WO2020167316A1

WO2020167316A1 - Machine learning-based generative design for process planning

Info

Publication number: WO2020167316A1
Application number: PCT/US2019/018222
Authority: WO
Inventors: Sanjeev SRIVASTAVA; Thomas Gruenewald; Stephan Grimm; Pavel Gocev; Livio Dalloro; Alois Haselböck; Giray HAVUR; Denis Smirnov; Martin RINGSQUANDL; Dian Song Cao
Original assignee: Siemens AG; Siemens Corp
Current assignee: Siemens AG; Siemens Corp
Priority date: 2019-02-15
Filing date: 2019-02-15
Publication date: 2020-08-20
Anticipated expiration: 2021-08-15

Abstract

For manufacturing design by a machine-learned network (46), a machine-learned model (44) is trained to generate (13) a bill of process (BOP) given information from a bill of materials (BOM). Rather than rely on manual processes and associated problems, the machine-learned model (44) provides a BOP given a new product design that is described in the BOM. The machine-learned model (44) may more likely generate an optimized BOP by using a structured input in the form of a knowledge graph. Other machine-learned models (44) may be used, such as for generating (31) the knowledge graph and/or application for selecting from various possible BOPs.

Description

MACHINE LEARNING-BASED GENERATIVE DESIGN FOR PROCESS

PLANNING

BACKGROUND

[0001] The present embodiments relate to generation of a bill of process for manufacturing a designed product, such as a gas turbine blade. The scope of the expertise required to build a turbine part involves not only the knowledge of current machinery but also the sequence of process steps to construct the product. An experienced engineer reviews documentation to design a process that can achieve the desired product.

[0002] For example, generation of a manufacturing process plan for a turbine and/or turbine parts (e.g., blades and/or vanes) is a manual process.

A part file is extracted from Bill of Materials (BOM) of a computer assisted design (CAD) file based on manufacturing requirements. Desired design and/or geometric features are captured in the part file. A manufacturing or shop engineer uses the CAD file to analyze the manufacturing requirements in the form of manufacturing features that need to be created. The various available resources (e.g., machines, tools, shop floor engineers, etc.) are analyzed for generation of required manufacturing features from the desired design and/or geometric features. Various sequential manufacturing operations to create each manufacturing feature are generated. These manufacturing operations are then stored in an enterprise resource planning (ERP) system and are in the form of human readable text. This information is downloaded from the ERP system and converted into a bill of process (BOP) by another shop floor engineer. The BOP is then passed on to the factory for generation of machine programs for actual implementation of determined manufacturing operations.

[0003] These steps are time consuming as the various steps may require interaction among human engineers that are distributed geographically. The human engineers use their tacit knowledge, different tools and exchange information in various formats, hence leading to more iterations in generating the BOP. These human engineers may make mistakes leading to

manufacture of a defective part. The tacit knowledge is distributed, so errors may be more likely to occur. The manufacturing steps are entered into the ERP system in a natural language format, so the engineer may make spelling mistakes leading to confusion, downstream problems, and defective part production.

SUMMARY

[0004] By way of introduction, the preferred embodiments described below include methods, systems, instructions, and computer readable media for manufacturing design by a machine-learned network. A machine-learned model is trained to generate a bill of process (BOP) given information from a bill of materials (BOM). Rather than rely on manual processes and

associated problems, the machine-learned model provides a BOP given a new product design that is described in the BOM. The machine-learned model may more likely generate an optimized BOP by using a structured input in the form of a knowledge graph. Other machine-learned models may be used, such as for generating the knowledge graph and/or application for selecting from various possible BOPs.

[0005] In a first aspect, a method is provided for generative design by an artificial intelligence processor. A BOM for a turbine or turbine part is received. The artificial intelligence processor generates a BOP for the turbine or turbine part. A first machine-learned network generates BOP in response to input of information from the BOM. The bill of process for the turbine or turbine part is stored for later use in manufacturing.

[0006] In a second aspect, a system is provided for machine-learning- based design. A processor is configured by instructions stored in a memory. The instructions, when executed by the processor, are to generate a routing solution for manufacture of a design from input of design parameters. The routing solution is generated by a first machine-learned model in response to the input. The routing solution for the design is output.

[0007] In a third aspect, a method is provided for generative design by a design processor. Past routing solutions for manufacture of a product are received. A routing solution includes manufacturing steps, resources used for the manufacturing steps, and timing of the various manufacturing steps. A knowledge graph is generated from the past routing solutions and a pre defined manufacturing ontology. The knowledge graph includes entities, properties of entities, and relationships between entities. A bill of process for an unseen product design may be generated using the knowledge graph.

[0008] The present invention is defined by the following claims, and nothing in this section should be taken as a limitation on those claims.

Further aspects and advantages of the invention are discussed below in conjunction with the preferred embodiments and may be later claimed independently or in combination.

BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The components and the figures are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the invention. Moreover, in the figures, like reference numerals designate corresponding parts throughout the different views.

[0010] Figure 1 is a flow chart diagram of one embodiment of a method for generative design by an artificial intelligence processor;

[0011] Figure 2 illustrates learning and generation of a routing solution from bill of materials information;

[0012] Figure 3 is a flow chart diagram of one embodiment of a method for generative design by a design processor;

[0013] Figure 4 is a block diagram of one embodiment of a system for machine-learning-based design.

DETAILED DESCRIPTION OF THE DRAWINGS AND PRESENTLY PREFERRED EMBODIMENTS

[0014] A generative manufacturing design approach is provided for turbines, turbine blades, turbine vanes, and/or other parts. A manufacturing process plan is generated automatically using a machine-learned model.

Various sequential manufacturing operations to create each required manufacturing feature and the specific manufacturing tasks (e.g., BOP) for each sequential manufacturing operation are generated based on the available manufacturing resources. Human involvement in BOP generation is reduced by using the machine-learned model, making the manufacturing planning process faster, more error-free, and more structured. [0015] A part to be manufactured may be described as a combination of manufacturing features. In turn, each manufacturing feature may be described as an outcome of execution of a set of manufacturing operations. Machine training is used to complete the link from manufacturing features to the operations. For training, routing solutions generated by human engineers and stored as natural language text are converted into knowledge modules in a knowledge base. The knowledge base may be enhanced through new routing solutions that are generated by searching the knowledge base itself and utilizing the past routing solution knowledge stored in the knowledge base. Machine learning uses the knowledge base to automate generation of routing solutions, making the manufacturing planning process faster, more error-free, and more structured.

[0016] The object or product for which a BOP is to be designed is of any type. In the examples below, a turbine or turbine part, such as a blade or vane, is designed. Other objects include, among others, a suspension, printed circuit board, axle, circuit, transformer, impeller, topology for any type of part, structure for any type of part (e.g., wheel or hull design).

[0017] Figure 1 shows one embodiment of a method for generative design by an artificial intelligence processor. Information from a bill of materials is input to a machine-learned model, and the artificial intelligence processor outputs a BOP.

[0018] The method is implemented by a machine, computer, and/or server. For example, given a BOM, the artificial intelligence processor generates one or many BOP(s) as output. A controller may use this output to operate one or more manufacturing machines.

[0019] The acts are performed in the order shown (i.e. , top to bottom or numerical order) or other orders. For example, act 13 is performed prior to act 12. As another example, act 15 is performed before act 14.

[0020] Additional, different, or fewer acts may be provided. For example, acts 14, 15, and/or 16 are not provided. As another example, act 11 is not performed. In yet another example, act 12 is not performed. In other examples, other intermediary acts to generate the input from the BOM are provided. [0021] In act 10, the artificial intelligence processor or other processor receives a bill of materials. Receipt is by loading from memory, transfer over a computer network, and/or manual entry.

[0022] The bill of materials is an excel file or of another file format. The requirements, part numbers, or other manufacturing information are extracted from the bill of materials. For example, a part file is extracted from a CAD file based on a part number of the bill of materials. For a turbine, a blade or vane design may be extracted from a CAD file for the turbine based on a part number in the bill of materials. The CAD file and/or the bill of materials are received for extracting the design of the part to be manufactured.

[0023] Design and/or geometric features of the extracted part may be identified. The part file or other extracted information defines the design of the part, including materials, size, geometry, joinings, and shape.

Alternatively, the bill of materials provides the design and/or geometric features.

[0024] The manufacturing features are values for how the design and/or geometric features are made. For example, it is a manufacturing feature of a cone to extrude the cone with specific dimensions. Where to cut, how to finish, or other manufacturing features for making the design and/or geometry are provided.

[0025] In act 11 , the processor or artificial intelligence processor extracts manufacturing features from the bill of materials. Manual extraction may be used.

[0026] A processor may extract these design and/or geometric features from a part file, or it may extract the part file from the bill of materials.

Alternatively, a user can manually extract the information.

[0027] The extraction may be direct from the bill of materials. Alternatively, the extraction is from design and/or geometric features. In yet other embodiments, the bill of materials or information provided in the bill of materials is used without extraction.

[0028] In act 12, the processor or artificial intelligence processor formats the information from the bill of materials. The manufacturing features are formatted for input to a machine-learned model. Since the machine-learned model is trained based on samples having a given format, the same format is used for application of the information to the model.

[0029] Programmed operations are used for reformatting, as needed. Alternatively, a machine-learned model is used to reformat. The machine- learned model is capable of receiving information in one or more formats and output the information in a standard format.

[0030] The format may be structured and/or unstructured. A table, computer file (e.g., CAD file for the part), spreadsheet (e.g., BOM), and/or another format may be used. For example, an excel spreadsheet with columns for information on various parts, subparts, process, and

manufacturing resources, includes entries for different turbine parts. An annotation, such as plain text with or without misspellings, may be linked to a column heading, row heading, or entry (e.g., value). The annotation may be design or manufacturing notes.

[0031] In one embodiment, the format to capture the manufacturing knowledge from the past routing solutions is a knowledge graph. The knowledge graph is structured based on a pre-defined manufacturing ontology. For example, a product, process, resource (PPR) format can be used for the ontology. Entities, the semantic types, properties, and

relationship between entities is provided in a node and edge format.

[0032] For generating a BOP, the ontology format is used for entry of the manufacturing features, product, resources or other information from the BOM into the machine-learned network. The machine-learned network is trained based on many samples of input information and corresponding ground truth outputs (e.g., many samples of BOM extracted information and linked or corresponding BOP as ground truth for each input sample). Both the BOM information and the BOP may be represented as a knowledge graph.

[0033] Figure 2 shows an example approach with a learning phase and a generation phase. For learning, past routing solutions 20 are provided to a knowledge distiller 22 to form the knowledge base 23 of knowledge graphs. A knowledge refiner 24 may add to or alter the knowledge base 23. A routing solutions generator 27 is then trained based on the knowledge base 23. For application, new product requirements 25 are received (e.g., act 10). Required features 26 (e.g., input feature vector/data) are extracted and formatted (e.g., acts 11 and/or 12). The routing solutions generator 27 (e.g., machine-learned model) generates (e.g., act 13) one or more routing solutions 28 for the new product of the new product requirements 25. A solution quantifier 29 ranks and selects (e.g., act 14) the routing solutions 29 to identify and store 15 one of the routing solutions for manufacturing the product.

[0034] Figure 3 shows one embodiment of a method for generative design by a design processor. The method is for creating the knowledge base 23 used to machine train and/or for formatting new product requirements 25 into the manufacturing features 26 as the input to the generator 27. The discussion below is in the context of training, but the same method may be used to reformat for application. The discussion below is in the context of formatting for routing solutions, but the same method may be used to reformat the input manufacturing features and/or other information extracted from the BOM.

[0035] The acts of Figure 3 are performed by a processor, such as a design processor of a workstation or computer. A memory and/or user interface may be used for interacting with and/or controlling any of the acts of Figure 3.

[0036] The acts of Figure 3 are performed in the order shown (e.g., numerical) or other orders. For example, acts 32 and 33 may be performed in reverse order or simultaneously. Act 34 may be performed once again after act 35.

[0037] Additional, different, or fewer acts may be provided. For example, acts 32 and/or 33 are not provided. As another example, act 34 is not provided. Act 30 may be for a BOM rather than BOP. In yet another example, act 35 is not provided.

[0038] For training a model, training data is received. This training data includes hundreds, thousands, tens of thousands, or other number of samples with ground truth . The generative model takes as input BOM information and outputs BOP. Many samples of BOM information and linked BOP outputs are provided for training. The formats for these samples are normalized, such as formatting as structured information in knowledge graphs.

[0039] In act 30, one or more routing solutions are received. The reception is by a machine used for machine learning via a network or loading from memory. Another processor, such as an engineering workstation processor, may receive the routing solution.

[0040] The routing solution is a BOP for manufacture of a product. For example, BOPs for many different turbine blades are received. The routing solution includes manufacturing steps, resources used for the manufacturing steps, and timing of the manufacturing steps. The routing solution may include other information converting the manufacturing features into

manufacturing steps using machines (e.g., equipment) and material to make the part.

[0041] In act 31 , the machine or other processor generates a knowledge graph. Hand coding and/or a machine-learned model as a knowledge distiller reformats the routing solution into a knowledge graph. A plurality of the structured knowledge graphs provides a knowledge base for machine training.

[0042] The knowledge graph includes entities, properties of entities, and contextual relationships between entities in an ontology. Other ERP data structures may be used. Past manufacturing routing solutions are used to extract useful information (knowledge). The past routing solutions, generated by human engineers with natural language annotations, may include various manufacturing steps, resources (machines) used in each manufacturing steps, and other information like time required for setting up machines and process time, etc.

[0043] The received routing solutions may be pre-processed (e.g., removal of erroneous information), reduced (e.g., unsupervised dimensionality reduction), and/or expanded (e.g., kernel approximation). The knowledge graph is then generated (e.g., feature vector extraction for training) as a feature representation. Past designs are used to derive information and patterns to achieve some specific objectives (e.g., less time, less

computational resources, faster design exploration, faster validation, generation of new design, etc.). The past instances for a given type of product (e.g., turbine vane or blade) are used to learn to provide BOP for an unseen design (e.g., BOM requirements). Performance measures may be included for the various samples in the training data in order to optimize the BOP learned to be output.

[0044] In act 32, the knowledge graph may be updated. An entry in the graph (e.g., a node and/or edge) is inferred from unstructured data. For example, a document, annotation, other natural language text, or graphics are provided with the BOP or BOM in the training data. Rather than ignoring the unstructured data, an entry is made or changed in the knowledge graph based on the unstructured data.

[0045] A machine-learned model may be used to infer from the unstructured data. Alternatively, a semantic processor or process is applied.

[0046] The unstructured data is scanned into a structured format. The unstructured data may be reformatted into the format of the knowledge graph directly. Alternatively, the unstructured data is reformatted into a structured format, which is then used to generate the entry into the knowledge graph.

[0047] In one embodiment, the unstructured data is formatted into a structured format, such as a“Product”,“Process”, or“Resource” category.

The contextual relationships between entities represented in the unstructured data may be used to add relationship links or edges in the knowledge graph. Entities and/or properties of entities expressed in the unstructured data may be used to add or change entries in the knowledge graph.

[0048] By generating knowledge graphs for different product designs, a knowledge base is created for machine training. The knowledge base stores the generated knowledge in a structured format, such as knowledge graphs. Other structured formats, such as a table or spreadsheet, may be used.

Knowledge graphs are networks of entities, their semantic types, properties and relationship between entities. The underlying basis of a knowledge graph is an ontology. An ontology specifies the semantics of the data and is typically based on logical formalisms that support some form of inference.

This inference and reasoning capability allow implicit information to be derived from explicitly asserted data. This facilitates inference of information that can be otherwise hard to discover. For manufacturing planning process, an ontology can be defined to capture the various entities (e.g., parts of a blade, geometric features, manufacturing features, machine, material, etc.), their properties (e.g., geometrical parameters of a feature, machine location, machine setup time, etc.), and relationship between entities (e.g.,“hasA”, “partOf”,“isCreatedBy”,“followedBy”, etc.). The past manufacturing routing solutions may be defined as instances of the ontology. Each may be its own knowledge graph. Alternatively, a knowledge graph is formed from a combination of graphs from the different samples.

[0049] In act 33, the processor or artificial intelligence processor infers an entry for the knowledge graph from other knowledge graphs. The collection of knowledge graphs is searched for commonalities. The knowledge base is refined. Useful knowledge (e.g., rules, constraints, logic, etc.) are extracted. This new knowledge is then added in the knowledge base, creating a self learning process. For example, if there are two constraints reflected in the knowledge graphs - 1 ) Process B is always preceded by Process A; and 2) Process C is always preceded by Process B; then a new constraint - Process C is always preceded by Process A, is derived and stored in the knowledge base. This knowledge is added to the applicable knowledge graphs.

[0050] The inference is by an inference engine that reasons on the knowledge stored in the knowledge base. The inference engine is a machine- learned model, such as a neural network (NN), clusterer, or support vector machine. In one embodiment, representation learning and corresponding architecture are used to machine train the inference engine to extract the knowledge and refine the knowledge base.

[0051] In act 34, the knowledge base is used in machine training. The artificial intelligence processor or other processor of a machine uses the knowledge base as training data. Other training data may be used, such as resources (e.g., manufacturing machines), resource properties (e.g., time to operate), costs, materials, performance indicator, or other information associated with a manufacturing facility and capability. The various data and the BOM of the knowledge base are used as input vector samples. The BOPs of the knowledge base for the corresponding BOMs are used as ground truth. In alternative embodiments, the BOPs and BOMs of the training data are not of the knowledge base and/or are in a format other than knowledge graphs. The machine training trains to convert BOM information to a BOP, such as a routing solution for manufacture of a product.

[0052] Any architecture and corresponding type of machine learning may be used. In one embodiment, deep learning of a neural network is used. In another embodiment, constrained learning, such as based on graph searching, is used with constraints from the knowledge base, such as inferred constraints. Machine learning architectures with memory (e.g., long term short term) and/or attention augmented neural networks may be used.

[0053] A machine trains the model (e.g., neural network) to generate one or more BOPs given unseen BOM information. The machine trains the network to receive values for BOM information and output values for a routing solution for manufacture.

[0054] Deep learning is used to train the artificial neural networks. Any deep learning approach or architecture may be used. The deep learning trains filter kernels, connections, weights, and/or other features that, in combination, indicate the values for outputs given the inputs. Features useable to determine the output may be learned using deep learning. The deep leaning provides a deep machine-learned network that outputs results given previously unseen inputs. The training data is used by the machine to learn to output values given unseen input values.

[0055] The architectures of the networks may include convolutional, sub sampling (e.g., max or average pooling), fully connected layers, recurrent, SoftMax, concatenation, dropout, residual, and/or other types of layers. Any combination of layers may be provided. Any arrangement of layers may be used. Skipping, feedback, or other connections within or between layers may be used. Hierarchical structures may be employed, either for learning features or representation or for classification or regression.

[0056] The architectures for deep learning may include a convolutional neural network (CNN) or convolution layer. The CNN defines one or more layers where each layer has a filter kernel for convolution with the input to the layer. Other convolutional layers may provide further abstraction, such as receiving the output from a previous layer and convolving the output with another filter kernel.

[0057] A fully connected neural network may be trained using a back- propagation learning approach. The architecture may include one or more dense layers. The dense layers connect various features from a previous layer to an output from the layer. In one embodiment, the dense layers are fully connected layers. One or more fully connected layers are provided. The dense layers may form a multi-layer perceptron.

[0058] The network is trained for any number of epochs (e.g., 300) using any optimization, such as Adam, a learning rate (e.g., 0.0001 ), and/or a loss (e.g., mean-squared error loss (L1 )). Other arrangements, layers, units, activation functions, architectures, learning rates, optimizations, loss functions, and/or normalization may be used. Other architectures may be used.

[0059] The training identifies kernels, weights, connections, and/or other information of the architecture relating the input to the ground truth output.

The deep learning trained neural network is stored. Once trained, the machine-learned artificial neural network may be applied by an artificial intelligence processor. The network is stored so that the network may be used in application for a particular design.

[0060] In act 35, the artificial intelligence processor generates a bill of process from the knowledge graph. The knowledge graphs are used to train the machine-learned model. The artificially intelligence processor applies the trained model to generate a BOP or other routing solution. Alternatively, a given knowledge graph representing a routing solution may be converted into a format usable by a manufacturing facility to manufacture a product. The knowledge graph is used to create the bill of process for manufacturing.

[0061] Returning to Figure 1 , in act 13, the artificial intelligence processor or other processor generates a BOP for the turbine or other produce design. The BOP is generated by application of the machine-learned network (e.g., application of the routing solutions generator 27). The BOM information is input to the machine-learned network. In response to input of the information from the BOM (e.g., input of extracted manufacturing features) for a product to be manufactured, one or more routing solutions are output by the machine- learned network.

[0062] The routing solutions generator 27 generates possible routing solutions that can lead to manufacturing of the turbine part (e.g., blades and/or vanes) that meet the product requirements. The product requirements may be translated into required manufacturing features for input and/or are themselves input. The routing solutions generator 27 takes these

manufacturing features as input and then searches the knowledge base through the machine-learned model to determine the manufacturing operations through which these features were manufactured in the past routing solutions.

[0063] The searching by application of the machine-learned model generates the BOP as a routing solution in manufacture of the product. A task list of a sequence of manufacturing operations and linked machines for performing the operations in the sequence is output.

[0064] The machine-learned model provided by the machine-learned network generates one or more routing solutions. There may be several ways (e.g., different manufacturing operation combinations) a manufacturing feature may be manufactured. There may be multiple machines that can do the same manufacturing operation. This may lead to generation of multiple routing solutions as the training data reflects the multiple options. A single one or multiple different possible BOPs are generated in response to one input.

[0065] The machine-learned network may be trained to output a single BOP. For example, values of performance indicators are provided as part of the ground truth and used to train for outputting a BOP associated with best meeting of the performance. Alternatively, the machine-learned network is trained to output any number of possible BOPs. Any number of routing solutions may be output, such as a fixed number or all possible routing solutions. The performance is then evaluated on the output BOPs.

[0066] In act 14, the artificial intelligence processor or another processor ranks the BOPs. The ranking uses one or more performance indicators, such as the cost, time, or resource optimization. Any performance indicator may be used. For ranking, each given BOP or routing solution is quantified based on certain use case specific metric(s). The generated routing solutions are validated and ranked to determine the most optimal feasible solution. The ranking of feasible solutions may be done based on one or more user defined key performance indicators. The parameters of the manufacturing

environment, such as machine set-up time, process time, energy

consumption, cost, total time for production, throughput, etc., may be used.

[0067] The quantification for ranking is performed by simulation or a data- driven method. One implementation is a physics simulation. A physics model of a virtual manufacturing environment simulates operation using the routing solution. The value or values of the performance indicator or indicators are calculated by simulating the various manufacturing steps and machines for the steps in the simulation environment.

[0068] In a data-driven approach, a machine-learned network is applied. Training data for input BOPs with ground truth values for the performance indicators are used to machine train the network. Machine training a neural network or other model to output a value or values for performance indicators provides the machine-learned network. The machine-learned network may output values given an unseen BOP. The output is generated in response to the input, so may be less time and processing consuming than simulation.

The various BOPs are ranked. The BOPs are input to the machine-learned network, which outputs the ranking.

[0069] The ranking may be an organization from highest to lowest or calculation of values for searching for a highest or lowest, depending on the performance indicator. The machine-learned network ranks by determining the values of the performance indicator for the different solutions.

Alternatively, the machine-learned network ranks by ordering the routing solutions.

[0070] The design routing solutions may be ranked, and the most optimal solution may be selected by the processor. Alternatively, the ranking is provided as an output so that a process engineer may select the BOP to use.

[0071] In act 15, the BOP is stored. The BOP for the product to be built (e.g., turbine or turbine part) is stored. The processor or artificial intelligence processor stores the BOP or BOPs in a memory. For example, the BOP is stored in an ERP system and/or at a manufacturing facility. The selected BOP, a number of highest-ranking BOPs, or all the possible BOPs are stored.

[0072] The stored BOP may be used directly for manufacturing.

Alternatively, the BOP is presented to an engineer for verification and/or modification.

[0073] In act 16, a manufacturing facility (e.g., factory or other collection of machines) manufactures the product (e.g., turbine or turbine part) according to the BOP. The tasks in the BOP are performed. The machines are set-up or configured. The materials are gathered and processed by the machines as configured. The tasks are performed in the sequence indicated in the BOP to manufacture the product. The tasks may be repeated to manufacture more than one product.

[0074] Figure 4 shows one embodiment of a system for machine-learning- based design. One or more machine-trained models (e.g., neural networks) 44, 46 are applied. The system implements the method of Figure 1 , the method of Figure 3, or the process of Figure 2. Other methods may be implemented. In alternative embodiments, the system is used to train the models 44, 46. The system may implement the method of Figure 3, but another machine learning may be used.

[0075] The system includes a design processor 40 and a memory 42. Additional, different, or fewer components may be provided. For example, a network or network connection is provided, such as for networking the design processor 40 and the memory 42. In another example, a display and/or user interface (e.g., display and user input device - mouse and keyboard) are provided for interaction with the learning, input BOM, and/or output BOP.

[0076] The design processor 40 and/or memory 42 are part of an engineering workstation. Alternatively, the design processor 40 and/or memory 42 are part of a separate computer or server.

[0077] The memory 42 is a graphics processing memory, a video random access memory, a random-access memory, system memory, cache memory, hard drive, optical media, magnetic media, flash drive, buffer, database, combinations thereof, or other now known or later developed non-transitory memory device for storing data. The memory 42 is part of the design processor 40, part of a computer associated with the design processor 40, part of a database, part of another system, or a standalone device.

[0078] The memory 42 stores the training data, a knowledge base, extracted BOM information, the BOM, BOPs, values for performance indicators, and/or the machine-learned models 44, 46. Only one, two, or three or more machine-learned models may be stored. The memory 42 may store other information. For example, the values of the input feature vector, the values of the output vector, the connections/nodes/weights/convolution kernels of the deep machine-learned models 44, 46, calculated values in applying the models 44, 46, selections, rankings, and/or other data from manufacturing design are stored. The design processor 40 may use the memory 42 to temporarily store information during performance of the methods of Figure 1 and/or 3.

[0079] The memory 42 or other memory is alternatively or additionally a non-transitory computer readable storage medium storing data representing instructions executable by the programmed design processor 40. The instructions for implementing the processes, methods and/or techniques discussed herein are provided on non-transitory computer-readable storage media or memories, such as a cache, buffer, RAM, removable media, hard drive, or other computer readable storage media. Non-transitory computer readable storage media include various types of volatile and nonvolatile storage media. The functions, acts or tasks illustrated in the figures or described herein are executed in response to one or more sets of instructions stored in or on computer readable storage media. The functions, acts or tasks are independent of the particular type of instructions set, storage media, processor or processing strategy and may be performed by software, hardware, integrated circuits, firmware, micro code and the like, operating alone, or in combination. Likewise, processing strategies may include multiprocessing, multitasking, parallel processing, and the like.

[0080] In one embodiment, the instructions are stored on a removable media device for reading by local or remote systems. In other embodiments, the instructions are stored in a remote location for transfer through a computer network or over telephone lines. In yet other embodiments, the instructions are stored within a given computer, CPU, GPU, or system.

[0081] The design processor 40 is an artificial intelligence processor, such as a general processor, central processing unit, control processor, graphics processor, digital signal processor, application specific integrated circuit, field programmable gate array, massively parallel processor, processor designed specifically to implement machine-learned networks, digital circuit, analog circuit, combinations thereof, or other now known or later developed device for generating manufacturing designs. The design processor 40 is a single device or multiple devices operating in serial, parallel, or separately. The design processor 40 may be a main processor of a computer, such as a laptop, server, or desktop computer, or may be a processor for handling some tasks in a larger system. The design processor 40 is configured by

instructions, firmware, design, hardware, and/or software to perform the acts discussed herein.

[0082] The design processor 40 is configured by instructions or software stored in the memory 42 to format design parameters as a knowledge graph in an ontology. Information from a BOM, such as requirements, are formatted for input to the machine-learned model 44. The knowledge graph may include nodes and edges reflecting the ontology. The ontology provides for an entity, entity property, or relationship in the knowledge graph. The nodes, edges, entity, property, and/or relationship may be determined, at least in part, by application of the other machine-learned model 46. For example, information for the knowledge graph is inferred from text linked to one or more design parameters. As another example, information for the knowledge graph is inferred from other knowledge graphs.

[0083] The design processor 40 is configured by instructions or software stored in the memory 42 to generate a routing solution for manufacture of a design from input of the design parameters. The machine-learned model 44 generates one or more routing solutions in response to the input. Where multiple routing solutions are output, the design processor 40 may be configured to select one of the routing solutions from the output solutions based on one or more performance indicators. The design processor may apply another machine-learned network (e.g., input routing solutions for output of values of the performance indicator) or may apply a physics simulation to determine values for the performance indicator. The design processor 40 outputs one or more routing solutions for the designed product.

[0084] While the invention has been described above by reference to various embodiments, it should be understood that many changes and modifications can be made without departing from the scope of the invention. It is therefore intended that the foregoing detailed description be regarded as illustrative rather than limiting, and that it be understood that it is the following claims, including all equivalents, that are intended to define the spirit and scope of this invention.

Claims

I (WE) CLAIM:

1. A method for generative design by an artificial intelligence processor (40), the method comprising:

receiving (10) a bill of materials for a turbine or turbine part;

generating (13), by the artificial intelligence processor (40), a bill of process for the turbine or turbine part, the generating (13) being by a first machine-learned network (46) in response to input of information from the bill of materials; and

storing (15) the bill of process for the turbine or turbine part.

2. The method of claim 1 wherein receiving (10) the bill of materials comprises receiving (10) a computer assisted design file and wherein generating (13) comprises generating (13) the bill of process as a routing in manufacture of the turbine design.

3. The method of claim 2 further comprising extracting (1 1 ) manufacturing features from the bill of materials, and wherein generating (13) in response to the input of the information comprises generating (13) in response to the input of the manufacturing features.

4. The method of claim 2 wherein generating (13) the routing comprises generating (13) a task list of a sequence of manufacturing operations and linked machines for performing the operations in the sequence.

5. The method of claim 1 wherein generating (13) comprises generating (13) with the machine-learned network (46) comprising a neural network.

6. The method of claim 1 wherein generating (13) comprises generating (13) with the machine-learned network (46) having been trained with training data comprising knowledge graphs for sample inputs and ground truths for the sample inputs.

7. The method of claim 6 wherein the knowledge graphs are in a structured format provided, at least in part, by a second machine-learned network (46) from unstructured text.

8. The method of claim 6 wherein at least one of the knowledge graphs includes a pattern identified from machine learning.

9. The method of claim 1 further comprising formatting (12) the

information from the bill of materials as a knowledge graph, and wherein generating (13) comprises inputting the information as the knowledge graph.

10. The method of claim 1 wherein generating (13) comprises generating (13) the bill of process as one of a plurality of bills of process, and

further comprising ranking (14) the bills of process of the plurality by a second machine-learned network (46), the one of the bills of process being a highest ranking (14) of the plurality.

11. The method of claim 10 wherein ranking (14) comprises ranking (14) by output of a performance indicator by the second machine-learned network (46) in response to input of the bills of process.

12. The method of claim 1 further comprises manufacturing (16) the turbine or turbine part according to the bill of process.

13. A system for machine-learning-based design, the system comprising: a processor (40) configured by instructions stored in a memory (42), the instructions when executed by the processor (40) being to:

generate a routing solution for manufacture of a design from input of design parameters, the routing solution generated by a first machine-learned model (44) in response to the input;

output the routing solution for the design.

14. The system of claim 13 wherein the processor (40) is configured to format the design parameters as a knowledge graph in an ontology.

15. The system of claim 14 wherein the processor (40) is configured to include an entity, entity property, or relationship in the knowledge graph by a second machine-learned model (44) in response to input of text linked to at least one of the design parameters.

16. The system of claim 13 wherein the processor (40) is configured to generate the routing solution as one of a plurality of routing solutions and to select the one routing solution from the plurality based on a performance indicator generated by a second machine-learned model (44) in response to input of the routing solutions of the plurality.

17. A method for generative design by a design processor (40), the method comprising:

receiving (30) past routing solutions for manufacture of a product, the past routing solutions including manufacturing steps, resources used for the manufacturing steps, and timing of the manufacturing steps;

generating (31 ) a knowledge graph from the past routing solutions and a pre-defined manufacturing ontology, the knowledge graph including entities, properties of entities, and relationships between entities; and

generating (35) a bill of process using the knowledge graph.

18. The method of claim 17 further comprising inferring (33) an entry for the knowledge graph with a machine-learned model (44) from other knowledge graphs.

19. The method of claim 17 further comprising machine training (34) a neural network to convert bill of material information to a bill of process, the training data including the past routing solutions.

20. The method of claim 17 further comprising inferring (32) an entry for the knowledge graph with a machine-learned model (44) from unstructured data.