objects, extracting processing object quality features and processing object material features from the historical welding processing data, and using the data as second training data to obtain a processing object classification model through training with a preset support vector machine model;

[0248] Step S103: constructing two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimizing the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network after model optimization, and integrating the shared underlying feature extraction network into a welding parameter configuration model based on a neural network;

[0249] Step S104: receiving a welding request sent by a user, extracting the industrial robot brand and processing object information in the welding request and inputting them into the welding parameter configuration model, obtaining corresponding optimal welding parameters through the welding parameter configuration model, and generating a welding machine control program for the industrial robot according to the optimal welding parameters.

[0250] From the above description, it can be seen that the computer-readable storage medium provided in the embodiment of the present application constructs two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimizes the task loss function of the task-specific classification output layer through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network, and integrates the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; receives a welding request sent by a user, extracts the industrial robot brand and processing object information in the welding request and inputs it into the welding parameter configuration model, obtains the corresponding optimal welding parameters through the welding parameter configuration model, and generates a welding machine control program for the industrial robot according to the optimal welding parameters, thereby effectively improving the automation and accuracy of the welding machine control program executable by the industrial robot.

[0251] The embodiments of the present application also provide a computer program product capable of implementing all the steps in the welding industrial robot parameter planning method in the above embodiments, where the execution subject is a server or a client. When the computer program / instruction is executed by a processor, the steps of the welding industrial robot parameter planning method are implemented. For example, the computer program / instruction implements the following steps:

[0252] Step S101: acquiring historical welding control data of industrial robots of different brands, extracting robot kinematic features and welding machine control features from the historical welding control data, and using the data as first training data to obtain an industrial robot classification model through training with a preset convolutional neural network model;

[0253] Step S102: acquiring historical welding processing data of different processing objects, extracting processing object quality features and processing object material features from the historical welding processing data, and using the data as second training data to obtain a processing object classification model through training with a preset support vector machine model;

[0254] Step S103: constructing two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimizing the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network after model optimization, and integrating the shared underlying feature extraction network into a welding parameter configuration model based on a neural network;

[0255] Step S104: receiving a welding request sent by a user, extracting the industrial robot brand and processing object information in the welding request and inputting them into the welding parameter configuration model, obtaining corresponding optimal welding parameters through the welding parameter configuration model, and generating a welding machine control program for the industrial robot according to the optimal welding parameters.

[0256] From the above description, it can be seen that the computer program product provided in the embodiment of the present application constructs two task-specific classification output layers corresponding to the industrial robot classification model and the processing object classification model and a shared underlying feature extraction network, and simultaneously optimizes the task loss function of the task-specific classification output layer through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network, and integrates the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; receives a welding request sent by a user, extracts the industrial robot brand and processing object information in the welding request and inputs it into the welding parameter configuration model, obtains the corresponding optimal welding parameters through the welding parameter configuration model, and generates a welding machine control program for the industrial robot according to the optimal welding parameters, thereby effectively improving the automation and accuracy of the welding machine control program executable by the industrial robot.

[0257] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0258] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (apparatus), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0259] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0260] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0261] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A parameter planning method for a welding industrial robot, characterized in that: The method comprises: Obtain historical welding control data of industrial robots of different brands, extract robot kinematic features and welding machine control features from the historical welding control data, and use them as first training data to obtain an industrial robot classification model through training with a preset convolutional neural network model; Acquire historical welding processing data of different processing objects, extract processing object quality features and processing object material features from the historical welding processing data, and use them as second training data to obtain a processing object classification model through training with a preset support vector machine model; Constructing two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, optimizing the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model simultaneously through a joint learning framework of multi-task learning, obtaining a shared underlying feature extraction network after model optimization, and fusing the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; A welding request sent by a user is received, the industrial robot brand and processing object information in the welding request are extracted and input into the welding parameter configuration model, the corresponding optimal welding parameters are obtained through the welding parameter configuration model, and a welding machine control program of the industrial robot is generated according to the optimal welding parameters.

2. The parameter planning method for a welding industrial robot according to claim 1, characterized in that: The method comprises: obtaining historical welding control data of industrial robots of different brands, extracting robot kinematic features and welding machine control features from the historical welding control data, and using the data as first training data to train an industrial robot classification model through a preset convolutional neural network model, including: Acquire historical welding control data of industrial robots of different brands, wherein the historical welding control data includes at least one of current, voltage, joint angle, and end tool posture parameters, and extract robot kinematic characteristics and welding machine control characteristics from the historical welding control data through video tracking algorithm and machine vision detection, and use them as first training data; By presetting a convolutional neural network model as an industrial robot classifier, the first training data is used as the input of the convolutional neural network model for model training to obtain an industrial robot classification model.

3. The parameter planning method for a welding industrial robot according to claim 1, characterized in that: The method of acquiring historical welding processing data of different processing objects, extracting processing object quality features and processing object material features from the historical welding processing data, and using the features as second training data to train a preset support vector machine model to obtain a processing object classification model includes: Acquire historical welding processing data of different processing objects, wherein the historical welding processing data includes at least one of material properties, size, shape, and surface quality parameters, and automatically extract potential quality features and material features of the processing objects from the historical welding processing data by an unsupervised learning autoencoder, and use the data as second training data; A support vector machine model is preset as a processing object classifier, and the second training data is used as an input of the support vector machine model for model training to obtain a processing object classification model.

4. The welding industrial robot parameter planning method according to claim 1, characterized in that: The constructing of two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model includes: Constructing two task-specific classification output layers corresponding to the industrial robot classification task and the processing object classification task respectively, wherein the task-specific classification output layers are responsible for the classification prediction of the respective tasks; A shared underlying feature extraction network is constructed as a shared component of multi-task learning, wherein the shared underlying feature extraction network is responsible for extracting a common feature representation from input data to provide support for two tasks.

5. The parameter planning method for a welding industrial robot according to claim 1, characterized in that: The joint learning framework through multi-task learning simultaneously optimizes the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model to obtain a shared underlying feature extraction network after model optimization, including: The prediction errors of the industrial robot classification task and the processing object classification task are calculated respectively, and the loss functions specific to the two tasks are obtained to update the shared underlying feature extraction network. The loss functions of the two tasks are jointly optimized through the back-propagation algorithm, and a shared underlying feature extraction network is obtained that can learn common feature representations that are effective for both tasks.

6. The parameter planning method for a welding industrial robot according to claim 1, characterized in that: The step of fusing the shared underlying feature extraction network into a neural network-based welding parameter configuration model comprises: The optimized shared underlying feature extraction network is used as the underlying input feature extraction module of the welding parameter configuration model; The welding parameter configuration model is end-to-end optimized and trained to obtain an optimized and trained welding parameter configuration model.

7. The parameter planning method for a welding industrial robot according to claim 1, characterized in that: The extracting the industrial robot brand and processing object information in the welding request and inputting them into the welding parameter configuration model, and obtaining the corresponding optimal welding parameters through the welding parameter configuration model, includes: Extracting the industrial robot brand and processing object information in the welding request and integrating them into a complete feature vector; The feature vector is received as input by a welding parameter configuration model integrated with a shared feature extraction network to obtain optimal welding parameters output by the welding parameter configuration model, wherein the optimal welding parameters include at least one of current, voltage, and welding speed.

8. A parameter planning device for a welding industrial robot, characterized in that: The device comprises: An industrial robot classification model building module is used to obtain historical welding control data of industrial robots of different brands, extract robot kinematic features and welding machine control features from the historical welding control data, and use them as first training data to obtain an industrial robot classification model through a preset convolutional neural network model training; A processing object classification model building module is used to obtain historical welding processing data of different processing objects, extract processing object quality characteristics and processing object material characteristics from the historical welding processing data, and use them as second training data to obtain a processing object classification model through training with a preset support vector machine model; A welding parameter configuration model construction module is used to construct two task-specific classification output layers and a shared underlying feature extraction network corresponding to the industrial robot classification model and the processing object classification model, and simultaneously optimize the task loss functions of the task-specific classification output layers of the industrial robot classification model and the processing object classification model through a joint learning framework of multi-task learning to obtain a shared underlying feature extraction network after model optimization, and fuse the shared underlying feature extraction network into a welding parameter configuration model based on a neural network; The automatic welding control module is used to receive a welding request sent by a user, extract the industrial robot brand and processing object information in the welding request and input them into the welding parameter configuration model, obtain the corresponding optimal welding parameters through the welding parameter configuration model, and generate a welding machine control program for the industrial robot according to the optimal welding parameters.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the welding industrial robot parameter planning method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the welding industrial robot parameter planning method according to any one of claims 1 to 7 are implemented.

Images