Method for automatically generating assembly process based on Inte3D and InteAI

Abstract

The invention discloses an automatic assembly process generation method based on Inte3D and InteAI collaboration, and the method comprises the steps: obtaining a related file of a to-be-processed part input by a user, carrying out the preprocessing of the related file of the to-be-processed part, so as to obtain a preprocessed related file, inputting the preprocessed related file into a pre-established process generation model, and carrying out the processing of the to-be-processed part. And the assembly process of the to-be-machined part is obtained. The technical problem that an existing method based on a three-dimensional process planning system excessively depends on geometric information of a model and is disjointed with an actual production scene can be solved; and the technical problems that an existing method based on an industrial intelligent platform neglects implicit knowledge and the robustness of a process scheme is insufficient are solved. And an existing method based on an expert system is difficult to adapt to rapid iteration of a product, and the novel scene rule matching success rate is low. And the technical problems of assembly interference and unreasonable parameters due to the lack of closed-loop optimization in the existing three methods are solved.

Description

Technical Field

[0001] This invention belongs to the fields of intelligent manufacturing and deep learning technology, and more specifically, relates to a method for automatically generating assembly processes based on Inte3D and InteAI. Background Technology

[0002] Against the backdrop of modern manufacturing's transformation towards intelligence and efficiency, assembly process design, as the core hub connecting product design and production execution, directly determines product launch cycle and manufacturing costs in terms of efficiency and quality. With increasingly complex product structures and continuously growing customer customization demands, traditional process design models relying on manual experience are no longer suitable for the market demands of flexible production and rapid iteration. The automation and intelligent generation of assembly processes has become a critical issue that the industry urgently needs to address—it not only breaks through the limitations of manual operation but also enables the systematic reuse of process knowledge and the optimized allocation of production resources, providing core technological support for cost reduction and efficiency improvement in the manufacturing industry.

[0003] Currently, various technical methods for the automatic generation of assembly processes have emerged in the industry, which can be mainly divided into three categories: The first category is the automated generation method based on 3D process planning systems. This method automatically generates visualized assembly processes and process documents by lightweighting the 3D digital model of the product, identifying assembly constraints, and planning process paths. Its core reliance is on model-driven rule-based reasoning. The second category is the data analysis-driven method based on industrial intelligent platforms. This method uses big data technology to mine the correlation patterns in historical process data and production resource data, and combines machine learning algorithms to achieve intelligent recommendation and decision-making of process parameters. The third category is the knowledge-driven method based on expert systems. This method builds an assembly process knowledge base, transforms expert experience into a reusable rule base, and automatically generates process solutions based on rule matching and logical reasoning.

[0004] However, the aforementioned existing methods for automatically generating assembly processes all have some significant drawbacks: First, automated generation methods based on 3D process planning systems rely excessively on the geometric and constraint information of the model itself, lacking deep integration with production resources (such as equipment capacity and tool status) and historical process data. This results in generated process schemes that are easily disconnected from actual production scenarios, leading to insufficient practicality. Second, data analysis-driven methods based on industrial intelligent platforms emphasize data pattern mining but neglect explicit constraint rules and implicit knowledge (such as expert-accumulated experience in handling special working conditions) in assembly processes, resulting in low rationality and robustness of the process schemes. Third, knowledge-driven methods based on expert systems are limited by the completeness and update speed of the knowledge base. When faced with new assembly scenarios for complex products, the rule matching success rate is low, making it difficult to adapt to the rapid iteration needs of products. Fourth, all three methods lack a closed-loop mechanism of "data-knowledge-model-simulation-optimization." The generated process schemes have not undergone simulation verification and iterative optimization, leading to problems such as assembly interference and unreasonable parameters, which in turn result in high rework rates in actual production. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for automatically generating assembly processes based on Inte3D and InteAI. The aim is to solve the technical problems of existing automated generation methods based on 3D process planning systems, which generate process schemes that are easily out of sync with actual production scenarios and lack practicality; the technical problems of existing data analysis-driven methods based on industrial intelligent platforms, which suffer from low rationality and robustness of process schemes; the technical problems of existing knowledge-driven methods based on expert systems, which suffer from low rule matching success rates and difficulty adapting to the rapid iteration needs of products when facing new assembly scenarios for complex products; and the technical problems of the three existing methods, which, due to the lack of simulation verification and iterative optimization, are prone to assembly interference, unreasonable parameters, and thus high rework rates in actual production.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for automatically generating assembly processes based on Inte3D and InteAI is provided, comprising the following steps: (1) Obtain the relevant files of the part to be processed input by the user, and preprocess the relevant files of the part to be processed to obtain the preprocessed relevant files; (2) Input the pre-processed relevant files obtained in step (1) into the pre-established process generation model to obtain the assembly process of the part to be processed.

[0007] Preferably, the relevant files for the parts to be processed in the steps are three-dimensional digital model files or part information, and the part information includes design requirements and available production resources; The preprocessing of the relevant files of the part to be processed in step (1) is as follows: First, the relevant files are converted into a lightweight model format using the Inte3D tool. The lightweight model file is then integrated with the scattered part information to obtain a uniform and efficiently parsable intermediate model file B. Then, the intermediate model file B is cleaned up for geometric noise and redundant data to obtain a processed intermediate model file C. Subsequently, the processed intermediate model file C is cleaned up for redundant data to obtain a secondary processed intermediate model file D. Finally, the secondary processed intermediate model file D is processed for compatibility standardization to obtain the preprocessed relevant files.

[0008] Preferably, the process generation model includes a feature extraction module, a knowledge fusion module, and a process decision module, with the following specific structure: The feature extraction module includes a Convolutional Neural Network (CNN), a Transformer encoder, and a data normalization module, wherein: The input to the convolutional neural network is a 3D digital model file of multiple parts to be processed. The convolutional neural network extracts the geometric feature vectors of all parts to be processed from the 3D digital model file and outputs them. The input to the Transformer encoder is unstructured data of multiple parts to be processed. The Transformer encoder extracts semantic feature vectors from the unstructured data and outputs them. The input to the data standardization module is the production resource data of multiple parts to be processed. The data standardization module extracts resource feature vectors from the production resource data and outputs them. The knowledge fusion module takes the geometric feature vector, semantic feature vector, and resource feature vector output by the feature extraction module as input. The knowledge fusion module performs cross-modal fusion processing on the input to obtain a fused feature matrix and output it. The process decision module takes the fusion feature matrix output by the knowledge fusion module as its input. The process decision module uses the genetic algorithm (GA) to process the fusion feature matrix to obtain the optimal set of process decision parameters that are suitable for the current production scenario. Preferably, the processing procedure of the knowledge fusion module is as follows: First, the weights of the geometric feature vector, semantic feature vector, and resource feature vector are calculated using an attention mechanism; Subsequently, for the high-weight feature vectors, the knowledge fusion module first normalizes them using the Z-score normalization algorithm, then preprocesses the normalized feature vectors with a Gaussian filter algorithm of 3×3 kernel size and standard deviation σ=0.8 for noise reduction, and finally processes each preprocessed feature vector using a weighted average fusion algorithm. Where Wᵢ is the i-th high weight output by the attention mechanism, and has , Let i be the preprocessed feature vectors, and n represent the total number of feature vectors participating in the fusion, to obtain the initial core fusion feature corresponding to each feature vector. Principal component analysis (PCA) was used to analyze all the initial core fusion features obtained. Dimensionality reduction is performed to obtain unified core fusion features across multiple dimensions. ; Meanwhile, for the medium-weighted feature vectors, the knowledge fusion module first preprocesses them using the Min-Max normalization algorithm, and then uses the Local Linear Embedding (LLE) algorithm to process the preprocessed feature vectors to obtain multiple medium-weighted feature vectors with optimized dimensions. ; Meanwhile, for low-weight feature vectors, the knowledge fusion module first uses the Min-Max normalization algorithm for preprocessing, and then uses the LLE algorithm to extract core effective information from the preprocessed feature vectors to obtain multiple low-weight feature vectors with optimized dimensions. ; Subsequently, the knowledge fusion module extracts weighted feature vectors from all the data. and all low-weight feature vectors Screening and core fusion features For multiple feature vectors with a correlation ≥ 0.6, an adaptive weighted fusion algorithm is used to sequentially integrate all the selected medium-weighted feature vectors and low-weighted features into the core fusion feature. During the integration process, the gradient descent algorithm is used to dynamically adjust the weights of the high-weight feature vectors and low-weight feature vectors during the integration process, ensuring that the proportion of high-weight feature vectors in the integrated feature vectors is greater than or equal to 60%, thereby obtaining the initial fused feature set. Then, the parameter matching rules in the knowledge graph and the pre-stored process association data are used to correct the initial fusion feature set. First, the association relationship of each feature parameter in the initial fusion feature set is checked one by one, and abnormal feature parameters A that do not meet the rules are marked. Then, abnormal feature parameters A are compared with historical process association data to filter out candidate parameter values ​​B that meet the rules. After that, based on the candidate parameter values ​​B, the abnormal parameters in the initial fusion feature set are replaced. Finally, this correction process is repeated until all parameters in the initial fusion feature set meet the rule requirements to obtain the corrected fusion feature set. Finally, conflict detection and resolution are performed on the corrected fusion feature set to form a fusion feature matrix and output it.

[0009] Preferably, the process of performing conflict detection and resolution on the corrected fused feature set to form a fused feature matrix and output it specifically involves: First, calling a pre-established conflict detection rule base to traverse the corrected fused feature set dimension by dimension to obtain all conflict features, as well as the dimension and type of each conflict feature, and recording the parameter value and weight of each conflict feature; then, sorting all conflict features in descending order of their weights, retaining the top 30% of conflict features in the sorting result as high-priority conflict features, and the remaining conflict features as low-priority conflict features; and performing weighted harmonic correction on all low-priority conflict features to obtain the weighted harmonic correction parameter values ​​of the low-priority conflict features. , Let be the parameter value of the i-th low-priority conflict feature. Let P be the weight of the j-th low-priority conflicting feature, and n represent the total number of low-priority conflicting features. Finally, P is the parameter value of all high-priority conflicting features and the weighted harmonic correction of the low-priority conflicting features. final Integrate by dimension order to obtain and output the fused feature matrix.

[0010] Preferably, the process decision module's processing procedure is as follows: First, the obtained fusion feature matrix is ​​mapped through a fully connected network to obtain a candidate set of process parameters. This candidate set includes the key parameter ranges corresponding to each processing method, as well as the candidate equipment models and candidate tool types corresponding to each processing method. Then, the obtained candidate set of process parameters is probabilistically sorted using the Softmax activation function, and the constraint satisfaction problem solving (CSPs) algorithm is used to perform constraint matching and feasibility verification on the probabilistic sorting results to obtain multiple effective process parameter combinations. Finally, a directed acyclic graph (DAG) is constructed based on all the obtained effective process parameter combinations and the process dependencies corresponding to each effective process parameter combination. The process sorting algorithm is then used to integrate and optimize the order of all effective process parameters contained in the DAG to obtain a complete assembly process plan for all parts to be processed, including process sequence, processing parameters, equipment selection, and matching tools.

[0011] Preferably, the process generation model is obtained through the following steps: (1-1) Obtain multi-source data of enterprises, including structured data and unstructured data, classify and organize the obtained multi-source data of enterprises, and use the classification and organization results as the original dataset; (1-2) Perform data preprocessing on the original dataset obtained in step (1-1) to obtain the preprocessed dataset; (1-3) Perform data annotation and association processing on the preprocessed dataset obtained in step (1-2) to obtain the secondary processed dataset and output it; (1-4) Divide the dataset obtained in step (1-3) into training set, validation set and test set in a ratio of 7:2:1, and perform data augmentation on the training set to obtain the data augmented training set. (1-5) Initialize the process generation model to obtain the initialized process generation model; (1-6) For each sample in the training set after data augmentation processing obtained in step (1-4), the sample is input into the convolutional neural network in the process generation model initialized in step (1-5) to obtain the three fixed-length local feature vector sequences corresponding to the sample. (1-7) For each sample in the training set after data augmentation processing obtained in step (1-4), the three fixed-length local feature vector sequences corresponding to the sample obtained in step (1-6) are input into the Transformer encoder in the process generation model to obtain a global feature encoding vector corresponding to the sample. (1-8) For the training set in the training set after data augmentation processing obtained in step (1-4), the global feature encoding vector corresponding to the sample obtained in step (1-7) is input into the output layer of the process generation model to obtain the combined loss value corresponding to the sample. (1-9) For each sample in the training set after data augmentation processing obtained in step (1-4), the process generation model is iteratively trained using the backpropagation algorithm based on the combined loss value corresponding to the sample obtained in step (1-8) until the process generation model reaches the preset number of iterations (200 times in this invention), and the optimal parameters of the process generation model at this time are obtained, thereby obtaining the initially trained process generation model. (1-10) Use the test set obtained in step (1-4) to test the process generation model initially trained in step (1-9) until the detection accuracy reaches the optimal level, thereby obtaining the final trained process generation model. Preferably, the enterprise's structured data includes 3D digital model files from the enterprise's Computer-Aided Design (CAD) / Product Lifecycle Management (PLM) system, production resource data from the Enterprise Resource Planning (ERP) / Manufacturing Execution System (MES) system, and historical process data from the process database; Unstructured data from enterprises includes technical manuals, expert experience records, and simulation logs; For the structured data in the original dataset, the data preprocessing process in step (1-2) specifically involves lightweighting the 3D model file, cleaning the production resource data, and standardizing the format of the historical process data. For the unstructured data in the original dataset, the data preprocessing process in steps (1-2) specifically involves: performing structured parsing and key information extraction on the technical manual to obtain structured entries; using Natural Language Processing (NLP) technology to perform text segmentation and entity extraction on expert experience records; and performing text structure conversion on simulation logs. The process of structured parsing and key information extraction of the technical manual is as follows: First, the technical manual is split into chapters and hierarchical sections, and the process procedure modules in each chapter are located and identified. Then, information in all process procedure modules is extracted through keyword matching, including assembly requirement clauses and equipment operation specification texts. Finally, the extracted information is organized into structured entries according to the predetermined field format.

[0012] Preferably, steps (1-3) are as follows: First, for the structured data in the preprocessed dataset, the correspondence between part features and historical processes is labeled; for the unstructured data in the preprocessed original dataset, semantic labels are labeled. Then, a mapping relationship is established between the labeled structured and unstructured data, the three-dimensional digital model file, the production resource data, and the historical process data through a data association engine. The initialization in steps (1-5) specifically involves setting the initial values ​​of the weight parameters to random values ​​output by a truncated normal distribution with a standard deviation of 0.1, selecting the Adam optimizer as the optimizer for model training, setting the initial value of the bias parameter to 0, setting the initial learning rate to lr=0.0003, adopting a step-wise learning decay strategy with a step size of stepsize=200 and a decay weight of gamma=0.1. Steps (1-6) are as follows: First, each sample in the training set after data augmentation in step (1-4) undergoes a three-layer convolution operation to obtain three local key features corresponding to the sample. The local key features output by the first convolution operation are the numerical gradient features of continuous parameters such as processing temperature and feed rate. The local key features output by the second convolution operation are the category encoding features of discrete parameters such as tool type and solder type. The local key features output by the third convolution operation are the unified representation features that fuse feature dimensions and redundant information. Then, the local key features output by each convolution operation are sequentially fed into the BatchNorm normalization layer for standardization, and then fed into the ReLU activation function to introduce nonlinear expressive power to obtain three high-dimensional local feature maps after normalization and nonlinear enhancement. Finally, the three high-dimensional local feature maps are compressed by global average pooling to obtain three fixed-length (256-dimensional) local feature vector sequences corresponding to the sample.

[0013] Preferably, the Transformer encoder is composed of 6 stacked coding units, each of which includes an 8-head multi-head self-attention mechanism and a feedforward neural network with a hidden layer dimension of 1024, and is configured with layer normalization and residual connections. Steps (1-8) are as follows: First, the Transformer encoder is used as the network structure A for adapting global feature encoding. Then, based on the multi-head self-attention mechanism in network structure A, the association weights of any two features in each local feature vector sequence obtained in step (1-6) are calculated as the interaction feature B corresponding to the local feature vector sequence after global association modeling. Then, the interaction feature B is input into the feedforward neural network and processed by two linear transformations and ReLU activation to obtain the semantically enhanced feature C corresponding to the local process feature vector sequence. Finally, layer normalization is performed on the semantically enhanced features C corresponding to all three local process feature vector sequences, and finally, a fixed-dimensional global feature encoding vector corresponding to the sample is output. The output layer is a two-branch output layer structure with a classification head and a regression head. The classification head is configured with a fully connected network containing a Softmax activation function, and the regression head is configured with a fully connected network containing a Linear activation function. Steps (1-8) are as follows: First, the classification head maps the features of the global feature encoding vector corresponding to each sample obtained in step (1-7) to the category space of discrete process parameters through linear transformation to obtain the predicted probability distribution of discrete process parameters in the sample. Then, the regression head performs multi-level linear transformation and numerical calibration on the global feature encoding vector corresponding to each sample obtained in step (1-7) to obtain the predicted value of continuous parameters in the sample. Then, based on the predicted probability distribution of discrete process parameters in the sample output by the classification head and the true discrete label of the sample, the classification error Loss_CE is obtained by using cross-entropy loss. Then, based on the predicted value of continuous parameters in the sample output by the regression head and the true continuous parameters of the sample, the regression bias Loss_MSE is obtained by using mean squared error loss, while setting the weight coefficient α=0.6. Finally, based on the classification error Loss_CE and the regression bias Loss_MSE, the combined loss value Loss=α・Loss_CE+(1−α)・Loss_MSE corresponding to the sample is obtained.

[0014] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. This invention employs the local feature extraction of convolutional neural networks in steps (1-6) and the global feature encoding of the Transformer encoder in steps (1-7). It accurately extracts local features such as continuous parameter gradients and discrete parameter encodings from process data through multi-scale convolution, and then achieves global correlation modeling across feature types through the Transformer encoder. It deeply integrates multi-source data such as 3D numerical models, production resources, and historical processes. Therefore, it can solve the technical problems of existing methods based on 3D process planning systems, such as over-reliance on model geometric information, disconnection from actual production scenarios, and insufficient practicality. 2. This invention employs data annotation and association processing in steps (1-3) and model iterative training in steps (1-9). It establishes a mapping relationship between three-dimensional digital models, production resources, and historical processes through a data association engine, transforms expert experience into labeled information and training data, and optimizes model parameters by combining backpropagation algorithm. This allows the model to learn data patterns and explicit constraint rules simultaneously. Therefore, it can solve the technical problem of neglecting implicit knowledge in existing methods based on industrial intelligent platforms, which leads to low rationality and robustness of process solutions. 3. This invention employs data augmentation and dataset partitioning in steps (1-4) and model testing, verification, and finalization in steps (1-10). By expanding the training samples to cover new assembly scenarios through data augmentation and continuously optimizing the model's generalization ability through a closed-loop iterative mechanism of "training-verification-testing," it can solve the technical problem that existing expert system-based methods have low rule matching success rates and are difficult to adapt to the rapid iteration needs of products when facing new assembly scenarios of complex products. 4. This invention employs global feature encoding in steps (1-7) and dual-branch output layer parameter generation in steps (1-8). It integrates multi-source information through global feature encoding, and then accurately outputs discrete and continuous process parameters through classification and regression heads. The generated process scheme can be directly imported into Inte3D for simulation verification, forming a complete closed loop of "data-knowledge-model-simulation-optimization". Therefore, it can solve the technical problem that the existing three types of methods are prone to assembly interference and unreasonable parameters due to the lack of closed-loop optimization, which leads to a high rework rate in actual production. 5. The process of this invention has a high degree of automation: relying on the standardized design of the entire process from data preprocessing in step (1-2) to model finalization in step (1-10), end-to-end automation is achieved from raw data input to final process solution output, without the need for manual intervention in the middle, which greatly reduces repetitive operations in process design and further improves design efficiency. 6. The present invention has strong resource adaptability: by establishing the mapping relationship between production resources and process data through steps (1-3), and by generating a scheme by combining equipment capacity parameter constraints in the output layer of steps (1-8), the process parameters are always adapted to the existing equipment models, processing precision and other resource conditions of the enterprise, avoiding the generation of a scheme that is "theoretically feasible but cannot be implemented in practice", and reducing production adjustment costs; 7. This invention can achieve dynamic iterative optimization: the model iterative training in steps (1-9) and the performance verification in steps (1-10) form a dynamic update mechanism. With the continuous input of new process data and new production resources from enterprises, the model can continuously optimize parameter weights, continuously improve the accuracy and adaptability of process solutions, and adapt to the long-term development needs of enterprises. 8. The present invention has good cross-scenario compatibility: Step (1-2) designs differentiated preprocessing schemes for structured and unstructured multi-source data, and the feature extraction and encoding modules in steps (1-6) and (1-7) support the process generation requirements of different types of products (such as precision electronic components and large mechanical parts), and are adapted to the production mode of multiple varieties and small batches. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the automatic assembly process generation method based on the collaboration of Inte3D and InteAI according to the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0017] To address the aforementioned shortcomings or improvement needs of existing automated assembly process generation methods, this invention provides an automated assembly process generation method based on the collaboration of Inte3D and InteAI. This method aims to solve the technical problems of existing methods based on 3D process planning systems that over-rely on model geometric information and are disconnected from actual production scenarios; existing methods based on industrial intelligent platforms that neglect tacit knowledge and have insufficient robustness of process solutions; existing methods based on expert systems that struggle to adapt to rapid product iteration and have low success rates in matching rules for new scenarios; and existing methods lacking closed-loop optimization and prone to assembly interference and unreasonable parameters. Simultaneously, it further enhances the automation level, resource adaptability, and dynamic iteration capabilities of assembly process design, enabling the systematic reuse and standardized implementation of process knowledge, reducing reliance on manual labor and production rework costs, and meeting the market demands of flexible production and rapid iteration in modern manufacturing.

[0018] like Figure 1 As shown, this invention provides a method for automatically generating assembly processes based on Inte3D and InteAI, including the following steps: (1) Obtain the relevant files of the part to be processed input by the user, and preprocess the relevant files of the part to be processed to obtain the preprocessed relevant files; In this step, the relevant files for the parts to be processed can be 3D digital model files or part information. The part information includes design requirements and available production resources.

[0019] Specifically, the preprocessing of the relevant files for the part to be processed in this step involves first converting the relevant files into a lightweight model format (such as STEP format) using the Inte3D tool, and then integrating this lightweight model file with the scattered part information to obtain a uniform and efficiently parsable intermediate model file B. Then, the intermediate model file B is cleaned up for geometric noise and redundant data (i.e., geometric noise such as small fragmented surfaces, repeated surfaces, and non-manifold edges are removed from the model using a geometric topology verification algorithm, and redundant descriptions unrelated to processing are deleted from the part information) to obtain the processed intermediate model file C (which is a file with a regular geometric structure and concise information). Subsequently, redundant data is cleaned from the processed intermediate model file C (i.e., invalid information irrelevant to production and processing is removed, and then feature consistency and data deduplication optimization are performed on the intermediate model file after removing invalid information to verify the logical consistency of geometric features. Finally, duplicate feature parameters and resource information entries are removed by hash comparison) to obtain the secondary processed intermediate model file D (which is a file with accurate features and no duplicate data). Finally, the secondary processed intermediate model file D is subjected to processing adaptation standardization processing (i.e., geometric features and key attributes are structurally encapsulated according to production and processing specifications, and compatibility annotations corresponding to available production resources are added) to obtain the preprocessed related files.

[0020] The advantage of this step (1) is that the four-level preprocessing process of "format conversion - noise cleaning - redundancy removal - standardization processing" has achieved the unified standardization of multiple types of input files, which not only reduces the computational complexity of model parsing, but also ensures that the preprocessing results are compatible with the subsequent process generation requirements through compatibility labeling, laying the foundation for accurate feature extraction of the model.

[0021] (2) Input the pre-processed relevant files obtained in step (1) into the pre-established process generation model to obtain the assembly process of the part to be processed.

[0022] The process generation model of the present invention includes a feature extraction module, a knowledge fusion module, and a process decision module. The specific structure of the process generation model is as follows: The feature extraction module includes a Convolutional Neural Network (CNN), a Transformer encoder, and a data normalization module, among which: The input to the convolutional neural network is a 3D digital model file of multiple parts to be processed. The convolutional neural network extracts the geometric feature vectors (such as hole depth, slot width, shaft diameter tolerance, etc.) of all the parts to be processed from the 3D digital model file and outputs them. The input to the Transformer encoder is unstructured data of multiple parts to be machined (including expert experience text and technical manuals). The Transformer encoder extracts semantic feature vectors (such as "Recommended milling speed for 45# steel is 1500 r / min") from the unstructured data and outputs them. The input to the data standardization module is the production resource data of multiple parts to be processed (including unified equipment parameter units and tool model naming rules). The data standardization module extracts resource feature vectors from the production resource data and outputs them. The knowledge fusion module takes the geometric feature vector, semantic feature vector, and resource feature vector output by the feature extraction module as input. The knowledge fusion module performs cross-modal fusion processing on the input to obtain a fused feature matrix and output it.

[0023] Specifically, the knowledge fusion module first calculates the weights of the geometric feature vector, semantic feature vector, and resource feature vector through an attention mechanism (the sum of the three weights is 1, such as geometric feature weight 0.4, semantic feature weight 0.3, and resource feature weight 0.3). Subsequently, for the high-weight feature vectors, the knowledge fusion module first normalizes them using the Z-score normalization algorithm, then preprocesses the normalized feature vectors with a Gaussian filter algorithm of 3×3 kernel size and standard deviation σ=0.8 for noise reduction. Finally, a weighted average fusion algorithm is used to process each preprocessed feature vector (the formula is...). Where Wᵢ is the i-th high weight output by the attention mechanism, and has , Given the preprocessed i feature vectors (where n represents the total number of feature vectors participating in the fusion), obtain the initial core fusion features corresponding to each feature vector. Principal Component Analysis (PCA) was used to analyze all the initial core fusion features obtained. Dimensionality reduction is performed (principal components with a cumulative variance contribution rate ≥ 95%) to obtain unified core fusion features across multiple dimensions. ; Meanwhile, for the medium-weighted feature vectors, the knowledge fusion module first preprocesses them using the Min-Max normalization algorithm. Then, it uses the Locally Linear Embedding (LLE) algorithm (with 12 nearest neighbors) to process the preprocessed feature vectors, resulting in multiple medium-weighted feature vectors with optimized dimensions. ; Meanwhile, for low-weight feature vectors, the knowledge fusion module first preprocesses them using the Min-Max normalization algorithm, and then extracts core effective information from the preprocessed feature vectors using the LLE algorithm (with the nearest neighbor number set to 12). This core effective information refers to the feature dimensions and corresponding values ​​in the low-weight feature vectors that, after removing redundant, noise, and irrelevant dimensions, can represent the core attributes of industrial process knowledge and have a practical contribution to subsequent retrieval / matching / decision-making tasks. This results in multiple low-weight feature vectors with optimized dimensions. ; Subsequently, the knowledge fusion module extracts weighted feature vectors from all the data. and all low-weight feature vectors Screening and core fusion features For multiple feature vectors with a correlation ≥ 0.6, an adaptive weighted fusion algorithm is used to sequentially integrate all the selected medium-weighted feature vectors and low-weighted features into the core fusion feature. During the integration process, the gradient descent algorithm (learning rate 0.01, iteration count 1000) is used to dynamically adjust the weights of the high-weight feature vectors and low-weight feature vectors during the integration process, ensuring that the proportion of high-weight feature vectors in the integrated feature vectors is greater than or equal to 60%, thereby obtaining the initial fused feature set. Then, the parameter matching rules in the knowledge graph (specifically, a set of logical rules used to constrain the relationships between industrial feature parameters, such as if the "tightening torque" feature parameter of a certain assembly process is 5 N·m, then its corresponding "bolt specification" parameter must match M6) and the pre-stored process association data are called to correct the initial fused feature set. (First, the relationship between each feature parameter in the initial fused feature set is checked one by one, and abnormal feature parameters A that do not conform to the rules are marked; then, abnormal feature parameters A are compared with historical process association data to filter out candidate parameter values ​​B that conform to the rules; then, based on candidate parameter values ​​B, abnormal parameters in the initial fused feature set are replaced; finally, this correction process is repeated until all parameters in the initial fused feature set meet the rule requirements) to obtain the corrected fused feature set. Finally, conflict detection and resolution are performed on the corrected fused feature set (to resolve feature conflicts such as mismatch between expert experience parameters and device capabilities in the corrected fused feature set) to form a fused feature matrix and output it: The above process is as follows: First, a pre-established conflict detection rule base is invoked to traverse the corrected fusion feature set dimension by dimension to obtain all conflicting features (which refer to features in the corrected fusion feature set that violate the actual logic of industrial processes and parameter correlation constraints, such as the feature "low-temperature special solder" appearing in "high-temperature welding process"), as well as the dimension and type of each conflicting feature, and to record the parameter value and weight of each conflicting feature; then, all conflicting features are sorted in descending order of their weights, and the top 30% of the conflicting features in the sorting results are retained as high-priority conflicting features, and the remaining conflicting features are retained as low-priority conflicting features; weighted harmonic correction is then applied to all low-priority conflicting features to obtain the weighted harmonic correction parameter value P of the low-priority conflicting features. final (specifically) , Let be the parameter value of the i-th low-priority conflict feature. Let P be the weight of the j-th low-priority conflicting feature, and n represent the total number of low-priority conflicting features. Finally, calculate the parameter values ​​P of all high-priority conflicting features and the weighted harmonic correction of the low-priority conflicting features. final Integrate by dimension order to obtain and output the fused feature matrix.

[0024] The process decision module takes as input the fusion feature matrix output by the knowledge fusion module. The process decision module uses a genetic algorithm (GA) to process the fusion feature matrix to obtain the optimal set of process decision parameters (including processing temperature, feed rate, tool type, and processing sequence) that are suitable for the current production scenario.

[0025] Specifically, the process decision module first maps the acquired fused feature matrix through a fully connected network to obtain a candidate set of process parameters. This candidate set includes the key parameter ranges corresponding to each machining method (e.g., "milling - speed 1200-1600 r / min, feed rate 0.15-0.3 mm / r, depth of cut 2-5 mm" and "drilling - speed 800-1200 r / min, feed rate 0.1-0.2 mm / r, depth of cut 15-30 mm"), as well as the candidate equipment model and candidate tool type corresponding to each machining method. Then, the obtained candidate set of process parameters is probabilistically sorted using the Softmax activation function, and the constraint satisfaction problem (C) is solved. The SPs algorithm performs constraint matching and feasibility verification on the probability sorting results to obtain multiple effective process parameter combinations (including specific parameter value combinations, suitable equipment combinations, and matching tool combinations; all effective process parameter combinations satisfy all hard constraints and are sorted according to probability priority). Finally, a directed acyclic graph (DAG) is constructed based on all the obtained effective process parameter combinations and the process dependencies corresponding to each effective process parameter combination. A process sorting algorithm (such as topology sorting) is then used to integrate and optimize the order of all effective process parameters contained in the DAG to obtain a complete assembly process plan for all parts to be processed, including process sequence, processing parameters, equipment selection, and matching tools.

[0026] The process generation model in this invention is obtained through the following steps: (1-1) Obtain multi-source data of enterprises, including structured data and unstructured data. Classify and organize the obtained multi-source data of enterprises (that is, classify structured data into one category and unstructured data into another category) to obtain the classification and organization results as the original dataset. Specifically, the structured data of an enterprise includes 3D digital model files from its Computer-Aided Design (CAD) / Product Lifecycle Management (PLM) system (which include the geometric dimensions and topology of parts), production resource data from its Enterprise Resource Planning (ERP) / Manufacturing Execution System (MES) system (which includes equipment models, machining accuracy, and tool parameters), and historical process data from its process database (including verified processes, parameters, quality inspection standards, etc.). The unstructured data of an enterprise includes technical manuals, expert experience records, and simulation logs. (1-2) Perform data preprocessing on the original dataset obtained in step (1-1) to obtain the preprocessed dataset; For the structured data in the original dataset, the data preprocessing process in this step specifically involves: lightweighting the 3D model files (including removing redundant geometric surfaces and simplifying surface features); cleaning the production resource data (including removing outliers, such as invalid records of equipment failures, and completing missing values, such as tool life parameters); and standardizing the format of historical process data (including unifying parameter units, such as "cutting speed 100m / min", unifying formats, such as "VC=100m / min", and unifying process naming rules). For the unstructured data in the original dataset, the data preprocessing in this step specifically involves: performing structured parsing and key information extraction on the technical manual to obtain structured entries; using Natural Language Processing (NLP) technology to perform text segmentation and entity extraction on the expert experience records (i.e., extracting the correlation between materials and parameters); and performing text structure transformation on the simulation logs (e.g., transforming "excessive assembly gap causes vibration" into an assembly gap-vibration intensity correlation table). The process of structured parsing and key information extraction from technical manuals involves the following steps: First, the manual is broken down into chapters (e.g., "Part Specifications," "Processing Technology," "Equipment Requirements," "Quality Standards"), and the process specification modules (e.g., welding process details, component assembly flow modules) within each chapter are located and identified. Then, information within all process specification modules is extracted through keyword matching, including assembly requirement clauses (e.g., "0402 packaged component welding temperature ≤ 260℃") and equipment operation specification texts (e.g., "Pick-and-place machine nozzle model compatibility rules"). Finally, the extracted information is organized into structured entries according to a predetermined field format (e.g., "process type - requirement content - parameter range - applicable scenario").

[0027] (1-3) Perform data annotation and association processing on the preprocessed dataset obtained in step (1-2) to obtain the secondary processed dataset and output it; Specifically, this step involves first labeling the structured data in the preprocessed dataset with the correspondence between part features and historical processes (e.g., "shaft diameter Φ30±0.005mm → grinding machine processing + speed 2000r / min"). Then, for the unstructured data in the preprocessed raw dataset, semantic labels are added (e.g., "abnormal working condition handling plan", "process optimization suggestions", etc.). Next, a data association engine is used to establish a mapping relationship between the labeled structured and unstructured data, including 3D digital model files, production resource data, and historical process data (e.g., a digital model of a certain part corresponds to the historical process of "equipment A + tool B + parameter C"). (1-4) Divide the dataset obtained in step (1-3) after secondary processing into a training set (for updating model parameters), a validation set (for adjusting hyperparameters), and a test set (for evaluating model performance) in a ratio of 7:2:1. Perform data augmentation on the training set (such as rotating and scaling the 3D model file to generate new samples, and making minor adjustments to the historical process data to generate candidate samples) to obtain the data augmented training set. (1-5) Initialize the process generation model to obtain the initialized process generation model; The initialization in this step specifically involves setting the initial values ​​of the weight parameters to random values ​​output by a truncated normal distribution with a standard deviation of 0.1, selecting the Adam optimizer as the optimizer for model training, setting the initial value of the bias parameter to 0, setting the initial learning rate to lr=0.0003, and adopting a step-by-step learning decay strategy with a step size of stepsize=200 and a decay weight of gamma=0.1, meaning that the learning rate is multiplied by 0.1 every 200 epochs. The advantages of the above sub-steps (1-1) to (1-5) are: to construct a complete training data and model preparation process of "data acquisition-preprocessing-labeling association-segmentation enhancement-model initialization", to provide high-quality data and a stable initial model foundation for subsequent model training, and to ensure the efficiency and reliability of the training process.

[0028] (1-6) For each sample in the training set after data augmentation processing obtained in step (1-4), the sample is input into the convolutional neural network in the process generation model initialized in step (1-5) to obtain the three fixed-length (256-dimensional) local feature vector sequences corresponding to the sample. Specifically, this step involves performing a three-layer convolution operation on each sample in the data-augmented training set obtained in steps (1-4) (with kernel sizes of 3×3, 3×3, and 1×1, stride of 1, and padding set to "same") to obtain three local key features corresponding to each sample. The first layer of convolution outputs numerical gradient features of continuous parameters such as machining temperature and feed rate. The second layer outputs category encoding features of discrete parameters such as tool type and solder type. The third layer outputs... The local key features are unified representation features that integrate feature dimensions and redundant information. Then, the local key features output from each convolutional layer are sequentially fed into the BatchNorm normalization layer for standardization (to reduce the impact of internal covariate shift on training), and then fed into the ReLU activation function to introduce non-linear expressive power (while suppressing gradient vanishing) to obtain three high-dimensional local feature maps after normalization and non-linear enhancement. Then, the three high-dimensional local feature maps are compressed through global average pooling to finally obtain three fixed-length (256-dimensional) local feature vector sequences corresponding to the sample. (1-7) For each sample in the training set after data augmentation processing obtained in step (1-4), the three fixed-length (256-dimensional) local feature vector sequences corresponding to the sample obtained in step (1-6) are input into the Transformer encoder in the process generation model to obtain a global feature encoding vector corresponding to the sample. Specifically, the Transformer encoder consists of six stacked coding units. Each coding unit includes an eight-head multi-head self-attention mechanism and a feedforward neural network with a hidden layer dimension of 1024, and is configured with layer normalization and residual connections.

[0029] Specifically, this step involves first using the Transformer encoder as a network structure A adapted for global feature encoding (all subsequent feature processing operations are based on this network structure A); then, based on the multi-head self-attention mechanism in this network structure A, calculating the association weights of any two features in each local feature vector sequence obtained in steps (1-6) (achieving global feature interaction), which serves as the interactive feature B corresponding to the local feature vector sequence after global association modeling; then, inputting the interactive feature B into a feedforward neural network, and processing it through two linear transformations and ReLU activation (performing nonlinear enhancement to strengthen the process semantic representation ability of the features and highlight key process information) to obtain the semantically enhanced feature C corresponding to the local process feature vector sequence; finally, performing layer normalization processing on the semantically enhanced feature C corresponding to all three local process feature vector sequences (combining residual connections to alleviate the gradient decay problem in deep network training), and finally outputting a fixed-dimensional (256-dimensional) global feature encoding vector corresponding to the sample. (1-8) For the training set in the training set after data augmentation processing obtained in step (1-4), the global feature encoding vector corresponding to the sample obtained in step (1-7) is input into the output layer of the process generation model to obtain the combined loss value corresponding to the sample. Specifically, the output layer is a dual-branch output layer structure consisting of a classification head and a regression head. The classification head is configured with a fully connected network containing a Softmax activation function, and the regression head is configured with a fully connected network containing a Linear activation function.

[0030] Specifically, this step involves the following steps: First, the classification head uses a linear transformation to map the features of the global feature encoding vector corresponding to each sample obtained in steps (1-7) to the category space of discrete process parameters, thereby obtaining the predicted probability distribution of discrete process parameters in that sample (including the predicted probability distribution of discrete parameters such as tool type, machining sequence, and solder type, used to achieve category probability modeling of discrete parameters); then, the regression head performs multi-level linear transformation and numerical calibration on the global feature encoding vector corresponding to each sample obtained in steps (1-7), to obtain the predicted values ​​of continuous parameters in that sample (including the predicted values ​​of continuous parameters such as machining temperature, feed rate, and depth of cut, used to adapt to the quantitative prediction requirements of continuous parameters); finally, based on the output of the classification head... The predicted probability distribution of discrete process parameters in the sample and the true discrete label of the sample are obtained, and the classification error Loss_CE is obtained by using cross-entropy loss. Then, based on the predicted values ​​of continuous parameters in the sample output by the regression head and the true continuous parameters of the sample, the regression bias Loss_MSE is obtained by using mean squared error loss, while setting the weight coefficient α=0.6 (to highlight the influence of discrete process parameters on the overall process scheme). Finally, based on the classification error Loss_CE and the regression bias Loss_MSE, the combined loss value Loss=α・Loss_CE+(1−α)・Loss_MSE corresponding to the sample is obtained (used to comprehensively reflect the prediction difference between the two types of parameters and provide a comprehensive loss feedback basis for model parameter updates). (1-9) For each sample in the training set after data augmentation processing obtained in step (1-4), the process generation model is iteratively trained using the backpropagation algorithm based on the combined loss value corresponding to the sample obtained in step (1-8) until the process generation model reaches the preset number of iterations (200 times in this invention), and the optimal parameters of the process generation model at this time are obtained, thereby obtaining the initially trained process generation model. Specifically, in this step, the gradient is calculated layer by layer using the backpropagation algorithm: from the gradient of the output layer... (Including classifier head weight gradient) Regression head weight gradient Begin by calculating the gradients of the Transformer encoder sequentially. (Attention layer weights) (Weights of the feedforward neural network), gradients of the convolutional layer (Convolution kernel weights) (Convolutional layer bias), and use the Adam optimizer configured in step (1-5) to update all parameters according to the gradient. The learning rate during the update is adjusted using a step decay strategy (stepsize=200, decay weight gamma=0.1). Finally, repeat the training loop from step (1-6) to (1-8) until the loss converges to the minimum value to obtain the initially trained process generation model.

[0031] (1-10) Use the test set obtained in step (1-4) to test the process generation model initially trained in step (1-9) until the detection accuracy reaches the optimal level, thereby obtaining the final trained process generation model.

[0032] The advantages of the above sub-steps (1-6) to (1-10) are: to construct a complete model training and optimization process of "local feature extraction - global feature encoding - loss calculation - iterative training - performance testing", to realize full-link control from feature learning to model finalization, and to ensure that the final generated process generation model has high-precision and strong generalization process parameter prediction capabilities.

[0033] The following uses "Automotive Transmission Bearing Assembly Process Generation" as an example to explain in detail the model training and process generation process: 1. Model Training Implementation (1-1) Obtain the 3D digital model of the gearbox bearing from the CAD system of an automobile factory (shaft diameter Φ25±0.003mm, bearing inner ring tolerance ±0.002mm), the grinding machine equipment data from the ERP system (model MG1432, machining accuracy 0.001mm, spindle speed range 1000-3000r / min), 100 bearing assembly history processes from the process library (including "rough grinding → fine grinding → cleaning → press fitting" process, fine grinding speed 2500r / min, feed rate 0.08mm / r), and expert experience "bearing press fitting force should be controlled at 6000-8000N", organize them into the original dataset and output them. (1-2) Preprocessing of the original dataset.

[0034] Specifically, redundant chamfer features are removed from the 3D digital model to reduce its weight; missing values ​​for "tool life 800h" are added to the grinding machine data; the format of "fine grinding speed 2500r / min" is standardized for historical processes; and the correlation between "pressing force and bearing model" is extracted from expert experience to obtain preprocessed data and output it. (1-3) Mark the correspondence between “shaft diameter Φ25±0.003mm → MG1432 grinding machine + fine grinding speed 2500r / min”, establish the digital model-equipment-process mapping, and obtain a dedicated dataset; (1-4) Divide the dataset into 7:2:1, rotate the numerical model of the training set to generate 50 new samples, and obtain training, validation and test data; (1-5) Initialize model parameters (CNN convolution kernel 3×3, Transformer layer number 6), Adam optimizer with a learning rate of 1e-4, combined loss function, iteratively train for 20 epochs to obtain the preliminary model and output it; (1-6) The validation set “process accuracy” is 88%. The learning rate is adjusted to 8e-5, the number of Transformer layers is increased to 8, and the accuracy is improved to 92% after retraining. The optimized model is obtained and output. (1-7) The test set accuracy was 93%, the parameter rationality was 96%, and the defect prediction rate was 87%, thus obtaining a qualified initial process generation model.

[0035] 2. Online process generation and implementation (1) Obtain the three-dimensional digital model of the new gearbox bearing (shaft diameter Φ28±0.004mm) uploaded by the user and available resources (MG1432 grinding machine, press fitting machine), preprocess it to obtain the data to be processed and output it; (2) The AI ​​large model extracts features such as "shaft diameter Φ28±0.004mm, material 45# steel" to obtain part feature data and output it; (3) Input the task scheduling engine, call the knowledge base "45# steel bearing process rules", drive the core AI service group to reason, obtain the preliminary process reasoning results and output them; (4) The process optimization algorithm optimizes and obtains the initial scheme of "rough grinding → fine grinding → cleaning → pressing" (fine grinding speed 2400r / min, pressing force 7500N) and outputs it; (5) Import the lightweight Inte3D model; (6) The simulation was performed to obtain the process data of "shaft diameter deformation of 0.003mm after press fitting"; (7) The analysis shows that "the deformation amount is within tolerance" and the simulation results are qualified; (8) No corrections are needed; proceed directly to the second simulation. (9) All indicators in the second simulation met the standards, and the optimized solution was obtained; (10) Store the solution and generate the work instruction, distribute it to the production system, and complete the process generation.

[0036] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatically generating assembly processes based on Inte3D and InteAI, characterized in that, Includes the following steps: (1) Obtain the relevant files of the part to be processed input by the user, and preprocess the relevant files of the part to be processed to obtain the preprocessed relevant files; (2) Input the pre-processed relevant files obtained in step (1) into the pre-established process generation model to obtain the assembly process of the part to be processed.

2. The method for automatically generating assembly processes based on Inte3D and InteAI according to claim 1, characterized in that, The relevant files for the parts to be processed in the steps are three-dimensional digital model files or part information, which includes design requirements and available production resources; The preprocessing of the relevant files of the part to be processed in step (1) is as follows: First, the relevant files are converted into a lightweight model format using the Inte3D tool. The lightweight model file is then integrated with the scattered part information to obtain a uniform and efficiently parsable intermediate model file B. Then, the intermediate model file B is cleaned up for geometric noise and redundant data to obtain a processed intermediate model file C. Subsequently, the processed intermediate model file C is cleaned up for redundant data to obtain a secondary processed intermediate model file D. Finally, the secondary processed intermediate model file D is processed for compatibility standardization to obtain the preprocessed relevant files.

3. The method for automatically generating assembly processes based on Inte3D and InteAI according to claim 1 or 2, characterized in that, The process generation model includes a feature extraction module, a knowledge fusion module, and a process decision module, with the following specific structure: The feature extraction module includes a Convolutional Neural Network (CNN), a Transformer encoder, and a data normalization module, wherein: The input to the convolutional neural network is a 3D digital model file of multiple parts to be processed. The convolutional neural network extracts the geometric feature vectors of all parts to be processed from the 3D digital model file and outputs them. The input to the Transformer encoder is unstructured data of multiple parts to be processed. The Transformer encoder extracts semantic feature vectors from the unstructured data and outputs them. The input to the data standardization module is the production resource data of multiple parts to be processed. The data standardization module extracts resource feature vectors from the production resource data and outputs them. The knowledge fusion module takes the geometric feature vector, semantic feature vector, and resource feature vector output by the feature extraction module as input. The knowledge fusion module performs cross-modal fusion processing on the input to obtain a fused feature matrix and output it. The process decision module takes the fusion feature matrix output by the knowledge fusion module as its input. The process decision module uses the genetic algorithm (GA) to process the fusion feature matrix to obtain the optimal set of process decision parameters that are suitable for the current production scenario.

4. The method for automatically generating assembly processes based on Inte3D and InteAI according to any one of claims 1 to 3, characterized in that, The specific processing procedure of the knowledge fusion module is as follows: First, the weights of the geometric feature vector, semantic feature vector, and resource feature vector are calculated using an attention mechanism; Subsequently, for the high-weight feature vectors, the knowledge fusion module first normalizes them using the Z-score normalization algorithm, then preprocesses the normalized feature vectors with a Gaussian filter algorithm of 3×3 kernel size and standard deviation σ=0.8 for noise reduction, and finally processes each preprocessed feature vector using a weighted average fusion algorithm. Where Wᵢ is the i-th high weight output by the attention mechanism, and has , Let i be the preprocessed feature vectors, and n represent the total number of feature vectors participating in the fusion, to obtain the initial core fusion feature corresponding to each feature vector. Principal component analysis (PCA) was used to analyze all the initial core fusion features obtained. Dimensionality reduction is performed to obtain unified core fusion features across multiple dimensions. ; Meanwhile, for the medium-weighted feature vectors, the knowledge fusion module first preprocesses them using the Min-Max normalization algorithm, and then uses the Local Linear Embedding (LLE) algorithm to process the preprocessed feature vectors to obtain multiple medium-weighted feature vectors with optimized dimensions. ; Meanwhile, for low-weight feature vectors, the knowledge fusion module first uses the Min-Max normalization algorithm for preprocessing, and then uses the LLE algorithm to extract core effective information from the preprocessed feature vectors to obtain multiple low-weight feature vectors with optimized dimensions. ; Subsequently, the knowledge fusion module extracts weighted feature vectors from all the data. and all low-weight feature vectors Screening and core fusion features For multiple feature vectors with a correlation ≥ 0.6, an adaptive weighted fusion algorithm is used to sequentially integrate all the selected medium-weighted feature vectors and low-weighted features into the core fusion feature. During the integration process, the gradient descent algorithm is used to dynamically adjust the weights of the high-weight feature vectors and low-weight feature vectors during the integration process, ensuring that the proportion of high-weight feature vectors in the integrated feature vectors is greater than or equal to 60%, thereby obtaining the initial fused feature set. Then, the parameter matching rules in the knowledge graph and the pre-stored process association data are called to correct the initial fusion feature set. That is, firstly, the association relationship of each feature parameter in the initial fusion feature set is checked one by one, and abnormal feature parameter A that does not conform to the rules is marked. Then, the abnormal feature parameter A is compared with historical process correlation data to select candidate parameter values ​​B that meet the rules; subsequently, based on the candidate parameter values ​​B, the abnormal parameters in the initial fusion feature set are replaced; finally, this correction process is repeated until all parameters in the initial fusion feature set meet the rule requirements to obtain the corrected fusion feature set. Finally, conflict detection and resolution are performed on the corrected fusion feature set to form a fusion feature matrix and output it.

5. The method for automatically generating assembly processes based on Inte3D and InteAI according to claim 4, characterized in that, The process of performing conflict detection and resolution on the corrected fused feature set to form a fused feature matrix and output it is as follows: First, a pre-established conflict detection rule base is called to traverse the corrected fused feature set dimension by dimension to obtain all conflict features, as well as the dimension and type of each conflict feature, and to record the parameter value and weight of each conflict feature. Then, all conflict features are sorted in descending order of their weights, and the top 30% of conflict features in the sorting result are retained as high-priority conflict features, while the remaining conflict features are designated as low-priority conflict features. Weighted harmonic correction is then applied to all low-priority conflict features to obtain the parameter values ​​of the weighted harmonic correction low-priority conflict features. , Let be the parameter value of the i-th low-priority conflict feature. Let be the weight of the j-th low-priority conflicting feature, and n represent the total number of low-priority conflicting features. Finally, all high-priority conflict features and the weighted harmonic correction of low-priority conflict feature parameter values ​​P final Integrate by dimension order to obtain and output the fused feature matrix.

6. The method for automatically generating assembly processes based on Inte3D and InteAI according to claim 5, characterized in that, The process decision module's processing steps are as follows: First, the obtained fusion feature matrix is ​​mapped through a fully connected network to obtain a candidate set of process parameters. This candidate set includes the key parameter ranges corresponding to each processing method, as well as the candidate equipment models and candidate tool types corresponding to each processing method. Then, the obtained candidate set of process parameters is probabilistically sorted using the Softmax activation function, and the constraint satisfaction problem solving (CSPs) algorithm is used to perform constraint matching and feasibility verification on the probabilistic sorting results to obtain multiple effective process parameter combinations. Finally, a directed acyclic graph (DAG) is constructed based on all the obtained effective process parameter combinations and the process dependencies corresponding to each effective process parameter combination. The process ordering algorithm is then used to integrate and optimize the order of all effective process parameters contained in the DAG to obtain a complete assembly process plan for all parts to be processed, including process sequence, processing parameters, equipment selection, and matching tools.

7. The method for automatically generating assembly processes based on Inte3D and InteAI according to claim 6, characterized in that, The process generation model is trained through the following steps: (1-1) Obtain multi-source data of enterprises, including structured data and unstructured data, classify and organize the obtained multi-source data of enterprises, and use the classification and organization results as the original dataset; (1-2) Perform data preprocessing on the original dataset obtained in step (1-1) to obtain the preprocessed dataset; (1-3) Perform data annotation and association processing on the preprocessed dataset obtained in step (1-2) to obtain the secondary processed dataset and output it; (1-4) Divide the dataset obtained in step (1-3) into training set, validation set and test set in a ratio of 7:2:1, and perform data augmentation on the training set to obtain the data augmented training set. (1-5) Initialize the process generation model to obtain the initialized process generation model; (1-6) For each sample in the training set after data augmentation processing obtained in step (1-4), the sample is input into the convolutional neural network in the process generation model initialized in step (1-5) to obtain the three fixed-length local feature vector sequences corresponding to the sample. (1-7) For each sample in the training set after data augmentation processing obtained in step (1-4), the three fixed-length local feature vector sequences corresponding to the sample obtained in step (1-6) are input into the Transformer encoder in the process generation model to obtain a global feature encoding vector corresponding to the sample. (1-8) For the training set in the training set after data augmentation processing obtained in step (1-4), the global feature encoding vector corresponding to the sample obtained in step (1-7) is input into the output layer of the process generation model to obtain the combined loss value corresponding to the sample. (1-9) For each sample in the training set after data augmentation processing obtained in step (1-4), the process generation model is iteratively trained using the backpropagation algorithm based on the combined loss value corresponding to the sample obtained in step (1-8) until the process generation model reaches the preset number of iterations (200 times in this invention), and the optimal parameters of the process generation model at this time are obtained, thereby obtaining the initially trained process generation model. (1-10) Use the test set obtained in step (1-4) to test the process generation model initially trained in step (1-9) until the detection accuracy reaches the optimal level, thereby obtaining the final trained process generation model.

8. The method for automatically generating assembly processes based on Inte3D and InteAI according to claim 7, characterized in that, The structured data of an enterprise includes 3D digital model files from its computer-aided design (CAD) / product lifecycle management (PLM) systems, production resource data from its enterprise resource planning (ERP) / manufacturing execution system (MES) systems, and historical process data from its process database. Unstructured data from enterprises includes technical manuals, expert experience records, and simulation logs; For the structured data in the original dataset, the data preprocessing process in step (1-2) specifically involves lightweighting the 3D model file, cleaning the production resource data, and standardizing the format of the historical process data. For the unstructured data in the original dataset, the data preprocessing process in steps (1-2) specifically involves: performing structured parsing and key information extraction on the technical manual to obtain structured entries; using Natural Language Processing (NLP) technology to perform text segmentation and entity extraction on expert experience records; and performing text structure conversion on simulation logs. The process of structured parsing and key information extraction of the technical manual is as follows: First, the technical manual is split into chapters and hierarchical sections, and the process procedure modules in each chapter are located and identified. Then, information in all process procedure modules is extracted through keyword matching, including assembly requirement clauses and equipment operation specification texts. Finally, the extracted information is organized into structured entries according to the predetermined field format.

9. The method for automatically generating assembly processes based on Inte3D and InteAI according to claim 8, characterized in that, Steps (1-3) are as follows: First, for the structured data in the preprocessed dataset, the correspondence between part features and historical processes is labeled; for the unstructured data in the preprocessed original dataset, semantic labels are labeled. Then, a mapping relationship is established between the labeled structured and unstructured data, the three-dimensional digital model file, the production resource data, and the historical process data through a data association engine. The initialization in steps (1-5) specifically involves setting the initial values ​​of the weight parameters to random values ​​output by a truncated normal distribution with a standard deviation of 0.1, selecting the Adam optimizer as the optimizer for model training, setting the initial value of the bias parameter to 0, setting the initial learning rate to lr=0.0003, adopting a step-wise learning decay strategy with a step size of stepsize=200 and a decay weight of gamma=0.

1. Steps (1-6) are as follows: First, each sample in the training set after data augmentation in step (1-4) undergoes a three-layer convolution operation to obtain three local key features corresponding to the sample. The local key features output by the first convolution operation are the numerical gradient features of continuous parameters such as processing temperature and feed rate. The local key features output by the second convolution operation are the category encoding features of discrete parameters such as tool type and solder type. The local key features output by the third convolution operation are the unified representation features that fuse feature dimensions and redundant information. Then, the local key features output by each convolution operation are sequentially fed into the BatchNorm normalization layer for standardization, and then fed into the ReLU activation function to introduce nonlinear expressive power to obtain three high-dimensional local feature maps after normalization and nonlinear enhancement. Finally, the three high-dimensional local feature maps are compressed by global average pooling to obtain three fixed-length (256-dimensional) local feature vector sequences corresponding to the sample.

10. The method for automatically generating assembly processes based on Inte3D and InteAI according to claim 9, characterized in that, The Transformer encoder consists of 6 stacked coding units. Each coding unit includes an 8-head multi-head self-attention mechanism and a feedforward neural network with a hidden layer dimension of 1024, and is configured with layer normalization and residual connections. Steps (1-8) are as follows: First, the Transformer encoder is used as the network structure A for adapting global feature encoding. Then, based on the multi-head self-attention mechanism in network structure A, the association weights of any two features in each local feature vector sequence obtained in step (1-6) are calculated as the interaction feature B corresponding to the local feature vector sequence after global association modeling. Then, the interaction feature B is input into the feedforward neural network and processed by two linear transformations and ReLU activation to obtain the semantically enhanced feature C corresponding to the local process feature vector sequence. Finally, layer normalization is performed on the semantically enhanced features C corresponding to all three local process feature vector sequences, and finally, a fixed-dimensional global feature encoding vector corresponding to the sample is output. The output layer is a two-branch output layer structure with a classification head and a regression head. The classification head is configured with a fully connected network containing a Softmax activation function, and the regression head is configured with a fully connected network containing a Linear activation function. Steps (1-8) are as follows: First, the classification head maps the features of the global feature encoding vector corresponding to each sample obtained in step (1-7) to the category space of discrete process parameters through linear transformation to obtain the predicted probability distribution of discrete process parameters in the sample. Then, the regression head performs multi-level linear transformation and numerical calibration on the global feature encoding vector corresponding to each sample obtained in step (1-7) to obtain the predicted value of continuous parameters in the sample. Then, based on the predicted probability distribution of discrete process parameters in the sample output by the classification head and the true discrete label of the sample, the classification error Loss_CE is obtained by using cross-entropy loss. Then, based on the predicted value of continuous parameters in the sample output by the regression head and the true continuous parameters of the sample, the regression bias Loss_MSE is obtained by using mean squared error loss, while setting the weight coefficient α=0.

6. Finally, based on the classification error Loss_CE and the regression bias Loss_MSE, the combined loss value Loss=α・Loss_CE+(1−α)・Loss_MSE corresponding to the sample is obtained.

Images