Automobile body-in-white welding process planning method

Abstract

The invention relates to an automobile body-in-white welding process planning method, which comprises the following steps of: performing data acquisition through terminal equipment, collecting welding process data in a welding process, and monitoring the running state of a welding robot in real time; edge computing nodes are arranged on a welding site and are responsible for preprocessing collected welding process data and extracting knowledge in real time; the cloud constructs a welding process knowledge graph according to the existing enterprise historical data and the welding process data extracted by the edge end knowledge; and after the cloud establishes a welding process knowledge graph, optimizing body-in-white welding process parameters and carrying out optimization solution on a body-in-white welding path and a body-in-white welding spot welding sequence on the basis of an improved artificial neural network-based automobile body-in-white welding process planning method. The method can adapt to different welding scenes and requirements, the method can adapt to shapes and materials of different workpieces, and efficient and accurate welding is achieved.

Description

Technical Field

[0001] The invention relates to the field of welding manufacturing, in particular to a method for planning welding process of an automobile body-in-white. Background Art

[0002] Automobile body-in-white welding process planning refers to the management activities of planning and arranging each link in the automobile body-in-white welding manufacturing process to ensure the efficiency and accuracy of the automobile manufacturing process after the automobile design is completed, so as to ensure the quality of the automobile, reduce the production cost of the automobile and improve its production efficiency. It is a bridge between automobile design and automobile manufacturing, aiming to develop the most effective welding sequence, select appropriate welding equipment and tools, and meet the requirements and specifications of automobile manufacturing.

[0003] As an important part of the welding manufacturing process of large automobile manufacturers, the process planning of automobile body-in-white welding is often included in the robot's welding process file, such as welding route and precision requirements. At the same time, due to the large number of shapes, materials, quantities and types of workpieces, the welding process often includes multiple processes and involves a huge amount of procedures, which makes it difficult for personnel to manage welding process knowledge and it is impossible to collect and effectively reuse the dispersed and heterogeneous welding process knowledge, further increasing the production time and cost of the enterprise. Therefore, it is urgent to study a method for automobile body-in-white welding process planning, improve the intelligence of welding robots, and shorten the time for automobile body-in-white welding process planning.

[0004] At present, in the traditional automobile body-in-white welding process planning, enterprise technicians usually follow a series of pre-set procedures and rules, which may include welding sequence, welding method, welding parameters (such as current, voltage, welding speed) and the use of welding tools. Traditional methods are highly dependent on manual experience and fixed operating procedures, and welding process knowledge lacks scientific management. There are problems such as low process planning efficiency, inability to adjust the process and balance resource load in real time, lack of flexibility, etc., which makes it difficult to adapt to the current complex and changeable automobile dynamic production and manufacturing process. Therefore, it is necessary to effectively organize and manage welding process knowledge files to improve the efficiency of process planning for automobile production.

[0005] As a branch of artificial intelligence, knowledge graph technology is an effective tool for knowledge management and a key technology for the transformation from traditional manufacturing to intelligent manufacturing. The application of knowledge graph technology in the manufacturing field has become a trend. Knowledge graphs can help engineers quickly acquire and understand process-related knowledge and reduce the time for searching and learning. By constructing a process knowledge graph, scattered knowledge can be integrated to facilitate sharing and reuse, thereby improving the utilization rate of process knowledge. Knowledge graphs can also store and process complex process knowledge in various forms. The use of process knowledge graphs can reveal the explicit and implicit relationships between process knowledge and connect seemingly isolated process knowledge to each other. Through this advanced information technology and management system, a high degree of automation of the production line can be achieved, and production efficiency and product quality can be improved.

[0006] However, in the manufacturing industry, knowledge graphs are still difficult to store, share, evolve, and match. With the development of computer science and information technology, the emergence of cloud-edge-end collaborative technology has provided new ideas for solving the above problems. Through cloud-edge-end collaborative technology, enterprises can monitor and analyze the operating status of equipment in real time, realize centralized management of process knowledge, reduce data transmission delays, and improve data processing efficiency. Edge devices can process initial data and perform underlying tasks, share the workload of cloud servers, and reduce the cost of equipment operation and data transmission as well as energy and bandwidth consumption.

[0007] In summary, the current level of white-body welding process planning of domestic automobile manufacturers is relatively low, and it is difficult to adapt to complex and changeable welding tasks. In order to improve the production efficiency of enterprises, it is very meaningful to formulate a more intelligent white-body welding process planning method. A white-body welding process planning method for automobiles proposed in the present invention combines knowledge graph technology and cloud-edge-end collaborative technology, detects welding process parameters at the terminal, extracts knowledge of welding process data at the edge, and establishes a welding process knowledge graph model in the cloud, and develops the representation, storage, update, sharing and use of welding process knowledge graphs, thereby realizing the white-body welding process planning for automobiles. This method not only improves the production efficiency of enterprises, but also reduces costs and resource consumption. It is one of the important technical means to promote intelligent manufacturing and industrial upgrading. Summary of the invention

[0008] The purpose of the present invention is to provide a method for automobile body-in-white welding process planning in order to solve the problems that process planners have difficulty in managing welding process knowledge, cannot effectively reuse heterogeneous welding process knowledge, cannot adjust welding process in real time and balance resource load.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A method for planning a welding process for an automobile body in white, the method comprising:

[0011] Step S1: Data is collected through terminal devices such as sensors to collect welding process data during the welding process, including welding parameters, robot motion trajectory, weld quality and other information, to monitor the operating status of the welding robot in real time, and to transmit the welding process data to the edge computing node;

[0012] Step S2: Setting up edge computing nodes at the welding site to preprocess the collected welding process data and extract knowledge in real time. The preprocessing operations include noise filtering, data calibration, data interpolation, data cleaning, outlier detection and processing, and data fusion. Knowledge extraction serves as the basis for the subsequent construction of a knowledge graph in the cloud.

[0013] Step S3: The cloud constructs a welding process knowledge graph based on the existing enterprise historical data and the welding process data extracted from the edge knowledge to realize the functions of welding process knowledge storage, update, sharing and use.

[0014] Step S4: After the welding process knowledge graph is established in the cloud, the welding process planning method of the automobile body in white is used to optimize the welding process parameters of the body in white, and the welding path of the body in white and the welding sequence of the body in white welding points are optimized and solved.

[0015] As a further technical solution of the present invention, in step S1, when the terminal device collects data, the following steps are included:

[0016] Current, voltage, temperature, speed and displacement sensors are used to collect welding process data (welding current, voltage, speed and power) during the welding process of the automobile body in white; encoders, gyroscopes and advanced visual systems are used to monitor the movement status of the welding robot in real time.

[0017] As a further technical solution of the present invention, in step S2, the steps of preprocessing welding process data and extracting knowledge by the edge computing node are as follows;

[0018] The preprocessing includes the following steps:

[0019] Noise filtering: removes high-frequency noise and outliers from sensor signals;

[0020] Data calibration: Calibrate the sensor readings to ensure data accuracy;

[0021] Data interpolation: interpolating missing or discontinuous data points to obtain a complete data set;

[0022] Data cleaning, outlier detection and processing: identifying and removing or correcting abnormal data points, data normalization: converting data to a uniform scale or range for easy comparison and analysis;

[0023] Data fusion: Fuse data from different sensors to improve data consistency and reliability;

[0024] Edge knowledge extraction includes entity extraction, relationship extraction, and attribute extraction. Entity extraction uses a named entity recognition model / CNN-LSTM-CRF model that combines a convolutional neural network (CNN), a long short-term memory network (LSTM), and a conditional random field (CRF). The basic components of the CNN-LSTM-CRF model are as follows:

[0025] Input embedding: Convert text data into numerical form through word embedding, which can capture semantic information at the word level;

[0026] CNN layer: The convolutional neural network layer is used to extract local features, that is, to extract useful features from the input embedding matrix. These features are usually implemented through a series of convolution filters and activation functions (such as ReLU);

[0027] LSTM layer: The long short-term memory network layer is used to process sequence data and capture long-distance dependencies. LSTM controls the flow of information through a gating mechanism, thereby effectively learning the temporal dynamic characteristics in text sequences;

[0028] CRF layer: The conditional random field layer is used to perform structured prediction on sequence data, that is, to predict the most likely label sequence of the entire sequence. The CRF layer takes into account the transition probability between labels to ensure that the predicted label sequence is grammatically and semantically reasonable.

[0029] As a further technical solution of the present invention, the CNN-LSTM-CRF mathematical model includes the following parts:

[0030] Convolutional neural network CNN part:

[0031] CNN is used to extract local features from the input sequence. Assume that the input sequence is X = [x1, x2, ..., x N ], where x i For each element in the sequence (such as the representation of a character or word), CNN generates a feature map F through a series of convolutional layers and pooling layers. cnn ;

[0032] F cnn =CNN(X);

[0033] Long short-term memory network LSTM part:

[0034] LSTM is used to process sequence data and capture long-term dependencies. It receives the feature map F generated by CNN.cnn As input, it is recursively calculated on the time step to generate the hidden state sequence H lstm ;

[0035] H lstm =LSTM(F cnn );

[0036] Conditional Random Field CRF part:

[0037] CRF is used to label the sequence, which takes into account the dependency between labels. Given the hidden state sequence H of LSTM lstm , the CRF layer calculates the probabilities of all possible label sequences and selects the sequence with the highest probability as output;

[0038] The energy function of the CRF layer is defined as:

[0039]

[0040] where Y = [y1,y2,...,y N ] is the label sequence, A is the transfer matrix, which represents the energy transferred from one label to another, and P is the emission matrix, which represents the energy of observing a certain label at a certain position. Where W and b are learnable parameters;

[0041] The CRF layer learns parameters by maximizing the log-likelihood function:

[0042]

[0043] Where L(θ) is the model parameter, Y * is the optimal label sequence, Y' is the predicted label sequence, and X is the input sequence;

[0044] Among them, (-E(Y * |X)) is the energy function, representing the optimal label sequence Y * And the energy value of the input sequence X under parameter θ. The exponential function exp(-E(Y * |X)) converts energy into probability so that sequences with lower energy have higher probability.

[0045] ∑Y': Sum all possible label sequences Y'. Here Y' represents all possible label sequences predicted by the model, including correct and incorrect predictions.

[0046] exp(-E(Y'|X)): is an exponential function, where -E(Y'|X) is an energy function, which means that the energy value of the predicted label sequence Y' and the input sequence X under the parameter θ is used to measure the matching degree of the predicted label sequence Y' and the input sequence X under the parameter θ. The lower the energy function, the higher the matching degree. The exponential function exp(-E(Y'|X)) converts energy into probability, so that sequences with lower energy have higher probabilities.

[0047] Find the optimal label sequence that minimizes the energy function:

[0048]

[0049] As a further technical solution of the present invention, in step S3, the steps of constructing a welding process knowledge graph are as follows;

[0050] Firstly, the welding task ontology is constructed by using the Protégé software using the top-down ontology model construction method. The welding process files of enterprises are taken as the data source, and entities, attributes and relationships between entities are extracted from welding tasks and historical process data.

[0051] Then, the extracted welding process knowledge is stored in the Neo4j graph database in the form of triples. The welding process knowledge includes structured data (such as relational database), semi-structured data (such as encyclopedia), and unstructured data (such as pictures, audio, and video). The knowledge in the welding field is acquired, described, and represented, and the description information about the same entity or concept from multiple sources is integrated to eliminate contradictions and ambiguities.

[0052] Finally, the welding process knowledge graph is stored in the cloud for easy retrieval.

[0053] As a further technical solution of the present invention, in step S4, the whale algorithm is used to optimize and improve the artificial neural network, and the specific steps include data collection and preprocessing, artificial neural network model construction and system verification and optimization;

[0054] Collect relevant data of automobile body-in-white welding, including weld point location, welding parameters and material properties, for training artificial neural network model so that the model can learn the rules and characteristics of the welding process; data preprocessing includes cleaning, standardization and feature extraction to improve the learning efficiency and prediction accuracy of the model;

[0055] Construct an improved artificial neural network algorithm that can predict the optimal welding parameters based on the input welding data; the design of the model should consider factors such as network structure, activation function, and optimization algorithm to ensure that the model has good generalization ability and fast convergence characteristics;

[0056] Improved artificial neural network algorithm implementation steps:

[0057] Initialize the population: randomly generate a certain number of candidate solutions (whale positions), which represent preliminary estimates of network weights and biases;

[0058] Evaluate fitness: Calculate the network performance index corresponding to each candidate solution, such as mean square error (MSE), as the value of the fitness function;

[0059] Update position: Update the position of candidate solutions according to the whale's hunting strategy, which usually includes a spiral update mechanism and a surround update mechanism. The surround update mechanism simulates the movement of whales around prey to narrow the search range;

[0060] Iterative optimization: Repeat the update steps until the stopping condition is met, such as reaching the preset maximum number of iterations or the fitness function value no longer improves significantly;

[0061] Output optimal solution: Finally, the algorithm returns the best candidate solution, that is, the optimized network parameters;

[0062] In mathematical expression, the improved artificial neural network can be expressed as:

[0063] The output y of the network is calculated from the input x through the activation function f, taking into account the network weights W and bias b: y = f(Wx + b); the fitness function J is usually in the form of an error function, such as the mean square error:

[0064]

[0065] Where N is the number of samples, y i is the predicted output of the network, t i is the real output;

[0066] The whale algorithm minimizes the fitness function J by iteratively updating weights and biases:

[0067] W new =W old +A·Db new =b old +A·E;

[0068] The meanings of the parameters in the formula are:

[0069] W old : Represents the position of each candidate solution (or "whale") in the whale algorithm before updating.

[0070] W new : Indicates the updated position of each candidate solution in the whale algorithm.

[0071] A: In the whale algorithm, A is a coefficient that controls the update amplitude of the candidate solution position during the search process. It can change dynamically according to the number of iterations or other parameters to simulate the predation behavior of whales.

[0072] D: represents the dimension of the search space, that is, the number of variables in the problem.

[0073] E: represents the energy function or fitness function of the algorithm, which is used to evaluate the quality of candidate solutions.

[0074] b old : Represents another parameter of each candidate solution in the whale algorithm before updating, which is related to weight or bias.

[0075] b new : Represents the updated parameters of each candidate solution in the whale algorithm, which is related to weight or bias.

[0076] In the whale algorithm, these parameters together determine the position update of candidate solutions and how the algorithm searches for the optimal solution. The position and related parameters of each candidate solution are updated according to the algorithm rules in each iteration to simulate the behavior of whales when hunting.

[0077] Through simulation experiments and actual welding tests, the effectiveness of the improved artificial neural network and whale optimization algorithm in automobile body-in-white welding is verified. According to the test results, necessary adjustments and optimizations are made to the improved artificial neural network and whale optimization algorithm to achieve the best welding effect.

[0078] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention combines knowledge graph technology and cloud-edge-end technology to realize the planning of automobile body-in-white welding process. Using the CNN-LSTM-CRF model for named entity recognition can provide powerful text processing capabilities for building welding process knowledge graphs, help extract and integrate welding process related information from text data, provide support for knowledge management and application in the field of welding technology, and can adapt to different welding scenarios and needs. The method of the present invention can adapt to the shapes and materials of different workpieces, realize efficient and accurate welding, greatly improve the automation level and production efficiency of welding production lines, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 The flowchart of a method for automobile body-in-white welding process planning is shown in FIG.

[0080] Figure 2 Workflow diagram of the cloud-based knowledge graph automobile body-in-white welding process planning method. DETAILED DESCRIPTION

[0081] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0082] See also Figure 1 to Figure 2 The present invention provides a method for planning the welding process of the automobile body in white. By combining knowledge graph technology and cloud-edge-end technology, the welding process planning of the automobile body in white is realized, and support is provided for knowledge management and application in the field of welding technology. The method can adapt to different welding scenarios and requirements. The method of the present invention can adapt to the shapes and materials of different workpieces, realize efficient and precise welding, improve the automation level and production efficiency of the welding production line, reduce dependence on manual experience, and has broad application prospects.

[0083] A method for planning a welding process for an automobile body in white, the method comprising:

[0084] Step S1: Data is collected through terminal devices such as sensors to collect welding process data during the welding process, including welding parameters, robot motion trajectory, weld quality and other information, to monitor the operating status of the welding robot in real time, and to transmit the welding process data to the edge computing node;

[0085] In step S1, when the terminal device collects data, the following steps are included:

[0086] Current, voltage, temperature, speed and displacement sensors are used to collect welding process data (welding current, voltage, speed and power) during the welding process of the automobile body in white; encoders, gyroscopes and advanced visual systems are used to monitor the movement status of the welding robot in real time.

[0087] Step S2: Setting up edge computing nodes at the welding site to preprocess the collected welding process data and extract knowledge in real time. The preprocessing operations include noise filtering, data calibration, data interpolation, data cleaning, outlier detection and processing, and data fusion. Knowledge extraction serves as the basis for the subsequent construction of a knowledge graph in the cloud.

[0088] The preprocessing includes the following steps:

[0089] Noise filtering: removes high-frequency noise and outliers from sensor signals;

[0090] Data calibration: Calibrate the sensor readings to ensure data accuracy;

[0091] Data interpolation: interpolating missing or discontinuous data points to obtain a complete data set;

[0092] Data cleaning, outlier detection and processing: identifying and removing or correcting abnormal data points, data normalization: converting data to a uniform scale or range for easy comparison and analysis;

[0093] Data fusion: Fuse data from different sensors to improve data consistency and reliability;

[0094] Edge knowledge extraction includes entity extraction, relationship extraction, and attribute extraction. Entity extraction uses a named entity recognition model (CNN-LSTM-CRF model) that combines convolutional neural network CNN, long short-term memory network LSTM, and conditional random field CRF. The basic composition of the CNN-LSTM-CRF model is as follows:

[0095] Input embedding: Convert text data into numerical form through word embedding, which can capture semantic information at the word level;

[0096] CNN layer: The convolutional neural network layer is used to extract local features, that is, to extract useful features from the input embedding matrix. These features are usually implemented through a series of convolution filters and activation functions (such as ReLU);

[0097] LSTM layer: The long short-term memory network layer is used to process sequence data and capture long-distance dependencies. LSTM controls the flow of information through a gating mechanism, thereby effectively learning the temporal dynamic characteristics in text sequences;

[0098] CRF layer: The conditional random field layer is used to perform structured prediction on sequence data, that is, to predict the most likely label sequence of the entire sequence. The CRF layer takes into account the transition probability between labels to ensure that the predicted label sequence is grammatically and semantically reasonable.

[0099] The CNN-LSTM-CRF mathematical model consists of the following parts:

[0100] Convolutional neural network CNN part:

[0101] CNN is used to extract local features from the input sequence. Assume that the input sequence is X = [x1, x2, ..., x N ], where x i For each element in the sequence (such as the representation of a character or word), CNN generates a feature map F through a series of convolutional layers and pooling layers. cnn ;

[0102] F cnn =CNN(X);

[0103] Long short-term memory network LSTM part:

[0104] LSTM is used to process sequence data and capture long-term dependencies. It receives the feature map F generated by CNN. cnn As input, it is recursively calculated on the time step to generate the hidden state sequence H lstm ;

[0105] H lstm =LSTM(F cnn );

[0106] Conditional Random Field CRF part:

[0107] CRF is used to label the sequence, which takes into account the dependency between labels. Given the hidden state sequence H of LSTM lstm , the CRF layer calculates the probabilities of all possible label sequences and selects the sequence with the highest probability as output;

[0108] The energy function of the CRF layer is defined as:

[0109]

[0110] where Y = [y1,y2,...,y N ] is the label sequence, A is the transfer matrix, which represents the energy transferred from one label to another, and P is the emission matrix, which represents the energy of observing a certain label at a certain position. Where W and b are learnable parameters;

[0111] The CRF layer learns parameters by maximizing the log-likelihood function:

[0112]

[0113] Where L(θ) is the model parameter, Y * is the optimal label sequence, Y' is the predicted label sequence, and X is the input sequence; where (-E(Y * |X)) is the energy function, representing the optimal label sequence Y * And the energy value of the input sequence X under parameter θ. The exponential function exp(-E(Y * |X)) converts energy into probability so that sequences with lower energy have higher probability.

[0114] ∑Y': Sum all possible label sequences Y'. Here Y' represents all possible label sequences predicted by the model, including correct and incorrect predictions.

[0115] exp(-E(Y'|X)): is an exponential function, where -E(Y'|X) is an energy function, which means that the energy value of the predicted label sequence Y' and the input sequence X under the parameter θ is used to measure the matching degree of the predicted label sequence Y' and the input sequence X under the parameter θ. The lower the energy function, the higher the matching degree. The exponential function exp(-E(Y'|X)) converts energy into probability, so that sequences with lower energy have higher probabilities.

[0116] Find the optimal label sequence that minimizes the energy function:

[0117]

[0118] Step S3: The cloud constructs a welding process knowledge graph based on the existing enterprise historical data and the welding process data extracted from the edge knowledge to realize the functions of welding process knowledge storage, update, sharing and use.

[0119] Firstly, the welding task ontology is constructed by using the Protégé software using the top-down ontology model construction method. The welding process files of enterprises are taken as the data source, and entities, attributes and relationships between entities are extracted from welding tasks and historical process data.

[0120] Then, the extracted welding process knowledge is stored in the Neo4j graph database in the form of triples. The welding process knowledge includes structured data (such as relational database), semi-structured data (such as encyclopedia), and unstructured data (such as pictures, audio, and video). The knowledge in the welding field is acquired, described, and represented, and the description information about the same entity or concept from multiple sources is integrated to eliminate contradictions and ambiguities.

[0121] Finally, the welding process knowledge graph is stored in the cloud for easy retrieval.

[0122] Step S4: After the welding process knowledge graph is established in the cloud, the welding process planning method of the automobile body in white is used to optimize the welding process parameters of the body in white, and the welding path of the body in white and the welding sequence of the body in white welding points are optimized and solved.

[0123] In step S4, the whale algorithm is used to optimize and improve the artificial neural network, and the specific steps include data collection and preprocessing, artificial neural network model construction and system verification and optimization;

[0124] Collect relevant data of automobile body-in-white welding, including weld point location, welding parameters and material properties, for training artificial neural network model so that the model can learn the rules and characteristics of the welding process; data preprocessing includes cleaning, standardization and feature extraction to improve the learning efficiency and prediction accuracy of the model;

[0125] Construct an improved artificial neural network algorithm that can predict the optimal welding parameters based on the input welding data; the design of the model should consider factors such as network structure, activation function, and optimization algorithm to ensure that the model has good generalization ability and fast convergence characteristics;

[0126] Improved artificial neural network algorithm implementation steps:

[0127] Initialize the population: randomly generate a certain number of candidate solutions (whale positions), which represent preliminary estimates of network weights and biases;

[0128] Evaluate fitness: Calculate the network performance index corresponding to each candidate solution, such as mean square error (MSE), as the value of the fitness function;

[0129] Update position: Update the position of candidate solutions according to the whale's hunting strategy, which usually includes a spiral update mechanism and a surround update mechanism. The surround update mechanism simulates the movement of whales around prey to narrow the search range;

[0130] Iterative optimization: Repeat the update steps until the stopping condition is met, such as reaching the preset maximum number of iterations or the fitness function value no longer improves significantly;

[0131] Output optimal solution: Finally, the algorithm returns the best candidate solution, that is, the optimized network parameters;

[0132] In mathematical expression, the improved artificial neural network can be expressed as:

[0133] The output y of the network is calculated from the input x through the activation function f, taking into account the network weights W and bias b: y = f(Wx + b); the fitness function J is usually in the form of an error function, such as the mean square error:

[0134]

[0135] Where N is the number of samples, y i is the predicted output of the network, t i is the real output;

[0136] The whale algorithm minimizes the fitness function J by iteratively updating weights and biases:

[0137] =W old +A·Db new =b old +A·E;

[0138] The meanings of the parameters in the formula are:

[0139] W old: Represents the position of each candidate solution (or "whale") in the whale algorithm before updating.

[0140] W new : Indicates the updated position of each candidate solution in the whale algorithm.

[0141] A: In the whale algorithm, A is a coefficient that controls the update amplitude of the candidate solution position during the search process. It can change dynamically according to the number of iterations or other parameters to simulate the predation behavior of whales.

[0142] D: represents the dimension of the search space, that is, the number of variables in the problem.

[0143] E: represents the energy function or fitness function of the algorithm, which is used to evaluate the quality of candidate solutions.

[0144] b old : Represents another parameter of each candidate solution in the whale algorithm before updating, which is related to weight or bias.

[0145] b new : Represents the updated parameters of each candidate solution in the whale algorithm, which is related to weight or bias.

[0146] In the whale algorithm, these parameters together determine the position update of candidate solutions and how the algorithm searches for the optimal solution. The position and related parameters of each candidate solution are updated according to the algorithm rules in each iteration to simulate the behavior of whales when hunting.

[0147] Through simulation experiments and actual welding tests, the effectiveness of the improved artificial neural network and whale optimization algorithm in automobile body-in-white welding is verified. According to the test results, necessary adjustments and optimizations are made to the improved artificial neural network and whale optimization algorithm to achieve the best welding effect.

[0148] In the present invention, the formulas are obtained based on the collection of actual data and software simulation, and are intended to reflect the actual situation of text processing. The parameters in these formulas, including weight coefficients and proportional coefficients, are set to ensure that they can reflect the relative importance of each parameter after quantization.

[0149] Specifically, the weight coefficients and scale coefficients are set based on the experience of domain experts and statistical analysis of a large amount of data. The size of these coefficients is crucial to the output of the model, as they determine how much the model values ​​different features when processing text. However, these coefficients are not fixed and can be adjusted and optimized based on actual conditions.

[0150] In practical applications, the output of the model will be used to guide the execution of related tasks. By comparing the output of the model with the judgment of experts, the parameters can be further adjusted to improve the accuracy and reliability of the model. This adjustment process is a dynamic process and can be continuously optimized according to the development of new data and technology.

[0151] It should be emphasized that although the present invention provides specific embodiments, these examples are not exhaustive of all possible solutions. The present invention provides a general method that can be flexibly adjusted and expanded according to different application scenarios and requirements. Therefore, the scope of the present invention is not limited to the disclosed embodiments, but also includes all possible variations and improvements that can be implemented under the principles of the present invention.

Claims

1. A method for planning the welding process of an automobile body in white, characterized in that: The method comprises: Step S1: Data is collected through the terminal device to collect welding process data during the welding process, including welding parameters, robot motion trajectory and weld quality, to monitor the operating status of the welding robot in real time, and to transmit the welding process data to the edge computing node; Step S2: Setting up edge computing nodes at the welding site to preprocess the collected welding process data and extract knowledge in real time. The preprocessing operations include noise filtering, data calibration, data interpolation, data cleaning, outlier detection and processing, and data fusion. Knowledge extraction serves as the basis for the subsequent construction of a knowledge graph in the cloud. Step S3: The cloud constructs a welding process knowledge graph based on the existing enterprise historical data and the welding process data extracted by the edge knowledge to realize the functions of storing, updating, sharing and using the welding process knowledge; Step S4: After the welding process knowledge graph is established in the cloud, the welding process planning method of the automobile body in white is used to optimize the welding process parameters of the body in white, and the welding path of the body in white and the welding sequence of the body in white welding points are optimized and solved.

2. The method for planning the welding process of an automobile body in white according to claim 1, characterized in that: In step S1, when the terminal device collects data, the following steps are included: Current, voltage, temperature, speed and displacement sensors are used to collect welding process data during the automobile body-in-white welding process; encoders, gyroscopes and visual systems are used to monitor the movement status of the welding robot in real time.

3. The method for planning the welding process of an automobile body in white according to claim 2, characterized in that: In step S2, the steps of preprocessing welding process data and extracting knowledge by the edge computing node are as follows; The preprocessing includes the following steps: Noise filtering: removes high-frequency noise and outliers from sensor signals; Data calibration: Calibrate the sensor readings to ensure data accuracy; Data interpolation: interpolating missing or discontinuous data points to obtain a complete data set; Data cleaning, outlier detection and processing: identifying and removing or correcting abnormal data points, data normalization: converting data to a uniform scale or range for easy comparison and analysis; Data fusion: Fuse data from different sensors to improve data consistency and reliability; Edge knowledge extraction includes entity extraction, relationship extraction, and attribute extraction. Entity extraction uses a named entity recognition model / CNN-LSTM-CRF model that combines a convolutional neural network (CNN), a long short-term memory network (LSTM), and a conditional random field (CRF). The basic components of the CNN-LSTM-CRF model are as follows: Input embedding: Convert text data into numerical form, which is achieved through word embedding, which captures semantic information at the word level; CNN layer: The convolutional neural network layer is used to extract local features, that is, to extract useful features from the input embedding matrix. These features are usually implemented through a series of convolution filters and activation functions. LSTM layer: The long short-term memory network layer is used to process sequence data and capture long-distance dependencies. LSTM controls the flow of information through a gating mechanism, thereby effectively learning the temporal dynamic characteristics in text sequences; CRF layer: The conditional random field layer is used to perform structured prediction on sequence data, that is, to predict the most likely label sequence of the entire sequence. The CRF layer takes into account the transition probability between labels to ensure that the predicted label sequence is grammatically and semantically reasonable.

4. The method for planning the welding process of an automobile body in white according to claim 3, characterized in that: The CNN-LSTM-CRF mathematical model consists of the following parts: Convolutional neural network CNN part: CNN is used to extract local features from the input sequence. Assume that the input sequence is X = [x1, x2, ..., x N ], where x i For each element in the sequence, CNN generates a feature map F through a series of convolutional layers and pooling layers. cnn ; F cnn =CNN(X); Long short-term memory network LSTM part: LSTM is used to process sequence data and capture long-term dependencies. It receives the feature map F generated by CNN. cnn As input, it is recursively calculated on the time step to generate the hidden state sequence H lstm ; H lstm =LSTM(F cnn ); Conditional Random Field CRF part: CRF is used to label the sequence, which takes into account the dependency between labels. Given the hidden state sequence H of LSTM lstm , the CRF layer calculates the probabilities of all possible label sequences and selects the sequence with the highest probability as output; The energy function of the CRF layer is defined as: where Y = [y1,y2,...,y N ] is the label sequence, A is the transfer matrix, which represents the energy transferred from one label to another, and P is the emission matrix, which represents the energy of observing a certain label at a certain position. Where W and b are learnable parameters; The CRF layer learns parameters by maximizing the log-likelihood function: Where L(θ) is the model parameter, Y * is the optimal label sequence, Y' is the predicted label sequence, and X is the input sequence; Among them, (-E(Y * |X)) is the energy function, representing the optimal label sequence Y * and the energy value of the input sequence X under parameter θ, the exponential function exp(-E(Y * |X)) converts energy into probability, making sequences with lower energy have higher probability; ∑Y' refers to the sum of all possible label sequences Y', where Y' represents all possible label sequences predicted by the model, including correct and incorrect predictions; exp(-E(Y'|X)) is an exponential function, where -E(Y'|X) is an energy function, which means that the energy value of the predicted label sequence Y' and the input sequence X under the parameter θ is used to measure the matching degree of the predicted label sequence Y' and the input sequence X under the parameter θ; the lower the energy function, the higher the matching degree; the exponential function exp(-E(Y'|X)) converts energy into probability, so that the sequence with lower energy has a higher probability; Find the optimal label sequence that minimizes the energy function:

5. The method for planning the automobile body-in-white welding process according to claim 1, characterized in that: In step S3, the steps of constructing a welding process knowledge graph are as follows; Firstly, the welding task ontology is constructed by using the Protégé software using the top-down ontology model construction method. The welding process files of enterprises are taken as the data source, and entities, attributes and relationships between entities are extracted from welding tasks and historical process data. . Then, the extracted welding process knowledge is stored in the Neo4j graph database in the form of triples, and the welding process knowledge includes structured data, semi-structured data, and unstructured data; the knowledge in the welding field is acquired, described, and represented, and the description information about the same entity or concept from multiple sources is integrated to eliminate contradictions and ambiguities; Finally, the welding process knowledge graph is stored in the cloud for easy retrieval.

6. The method for planning the welding process of an automobile body in white according to claim 1, characterized in that: In step S4, the whale algorithm is used to optimize and improve the artificial neural network, and the specific steps include data collection and preprocessing, artificial neural network model construction and system verification and optimization; Collect relevant data of automobile body-in-white welding, including weld point location, welding parameters and material properties, for training artificial neural network model so that the model can learn the rules and characteristics of welding process; data preprocessing includes cleaning, standardization and feature extraction to improve the learning efficiency and prediction accuracy of the model; Constructing an improved artificial neural network algorithm that predicts optimal welding parameters based on input welding data; Improved artificial neural network algorithm implementation steps: Initialize the population: randomly generate a certain number of candidate solutions, which represent preliminary estimates of network weights and biases; Evaluate fitness: Calculate the network performance index corresponding to each candidate solution as the value of the fitness function; Update position: According to the whale's hunting strategy, the position of the candidate solution is updated, including the spiral update mechanism and the encirclement update mechanism. The encirclement update mechanism simulates the movement of whales around prey to narrow the search range; Iterative optimization: repeat the update steps until the stopping condition is met; Output optimal solution: Finally, the algorithm returns the best candidate solution, that is, the optimized network parameters; In mathematical expression, the improved artificial neural network is expressed as: The output y of the network is calculated from the input x through the activation function f, taking into account the network's weights W and bias b: y = f(Wx + b); The fitness function J is usually in the form of an error function: Where N is the number of samples, y i is the predicted output of the network, t i is the real output; The whale algorithm minimizes the fitness function J by iteratively updating weights and biases: W new =W old +A·Db new =b old +A·E; The meanings of the parameters in the formula are: W old : represents the position of each candidate solution in the whale algorithm before updating; W new : represents the updated position of each candidate solution in the whale algorithm; A: In the whale algorithm, A is a coefficient used to control the update amplitude of the candidate solution position during the search process; D: represents the dimension of the search space, that is, the number of variables in the problem; E: represents the energy function or fitness function of the algorithm, which is used to evaluate the quality of candidate solutions; b old : represents another parameter of each candidate solution in the whale algorithm before updating, which is related to weight or bias; b new : represents the updated parameters of each candidate solution in the whale algorithm, which is related to weight or bias; Through simulation experiments and actual welding tests, the effectiveness of the improved artificial neural network and whale optimization algorithm in automobile body-in-white welding is verified. According to the test results, the improved artificial neural network and whale optimization algorithm are adjusted and optimized to achieve the best welding effect.

Images