A ship section operation package intelligent generation method and system based on a knowledge graph and BOM analysis

Abstract

This invention discloses an intelligent generation method and system for ship segment work packages based on knowledge graphs and BOM parsing. The method includes constructing a ship process knowledge graph; retrieving all quantity nodes associated with the "work time" node to form a candidate pool; calculating and ranking the relevance between all nodes in the candidate pool and the "work time" node; selecting data from nodes with relevance higher than a set threshold as model input; training an initial weight and threshold using a particle swarm optimization algorithm to find a determined backpropagation neural network, resulting in a hybrid prediction model; based on the generation of specific segment work package instructions, acquiring quantity data and using the hybrid prediction model to predict standard work hours; and then, based on the ship process knowledge graph, querying the standard process information and resource scheduling information for the corresponding segment to form a standardized digital work package. This invention integrates heterogeneous process knowledge and accurately predicts work hours, achieving refined and automated production planning.

Description

Technical Field

[0001] This invention relates to a method for generating ship section work packages, and in particular to an intelligent method and system for generating ship section work packages based on knowledge graphs and BOM parsing. Background Technology

[0002] In the shipbuilding industry, existing production models generally suffer from data barriers and "information silos" between design, production, and management, resulting in incomplete process data, inconsistent information sources, and difficulty in reusing knowledge.

[0003] Furthermore, the calculation of key time quotas in production planning has long relied on traditional estimation and table lookup methods. These methods are not only inefficient but also inaccurate, deviating significantly from actual production conditions. This disconnect between planning and execution has led to problems such as unstable production rhythms, severely hindering the improvement of ship assembly and construction efficiency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent generation method for ship section work packages based on knowledge graphs and BOM parsing, solving the problems of scattered shipbuilding process data, inaccurate time prediction, and disconnect between plans and actual work. This invention also provides an intelligent generation system for ship section work packages based on knowledge graphs and BOM parsing.

[0005] The technical solution of the present invention is as follows:

[0006] A method for intelligent generation of ship section work packages based on knowledge graph and BOM parsing includes:

[0007] S1. Construct a knowledge graph of shipbuilding processes;

[0008] S2. Based on the ship process knowledge graph, retrieve all quantity nodes that are associated with the "working time" node to form a candidate pool;

[0009] S3. Calculate and sort the correlation between all nodes in the candidate pool and the "working hours" node, and select the data of nodes with a correlation higher than the set threshold as the model input.

[0010] S4. Train a BP neural network based on historical data, with the model input as input and working hours as output, to obtain a hybrid prediction model. The initial weights and thresholds of the BP neural network are determined by particle swarm optimization algorithm.

[0011] S5. Based on the instructions to generate specific segmented work packages, extract structured and unstructured bill of materials information to obtain quantity data, and based on the ship process knowledge graph, query the standard process information and resource scheduling information of the corresponding segment;

[0012] S6. Based on the material quantity data obtained from S5, input the mixed prediction model of S4 to predict the standard working hours, and then combine it with the information obtained from S5 to form a standardized digital work package.

[0013] Furthermore, within a set period, the deviation between the standard working hours predicted by the hybrid prediction model and the actual working hours is calculated. When the deviation exceeds a set update threshold, the node data of the ship's process knowledge graph is updated based on the actual working hours. The production data causing the deviation is then added as new samples to the original training dataset of the BP neural network, and the BP neural network is retrained to obtain the hybrid prediction model. Through this dynamic iterative optimization mechanism, the hybrid prediction model can learn and correct itself based on actual feedback from the production site, dynamically adapting to changes in team capabilities, process changes, etc., ensuring the long-term accuracy and practical guidance of the working hour quota, and overcoming the fundamental limitations of traditional static models.

[0014] Furthermore, the deviation is the average working hour deviation rate.

[0015] ,

[0016] Where n is the number of monitoring times within a set period, and the update threshold is 10% to 15%.

[0017] Furthermore, the time label in the new sample is calculated as actual working hours multiplied by the team efficiency coefficient. By correcting for the efficiency coefficient, the impact of team production capacity on working hours is reduced, allowing the model to learn a purer objective law between process and working hours.

[0018] Furthermore, the data in the shipbuilding process knowledge graph includes at least data from the product data management system and the enterprise resource planning system, as well as unstructured text knowledge from process manuals and technical specifications.

[0019] Another technical solution of the present invention is: an intelligent generation system for ship section work packages based on knowledge graphs and BOM analysis, comprising:

[0020] The knowledge graph module is used to construct a shipbuilding technology knowledge graph.

[0021] The node candidate module, based on the ship technology knowledge graph, retrieves all quantity nodes that are associated with the "working time" node to form a candidate pool;

[0022] The relevance selection module is used to calculate and sort the relevance between all nodes in the candidate pool and the "working hours" node, and select the data of nodes with a relevance higher than a set threshold as the model input.

[0023] The model building module trains a BP neural network based on historical data, using the model input as input and working hours as output, to obtain a hybrid prediction model;

[0024] The instruction extraction model extracts structured and unstructured bill of materials information to obtain quantity data based on the generation of specific segmented work packages. Based on the ship process knowledge graph, it queries the standard process information and resource scheduling information of the corresponding segment.

[0025] Additionally, the work package generation module, based on the material quantity data obtained from the instruction extraction model, inputs it into the hybrid prediction model of the model building module to predict standard working hours, and then combines it with the information obtained from the instruction extraction model to form a standardized digital work package.

[0026] Furthermore, it includes a model update module, which is used to calculate the deviation between the standard working hours predicted by the hybrid prediction model and the actual working hours within a set period. When the deviation exceeds the set update threshold, the node data of the ship process knowledge graph is updated based on the actual working hours, and the production data that caused the deviation is added to the training dataset of the original training BP neural network as a new sample. The BP neural network is then retrained to obtain the hybrid prediction model.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] This invention utilizes knowledge graph technology to break down "information silos," unifying and semantically modeling scattered and heterogeneous process knowledge to construct a computable knowledge foundation, providing reliable data support for intelligent prediction and decision-making.

[0029] This invention constructs a hybrid prediction model using a BP neural network with initial weights and thresholds determined by a particle swarm optimization algorithm. This transforms traditional time estimation, which relies on human experience, into data-driven intelligent prediction. It can accurately capture the complex nonlinear relationship between time and multidimensional influencing factors, significantly improving the calculation efficiency and accuracy of time quotas.

[0030] This invention utilizes a Bill of Materials (BOM)-driven intelligent generation process to transform the tedious manual planning process into an automated workflow. This ensures the completeness, accuracy, and consistency of work package contents, thereby improving production preparation efficiency and the standardization of on-site operations. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the intelligent generation method for ship section work packages based on knowledge graph and BOM analysis according to the present invention.

[0032] Figure 2 This is a schematic diagram of a portion of the shipbuilding technology knowledge graph, with "flat section" as the core node.

[0033] Figure 3This is a flowchart illustrating the process of training a BP neural network using the PSO algorithm for global optimization.

[0034] Figure 4 This is a fitness curve diagram for the PSO algorithm.

[0035] Figure 5 The training curve for the hybrid prediction model MSE is shown.

[0036] Figure 6 This is a comparison chart of the predictions and actual values ​​from the hybrid prediction model.

[0037] Figure 7 Time Deviation Monitoring Chart for Elite and Novice Work Teams

[0038] Figure 8 This is a comparison chart of the prediction accuracy of the hybrid prediction model before and after iteration. Detailed Implementation

[0039] The present invention will be further described below with reference to embodiments, but these are not intended to limit the scope of the invention.

[0040] The intelligent generation system for ship section work packages based on knowledge graphs and BOM parsing, according to embodiments of the present invention, includes:

[0041] The knowledge graph module is used to construct a shipbuilding process knowledge graph, specifically a process knowledge graph for the ship section manufacturing field. This shipbuilding process knowledge graph defines an ontology model covering core elements such as ship type, typical sections, standard procedures, work teams, job types, equipment, and quantities. It employs knowledge extraction, knowledge fusion, and knowledge instantiation technologies to integrate multi-source heterogeneous data from systems such as Product Data Management (PDM) and Enterprise Resource Planning (ERP), as well as discrete knowledge from unstructured texts such as process manuals and technical specifications, into a unified, interconnected semantic knowledge network. This network is then stored using a graph database (such as Neo4j).

[0042] The node candidate module, based on the ship technology knowledge graph, retrieves all quantity nodes that are related to the "working time" node to form a candidate pool.

[0043] The relevance selection module is used to calculate and sort the relevance between all nodes in the candidate pool and the "working hours" node, and select the data of nodes with a relevance higher than a set threshold as the model input.

[0044] The model building module trains a BP neural network based on historical data, with model input as input and working hours as output, to obtain a hybrid prediction model. This BP neural network uses the global optimization capability of the PSO algorithm to find the optimal initial weights and thresholds for the BP neural network, and then performs fine-tuning training. It uses the nonlinear mapping capability of the BP algorithm to achieve high-precision prediction of standard working hours.

[0045] As a preferred embodiment, to address dynamic factors on-site such as changes in team capabilities and process modifications, the system also includes a model update module to iteratively update the hybrid prediction model. This module calculates the deviation between the standard working hours predicted by the hybrid prediction model and the actual working hours within a set period. When the deviation exceeds a set update threshold, it updates the node data of the ship's process knowledge graph based on the actual working hours. The production data causing the deviation is then added as new samples to the original training dataset of the BP neural network. The BP neural network is retrained to obtain a new hybrid prediction model that better reflects the current production reality, and this model is deployed to the system for subsequent working hour predictions.

[0046] The instruction extraction model extracts structured and unstructured bill of materials information to obtain quantity data based on the generation of specific segmented work packages. Based on the ship's process knowledge graph, it queries the standard process information and resource scheduling information of the corresponding segment.

[0047] Additionally, the work package generation module, based on the material quantity data obtained from the instruction extraction model, inputs it into the hybrid prediction model of the model building module to predict standard working hours, and then combines it with the information obtained from the instruction extraction model to form a standardized digital work package.

[0048] Please combine Figure 1 As shown in the above embodiment, the intelligent generation method for ship section work packages implemented by the knowledge graph and BOM parsing-based intelligent generation system specifically includes the following processes:

[0049] Step S1: Construct a shipbuilding technology knowledge graph.

[0050] Neo4j graph database was chosen as the storage backend. First, the ontology (Schema) of the knowledge graph was defined. Core entity types include: [Project and Product Entities] (e.g., ship type, ship project, typical section), [Process Flow Entities] (e.g., standard process, job type), [Resource and Organization Entities] (e.g., work team, job type, equipment), and [Quantity and Measurement Entities] (e.g., quantity, working hours). Core relationship types include: (Typical Section) - [Own Process] -> (Standard Process), (Standard Process) - [Required Skill] -> (Job Type), (Work Team) - [Execution] -> (Standard Process), etc. Through knowledge extraction, fusion, and instantiation processes, combined with the Product Data Management System (PDM) and Enterprise Resource Planning System (ERP), knowledge from historical project data, process documents, BOM tables, and other data sources is transformed into nodes and edges in the knowledge graph, forming a knowledge network covering the entire segment manufacturing process. An example of a knowledge graph for a flat panel segment is shown below. Figure 2 As shown.

[0051] S2. Based on the ship technology knowledge graph, all material quantity nodes that are related to the "working time" node are retrieved through the query statement of the graph database to form a candidate pool, and the attribute data of the material quantity nodes are also introduced.

[0052] S3. Calculate and sort the correlation between all nodes in the candidate pool and the "working hours" node. The calculation is based on the historical actual working hours data corresponding to the "working hours" node and the attribute data of the material quantity node. Select the data of the nodes with a correlation higher than the set threshold as the model input.

[0053] A specific relevance calculation method employs XGBoost machine learning feature importance assessment, using XGBRegressor to build a prediction model. The dataset is divided into training and validation sets, and early stopping and cross-validation are used to prevent overfitting. This prediction model learns the weights of each feature (node ​​data) on work time, calculates feature importance scores based on feature split gain, sorts all features by score from highest to lowest importance, and determines a score value as a threshold. Features with scores higher than this threshold are used as subsequent model inputs. The parameters of the XGBRegressor prediction model are: number of base learners: 300, maximum tree depth: 5, learning rate: 0.05, subsampling ratio: 80%, column sampling ratio per tree: 80%, and feature importance calculation type: gain (feature split gain).

[0054] In this embodiment, 19 quantity nodes that significantly affect working hours were obtained, specifically:

[0055] Number of small assembly units N1, number of medium assembly units N2, weight of small assembly units W1, weight of medium assembly units W2, weight of large assembly units W3, flat welding length L1, horizontal welding length L2, vertical welding length L3, overhead welding length L4, number of turning times Nt;

[0056] The equipment unit weight Wq, iron outfitting component weight Wh, pipe outfitting component weight Wp, and electrical outfitting component weight We belong to the segmented pre-outfitting information;

[0057] The information pertains to the tire frame, including the tire frame type M and the tire area S.

[0058] The number N of the lifting equipment information;

[0059] The weight of the erected scaffolding, Wf, is part of the scaffolding information.

[0060] And the segmentation type K (represented by the proportion of surface area).

[0061] S4. Based on historical data, train a BP neural network with model input as input and working hours as output to obtain a hybrid prediction model. The initial weights and thresholds of the BP neural network are determined by particle swarm optimization algorithm.

[0062] Based on the material quantity node determined in step S3 and the "working time" node, relevant historical data are obtained. Among the 200 segmented data, 150 groups are randomly selected as training samples and 50 groups are selected as test samples for training the hybrid prediction model.

[0063] The BP neural network is a three-layer network with 19 neurons in the input layer, 8 neurons in the hidden layer, and 1 neuron in the output layer. The PSO (Particle Swarm Optimization) algorithm parameters are set as follows: population size 50, maximum number of iterations 100, learning factors c1=2.5, c2=1.0. The BP network parameters are set as follows: number of iterations 500, expected root mean square error 10⁻⁶. -7 The learning rate is 0.01.

[0064] Please combine Figure 3 As shown, the training process first runs the PSO algorithm for 100 iterations, using MSE as the fitness, to find the globally optimal initial weights and threshold of the BP network. The PSO fitness curve is shown below. Figure 4 As shown, its optimal fitness is 0.01753. Then, these optimal parameters are assigned to the BP network, and it is trained using 150 training samples. The MSE training curve during the training process is shown below. Figure 5 As shown.

[0065] After training, model validation was performed using the remaining 50 test samples to verify the trained model through simulation. The comparison results between predicted and actual values ​​are as follows: Figure 6 As shown, the results indicate that the error is within the allowable range for production, proving that the initial model is feasible and effective.

[0066] To address dynamic factors on-site, such as changes in team capabilities and process modifications, a closed-loop iterative optimization mechanism is established. Within a set period, the deviation between the "actual working hours" and the "standard working hours" predicted by the model is continuously monitored. When the average relative deviation rate of a certain type of operation continuously exceeds a preset threshold (e.g., ±15%) within a specific evaluation period, the iterative optimization process is automatically triggered.

[0067] ,

[0068] Where n is the number of monitoring times within the set period.

[0069] The system updates relevant attributes in the knowledge graph based on the latest actual working hours data. Simultaneously, the latest batch of production data that caused the triggering bias is added as new samples to the existing model training dataset.

[0070] Specifically, the 19 input features in the new samples remain unchanged. The output labels cannot be directly based on actual working hours; instead, they need to be corrected using an efficiency coefficient: the standard working hour label for the newly added training samples = actual working hours × the efficiency coefficient of the work group. For example, if a novice work group actually used 120 hours with an efficiency coefficient of 0.8, then the "true standard working hour label" for this batch of tasks in the augmented dataset should be 120 × 0.8 = 96 hours. The hybrid prediction model is then retrained using the augmented dataset to generate a new version that better reflects the current production reality. This new version is then deployed into the system for subsequent working hour predictions, thus completing one closed-loop iteration.

[0071] This embodiment uses a six-month simulation experiment.

[0072] 1) Scenario setting: The first 3 months are a stable production period. Starting from the 4th month, a high-efficiency "elite team" (efficiency coefficient η_elite=1.2) and a low-efficiency "novice team" (efficiency coefficient η_novice=0.8) are introduced to simulate changes in the capabilities of production personnel.

[0073] 2) Monitoring and Triggering: In the fourth month, the system monitored an average time deviation rate of -16.4% for the elite work team and +25.4% for the novice work team. Both consistently exceeded the 15% trigger threshold. Figure 7 As shown, therefore, at the end of the 4th month, the iterative optimization process is triggered.

[0074] 3) Iteration and correction: The production data of the two work groups in the fourth month are used as new samples and combined with their efficiency coefficients to enhance the original training set. Then, the enhanced dataset is used to retrain the hybrid prediction model to generate a new version of the model.

[0075] 4) Performance Verification: Starting from the 5th month, the system uses the new model for prediction. For example... Figure 8 As shown, the time prediction deviation rates for both work groups quickly returned to the normal range of [-5%, +5%]. This demonstrates that the dynamic iteration mechanism of this invention can effectively capture production changes and restore prediction accuracy through self-learning.

[0076] S5. Based on the instructions to generate specific segmented work packages, extract structured and unstructured bill of materials information to obtain quantity data, and query the standard process information and resource scheduling information of the corresponding segment based on the ship process knowledge graph.

[0077] Taking the generation of a job package for segment "DC11" as an example, after receiving the request to generate the job package for segment "DC11", the structural BOM of the segment is extracted from the PLM system, and the process BOM such as welding length is extracted from the associated process card using regular expression technology.

[0078] The structural BOM includes, but is not limited to, objective physical attributes such as part number, part name, material, dimensions (length / width / thickness), weight, quantity, and assembly hierarchy (e.g., the specific sub-assembly / intermediate assembly number to which it belongs).

[0079] Process BOM (extracted from process card regular expressions): including but not limited to bevel type, welding type (flat welding, vertical welding, etc.), welding length, assembly clearance requirements, special tooling and fixture requirements, machining allowance and other process execution attributes.

[0080] Knowledge Query: Starting with "DC11", query information such as its ship type, standard process flow, required equipment and job type in the knowledge graph.

[0081] S6. Based on the material quantity data obtained from S5, input the mixed prediction model of S4 to predict the standard working hours, and then combine it with the information obtained from S5 to form a standardized digital work package.

[0082] Specifically, all retrieved information, including basic information (project number, segment number), task schedule (standard operating procedures, job types, number of workers, standard working hours), 3D model view, pallet information (bill of materials), construction operation standards, and safety management requirements, is automatically integrated into a standardized "process information integration page," which is then distributed to the production team as a complete digital work package. The work package in this embodiment includes the following data:

[0083] Project number and segment number: These are derived from the planner's initial input (trigger command) on the system's front-end interface.

[0084] 3D model view: Obtained directly from the lightweight model interface of PLM automatically called by the system based on the segment number.

[0085] Pallet information (bill of materials): derived from the extraction results of the PLM system and process cards by the BOM data parsing module.

[0086] Standard operating procedures and job types: These are derived from the pre-defined relationships between ship type, typical section, operating procedures, and job types in the system.

[0087] Work teams: These are derived from the team management section of the MES system.

[0088] Number of workers: This is derived from the number of workers required for the standard process set when the standard process is preset in the system.

[0089] Standard working hours: derived from the calculation and prediction results of the hybrid prediction model based on the aforementioned extracted material quantity data.

[0090] Construction operation standards and safety management requirements: These are derived from the system's preset standards associated with standard operating procedures. When a standard operating procedure is required in the operation of this segment, the relevant construction operation standards and safety management requirements will be automatically retrieved.

[0091] Finally, it should be noted that the specific methods of the above embodiments can form a computer program product. At the same time, the computer program product implemented in this application can be stored on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.).

Claims

1. A ship section job package intelligent generation method based on knowledge graph and BOM analysis, characterized in that, include: S1. Construct a knowledge graph of shipbuilding processes; S2. Based on the ship technology knowledge graph, retrieve all quantity nodes that are associated with the "working time" node to form a candidate pool; S3. Calculate and sort the correlation between all nodes in the candidate pool and the "working hours" node, and select the data of nodes with a correlation higher than the set threshold as the model input. S4. Train a BP neural network based on historical data, with the model input as input and working hours as output, to obtain a hybrid prediction model. The initial weights and thresholds of the BP neural network are determined by particle swarm optimization algorithm. S5. Based on the instructions to generate specific segmented work packages, extract structured and unstructured bill of materials information to obtain quantity data, and based on the ship process knowledge graph, query the standard process information and resource scheduling information of the corresponding segment; S6. Based on the material quantity data obtained from S5, input the mixed prediction model of S4 to predict the standard working hours, and then combine it with the information obtained from S5 to form a standardized digital work package.

2. The ship section job package intelligent generation method based on knowledge graph and BOM analysis according to claim 1, characterized in that, Within a set period, the deviation between the standard working hours predicted by the hybrid prediction model and the actual working hours is calculated. When the deviation exceeds the set update threshold, the node data of the ship process knowledge graph is updated based on the actual working hours. The production data that caused the deviation is added to the training dataset of the original training BP neural network as a new sample, and the BP neural network is retrained to obtain the hybrid prediction model.

3. The ship section job package intelligent generation method based on knowledge graph and BOM analysis according to claim 2, characterized in that, The deviation is the average working hour deviation rate. ， Where n is the number of monitoring times within a set period, and the update threshold is 10% to 15%.

4. The ship section job package intelligent generation method based on knowledge graph and BOM analysis according to claim 2, characterized in that, The working hour label in the new sample = actual working hours × team efficiency coefficient.

5. The ship section job package intelligent generation method based on knowledge graph and BOM analysis according to claim 1, characterized in that, The data in the ship technology knowledge graph includes at least data from the product data management system and enterprise resource planning system, as well as unstructured text knowledge from process manuals and technical specifications.

6. The ship section job package intelligent generation method based on knowledge graph and BOM analysis according to claim 1, characterized in that, The data in the ship technology knowledge graph includes at least data from the product data management system and enterprise resource planning system, as well as unstructured text knowledge from process manuals and technical specifications.

7. A ship section job package intelligent generation system based on a knowledge graph and BOM analysis, characterized in that, include: The knowledge graph module is used to construct a shipbuilding technology knowledge graph. The node candidate module, based on the ship technology knowledge graph, retrieves all quantity nodes that are associated with the "working time" node to form a candidate pool; The relevance selection module is used to calculate and sort the relevance between all nodes in the candidate pool and the "working hours" node, and select the data of nodes with a relevance higher than a set threshold as model input. The model building module trains a BP neural network based on historical data, with the model input as input and working hours as output, to obtain a hybrid prediction model. The initial weights and thresholds of the BP neural network are determined by the particle swarm optimization algorithm. The instruction extraction model extracts structured and unstructured bill of materials information to obtain quantity data based on the generation of specific segmented work packages. Based on the ship process knowledge graph, it queries the standard process information and resource scheduling information of the corresponding segment. Additionally, the work package generation module, based on the material quantity data obtained from the instruction extraction model, inputs it into the hybrid prediction model of the model building module to predict standard working hours, and then combines it with the information obtained from the instruction extraction model to form a standardized digital work package.

8. The ship section job package intelligent generation system based on knowledge graph and BOM analysis according to claim 7, characterized in that, The system includes a model update module, which calculates the deviation between the standard working hours and the actual working hours predicted by the hybrid prediction model within a set period. When the deviation exceeds the set update threshold, the system updates the node data of the ship process knowledge graph based on the actual working hours and adds the production data that caused the deviation as a new sample to the training dataset of the original training BP neural network. The BP neural network is then retrained to obtain the hybrid prediction model.

9. The ship section job package intelligent generation system based on knowledge graph and BOM analysis according to claim 8, characterized in that, The deviation is the average working hour deviation rate. ， Where n is the number of monitoring times within a set period, and the update threshold is 10% to 15%.

10. The ship section job package intelligent generation system based on knowledge graph and BOM analysis according to claim 8, characterized in that, The working hour label in the new sample = actual working hours × team efficiency coefficient.

Images