Display screen production flow management system and method based on artificial intelligence
By constructing a process flow decision tree and a production process topology diagram, and combining the C4.5 algorithm and the random forest model, the problems of imbalance in capacity prediction between core and auxiliary components and cross-batch data interference in display production were solved, achieving precise and stable capacity management and improving production efficiency and quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 江苏锦花电子股份有限公司
- Filing Date
- 2026-02-09
- Publication Date
- 2026-05-15
AI Technical Summary
In the existing display production process management, imbalances in the production capacity forecasting of core and auxiliary components, cross-batch data interference, poor model robustness, redundant feature selection, and difficulty in accurately mapping assembly relationships lead to low production efficiency and unstable quality.
By employing AI-based process modeling, feature optimization, and batch adaptation methods, and constructing process flow decision trees and production flow topology diagrams, combined with the C4.5 algorithm and random forest model, differentiated capacity prediction of core and auxiliary components is achieved. A two-branch random forest model is constructed to improve prediction accuracy and robustness.
It achieves precise matching of production capacity between core and auxiliary components, reduces model complexity, improves the prediction accuracy and stability of production batches, avoids production capacity imbalance and waste, and improves production efficiency and quality.
Smart Images
Figure CN121684532B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of production management technology, specifically to an artificial intelligence-based display screen production process management system and method. Background Technology
[0002] As a core component of electronic terminals, the production process of displays is characterized by a wide variety of component types, complex assembly processes, and significant batch production. With the accelerated iteration of display technology, display products are developing towards higher resolution, flexibility, and integration, requiring dozens or even hundreds of component models. Moreover, the production rhythm and consumption patterns of core components (such as display panels and driver chips) and auxiliary components (such as mounting brackets and connecting cables) differ significantly, making the requirements for production capacity matching increasingly stringent.
[0003] In current display production process management, existing technologies face several pressing issues: First, capacity forecasting often relies on a single model to predict the entire process components uniformly, failing to differentiate the functional priorities of core and auxiliary components. This leads to frequent imbalances, such as capacity shortages in core components or overcapacity in auxiliary components, severely impacting production progress. Second, after production batches are divided, no batch-specific process models or sample libraries are established. Cross-batch data interference results in insufficient prediction accuracy, making it difficult to adapt to the production needs of different batch types, such as mass production, trial production, and customization. Third, training samples only cover ideal production conditions, failing to consider unforeseen circumstances such as component defects or equipment failures. This results in poor model robustness and significant prediction errors in actual production. Fourth, feature selection relies on manual experience or simple statistical methods, without combining algorithms to quantify feature importance. This leads to high model redundancy, low computational efficiency, and an inability to quickly respond to real-time production scheduling needs. Fifth, production process modeling often employs linear programming, failing to reflect the assembly relationships between components through a combination of decision trees and topology graphs, making it difficult to accurately map actual production logic.
[0004] Furthermore, in existing technologies, the application of AI models such as random forests in production capacity prediction is mostly limited to a global, unified prediction level, without customized design for the batch characteristics and node functional differences in display production. This results in insufficient model adaptability to the production process. Simultaneously, most solutions ignore the impact of component defects on production capacity consumption, and the training samples cover only a single scenario, failing to cope with the dynamic production capacity adjustment needs in complex production environments. These problems lead to frequent imbalances in component supply and demand, batch delivery delays, and wasted production costs during display production, hindering the improvement of production efficiency and product quality. Summary of the Invention
[0005] The purpose of this invention is to provide an artificial intelligence-based display screen production process management system and method to solve the problems mentioned in the background art. The technical solution of this invention can be summarized as a full-link intelligent logic of "process modeling - feature optimization - batch adaptation - integrated prediction", including:
[0006] At the process modeling level (S1-S3): Based on the sequential logic of the display assembly process flow, a process flow decision tree is constructed, so that each tree node corresponds one-to-one with the component model and carries the capacity parameters. Then, the main trunk (core component) and leaf (auxiliary component) nodes are marked through cross-flow relationships to generate a visual production process topology diagram and realize the digital mapping of the production process.
[0007] Batch adaptation level (S4-S5): Divide tasks according to production batches, build an independent stage topology map and sample set for each batch to avoid cross-batch data interference, and achieve batch-specific adaptation of the prediction model for each batch's production characteristics (such as the scale of mass production batches and the personalization of customized batches).
[0008] Feature optimization level (S6): The C4.5 algorithm is used to quantify the feature information gain ratio, select the core features most critical to capacity prediction (real-time value of production demand and consumption rate parameters), eliminate redundant information, reduce the computational complexity of the model, and at the same time ensure the strong correlation between features and prediction targets.
[0009] Ensemble prediction layer (S7-S8): Generate diverse training subsets through bootstrap resampling, construct multiple C4.5 decision tree-based learners, and integrate them into a two-branch random forest using weighted voting. The trunk branch focuses on core components, and the leaf branches focus on auxiliary components. Combined with the weight allocation of node importance, the specificity and generalization ability of the prediction are improved.
[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0011] This is an AI-based display screen production process management system, comprising: a component and task management module, a process flow decision tree construction module, a production process topology graph generation module, a production batch partitioning and sample set construction module, a feature processing module, a feature importance calculation module, a base learner construction module, and a random forest capacity prediction module; among which:
[0012] The component and task management module is used to build a display screen component type library and record the model information of each component; build a display screen component production database and record component production data; create production tasks and encapsulate them into the production database, record the quantity of a single component model, and publish the upper limit of production demand to each production end;
[0013] The process flow decision tree construction module creates a process flow decision tree based on the assembly process flow of the finished display screen, so that each tree node corresponds to a component model and carries the real-time value of the production demand and consumption rate parameters of that component.
[0014] The production process topology generation module is used to establish cross-flow relationships between various process flow decision trees, mark the main nodes and leaf nodes, and construct a complete production process topology under the display screen production task based on the two types of nodes.
[0015] The production batch division and sample set construction module is used to divide the production task into production batches, construct a complete production process topology map for each batch stage, and realize independent storage and associated calling; for the topology map of each batch, a first type of training sample set in the state without missing components and a second type of training sample set in the state with missing components are constructed.
[0016] The feature processing module is used to extract two types of core feature parameter samples from two types of training sample sets: real-time value of production demand and consumption rate parameter. The core feature parameter samples are standardized to eliminate the difference in units, and the main node comprehensive sample set and the leaf node comprehensive sample set are obtained.
[0017] The feature importance calculation module is used to construct independent C4.5 algorithm execution models for the main node and the leaf node respectively. Taking "whether the component production capacity meets the assembly requirements" as the target classification label, it quantifies the information entropy, conditional entropy, information gain and information gain ratio of the target features, and obtains the optimal feature set of the main node and the optimal feature set of the leaf node based on the information gain ratio sorting and filtering.
[0018] The base learner construction module performs bootstrap resampling on the comprehensive sample set of the main nodes and the comprehensive sample set of the leaf nodes, based on the optimal feature set of the main nodes and the optimal feature set of the leaf nodes, to generate k training subsets of the main nodes and k training subsets of the leaf nodes. For each training subset, a C4.5 decision tree model is constructed based on the corresponding optimal feature set. During the construction process, m tree nodes are randomly selected from the optimal feature set to participate in the node splitting decision.
[0019] The random forest capacity prediction module is used to integrate the C4.5 decision tree model of the trunk node and the leaf node into a two-branch random forest model by using a weighted voting method based on the importance ratio of the tree node in the production process; it receives the real-time feature data of each tree node in the current production batch after standardization, inputs them into the corresponding random forest branches, and outputs the capacity prediction results of each component.
[0020] An AI-based method for managing the production process of displays includes the following steps:
[0021] Step S1: Build a display component type library and production database, create production tasks and register component models and quantities, and publish the upper limit of production demand to the production end;
[0022] Step S2: Create a process flow decision tree based on the display assembly process flow, so that the tree nodes carry the real-time value of the production demand and consumption rate parameters of the corresponding components;
[0023] Step S3: Establish cross-flow relationships between process flow decision trees, mark the main nodes and leaf nodes, and form a complete production process topology diagram;
[0024] Step S4: Divide the production batches and construct a complete production process topology diagram for each batch. For each batch, construct two training sample sets: one with no defective components and one with defective components.
[0025] Step S5: Extract core feature parameters from the two types of training sample sets and perform standardization processing to separate the comprehensive sample set of the main node and the leaf node;
[0026] Step S6: Calculate the feature information gain ratio of the trunk node and the leaf node respectively based on the C4.5 algorithm, and select the optimal feature set;
[0027] Step S7: Perform bootstrap resampling on the comprehensive sample set, and construct multiple C4.5 decision tree base learners based on the optimal feature set. A random feature selection strategy is adopted during the construction process.
[0028] Step S8: Integrate the C4.5 decision tree model of the trunk node and the leaf node through weighted voting to form a two-branch random forest model. Input the real-time feature data of the current production batch and output the capacity prediction results of each component.
[0029] As a preferred embodiment of the present invention, the specific implementation process of step S1 includes:
[0030] Build a display component type library to record the model information of each component in the assembled finished display screen, and encapsulate the data corresponding to one component model information into one display component type;
[0031] Establish a production database for display screen components to record production data for each display screen component;
[0032] Create a display screen production task and encapsulate it into the display screen component production database;
[0033] Under the aforementioned display screen production task, the number of single component models for assembling the finished display screen is recorded;
[0034] Under the display screen production task, an upper limit value of the production demand is issued to each display screen component production end. The upper limit value of the production demand is the product of the number of finished display screens assembled and the number of single component models.
[0035] As a preferred embodiment of the present invention, the specific implementation process of step S2 includes:
[0036] Based on the assembly process flow of the finished display screen, a process flow decision tree is created. One assembly process flow corresponds to one process flow decision tree. Each assembly process flow generated according to the assembly order logic between single component models corresponds to each process flow decision tree. A tree node in the process flow decision tree corresponds to a flow node in the assembly process flow, and a flow node corresponds to a component model.
[0037] In the created process flow decision tree, each tree node carries the real-time value of the production demand for a single component model. The real-time value of the production demand is the difference between the upper limit of the production demand and the production consumption value. The production consumption value is the number of single component models that have been assembled in the assembly state.
[0038] In the created process flow decision tree, each tree node also carries the consumption rate parameter for a single component model.
[0039] As a preferred embodiment of the present invention, the specific implementation process of step S3 includes:
[0040] Establish cross-flow relationships between the created process flow decision trees. The cross-flow relationship is an assembly relationship between a tree node in one process flow decision tree and a tree node in another process flow decision tree.
[0041] Based on the cross-flow relationship, tree nodes with assembly relationships are marked as trunk nodes, and tree nodes without assembly relationships are marked as leaf nodes;
[0042] Based on the trunk nodes and leaf nodes, a complete production process topology diagram is constructed for the display screen production task.
[0043] As a preferred embodiment of the present invention, the specific implementation process of step S4 includes:
[0044] The production tasks of the display screen are divided into production batches. Each production batch corresponds to a complete production process topology diagram of a stage. The complete production process topology diagrams of the stages corresponding to the production batch are stored independently and called in association.
[0045] For each production batch, a complete production process topology diagram representing each stage is constructed, and two types of training sample sets are built, including:
[0046] The first type of training sample set under the condition of no missing components is used to record the real-time value of the output requirement and consumption rate parameter of each tree node under the condition of no missing components;
[0047] The second type of training sample set under the condition of missing components is used to record the real-time value of the output requirement and consumption rate parameter of each tree node under the condition of missing components.
[0048] As a preferred embodiment of the present invention, the specific implementation process of step S5 includes:
[0049] From the first type of training sample set and the second type of training sample set, the core feature parameter samples of the tree nodes are extracted respectively, and are denoted as tree node samples. The core feature parameter samples include the real-time value of output demand, denoted as feature F1, and the consumption rate parameter, denoted as feature F2.
[0050] The core feature parameter samples in the two training sample sets are standardized to eliminate the difference in dimensions, resulting in a standardized feature sample set, which includes the first standardized feature sample set corresponding to the first training sample set, denoted as SS1, and the second standardized feature sample set corresponding to the second training sample set, denoted as SS2.
[0051] The tree node samples marked as trunk nodes are separated from the first type of standardized feature sample set and the second type of standardized feature sample set to form a comprehensive trunk node sample set Smain.
[0052] Tree node samples marked as leaf nodes are separated from the first type of normalized feature sample set and the second type of normalized feature sample set to form a comprehensive leaf node sample set Sleaf.
[0053] As a preferred embodiment of the present invention, the specific implementation process of step S6 includes:
[0054] Independent C4.5 algorithm execution models are constructed for the trunk node and leaf node respectively;
[0055] For the comprehensive sample set of the main nodes, the target classification label is "whether the component production capacity meets the assembly requirements", including the target meets the label and the target does not meet the label. The information entropy H(D) and conditional entropy H(D|A) of each tree node sample (F1, F2) are calculated to obtain the information gain Gain(D, A) of the target feature A, where D is the total sample domain formed by the comprehensive sample set of the main nodes, and the target feature A is feature F1 or feature F2.
[0056] Based on the information gain Gain(D, A), the information gain ratio Gain_ratio(D, A) of the target feature A is obtained as Gain(D, A) / H_A(D), where H_A(D) is the feature entropy of the target feature A, and the feature entropy is an intrinsic value.
[0057] For the comprehensive sample set of leaf nodes, the same calculation logic is used to calculate the information gain ratio of features F1 and F2 with the same target classification label;
[0058] The information gain ratio is used as a quantitative indicator of the feature importance of tree nodes. The feature importance of trunk nodes and leaf nodes is sorted in descending order. The tree node samples with the top N feature importance are selected to obtain the optimal feature set Fmain for trunk nodes and the optimal feature set Fleaf for leaf nodes, thus completing the feature selection of tree nodes.
[0059] As a preferred embodiment of the present invention, the specific implementation process of step S7 includes:
[0060] Based on the optimal feature set Fmain of the backbone nodes and the optimal feature set Fleaf of the leaf nodes, bootstrap resampling is performed on the comprehensive sample set Smain of the backbone nodes and the comprehensive sample set Sleaf of the leaf nodes, respectively:
[0061] Randomly select k tree node samples from the main node comprehensive sample set Smain to form k main node training subsets;
[0062] Randomly select k tree node samples from the comprehensive sample set Sleaf of leaf nodes to form a k leaf node training subset;
[0063] Where k is the number of base learners in the random forest;
[0064] For each backbone node, a training subset is used to construct the corresponding C4.5 decision tree model based on the optimal feature set Fmain of the backbone node;
[0065] For each leaf node, a subset is trained, and the corresponding C4.5 decision tree model is constructed based on the optimal feature set Fleaf of the leaf node;
[0066] During the construction of each base learner, a random feature selection strategy is adopted to randomly select m tree node samples from the optimal feature set Fmain of the trunk node or the optimal feature set Fleaf of the leaf node to participate in the splitting decision of the base learner.
[0067] As a preferred embodiment of the present invention, the specific implementation process of step S8 includes:
[0068] Using a weighted voting method, based on the importance ratio of tree nodes in the production process, the C4.5 decision tree model constructed from the optimal feature set Fmain of the trunk node and the C4.5 decision tree model constructed from the optimal feature set Fleaf of the leaf node are integrated to form a two-branch random forest model. The trunk node branch is responsible for predicting the capacity matching degree of the core assembly components, and the leaf node branch is responsible for predicting the capacity matching degree of the auxiliary assembly components.
[0069] Input the real-time feature data of each tree node in the current production batch, after data standardization, into the corresponding random forest branch, and output the production capacity prediction results of each component.
[0070] Compared with the prior art, the beneficial effects achieved by the present invention are:
[0071] By adopting a collaborative design of "trunk / sub-leaf nodes + dual-branch random forest", the functional priority of components is deeply integrated with the ensemble learning model, which solves the technical problem of confusion between core and auxiliary component capacity prediction in the existing technology and realizes differentiated and accurate capacity prediction.
[0072] A three-in-one adaptation scheme of "production batch - topology map - sample set" is proposed. Each batch independently constructs a stage topology map and a dedicated sample set, which not only avoids cross-batch data interference, but also realizes batch-level dynamic model optimization, solves the problem of unstable prediction accuracy in multi-batch production, and helps to improve batch capacity matching compared with the existing cross-batch hybrid prediction scheme.
[0073] By constructing two training sample sets with and without defective components, and combining feature quantization and screening using the C4.5 algorithm, the robustness of the model to complex production environments is improved, enabling the model to maintain stable predictions even under sudden situations such as component failures and supply shortages. It also solves the pain point of "feature redundancy leading to computational inefficiency" in existing models. While reducing model complexity, it helps to improve prediction accuracy, avoids confusion between core and auxiliary component capacity predictions, and prevents the spread of prediction errors for core component capacity gaps and the waste of auxiliary components. Attached Figure Description
[0074] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.
[0075] Figure 1 This is a schematic diagram of the steps of the display screen production process management method based on artificial intelligence of the present invention. Detailed Implementation
[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0077] In this first embodiment: an artificial intelligence-based display screen production process management system is provided, the system including:
[0078] The component and task management module is used to build a display screen component type library and record the model information of each component; build a display screen component production database and record component production data; create production tasks and encapsulate them into the production database, record the quantity of a single component model, and publish the upper limit of production demand to each production end;
[0079] The process flow decision tree construction module creates a process flow decision tree based on the assembly process flow of the finished product display screen, so that each tree node corresponds to a component model and carries the real-time value of the production demand and consumption rate parameters of that component.
[0080] The production process topology generation module is used to establish cross-flow relationships between various process flow decision trees, mark the main nodes and leaf nodes, and construct a complete production process topology under the display screen production task based on the two types of nodes.
[0081] The production batch segmentation and sample set construction module is used to segment production tasks into production batches, construct a complete production process topology map for each batch stage, and realize independent storage and associated retrieval; for each batch's topology map, it constructs a first type of training sample set in the state without missing components and a second type of training sample set in the state with missing components.
[0082] The feature processing module is used to extract two types of core feature parameter samples from two types of training sample sets: real-time value of production demand and consumption rate parameter. The core feature parameter samples are standardized to eliminate the difference in units, and the main node comprehensive sample set and the leaf node comprehensive sample set are obtained.
[0083] The feature importance calculation module is used to build independent C4.5 algorithm execution models for the main node and the leaf node respectively. With "whether the component production capacity meets the assembly requirements" as the target classification label, it quantifies the information entropy, conditional entropy, information gain and information gain ratio of the target features. Based on the information gain ratio, it sorts and filters to obtain the optimal feature set of the main node and the optimal feature set of the leaf node.
[0084] The base learner construction module performs bootstrap resampling on the comprehensive sample sets of the main nodes and the comprehensive sample sets of the leaf nodes, based on the optimal feature sets of the main nodes and the optimal feature sets of the leaf nodes, to generate k training subsets of the main nodes and k training subsets of the leaf nodes. For each training subset, a C4.5 decision tree model is constructed based on the corresponding optimal feature set. During the construction process, m tree nodes are randomly selected from the optimal feature set to participate in the node splitting decision.
[0085] The Random Forest Capacity Prediction Module uses a weighted voting method to integrate the C4.5 decision tree model of the trunk node and the leaf node into a two-branch random forest model based on the importance ratio of the tree node in the production process. It receives the real-time feature data of each tree node in the current production batch after standardization, inputs them into the corresponding random forest branches, and outputs the capacity prediction results of each component.
[0086] Please see Figure 1 In this second embodiment, an AI-based display screen production process management method is provided to be applicable to the first embodiment. This embodiment takes the mass production of a 65-inch 4K LCD display screen from a display screen manufacturer as an example. The production task is 10,000 finished products per month, divided into 10 production batches of 1,000 units per batch, involving 28 types of components (including 6 core components: display panel, driver chip, backlight module, control circuit board, power module, touch sensor; and 22 auxiliary components: fixing bracket, connecting wires, dustproof stickers, etc.).
[0087] The method includes the following steps:
[0088] Step S1: Build a display component type library and production database, create production tasks and register component models and quantities, and publish the upper limit of production demand to the production end;
[0089] Specifically, a display component type library is constructed to record the model information of each component in the assembled finished display screen, and the data corresponding to one component model information is encapsulated into one display component type;
[0090] Establish a production database for display screen components to record production data for each display screen component;
[0091] Create a display screen production task and encapsulate it into the display screen component production database;
[0092] Under the display screen production task, record the quantity of a single component model for assembling the finished display screen;
[0093] Under the display production task, the upper limit of production demand is issued to each display component production end. The upper limit of production demand is the product of the number of finished display assembly units and the number of single component models.
[0094] For example, the finished product assembly quantity is 1000 units / batch, the quantity of a single core component (such as a display panel) is 1 unit / set, and the quantity of auxiliary components (such as connecting wires) is 4 pieces / set;
[0095] Production demand ceiling = 1000 units × 1 = 1000 (core components), 1000 units × 4 = 4000 (auxiliary components).
[0096] Step S2: Create a process flow decision tree based on the display assembly process flow, so that the tree nodes carry the real-time value of the production demand and consumption rate parameters of the corresponding components;
[0097] Specifically, based on the assembly process flow of the finished display screen, a process flow decision tree is created. One assembly process flow corresponds to one process flow decision tree. Each assembly process flow generated according to the assembly order logic between single component models corresponds to each process flow decision tree. A tree node in the process flow decision tree corresponds to a flow node in the assembly process flow, and a flow node corresponds to a component model.
[0098] In the created process flow decision tree, each tree node carries the real-time value of the production demand for a single component model. The real-time value of the production demand is the difference between the upper limit of the production demand and the production consumption value. The production consumption value is the number of single component models that have been assembled in the assembly state.
[0099] In the created process flow decision tree, each tree node also carries the consumption rate parameter for a single component model;
[0100] For example, following the process flow of "display panel → backlight module → driver chip → control circuit board → power module → touch sensor → auxiliary component assembly", a main process flow decision tree is created, with each tree node corresponding to one component type.
[0101] The tree nodes are updated in real time with the real-time production demand value (e.g., the initial display panel is 1000, and after assembling 500 units, it becomes 500) and the consumption rate parameter (core components average 0.5 units / minute, auxiliary components average 2 units / minute).
[0102] Step S3: Establish cross-flow relationships between process flow decision trees, mark the main nodes and leaf nodes, and form a complete production process topology diagram;
[0103] Specifically, cross-flow relationships are established between the created process flow decision trees. A cross-flow relationship is an assembly relationship between a tree node in one process flow decision tree and a tree node in another process flow decision tree.
[0104] Based on the cross-flow relationship, tree nodes with assembly relationships are marked as trunk nodes, and tree nodes without assembly relationships are marked as leaf nodes;
[0105] Based on the trunk nodes and leaf nodes, a complete production process topology diagram is constructed for the display screen production task.
[0106] For example, the driver chip node and the control board node have an assembly dependency (cross-flow relationship) and are marked as core backbone nodes; the fixed bracket node has no cross-tree assembly relationship and is marked as a leaf node; a batch-specific topology map containing 6 backbone nodes and 22 leaf nodes is generated to visually display the component assembly association.
[0107] Step S4: Divide the production batches and construct a complete production process topology diagram for each batch. For each batch, construct two training sample sets: one with no defective components and one with defective components.
[0108] Specifically, the production tasks of the display screen are divided into production batches. Each production batch corresponds to a complete production process topology diagram of a stage, and the complete production process topology diagram of the stage corresponding to the production batch is stored independently and called in association.
[0109] For each production batch, a complete production process topology diagram representing each stage is constructed, and two types of training sample sets are built, including:
[0110] The first type of training sample set under the condition of no missing components is used to record the real-time value of the output requirement and consumption rate parameter of each tree node under the condition of no missing components;
[0111] The second type of training sample set under the condition of missing components is used to record the real-time value of the output requirement and consumption rate parameter of each tree node under the condition of missing components;
[0112] For example, 1000 units / batch, a total of 10 batches, with the topology map of each batch stored independently; for each batch, a first type of sample set (without defects) of 1000 data points and a second type of sample set (with defects, simulating 3% component failure) of 300 data points are constructed, and nodes F1 (real-time value of production demand) and F2 (consumption rate parameter) are recorded.
[0113] Step S5: Extract core feature parameters from the two types of training sample sets and perform standardization processing to separate the comprehensive sample set of the main node and the leaf node;
[0114] Specifically, core feature parameter samples of tree nodes are extracted from the first type of training sample set and the second type of training sample set, respectively, and are denoted as tree node samples. The core feature parameter samples include the real-time value of output demand, denoted as feature F1, and the consumption rate parameter, denoted as feature F2.
[0115] The core feature parameter samples in the two training sample sets are standardized to eliminate the difference in dimensions, resulting in a standardized feature sample set, which includes the first standardized feature sample set corresponding to the first training sample set, denoted as SS1, and the second standardized feature sample set corresponding to the second training sample set, denoted as SS2.
[0116] The tree node samples marked as trunk nodes are separated from the first type of standardized feature sample set and the second type of standardized feature sample set to form a comprehensive trunk node sample set Smain.
[0117] Tree node samples marked as leaf nodes are separated from the first type of normalized feature sample set and the second type of normalized feature sample set to form a comprehensive leaf node sample set Sleaf;
[0118] For example, the Z-score standardization formula is used to standardize the data, eliminate differences in units, and classify the sample set according to node type.
[0119] Step S6: Calculate the feature information gain ratio of the trunk node and the leaf node respectively based on the C4.5 algorithm, and select the optimal feature set;
[0120] Specifically, separate C4.5 algorithm execution models are built for the trunk node and the leaf node respectively;
[0121] For the comprehensive sample set of the main nodes, the target classification label is "whether the component production capacity meets the assembly requirements", including the target meets the label and the target does not meet the label. The information entropy H(D) and conditional entropy H(D|A) of each tree node sample (F1, F2) are calculated to obtain the information gain Gain(D, A) of the target feature A, where D is the total sample domain formed by the comprehensive sample set of the main nodes, and the target feature A is feature F1 or feature F2.
[0122] Based on the information gain Gain(D, A), the information gain ratio Gain_ratio(D, A) of the target feature A is obtained as Gain(D, A) / H_A(D), where H_A(D) is the feature entropy of the target feature A, and the feature entropy is an intrinsic value.
[0123] For the comprehensive sample set of leaf nodes, the same calculation logic is used to calculate the information gain ratio of features F1 and F2 with the same target classification label;
[0124] The information gain ratio is used as a quantitative indicator of the feature importance of tree nodes. The feature importance of trunk nodes and leaf nodes is sorted in descending order. The tree node samples with the top N feature importance are selected to obtain the optimal feature set Fmain for trunk nodes and the optimal feature set Fleaf for leaf nodes, thus completing the feature selection of tree nodes.
[0125] For example, the C4.5 constraints are set as follows: minimum information gain ratio threshold of 0.1, termination condition is that the number of node samples is less than 5% of the total number of samples; information gain ratio of F1 for backbone nodes is 0.18, information gain ratio of F2 is 0.12, information gain ratio of F1 for leaf nodes is 0.15, and information gain ratio of F2 is 0.09.
[0126] Step S7: Perform bootstrap resampling on the comprehensive sample set, and construct multiple C4.5 decision tree base learners based on the optimal feature set. A random feature selection strategy is adopted during the construction process.
[0127] Specifically, based on the optimal feature set Fmain of the backbone nodes and the optimal feature set Fleaf of the leaf nodes, bootstrap resampling is performed on the comprehensive sample set Smain of the backbone nodes and the comprehensive sample set Sleaf of the leaf nodes, respectively:
[0128] Randomly select k tree node samples from the main node comprehensive sample set Smain to form k main node training subsets;
[0129] Randomly select k tree node samples from the comprehensive sample set Sleaf of leaf nodes to form a k leaf node training subset;
[0130] Where k is the number of base learners in the random forest;
[0131] For each backbone node, a training subset is used to construct the corresponding C4.5 decision tree model based on the optimal feature set Fmain of the backbone node;
[0132] For each leaf node, a subset is trained, and the corresponding C4.5 decision tree model is constructed based on the optimal feature set Fleaf of the leaf node;
[0133] During the construction of each base learner, a random feature selection strategy is adopted to randomly select m tree node samples from the optimal feature set Fmain of the trunk node or the optimal feature set Fleaf of the leaf node to participate in the splitting decision of the base learner.
[0134] For example, set k=50 (number of base learners), extract 50 training subsets from Smain and 50 training subsets from Sleaf; construct 100 C4.5 decision tree base learners (50 for trunk nodes and 50 for leaf nodes).
[0135] Step S8: Integrate the C4.5 decision tree model of the trunk node and the leaf node through weighted voting to form a two-branch random forest model. Input the real-time feature data of the current production batch and output the capacity prediction results of each component.
[0136] Specifically, a weighted voting method is adopted, based on the importance ratio of tree nodes in the production process, to integrate the C4.5 decision tree model constructed from the optimal feature set Fmain of the trunk node and the C4.5 decision tree model constructed from the optimal feature set Fleaf of the leaf node to form a two-branch random forest model. The trunk node branch is responsible for predicting the capacity matching degree of the core assembly components, and the leaf node branch is responsible for predicting the capacity matching degree of the auxiliary assembly components.
[0137] Input the real-time feature data of each tree node in the current production batch after data standardization into the corresponding random forest branch, and output the production capacity prediction results of each component.
[0138] For example, the branch weight of the trunk node can be set to 0.7, and the branch weight of the leaf node can be set to 0.3.
[0139] It is worth mentioning that the C4.5 algorithm is a classic decision tree generation algorithm. Its core goal is to construct the optimal decision tree by quantifying the contribution of features to the classification results, so as to achieve accurate classification or prediction.
[0140] In this embodiment, two training sample sets are constructed: one with "no defective components" and the other with "defective components". These sets cover normal and abnormal scenarios in display production, solving the C4.5 algorithm's dependence on "single-state samples" and making the algorithm training data more consistent with unexpected situations such as component failures and supply shortages in actual production.
[0141] Two comprehensive sample sets (Smain and Sleaf) are separated according to "main nodes (core components)" and "sub-leaf nodes (auxiliary components)" to provide "classification training data" for the C4.5 algorithm, avoid confusion between the production capacity characteristics of core and auxiliary components, and adapt to the "single node single split logic" characteristic of the C4.5 algorithm;
[0142] The real-time output demand (F1) and consumption rate parameter (F2) are standardized to eliminate the difference in units (e.g., F1 is in "pieces" and F2 is in "pieces / minute"), ensuring that the feature weights are balanced when the C4.5 algorithm calculates the information entropy, and avoiding feature selection bias caused by inconsistent units.
[0143] In addition, this embodiment extracts two core features (F1, F2) that are strongly correlated with capacity prediction, instead of blindly inputting multi-dimensional data. This adapts to the C4.5 algorithm's "feature quality takes precedence over quantity" operating logic, reducing the impact of redundant features on the algorithm's computational efficiency.
[0144] Independent C4.5 algorithm execution models are constructed for the main nodes and leaf nodes respectively. The information gain ratios of F1 and F2 are calculated separately with "whether the component capacity meets the assembly requirements" as the unified classification label. This "independent calculation for each node" method is adapted to the characteristic of the C4.5 algorithm of "single sample set and single classification target". At the same time, the importance of features is quantified by information gain ratio to ensure that the optimal feature set (Fmain, Fleaf) selected is accurately matched with the capacity pattern of the node type (core / auxiliary component).
[0145] In the C4.5 decision tree construction process, the importance ratio of tree nodes in the production process is incorporated (0.7 for trunk nodes and 0.3 for leaf nodes). This allows the algorithm to prioritize the feature contribution of core components when splitting nodes, adapting to the actual need of "prioritizing the production capacity of core components" in display production. The termination condition of the C4.5 algorithm is set as "the number of tree node samples is less than 5% of the total number of samples". Combined with the sample size of display production batches, this ensures both the splitting depth of the decision tree and avoids algorithm timeouts due to excessive sample size.
[0146] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0147] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A display screen production process management method based on artificial intelligence, characterized in that, The method includes the following steps: Step S1: Build a display component type library and production database, create production tasks and register component models and quantities, and publish the upper limit of production demand to the production end; Step S2: Create a process flow decision tree based on the display assembly process flow, so that the tree nodes carry the real-time value of the production demand and consumption rate parameters of the corresponding components; The specific implementation process of step S2 includes: Based on the assembly process flow of the finished display screen, a process flow decision tree is created. One assembly process flow corresponds to one process flow decision tree. Each assembly process flow generated according to the assembly order logic between single component models corresponds to each process flow decision tree. A tree node in the process flow decision tree corresponds to a flow node in the assembly process flow, and a flow node corresponds to a component model. Step S3: Establish cross-flow relationships between process flow decision trees, mark the main nodes and leaf nodes, and form a complete production process topology diagram; The specific implementation process of step S3 includes: Establish cross-flow relationships between the created process flow decision trees. The cross-flow relationship is an assembly relationship between a tree node in one process flow decision tree and a tree node in another process flow decision tree. Based on the cross-flow relationship, tree nodes with assembly relationships are marked as trunk nodes, and tree nodes without assembly relationships are marked as leaf nodes; Based on the trunk nodes and leaf nodes, a complete production process topology diagram is constructed for the display screen production task. Step S4: Divide the production batches and construct a complete production process topology diagram for each batch. For each batch, construct two training sample sets: one with no defective components and one with defective components. Step S5: Extract core feature parameters from the two types of training sample sets and perform standardization processing to separate the comprehensive sample set of the main node and the leaf node; Step S6: Calculate the feature information gain ratio of the trunk node and the leaf node respectively based on the C4.5 algorithm, and select the optimal feature set; The specific implementation process of step S6 includes: Independent C4.5 algorithm execution models are constructed for the trunk node and leaf node respectively; For the comprehensive sample set of the main nodes, the target classification label is "whether the component production capacity meets the assembly requirements", including the target meets the label and the target does not meet the label. The information entropy H(D) and conditional entropy H(D|A) of each tree node sample (F1, F2) are calculated to obtain the information gain Gain(D, A) of the target feature A, where D is the total sample domain formed by the comprehensive sample set of the main nodes, and the target feature A is feature F1 or feature F2. Based on the information gain Gain(D, A), the information gain ratio Gain_ratio(D, A) of the target feature A is obtained as Gain(D, A) / H_A(D), where H_A(D) is the feature entropy of the target feature A, and the feature entropy is an intrinsic value. For the comprehensive sample set of leaf nodes, the same calculation logic is used to calculate the information gain ratio of features F1 and F2 with the same target classification label; The information gain ratio is used as a quantitative indicator of the feature importance of tree nodes. The feature importance of trunk nodes and leaf nodes is sorted in descending order. The tree node samples with the top N feature importance are selected to obtain the optimal feature set Fmain for trunk nodes and the optimal feature set Fleaf for leaf nodes, thus completing the feature selection of tree nodes. Step S7: Perform bootstrap resampling on the comprehensive sample set, and construct multiple C4.5 decision tree base learners based on the optimal feature set. A random feature selection strategy is adopted during the construction process. Step S8: Integrate the C4.5 decision tree model of the trunk node and the leaf node using a weighted voting method to form a two-branch random forest model. The trunk node branch is responsible for predicting the capacity matching degree of the core assembly components, and the leaf node branch is responsible for predicting the capacity matching degree of the auxiliary assembly components. Input the real-time feature data of the current production batch and output the capacity prediction results of each component.
2. The display screen production process management method based on artificial intelligence according to claim 1, characterized in that, The specific implementation process of step S1 includes: Build a display component type library to record the model information of each component in the assembled finished display screen, and encapsulate the data corresponding to one component model information into one display component type; Establish a production database for display screen components to record production data for each display screen component; Create a display screen production task and encapsulate it into the display screen component production database; Under the aforementioned display screen production task, the number of single component models for assembling the finished display screen is recorded; Under the display screen production task, an upper limit value of the production demand is issued to each display screen component production end. The upper limit value of the production demand is the product of the number of finished display screens assembled and the number of single component models.
3. The display screen production process management method based on artificial intelligence according to claim 1, characterized in that, The specific implementation process of step S2 includes: In the created process flow decision tree, each tree node carries the real-time value of the production demand for a single component model. The real-time value of the production demand is the difference between the upper limit of the production demand and the production consumption value. The production consumption value is the number of single component models that have been assembled in the assembly state. In the created process flow decision tree, each tree node also carries the consumption rate parameter for a single component model.
4. The display screen production process management method based on artificial intelligence according to claim 1, characterized in that, The specific implementation process of step S4 includes: The production tasks of the display screen are divided into production batches. Each production batch corresponds to a complete production process topology diagram of a stage. The complete production process topology diagrams of the stages corresponding to the production batch are stored independently and called in association. For each production batch, a complete production process topology diagram representing each stage is constructed, and two types of training sample sets are built, including: The first type of training sample set under the condition of no missing components is used to record the real-time value of the output requirement and consumption rate parameter of each tree node under the condition of no missing components; The second type of training sample set under the condition of missing components is used to record the real-time value of the output requirement and consumption rate parameter of each tree node under the condition of missing components.
5. The display screen production process management method based on artificial intelligence according to claim 1, characterized in that, The specific implementation process of step S5 includes: From the first type of training sample set and the second type of training sample set, the core feature parameter samples of the tree nodes are extracted respectively, and are denoted as tree node samples. The core feature parameter samples include the real-time value of output demand, denoted as feature F1, and the consumption rate parameter, denoted as feature F2. The core feature parameter samples in the two training sample sets are standardized to eliminate the difference in dimensions, resulting in a standardized feature sample set, which includes the first standardized feature sample set corresponding to the first training sample set, denoted as SS1, and the second standardized feature sample set corresponding to the second training sample set, denoted as SS2. The tree node samples marked as trunk nodes are separated from the first type of standardized feature sample set and the second type of standardized feature sample set to form a comprehensive trunk node sample set Smain. Tree node samples marked as leaf nodes are separated from the first type of normalized feature sample set and the second type of normalized feature sample set to form a comprehensive leaf node sample set Sleaf.
6. The artificial intelligence-based display screen production process management method according to claim 1, characterized in that, The specific implementation process of step S7 includes: Based on the optimal feature set Fmain of the backbone nodes and the optimal feature set Fleaf of the leaf nodes, bootstrap resampling is performed on the comprehensive sample set Smain of the backbone nodes and the comprehensive sample set Sleaf of the leaf nodes, respectively: Randomly select k tree node samples from the main node comprehensive sample set Smain to form k main node training subsets; Randomly select k tree node samples from the comprehensive sample set Sleaf of leaf nodes to form a k leaf node training subset; Where k is the number of base learners in the random forest; For each backbone node, a training subset is used to construct the corresponding C4.5 decision tree model based on the optimal feature set Fmain of the backbone node; For each leaf node, a subset is trained, and the corresponding C4.5 decision tree model is constructed based on the optimal feature set Fleaf of the leaf node; During the construction of each base learner, a random feature selection strategy is adopted to randomly select m tree node samples from the optimal feature set Fmain of the trunk node or the optimal feature set Fleaf of the leaf node to participate in the splitting decision of the base learner.
7. The artificial intelligence-based display screen production process management method according to claim 1, characterized in that, The specific implementation process of step S8 includes: Using a weighted voting method, based on the importance ratio of tree nodes in the production process, the C4.5 decision tree model corresponding to the optimal feature set Fmain of the trunk node and the C4.5 decision tree model corresponding to the optimal feature set Fleaf of the leaf node are integrated to form a two-branch random forest model. Input the real-time feature data of each tree node in the current production batch, after data standardization, into the corresponding random forest branch, and output the production capacity prediction results of each component.
8. A system for implementing the artificial intelligence-based display screen production process management method as described in claim 1, characterized in that, The system includes a component and task management module, a process flow decision tree construction module, a production process topology graph generation module, a production batch partitioning and sample set construction module, a feature processing module, a feature importance calculation module, a base learner construction module, and a random forest capacity prediction module; wherein: The component and task management module is used to build a display screen component type library and record the model information of each component; build a display screen component production database and record component production data; create production tasks and encapsulate them into the production database, record the quantity of a single component model, and publish the upper limit of production demand to each production end; The process flow decision tree construction module creates a process flow decision tree based on the assembly process flow of the finished display screen, so that each tree node corresponds to a component model and carries the real-time value of the production demand and consumption rate parameters of that component. The production process topology generation module is used to establish cross-flow relationships between various process flow decision trees, mark the main nodes and leaf nodes, and construct a complete production process topology under the display screen production task based on the two types of nodes. The production batch division and sample set construction module is used to divide the production task into production batches, construct a complete production process topology map for each batch stage, and realize independent storage and associated calling; for the topology map of each batch, a first type of training sample set in the state without missing components and a second type of training sample set in the state with missing components are constructed. The feature processing module is used to extract two types of core feature parameter samples from two types of training sample sets: real-time value of production demand and consumption rate parameter. The core feature parameter samples are standardized to eliminate the difference in units, and the main node comprehensive sample set and the leaf node comprehensive sample set are obtained. The feature importance calculation module is used to construct independent C4.5 algorithm execution models for the main node and the leaf node respectively. With "whether the component production capacity meets the assembly requirements" as the target classification label, it quantifies the information entropy, conditional entropy, information gain and information gain ratio of the target features, and obtains the optimal feature set of the main node and the optimal feature set of the leaf node based on the information gain ratio sorting and filtering. The base learner construction module performs bootstrap resampling on the comprehensive sample set of the main nodes and the comprehensive sample set of the leaf nodes, based on the optimal feature set of the main nodes and the optimal feature set of the leaf nodes, to generate k training subsets of the main nodes and k training subsets of the leaf nodes. For each training subset, a C4.5 decision tree model is constructed based on the corresponding optimal feature set. During the construction process, m tree nodes are randomly selected from the optimal feature set to participate in the node splitting decision. The random forest capacity prediction module is used to integrate the C4.5 decision tree model of the trunk node and the leaf node into a two-branch random forest model by using a weighted voting method based on the importance ratio of the tree node in the production process; it receives the real-time feature data of each tree node in the current production batch after standardization, inputs them into the corresponding random forest branches, and outputs the capacity prediction results of each component.