An artificial intelligence-based outfitting piece production quality control system and a running method thereof

Abstract

The application provides an artificial intelligence-based outfitting piece production quality control system and a running method thereof, and contains a data acquisition module, an AI quality analysis module and a real-time control module. The data acquisition module collects full-process data by means of multiple sensors, industrial cameras and 3D scanners. The AI quality analysis module adopts multiple models to detect the collected process data in parallel, and outputs a comprehensive quality inspection conclusion after weighted fusion. The real-time control module completes early warning, automatic parameter adjustment and artificial maintenance closed-loop disposal in stages. The outfitting piece production quality control system further includes a traceability module and a self-adaptive optimization module. The traceability module archives the production full-link data of the workpiece by taking a two-dimensional code + RFID as an index, and realizes rapid tracing. The self-adaptive optimization module periodically iterates the model and the control process parameter threshold. The outfitting piece production quality control system can realize full-automatic intelligent quality inspection, and the defect recognition accuracy rate reaches more than 99%, thereby greatly reducing the labor cost. The production abnormality can be disposed in seconds, and the rework rate is significantly reduced.

Description

Technical Field

[0001] This invention belongs to the field of shipbuilding technology, specifically relating to an artificial intelligence-based outfitting component production control system and its operation method. Background Technology

[0002] In the shipbuilding process, self-made outfitting components (such as ship piping, equipment supports, cable trays, and compartment housings) are key parts that ensure the structural integrity and functional reliability of the ship. Their production quality directly affects the ship's navigation safety and service life. Currently, shipbuilding plants mainly rely on traditional manual inspection and experience-based management for the production quality control of self-made outfitting components, which has the following technical shortcomings: 1. Low inspection efficiency and large error: Traditional control adopts the "manual inspection + sampling inspection" mode. For indicators such as the dimensional accuracy and surface defects (such as cracks, deformation, and welding porosity) of outfitting parts, it relies on the visual observation of the inspectors and the measurement of hand-held measuring tools. The inspection time for a single part can be as long as 5 to 10 minutes. Moreover, it is affected by the experience and fatigue of the personnel, and the defect identification accuracy is less than 85%, which can easily lead to unqualified products flowing into the subsequent assembly process.

[0003] 2. Fragmented process data makes traceability difficult: Outfitting production involves multiple processes such as material preparation, forming, welding, and surface treatment. Data from each process (such as welding current, forming pressure, and processing temperature) is stored in different equipment control systems or paper records without a unified data link. When quality problems occur, each process's data must be checked manually, with a traceability cycle of 24 to 48 hours, making it impossible to quickly locate the root cause of the problem.

[0004] 3. Delayed early warning and passive control: Traditional control only addresses quality issues after they are detected during the inspection stage, failing to monitor abnormal production process parameters in real time. For example, if current fluctuations exceed a threshold during welding, weld defects are only discovered during finished product inspection, resulting in batches of substandard products and a rework rate as high as 12% to 15%, significantly increasing production costs and time.

[0005] 4. Lack of adaptive optimization in control strategies: The production processes for different types of outfitting components (such as carbon steel pipes and stainless steel supports) vary significantly. Traditional control measures require manual formulation of different testing standards and parameter thresholds, and cannot be dynamically adjusted based on historical quality data. When quality fluctuates due to factors such as changes in raw material batches or equipment aging, the control strategies lag behind, making it difficult to maintain stable production quality.

[0006] In the existing technology, although some shipbuilding companies have introduced industrial control systems (such as MES) for production management, they only realize data collection and simple statistics, without combining artificial intelligence technology for quality analysis and intelligent control; other technologies focus on AI inspection of a single process (such as welding), lack full-process quality control capabilities, and cannot be adapted to the production scenarios of multiple types of outfitting parts. Summary of the Invention

[0007] To achieve the above objectives, the present invention provides the following technical solution: An AI-based outfitting component production control system includes: a data acquisition module, comprising a process parameter acquisition unit and an appearance data acquisition unit; the process parameter acquisition unit includes several sensors arranged in various production processes, which are used to acquire process parameters of each production process in real time; the appearance data acquisition unit is used to acquire two-dimensional appearance images and three-dimensional point cloud data of the workpiece; and an AI quality analysis module, comprising a data preprocessing unit, a model storage and retrieval unit, and a quality result generation unit, wherein the data preprocessing unit is used to convert the process parameters, two-dimensional appearance images, and three-dimensional point cloud data detected by the sensors into standardized features; and the model storage and retrieval unit stores a process parameter anomaly detection model, etc. An appearance defect identification model and a dimensional accuracy judgment model are used to perform parallel calculations on each of the standardized features and output sub-item inspection results. The quality result generation unit uses a weighted voting fusion mechanism to integrate the sub-item inspection results and generate the final quality conclusion of the workpiece. A real-time control module is used to receive the final quality conclusion and to proactively intervene and handle abnormalities in the production process in a graded closed-loop manner. The real-time control module includes an abnormality warning unit and a parameter automatic adjustment unit. The abnormality warning unit is used to trigger an early warning when the judgment result of the final quality conclusion is unqualified or the process parameters are abnormal. The parameter automatic adjustment unit is used to automatically generate defects caused by fluctuations in the process parameters and issue process correction instructions to relevant production equipment.

[0008] In some embodiments, the outfitting component production control system further includes a traceability module, which comprises a unique identifier generation unit, a data association storage unit, and a traceability query unit. The unique identifier generation unit equips each workpiece with a traceability identifier consisting of a QR code and a radio frequency identification chip. The data association storage unit uses a MySQL database and a MongoDB database, and uses the traceability identifier as an index to bind and archive all input and output data generated by a single workpiece in the data acquisition module, the AI ​​quality analysis module, and the real-time control module. The traceability query unit provides production personnel with traceability query methods including QR code scanning, multi-dimensional conditional query, and visual timeline display.

[0009] In some embodiments, the outfitting component production control system further includes: an adaptive optimization module, which comprises a historical data filtering unit, a model parameter optimization unit, and a control threshold optimization unit. The historical data filtering unit is used to filter valid training samples from the MySQL database and the MongoDB database. The model parameter optimization unit uses a gradient descent algorithm with anomaly detection accuracy or defect recall as the objective function to iteratively optimize the process parameter anomaly detection model, appearance defect identification model, and dimensional accuracy determination model. The control threshold optimization unit is used to construct a cost-benefit analysis model by combining historical defect data and rework loss data, and dynamically adjust the warning threshold and parameter correction threshold of the real-time control module.

[0010] In some implementations, the process parameter acquisition unit further includes an edge computing gateway, which is deployed on the production workshop site to perform outlier removal and data normalization preprocessing on the raw data acquired by the sensors, and output regularized time-series feature data to the AI ​​quality analysis module.

[0011] In some embodiments, the appearance data acquisition unit includes multiple multi-view industrial camera arrays and multiple 3D scanners arranged in each of the production processes; the industrial cameras are used to capture full-coverage two-dimensional high-definition images of the workpiece and capture appearance defects of the workpiece; the 3D scanners are used to generate a three-dimensional point cloud model of the workpiece and calculate the dimensional deviation between the workpiece and the standard design drawings.

[0012] In some embodiments, the data preprocessing unit includes: a time-series process parameter processing subunit, which sets a sliding window, extracts the mean, variance, and peak value of the process parameters within the sliding window, and generates time-series process features; a two-dimensional appearance image processing subunit, which performs histogram equalization enhancement and Gaussian filtering noise reduction on the two-dimensional appearance image, and divides the defect region of the two-dimensional appearance image; and a three-dimensional point cloud data processing subunit, which uses the ICP algorithm to register the three-dimensional point cloud data, aligns the actual three-dimensional model of the workpiece with the design theoretical model, and automatically calculates the dimensional deviation value.

[0013] In some implementations, the process parameter anomaly detection model is based on the random forest algorithm, which is used to output the probability of process parameter anomalies after inputting the process timing features; the appearance defect recognition model is based on the improved YOLOv8 algorithm, which is used to output the coordinate position, defect area and defect severity type of the defect after inputting the two-dimensional appearance image; and the dimensional accuracy determination model is based on the support vector machine algorithm, which is used to input the three-dimensional dimensional deviation value.

[0014] In some implementations, the real-time control module further includes a manual intervention scheduling unit, which is configured to: trigger when automatic parameter adjustment fails to eliminate defects, automatically generate a task order containing a repair plan, and push it to a handheld terminal on the production site.

[0015] This invention also provides a method for operating an AI-based outfitting component production control system as described above, comprising the following steps: the workpiece sequentially flows through each of the production processes, and the process parameters, the two-dimensional appearance image, and the three-dimensional point cloud data are synchronously collected through a process parameter acquisition unit and an appearance data acquisition unit; the process parameters, the two-dimensional appearance image, and the three-dimensional point cloud data are converted into standardized features and then fed into the process parameter anomaly detection model, the appearance defect identification model, and the dimensional accuracy judgment model for parallel reasoning, and then weighted voting is used to generate the final quality conclusion of the workpiece; graded handling is performed according to the final quality conclusion: when the process parameters are abnormal, an early warning is triggered; when the process parameters fluctuate, the process correction instruction is automatically issued to the relevant production equipment.

[0016] In some embodiments, the operating method further includes: when the workpiece enters the unloading process, generating a traceability identifier containing a QR code and an RFID chip for each workpiece, the traceability identifier containing workpiece model, batch, order, and workstation information; uniformly archiving and associating the process data, quality analysis results, and handling records of each production process with the traceability identifier for production personnel to trace; simultaneously, periodically selecting effective training samples from historical archived data, using a gradient descent algorithm to iteratively optimize the process parameter anomaly detection model, the appearance defect identification model, and the dimensional accuracy judgment model, and dynamically adjusting the warning threshold and correction threshold based on historical defect data and rework loss costs.

[0017] Compared with existing technologies, the AI-based outfitting component production control system and its operation method provided by this invention have the following beneficial effects: 1. The outfitting component production control system provided by this invention relies on sensors, industrial cameras, and 3D scanners to complete the fully automatic acquisition of process parameters, appearance images, and three-dimensional dimensions. Combined with multi-model fusion AI algorithms, it replaces manual visual inspection and measuring tool inspection, reducing the inspection time of a single workpiece to less than 1 minute and improving the defect identification accuracy rate to over 99%. It completely avoids the problems of missed inspections and false inspections caused by subjective human judgment and visual fatigue. It can directly reduce the manual input of quality inspection positions by 30% to 40%, and reduce the shipyard's quality inspection labor costs in the long term.

[0018] 2. The outfitting component production control system provided by this invention relies on a real-time management module to achieve triple synchronous early warning. Slight fluctuations in process parameters can be automatically adjusted remotely without the need for machine shutdown. Complex defects are automatically dispatched to form a closed-loop handling system, reducing the overall rework rate of outfitting components from the traditional 12%~15% to below 3%, reducing scrapped raw materials and repeated processing time losses, and shortening the overall production cycle by 5%~8%.

[0019] 3. The outfitting component production control system provided by this invention equips each workpiece with a dual identification of QR code + RFID chip. Data collected, analyzed and processed throughout the entire process are uniformly linked and archived. The traceability query time is shortened from 24-48 hours of traditional manual investigation to less than 10 minutes. Key quality records are encrypted and uploaded to the blockchain, ensuring that the data is tamper-proof and leaves a complete trace, fully supporting ship delivery archiving, third-party ship inspection verification, and quality responsibility determination.

[0020] 4. The outfitting component production control system provided by this invention relies on the full-cycle historical production data of the traceability module. The adaptive optimization module periodically iterates and optimizes the AI ​​recognition model, dynamically updates process warnings and parameter correction thresholds, and can adapt to the synchronous production control needs of more than ten types of self-made outfitting components with different structures and processes, such as carbon steel pipelines, stainless steel supports, and cabin boxes. The entire optimization process runs automatically, without the need for process personnel to manually calibrate and modify quality control standards repeatedly. It can proactively adapt to various dynamic changes in on-site conditions, such as raw material batch replacement, long-term aging and wear of processing equipment, and fluctuations in production capacity during peak and off-peak seasons. It avoids the hidden quality risks caused by the lag in manual standard updates from the root, and maintains a stable and accurate full-process quality control capability in the long term.

[0021] 5. The outfitting production control system provided by this invention can mine the patterns of equipment wear and tear by accumulating long-term, multi-dimensional production data, predict the aging of tooling such as welding robots and bending equipment and the wear cycle of accessories in advance, arrange maintenance and repair in advance, reduce equipment downtime by 15% to 20%, and ensure continuous and stable production of the production line. Attached Figure Description

[0022] Figure 1 A schematic diagram of the structure of the outfitting component production control system based on artificial intelligence provided by the present invention. Detailed Implementation

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0024] To keep the drawings simple, each drawing only schematically shows the parts relevant to the invention, and they do not represent the actual structure of the product.

[0025] The present invention provides an artificial intelligence-based outfitting component production control system, combined with Figure 1 As shown, it includes: a data acquisition module, an AI quality analysis module, and a real-time control module, specifically: The aforementioned data acquisition module is responsible for the unified collection of production data across the entire process and for front-end cleaning and preprocessing, thus solving the problems of low efficiency, scattered data, and high noise interference in the raw data associated with traditional manual data collection. Specifically, it includes the following units: (1) Process parameter acquisition unit Hardware deployment: Various sensors, such as temperature sensors, pressure sensors, current and voltage sensors, and encoders, are deployed in batches at key points of critical production equipment such as feeding machines, bending machines, welding robots, and sandblasting machines to collect process parameters of each production process in real time, such as welding temperature and welding current, so as to realize the dynamic collection and recording of key data in the production process at all times.

[0026] For example, a cutting temperature sensor (model PT100, measuring range 50~400℃) and a cutting speed encoder (resolution 1000 pulses / revolution) are installed on the blanking machine; a pressure sensor (model CYB~20S, measuring range 0~1500kN) and an angle sensor (model WDD35D4, measuring range 0~360°) are installed on the bending machine; a current / voltage sensor (model CSM025, measuring range 0~500A / 0~50V) and a weld temperature sensor (model K thermocouple, measuring range 0~1000℃) are installed on the welding robot (model ABBIRB 1410); and a pressure sensor (model MPM480, measuring range 0~1MPa) and a coating thickness sensor (model TT220, measuring range 0~1000μm) are installed on the sandblasting machine.

[0027] The process parameter acquisition unit collects parameters covering the following four major process parameters: Material cutting process: cutting speed of sheet / pipe (0~500mm / min), cutting temperature (20~300℃), cut flatness error (0~0.5mm).

[0028] Forming process: bending pressure (0~1000kN), bending angle (0~180°), single forming time (0~60s).

[0029] Welding process: welding current (50~500A), welding voltage (10~40V), welding speed (5~50mm / s), weld temperature (20~800℃).

[0030] Surface treatment process: sandblasting pressure (0.1~0.8MPa), sandblasting time (0~300s), coating thickness (0~200μm).

[0031] Furthermore, the raw data directly collected by the aforementioned sensors has issues such as noise, abnormal jumps, inconsistent formats, and data redundancy. Therefore, it is necessary to preprocess the data collected by the sensors. The preprocessing is completed by the Advantech UNO-2484G edge computing gateway deployed on-site in the workshop. There is no need to transmit massive amounts of raw sensor data to the backend server for further processing. Relying on the computing power of the edge computing gateway, on-site processing is achieved, completing two preprocessing steps: outlier removal and data normalization. This filters out invalid data generated by instantaneous equipment jitter and outputs regular time-series feature data, reducing the load on the backend AI quality analysis module.

[0032] (2) Appearance data acquisition unit Hardware Deployment: Multi-view industrial camera arrays and 3D scanners are deployed at the outlets of each process (material cutting, forming, welding, and surface treatment processes). For example, one industrial camera (Hikvision MV-CA020-10GM, 20-megapixel resolution, 30fps) is deployed at the outlet of the material cutting process; two industrial cameras are deployed on the top and sides of the outlet of the bending process; three industrial cameras and one 3D scanner (Keyence LK-G80, ±0.02mm scanning accuracy) are deployed on the top and sides of the outlet of the welding process; and one industrial camera and one 3D scanner are deployed at the outlet of the sandblasting process.

[0033] Data Acquisition: Industrial cameras capture full coverage of high-resolution 2D images of workpieces (workpieces are manufactured through processes such as blanking, forming, welding, and surface treatment. For ease of distinction, workpieces that are not yet completed in the process flow are referred to as workpieces), capturing 8 types of appearance defects such as cracks, dents, weld porosity, and surface deformation; 3D scanners generate 3D point cloud models of workpieces, automatically comparing them with the theoretical dimensions of the workpiece design drawings, and calculating key dimensional deviations such as pipe inner diameter, support hole diameter, and box verticality.

[0034] The process parameter acquisition unit and appearance data acquisition unit can completely replace the traditional manual visual observation and handheld measuring tool single-point measurement mode, and realize the fully automatic synchronous acquisition of three types of data: process parameters, appearance images, and three-dimensional dimensions.

[0035] The aforementioned AI quality analysis module receives standardized multi-dimensional data (process parameters collected by the process parameter acquisition unit and appearance images and 3D images collected by the appearance data acquisition unit) from the data acquisition module. Based on a multi-fusion machine learning algorithm, it completes quality judgment, eliminating subjective biases in manual quality inspection. This unit includes a data preprocessing unit, a model storage and retrieval unit, and a quality result generation unit. Specifically: (1) Data preprocessing unit, used to extract differentiated features from the three different types of data output by the data acquisition module and convert them into standardized features for the adaptation model: ① Time-series process parameter processing subunit: For the process parameters normalized by the edge gateway, a 10-second sliding window is set to extract the time-series features such as the mean, variance, and peak value of the process parameters within the sliding window. The continuously fluctuating raw parameters output by the sensor are transformed into time-series process features that can be recognized by the model (the process parameter anomaly detection model of the model storage and retrieval unit), which facilitates the backend random forest model in determining whether there are abnormal fluctuations in the process parameters.

[0036] ② Two-dimensional appearance image processing subunit: Histogram equalization (image enhancement, brightening of fine cracks and pore defects) and Gaussian filtering (removing dust and light noise) are performed on the two-dimensional images of the workpiece captured by the industrial camera. Then, the defect region is divided by the U-Net semantic segmentation model to eliminate interference terms in the two-dimensional image, accurately select the area where the workpiece defect is located, reduce the recognition difficulty of the model (appearance defect recognition model of model storage and retrieval unit) and improve the detection rate of small defects.

[0037] ③ 3D Point Cloud Data Processing Subunit: This subunit uses the ICP algorithm to register the point cloud generated by the 3D scanner, aligning the actual 3D model of the workpiece with the theoretical design model. It automatically calculates dimensional deviations such as hole diameter, pipe inner diameter, and perpendicularity. By quantifying dimensional errors, it outputs deviation values, providing quantitative input for the model (the dimensional accuracy determination model of the model storage and retrieval unit), thus achieving automatic dimensional accuracy determination.

[0038] (2) Model storage and retrieval unit, including process parameter anomaly detection model, appearance defect identification model and dimensional accuracy judgment model. The model storage and retrieval unit identifies the workpiece material and type according to the production order and automatically matches the corresponding model for parallel operation: ① Process parameter anomaly detection model: Based on the random forest algorithm, the input is the preprocessed process time sequence features, and the output is the probability of process parameter anomalies (0~1 interval, with the judgment threshold set to 0.8).

[0039] ② Appearance Defect Recognition Model: Based on the improved YOLOv8 algorithm, the model automatically identifies surface defects (including cracks, pores, deformation, inclusions, lack of fusion, undercut, burn-through, etc.) by taking a pre-processed two-dimensional appearance image as input, and outputs the defect coordinates, defect area, and defect severity type.

[0040] ③ Dimensional accuracy judgment model: Based on the SVM support vector machine algorithm, the model takes the deviation value of the three-dimensional dimension as input and automatically outputs the judgment result of qualified / unqualified.

[0041] (3) Quality Result Generation Unit After the process parameter anomaly detection model, appearance defect identification model, and dimensional accuracy judgment model each independently output their sub-item inspection results, they are uniformly sent to the quality result generation unit. The quality result generation unit uses a weighted voting fusion mechanism to integrate the output results of the three types of models and generate the final overall pass / fail judgment for the workpiece.

[0042] The weighting rules in the quality result generation unit are as follows: the weight of process parameters is 0.3, the weight of appearance defect identification is 0.4, and the weight of dimensional accuracy judgment is 0.3. The final quality conclusion of the workpiece is output comprehensively. If it is judged as unqualified, a standardized defect report is automatically generated, clearly indicating the defect type, severity (minor / moderate / serious), related processes, and the negative impact of the defect on the ship's performance (strength, sealing, etc.).

[0043] The AI ​​quality analysis module replaces manual subjective quality inspection with AI algorithms, reducing the inspection time for a single workpiece to less than 1 minute, increasing the defect identification accuracy to over 99%, and significantly reducing the cost of manual inspection.

[0044] The aforementioned real-time control module is the workshop-level anomaly handling execution terminal. It receives the final quality conclusions from the upstream AI quality analysis module and proactively intervenes in and handles anomalies in the production process through a tiered, closed-loop system. This reverses the traditional passive quality control model of "rework after completion," effectively reducing batch rework losses. This module includes an anomaly early warning unit, an automatic parameter adjustment unit, and a manual intervention scheduling unit. Specifically: (1) Abnormal early warning unit When a workpiece is deemed unqualified or a process parameter becomes abnormal (e.g., the probability of abnormal process parameter exceeds the set threshold of 0.8), the abnormality warning unit will simultaneously trigger a triple warning mechanism within 10 seconds: a sound and light alarm on the workshop floor, a pop-up window in the production management terminal system, and a text message push to the process team leader.

[0045] The warning information is linked to the unique identifier of the workpiece. The unique identifier clearly marks the workpiece number, the process to which the defect belongs, and a detailed description of the defect, which makes it easy for managers to quickly locate the problematic workpiece.

[0046] (2) Automatic parameter adjustment unit For minor defects caused by simple fluctuations in process parameters, the automatic parameter adjustment unit automatically generates process correction instructions, which are then sent in real time to the PLC control systems of production equipment such as welding robots, bending machines, and sandblasting machines via the workshop's industrial Ethernet, enabling remote and autonomous parameter adjustment of the equipment.

[0047] This unit has built-in fixed parameter correction rules, strictly following the proportional logic of "adjusting 5 units for every 10 units of deviation" to precisely adjust process parameters. For example, when the measured welding current exceeds the standard threshold of 10A, the automatic parameter adjustment unit will automatically issue a control command to the welding robot to "adjust the current to 5A". Similarly, when other process parameters such as bending pressure and sandblasting pressure show positive or negative deviations, they are all corrected in the opposite direction according to the same proportion, quickly pulling each process parameter back to the standard allowable range.

[0048] The entire parameter adjustment process is fully automated, requiring no equipment downtime or manual parameter modification by operators at the control panel, thus maintaining continuous production. The automatic parameter adjustment unit also retains complete logs, including parameter values ​​before and after adjustment, adjustment command issuance time, and equipment feedback status, all synchronously linked to a unique identifier on the workpiece. If the process parameters still cannot return to the standard range after multiple rounds of automatic parameter adjustment, the process is automatically transferred to the manual intervention scheduling unit for handling.

[0049] (3) Manual intervention scheduling unit For two types of complex quality problems that cannot be eliminated by automatic parameter adjustment: first, defects such as large-area deformation, deep cracks, large-area porosity, and coating peeling in outfitting components; second, process parameters still exceeding the standard and defects that cannot be eliminated even after the automatic parameter adjustment unit issues correction commands multiple times. When either of these two mining-related issues triggers the manual intervention scheduling unit, a manual intervention task order is automatically generated and pushed to the handheld PDA mobile terminals of on-site quality inspectors and maintenance personnel.

[0050] The manual intervention task sheet includes a unique identifier for the workpiece, a defect location map, the corresponding production process for the defect, the severity level of the defect, and a standardized repair plan (rework welding, reshaping repair, re-sandblasting and spraying, secondary dimensional grinding, etc.). After the maintenance personnel complete the defect handling, they can upload images of the repair site and fill in the operation process on the PDA; after the quality inspector completes the re-inspection, the acceptance result is entered, and the task sheet is officially closed.

[0051] The manual intervention scheduling unit automatically records information such as the time of manual intervention task assignment, the personnel handling the task, the entire operation process, the re-inspection results, and the scrapping determination. All records are synchronously linked to the unique identifier of the workpiece, forming a complete closed-loop management.

[0052] In some implementations, under traditional methods, when a workpiece has a problem, manual review of paper records and historical data from different equipment is required, resulting in a traceability cycle of 24-48 hours. To address this issue, the outfitting component production control system also includes a traceability module. This module serves as the data storage and traceability query platform for the entire process, uniformly receiving all production, quality inspection, and disposal data from the three modules: data collection, AI quality analysis, and real-time control. It establishes a full lifecycle data association system for each workpiece, completely resolving the pain points of fragmented data in traditional processes and the extremely long time required for quality problem investigation. This module includes a unique identifier generation unit, a data association storage unit, and a traceability query unit, specifically: (1) Unique Identifier Generation Unit Each workpiece is equipped with a dual unique traceability identifier consisting of a QR code and a radio frequency identification chip (RFID chip), which is the unique identification of the workpiece mentioned above. The QR code and RFID chip contain core information such as workpiece model, serial number, production batch, order number, planned production time, processing station, detailed description of defects, and the process to which the defects belong.

[0053] After the workpiece blanking process is completed, a QR code is permanently engraved on the surface of the workpiece using a laser marking machine, and an RFID contactless tag is bound simultaneously, realizing a unique identity binding for each workpiece throughout its entire life cycle. On-site scanning of the code allows for quick retrieval of the corresponding full-process information.

[0054] (2) Data association storage unit A hybrid distributed storage architecture using MySQL and MongoDB databases is constructed, which can adapt to the data storage needs of various types such as structured process parameters, unstructured images, and defect reports. The data storage period is no less than 5 years, meeting the long-term quality archive retention standards for ship construction.

[0055] The data association storage unit uses the QR code and RFID chip of each workpiece as an index to bind and archive three types of core data generated throughout the entire business chain: The first type is process data: the original process parameters of each process collected by the data acquisition module, the standardized time-series feature data generated by the edge gateway preprocessing, and image data; the second type is analysis data: the initial judgment results, final judgment results, various defect locations, defect levels, and other quality inspection records output by the AI ​​quality analysis module; the third type is control data: the early warning logs, automatic equipment parameter adjustment instructions, manual repair and rework records generated by the real-time control module, and other records of the entire process.

[0056] Furthermore, for key traceability data (such as defect judgment conclusions and rework rectification records), the data association storage unit also introduces blockchain technology to complete on-chain encryption and evidence storage. By utilizing the immutable and fully traceable technical characteristics of blockchain, the risk of human modification or deletion of quality records is eliminated, and the true production and quality inspection information is completely retained. It can directly support the use of scenarios such as ship delivery archiving, third-party ship inspection verification, and quality responsibility traceability.

[0057] (3) Traceability query unit This unit provides workshop managers, quality inspectors, and process engineers with three convenient traceability query methods, covering different business scenarios such as rapid on-site verification, batch problem screening, and full-process review: ① Scan code query: Workshop managers and quality inspectors can scan the QR code on the surface of the workpiece with a handheld terminal to instantly retrieve complete full-process data of all processes from material cutting to surface treatment of the workpiece.

[0058] ② Conditional query: Supports precise retrieval by inputting multiple dimensions such as production batch, order number, defect type, processing station, and production time period, and quickly batch filters out workpieces with similar quality problems.

[0059] ③ Visualized display: The system presents the production time of each process, changes in process parameters, quality inspection results, and records of handling various anomalies in a timeline format. It also supports one-click export of traceability reports in PDF format. The reports fully include workpiece identification information, complete process parameters, defect details, and rectification records, and can be directly printed and archived.

[0060] The entire traceability and query process can be completed within 10 minutes, which is a significant improvement in efficiency compared to the traditional 24-48 hour manual investigation.

[0061] In some implementations, traditional quality control relies on manually setting various outfitting component process thresholds and inspection standards. When raw material batches change, equipment ages over time, or production conditions fluctuate during peak and off-peak seasons, quality is prone to fluctuations. Manually updating standards is lagging and cannot adapt to real-time production changes. Furthermore, traditional single AI inspection models are fixed and have limited ability to identify low-resolution defects such as tiny cracks and minor dimensional deviations, requiring manual re-labeling and adjustment of model parameters, resulting in high optimization costs and long cycles. Therefore, to address these issues, the outfitting component production quality control system provided by this invention also includes an adaptive optimization module. Based on historical production data stored in the traceability module, it periodically optimizes the model in the model storage and retrieval unit, solving the problems of traditional control standards being manually fixed and unable to adapt to dynamic production environments such as raw material batch changes and equipment aging. This module includes a historical data filtering unit, a model parameter optimization unit, and a control threshold optimization unit, specifically: (1) The historical data filtering unit can automatically filter valid training samples from the database of the traceability module and remove abnormal data that have no reference value, such as equipment failure shutdown and personnel misoperation.

[0062] The historical data filtering unit uses the following three criteria for filtering data: ① Data integrity: Process data, analysis data, and control data are all complete.

[0063] ② Sample representativeness: Covering different types of outfitting parts and different production periods (such as peak season / off-season).

[0064] ③ Data volume requirement: The historical data sample size for a single type of outfitting component is ≥1000 pieces.

[0065] (2) Model parameter optimization unit The gradient descent algorithm is used to iteratively optimize the model parameters in the model storage and retrieval unit. For example, for the process parameter anomaly detection model, the number of decision trees (current range 50~200 trees) and the maximum depth of decision trees (current range 10~30 layers) are adjusted with the anomaly recognition accuracy as the objective function. For the appearance defect recognition model, the anchor box size and learning rate (current range 0.001~0.01) are adjusted with the defect recall rate as the objective function.

[0066] In addition, a round of model optimization is automatically executed every quarter. After the optimization is completed, a quantitative amount (50 pieces) of workpieces are selected to carry out small-batch trial production verification to ensure that the overall recognition accuracy of the model is improved by more than 2%.

[0067] (3) The control threshold optimization unit can combine historical defect data and rework loss costs to build a cost-benefit analysis model and dynamically adjust the warning threshold and parameter correction threshold of the real-time control module. For example, if the original welding current warning threshold is set to 450A, data analysis shows that although the 440~450A range does not directly cause defects, the rework rate of subsequent processes increases by 5%. The control threshold optimization unit automatically sends an adjustment instruction to the abnormal warning unit to lower the warning threshold to 440A.

[0068] After optimization, simulation tests are conducted to simulate different parameter fluctuation scenarios. While ensuring that the overall defect rate decreases by more than 1% after optimization, the threshold is not too strict, which may cause frequent warnings and production interruptions.

[0069] The complete operation flow of the outfitting component production quality control system is as follows: S1. When a workpiece enters the unloading process, the unique identifier generation unit of the traceability module generates a unique traceability identifier for each workpiece, including a QR code marked on the workpiece surface and an RFID chip attached to the workpiece. The traceability identifier contains information such as workpiece model, production batch, order number, planned production time, and processing station, realizing a unique identity binding for each workpiece throughout its entire lifecycle and providing an indexing basis for the associated storage of all subsequent production and quality data.

[0070] S2. Workpieces carrying unique identifiers sequentially undergo processes such as blanking, forming, welding, and surface treatment. During each process, the data acquisition module simultaneously and automatically collects three types of data: process parameters, appearance images, and three-dimensional dimensions, as detailed below: The process parameter acquisition unit collects process parameters in real time at a sampling frequency of 10Hz through sensor arrays deployed at key points of production equipment in each process. The raw data collected by the sensors is processed by an edge computing gateway deployed on the workshop floor to remove equipment anomalies and normalize the data before transmitting the normalized valid data to the backend server. Simultaneously, the appearance data acquisition unit operates in sync with a multi-view industrial camera array and a 3D structured light scanner deployed at the exit of each process. The industrial cameras are used to capture defects such as cracks, dents, weld porosity, and surface deformation; the 3D scanner scans the workpiece and generates a 3D point cloud model of the workpiece, which is automatically compared with the theoretical dimensions of the design drawings to calculate the dimensional deviation value.

[0071] S3. The standardized multi-dimensional data output from the data acquisition module is uniformly sent to the AI ​​quality analysis module, where it undergoes three stages: data preprocessing, model matching and parallel inference, and weighted voting to complete the quality assessment, as detailed below: The data preprocessing unit performs differentiated feature extraction on three types of data: a 10-second sliding window is set for the time-series process parameters, and statistical features such as mean, variance, and peak value are extracted within the window; histogram equalization and Gaussian filtering are performed on the two-dimensional appearance image for enhancement and Gaussian filtering for noise reduction, and then the defect area is accurately divided by the U-Net semantic segmentation model; the ICP registration algorithm is used for the three-dimensional point cloud data to align the actual scanning model with the design theoretical model and automatically calculate various dimensional deviations.

[0072] After preprocessing, the model storage and retrieval unit automatically matches the corresponding model combination based on the workpiece type identified in the production order. Then, the three models are processed in parallel, and the result of passing or failing is output.

[0073] The quality result generation unit uses a weighted voting fusion mechanism to generate the final overall pass or fail determination of the workpiece.

[0074] After the final judgment output by the S4 AI quality analysis module is transmitted to the real-time control module, the module performs a three-level tiered handling based on the nature and severity of the problem: When a workpiece is deemed unqualified or the probability of abnormal process parameters exceeds a set threshold, the anomaly warning unit simultaneously triggers a triple warning mechanism. Warning information is bound to a unique workpiece identifier, facilitating rapid identification of problematic workpieces by management personnel. If a simple fluctuation in process parameters occurs, the system automatically generates a correction command and sends it to the PLC control system of the supporting equipment. If complex defects such as large-area deformation or deep cracks occur, or if the parameters fail to return to the standard range after multiple rounds of correction commands from the automatic parameter adjustment unit, the manual intervention scheduling unit automatically generates a manual intervention task order and pushes it to the handheld PDA terminals of on-site quality inspectors and maintenance personnel.

[0075] During steps S1 to S4, the system's traceability module archives all data generated by the workpiece at each production stage—including process parameters and appearance data acquired by the data acquisition module, judgment results and defect reports output by the AI ​​quality analysis module, and early warning logs and parameter adjustment intervention records generated by the real-time control module—into the data association storage unit and binds them to the traceability tag.

[0076] Workshop managers and quality inspectors can access the entire process history of a workpiece simply by scanning a code; managers can also search by batch, defect type, and other multi-dimensional conditions, and view parameter changes and quality results in a timeline format, and export a PDF report with one click.

[0077] Furthermore, the adaptive optimization module automatically reviews historical production data quarterly, performs a round of global optimization, selects effective training samples from the traceability module's database, and uses a gradient descent algorithm with anomaly identification accuracy or defect recall as the objective function to iteratively optimize the model parameters in the AI ​​quality analysis module. Simultaneously, a cost-benefit analysis model is built based on historical defect data and rework loss costs, dynamically adjusting the warning thresholds and parameter correction thresholds of the real-time control module to ensure the system continuously adapts to dynamic production environments such as raw material batch changes and equipment aging.

[0078] To make the technical solution of this invention clearer and easier to understand, the deployment method and workflow of the outfitting component production control system of this invention will be described in detail below with reference to specific embodiments, taking the actual production process of carbon steel pipeline outfitting components (specification: DN100, material Q235-B) in a ship assembly plant as an example: (a) Implementation Scenarios This embodiment focuses on the production process of carbon steel pipeline outfitting components, involving four processes: material cutting, bending, welding, and sandblasting. The production batch is 100 pieces, and the entire process is managed using the outfitting component production control system of this invention.

[0079] (II) System Deployment (1) Hardware deployment: ① Process parameter acquisition: A cutting temperature sensor (model PT100, measurement range: 50~400℃) and a cutting speed encoder (resolution 1000 pulses / revolution) are installed on the blanking machine; a pressure sensor (model CYB~20S, measurement range 0~1500kN) and an angle sensor (model WDD35D4, measurement range 0~360°) are installed on the bending machine; a current / voltage sensor (model CSM025, measurement range 0~500A / 0~50V) and a weld temperature sensor (model K thermocouple, measurement range 0~1000℃) are installed on the welding robot (model ABB IRB1410); a pressure sensor (model MPM480, measurement range 0~1MPa) and a coating thickness sensor (model TT220, measurement range 0~1000μm) are installed on the sandblasting machine.

[0080] ② Appearance data acquisition: One industrial camera (Hikvision MV-CA020-10GM) is deployed at the exit of the blanking process, two industrial cameras (top + side) are deployed at the exit of the bending process, three industrial cameras (top + both sides) and one 3D scanner (Keyence LK-G80) are deployed at the exit of the welding process, and one industrial camera and one 3D scanner are deployed at the exit of the sandblasting process.

[0081] ③ Edge computing and storage: Deploy an edge computing gateway (Advantech UNO-2484G) to perform data preprocessing; use a distributed server (CPU Intel Xeon Gold 6330, memory 64GB, hard disk 10TB) to deploy the database and three types of models.

[0082] (2) Software deployment The data acquisition module uses Advantech WebAccess software to achieve real-time acquisition and transmission of sensor data; the three models are developed based on the TensorFlow 2.8 framework, and the Random Forest (process parameter anomaly detection), improved YOLOv8 (appearance defect recognition), and SVM (dimensional accuracy determination) models are deployed on an NVIDIA A100 GPU server; the real-time control module and traceability software adopt a Spring Boot + Vue architecture, with the front end implementing data visualization and operation interface, and the back end completing data association and query logic.

[0083] (III) Work Process (1) Data acquisition stage (taking welding process as an example) After the welding robot starts, current / voltage sensors collect welding current and welding voltage in real time at a sampling frequency of 10Hz, while weld temperature sensors simultaneously collect weld temperature. At the same time, three industrial cameras at the welding process exit capture two-dimensional images of the weld area at 30fps, and a 3D scanner scans the pipe's inner diameter and wall thickness (scanning time 2 seconds per piece). The collected raw data is preprocessed directly via an edge computing gateway, removing outliers exceeding 500A or 40V and performing noise reduction preprocessing on the image data. The normalized data is then transmitted to the AI ​​quality analysis module.

[0084] (2) Quality analysis stage (taking welding process as an example) First, data preprocessing is performed. The data preprocessing unit extracts the mean (420A) and variance (5A) of the welding current within a 10-second sliding window. 2 The defect area of ​​the weld image is segmented, and the actual value of the pipe inner diameter (100.2 mm, theoretical value is 100 ± 0.3 mm) is calculated based on the 3D scanning data.

[0085] The model calling unit automatically matches and calls three corresponding models for parallel analysis based on the "carbon steel pipeline" type: the process parameter anomaly detection model outputs an abnormal current with a probability of 0.1 (less than the threshold of 0.8, considered normal); the appearance defect identification model detects one pore in the weld (area 2mm). 2 The defect was determined to be minor. The dimensional accuracy assessment model calculated that the inner diameter deviation of the pipeline was 0.2 mm (less than the allowable deviation of 0.3 mm, so it was deemed acceptable).

[0086] The quality result generation unit uses a weighted voting method to combine the three results and outputs a "qualified" judgment. The defect report records "minor porosity in the weld, which does not affect the performance of use".

[0087] (3) Real-time control phase Since the workpiece was deemed qualified, the abnormality warning unit did not trigger a warning, the automatic parameter adjustment unit did not need to send parameter adjustment instructions, and the manual intervention scheduling unit did not push any tasks.

[0088] After the system completes the welding quality assessment of the pipeline workpiece and archives and stores the information, it allows it to be transferred to the next sandblasting process. After the workpiece enters the sandblasting station, the system will again perform the data acquisition, quality analysis and real-time control operations corresponding to the sandblasting process.

[0089] Throughout the entire process of workpiece blanking, bending, welding, and sandblasting, the data association and storage unit continuously binds the data and images collected from the entire process of blanking, bending, welding, and sandblasting, as well as the quality inspection conclusions, to the QR code on the surface of the pipe fittings, and stores them synchronously in a hybrid distributed database of MySQL and MongoDB. It also completes the on-chain storage of key information such as defects and quality inspection on the blockchain.

[0090] Production and quality control personnel can scan the workpiece QR code with a handheld terminal at any stage of the pipe fitting production to retrieve the entire process data. The system intuitively displays the production parameters and quality inspection conclusions of each process in a timeline: blanking process (cutting speed of 300mm / min, processing temperature of 250℃) → bending process (pressure of 500kN, bending angle of 90°) → welding process (current of 420A, voltage of 25V) → sandblasting process (pressure of 0.5MPa, sandblasting time of 60s), presenting the complete production process.

[0091] The data filtering unit screened historical data from 1000 carbon steel pipes within the past three months from the traceability module (excluding 20 abnormal data points caused by equipment failure). The model parameter optimization unit analysis revealed that the appearance defect identification model was effective for defects smaller than 1 mm². 2 The recall rate for microcracks was only 90%. Therefore, the learning rate of YOLOv8 was reduced from 0.005 to 0.003 using the gradient descent algorithm, which improved the recall rate to 95%. Analysis of the control threshold optimization unit revealed that when the welding current exceeded 440A, the rework rate of subsequent processes increased from 3% to 8%. The current warning threshold of the process parameter anomaly detection model was reduced from 450A to 440A, which reduced the rework rate to 2%.

[0092] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A production control system for outfitting components based on artificial intelligence, characterized in that, include: The data acquisition module includes a process parameter acquisition unit and an appearance data acquisition unit. The process parameter acquisition unit includes several sensors arranged in each production process, which are used to acquire the process parameters of each production process in real time. The appearance data acquisition unit is used to acquire two-dimensional appearance images and three-dimensional point cloud data of the workpiece. The AI ​​quality analysis module includes a data preprocessing unit, a model storage and retrieval unit, and a quality result generation unit. The data preprocessing unit converts the process parameters, two-dimensional appearance images, and three-dimensional point cloud data detected by the sensors into standardized features. The model storage and retrieval unit stores a process parameter anomaly detection model, an appearance defect recognition model, and a dimensional accuracy judgment model, which are used to perform parallel calculations on each of the standardized features and output sub-item detection results. The quality result generation unit uses a weighted voting fusion mechanism to synthesize the sub-item detection results and generate the final quality conclusion for the workpiece. The real-time control module is used to receive the final quality conclusion and to proactively intervene in and handle abnormalities in the production process in a graded closed-loop manner. The real-time control module includes an abnormality early warning unit and an automatic parameter adjustment unit. The abnormality early warning unit is used to trigger an early warning when the final quality conclusion is unqualified or the process parameters are abnormal. The automatic parameter adjustment unit is used to automatically generate and issue process correction instructions to relevant production equipment for defects caused by fluctuations in the process parameters.

2. The outfitting component production control system based on artificial intelligence according to claim 1, characterized in that, Also includes: The traceability module includes a unique identifier generation unit, a data association storage unit, and a traceability query unit. The unique identifier generation unit equips each workpiece with a traceability identifier consisting of a QR code and a radio frequency identification chip. The data association storage unit uses MySQL and MongoDB databases, and uses the traceability identifier as an index to bind and archive all input and output data generated by the data acquisition module, the AI ​​quality analysis module and the real-time control module for a single workpiece. The traceability query unit is used to provide production operators with traceability query methods such as QR code query, multi-dimensional condition query, and visual timeline display.

3. The outfitting component production control system based on artificial intelligence according to claim 2, characterized in that, Also includes: An adaptive optimization module, comprising a historical data filtering unit, a model parameter optimization unit, and a control threshold optimization unit, wherein the historical data filtering unit is used to filter valid training samples from the MySQL database and the MongoDB database; The model parameter optimization unit uses the gradient descent algorithm with anomaly recognition accuracy or defect recall as the objective function to iteratively optimize the process parameter anomaly detection model, appearance defect recognition model, and dimensional accuracy determination model. The control threshold optimization unit is used to construct a cost-benefit analysis model by combining historical defect data and rework loss data, and dynamically adjust the early warning threshold and parameter correction threshold of the real-time control module.

4. The outfitting component production control system based on artificial intelligence according to claim 1, characterized in that, The process parameter acquisition unit also includes an edge computing gateway, which is deployed on the production workshop site to perform outlier removal and data normalization preprocessing on the raw data collected by the sensors, and output regular time-series feature data to the AI ​​quality analysis module.

5. The outfitting component production control system based on artificial intelligence according to claim 1, characterized in that, The appearance data acquisition unit includes multiple multi-view industrial camera arrays and multiple 3D scanners arranged in each of the production processes; the industrial cameras are used to capture full-coverage two-dimensional high-definition images of the workpiece and capture appearance defects of the workpiece. The 3D scanner is used to generate a three-dimensional point cloud model of the workpiece and to calculate the dimensional deviation between the workpiece and the standard design drawings.

6. The outfitting component production control system based on artificial intelligence according to claim 1, characterized in that, The data preprocessing unit includes: The timing process parameter processing subunit sets a sliding window, extracts the mean, variance, and peak value of the process parameters within the sliding window, and generates timing process features; A two-dimensional appearance image processing subunit is used to perform histogram equalization enhancement and Gaussian filtering denoising on the two-dimensional appearance image, and to divide the defect region of the two-dimensional appearance image. The three-dimensional point cloud data processing subunit uses the ICP algorithm to register the three-dimensional point cloud data, aligning the actual three-dimensional model of the workpiece with the design theoretical model, and automatically calculating the dimensional deviation value.

7. The outfitting component production control system based on artificial intelligence according to claim 6, characterized in that, The process parameter anomaly detection model is based on the random forest algorithm and outputs the probability of process parameter anomalies after inputting the process timing features; the appearance defect recognition model is based on the improved YOLOv8 algorithm and outputs the coordinate position, defect area, and defect severity type of the defect after inputting the two-dimensional appearance image; the dimensional accuracy determination model is based on the support vector machine algorithm and inputs the three-dimensional dimensional deviation value.

8. The outfitting component production control system based on artificial intelligence according to claim 1, characterized in that, The real-time control module also includes a manual intervention scheduling unit, which is configured to be triggered when automatic parameter adjustment fails to eliminate defects, automatically generate a task order containing a repair plan and push it to a handheld terminal on the production site.

9. A method for operating an outfitting component production control system based on artificial intelligence as described in any one of claims 1 to 8, characterized in that, Includes the following steps: The workpiece flows through each of the production processes in sequence, and the process parameters, the two-dimensional appearance image and the three-dimensional point cloud data are collected synchronously by the process parameter acquisition unit and the appearance data acquisition unit. After the process parameters, the two-dimensional appearance image, and the three-dimensional point cloud data are converted into standardized features, they are respectively fed into the process parameter anomaly detection model, the appearance defect identification model, and the dimensional accuracy judgment model for parallel reasoning, and then weighted voting fusion is used to generate the final quality conclusion of the workpiece. Based on the final quality conclusion, a graded handling procedure is implemented: when the process parameters are abnormal, an early warning is triggered; when the process parameters fluctuate, the process correction instruction is automatically sent to the relevant production equipment.

10. The operation method of the outfitting component production control system based on artificial intelligence according to claim 9, characterized in that, Also includes: When the workpiece enters the unloading process, a traceability identifier containing a QR code and an RFID chip is generated for each workpiece. The traceability identifier contains the workpiece model, batch, order and workstation information. The process data, quality analysis results, and handling records of each production process are uniformly archived and linked to the traceability identifier for production personnel to trace. At the same time, effective training samples are periodically selected from historical archived data, and the gradient descent algorithm is used to iteratively optimize the process parameter anomaly detection model, the appearance defect identification model, and the dimensional accuracy judgment model. The warning threshold and correction threshold are dynamically adjusted based on historical defect data and rework loss costs.

Images