Twin decision-making method for machine tool processing based on knowledge enhancement of large language model

Abstract

The present application discloses a twin decision method for machine tool processing technology based on large language model knowledge enhancement, which relates to the field of machine tool processing technology. The method includes: constructing a processing technology knowledge graph based on the processing technology historical data and using a large language model; updating the processing technology knowledge graph using digital twin data to obtain a process knowledge graph enhanced by the digital twin data; determining the processing technology solution set under the current processing task based on the enhanced process knowledge graph according to the current processing task; adjusting the weight value of the optimization target according to the acquired tool wear state to obtain the optimal weight value of the optimization target; determining the optimal processing technology solution under the current processing task based on the processing technology solution set according to the optimal weight value. The present application constructs and utilizes multimodal digital twin data to enhance the process knowledge graph, and adaptively determines the optimal process solution in combination with the real-time tool wear state, thereby ensuring the processing quality of the machine tool.

Description

Technical Field

[0001] The present application relates to the field of machine tool processing technology, and in particular to a twin decision method for machine tool processing technology based on large language model knowledge enhancement. Background Art

[0002] As the manufacturing industry transforms towards intelligence and digitalization, the optimization of CNC machine tool processing technology has become a key factor in improving product quality and reducing costs. Traditional processing technology optimization methods mainly rely on engineers' experience and repeated experiments, which is not only time-consuming and labor-intensive, but also difficult to adapt to complex and changing production environments. In recent years, digital twin technology has been widely used in the manufacturing industry, providing important support for the realization of intelligent manufacturing and Industry 4.0. Digital twins build virtual models of physical entities to achieve real-time data interaction and state synchronization, providing new ideas for processing technology optimization. At the same time, knowledge graphs, as an advanced knowledge representation and management method, have gradually attracted attention in the manufacturing field. Knowledge graphs can systematically represent and organize complex process knowledge to provide support for intelligent decision-making.

[0003] At present, there are still the following problems in the optimization of CNC machine tool processing technology: 1) It is difficult to effectively integrate multi-source heterogeneous data; 2) There is a lack of systematic representation and utilization of process knowledge; 3) Process parameter optimization lacks real-time and adaptability; 4) There is a lack of means to optimize the process by using tool wear state quantity. These problems have become the bottleneck of CNC machine tool processing accuracy and restricted the further development of intelligent manufacturing. Summary of the invention

[0004] The purpose of this application is to provide a twin decision method for machine tool processing technology based on large language model knowledge enhancement, which can improve the processing quality of CNC machine tools.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a twin decision method for machine tool processing technology based on large language model knowledge enhancement, and the twin decision method for machine tool processing technology includes:

[0007] According to the historical processing technology data, a large language model is used to construct a processing technology knowledge graph of the machine tool; the processing technology knowledge graph is used to provide processing technology solutions for historical processing tasks.

[0008] The digital twin data is used to update the processing technology knowledge graph to obtain a processing technology knowledge graph enhanced by the digital twin data.

[0009] According to the current processing task of the machine tool, the processing technology solution set under the current processing task is determined based on the process knowledge graph enhanced by the digital twin data; the processing technology solution set is a solution set obtained with expert rules as constraints and material removal rate and surface roughness as optimization targets.

[0010] According to the acquired real-time tool wear status of the machine tool, the weight value of the optimization target is dynamically adjusted to determine the optimal weight value of the optimization target under the current machining task.

[0011] According to the optimal weight value, based on the processing solution set, the optimal processing solution for the current processing task is determined.

[0012] Optionally, the processing technology history data is a process specification of a typical workpiece; the process specification includes a process card, a process card and a procedure card.

[0013] Optionally, based on the historical processing data, a large language model is used to construct a processing technology knowledge graph for machine tools, including:

[0014] Preprocess the processing technology history data.

[0015] By fine-tuning the large language model, named entity recognition and relationship extraction are performed on the preprocessed machining process history data to obtain the key entities of the machine tool machining process and the relationships between the key entities; the key entities are represented by workpiece information, tool information, process parameters and machining accuracy requirements; the relationships include the adaptation relationship between process parameters and workpiece information, the dependency relationship between tool information and process parameters, and the influence relationship between process parameters and machining accuracy requirements.

[0016] Based on the key entities and the relationships between the key entities, a processing technology knowledge framework that reflects the association between process flows and process parameters is constructed.

[0017] Based on the processing technology historical data, data preprocessing and natural language processing analysis are used to store the key information of the processing technology historical data in the processing technology knowledge framework to obtain the processing technology knowledge graph.

[0018] Optionally, the processing technology knowledge graph is updated using the digital twin data to obtain a process knowledge graph enhanced by the digital twin data, specifically including:

[0019] The digital twin data is preprocessed; the digital twin data is multimodal digital twin data; the multimodal digital twin data includes text data, time series sensor data and point cloud data.

[0020] The multimodal digital twin data is encoded using a data encoding model to obtain semantic information of the multimodal digital twin data; the data encoding model includes Transformer Bidirectional encoding model, convolutional neural network-long short-term memory model, point cloud classification and segmentation model and autoencoder based on Transformer The bidirectional encoding model is used for text data processing; the convolutional neural network-long short-term memory model is used for time series sensor data processing; the point cloud classification and segmentation model is used for point cloud data processing; the autoencoder is used for low-dimensional representation learning of multimodal digital twin data.

[0021] According to the semantic information, a large language model is used to parse and integrate the multimodal digital twin data to obtain new key entities and the relationships between the key entities.

[0022] Based on the new key entities and the relationships between the key entities, the processing technology knowledge graph is dynamically updated to generate a process knowledge graph enhanced by digital twin data.

[0023] Optionally, the process of determining the current processing task of the machine tool is as follows:

[0024] The large language model is used to semantically encode the processing parameters, workpiece information, and tool information of the current machine tool.

[0025] A semantically enhanced retrieval method is adopted to retrieve nodes close to the current semantic encoding in the process knowledge graph enhanced by digital twin data by utilizing semantic similarity calculation and matching technology.

[0026] A large language model is used to generate a process description of the current machining task of the machine tool based on the retrieved nodes that are close to the current semantic coding; the process description is expressed based on the spindle speed, feed speed, and cutting depth.

[0027] Optionally, according to the current processing task of the machine tool, based on the process knowledge graph enhanced by the digital twin data, the processing technology solution set under the current processing task is determined, specifically including:

[0028] According to the current processing task of the machine tool, an optimization framework based on a multi-objective proximal strategy optimization algorithm is constructed; the optimization framework based on the multi-objective proximal strategy optimization algorithm is a hybrid model based on a deep reinforcement learning method and a knowledge graph; in the optimization framework, material removal rate and surface roughness are used as components of the reward function of the deep reinforcement learning method.

[0029] The strategy network and value network of deep reinforcement learning in the multi-objective proximal strategy optimization algorithm are initialized using the processing technology history data contained in the process knowledge graph enhanced by the digital twin data.

[0030] The current processing task information is encoded as the input of the reinforcement learning model, and a multi-objective proximal strategy optimization algorithm is used to iteratively optimize the process parameters.

[0031] Combined with the expert rule constraints of process parameter selection and process path planning in the knowledge graph, the solution set of processing technology solutions for the current processing task is determined.

[0032] Optionally, the reward function is specifically:

[0033] ].

[0034] in, The weighting factor of material removal rate, is the weight coefficient of surface roughness, , is the material removal rate, is the surface roughness, R is the reward function, R The values ​​of are expressed in vector form.

[0035] Optionally, the weighted adaptive solution process for tool wear status is:

[0036] Collect multi-source sensor data during the CNC machine tool processing process, record the tool wear after each processing, and draw a tool wear curve; the multi-source sensor data includes acoustic emission, current, vibration and force.

[0037] According to the tool wear curve, the tool wear is divided into multiple tool wear states.

[0038] Construct a bidirectional long short-term memory network integrating a deep residual mechanism.

[0039] Based on a bidirectional long short-term memory network integrated with a deep residual mechanism, a mapping relationship between multi-source sensor data and tool wear status is established, and each category of tool wear status is quantified; the mapping relationship is used to classify tool wear status in the actual service process of CNC machine tools.

[0040] Optionally, a multi-objective proximal strategy optimization algorithm is used to iteratively optimize process parameters, including:

[0041] Constructing state space based on the definition of knowledge graph S and action space A The state space S Including the current machine tool operating parameters; the operating parameters include temperature, grinding force, processing speed and tool wear status; the action space A The invention comprises adjustable machining process parameters; the machining process parameters include cutting speed, feed rate and cutting depth.

[0042] In each iteration, several trajectories are sampled ;in, s t Represents the time step t The state of is the state space S An element in a t Represents the time step t The action is the action space A An element in Indicates that at time step t Reward.

[0043] Based on each optimization objective, calculate the corresponding advantage estimate and ; The optimization objectives are material removal rate and surface roughness, is the advantage function of material removal rate, is the dominant function of surface roughness.

[0044] According to minimizing the loss function , update the policy network ;in, is the weighted advantage function, is the probability ratio of the new and old strategies, is the clipping range parameter; E Indicates expected value.

[0045] Update the value network and adjust the weights and ,generate Pareto Cutting-edge solutions.

[0046] Combined with the process parameter selection, process path planning expert rules and domain knowledge in the knowledge graph, the generated Pareto The frontier solutions are sorted and the process parameters after iteration are determined.

[0047] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0048] The present application provides a twin decision method for machine tool processing technology based on large language model knowledge enhancement, the method comprising: constructing a processing technology knowledge graph of a machine tool based on the processing technology historical data and using a large language model; the processing technology knowledge graph is used to provide a processing technology solution for historical processing tasks; using digital twin data, the processing technology knowledge graph is updated to obtain a process knowledge graph enhanced by digital twin data; according to the current processing task of the machine tool, based on the process knowledge graph enhanced by digital twin data, the processing technology solution set under the current processing task is determined; the processing technology solution set is a solution set obtained by taking expert rules as constraints and material removal rate and surface roughness as optimization targets; according to the acquired tool wear state, the weight value of the optimization target is adjusted to determine the optimal weight value of the optimization target; according to the optimal weight value, based on the processing technology solution set, the optimal processing technology solution under the current processing task is determined. The present application constructs and utilizes multimodal digital twin data to enhance the process knowledge graph, and adaptively determines the optimal process solution in combination with the real-time tool wear state, thereby ensuring the processing quality of the machine tool. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 A flow chart of a twin decision method for machine tool processing technology based on large language model knowledge enhancement provided in one embodiment of the present application.

[0051] Figure 2 A twin enhanced knowledge graph provided for one embodiment of the present application.

[0052] Figure 3 A structural diagram of a hybrid model based on a deep reinforcement learning method and a knowledge graph provided in one embodiment of the present application. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0054] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0055] Embodiment 1

[0056] like Figure 1 As shown, a twin decision method for machine tool processing technology based on large language model knowledge enhancement is characterized in that the twin decision method for machine tool processing technology includes:

[0057] Step 101: Based on the historical processing technology data, a large language model is used to construct a processing technology knowledge graph of the machine tool; the processing technology knowledge graph is used to provide processing technology solutions for historical processing tasks.

[0058] Step 102: Use the digital twin data to update the processing technology knowledge graph to obtain a processing technology knowledge graph enhanced by the digital twin data.

[0059] Step 103: According to the current processing task of the machine tool, based on the process knowledge graph enhanced by the digital twin data, determine the processing technology solution set under the current processing task; the processing technology solution set is a solution set obtained by taking expert rules as constraints and material removal rate and surface roughness as optimization targets.

[0060] Step 104: According to the acquired real-time tool wear status of the machine tool, the weight value of the optimization target is dynamically adjusted to determine the optimal weight value of the optimization target under the current processing task.

[0061] Step 105: According to the optimal weight value and based on the processing solution set, determine the optimal processing solution for the current processing task.

[0062] Among them, the processing technology history data is the process specification of a typical workpiece. The process specification is determined by text, charts and other carriers to guide product processing and workers' operation. It is the basic basis for enterprises to plan, organize and control production, and is an important guarantee for enterprises to ensure product quality and improve labor productivity. In machinery manufacturing enterprises, there are three main forms of process specifications: (1) Process card (or process route card). It is compiled according to parts, and stipulates the process route, process name, equipment and process equipment used for each part in the manufacturing process. It is a comprehensive document that guides the processing of parts. (2) Process card. It is compiled according to the workshop (process stage) of parts, and stipulates the various processes that parts must go through in a workshop (process stage) as well as the equipment, process equipment and processing specifications used in each process. It is the basis for each workshop to prepare for work and organize production. (3) Process card. It is compiled according to each process of the parts, and specifies in detail the operating methods, technical requirements and precautions of each process, and is accompanied by a processing sketch. It is a process document used to specifically guide workers' operations; the process card includes processing steps and process requirements (such as processing sequence, dimensional tolerance, surface treatment, etc.), machine tool operating parameters (such as spindle speed, feed speed, cutting depth), process monitoring data (such as temperature, grinding force, vibration, machine tool load), processing result data (such as surface roughness, material removal rate), tool status (such as tool wear, chip morphology), etc.

[0063] Wherein, when executing step 101, if Figure 2 As shown, specifically including:

[0064] Step 111: pre-processing the processing technology history data.

[0065] Step 112: By fine-tuning the large language model, named entity recognition and relationship extraction are performed on the preprocessed processing technology history data to obtain the key entities of the machine tool processing technology and the relationship between the key entities; the key entities are represented by workpiece information, tool information, process parameters and processing accuracy requirements; the relationship includes the adaptation relationship between process parameters and workpiece information, the dependency relationship between tool information and process parameters, and the influence relationship between process parameters and processing accuracy requirements.

[0066] Step 113: Based on the key entities and the relationships between the key entities, a processing technology knowledge framework that reflects the association between process flow and process parameters is constructed.

[0067] Step 114: Based on the processing technology historical data, data preprocessing and natural language processing analysis are used to store the key information of the processing technology historical data in the processing technology knowledge framework to obtain a processing technology knowledge graph.

[0068] Among them, the processing technology historical data is preprocessed, specifically:

[0069] Data extraction and classification: Optical character recognition ( OCR ) technology or structured data analysis methods to extract information from the processing procedures and preliminarily classify them into the following categories: a ) Workpiece information: material, size, shape, etc.; b ) Processing parameters: cutting speed, feed rate, cutting depth, etc.; c ) Tool information: tool type, material, geometric parameters, etc.; d )Quality requirements: surface roughness, dimensional tolerance, etc.; e )Process information: processing sequence, process route, etc.

[0070] Data standardization: Standardize the extracted data, including: a )Unit unification: Convert data in different unit systems (such as metric and imperial) to a unified standard. b )Terminology unification: standardize different expressions of the same concept, such as "speed" and "spindle speed" into one expression.

[0071] Missing value handling: a ) The missing key parameters (such as main cutting parameters) can be supplemented by referring to similar process procedures or consulting experts. b ) For missing non-critical parameters, statistical methods (median filling) or marking as "unknown" can be used.

[0072] Outlier detection and processing: a ) Use domain knowledge to set reasonable parameter ranges and detect outliers that exceed the range. b ) The detected outliers are corrected or marked as pending confirmation based on expert knowledge.

[0073] Specifically, in some implementations of this embodiment, the large language model constructs and / or completes the processing technology knowledge graph as follows:

[0074] 1) Multimodal processing technology knowledge fusion, the specific steps are as follows:

[0075] Image processing: Using improved ResNet 50- FPN The model processes the thermal image of the grinding process, identifies the temperature distribution on the workpiece surface, and locates the overheated area. At the same time, it analyzes the grinding wheel surface image and detects the wear state of the grinding wheel.

[0076] Text processing: using BERT - BiLSTM - CRFThe model extracts key parameters from the grinding process documentation, such as wheel speed, table feed rate, depth of cut, etc. Identify workpiece material properties, such as hardness ( HRC 60-62) Surface roughness requirements ( Ra 0.4 μm ), etc. Extraction processing steps, such as rough grinding, fine grinding, spark ejection, etc.

[0077] Numerical Analysis: Applications LSTM The model analyzes sensor data such as grinding force and vibration to identify abnormal machining conditions.

[0078] Through the multi-head attention mechanism and improved Transformer The architecture is used for cross-modal information fusion: linear projection is performed on thermal image features, grinding wheel image features, text embeddings, and sensor data features. The self-attention mechanism is used to calculate the correlation within each modality, such as the relationship between temperature distribution and grinding wheel wear. The cross-attention mechanism is applied to capture the mutual influence between different modalities, such as the relationship between grinding parameters and chip morphology. Multimodal information is integrated through residual connections and layer normalization. Finally, a unified knowledge representation is generated through a feedforward network to form a complete grinding process knowledge graph node and relationship.

[0079] 2) Dynamic knowledge reasoning, the specific steps are as follows:

[0080] Based on graph neural network ( GNN )Build a knowledge reasoning framework and design reasoning tasks prompt Template, such as "Grinding wheel linear speed is given as 30 m / s , the workpiece hardness is HRC 61, infer the appropriate range of workbench feed speed". Combined with the semantic understanding ability of the large language model, it is inferred that "the appropriate range of workbench feed speed is 18-22 mm / min ". Use based VAE The generative model generates new hypotheses, such as "increasing the coolant flow rate may further improve the surface roughness", based on the known "grinding wheel linear speed-table feed speed-surface roughness" relationship.

[0081] 3) Adaptive update of knowledge graph. The specific steps are as follows:

[0082] Use rules to check logical consistency in the knowledge graph, such as "the cutting depth in the fine grinding stage should not be greater than that in the rough grinding stage." Design a conflict resolution algorithm based on confidence propagation. When a conflict is detected, such as the description of the optimal grinding wheel linear speed from two sources is inconsistent, automatically adjust the weights based on the importance of each node and the credibility of the relationship. Use graph pruning technology to delete low-confidence or outdated knowledge nodes. For example, delete the relevant parameter nodes of the old model grinding wheel. Dynamically merge similar nodes, such as merging "coolant flow" and "coolant pressure" into a "cooling effect" node.

[0083] 4) Semantic enhanced retrieval, the specific steps are as follows:

[0084] Use pre-trained BERT The model encodes user queries. For example, "How to improve the surface quality of workpieces" is encoded as a semantic vector. A graph neural network based on an attention mechanism is used to calculate the relevance of the query and the knowledge graph nodes. The most relevant subgraph is retrieved, including nodes such as "surface roughness", "grinding wheel selection", and "grinding parameters". A templated approach is used to generate natural language answers, such as "To improve the surface quality of workpieces, it is recommended to use a fine-grained grinding wheel and increase the grinding wheel linear speed to 32 m / s , reduce the feed speed to 18 mm / min and ensure adequate coolant supply.

[0085] Among them, in some implementations of this embodiment, when executing step 102, if Figure 2 As shown, it may specifically include:

[0086] Step 201: perform data preprocessing on the digital twin data; the digital twin data is multimodal digital twin data; the multimodal digital twin data includes text data, time series sensor data and point cloud data.

[0087] Step 202: Use a data encoding model to encode the multimodal digital twin data to obtain semantic information of the multimodal digital twin data; the data encoding model includes a Transformer Bidirectional encoding model, convolutional neural network-long short-term memory model, point cloud classification and segmentation model and autoencoder based on Transformer The bidirectional encoding model is used for text data processing; the convolutional neural network-long short-term memory model is used for time series sensor data processing; the point cloud classification and segmentation model is used for point cloud data processing; the autoencoder is used for low-dimensional representation learning of multimodal digital twin data.

[0088] Step 203: Based on the semantic information, a large language model is used to parse and integrate the multimodal digital twin data to obtain new key entities and the relationships between the key entities.

[0089] Step 204: Based on the new key entities and the relationships between the key entities, the processing technology knowledge graph is dynamically updated to generate a process knowledge graph enhanced by the digital twin data.

[0090] Specifically, data preprocessing includes:

[0091] a )Structured database standardization: SQL Standardize the historical machining parameters in the database, such as converting cutting speeds in different units into m / s .

[0092] b ) Temperature sensor data processing: Convert the workpiece surface temperature sensor data into a time series format, where each data point contains a timestamp and temperature value ( °C ).

[0093] c ) Infrared 2D data processing: Gaussian filtering is used to reduce noise on the infrared images collected by the thermal imager. The adaptive threshold method is used to extract hot spot features and identify high temperature areas.

[0094] d ) Product quality evaluation data processing: Segment the text-based quality inspection report and use named entity recognition ( NER ) technology to extract key entities such as "surface roughness", "dimensional accuracy", etc.

[0095] e ) Point cloud data processing: Reduce the dimension of the workpiece surface morphology point cloud data. Use principal component analysis ( PCA ) method to extract the main features.

[0096] f ) Vernier caliper measurement data standardization: multiple measurements of the size data z-score standardization.

[0097] g ) Structuring of CNC system parameter data: converting unstructured CNC The system log is converted into a structured format. Key parameters such as spindle speed, feed speed, etc. are extracted to form key-value pairs.

[0098] Specifically, data encoding includes:

[0099] a )Text data processing: Use pre-trained BERT - base - uncased The model processes quality assessment reports and operator notes to generate text embedding vectors.

[0100] b )Time Series Data Processing: Design CNN - LSTM The hybrid model processes temperature and grinding force data. CNN Some are set to 3 convolutional layers to extract local features. LSTM Section: 2 floors LSTM , 128 units per layer to capture long-term dependencies. Output 300-dimensional feature vector.

[0101] c ) Point cloud data processing: Using improved PointNet ++Model processes workpiece surface morphology data.

[0102] d )Structured data processing: Use stacked autoencoders to reduce the dimension and extract features of the standardized processing parameters.

[0103] Specifically, the digital twin data includes processing materials, tool types, processing parameters and processing results; the processing results include material removal rate and surface roughness.

[0104] Specifically, semantic mapping includes:

[0105] Designed specifically prompt The template guides the large language model to understand the semantics of different types of data, including the following steps:

[0106] a )Temperature data mapping:

[0107] prompt Template: "The workpiece surface temperature data[ data ] is mapped to the 'heat effect' node in the knowledge graph. The temperature exceeds 350 °C , it is marked as 'heat damage risk' state".

[0108] b )Quality evaluation data mapping:

[0109] prompt Template: "Based on quality inspection report[ data ], update the 'Product Quality' node. If there is a description of 'Surface Scratches', add a 'Surface Defects' subnode to the map.

[0110] c ) Point cloud data mapping:

[0111] prompt Template: "Use the surface morphology of the workpiece[ data ], update the 'Surface Integrity' node. When the waviness exceeds 2 μm, associated to the 'Vibration Problem' node".

[0112] A multi-head attention mechanism is used to identify the correlation between different data sources. For example, temperature data, grinding force data, and surface quality data are linked to identify their mutual influence.

[0113] Specifically, the knowledge graph expansion includes the following steps:

[0114] a ) Template-based triple generation:

[0115] Design template: "(Entity 1, Relationship, Entity 2)".

[0116] For example, from temperature data a triple is generated: ("thermal effect", impact, "surface quality").

[0117] b ) Generate new relations and attributes using large language models:

[0118] prompt Template: "Based on [ entity 1] and [ entity 2], and infer the possible existence of new relationships".

[0119] For example, by inputting "grinding wheel wear" and "surface roughness", the model may infer a new related entity "machining accuracy".

[0120] c )Incremental learning strategy:

[0121] Set the knowledge graph update threshold, for example, when the cosine similarity between new data and existing knowledge is less than 0.8, trigger the update. Use the sliding window method to retain the most recent 1,000 processed data and dynamically update the graph.

[0122] In some implementations of this embodiment, before executing step 103, the current processing task of the machine tool is also determined, and the process is as follows:

[0123] Step 211: Use the large language model to semantically encode the processing parameters, workpiece information, and tool information of the current machine tool.

[0124] Step 212: Using a semantic enhancement retrieval method, by utilizing semantic similarity calculation and matching technology, retrieve nodes that are close to the current semantic encoding in the process knowledge graph enhanced by the digital twin data.

[0125] Step 213: Using a large language model, based on the retrieved nodes close to the current semantic code, a process description of the current machining task of the machine tool is generated; the process description is expressed based on the spindle speed, feed speed, cutting depth and machining stage.

[0126] Specifically, 1) Semantic encoding: Use the pre-trained large language model to semantically encode the current grinding machine processing data. Use dedicated models to process different types of data (temperature sensor data, infrared two-dimensional data, grinding force data, and processing parameters). Image data uses image processing models, text data uses natural language processing models, and numerical data uses numerical analysis models. All processed data are input into the large language model for semantic encoding to generate a unified semantic representation. 2) Semantic enhancement retrieval: Use the semantic enhancement retrieval method to retrieve nodes in the twin knowledge graph that are semantically similar to the current processing data. The large language model is used to calculate the similarity between the semantic representation of the current data and each node in the knowledge graph, and the most relevant knowledge nodes are selected. 3) Process reasoning: Use the reasoning ability of the large language model to generate a process description based on the retrieved knowledge graph substructure. For example, if the current grinding force fluctuates greatly and the surface temperature distribution of the workpiece is abnormal, the large language model can infer and generate a description of the "grinding wheel wear process". 4) Multi-head self-attention mechanism fusion: The multi-head self-attention mechanism is used to fuse the real-time data collected by different sensors with the historical data in the knowledge graph. This step obtains a more accurate process description by weighted averaging the attention of data from different sources. 5) Process analysis and interpretation: Design specific prompt Templates guide the large language model to interpret the fused information. prompt A template could be “Analyze possible processes based on current grinding force and temperature data.” The large language model generates detailed process analysis results based on this template.

[0127] Wherein, when executing step 103, if Figure 3 As shown, it may specifically include:

[0128] Step 301: According to the current processing task of the machine tool, an optimization framework based on a multi-objective proximal strategy optimization algorithm is constructed; the optimization framework based on the multi-objective proximal strategy optimization algorithm is a hybrid model based on a deep reinforcement learning method and a knowledge graph; in the optimization framework, material removal rate and surface roughness are used as components of the reward function of the deep reinforcement learning method.

[0129] Step 302: Using the processing technology history data contained in the process knowledge graph enhanced by the digital twin data, the policy network and value network of deep reinforcement learning in the multi-objective proximal strategy optimization algorithm are initialized.

[0130] Step 303: Encode the current processing task information as the input of the reinforcement learning model, adopt a multi-objective proximal strategy optimization algorithm, and iteratively optimize the process parameters.

[0131] Step 304: Determine the set of processing solution for the current processing task by combining the process parameter selection and the expert rule constraints of process path planning in the knowledge graph.

[0132] Specifically, according to the current processing tasks and process knowledge graph of the machine tool, a hybrid model is constructed to combine deep reinforcement learning, knowledge graph and multi-objective optimization methods to form a multi-objective proximal strategy optimization algorithm ( Proximal Policy Optimization ). The framework uses the knowledge graph to initialize the state space and action space, transforms the multi-objective optimization problem into a reward function for reinforcement learning, and uses the rules provided by the knowledge graph to constrain the exploration process of reinforcement learning. The multi-objective optimization framework converts the material removal rate ( MRR ) and surface roughness ( Ra ) as a component of the reward function.

[0133] Reward Function R Defined as a weighted combination of material removal rate and surface roughness:

[0134] .

[0135] In the formula, The weighting factor of material removal rate, is the weight coefficient of surface roughness, , is the material removal rate, is the surface roughness, R is the reward function, R The values ​​of are expressed in vector form.

[0136] Initialize the historical processing data in the process knowledge graph enhanced by digital twin data PPO The policy network of . The formula of the policy network is:

[0137] .

[0138] in, π is the policy function, which indicates the selection of actions in state s a The probability distribution of f is the neural network function, is the network weight, is the current state, ais the action (i.e., process parameter adjustment). A task-based state representation method is designed to encode the current processing task information as the input of the reinforcement learning model, where the state representation formula is:

[0139] s t =[Working condition description, processing parameters, sensor data].

[0140] When executing step 303, the specific steps may be as follows:

[0141] Constructing state space based on the definition of knowledge graph S and action space A The state space S Including the current machine tool operating parameters; the operating parameters include temperature, grinding force, processing speed and tool wear status; the action space A The invention comprises adjustable machining process parameters; the machining process parameters include cutting speed, feed rate and cutting depth.

[0142] In each iteration, several trajectories are sampled ;in, s t Represents the time step t The state of is the state space S An element in a t Represents the time step t The action is the action space A An element in Indicates that at time step t Reward.

[0143] Based on each optimization objective, calculate the corresponding advantage estimate and ; The optimization objectives are material removal rate and surface roughness, is the advantage function of material removal rate, is the dominant function of surface roughness.

[0144] According to minimizing the loss function , update the policy network ;in, is the weighted advantage function, is the probability ratio of the new and old strategies, is the clipping range parameter; E Indicates expected value.

[0145] Update the value network and adjust the weights and ,generate Pareto Cutting-edge solutions.

[0146] Combined with the process parameter selection, process path planning expert rules and domain knowledge in the knowledge graph, the generated Pareto The frontier solutions are sorted and the process parameters after iteration are determined.

[0147] Specifically, when executing step 304, based on the output of the hybrid model, generate Pareto The optimal solution set shows different trade-offs between material removal rate and surface roughness. Pareto The solutions are screened and sorted to determine the final processing solution set for the current processing task.

[0148] Use the process specifications and safety requirements in the knowledge graph to generate corresponding constraints. For example, "the grinding wheel linear speed shall not exceed 40 m / s , cutting depth does not exceed 0.02 mm ". Make sure the optimization results meet these constraints.

[0149] Wherein, when executing step 104, the process of acquiring the tool wear state is:

[0150] Step 401: Collect multi-source sensor data during the CNC machine tool processing process, record the tool wear after each processing, and draw a tool wear curve; the multi-source sensor data includes acoustic emission, current, vibration and force.

[0151] Step 402: According to the tool wear curve, the tool wear is divided into a plurality of tool wear states.

[0152] Step 403: Construct a bidirectional long short-term memory network integrating a deep residual mechanism.

[0153] Step 404: Based on a bidirectional long short-term memory network fused with a deep residual mechanism, a mapping relationship between multi-source sensor data and tool wear status is established, and each category of tool wear status is quantified; the mapping relationship is used for the classification of tool wear status in the actual service process of CNC machine tools.

[0154] When executing step 104, the optimal processing scheme may be determined as follows:

[0155] According to the wear state of the tool, the weight value of each optimization target is set, and based on this, the most optimized process combination under the current processing task is derived. In order to achieve the dynamic adjustment of this optimized process combination, the OPC UA The protocol controls the machining process of CNC machine tools in real time.

[0156] As tool wear continued to increase, we took corresponding measures, namely gradually increasing the weight of surface roughness in the optimization target. This adjustment is aimed at ensuring that the machining quality of the workpiece is always maintained at a high standard.

[0157] In summary, this application has the following technical effects:

[0158] 1. Through the fusion association and deduplication of the large language model, the effective integration of different machine tool processing technology data types is achieved. The comprehensiveness and accuracy of the knowledge graph are ensured, and the processing technology knowledge graph is continuously optimized and improved through the real-time updated digital twin multi-source data.

[0159] 2. By combining the large language model with the constructed twin knowledge graph, the task can be accurately determined based on the current digital twin data, and the corresponding task solution set can be given. With material removal rate and surface roughness as the optimization targets, the optimal combination of CNC machine tool process parameters is determined, thereby improving the processing quality.

[0160] 3. After optimizing the processing parameters on the virtual debugging platform, the optimized parameters are fed back to the actual machine tool CNC system in real time to form a closed-loop feedback mechanism. Through the digital twin perception and quantification of wear status, the present invention can realize the adaptive optimization of CNC machine tool processing parameters, thereby improving processing quality and reducing production costs.

[0161] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0162] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A twin decision method for machine tool processing technology based on large language model knowledge enhancement, characterized in that: The twin decision method for machine tool processing technology includes: Based on the historical processing technology data, a large language model is used to construct a processing technology knowledge graph of the machine tool; the processing technology knowledge graph is used to provide processing technology solutions for historical processing tasks; Using the digital twin data, the processing technology knowledge graph is updated to obtain a process knowledge graph enhanced by the digital twin data; According to the current processing task of the machine tool, based on the process knowledge graph enhanced by the digital twin data, determine the processing solution set under the current processing task; the processing solution set is a solution set obtained by taking expert rules as constraints and material removal rate and surface roughness as optimization targets; According to the acquired real-time tool wear status of the machine tool, the weight value of the optimization target is dynamically adjusted to determine the optimal weight value of the optimization target under the current processing task; According to the optimal weight value, based on the processing solution set, determine the optimal processing solution for the current processing task; Among them, based on the historical data of processing technology, the large language model is used to build the processing technology knowledge graph of machine tools, including: Preprocess the processing technology history data; By fine-tuning the large language model, named entity recognition and relationship extraction are performed on the pre-processed processing technology history data to obtain the key entities of the machine tool processing technology and the relationship between the key entities; the key entities are represented by workpiece information, tool information, process parameters and processing accuracy requirements; the relationships include the adaptation relationship between process parameters and workpiece information, the dependency relationship between tool information and process parameters, and the influence relationship between process parameters and processing accuracy requirements; Based on the key entities and the relationships between the key entities, a processing technology knowledge framework reflecting the association between the process flow and the process parameters is constructed; Based on the processing technology historical data, data preprocessing and natural language processing analysis methods are used to store the key information of the processing technology historical data in the processing technology knowledge framework to obtain a processing technology knowledge graph; The digital twin data is used to update the processing technology knowledge graph to obtain the process knowledge graph enhanced by the digital twin data, which specifically includes: Performing data preprocessing on the digital twin data; the digital twin data is multimodal digital twin data; the multimodal digital twin data includes text data, time series sensor data and point cloud data; The multimodal digital twin data is encoded using a data encoding model to obtain semantic information of the multimodal digital twin data; the data encoding model includes Transformer Bidirectional encoding model, convolutional neural network-long short-term memory model, point cloud classification and segmentation model and autoencoder based on Transformer The bidirectional encoding model is used for text data processing; the convolutional neural network-long short-term memory model is used for time series sensor data processing; the point cloud classification and segmentation model is used for point cloud data processing; the autoencoder is used for low-dimensional representation learning of multimodal digital twin data; According to the semantic information, a large language model is used to parse and integrate the multimodal digital twin data to obtain new key entities and the relationships between the key entities; Based on the new key entities and the relationships between the key entities, the processing technology knowledge graph is dynamically updated to generate a process knowledge graph enhanced by the digital twin data; Among them, according to the current processing task of the machine tool, based on the process knowledge graph enhanced by the digital twin data, the processing solution set under the current processing task is determined, including: According to the current processing task of the machine tool, an optimization framework based on a multi-objective proximal strategy optimization algorithm is constructed; the optimization framework based on the multi-objective proximal strategy optimization algorithm is a hybrid model based on a deep reinforcement learning method and a knowledge graph; in the optimization framework, material removal rate and surface roughness are used as components of a reward function of the deep reinforcement learning method; Initialize the strategy network and value network of deep reinforcement learning in the multi-objective proximal strategy optimization algorithm using the processing technology history data contained in the process knowledge graph enhanced by the digital twin data; The current processing task information is encoded as the input of the reinforcement learning model, and the multi-objective proximal strategy optimization algorithm is used to iteratively optimize the process parameters; Combined with the expert rule constraints of process parameter selection and process path planning in the knowledge graph, the solution set of the processing technology under the current processing task is determined; Among them, a multi-objective proximal strategy optimization algorithm is used to iteratively optimize process parameters, including: Constructing state space based on the definition of knowledge graph S and action space A The state space S Including the current machine tool operating parameters; the operating parameters include temperature, grinding force, processing speed and tool wear status; the action space A Including adjustable machining process parameters; the machining process parameters include cutting speed, feed rate and cutting depth; In each iteration, several trajectories are sampled ;in, Represents the time step t The action is the action space A An element in Indicates that at time step t Rewards; Based on each optimization objective, calculate the corresponding advantage estimate and ; The optimization objectives are material removal rate and surface roughness, is the advantage function of material removal rate, is the dominant function of surface roughness; According to minimizing the loss function , update the policy network ;in, is the weighted advantage function, is the probability ratio of the new and old strategies, is the clipping range parameter; E Indicates expected value; Update the value network and adjust the weights and ,generate Pareto Frontier solutions; Combined with the process parameter selection, process path planning expert rules and domain knowledge in the knowledge graph, the generated Pareto The frontier solutions are sorted and the process parameters after iteration are determined.

2. According to claim 1, a twin decision method for machine tool processing technology based on large language model knowledge enhancement is characterized in that: The processing technology history data is the process specification of a typical workpiece; the process specification includes a process card, a process card and a procedure card.

3. According to claim 1, a twin decision method for machine tool processing technology based on large language model knowledge enhancement is characterized in that: The process of determining the current processing task of the machine tool is as follows: Use the large language model to semantically encode the current machine tool's processing parameters, workpiece information, and tool information; The semantic enhancement retrieval method is adopted to retrieve nodes similar to the current semantic coding in the process knowledge graph enhanced by digital twin data by using semantic similarity calculation and matching technology; Using a large language model, the process description of the current machining task of the machine tool is generated based on the retrieved nodes that are close to the current semantic encoding; The process description is expressed based on spindle speed, feed rate, and cutting depth.

4. According to claim 1, a twin decision method for machine tool processing technology based on large language model knowledge enhancement is characterized in that: The reward function is specifically: ]； in, The weighting factor of material removal rate, is the weight coefficient of surface roughness, + =1, is the material removal rate, is the surface roughness, R is the reward function, R The values ​​of are expressed in vector form.

5. According to claim 4, a twin decision method for machine tool processing technology based on large language model knowledge enhancement is characterized in that: The weighted adaptive solution process of tool wear state is: Collect multi-source sensor data during the CNC machine tool processing process, record the tool wear after each processing, and draw a tool wear curve; the multi-source sensor data includes acoustic emission, current, vibration and force; According to the tool wear curve, the tool wear is divided into multiple tool wear states; Construct a bidirectional long short-term memory network integrating a deep residual mechanism; Based on a bidirectional long short-term memory network integrated with a deep residual mechanism, a mapping relationship between multi-source sensor data and tool wear status is established, and each category of tool wear status is quantified; the mapping relationship is used to classify tool wear status in the actual service process of CNC machine tools.

Images