A data-driven forging process autonomous decision method

Abstract

The application discloses a data-driven forging process autonomous decision-making method, comprising the following steps: (1) collecting process data and forging quality inspection data in a forging production process, and establishing a forging process database; (2) mining frequent item sets from the forging process database, and abstracting nodes, node attributes and ternary relationships; (3) constructing a forging process knowledge graph based on the node relationships, and completing the forging process knowledge graph through knowledge reasoning; and (4) constructing and training a forging process reinforcement learning decision-making network based on the forging process knowledge graph. The method provided by the application can not only directly mine the correlation between the forging process and the performance of forgings from the forging process data and establish a forging process knowledge graph, but also obtain a reinforcement learning decision-making network of action strategies at different stages of the forging process, so that precise autonomous decision-making of the forging process is realized, and the intelligent level of the forging process is significantly improved.

Description

Technical Field

[0001] This invention belongs to the fields of intelligent manufacturing and material plastic processing technology, and specifically relates to a data-driven autonomous decision-making method for forging processes. Background Technology

[0002] The intelligent development of forging technology plays a crucial role in the development of my country's strategic emerging industries. Vigorously developing new intelligent forging technologies and equipment, and striving to transform from a labor-intensive, high-energy-consuming, and high-polluting traditional industry into a knowledge-intensive, low-energy-consuming, and environmentally friendly new industry, is the major environment and development trend facing my country's forging industry. However, the forging process is extremely complex, requiring a high degree of coupling between controllable variables, which has led to slow progress in the research and development of technologies related to autonomous decision-making in the forging process, thus hindering the rapid development of intelligent forging.

[0003] With the rise of artificial intelligence, data-driven and reinforcement learning methods are increasingly being used for decision-making in various processes due to their strong adaptability. Data-driven approaches can directly extract implicit features and correlations from data, avoiding traditional regression modeling methods; reinforcement learning can continuously learn from experience data to adjust the next action and obtain higher action value. Therefore, adopting data-driven and reinforcement learning in the forging process can continuously improve decision-making efficiency and product quality.

[0004] The patent specification with announcement number CN114254915A discloses a method for decision-making and optimization of the quality qualification status of shaft parts throughout the entire process. This method is based on the surrogate model function of key processing steps, and constructs and trains a hidden Markov model of key features of the qualification status of the parts. However, this method only considers three key geometric features: wall thickness difference, straightness and runout, and therefore is only applicable to the processing of slender shaft parts for aero-engines.

[0005] The patent specification with announcement number CN115564029A discloses a highly consistent human-machine hybrid decision-making method based on hybrid augmented intelligence. This method constructs a human-machine hybrid augmented reasoning decision-making model based on integrated multimodal "human-traffic" data, forming an online decision-making knowledge base for human-machine co-driving. However, this method first establishes the decision-making model and then establishes the knowledge base, and finally only outputs the driving rights subdivision rules, which cannot make decisions on complex forging processes.

[0006] The patent specification with announcement number CN115600497A discloses a knowledge-driven method for generating the main forging process flow. This method constructs a forging process knowledge base by mining process knowledge in forging examples, obtains forging feature step information based on retrieval and matching, and connects the process chain to form the main forging process flow. However, this method mainly obtains key forging process features through clustering and needs to calculate the proximity relationship between the process to be determined and each process example. It cannot adjust the process in time in the face of real-time changes in the forging process, and its scope of application is narrow.

[0007] The patent specification with announcement number CN115983387A discloses a process decision-making method for processing deformation control driven by a combination of data and causal knowledge. This method expresses the causal relationship between process variables, processing state, and processing deformation amount through an adjacency matrix and realizes causal intervention based on mask operation. However, this method only considers the processing geometric deformation control process and does not consider control variables such as part temperature and microstructure. The causal reasoning method is simple and cannot be applied to forging process decision-making. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a data-driven autonomous decision-making method for forging processes, comprising the following steps:

[0009] Step 1: Collect process data and forging quality inspection data during the forging production process, and establish a forging process database;

[0010] Step 2: Mine frequent itemsets from the forging process database, and abstract the key entity nodes, node attributes, and ternary relations of the forging process. This is specifically divided into the following sub-steps:

[0011] Step 2.1: Traverse the forging process database, record frequent items with support higher than the threshold, and generate frequent itemsets that meet the item count threshold;

[0012] Step 2.2: Traverse the candidate rules in the set of most frequent itemsets and filter out the rules with confidence scores higher than the threshold;

[0013] Step 2.3: Transform all rules into ternary relations of the forging process using first-order predicate logic. <N i E k N j >, where N is i N j All are physical nodes in the forging process, E k The relationships between nodes;

[0014] Step 3: Construct a forging process knowledge graph based on the node relationships in the forging process, and complete the forging process knowledge graph through knowledge reasoning. This is specifically divided into the following sub-steps:

[0015] Step 3.1: Use similarity calculation to distinguish or merge entity nodes with the same name abstracted from data of different stages of the forging process in Step 2, and eliminate ambiguous nodes;

[0016] Step 3.2: Transform the ternary relationship of the forging process after deambiguity into a forging process knowledge ontology, and automatically construct a forging process knowledge graph;

[0017] Step 3.3: Based on graph and logic reasoning, complete the missing values ​​of graph relationships and update the forging process knowledge graph;

[0018] Step 4: Construct and train a reinforcement learning decision network for the forging process based on the forging process knowledge graph. This step consists of the following sub-steps:

[0019] Step 4.1: Randomly select forging state S, select process action A based on knowledge graph and input it into process environment. The environment queries and outputs the corresponding action reward R and next state S′ according to knowledge graph. Put the quadruple T(S,A,R,S′) into experience pool and calculate action value Q(S,A) until the number of data in experience pool reaches the set threshold.

[0020] Step 4.2: Train the action network A and the value network Q simultaneously using the data in the experience pool. Update the action network parameters θ and the value network parameters w in each training round. The update method is θ'=tθ+(1-t)θ', w'=tw+(1-t)w', where t is a user-defined hyperparameter.

[0021] Step 4.3: Use the currently trained reinforcement learning decision network to make decisions in the actual forging process, continue to accumulate forging decision experience data, and when the success rate of autonomous decision-making is lower than the threshold, return to step 4.2 to update the decision network for the next time.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. The autonomous decision-making method for the forging process provided by this invention can not only directly mine the correlation between forging equipment, forging process, and forging performance from forging production process data and establish a forging process knowledge graph, but also obtain the process strategy value function corresponding to the process actions at different forging stages through training a reinforcement learning decision network, thereby realizing accurate autonomous decision-making in the forging process, which is conducive to the development of digital and intelligent forging technology.

[0024] 2. This invention can save on production trial and error costs and has wide applicability. For different types of small-batch forgings, it can make full use of existing production process data and knowledge to quickly explore new forging processes, reduce time costs, and effectively reduce material waste in small-batch forging production. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of a data-driven autonomous decision-making method for the forging process in a preferred embodiment of the present invention;

[0027] Figure 2 This is a knowledge graph of the TC18 titanium alloy forging process constructed in a preferred embodiment of the present invention;

[0028] Figure 3 This is a comparison curve of the forging pressure change during the fourth forging process of TC18 titanium alloy bars in a preferred embodiment of the present invention, showing the difference between the autonomous decision-making method and the traditional manual operation method. Detailed Implementation

[0029] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 As shown, this invention is a data-driven autonomous decision-making method for forging processes, comprising the following steps:

[0031] Step 1: Collect process data and forging quality inspection data during the forging production process, and establish a forging process database;

[0032] Step 2: Mine frequent itemsets from the forging process database, and abstract the key entity nodes, node attributes, and ternary relations of the forging process. This is specifically divided into the following sub-steps:

[0033] Step 2.1: Traverse the forging process database, record frequent items with support higher than the threshold, and generate frequent itemsets that meet the item count threshold;

[0034] Step 2.2: Traverse the candidate rules in the set of most frequent itemsets and filter out the rules with confidence scores higher than the threshold;

[0035] Step 2.3: Transform all rules into ternary relations of the forging process using first-order predicate logic. <N i E k N j >, where N is i N j All are physical nodes in the forging process, E k The relationships between nodes;

[0036] Step 3: Construct a forging process knowledge graph based on the node relationships in the forging process, and complete the forging process knowledge graph through knowledge reasoning. This is specifically divided into the following sub-steps:

[0037] Step 3.1: Use similarity calculation to distinguish or merge entity nodes with the same name abstracted from data of different stages of the forging process in Step 2, and eliminate ambiguous nodes;

[0038] Step 3.2: Transform the ternary relationship of the forging process after deambiguity into a forging process knowledge ontology, and automatically construct a forging process knowledge graph;

[0039] Step 3.3: Based on graph and logic reasoning, complete the missing values ​​of graph relationships and update the forging process knowledge graph;

[0040] Step 4: Construct and train a reinforcement learning decision network for the forging process based on the forging process knowledge graph. This step consists of the following sub-steps:

[0041] Step 4.1: Randomly select forging state S, select process action A based on knowledge graph and input it into process environment. The environment queries and outputs the corresponding action reward R and next state S′ according to knowledge graph. Put the quadruple T(S,A,R,S′) into experience pool and calculate action value Q(S,A) until the number of data in experience pool reaches the set threshold.

[0042] Step 4.2: Train the action network A and the value network Q simultaneously using the data in the experience pool. Update the action network parameters θ and the value network parameters w in each training round. The update method is θ'=tθ+(1-t)θ', w'=tw+(1-t)w', where t is a user-defined hyperparameter.

[0043] Step 4.3: Use the currently trained reinforcement learning decision network to make decisions in the actual forging process, continue to accumulate forging decision experience data, and when the success rate of autonomous decision-making is lower than the threshold, return to step 4.2 to update the decision network for the next time.

[0044] Preferred embodiment:

[0045] In this preferred embodiment, a TC18 titanium alloy bar with an initial size of Φ700mm×1200mm is used for forging, specifically including the following steps:

[0046] Step 1: Collect equipment and process data during the forging process of TC18 titanium alloy bars and establish a database of the TC18 titanium alloy bar forging process.

[0047] Step 2: Traverse the forging process database mentioned above, record frequent items with a support greater than 0.8, generate a frequent itemset that meets the item count requirement, traverse the candidate rules in the largest frequent itemset, filter out the rules with a confidence greater than 0.9, and transform all rules into a forging process ternary relation <head node, predicate relation, tail node> through first-order predicate logic;

[0048] Step 3: Similarity calculations are used to distinguish (similarity < 0.95) or merge (similarity ≥ 0.95) identical entity nodes abstracted from data at different stages of the forging process. The unambiguous ternary relationships of the forging process are transformed into a forging process knowledge ontology, automatically constructing a forging process knowledge graph. Based on graph and logical reasoning, missing values ​​in the graph relationships are filled in, resulting in the TC18 titanium alloy bar forging process knowledge graph. Figure 2 As shown.

[0049] Step 4: Randomly select a forging state S from the database. Based on the aforementioned knowledge graph, select a process action A and input it into the process environment. The environment queries the knowledge graph and outputs the corresponding action reward R and the next state S′. Add the quadruple T(S,A,R,S′) to the experience pool and calculate the action value Q(S,A) until the number of data entries in the experience pool reaches 10,000. Simultaneously train the action network A and the value network Q. In each round of training, update the action network parameters θ and the value network parameters w. The update method is θ' = 0.7θ + 0.3θ' and w' = 0.7w + 0.3w'. Use the currently trained reinforcement learning decision network to make decisions in the actual forging process and continue to accumulate forging decision experience data. When the success rate of autonomous decision-making is lower than 0.8, continue to train and update the decision network.

[0050] Figure 3 The diagram shows a comparison of the forging pressure variation curves during the fourth forging of TC18 titanium alloy bars in this preferred embodiment, using an autonomous decision-making method versus a traditional manual operation method. It can be seen that using the autonomous decision-making method saves approximately 60 seconds of forging time and reduces the average forging pressure by 17.6%, indicating that the present invention has good practical effects.

[0051] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A data-driven autonomous decision-making method for forging processes, characterized in that... By mining correlations from forging process data and establishing a knowledge graph, and then training a reinforcement learning decision network to achieve precise autonomous decision-making in the forging process, the following steps are included: Step 1: Collect process data and forging quality inspection data during the forging production process, and establish a forging process database; Step 2: Mine frequent itemsets from the forging process database, and abstract the key entity nodes, node attributes, and ternary relations of the forging process. This is specifically divided into the following sub-steps: Step 2.1: Traverse the forging process database, record frequent items with support higher than the threshold, and generate frequent itemsets that meet the item count threshold; Step 2.2: Traverse the candidate rules in the set of most frequent itemsets and filter out the rules with confidence scores higher than the threshold; Step 2.3: Transform all rules into ternary relations of the forging process using first-order predicate logic. <N i E k N j >, where N is i N j All are physical nodes in the forging process, E k The relationships between nodes; Step 3: Construct a forging process knowledge graph based on the node relationships in the forging process, and complete the forging process knowledge graph through knowledge reasoning. This is specifically divided into the following sub-steps: Step 3.1: Use similarity calculation to distinguish or merge entity nodes with the same name abstracted from data of different stages of the forging process in Step 2, and eliminate ambiguous nodes; Step 3.2: Transform the ternary relationship of the forging process after deambiguity into a forging process knowledge ontology, and automatically construct a forging process knowledge graph; Step 3.3: Based on graph and logic reasoning, complete the missing values ​​of graph relationships and update the forging process knowledge graph; Step 4: Construct and train a reinforcement learning decision network for the forging process based on the forging process knowledge graph. This step consists of the following sub-steps: Step 4.1: Randomly select forging state S, select process action A based on knowledge graph and input it into process environment. The environment queries and outputs the corresponding action reward R and next state S′ according to knowledge graph. Put the quadruple T(S,A,R,S′) into experience pool and calculate action value Q(S,A) until the number of data in experience pool reaches the set threshold. Step 4.2: Train the action network A and the value network Q simultaneously using the data in the experience pool. Update the action network parameters θ and the value network parameters w in each training round. The update method is θ'=tθ+(1-t)θ', w'=tw+(1-t)w', where t is a user-defined hyperparameter. Step 4.3: Use the currently trained reinforcement learning decision network to make decisions in the actual forging process, continue to accumulate forging decision experience data, and when the success rate of autonomous decision-making is lower than the threshold, return to step 4.2 to update the decision network for the next time.

Images