Discrete manufacturing full-life-cycle intelligent computing system and method

Abstract

The invention discloses a discrete manufacturing full-life-cycle intelligent computing system and method, the system comprises a data knowledge layer, a computing adaptation layer and a parallel service layer, a virtual matching mechanism is used between the data knowledge layer and the computing adaptation layer, that is, primary mapping is formed according to task demand description and manufacturing resource description in a twin world, an optimal algorithm is called for a specific task, the most reasonable resources are adapted, an optimal service strategy is obtained, and a mapping process from a service demand to the service strategy is realized; a virtual-real parallel computing mechanism is used between a computing adaptation layer and a parallel service layer, that is, a mutual linkage relation is established in a physical world according to a service execution strategy obtained in a twin world, so that the requirement of discrete manufacturing full-life-cycle intelligent service is met, and the mapping process from a service strategy to service execution is achieved. And through two times of mapping, autonomous computing adaptation of task-knowledge-algorithm-resource in the whole life cycle of discrete manufacturing is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of discrete manufacturing, and in particular relates to a discrete manufacturing full life cycle intelligent computing system and method. Background Art

[0002] Discrete manufacturing, characterized by discrete processes, high-variety, small-batch production, mixed-line production, and technologically advanced processes, is a key focus of the manufacturing industry's digital transformation. However, my country's discrete manufacturing industry started late, facing challenges such as weak foundational technologies, a lack of core technologies, and high production costs. Digital technologies are urgently needed to empower the transformation and upgrading of traditional discrete industries. Improving the digitalization of the discrete industry will address the pain points, difficulties, and bottlenecks in the process of high-quality development, supporting discrete manufacturing companies in improving quality, reducing costs, and increasing efficiency. Therefore, vigorously developing discrete manufacturing is an inevitable choice for enhancing my country's core industrial competitiveness and is of great strategic significance for accelerating China's economic development model and achieving the transformation from a manufacturing power to a manufacturing powerhouse.

[0003] Intelligent computing for the entire discrete manufacturing lifecycle reshapes the entire lifecycle by combining virtual and real approaches, establishing a new "physical-twin" alternating and parallel process and computational approach. For example, traditional R&D and design are followed by production according to blueprints, which can result in products failing to achieve expected performance. Adding a virtual dynamic simulation of design results between the design and manufacturing stages can effectively prevent substandard product designs. Traditional manufacturing for large-scale equipment is followed by integrated assembly and commissioning, followed by disassembly and transportation to on-site assembly and operation. This is costly and inefficient. Replacing these physical steps with virtual integrated assembly and commissioning can significantly improve production efficiency and reduce costs. Therefore, leveraging technologies such as the Industrial Internet, big data, digital twins, and machine learning to build a data-knowledge-driven intelligent computing paradigm, enabling intelligent and collaborative reshaping of all lifecycle stages, can significantly enhance the lean management and control capabilities of discrete manufacturing and the intelligent operation and maintenance services of products.

[0004] Intelligent computing is a broad concept. It generally refers to the autonomous execution of computational adaptation across the entire lifecycle, including product design, manufacturing, integrated assembly, and operations and maintenance services, based on the "task-knowledge-algorithm-resource" framework, supported by production data, prior knowledge, and process flows. This allows for the provision of parallel intelligent services to complete designated tasks at the lowest cost. Therefore, intelligent computing in discrete manufacturing involves all aspects and elements of the entire lifecycle, requiring comprehensive consideration of three key issues: 1. How to collaboratively analyze, integrate, evaluate, and predict multi-source, massive, and multimodal information throughout the entire lifecycle, and combine prior knowledge to construct a full-lifecycle knowledge graph to provide data and knowledge support for service-oriented manufacturing of heavy mining equipment; 2. How to leverage fused data and knowledge graphs to design intelligent computing paradigms to address the autonomous computational adaptation of the "task-knowledge-algorithm-resource" framework throughout the entire lifecycle, ensuring the smooth implementation of service-oriented manufacturing for heavy mining equipment; and 3. How to leverage digital twins and intelligent computing technologies to provide parallel intelligent services for product design, production reconfiguration, assembly testing, and predictive maintenance, thereby improving the efficiency and effectiveness of discrete manufacturing. Summary of the Invention

[0005] In order to overcome the above shortcomings, the present invention provides a discrete manufacturing full life cycle intelligent computing system and method, which utilizes physical-twin parallel mapping, based on production fusion information and process knowledge, and only requires virtual assembly and trial operation in the twin world to ensure product quality, greatly improve production efficiency and reduce costs.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: A discrete manufacturing full lifecycle intelligent computing system, including a data knowledge layer, a computing adaptation layer, a parallel service layer, and a mapping of each link in the full lifecycle into the physical world and the digital twin world; The data knowledge layer is used to collaboratively integrate multi-source and multi-modal data throughout the entire life cycle of discrete manufacturing and build a dynamic and scalable knowledge graph for the entire life cycle; The computing adaptation layer is driven by the knowledge graph and large model to achieve autonomous computing adaptation of tasks, knowledge, algorithms and resources, including task requirement perception, dynamic algorithm generation and resource matching; The parallel service layer performs intelligent service tasks of virtual-reality mapping through digital twin technology, including dynamic design deduction, virtual production reconstruction and virtual assembly trial operation. Further optimization, the computation adaptation layer implements cross-layer interaction through the following mapping function: The intelligent mapping function from the data knowledge layer to the computing adaptation layer is: Among them, s i represents the information representation vector of the i-th task, r represents the production resource information vector, Dt represents the data fusion information of round t, K t represents the knowledge graph constructed in round t, π i represents the service strategy of task i, Φ(·) is the autonomous computing adaptation mapping function, which enables the completion of intelligent computing tasks at the minimum cost; The intelligent mapping function from the computing adaptation layer to the parallel service layer is: Among them, r p Indicates the actual resources available, r t represents the discrete manufacturing parallel service strategy in round t, Ψ(·) is the parallel intelligent service mapping function; The intelligent mapping function from the computation adaptation layer to the data knowledge layer is defined as: Among them, D new Represents the newly generated data in each link of the whole life cycle, K new represents the new knowledge generated in the parallel intelligent service process, and Θ(·) represents the data-knowledge intelligent mapping function.

[0007] Further optimization, the data knowledge layer includes a multimodal data collaborative fusion module and a knowledge graph construction module, wherein the multimodal data collaborative fusion module cleans data based on federated learning and Shapley value evaluation, and generates a high-precision data set through an adaptive filtering algorithm, and the knowledge graph construction module uses a mask contrast learning model to identify full life cycle events and relationships, and completes the knowledge graph through a graph neural network.

[0008] Further optimization, the computing adaptation layer includes a large model fusion module and an autonomous adaptation module, wherein the large model fusion module aligns the probability distribution of multiple models through dynamic programming and fuses representations based on minimum cross entropy, and the autonomous adaptation module realizes task and knowledge matching based on knowledge graph search, realizes knowledge and algorithm adaptation based on graph reasoning, and realizes algorithm and resource adaptation based on joint reasoning.

[0009] Further optimization, the parallel service layer includes a twin scenario reconstruction module and a virtual-reality execution module, wherein the twin scenario reconstruction module uses three-dimensional reconstruction and motion model estimation to construct a digital mapping of the physical world and the twin world, and the virtual-reality execution module simulates and deduces in the twin world according to the service strategy, and dynamically adjusts the physical world execution plan through perturbation method and causal response optimization strategy.

[0010] Further optimization also includes a resource dynamic scheduling module, which is based on the resource demand prediction model of deep reinforcement learning, and completes the resource demand matrix in real time through matrix decomposition. It also uses the resource scheduling strategy based on the Stackelberg multi-objective game, combined with causal discovery and graph generation models to supplement abnormal resources.

[0011] Further optimization, the Stackelberg game model of the resource scheduling strategy is defined as: Γ{Υ task , γ resource ;π={y task ,y resource}; U1, ..., U m ;M resource} Among them, task and Y resource Represents task controller and resource controller respectively, y task represents the resource demand planning strategy formulated based on resource demand forecast, y resource represents the resource allocation strategy, U i is the i-th utility objective function, i∈{1,...,m}, m represents the number of utility objectives, M resource Represents existing resource information.

[0012] Further optimization, the execution process of the virtual-to-real mapping strategy includes: (1) Fusion of point cloud features from the physical world and the twin world through an attention mechanism; (2) Optimize the 3D reconstruction model based on geometric loss and motion joint loss to generate dynamic service scenarios.

[0013] A calculation method for a discrete manufacturing full life cycle intelligent computing system includes the following steps: S1. Data knowledge layer: Collects multi-source heterogeneous data from the physical world, transforms raw data into structured knowledge through deep multimodal data collaborative fusion technology, and constructs a knowledge graph covering the entire life cycle; S2, Computing Adaptation Layer: Based on the pre-trained large model, it analyzes task requirements, combines structured knowledge in the knowledge graph, dynamically matches algorithm libraries and resource pools, and designs a "task-knowledge-algorithm-resource" adaptation mechanism. Through dynamic resource scheduling strategies, it achieves flexible allocation of computing resources and optimal algorithm selection. S3, Parallel Service Layer: Based on the computational adaptation results, a twin world of the physical world is constructed to achieve virtual mapping and real-time synchronization of equipment, processes, and procedures. Through simulation deduction and causal response mechanisms, service strategies are dynamically optimized and optimization instructions are fed back to the physical world.

[0014] Further optimization will feed back the execution results of the physical world to the twin world in real time, triggering knowledge graph updates and computing adaptation strategy adjustments.

[0015] The beneficial effects of the present invention are: 1. Aiming at intelligent computing issues throughout the entire life cycle of discrete manufacturing, a three-layer, two-world intelligent computing model has been constructed. This model decomposes intelligent computing into a data knowledge layer, a computing adaptation layer, and a parallel service layer. The entire life cycle is mapped into the physical world and the twin world. This model enables autonomous computing adaptation and parallel intelligent services across the entire discrete manufacturing life cycle, encompassing tasks, knowledge, algorithms, and resources. This model completes computing tasks and service requirements across all stages of the life cycle at minimal cost, improving the efficiency and effectiveness of discrete manufacturing. 2. Through two-way closed-loop interaction between the physical world and the twin world, this method connects the entire chain of data management, intelligent computing and smart services, solves the problems of data silos, low computing efficiency and service lag in discrete manufacturing, and promotes the upgrading of the manufacturing process to intelligence and service. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of the technical framework of the intelligent computing model for the entire life cycle of discrete manufacturing; Figure 2 This is a schematic diagram of the "three-layer-two-world" intelligent computing model architecture for the discrete full life cycle; Figure 3 This is a schematic diagram of the big model architecture for intelligent computing of data throughout the entire life cycle of discrete manufacturing; Figure 4 Schematic diagram of the smart service strategy solution driven by digital twins. DETAILED DESCRIPTION

[0017] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is described in detail below in conjunction with specific embodiments. The following embodiments are implemented based on the technical solutions of the present invention, and provide detailed implementation methods and specific operating procedures. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the following embodiments.

[0018] A full-lifecycle intelligent computing system for discrete manufacturing includes a data knowledge layer, a computing adaptation layer, a parallel service layer, and a method for mapping each stage of the full lifecycle into the physical world and the digital twin world. The data knowledge layer: The intelligent management and control method for full-lifecycle data is based on data collected from each stage of the full lifecycle, existing production and manufacturing processes, and prior knowledge. A deep multimodal data collaborative fusion method is designed to establish a full-lifecycle knowledge graph based on masked contrast learning and large models. The computing adaptation layer: The large-model-driven autonomous computing adaptation method, based on the full-lifecycle fusion data and knowledge graph, designs a full-lifecycle intelligent computing large model, a task-knowledge-algorithm-resource autonomous computing adaptation method, and a dynamic resource scheduling strategy. The parallel service layer: The digital twin-driven intelligent service strategy: Mainly based on the results of intelligent computing adaptation, a digital twin-based intelligent service reconstruction model is constructed, and a service execution strategy based on virtual-reality mapping and a service execution strategy optimization method based on causal response are designed. The physical world and the twin world collaborate with each other through virtual comparison mapping to complete the established tasks at the lowest cost. That is, by utilizing physical-twin parallel mapping, based on production fusion information and process knowledge, only virtual assembly and trial operation are required in the twin world to ensure product quality, greatly improve production efficiency and reduce costs.

[0019] The three-layer-two-world intelligent computing model is formally expressed as follows: the intelligent mapping function from the data knowledge layer to the computing adaptation layer is defined as Among them, s i represents the information representation vector of the i-th task, r represents the production resource information vector, D t represents the data fusion information of round t, K t represents the knowledge graph constructed in round t, π i represents the service strategy of task i, and Φ(·) is the autonomous computing adaptation mapping function, which enables the completion of intelligent computing tasks at the minimum cost.

[0020] The intelligent mapping function from the computation adaptation layer to the parallel service layer is defined as Among them, r p Represents the existing actual resources, R t represents the discrete manufacturing parallel service strategy in round t, and Ψ(·) is the parallel intelligent service mapping function.

[0021] The intelligent mapping function from the computation adaptation layer to the data knowledge layer is defined as Among them, D new Represents the newly generated data in each link of the whole life cycle, K newrepresents the new knowledge generated in the parallel intelligent service process, and Θ(·) represents the data-knowledge intelligent mapping function.

[0022] A computational method for a discrete manufacturing full-lifecycle intelligent computing system operates as follows: A virtual matching mechanism is used between the data knowledge layer and the computational adaptation layer. This involves mapping the task requirements and manufacturing resource descriptions in the twin world. This allows the optimal algorithm to be called for a specific task, adapting the most appropriate resources to obtain the optimal service strategy, and implementing the mapping process from service requirements to service strategies. Furthermore, a virtual-real parallel computing mechanism is used between the computational adaptation layer and the parallel service layer. This establishes a mutual linkage in the physical world based on the service execution strategy obtained in the twin world to meet the intelligent service requirements of discrete manufacturing throughout its lifecycle and implement the mapping process from service strategy to service execution. Through these two mappings, autonomous computational adaptation of tasks, knowledge, algorithms, and resources throughout the discrete manufacturing lifecycle is achieved. A detailed explanation is provided below.

[0023] S1. Data knowledge layer Aiming at the multi-source, massive, and multi-modal production data of discrete manufacturing, a deep multimodal data collaborative fusion method for the entire life cycle is designed; based on existing production and manufacturing processes, prior knowledge, etc., a dynamic and scalable full-life cycle knowledge graph is established, laying the data knowledge foundation for discrete manufacturing intelligent computing.

[0024] (1) Data collaborative fusion method for the entire life cycle First, design a data cleaning method for the entire discrete manufacturing life cycle: Step 1: Based on the multi-source, massive, multi-modal full life cycle dataset D = {x1,...,x n},in Indicates the Data of each link, The Shapley value is used to represent the value of data in the large-scale intelligent computing model for the entire discrete manufacturing lifecycle. The larger the Shapley value, the greater the contribution of the data to the accuracy of the large-scale intelligent computing model. If data cleaning significantly improves the Shapley value, the value of the cleaned data is also higher. Step 2: Due to the dispersion of data throughout its life cycle, the Shapley value of the data is calculated using federated learning. calculate The Shapley values ​​of all candidate values ​​and the mean square error between them. The mean square error is used to measure the data The contribution of different candidate values ​​to the accuracy of the intelligent computing large model is different. The larger the value, the greater the difference in contribution. Step 3: Obtain a high-quality full-lifecycle dataset D * , providing a data foundation for large intelligent computing models.

[0025] Secondly, a filtering method for discrete manufacturing full life cycle data is designed. The cleaned full life cycle data still contains a lot of interference information. In order to improve data accuracy, it is necessary to filter the data of each link. Specifically, based on each link Based on the data filtering requirements description and historical filtering algorithm evaluation results, the intelligent computing model is used to generate the optimal filtering algorithm for this link. Obtain high-precision data. Since the full life cycle data is multi-source and multi-modal, the adaptive lifting algorithm is used to process the filtered data of each link to obtain high-precision full life cycle data, providing a more reliable data set for data fusion.

[0026] Finally, a collaborative fusion method for discrete manufacturing full life cycle data is designed: Step 1: Filtered dataset No. The feature set of each link data is expressed as in Representation link In order to ensure the effectiveness of data fusion, a nonlinear function is used to Extract high-level information of each modality, i.e. in Indicates the The j-th layer output of the high-level information in the data mode i of each link; Step 2: In order to fuse the full life cycle data in each modality, the soft control function H(·,·) is used to dynamically establish the characteristic correlation between the full life cycle data, that is, in, represents the full life cycle fusion feature of the jth layer, θ β and b β denote weights and biases respectively. Step 3: According to the correlation coefficient β of mode i i , distillation purification fusion features Right now Where "⊙" represents element-wise multiplication; Step 4: Further integrate the data of different modalities throughout the life cycle and use the soft control function G(·, ) to dynamically construct the feature correlation of each modality, that is, in, represents the full life cycle fusion feature of the jth layer, θ α and b αdenote weight and bias respectively; Step 5: Utilize alpha i Refining Fusion Features Right now Combine the features of other modalities to obtain the features of modal fusion; Step 6: Iterate in sequence to obtain the fusion feature z of the whole life cycle data * , to obtain high-quality and reliable data sets.

[0027] (2) Knowledge graph construction method for the entire life cycle First, we study the event recognition method for the entire life cycle: Step 1: Based on the full life cycle data set, process knowledge, prior knowledge, etc., use intelligent computing to pre-train a large model to obtain the event representation set in Indicates the The event representation set of each link, M represents the number of links in the discrete manufacturing life cycle; Step 2: Classify events based on their similarity. In order to learn the similarity of events, given an event pair in Representation link The i-th event of The event label matrix is ​​calculated using one-hot encoding make Indicates event pair Tags, Representing an event and Unknown type, mask value The calculation is as follows: Among them, Q and K represent query and key vectors respectively, σ(·) is the sigmoid function, m represents the marginal value, “∧” and “∨” represent logical AND and logical OR operations respectively, and the loss function of contrastive learning is in is the cross entropy loss function; Step 3: Use adaptive gradient descent to optimize the loss function to train the mask contrast learning model to obtain the event classification results, thereby realizing the recognition of full life cycle events.

[0028] Secondly, build a full life cycle event relationship extraction and fusion model: Step 1: Based on the recognition results of events, use prompt learning to guide the full life cycle intelligent computing model to generate the relationship between events. Specifically, design an event relationship extraction prompter to and The relationship between events is designed as prompt questions and input into the intelligent computing model. The intelligent computing model combines prior knowledge to determine whether there is a causal relationship, time relationship, constraint relationship, task-knowledge matching relationship, knowledge-algorithm matching relationship, algorithm-resource matching relationship, etc. between events; Step 2: For the relationship of a certain event, use the evaluation function to evaluate it, select the relationship with the evaluation value exceeding the threshold as the relationship of the event pair, generate the event pair relationship, and complete the extraction of the relationship between events; Step 3: To eliminate the semantic diversity problem, we use graph neural networks to fuse and align event information. In the graph neural network, the hidden representation of each node in the l+1 layer is calculated as: in, represents the adjacency matrix of a graph containing cycles, express The logarithmic matrix, φ(·) is the ReLU function, W l Represents the learnable parameters of layer l; Step 4: Obtain a relationship graph between events by training the graph neural network

[0029] Finally, we study the completion method of the knowledge graph of the entire life cycle: Step 1: According to the event relationship diagram Use self-supervised structural embedding pre-training to extract the structural information of events and inter-event relations, and learn their structural embedding vectors for event entities e and inter-event relations r. and where d e and d r They represent the embedding dimensions respectively. The loss function of self-supervised pre-training is defined as: Among them, γ represents the marginal value, F(h, r, t) is the scoring function of the rationality of the triple (h, r, t), σ(·) represents the sigmoid activation function, (h′ i , r′ i , t′ i ) represents the K negative samples of the triple (h, r, t), i = 1, ..., K, w i is the weight of self-supervised learning; Step 2: By minimizing the loss function, the optimal structural embedding triple (h, r, t) of event entities and inter-event relations is obtained; Step 3: In order to make the intelligent computing model better understand the structure of the embedded triples, use the knowledge prefix adapter Convert these embedding triples into virtual knowledge tokens, i.e. in, Indicates text token concatenation operation; Step 4: Placed in the original input sequence As instructions and triplet hints, we get the prefix sequence, i.e. in Indicates instruction prompts. Tip for triples; Step 5: Input this prefix sequence into the intelligent computing model, and use external knowledge sources to generate new triples, complete the event knowledge graph, and update the full life cycle knowledge graph in real time.

[0030] S2, Computation Adaptation Layer (1) Construction of a large model for discrete manufacturing intelligent computing Based on the existing large models of multimodality, prediction, scientific computing, etc., explore the alignment method based on token dimension and distribution dimension, and build a large model of intelligent computing for the entire life cycle; consider the cost of large model pre-training, design a large model pre-training method based on transfer learning; for the problem of precise adaptation of downstream tasks, design a model fine-tuning method based on pattern approximation, form a large model of intelligent computing for the entire life cycle of discrete manufacturing, and provide a basic computing engine for autonomous adaptation of tasks-knowledge-algorithms-resources. The architecture of the intelligent computing large model is as follows: Figure 3 shown.

[0031] First, build a large-scale model for full-life cycle intelligent computing based on knowledge fusion: Step 1: Given a large set of K models with different architectures and functions And each large model has been pre-trained on different data sets; Step 2: Calculate the probability distribution matrix using K large models where θ j represents the parameters of the j-th large model, and t represents the parameters from the data set Sampling obtains a data sequence of length N; due to the differences in function and structure between large models, the probability distribution matrix To solve this problem, we align different probability distribution matrices, including the token dimension and the distribution dimension. Step 3: In the token dimension, s 1：L and t 1：N t 1：N Represents two token sequences, then align the token sequence s 1：i and t1：j The total cost is f(i,j)=min{f(i-1,j)+c(s i ,t j ),f(i,j-1)+c(s i , t j ),f(i-1,j-1)+c(s i ,t j )} Among them, c(s i , t j ) indicates token s i and t j The predefined string edit distance is used to minimize the total cost f(i,j) using dynamic programming to achieve alignment in the token dimension; Step 4: In terms of distribution dimension, consider two different tasks S according to the tokenizers of the two large models i and T j The probability distribution of is aligned, and the total cost is F(i,j)=min{F(i-1,j)+C(S i ,T j ),F(i,j-1)+C(S i , T j ),F(i-1,j-1)+C(s i ,t j )} Among them, C(S i ,T j ) is task S i and T j The predefined editing matrix of is also used to minimize the total cost F(i, j) using the dynamic programming algorithm to achieve alignment in the distribution dimension; Step 5: Through the above alignment, we get the alignment probability distribution matrix Alignment-based probability distribution matrix Using the fusion strategy based on the minimum cross entropy score, the fused representation matrix is ​​obtained: Among them, Fusion(·) represents the fusion strategy; Step 6: In order to transfer the computing power of the large model to the target large model, it is necessary to align the prediction and fusion representation matrix P of the target large model t , for data t, Q t Represents the output distribution matrix of the target large model, and the loss function of the fusion target is defined as: in, Represents the matrix Pt and Q t At the same time, θ represents the parameters of the target large model, and the causal language inference method is used to reduce With Q t The difference between the two is directly proportional to the loss function. Therefore, the loss function of the overall objective is defined as Among them, λ represents the weight parameter; Step 7: By minimizing the overall loss function Update the parameter θ to obtain the target large model Provide a basic large model for autonomous computing adaptation throughout the entire life cycle of discrete manufacturing.

[0032] Secondly, build a large pre-trained model for intelligent computing throughout the entire life cycle based on transfer learning: Step 1: Divide the fused full lifecycle data into multiple small blocks and input them into the teacher model to obtain the target representation of complete semantic information; Step 2: Randomly mask the θ part of each data and block the remaining 1-r% visible data Input to the student model In the process, the aggregated information between small blocks is exchanged through the self-attention mechanism of Transformer; Step 3: Since the student model only uses part of the information, the student model and the teacher model are inconsistent at each layer. The dynamic alignment method is used to align the student model and the teacher model. i Afterwards, an adapter A is introduced i , project the feature space of the student model into the feature space of the teacher model; Step 4: In order to dynamically aggregate student model information and align teacher model features, a dynamic alignment matrix W = [w ij ] S×T , where S and T represent the number of blocks of the student model and the teacher model respectively. The dynamic alignment module is formalized as in, represents the set of linear combinations of the student model outputs; Step 5: In order to suppress the feature amplitude of the teacher model, normalize each feature of the teacher model, that is, Therefore, the loss function between student and teacher features is defined as Step 6: By optimizing the loss function Train to get a pre-trained large model Achieve compression and acceleration of large models of intelligent computing of data throughout the entire life cycle, and improve the reasoning speed of intelligent computing models.

[0033] Finally, we study the fine-tuning method of large-scale intelligent computing models throughout the entire life cycle based on pattern approximation: Step 1: Build a Big Model for Smart Computing Better adapt to downstream tasks, leveraging labeled data from the entire discrete manufacturing lifecycle and using pattern approximation to update attention-based frozen weight tensors Including large models The self-attention module and cross-attention module in ; Step 2: In these modules, initialize three mode factors U = [u1, ..., u R ]、V=[v1,...,v R ] and Z=[z1,...,z R ] approximates attention-based weights, where U and Z are initialized according to Gaussian distribution, and V is initialized to 0. U and V are used as global factors in pattern approximation to achieve cross-modal interaction and knowledge sharing between weight matrices. In order to further capture the discriminative features of each modality, the learnable coefficient vector is randomly initialized for each weight matrix on the modality m. Based on the above three pattern factors, pattern approximation in forward propagation is achieved through reversible decomposition; Step 3: In this way, approximate the pre-trained weight parameters and trainable parameters And decompose and fine-tune large models on downstream data Parameters, to obtain a large intelligent computing model suitable for downstream tasks It improves the autonomous computing and adaptability capabilities of large-scale intelligent computing models throughout their entire life cycle.

[0034] (2) Autonomous computing adaptation method based on knowledge graph and large model In view of the characteristics of the entire life cycle of discrete manufacturing, such as complex design structure, sophisticated production process, diverse scheduling tasks, numerous assembly procedures, and changeable operation and maintenance failures, based on the fusion of data and knowledge graphs throughout the entire life cycle, a task-knowledge-algorithm-resource adaptation method based on the intelligent computing big model is designed, including a task-knowledge autonomous computing adaptation method based on knowledge graph search, a knowledge-algorithm autonomous computing adaptation method based on graph reasoning, and an algorithm-resource autonomous computing adaptation method based on the joint reasoning of big models and knowledge graphs, to complete the intelligent computing tasks of the entire life cycle.

[0035] First, we design a task-knowledge autonomous computing adaptation method based on knowledge graph search: Step 1: Discrete manufacturing full life cycle knowledge graph is in Represents a collection of entity event nodes, represents the set of edges connecting event nodes, Represents a set of relationship types. Given a task knowledge matching pair τ represents a given task, represents the set of optional knowledge, N represents the number of optional knowledge, and all task entities are represented as The knowledge entity is According to matching pairs and knowledge graph Retrieve related subgraphs It contains nodes With node The first H hop nodes on the path between them, therefore, Contains all task entities Knowledge Entity The corresponding relationship between intermediate entities and entities; Step 2: Right Encoding converted to text prompt T K and input it into the intelligent computing model In the process, the generative knowledge and generalization ability of the large model are used to complete the autonomous adaptation between "task-knowledge"; Step 3: To initialize the large model The task-knowledge adaptation process randomly selects a task entity Driving large models Based on its own judgment, it chooses a path and eventually finds the appropriate knowledge entity. Specifically, the computational adaptation process of the large model is divided into multiple rounds. In each round, the current head entity v in the text prompt is h , the head entity connected to it And the corresponding relationship Inform the big model According to the large model Based on the prior knowledge of the entity and the external knowledge of the text prompt, the most appropriate tail entity is selected as the head entity of the next round; Step 4: Repeat this entity selection process until the appropriate knowledge entity v is matched. k ; Step 5: Big Model Generate knowledge adapted to specific tasks, form the optimal task-knowledge adaptation relationship, and update the intelligent computing knowledge graph model based on event entities and the relationship between event entities.

[0036] Then, we design a graph-based knowledge-algorithm autonomous computation adaptation method. Given an adapted task-knowledge pair, we use meta-hints to convert specific task knowledge into a description of the optimization problem, including task objectives, resource requirements, constraints, etc. In each round of adaptation, based on the optimization problem description of the meta-hint and the evaluated algorithm, the large model Generate a candidate algorithm; evaluate the algorithm and add it to the meta hint for the next round of algorithm adaptation process. Iterate the adaptation generation algorithm

[0037] Step 1: Given a description of an optimization problem, a large model Determine the knowledge graph The algorithm adapts the initial entity of the path and automatically extracts the entity to obtain the top-K topic entities related to this problem. Step 2: In round M When each path P n Contains M-1 triplets, namely in represents the optimization problem entity, Represents the algorithm entity, are their specific relations, so the tail entity set and relationship set on all paths P are expressed as and Step 3: In the relationship exploration phase, the big model Search to connect each path P n Upper tail entity Candidate relationship And aggregate these relationships into a relationship set Once the candidate relationship set is obtained and the expanded candidate adaptation path P cand , using large models Select new top-K adaptation paths, evaluate the obtained relationships, and prune candidate relationships between entities based on the evaluation scores; Step 4: Similar to relationship exploration, through intelligent calculation of large models Search and prune algorithm entities to generate algorithms that are adapted to task knowledge Algorithm The optimization configuration process is as follows: For the algorithm A set of feasible configurations, each of which Encoded as a vector, F Π represents the set of feature vectors used to describe the example, represents the performance function that maps an example feature vector and configuration pair (f, c) to an execution algorithm Output; Step 5: To get the algorithm Performance function Parameterize the performance function with parameter θ Get its prediction model Constructing a training set in Label the training set; use intelligent computing large models Finding θ * Optimal approximation At the same time, given the example feature f and the optimal parameter vector θ * ,algorithm The parameter configuration problem of is modeled as the following optimization problem: Use mathematical programming methods to solve the algorithm The optimal parameter configuration c * ; Step 6: Combining the above steps, calculate the optimal algorithm based on the task-knowledge pair Complete autonomous computing adaptation of knowledge and algorithms.

[0038] Finally, we designed a resource-autonomous computing adaptation method based on a joint reasoning algorithm based on a large model and knowledge graph: Step 1: Given knowledge - algorithm adaptation algorithm and resource selection scheme set Using large models The algorithm and matching resource selection Encoded as a vector sequence in, is the final hidden layer vector of each token in the query; Step 2: Input the token representation into the nonlinear layer to achieve alignment between the text representation space and the entity representation space, that is, where f s : is a linear transformation function, and σ(·) is an activation function. Therefore, we get the token representation Step 3: Based on the knowledge graph The initial representation of each entity is The entity representation is updated through the graph neural network. The specific update calculation of each layer is: Step 4: In order to take advantage of the knowledge graph and the big model at the same time, design a joint adaptation module to fuse the big model The encoding output and the output of the graph neural network. Specifically, according to the input and Constructing the Matrix in, represents the learnable weight matrix, Indicates element-wise multiplication, and [·;·] indicates that vectors are concatenated row by row. The rows and columns of and and calculate and The accompanying representation: in represents the matrix multiplication operator; the adjoint representation and other original features are fused to obtain Among them, W Q 、W X is a learnable weight matrix; Step 5: Tokenize the updated token Input to the first layer joint adaptation module to continue joint adaptation, and at the same time, the updated knowledge graph is dynamically pruned to obtain Step 6: After N layers of iterative updates, the algorithm-adapted representation Q of the fused knowledge information is obtained N and graph representation X that integrates adaptation information N .algorithm The degree of fit between resources is calculated as follows: Where s is Q N Average pooling, g represents X N Based on attention pooling. According to the degree of adaptation, the algorithm is obtained The optimal resource adaptation solution cs * , realizing autonomous computing adaptation of “algorithm-resources”.

[0039] (3) Dynamic resource scheduling strategy based on deep reinforcement learning Based on the full life cycle fusion data and autonomous computing adaptation results, the resource consumption relationship for specific tasks is clarified, and a real-time resource demand prediction model based on deep reinforcement learning matrix decomposition is constructed; according to the existing resource status, the supply and demand equilibrium relationship based on the double-layer Stackelberg multi-objective game is analyzed, and a resource scheduling strategy based on deep determination of multi-objective policy gradients is designed; when resources are insufficient, a resource supplement and reconstruction method based on causal discovery and graph generation models is designed to ensure the smooth implementation of full life cycle intelligent computing.

[0040] First, we build a real-time resource demand prediction model based on deep reinforcement learning matrix decomposition: Step 1: According to the adaptation results of the computing adaptation layer, each specific task needs to consume different resources, including computing, materials, storage and other resources. Each type of task τ i In time slot t j The required resources are recorded as a vector Where L represents the number of resource types, l∈{1,...,L}; Step 2: m types of production tasks τ1, ..., τ m In n time slots t1, ..., t n The resource requirements are expressed in the following matrix: Among them, the i-th row represents the task τ i In time slots t1..., t n The resource requirements of column j are in time slot t j Resource requirements for various tasks; Step 3: At time slot t n The resource demand is to predict the value of the nth column element of matrix D based on the observation values ​​of the first n-1 columns, so the problem is transformed into a matrix completion problem. To solve this problem, the deep reinforcement learning algorithm is used to decompose the matrix D into two matrices A and The product of In order to achieve accurate prediction, the loss function is introduced Where W represents the weight matrix; Step 4: To obtain the optimal decomposition matrix A * and B * , using the adaptive momentum algorithm to minimize the loss function According to the matrix A * and B * Calculate the matrix The value of Then we can get the value of the nth column element of matrix D and accurately predict the time slot t n resource requirements.

[0041] Secondly, we study the resource scheduling strategy based on the two-layer Stackelberg game: Step 1: Based on the resource demand forecast results and existing resource information, a two-layer Stackelberg game model consisting of the general controller, task controller, and resource controller is established. Γ={Υ task ,,Υ resource ;π={y task ,,y resource}; U1,...,U m ;M resource} Among them, task and Y resource Represents task controller and resource controller respectively, y task represents the resource demand planning strategy formulated based on resource demand forecast, y resource represents the resource allocation strategy, U i is the i-th utility objective function, i∈{1,...,m}, m represents the number of utility objectives, M resource Represents existing resource information; Step 2: Convert the resource coordination problem into a two-level Stackelberg game problem. To solve this game problem, a multi-objective Markov decision process is constructed. Represents the state space (including device status, resource status, task completion rate, etc.), represents the action space taken by the task and resource controller, represents the state transition probability, Indicates real-time feedback. represents the constraint regularization term, λ i is the penalty coefficient; Step 3: Find the optimal resource scheduling strategy by maximizing the long-term cumulative reward function, taking the task’s resource requirements and existing resource status information as input parameters of the deep neural network, designing a deterministic multi-objective policy gradient algorithm to update the resource scheduling strategy, and converge to the Pareto solution, that is, to obtain the optimal resource scheduling strategy π * , achieving supply and demand balance in resource scheduling.

[0042] Finally, we build a resource supplement and reconstruction method based on causal discovery and graph generation models: Step 1: To address the problem of insufficient resource supply, based on the autonomous computing adaptation results and existing resource conditions, define the normal and abnormal resource adaptation sets as follows: and in represents the set of adaptation conditions of resource n at time t, where t is the time of the first resource adaptation anomaly, N is the resource type, and T is the time to repair the resource adaptation anomaly; Step 2: Let F = 0 represent the normal adaptation dataset intervention node, F = 1 represent the abnormal adaptation dataset intervention node, and construct the resource adaptation effect node and satisfy Among them, P * represents the probability distribution of normal and abnormal adaptation data combined, P N represents the normal adaptation data distribution, P A Indicates abnormal adaptation to data distribution; Step 3: Based on causal discovery and hierarchical learning theory, the problem of tracing the source of resource adaptation anomalies is transformed into the problem of identifying the optimal causal graph G, that is, Step 4: Adapting effect nodes based on resources Neighborhood Construction Root Cause Node List Use resource scheduling models to calculate resource scheduling solutions and trace resource adaptation to abnormal nodes Analyze resource replenishment plans Among them, b′ represents the abnormal resource supplement part, Indicates the total number of resources after resource replenishment; Step 5: Based on the data collaborative fusion model and the knowledge graph model, a resource reconstruction method based on the graph generation model is constructed to obtain a real-time supplement and reconstruction solution for the entire life cycle of discrete manufacturing resources, effectively solving the problem of insufficient resources.

[0043] S3, Parallel Service Layer In view of the significant impact of digital twins on the intelligent process reengineering of discrete manufacturing, the dynamic evolution law of the entire life cycle of discrete manufacturing is analyzed, and a smart service reconstruction scenario based on digital twins is constructed; combined with the results of intelligent computing adaptation, a service execution strategy for physical-twin parallel world mapping is designed; an execution strategy evaluation mechanism based on perturbation method is proposed, and a real-time dynamic optimization and adjustment method for execution strategy based on causal response is designed, such as Figure 4 shown.

[0044] The above shows and describes the main features, methods of use, basic principles, and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and improvements may be made to the present invention based on actual circumstances without departing from the spirit and scope of the present invention. Such changes and improvements are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A discrete manufacturing full life cycle intelligent computing system, characterized by: It includes a data knowledge layer, a computing adaptation layer, a parallel service layer, and maps each link of the entire life cycle into the physical world and the digital twin world; the data knowledge layer is used to collaboratively integrate multi-source and multi-modal data of the entire life cycle of discrete manufacturing, and build a dynamic and scalable full-life cycle knowledge graph; the computing adaptation layer is based on the knowledge graph and large model drive to achieve autonomous computing adaptation of tasks-knowledge-algorithms-resources, including task requirement perception, dynamic algorithm generation and resource matching; the parallel service layer executes intelligent service tasks of virtual-reality mapping through digital twin technology, including dynamic design deduction, virtual production reconstruction and virtual assembly trial operation.

2. The discrete manufacturing full life cycle intelligent computing system according to claim 1, characterized in that: The computation adaptation layer implements cross-layer interaction through the following mapping functions: The intelligent mapping function from the data knowledge layer to the computing adaptation layer is: Among them, s i represents the information representation vector of the i-th task, r represents the production resource information vector, D t represents the data fusion information of round t, K t represents the knowledge graph constructed in round t, π i represents the service strategy of task i, Φ(·) is the autonomous computing adaptation mapping function, which enables the completion of intelligent computing tasks at the minimum cost; The intelligent mapping function from the computing adaptation layer to the parallel service layer is: Among them, r p Represents the existing actual resources, R t represents the discrete manufacturing parallel service strategy in round t, ψ(·) is the parallel intelligent service mapping function; The intelligent mapping function from the computation adaptation layer to the data knowledge layer is defined as: Among them, D new Represents the newly generated data in each link of the whole life cycle, K new represents the new knowledge generated in the parallel intelligent service process, and Θ(·) represents the data-knowledge intelligent mapping function.

3. The discrete manufacturing full life cycle intelligent computing system according to claim 1, characterized in that: The data knowledge layer includes a multimodal data collaborative fusion module and a knowledge graph construction module. The multimodal data collaborative fusion module cleans data based on federated learning and Shapley value evaluation, and generates a high-precision data set through an adaptive filtering algorithm. The knowledge graph construction module uses a mask contrast learning model to identify full life cycle events and relationships, and completes the knowledge graph through a graph neural network.

4. The discrete manufacturing full life cycle intelligent computing system according to claim 1, characterized in that: The computational adaptation layer includes a large model fusion module and an autonomous adaptation module. The large model fusion module aligns the probability distribution of multiple models through dynamic programming and fuses representations based on minimum cross entropy. The autonomous adaptation module realizes task and knowledge matching based on knowledge graph search, knowledge and algorithm adaptation based on graph reasoning, and algorithm and resource adaptation based on joint reasoning.

5. The discrete manufacturing full life cycle intelligent computing system according to claim 1, characterized in that: The parallel service layer includes a twin scenario reconstruction module and a virtual-reality execution module. The twin scenario reconstruction module uses three-dimensional reconstruction and motion model estimation to construct a digital mapping between the physical world and the twin world. The virtual-reality execution module simulates and deduces in the twin world according to the service strategy, and dynamically adjusts the physical world execution plan through perturbation method and causal response optimization strategy.

6. The discrete manufacturing full life cycle intelligent computing system according to claim 1, characterized in that: It also includes a resource dynamic scheduling module, which is based on a resource demand prediction model based on deep reinforcement learning. It completes the resource demand matrix in real time through matrix decomposition. It also uses a resource scheduling strategy based on the Stackelberg multi-objective game, combined with causal discovery and graph generation models to supplement abnormal resources.

7. The discrete manufacturing full life cycle intelligent computing system according to claim 6, characterized in that: The Stackelberg game model of the resource scheduling strategy is defined as: C={Y task ,Y resource ;π={y task ,y resource };U1,...,U m ;M resource } Among them, task and Y resource Represents task controller and resource controller respectively, y task represents the resource demand planning strategy formulated based on resource demand forecast, y tesource represents the resource allocation strategy, U i is the i-th utility objective function, i∈{1,...,m}, m represents the number of utility objectives, M resource Represents existing resource information.

8. The discrete manufacturing full life cycle intelligent computing system according to claim 1, characterized in that: The execution process of the virtual-to-real mapping strategy includes: (1) Fusion of point cloud features from the physical world and the twin world through an attention mechanism; (2) Optimize the 3D reconstruction model based on geometric loss and motion joint loss to generate dynamic service scenarios.

9. A calculation method for a discrete manufacturing full lifecycle intelligent computing system according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1. Data knowledge layer: Collects multi-source heterogeneous data from the physical world, transforms raw data into structured knowledge through deep multimodal data collaborative fusion technology, and constructs a knowledge graph covering the entire life cycle; S2, Computing Adaptation Layer: Based on the pre-trained large model, it analyzes task requirements, combines structured knowledge in the knowledge graph, dynamically matches algorithm libraries and resource pools, and designs a task-knowledge-algorithm-resource adaptation mechanism. Through dynamic resource scheduling strategies, it achieves flexible allocation of computing resources and optimal algorithm selection. S3, Parallel Service Layer: Based on the computational adaptation results, a twin world of the physical world is constructed to achieve virtual mapping and real-time synchronization of equipment, processes, and procedures. Through simulation deduction and causal response mechanisms, service strategies are dynamically optimized and optimization instructions are fed back to the physical world.

10. The calculation method of the discrete manufacturing full life cycle intelligent computing system according to claim 9, characterized in that: Feedback the execution results of the physical world to the twin world in real time, triggering knowledge graph updates and computing adaptation strategy adjustments.

Images