Complex part process route generation and distribution method based on dynamic knowledge graph

Abstract

In order to solve the problem that when an existing disclosed knowledge modeling related method is used for carrying out knowledge modeling and integration on a complex part and generating a process route, the precision manufacturing requirement of the complex part is difficult to meet, the invention provides a complex part process route generation and distribution method based on a dynamic knowledge graph, which comprises the following steps of: extracting static knowledge from a process image and a text; according to the method, a global knowledge set is constructed, dynamic attribute data associated with a factory, a production line and key equipment are integrated into a global attribute set of the global knowledge set to update the global attribute set, and a dynamic knowledge graph is constructed based on the updated global knowledge set, so that a solid data basis is provided for subsequent process generation and optimization; the GRNN model is introduced in the process route generation process, the historical process scheme and the test set are used for cross validation, the trained GRNN model not only can effectively identify the reasonable range of the process parameters, but also can make full use of graph structure information in the knowledge graph, and the accuracy of process route generation is improved.

Description

Technical Field

[0001] The invention relates to the technical field of mechanical processing, and in particular to a method for generating and allocating a process route for a complex component. Background Art

[0002] Complex components refer to industrial products with complex structures, diverse functions, high manufacturing difficulty and extremely high precision requirements. These components include multiple sub-assemblies and systems, which require high coordination and precise cooperation with each other. As a core component of the modern industrial field, its design, manufacturing, assembly and troubleshooting are all topics that engineers need to study in depth. With the rapid development of technology, the complexity and precision requirements of these components are constantly rising, and each part carries extremely high processing requirements. The research and development of new materials, precision manufacturing processes and the application of advanced digital technologies have made the design and manufacturing process of complex components more complicated.

[0003] At the same time, the manufacturing of complex components involves knowledge from multiple professional fields, such as material science, mechanical engineering, and automated control, requiring engineers to have rich professional knowledge and practical experience. However, this professional knowledge is usually scattered in various documents, and there are many challenges in knowledge inheritance and talent training, which undoubtedly brings difficulties to the research and development and customization of complex components.

[0004] Therefore, in order to better cope with these challenges, knowledge modeling and integration of complex components is particularly important. Through in-depth research and understanding of complex components, combined with advanced knowledge management technology, we can more effectively integrate and utilize various professional knowledge, improve the efficiency of design, manufacturing and maintenance of complex components, and promote the sustainable development of the industrial field.

[0005] In recent years, the research on knowledge modeling and program planning mainly includes:

[0006] The invention patent with publication number CN117010497A proposes a method for modeling airworthiness knowledge of general aviation aircraft landing gear shock absorbers. By constructing a shock absorber product structure tree node model, an airworthiness standard system triangle model, and a shock absorber airworthiness design process model, the shock absorber airworthiness knowledge modeling is achieved. This patent improves the efficiency and accuracy of determining the basis for airworthiness certification, but lacks the decision-making ability to deal with complex and changeable application scenarios, and cannot be applied to the generation of highly professional process routes for complex components.

[0007] The invention patent with publication number CN111274409A proposes a method and system for controlling the assembly process of an engine valve train based on a knowledge graph. It uses a parameter history database and fuzzy theory to collect and calculate the threshold of test parameters to ensure the quality of the engine valve train assembly. This solution improves assembly quality and reduces false positives through fuzzy theory, but there is a problem of inaccurate parameter design, which makes it difficult to meet the precision manufacturing requirements of complex components.

[0008] The existing technology has deficiencies in automation, decision-making, adaptability, efficiency and accuracy when dealing with complex component knowledge modeling and process planning. Although the above patents provide some solutions in shock absorber airworthiness and engine valve mechanism assembly, they are unable to effectively handle the generation of process routes for complex components, which require high professionalism. Summary of the invention

[0009] In order to solve the technical problem that the existing disclosed knowledge modeling related methods are difficult to meet the precision manufacturing requirements of complex components when performing knowledge modeling and integration on complex components and generating process routes, the present invention proposes a complex component process route generation method based on dynamic knowledge graph. On this basis, the present invention further proposes a complex component process route allocation method.

[0010] The inventive concept of the present invention is:

[0011] First, the knowledge scattered in various documents, real-time production sensor data and enterprise management information are organized in a unified manner to build a dynamic knowledge graph, realize the effective acquisition and integration of knowledge, and provide support for subsequent process generation.

[0012] Next, a demand knowledge graph is constructed based on the manufacturing requirements of complex components.

[0013] Then, a graph recurrent network process model is constructed and used to generate a process route. Graph recurrent networks have advantages in processing complex relationships and data structures, ensuring that the generated process route not only meets the design requirements of complex components, but also has the characteristics of efficiency and reliability.

[0014] On the basis of generating the process route, the ant colony algorithm is used to dynamically allocate the processing tasks. With its powerful optimization ability and adaptability, the ant colony algorithm can find the optimal processing task allocation plan in a short time, ensuring that the manufacturing process of complex parts can be carried out efficiently and orderly.

[0015] The technical solution adopted by the present invention is:

[0016] The method for generating a complex component process route based on a dynamic knowledge graph is special in that it includes the following steps:

[0017] Step 1: Construct a dynamic knowledge graph G based on the static knowledge extracted from the historical process image set and historical process text set of complex components, as well as the updated dynamic attribute set associated with the factory entity, production line entity, key equipment entity, and equipment component entity. dynamic ;

[0018] Step 2: Based on the demand text of the complex component, construct the demand knowledge graph G Ne ;

[0019] Step 3: Construct the GRNN network model and use X = [G dynamic ,G Ne ] to train it;

[0020] Step 4: Process route prediction;

[0021] Based on the process image set and process text set of the complex components corresponding to the current task, as well as the updated dynamic attribute set associated with the current factory entity, production line entity, key equipment entity and equipment component entity, the task dynamic knowledge graph G of the current task is constructed. dynamic_task ; Based on the requirement documents of the complex components corresponding to the current task, build the task requirement graph G of the current task Ne_task ;

[0022] Set X_task = [G dynamic_task ,G Ne_task ]The trained GRNN network model is input to cyclically predict the next process step, and the GRNN network model outputs the process route prediction result of the complex component corresponding to the current task.

[0023] Furthermore, the dynamic knowledge graph G in step 1 dynamic =(V,E,A′ global ,U); V is a node set, which is constructed based on the entity set in the dynamically updated global knowledge set; E is an edge set, which is constructed between nodes based on the relationship set in the dynamically updated global knowledge set; A′ global is the global attribute set in the dynamically updated global knowledge set; U is the update function used to update the attributes of the entity / node corresponding to the entity;

[0024] The method for acquiring the dynamically updated global knowledge set is:

[0025] Firstly, static knowledge is extracted from the historical process images and process texts of complex parts respectively, and then merged to form a global knowledge set, which includes a global entity set, a global relationship set and a global attribute set.

[0026] Then, a dynamic attribute set is defined, and the dynamic data associated with the factory entity, production line entity, key equipment entity and equipment component entity are integrated into the dynamic attribute set to update it. The updated dynamic attribute set is merged, deduplicated and standardized with the global attribute set in the global knowledge set to obtain a dynamically updated global knowledge set.

[0027] Furthermore, the historical process image collection described in step 1 includes part drawings, process tabs and design manuals in image form, from which static knowledge is extracted using visual processing technology; the historical process text collection includes process documents and process manuals in natural language form, from which static knowledge is extracted using natural language processing technology; the static knowledge includes entities, relationships and attributes; the process image collection of the complex component corresponding to the current task described in step 4 includes part drawings, process tabs and design manuals in image form, wherein the process tab is empty; the process text collection of the complex component corresponding to the current task includes process documents and process manuals in natural language form, wherein the process documents are empty.

[0028] Furthermore, the visual processing technology is OCR or Faster R-CNN; the natural language processing technology is Transformer.

[0029] Furthermore, the demand knowledge graph G in step 2 Ne =(V Ne ,E N ,A Ne ,U=1);V Ne is a node set, constructed according to the entity set extracted from the requirement document; N A is an edge set, constructed between nodes based on the relationship set extracted from the requirement document, representing the relationship between entities; Ne is a set of attributes extracted from the requirement document, U is an update function used to update the attributes of the entity / node corresponding to the entity, and U=1 indicates that the attributes of the entity / node corresponding to the entity are not updated.

[0030] Furthermore, the training objective of the GRNN network model in step 3 is to minimize the prediction error of the process route. An end-to-end training method is adopted, teacher forcing technology is used, and ground truth is used as input during training to improve the convergence speed.

[0031] The present invention also provides a complex component process route generation and allocation method based on a dynamic knowledge graph, which is characterized by comprising the following steps:

[0032] Step 1: Generate a process route for a complex component to be manufactured using the above-mentioned complex component process route generation method;

[0033] Step 2: Use the ant colony algorithm to optimize the process route generated in the first step and generate the best process route.

[0034] Furthermore, the second step is specifically as follows:

[0035] Step 1) Define the objective function f(y,t):

[0036]

[0037] Where t is the current time; w1, w2, w3 are weights used to balance the importance of different items; Parameter P used to calculate the current process step i With ideal parameters The function of the difference between i ,L f ,P l ,S d ,t) is used to calculate the current process step S i Operation type O i Compatibility or impact with real-time data; f is the production load, P l is the production progress, S d is the critical equipment status, L f , P l and S d From And at a certain time interval t, use the function U(e,a dynamic )renew; They are factory entity attribute set, production line attribute set and key equipment attribute set;

[0038] Step 2) Define the transition probability:

[0039]

[0040] Among them, Ω i (t) is the time at which the current process step S i The set of all process steps that can be reached, i.e. part of the solution space; τ ij (t) is the time from S i To S j The pheromone concentration of ij (t) is the time from S i To S j The heuristic information is calculated based on static data and real-time data; α and β are parameters that control the impact of pheromone and heuristic information on the transition probability;

[0041] Step 3) Update the pheromone as follows:

[0042] τ ij (t+1)=(1-ρ)·τ ij (t)+ρ·Δτ ij (t)

[0043] Where: ρ is the pheromone volatilization rate, which is used to control the decay rate of pheromone; Δτ ij (t) is the time from S i To S j The pheromone increment is defined as Δτ ij (t) = 1 / f(y*(t));

[0044] Step 4) Iterative optimization:

[0045] Repeat the steps of constructing the solution space and updating the pheromone for multiple iterations;

[0046] Step 5) Output results:

[0047] When the maximum number of iterations is reached or the set iteration termination condition is met, the iteration is stopped and the current optimal process task allocation plan and its corresponding quality indicators are output, including total processing time and total cost.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The present invention extracts static knowledge from process images and process texts through visual processing and natural language processing technologies, constructs a global knowledge set, and integrates dynamic attribute data associated with factories, production lines and key equipment into the global knowledge set. The global attribute set in the global knowledge set is updated, and a comprehensive, accurate and easy-to-query dynamic knowledge graph is constructed based on the updated global knowledge set, which provides a solid data foundation for subsequent process generation and optimization. Compared with the existing technology, the efficiency and accuracy of data integration are greatly improved.

[0050] 2. The present invention introduces a graph recurrent neural network (GRNN) in the process of process route generation, and uses historical process plans and test sets for cross-validation. The trained GRNN model can not only effectively identify the reasonable range of process parameters and ensure that these parameters have higher credibility in practical applications, but also make full use of the graph structure information in the knowledge graph, effectively extract graph features, and model the timing dependencies between process steps, thereby improving the accuracy of process route generation.

[0051] 3. The present invention can update entities, attributes and relationships in the knowledge graph in real time by integrating dynamic data, so that the generated process route can reflect the actual production status of the factory in real time. This dynamic adaptability makes the present invention more flexible and adaptable in practical applications and can meet the needs of different production scenarios.

[0052] 4. The present invention further uses an ant colony algorithm to optimize and allocate the generated process routes. By defining the objective function and comprehensively considering factors such as the estimated time, parameter differences, and real-time data compatibility, a process route that meets multiple constraints can be generated. At the same time, the ant colony algorithm is used for iterative optimization, which can gradually converge to the optimal or approximately optimal process route, further improving the quality and efficiency of the process route. Compared with the prior art, the optimization algorithm of the present invention is more intelligent, efficient, and flexible. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the overall flow chart of the present invention.

[0054] Figure 2 This is the process of constructing the dynamic knowledge graph involved in the method of the present invention.

[0055] Figure 3 It is the GRNN network architecture involved in the method of the present invention.

[0056] Figure 4 It is a flow chart of the ant colony algorithm involved in the method of the present invention. DETAILED DESCRIPTION

[0057] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0058] like Figure 1 As shown, the complex component process route generation method based on dynamic knowledge graph proposed in the present invention includes the following steps:

[0059] Step 1: Build a dynamic knowledge graph G dynamic .

[0060] Step 1.1 Extract static knowledge and build a global knowledge set.

[0061] Through visual processing such as OCR and Faster R-CNN and natural language processing technologies such as Transformer, static knowledge (entities, relationships and attributes) are extracted from the process image collection and the process text collection respectively, and merged to form a global knowledge set, which includes a global entity set, a global relationship set and a global attribute set.

[0062] Step 1.2 Dynamic data integration.

[0063] Define a dynamic attribute set, and integrate the dynamic data associated with the factory entity, production line entity, key equipment entity and equipment component entity into the dynamic attribute set to update it, merge, deduplicate and standardize the updated dynamic attribute set with the global attribute set in the global knowledge set obtained in step 1.1 to obtain a dynamically updated global knowledge set.

[0064] Step 1.3 Build a dynamic knowledge graph.

[0065] Use the dynamically updated global knowledge set (integrated with dynamic data) to build a dynamic knowledge graph G dynamic =(V,E,A ′ global ,U), where V is a node set, which is constructed based on the entity set in the dynamically updated global knowledge set, and each entity in the entity set corresponds to a node; E is an edge set, which is constructed between nodes based on the relationship set in the dynamically updated global knowledge set, representing the relationship between entities; A ′ global is the global attribute set in the global knowledge set after dynamic update; U is the update function used to update the attributes of the entity / node corresponding to the entity.

[0066] Step 2: Build the demand knowledge graph G N .

[0067] Constructing demand knowledge graph G Ne =(V Ne ,E N ,A Ne ,U=1), where V Ne is a node set, constructed according to the entity set extracted from the requirement document, and each entity in the entity set corresponds to a node; N A is an edge set, constructed between nodes based on the relationship set extracted from the requirement document, representing the relationship between entities; Ne is a set of attributes extracted from the requirement document, U is an update function used to update the attributes of the entity / node corresponding to the entity, and U=1 indicates that the attributes of the entity / node corresponding to the entity are not updated.

[0068] By building the demand knowledge graph G N , which can clearly map the entities, relationships and attributes extracted from the requirement documents to the requirement knowledge graph G N The nodes, edges and node attributes of the graph are convenient for subsequent graph neural network model processing. Demand knowledge graph G based on entities and relationships N It can be dynamically adjusted according to different process requirements to meet a variety of process generation scenarios.

[0069] Step 3: Build and train the GRNN network model.

[0070] Step 3.1: Input and output definition: transform the dynamic knowledge graph G dynamic and demand knowledge graph G N As input, the output is a complete process route.

[0071] Step 3.2: Network design: Construct a GRNN network to extract graph features and model the temporal dependencies between process steps.

[0072] Step 3.3: Model training: The graph neural network model is trained using end-to-end and teacherforcing techniques by minimizing the prediction error and optimizing the network parameters.

[0073] Step 4: Generate process route.

[0074] Based on the trained GRNN network model, the process steps are cyclically predicted according to the input graph to generate a complete process route.

[0075] In addition to the above-mentioned complex component process route generation method, the present invention also provides a complex component process route generation and allocation method based on a dynamic knowledge graph, which specifically includes the following steps:

[0076] The first step: generating a process route using the complex component process route generation method of the present invention.

[0077] Step 2: Use the ant colony algorithm to optimize the process route generated in the first step and generate the best process route.

[0078] The second step is as follows:

[0079] Step 1) Define the objective function

[0080] Based on the estimated time or cost of the process steps, the difference between the process parameters and the ideal parameters, and the compatibility of the operation type with real-time production data, a weighted objective function is constructed. This weighted objective function comprehensively considers the processing efficiency, process quality, and real-time changes in the production environment.

[0081] Step 2) Define the transition probability

[0082] In the ant colony algorithm, ants choose the next process step based on the transition probability containing pheromone concentration and heuristic information. This mechanism ensures that ants can balance historical information and real-time data of the current environment when choosing a path.

[0083] Step 3) Pheromone Update

[0084] After each iteration, the pheromone concentration on the path is updated according to the quality of the optimal process route found by the ants. A high-quality solution will increase the pheromone on its path, thereby increasing its probability of being selected in subsequent iterations.

[0085] Step 4) Iterative Optimization

[0086] By repeatedly constructing the solution space and updating the pheromone, multiple iterations are performed. This process enables the ant colony to gradually converge to the optimal or approximately optimal process route.

[0087] Step 5) Output the results

[0088] When the iteration termination condition is reached, the currently found optimal process route and its corresponding quality indicators, such as total processing time and total cost, are output as the final result of the entire optimization process.

[0089] Implementation details:

[0090] The process route generation and allocation method of the present invention is further described in detail below in conjunction with two examples of a gear transmission and an automobile engine and the accompanying drawings. This process already includes the process route generation method of the present invention, so the process route generation method of the present invention will not be described separately.

[0091] The process route generation and allocation method of the present invention specifically comprises the following steps:

[0092] Step 1: Build a dynamic knowledge graph.

[0093] Through knowledge extraction, a dynamic knowledge graph is established to provide a data basis for the subsequent construction of the GRNN network model.

[0094] Step 1.1 Extract static knowledge and build a global knowledge set.

[0095] Through visual processing and natural language processing techniques, static knowledge (entities, relationships and attributes) are extracted from historical process image collections (including part drawings, process tabs, design manuals, etc. in image form) and historical process text collections (including process documents, process manuals, etc. in natural language form), and integrated to form a global knowledge collection.

[0096] Step 1.1.1) Visual processing:

[0097] Assume that the process image set is I = {I1, I2, ..., I n}, n is the total number of images. Through visual processing technologies such as OCR and Faster R-CNN, from each image I i Extract the entity set E i , relation set R i and attribute set A iThese extraction operations can be expressed as functions:

[0098] E i =f E (I i ),R i =f R (I i ),A i =f A (I i )

[0099] Among them, f E ,f R ,f A are visual processing functions used to extract entities, relations, and attributes respectively.

[0100] Specific:

[0101] ①Take the gearbox as an example:

[0102] Craft image collection for I gearbox = {I gearbox,1 ,I gearbox,2 ,…,I gearbox,n}, n is the total number of gear box process images.

[0103] Use the visual processing function f E ,f R ,f A From Craft Image Collection I gearbox Each image I gearbox,i ∈I gearbox Extract the corresponding entities, relations and attribute sets from it.

[0104] The extracted entity set is E gearbox,i =f E (I gearbox,i )={"gear","shaft","bearing",...}; the relation set is R gearbox,i =f R (I gearbox,i ) = {"coordination relationship between gear and shaft", "installation position of bearing", ...}; attribute set is A gearbox,i =f A (I gearbox,i )={“number of teeth on the gear”=24,“diameter of the shaft”=10mm,…}.

[0105] ②Take the automobile engine as an example:

[0106] Craft image collection for I engine = {I engine,1 ,I engine,2 ,…,I engine,m}, m is the total number of automobile engine process images.

[0107] Use the visual processing function f E ,f R ,f A From Craft Image Collection I engine Each image I engine,j ∈I engine Extract the corresponding entities, relations and attribute sets from it.

[0108] The extracted entity set is E engine,j =f E (I engine,j )={“cylinder”, “piston”, “crankshaft”,…}; the relation set is R engine,j =f R (I engine,j ) = {"reciprocating motion of the piston in the cylinder", "rotational motion of the crankshaft", ...}; the attribute set is A engine,j =f A (I engine,j )={“diameter of cylinder”=80mm, “stroke of piston”=100mm,…}.

[0109] Step 1.1.2) Natural Language Processing:

[0110] Assume that the process text set is D = {D1, D2, ..., D h}, where h is the total number of process texts. Through natural language processing technology, from each text D j Extract the entity set E j , relation set R j and attribute set A j These extraction operations can be expressed as functions:

[0111] E j =g E (D j ),R j =g R (D j ),A j =g A (D j )

[0112] Among them, g E ,g R ,g A are natural language processing functions used to extract entities, relations, and attributes, respectively.

[0113] Specific:

[0114] ①Take the gearbox as an example:

[0115] Craft Text Collection for D gearbox ={D gearbox,1 ,D gearbox,2 ,…,D gearbox,p}, p is the total number of gear box process texts.

[0116] Using natural language processing function g E ,g R ,g A From the text collection D gearbox Each text in D gearbox,k ∈D gearbox Extract the corresponding entities, relations and attribute sets from it.

[0117] The extracted entity set E gearbox,k =g E (D gearbox,k )={"gear","shaft","bearing",...}; the relation set is R gearbox,k =g R (D gearbox,k ) = {"gear drives the shaft to rotate", "bearing supports the shaft", ...}; the attribute set is A gearbox,k =g A (D gearbox,k ) = {“Gear material” = “Steel”, “Shaft length” = 120cm,…}.

[0118] ②Take the automobile engine as an example:

[0119] Craft Text Collection for D engine ={D engine,1 ,D engine,2 ,…,D engine,q}, q is the total number of automobile engine process texts.

[0120] Using natural language processing function g E ,g R ,g A From the text collection D engine Each text in D engine,l ∈D engine Extract the corresponding entities, relations and attribute sets from it.

[0121] The extracted entity set E engine,l =g E (D engine,l )={“cylinder”, “piston”, “crankshaft”,…}; the relation set is R engine,l =g R (D engine,l ) = {"The piston reciprocates in the cylinder", "The crankshaft is connected to the piston by a connecting rod", ...}; the attribute set is A engine,l =g A (Dengine,l )={“number of cylinders”=4,“rotation speed of crankshaft”=3000rpm,…}.

[0122] The above steps 1.1.1) and 1.1.2) can be interchanged in order.

[0123] Step 1.1.3) Knowledge merging:

[0124] The entities, relationships and attributes extracted by visual processing and natural language processing are correspondingly merged to obtain a global knowledge set, which includes a global entity set, a global relationship set and a global attribute set.

[0125] Let the integrated global entity set be E global , the global relationship set is R global , the global attribute set is A global . These collections can be constructed in the following ways:

[0126]

[0127]

[0128]

[0129] Among them, dedup is a deduplication function used to eliminate duplicates.

[0130] Specific:

[0131] ①Take the gearbox as an example:

[0132] E global_gear =dedup(E gearbox,1 ∪E gearbox,2 ∪…∪E gearbox,p )={gear, shaft, bearing,…}

[0133] R global_gear =dedup(R gearbox,1 ∪R gearbox,2 ∪…∪R gearbox,p )={the matching relationship between gear and shaft, gear drives shaft to rotate, bearing installation position, bearing supports shaft, ...}

[0134] A global_gear =dedup(A gearbox,1 ∪A gearbox,2 ∪…∪A gearbox,p ) = {number of teeth of gear = 24, gear material = steel, shaft diameter = 10mm, shaft length = 120cm, ...}

[0135] ②Take the automobile engine as an example:

[0136] E global_engine =dedup(E engine,1 ∪E engine,2 ∪…∪E engine,q ) = {cylinder, piston, crankshaft, ...}

[0137] R global_engine =dedup(R engine,1 ∪R engine,2 ∪…∪R engine,q ) = {The piston reciprocates in the cylinder, the crankshaft rotates, the crankshaft is connected to the piston by a connecting rod, ...}

[0138] A global_engine =dedup(A engine,1 ∪A engine,2 ∪…∪A engine,q ) = {number of cylinders = 4, diameter of cylinder = 80 mm, stroke of piston = 100 mm, rotation speed of crankshaft = 3000 rpm, ...}

[0139] Step 1.2 Dynamic data integration.

[0140] On the basis of static knowledge extraction, dynamic data is further integrated to improve the attribute sets of factory entities, production line entities and key equipment entities.

[0141] Step 1.2.1) Define a dynamic attribute set.

[0142] Factory Entity Factory ) refers to a series of production-related characteristics, such as production load, available capacity, number of tasks to be processed, etc. These attributes together constitute the factory entity attribute set The specific performance is

[0143]

[0144] Production line entity Line ) refers to the characteristics that describe the operating status of the production line, such as production progress, line speed, etc. These attributes together form the production line attribute set The specific performance is

[0145]

[0146] Key equipment entity (e Device ) covers the operating status of the equipment, such as operating time, failure rate, maintenance records, etc. These attributes constitute the key equipment attribute set The specific performance is In addition, key equipment also contains the properties of its components, such as temperature, pressure, speed, and vibration, which constitute the equipment component property set. Specifically expressed as

[0147]

[0148] Step 1.2.2) Integrate the dynamic data into the dynamic attribute set and update it.

[0149] Get dynamic data from the factory's information system and industrial field sensors, and update the corresponding entity attributes in the dynamic attributes.

[0150] Step 1.2.3) Global attribute set update.

[0151] Combine the dynamic attribute set after dynamic data integration in step 1.2.2) with the global attribute set A obtained in step 1.1.3) global Merge and update the global attribute set A global , the updated global attribute set is A′ global :

[0152]

[0153] The merge operation also includes deduplication and standardization steps to ensure the uniqueness and consistency of attributes.

[0154] Specific:

[0155] ①Take the gearbox as an example:

[0156] Updated global properties set for gearboxes is the dynamic property set of the gearbox,

[0157] ②Take the automobile engine as an example:

[0158] Updated global attribute set for car engine is the dynamic attribute set of the car engine,

[0159] Step 1.3 constructs a dynamic knowledge graph using the dynamically updated global knowledge set.

[0160] Dynamic Knowledge Graph G dynamic =(V,E,A′ global, The construction of U) depends on the global entity set E obtained in step 1.1 global and the global relation set R global And the updated global attribute set A′ obtained in step 1.2global .

[0161] Step 1.3.1) Node construction.

[0162] For the global entity set E global For each entity e in the dynamic knowledge graph G, add it as a node v dynamic The node set V is:

[0163]

[0164] Step 1.3.2) side build.

[0165] For the global relation set R global Each relation r=(e1,rel,e2)(where e1,e2∈E global rel is the relationship type), in the dynamic knowledge graph G dynamic Add an edge e′ from node v1 of e1 to node v2 of e2 to the edge set E of e1, and mark it as relation rel. That is, the edge set is:

[0166]

[0167] Step 1.3.3) Attribute addition and update.

[0168] For the updated global attribute set A′ global Each attribute a=(e,attr,value)(where e∈E global , attr is the attribute name, value is the attribute value), add the attribute attr and its value value to the node v corresponding to the entity e.

[0169] For dynamic attributes, define the update function U(e,a dynamic )=update(e,a dynamic ) to update the properties of the node. This function will update the properties of the node based on the latest dynamic data a obtained from the data source dynamic To update the attributes on entity e (or its corresponding node v).

[0170] Specific:

[0171] ①Take the gearbox as an example:

[0172] According to the global entity set E of the gearbox global_gear To construct the node set V gear :

[0173]

[0174] Among them, gear iis the global entity set E global_gear An element in It is with gear i The corresponding node.

[0175] According to the global relationship set R of the gearbox global_gear To construct the edge set E gear :

[0176]

[0177] Among them, r is the global relationship set R global_gear rel is the name of the relationship, and v1 and v2 are the nodes corresponding to e1 and e2 respectively.

[0178] Dynamic properties of gearboxes The update process is expressed as:

[0179]

[0180] Finally, the dynamic knowledge graph G of the gearbox dynamic_gear ,G dynamic_engine for:

[0181] G dynamic_gear =(V gear ,E gear ,A′ global_gear ,U).

[0182] ②Take the automobile engine as an example:

[0183] According to the global entity set E of the car engine global_engine To construct the node set V engine :

[0184]

[0185] Among them, engine i is the global entity set E global_engine An element in Is with engine i The corresponding node.

[0186] According to the global relationship set R of the automobile engine global_engine To construct the edge set E engine :

[0187]

[0188] Among them, r is the global relationship set R global_engine rel is the name of the relationship, and v1 and v2 are the nodes corresponding to e1 and e2 respectively.

[0189] Dynamic properties of car engines The update process is:

[0190]

[0191] Finally, the dynamic knowledge graph G of the automobile engine dynamic_engine It is expressed as:

[0192] G dynamic_engine =(V engine ,E engine ,A′ global_engine ,U).

[0193] Step 2: Build the demand knowledge graph GN e =(V Ne ,E N ,A Ne ,U=1).

[0194] Step 2.1 Knowledge extraction from requirement documents.

[0195] Define the historical requirements document as D Ne . Requirements Document D Ne The document should include the functional and technical requirements of all complex components involved in building a dynamic knowledge graph, and further include the purpose, main components, and operating environment of these complex components. Ne Extract the entity set E Ne , relation set R Ne and attribute set A Ne These extraction operations can be expressed as functions:

[0196] E Ne =g E (D Ne ),R Ne =g R (D Ne ),A Ne =g A (D Ne )

[0197] Step 2.1 Node construction.

[0198] For the requirements document E Ne Each entity e in ne , taking it as a node v ne Add to demand knowledge graph G Ne The node set V Ne In. That is, the node set is:

[0199]

[0200] Step 2.2 Edge construction.

[0201] For the relation set R Ne Each relation r=(e ne1 ,rel,e ne2 )(where e ne1 ,e ne2 ∈E Ne rel is the relationship type), in the demand knowledge graph G Ne The edge set E N Add a line from e ne1 The node v ne1 To e ne2 The node v ne2 The edge ne ′ , marked as relation rel. That is, the edge set is:

[0202]

[0203] Step 2.3 Add attributes.

[0204] For the attribute set A Ne Each attribute a in ne =(e ne ,attr,value)(where e ne ∈E Ne , attr is the attribute name, value is the attribute value), add the attribute attr and its value value to the entity e ne The corresponding node v ne superior.

[0205] Specific:

[0206] ①Take the gearbox as an example:

[0207] Gearbox requirement document D Ne-gear The following requirements are described:

[0208] D Ne-gear =“The gearbox needs to meet the following conditions: Gear ratio adjustable range: 2.0-4.5;

[0209] Maximum torque bearing capacity: 800Nm; Material: high-strength alloy steel

[0210] Using natural language technology to process requirements documents Ne-gear After that, we get the following set:

[0211] Entity Set E Ne-gear={gearbox, gear ratio adjustable range, maximum torque bearing capacity, material};

[0212] Relation set R Ne-gear ={(gear transmission, required, gear ratio adjustable range),(gear transmission, required, maximum torque bearing capacity),(gear transmission, made of material, material)};

[0213] Attribute set A Ne-gear ={(gearbox, gear ratio adjustable range, 2.0-4.5),(gearbox, maximum torque bearing capacity, 800Nm),(gearbox, material, high-strength alloy steel)};

[0214] Node Set V Ne-gear = {v 齿轮变速箱 ,v 齿轮比可调范围 ,v 最大扭矩承受能力 ,v 材料};

[0215] Edge set E N-gear ={(v 齿轮变速箱 ,need,v 齿轮比可调范围 ),(v 齿轮变速箱 ,need,v 最大扭矩承受能力 ),(v 齿轮变速箱 , made of material,v 材料 )}

[0216] At node v 齿轮变速箱 Add the following attributes: gear ratio adjustable range = 2.0-4.5, maximum torque bearing capacity = 800Nm, material = high-strength alloy steel. Thus, the demand knowledge graph of gearbox is obtained as follows:

[0217] G Ne-gear =(V Ne-gear ,E N-gear ,A Ne-gear ,U gear =1).

[0218] ②Take the automobile engine as an example:

[0219] Assume that the requirement document D for the automobile engine Ne-engine The following requirements are described:

[0220] D Ne-engine = "The car engine needs to meet the following conditions: displacement: 2.0L; maximum power: 200hp; fuel type: gasoline; number of cylinders: 4; emission standard: National VI"

[0221] Using natural language technology to process requirements documents Ne-engine After that, we get the following set:

[0222] Entity Set E Ne-engine ={car engine, displacement, maximum power, fuel type, number of cylinders, emission standard};

[0223] Relation set R Ne-engine ={(car engine, has, displacement),(car engine, has, maximum power),(car engine, uses, fuel type),(car engine, has, number of cylinders),(car engine, complies with, emission standards)};

[0224] Attribute set A Ne-engine ={(car engine, displacement, 2.0L),(car engine, maximum power, 200hp),(car engine, fuel type, gasoline),(car engine, number of cylinders, 4),(car engine, emission standard, National VI)};

[0225] Node Set V Ne-engine = {v 汽车发动机 ,v 排量 ,v 最大功率 ,v 燃油类型 ,v 气缸数 ,v 排放标准}

[0226] Edge set E N-engine ={(v 汽车发动机 , with,v 排量 ),(v 汽车发动机 , with,v 最大功率 ),(v 汽车发动机 ,Use,v 燃油类型 ),(v 汽车发动机 , with,v 气缸数 ),(v 汽车发动机 , in line with,v 排放标准 )};

[0227] At node v 汽车发动机 Add the following attributes: displacement = 2.0L, maximum power = 200hp, fuel type = gasoline, number of cylinders = 4, emission standard = National VI. Thus, the demand knowledge graph of automobile engines is obtained as follows:

[0228] G Ne-engine =(V Ne-engine ,E N-engine ,A Ne-engine ,U engine =1).

[0229] Step 3: Build and train the GRNN network model.

[0230] Step 3.1 Input-output representation.

[0231] During training, the input of the GRNN network is the historical dynamic knowledge graph G dynamic and historical demand knowledge graph G Ne Since each process step S i = {O i ,P i} can be used with discrete operation types O i and the continuous parameter vector P i Therefore, the entire input can be expressed as X = [G dynamic ,G N ], the output of the GRNN network is the complete process route y=[S1,S2,…,S n ].

[0232] Step 3.2 Determine the network architecture (network structure such as Figure 3 As shown, it is the existing publicly available network structure).

[0233] The GRNN network includes:

[0234] Input layer: receives X = [G dynamic ,G Ne ];

[0235] Graph convolution layer: Using graph convolution operation H^{(l+1)}=σ(A)H^{(l)}W^{(l)}), from X=[G dynamic ,G Ne ] to extract graph features;

[0236] Graph pooling layer: GraphSAGE graph pooling technology is applied to reduce the dimension of graph features extracted by the graph convolution layer;

[0237] Recurrent layer: LSTM units are introduced to model the temporal dependencies between process steps;

[0238] Output layer: Use a fully connected layer to predict the next process step

[0239] Step 3.3 Model training.

[0240] The training goal is to minimize the prediction error of the process route An end-to-end training method is adopted, teacher forcing technology is used, and ground truth is used as input during training to improve the convergence speed.

[0241] Step 4: Generate process route.

[0242] Using the same method as steps 1-2, process the process image set and process text set of the current task, the updated dynamic attribute set associated with the current factory entity, production line entity, key equipment entity and equipment component entity, and the requirement document of the current task to construct the task dynamic knowledge graph G of the current task. dynamic_task and the task requirement graph G of the current task Ne_task Different from the historical process image collection and historical process text collection, the process tab in the process image collection of the current task is empty, and the process document in the process text collection of the current task is empty.

[0243] Set X_task = [G dynamic_task ,G Ne_task ] Input the trained GRNN network model to cyclically predict the next process step Until a complete process route is generated

[0244] Taking complex parts such as gear transmissions and automobile engines as examples, the generated process routes are as follows:

[0245] [{"id":"0001","Operation type":"Material preparation","Operation object":"Blank","Parameters":{"Material":"High-strength alloy steel"},"Estimated time":"2 hours","Estimated cost":"100 yuan"},{"id":"0002","Operation type":"Forging","Operation object":"0001","Parameters":{"Temperature":"1100℃","Pressure":"300MPa"},"Estimated time":"4 hours","Estimated cost":"300 yuan"},...]

[0246] [{"id":"0001","Operation type":"Material preparation","Operation object":"Blank","Parameters":{"Material":"Chrome-nickel alloy steel"},"Estimated time":"1.5 hours","Estimated cost":"120 yuan"},{"id":"0002","Operation type":"Forging","Operation object":"0001","Parameters":{"Temperature":"1050℃","Pressure":"320MPa"},"Estimated time":"3.5 hours","Estimated cost":"280 yuan"},…]

[0247] Step 5: Generate the best process task allocation plan.

[0248] Step 5.1 Define the objective function.

[0249] A company not only processes one type of parts, but also processes multiple parts at the same time, resulting in multiple corresponding processing routes. Suppose the process route set generated in step 4 is C i It can be a process step S i It can be the value of "estimated time" or the value of "estimated cost".

[0250] Objective Function

[0251] Where: t is the current time; w1, w2, w3 are weights used to balance the importance of different items; Parameter P used to calculate the current process step i With ideal parameters The function of the difference between i ,L f ,P l ,S d ,t) is used to calculate the current process step S i Operation type O i Compatibility or impact with real-time data (such as production load, production progress and equipment status); L f is the production load, P l is the production progress, S d is the critical equipment status, L f , P l and S d From And at a certain time interval t, use the function U(e,a dynamic )renew.

[0252] Step 5.2 defines the transition probability.

[0253] In the ant colony algorithm, ants are assigned to the ij (t) Select the next process step S j . This transition probability can be defined as:

[0254]

[0255] Where: Ω i (t) is the time at which the current process step S i The set of all process steps that can be reached, i.e. part of the solution space; τ ij (t) is the time from S i To S j The pheromone concentration of ij (t) is the time from S i To S jThe heuristic information is calculated based on static data and real-time data; α and β are parameters that control the impact of pheromone and heuristic information on the transition probability.

[0256] Step 5.3 Pheromone Update.

[0257] After each iteration, the pheromone concentration is updated according to the quality of the process route found by the ants. Assuming y*(t) is the optimal process route found at time t, the pheromone concentration can be updated as follows:

[0258] τ ij (t+1)=(1-ρ)·τ ij (t)+ρ·Δτ ij (t)

[0259] Where: ρ is the pheromone volatilization rate, which is used to control the decay rate of pheromone; Δτ ij (t) is the time from S i To S j The pheromone increment is defined as Δτ i j(t)=1 / f(y*(t)). Through the pheromone update mechanism, the high-quality solutions in the solution space will gradually enhance their attractiveness.

[0260] Step 5.4: Iterative optimization.

[0261] Repeat the steps of constructing the solution space and updating the pheromone for multiple iterations. As the iterations proceed, the pheromone concentration will gradually concentrate on the path corresponding to the better process task allocation solution, so that the entire ant colony gradually converges to the optimal solution or a near-optimal solution.

[0262] Step 5.5 outputs the results.

[0263] When the maximum number of iterations is reached or other termination conditions are met, the iteration stops. The current optimal process task allocation plan and its corresponding quality indicators including total processing time and total cost are output.

[0264] Example description:

[0265] Suppose an automobile company has multiple factories, each factory has multiple production lines, and each production line has multiple equipment. Now the company needs to produce automobile products. Automobiles are composed of multiple components such as engines, gearboxes, etc. Each component is composed of multiple parts, and the process route of each part is All are generated in step 4. In step 1.2, the component attributes of factories, production lines, equipment and some key equipment and The four are assigned and updated in real time through information systems and factory sensor data.

[0266] The output result of using the ant colony algorithm for one allocation is as follows:

[0267] {"Factory 1":{"Production load":70,"Available capacity":30,"Number of tasks to be processed":5,"Production line A":{"Production progress":60,"Line speed":50,"Equipment E1":{"Running time":100,"Failure rate":2,"Maintenance record":"2024-07-01","Current task":[{"Time period":"08:00-09:30","Input":"Gearbox blank","Output":"Gearbox blank with material preparation completed","Process":"Material preparation", "Parameters":{"Material":"Chrome-nickel alloy steel"}}]},"Equipment E2":{"Running time":150,"Failure rate":5,"Maintenance record":"2024-07-05","Current task":[{"Time period":"09:30-12:30","Input":"Gearbox blank with material preparation","Output":"Gearbox blank with forging","Process":"Forging","Parameters":{"Temperature":"1050℃","Pressure":"320MPa"}}]}},…}

[0268] Due to the properties of factories, production lines, equipment and some key equipment components and It will be updated in real time, and the allocation plan will be updated regularly as needed.

Claims

1. A complex component process route generation method based on dynamic knowledge graph, characterized in that: The following steps are involved: Step 1: Construct a dynamic knowledge graph G based on the static knowledge extracted from the historical process image set and historical process text set of complex components, as well as the updated dynamic attribute set associated with the factory entity, production line entity, key equipment entity, and equipment component entity. dynamic ; Step 2: Based on the demand text of the complex component, construct the demand knowledge graph G Ne ; Step 3: Construct the GRNN network model and use X = [G dynamic , G Ne ] to train it; Step 4: Process route prediction; Based on the process image set and process text set of the complex components corresponding to the current task, as well as the updated dynamic attribute set associated with the current factory entity, production line entity, key equipment entity and equipment component entity, the task dynamic knowledge graph G of the current task is constructed. dynamic_task ; Based on the requirement documents of the complex components corresponding to the current task, build the task requirement graph G of the current task Ne_task ; Set X_task = [G dynamic_task , G Ne_task ]The trained GRNN network model is input to cyclically predict the next process step, and the GRNN network model outputs the process route prediction result of the complex component corresponding to the current task.

2. The method for generating a complex component process route based on a dynamic knowledge graph according to claim 1, characterized in that: The dynamic knowledge graph G in step 1 dynamic =(V, E, A′ global , U); V is a node set, which is constructed based on the entity set in the dynamically updated global knowledge set; E is an edge set, which is constructed between nodes based on the relationship set in the dynamically updated global knowledge set; A′ global is the global attribute set in the dynamically updated global knowledge set; U is the update function used to update the attributes of the entity / node corresponding to the entity; The method for acquiring the dynamically updated global knowledge set is: Firstly, static knowledge is extracted from the historical process images and process texts of complex parts respectively, and then merged to form a global knowledge set, which includes a global entity set, a global relationship set and a global attribute set. Then, a dynamic attribute set is defined, and the dynamic data associated with the factory entity, production line entity, key equipment entity and equipment component entity are integrated into the dynamic attribute set to update it. The updated dynamic attribute set is merged, deduplicated and standardized with the global attribute set in the global knowledge set to obtain a dynamically updated global knowledge set.

3. The complex component process route generation method based on dynamic knowledge graph according to claim 1 or 2, characterized in that: The historical process image collection described in step 1 includes part drawings, process tabs and design manuals in the form of images, from which static knowledge is extracted using visual processing technology; the historical process text collection includes process documents and process manuals in the form of natural language, from which static knowledge is extracted using natural language processing technology; the static knowledge includes entities, relationships and attributes; The process image set of the complex component corresponding to the current task described in step 4 includes part drawings, process tabs and design manuals in image form, where the process tab is empty; the process text set of the complex component corresponding to the current task includes process documents and process manuals in natural language form, where the process documents are empty.

4. The method for generating a complex component process route based on a dynamic knowledge graph according to claim 3 is characterized in that: The visual processing technology is OCR or Faster R-CNN; the natural language processing technology is Transformer.

5. The method for generating a complex component process route based on a dynamic knowledge graph according to claim 3 is characterized in that: Demand knowledge graph G in step 2 Ne =(V Ne , E N , A Ne , U=1); V Ne is a node set, constructed according to the entity set extracted from the requirement document; N A is an edge set, constructed between nodes based on the relationship set extracted from the requirement document, representing the relationship between entities; Ne is a set of attributes extracted from the requirement document, U is an update function used to update the attributes of the entity / node corresponding to the entity, and U=1 indicates that the attributes of the entity / node corresponding to the entity are not updated.

6. The method for generating a complex component process route based on a dynamic knowledge graph according to claim 5, characterized in that: The training goal of the GRNN network model in step 3 is to minimize the prediction error of the process route. An end-to-end training method is adopted, teacher forcing technology is used, and ground truth is used as input during training to improve the convergence speed.

7. A complex component process route generation and allocation method based on dynamic knowledge graph, characterized in that: The following steps are involved: Step 1: Generate a process route for a complex component to be manufactured using the complex component process route generation method described in any one of claims 1 to 6; Step 2: Use the ant colony algorithm to optimize the process route generated in the first step and generate the best process route.

8. The complex component process route generation and allocation method based on dynamic knowledge graph according to claim 7 is characterized in that: The second step is as follows: Step 1) Define the objective function f(y, t): Where t is the current time; w1, w2, w3 are weights used to balance the importance of different items; Parameter P used to calculate the current process step i With ideal parameters The function of the difference between i , L f , P l , S d , t) is used to calculate the current process step S i Operation type O i Compatibility or impact with real-time data; f is the production load, P l is the production progress, S d is the critical equipment status, L f , P l and S d From And at a certain time interval t, use the function U(e, a dynamic )renew; They are factory entity attribute set, production line attribute set and key equipment attribute set; Step 2) Define the transition probability: Among them, Ω i (t) is the time at which the current process step S i The set of all process steps that can be reached, i.e. part of the solution space; τ ij (t) is the time from S i To S j The pheromone concentration of ij (t) is the time from S i To S j The heuristic information is calculated based on static data and real-time data; α and β are parameters that control the impact of pheromone and heuristic information on the transition probability; Step 3) Update the pheromone as follows: t ij (t+1)=(1-ρ)·τ ij (t)+ρ·Δτ ij (t) Where: ρ is the pheromone volatilization rate, which is used to control the decay rate of pheromone; Δτ ij (t) is the time from S i To S j The pheromone increment is defined as Δτ ij (t) = 1 / f(y*(t)); Step 4) Iterative optimization: Repeat the steps of constructing the solution space and updating the pheromone for multiple iterations; Step 5) Output results: When the maximum number of iterations is reached or the set iteration termination condition is met, the iteration is stopped and the current optimal process task allocation plan and its corresponding quality indicators are output, including total processing time and total cost.

Images