Deep-water explosion vessel full life cycle health management method based on digital twinning

Abstract

This invention discloses a method for full lifecycle health management of deep-water explosive containers based on digital twins, belonging to the field of deep-water explosive container health management technology. Based on an updated digital twin and corresponding historical and real-time data, a machine learning model integrating LSTM, CNN, and SVM is established. The model assesses the health status of the deep-water explosive container based on input real-time multi-source data, outputting a health status assessment result including container health status classification and remaining life prediction. Based on the health status assessment result, combined with a preset rule engine and a Q-learning reinforcement learning model, an optimal maintenance strategy is generated, considering maintenance costs, downtime losses, and failure risks. This invention enables effective management of equipment through a full lifecycle health management system for deep-water explosive containers, significantly reducing the workload of engineering technicians in equipment maintenance. Relying on a large amount of design and usage data, the accuracy of assessing the health status of deep-water explosive containers is improved.

Description

Technical Field

[0001] This invention relates to the field of deep-sea explosion container health management technology, and in particular to a method for full life-cycle health management of deep-sea explosion containers based on digital twins. Background Technology

[0002] Traditional deep-water explosion containers lack systematic health management. Due to limitations in technology at the time, they suffer from multi-dimensional systemic defects, directly impacting their safety, reliability, and life-cycle cost. These defects primarily include: 1. Insufficient dynamic stress and fatigue damage monitoring: Deep-water explosion containers endure long-term high pressure, impact loads, and cyclic stress. Traditional designs rely on empirical formulas and static strength checks, lacking real-time stress monitoring systems. 2. Limited health monitoring methods: They rely heavily on manual inspections and periodic checks, with limited sensor deployment and data acquisition. Traditional containers only install a few pressure or temperature sensors in critical areas, monitoring parameters are limited and coverage is insufficient, failing to obtain real-time physical and mechanical parameters for all parts of the container. 3. Data silos and lack of intelligent analysis: Traditional health management relies on manual recording and offline data analysis, lacking a unified data platform. Data from the design and manufacturing stages is disconnected from the usage stage, making full-process traceability and optimization difficult. 4. Lack of preventative maintenance mechanisms: Traditional containers follow a "periodic maintenance" model, neglecting the dynamic changes in equipment health status. Furthermore, existing life predictions and remaining strength assessments are outdated, largely based on experience-based judgments, and lack precise analysis of the degradation patterns of container performance, resulting in high maintenance costs and significant safety hazards.

[0003] To address the aforementioned issues, this invention proposes a digital twin-based full lifecycle health management method for deep-water explosive containers. The concept of digital twins was first introduced by Professor Michael Grieves in his product lifecycle management course at the University of Michigan. It was initially applied to the health maintenance and support of aerospace vehicles, and subsequently, various cutting-edge technologies, including computer-aided design (CAD), finite element analysis (CAE), and structural vibration analysis, were applied to the study of real-world entities. Currently, digital twin technology is widely used in industrial manufacturing, commercial layout, and municipal management, demonstrating its significant value. This invention constructs a digital twin model of the deep-water explosive container, integrating real-time data and physical simulation to achieve remaining life prediction and maintenance strategy optimization; it also utilizes machine learning algorithms to identify early failure characteristics, improving early warning accuracy. The digital twin-based full lifecycle health management method proposed in this invention effectively solves the problem of the lack of systematic health management in traditional deep-water explosive containers, achieving data fusion and automated decision-making throughout the entire lifecycle of operation, maintenance, and decommissioning. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for full lifecycle health management of deep-sea explosion containers based on digital twins. The technical solution adopted is as follows: A digital twin-based approach to the full lifecycle health management of deep-sea explosion containers includes the following steps: Step 1: Collect point cloud data of the deep-water explosion container based on 3D laser scanning technology, and construct a digital twin of the deep-water explosion container based on the point cloud data; Step 2: Stress and strain sensors and acceleration sensors are installed on the sides and ends of the deep-water explosion container, and pressure sensors are installed inside the deep-water explosion container to collect physical and mechanical parameters in real time and form an automated monitoring data source. Step 3: Establish a manual inspection mechanism that combines regular and event-triggered inspections to obtain information on the surface condition and macroscopic defects of containers that cannot be covered by sensors, which will complement and verify the automated monitoring data from Step 2. Step 4: Construct a multi-source data acquisition and transmission network architecture to upload the physical and mechanical parameters from Step 2 and the inspection results from Step 3 to the server. Step 5: The server dynamically corrects and optimizes the digital twin based on real-time multi-source data to maintain the consistency between the state of the digital twin and the physical entity. Step 6: Based on the updated digital twin from Step 5 and the corresponding historical and real-time data, establish a machine learning model that integrates LSTM, CNN, and SVM. Based on the input real-time multi-source data, assess the health status of the deep-sea explosion container and output a health status assessment result that includes container health status classification and remaining life prediction. Step 7: Based on the health status assessment results, and combining the preset rule engine and Q-learning reinforcement learning model, generate the optimal maintenance strategy while considering maintenance costs, downtime losses and failure risks.

[0005] Optionally, step 1 may include the following sub-steps: Step 11: Remove rust, oil, and biological deposits from the surface of the deep-water explosion container; Step 12: Attach high-reflectivity circular targets to the surface of the container, with a target spacing of no more than 10cm; Step 13: Scan the container using a handheld 3D laser scanner, ensuring a surface point cloud density of no less than 50 points / cm². 2 The point cloud density in the welding area is not less than 100 points / cm². 2 ; Step 14: Use point cloud processing software to denoise, register and fuse the collected point cloud data, and use the Poisson reconstruction algorithm to reconstruct the triangular mesh to generate a surface mesh model. Step 15: Import the mesh model into CAD software, extract geometric features through feature recognition, and generate a parametric solid model; Step 16: Import the physical model into the 3D engine to build an interactive and updatable digital twin.

[0006] Optionally, in step 2, the stress-strain sensor and acceleration sensor are arranged on the outer surface of the container. One measuring point is arranged at the middle of each of the two hemispheres at the ends of the capsule-shaped container, and no less than six measuring points are evenly arranged circumferentially in the middle of the cylindrical surface on the side. The density of measuring points on the entire outer surface is no less than 4 per m. 2 ; At each measuring point on the outer surface, at least one stress sensor, one triaxial 60° strain rosette, and one accelerometer are arranged. Pressure sensors are placed inside the container, 5-15 cm from the inner wall. One measuring point is placed at the midpoint of each of the two hemispheres, and at least four measuring points are evenly distributed circumferentially at the midpoint of the cylindrical surface on the side. The density of measuring points on the entire inner surface is no less than 1 sensor / m². 2 .

[0007] Optionally, the manual inspection mechanism in step 3 is specifically as follows: When the deep-water explosion container is not in use, a regular inspection shall be carried out once a month. When the deep-water explosive container is in use, an event-triggered inspection shall be carried out before and after each use. The inspection record should include at least the container's sealing performance, the condition of rust on the inner and outer surfaces, and any visible cracks or defects.

[0008] Optionally, the multi-data acquisition and transmission network architecture constructed in step 4 is a three-level structure, including: Access layer: Consists of various sensors and manual inspection terminals, responsible for raw data collection; Aggregation layer: Using wired, Bluetooth, 4G or 5G wireless transmission methods, it is responsible for aggregating and uploading data from the access layer; Core layer: Deployed on the server, responsible for receiving, storing and integrating the data uploaded from the aggregation layer.

[0009] Optionally, the specific process of building the fusion machine learning model in step 6 includes the following sub-steps: Step 61: Obtain multi-sensor time-series data of the deep-water explosion container sample from its intact state to failure through laboratory accelerated degradation tests; Step 62: Preprocess the time series data and extract time-domain features, frequency-domain features, and time-frequency-domain features; Step 63: Use the time-domain and frequency-domain feature data to train the LSTM model, use the time-frequency domain feature data to train the CNN model, and use the structured feature data to train the SVM model. Step 64: Integrate the trained LSTM, CNN, and SVM models, and use a weighted fusion method to perform collaborative prediction of remaining lifetime on real-time monitoring data. Step 65: Classify the health status level according to the predicted remaining lifespan value: more than 100 times is normal, more than 50 times and less than or equal to 100 times is minor damage, more than 20 times and less than or equal to 50 times is moderate damage, and less than or equal to 20 times is severe damage.

[0010] Optionally, the preset rule engine in step 7 shall include at least the following rules: If the health status assessment result is severe damage, then maintenance action of immediately replacing parts is triggered; If the health status assessment result is moderate damage, then immediate repair action is triggered; If the health assessment indicates minor injury, an immediate examination will be initiated.

[0011] Optionally, the Q-learning reinforcement learning model in step 7 is constructed and run as follows: Define the state space S=(H,P,K), where H represents the health status level, P represents the production plan status, and K represents the spare parts inventory status. Action space A is defined as including: immediate full maintenance, planned off-peak maintenance, partial repair, no maintenance for the time being, and maintenance after emergency procurement; Design reward function ,in To maintain direct costs, To avoid downtime losses, Costs related to potential failure risks; By iteratively updating the Q-table, the system learns a strategy to select the optimal maintenance action under different states.

[0012] Optionally, it also includes: Step 8, constructing a visual human-computer interaction interface to display the digital twin model, real-time monitoring data, health status assessment results and maintenance decision suggestions, and providing data filtering, historical query and report generation functions.

[0013] Optionally, it also includes: Step 9, when the health status assessment results determine that the container cannot meet the usage requirements, initiate the scrapping process, and use a digital twin to record the container's disassembly process, component recycling status and scrapping reasons, to complete the full life cycle information closed loop.

[0014] In summary, the present invention has at least one of the following beneficial technical effects: it can effectively manage the equipment through the deep-water explosion container full life cycle health management system, which greatly reduces the workload of engineering technicians in equipment maintenance.

[0015] Relying on a large amount of design and usage data, the accuracy of assessing the health status of deep-water explosion containers has been improved.

[0016] This invention can effectively monitor, assess, predict, and make maintenance decisions for the entire life cycle of deep-water explosion containers, from design, manufacturing, use to disposal. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall architecture of the digital twin health management system used in this invention; Figure 2 A schematic diagram illustrating the modules for building a digital twin model; Figure 3 This is a schematic diagram of a multi-source data acquisition and transmission module; Figure 4 A schematic diagram of the health status assessment and prediction module; Figure 5 This is a schematic diagram of the decision-making system modules. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings.

[0019] This invention discloses a method for full life-cycle health management of deep-sea explosion containers based on digital twins.

[0020] Reference Figures 1-5 Example 1, a method for full lifecycle health management of deep-sea explosion containers based on digital twins, includes the following steps: Step 1: Collect point cloud data of the deep-water explosion container based on 3D laser scanning technology, and construct a digital twin of the deep-water explosion container based on the point cloud data; Step 2: Stress and strain sensors and acceleration sensors are installed on the sides and ends of the deep-water explosion container, and pressure sensors are installed inside the deep-water explosion container to collect physical and mechanical parameters in real time and form an automated monitoring data source. Step 3: Establish a manual inspection mechanism that combines regular and event-triggered inspections to obtain information on the surface condition and macroscopic defects of containers that cannot be covered by sensors, which will complement and verify the automated monitoring data from Step 2. Step 4: Construct a multi-source data acquisition and transmission network architecture to upload the physical and mechanical parameters from Step 2 and the inspection results from Step 3 to the server. Step 5: The server dynamically corrects and optimizes the digital twin based on real-time multi-source data to maintain the consistency between the state of the digital twin and the physical entity. Step 6: Based on the updated digital twin from Step 5 and the corresponding historical and real-time data, establish a machine learning model that integrates LSTM, CNN, and SVM. Based on the input real-time multi-source data, assess the health status of the deep-sea explosion container and output a health status assessment result that includes container health status classification and remaining life prediction. Step 7: Based on the health status assessment results, and combining the preset rule engine and Q-learning reinforcement learning model, generate the optimal maintenance strategy while considering maintenance costs, downtime losses and failure risks.

[0021] Optionally, step 1 may include the following sub-steps: Step 11: Remove rust, oil, and biological deposits from the surface of the deep-water explosion container; Step 12: Attach high-reflectivity circular targets to the surface of the container, with a target spacing of no more than 10cm; Step 13: Scan the container using a handheld 3D laser scanner, ensuring a surface point cloud density of no less than 50 points / cm². 2 The point cloud density in the welding area is not less than 100 points / cm². 2 ; Step 14: Use point cloud processing software to denoise, register and fuse the collected point cloud data, and use the Poisson reconstruction algorithm to reconstruct the triangular mesh to generate a surface mesh model. Step 15: Import the mesh model into CAD software, extract geometric features through feature recognition, and generate a parametric solid model; Step 16: Import the physical model into the 3D engine to build an interactive and updatable digital twin.

[0022] Optionally, in step 2, the stress-strain sensor and acceleration sensor are arranged on the outer surface of the container. One measuring point is arranged at the middle of each of the two hemispheres at the ends of the capsule-shaped container, and no less than six measuring points are evenly arranged circumferentially in the middle of the cylindrical surface on the side. The density of measuring points on the entire outer surface is no less than 4 per m. 2 ; At each measuring point on the outer surface, at least one stress sensor, one triaxial 60° strain rosette, and one accelerometer are arranged. Pressure sensors are placed inside the container, 5-15 cm from the inner wall. One measuring point is placed at the midpoint of each of the two hemispheres, and at least four measuring points are evenly distributed circumferentially at the midpoint of the cylindrical surface on the side. The density of measuring points on the entire inner surface is no less than 1 sensor / m². 2 .

[0023] Optionally, the manual inspection mechanism in step 3 is specifically as follows: When the deep-water explosion container is not in use, a regular inspection shall be carried out once a month. When the deep-water explosive container is in use, an event-triggered inspection shall be carried out before and after each use. The inspection record should include at least the container's sealing performance, the condition of rust on the inner and outer surfaces, and any visible cracks or defects.

[0024] Optionally, the multi-data acquisition and transmission network architecture constructed in step 4 is a three-level structure, including: Access layer: Consists of various sensors and manual inspection terminals, responsible for raw data collection; Aggregation layer: Using wired, Bluetooth, 4G or 5G wireless transmission methods, it is responsible for aggregating and uploading data from the access layer; Core layer: Deployed on the server, responsible for receiving, storing and integrating the data uploaded from the aggregation layer.

[0025] Optionally, the specific process of building the fusion machine learning model in step 6 includes the following sub-steps: Step 61: Obtain multi-sensor time-series data of the deep-water explosion container sample from its intact state to failure through laboratory accelerated degradation tests; Step 62: Preprocess the time series data and extract time-domain features, frequency-domain features, and time-frequency-domain features; Step 63: Use the time-domain and frequency-domain feature data to train the LSTM model, use the time-frequency domain feature data to train the CNN model, and use the structured feature data to train the SVM model. Step 64: Integrate the trained LSTM, CNN, and SVM models, and use a weighted fusion method to perform collaborative prediction of remaining lifetime on real-time monitoring data. Step 65: Classify the health status level according to the predicted remaining lifespan value: more than 100 times is normal, more than 50 times and less than or equal to 100 times is minor damage, more than 20 times and less than or equal to 50 times is moderate damage, and less than or equal to 20 times is severe damage.

[0026] Optionally, the preset rule engine in step 7 shall include at least the following rules: If the health status assessment result is severe damage, then maintenance action of immediately replacing parts is triggered; If the health status assessment result is moderate damage, then immediate repair action is triggered; If the health assessment indicates minor injury, an immediate examination will be initiated.

[0027] Optionally, the Q-learning reinforcement learning model in step 7 is constructed and run as follows: Define the state space S=(H,P,K), where H represents the health status level, P represents the production plan status, and K represents the spare parts inventory status. Action space A is defined as including: immediate full maintenance, planned off-peak maintenance, partial repair, no maintenance for the time being, and maintenance after emergency procurement; Design reward function ,in To maintain direct costs, To avoid downtime losses, Costs related to potential failure risks; By iteratively updating the Q-table, the system learns a strategy to select the optimal maintenance action under different states.

[0028] By adopting the above technical solutions, the core value of digital twins lies in achieving accurate mapping between physical entities and virtual models, providing a virtual carrier for full life-cycle health management.

[0029] First, removing rust, oil, and biological deposits from the container surface is crucial to prevent impurities from obscuring critical structures or creating noise points, ensuring the purity of subsequent point cloud data acquisition. High-reflectivity circular targets are then affixed to the container surface. The global positioning function of these targets helps resolve point cloud stitching discrepancies during multi-view scanning. Setting the target spacing to no more than 10cm ensures sufficient common targets for adjacent scanning views, improving stitching accuracy. A handheld 3D laser scanner is used, with a minimum target density of 50 targets per centimeter. 2 Surface point cloud density, not less than 100 points / cm 2 The point cloud density in the welding area is increased because the welding area is a stress concentration and easily damaged area, requiring denser data to restore details. Noise reduction, registration, and fusion using point cloud processing software can eliminate invalid data caused by environmental interference. The Poisson reconstruction algorithm can generate a continuous and smooth surface mesh model based on discrete point clouds, maximizing the restoration of the container's geometry. Importing the mesh model into CAD software to extract geometric features is to generate an editable parametric solid model to meet the needs of subsequent dynamic correction. Finally, it is imported into a 3D engine to build an interactive and updatable digital twin, realizing the visualization association and state synchronization between the virtual model and the physical entity.

[0030] Sensor placement must be based on the container's structural mechanical properties and monitoring requirements to ensure the comprehensiveness and relevance of physical and mechanical parameter acquisition. Stress-strain sensors and acceleration sensors are placed on the outer surface because it facilitates installation and allows for direct capture of the container's structural response under load. The midpoints of the two hemispherical ends and the midpoints of the cylindrical surfaces on the sides of the capsule-shaped container are stress concentration areas; therefore, these areas should have at least four measuring points per meter. 2The density of measuring points on the outer surface enables comprehensive monitoring of critical areas without blind spots. Each measuring point integrates a stress sensor, a triaxial 60° strain rosette, and an accelerometer to simultaneously acquire stress magnitude, multi-directional strain, and vibration impact data, comprehensively reflecting the local structural condition. Pressure sensors are positioned inside the container, 5-15 cm from the inner wall, accurately capturing the shock wave pressure generated by underwater explosions while avoiding sensor damage from direct contact with the inner wall. The arrangement and density requirements of the measuring points on the inner surface ensure the uniformity and effectiveness of internal pressure field monitoring. Real-time data acquisition through various sensors forms an automated monitoring data source, providing continuous and objective basic data for health assessment.

[0031] While sensor monitoring can acquire continuous physical and mechanical parameters, it has blind spots in monitoring non-quantitative indicators such as macroscopic defects and sealing performance. Manual inspection mechanisms can compensate for this limitation. Monthly regular inspections are conducted when there are no usage tasks to continuously monitor changes in the container's condition under normal circumstances. Event-triggered inspections are performed before and after use when there are usage tasks. This is because the container is subjected to explosive loads during use, which may cause instantaneous damage or sudden changes in condition. Pre-use inspections can identify potential hazards and prevent operation with faults, while post-use inspections can promptly detect new defects that arise during use. Inspection records are kept of the container's sealing performance, the condition of internal and external surface corrosion, and visible cracks and defects. This information is directly related to the safe operation of the container and complements the automated monitoring data, reducing errors from a single data source and improving data reliability.

[0032] Efficient transmission and integration of multi-source data (sensor data, inspection data) is a prerequisite for health management. The three-tier network architecture is designed following a logical division of labor: acquisition, aggregation, and processing. The access layer consists of various sensors and manual inspection terminals, enabling direct acquisition of different types of raw data and adapting to diverse data sources. The aggregation layer employs multiple transmission methods such as wired, Bluetooth, 4G, or 5G, allowing for flexible selection based on the usage environment. Wired transmission ensures data stability, while wireless transmission enhances deployment flexibility, ensuring efficient data aggregation in different scenarios. The core layer is deployed on servers, responsible for data reception, storage, and integration processing. Centralized processing improves data processing efficiency, enables unified management of multi-source data, and provides data support for subsequent digital twin updates and model analysis.

[0033] Physical entities are subject to factors such as explosive loads and environmental corrosion during use, causing their states to continuously change. If a digital twin remains in a fixed state, it will lose its mapping value. Servers dynamically correct and optimize the digital twin based on real-time multi-source data. Essentially, this involves data-driven synchronization of the virtual model's state with the physical entity's. Real-time data reflects the physical entity's current state. By inputting this data into the digital twin, the model's geometric parameters and mechanical properties are adjusted to ensure the virtual model accurately replicates the physical entity's structural state and performance changes, providing a reliable virtual simulation foundation for subsequent health assessments and predictions.

[0034] The core of health status assessment is to accurately identify the degradation patterns of containers and predict their remaining lifespan. The design that integrates LSTM, CNN, and SVM models is based on the complementary advantages of each algorithm. First, accelerating degradation tests in the laboratory to obtain full lifecycle data of the container from its intact state to failure allows for the simulation of the long-term degradation process of the container in a short time, solving the problem of difficulty in obtaining full lifecycle data in actual use. Preprocessing the time-series data and extracting time-domain, frequency-domain, and time-frequency-domain features is to condense the degradation patterns in the data, transforming high-dimensional raw data into low-dimensional features that are easier for the model to learn. LSTM models excel at processing time-series data and can capture the time trend of lifespan evolution, thus they are used to train time-domain and frequency-domain features. CNN models have significant advantages in feature extraction, effectively identifying abnormal signals in time-frequency domain features, and are suitable for the analysis needs of vibration and shock data. SVM models are suitable for processing structured feature data and can establish a mapping relationship between engineering parameters and health status. Using a weighted fusion approach for collaborative prediction can integrate the prediction results of various models, reduce the error of a single model, and improve prediction accuracy. Classifying health status levels based on remaining lifespan values ​​clarifies the risk level of different states according to safety thresholds in engineering practice, providing a clear basis for maintenance decisions.

[0035] The goal of maintenance decisions is to achieve the optimal balance between cost and efficiency while ensuring safety. A pre-defined rule engine establishes basic decision logic based on risk priority. In a severely damaged state, the container poses an extremely high safety hazard, requiring immediate replacement of parts to prevent further damage. Moderate damage, if not repaired promptly, may quickly deteriorate into severe damage, thus triggering immediate repair actions. Minor damage requires inspection to confirm the defect's development trend and prevent small defects from escalating into major failures; the rule engine ensures rapid response in emergencies. The Q-learning reinforcement learning model is used to optimize long-term maintenance strategies. Its state space encompasses health status, production plans, and spare parts inventory, comprehensively considering factors influencing decision-making. The action space provides multiple maintenance options to adapt to different scenario requirements. The reward function quantifies direct maintenance costs, downtime losses, and potential failure risks, guiding the model to select the action with the lowest cost and least risk. Through iterative updates to the Q-table, the model learns from experience, continuously optimizing the state-action mapping relationship to achieve long-term optimal maintenance decisions.

[0036] Optionally, it also includes: Step 8, constructing a visual human-computer interaction interface to display the digital twin model, real-time monitoring data, health status assessment results and maintenance decision suggestions, and providing data filtering, historical query and report generation functions.

[0037] Optionally, it also includes: Step 9, when the health status assessment results determine that the container cannot meet the usage requirements, initiate the scrapping process, and use a digital twin to record the container's disassembly process, component recycling status and scrapping reasons, to complete the full life cycle information closed loop.

[0038] By adopting the above technical solutions, the core value of the visualized human-computer interaction interface is to build an efficient communication bridge between users and the digital twin health management system, breaking down information transmission barriers between multi-source data, complex models, and user operational needs. Displaying the digital twin model is crucial because the virtual model can intuitively present the container's geometric structure, sensor deployment locations, and real-time status distribution, allowing users to quickly grasp the overall equipment status without relying on specialized technical knowledge. The visualization of real-time monitoring data transforms discrete physical and mechanical parameters into intuitive charts, facilitating users to quickly identify abnormal data fluctuations and capture trends in equipment status changes. Presenting health status assessment results and maintenance decision recommendations transforms the abstract conclusions of model analysis into clear text, level labels, and other easily understandable information, directly providing users with decision-making support and shortening response time.

[0039] The data filtering function is designed based on users' differentiated data needs in different scenarios. For example, maintenance personnel may need to focus on data from specific sensors or within a specific time period, while managers may focus on overall health trends. Filtering allows for precise location of the required information, improving data utilization efficiency. The historical query function retains monitoring data, evaluation results, and maintenance records throughout the entire lifecycle, providing data support for equipment status tracing, fault cause analysis, and maintenance effectiveness verification, helping users summarize experience and patterns. The report generation function integrates key information in a structured manner, forming standardized documents to meet the written record requirements of scenarios such as equipment management and compliance inspections, further enhancing the practicality and operability of the digital twin health management system. The overall interface design revolves around the principles of "intuitive, convenient, and accurate," lowering the barrier to entry for system use and ensuring that users in different roles can efficiently utilize the system's functions.

[0040] The scrapping process is initiated based on health status assessment results. The core objective is to scientifically determine whether the equipment has lost its safe operating capability, preventing safety accidents caused by equipment exceeding its service life or failing to meet performance standards, and ensuring the safety of the testing process and personnel. Using a digital twin to record the disassembly process is beneficial because it allows for full traceability of the disassembly sequence, key steps, and operational details. This provides standardized guidance for disassembly operations, preventing secondary damage to parts or waste of resources due to improper disassembly. Furthermore, the retained disassembly data can serve as a reference for optimizing disassembly solutions for similar equipment.

[0041] Recording the recycling status of parts is to achieve the rational recycling of resources. By clarifying the residual value of parts (such as direct reuse, reuse after repair, or material recycling), we can maximize resource utilization and reduce the total life cycle cost of equipment. Recording the reasons for scrapping equipment preserves and archives the key factors of equipment failure (such as fatigue damage, corrosion aging, structural defects, etc.), providing important feedback for the design, manufacturing, and maintenance of similar equipment in the future. This helps to optimize design schemes, improve manufacturing processes, and refine maintenance strategies, thereby reducing the occurrence of similar failure problems from the source.

[0042] The entire process utilizes digital twins to achieve complete retention of information throughout the entire lifecycle, forming a closed-loop management system of "design-manufacturing-use-maintenance-scrapping-feedback". This breaks the limitations of information gaps after traditional scrapping, allowing data from every stage of the entire lifecycle to play a valuable role and promoting continuous optimization of equipment health management.

[0043] The following specific embodiments illustrate the implementation principle of the present invention: Step 1: Remove rust, oil, and biofouling from the surface of the deep-water explosion container to avoid obscuring critical structures or generating noise points; attach high-reflectivity circular targets (7mm in diameter) to the container surface with a spacing of 5cm; collect point cloud data of the deep-water explosion container using a handheld 3D laser scanner MIRACO. This handheld 3D laser scanner is compatible with all sizes and has a single-frame accuracy of 0.02mm, meeting the requirements; preprocess and optimize the point cloud data using CloudCompare software, and perform "triangular mesh reconstruction" on the complete point cloud using the Poisson reconstruction algorithm with a mesh resolution of 0.1mm; import the mesh model into SolidWorks software, and automatically extract parametric features such as cylinders, planes, and holes through "feature recognition" to generate an editable CAD solid model; import the solid model into Unity to generate a digital twin model; Step 2: One measuring point should be arranged at the midpoint of each of the two hemispheres of the capsule-shaped deep-water explosion container, with measuring points arranged outwards from the midpoint at a density of 5 points / m². Six measuring points should be evenly arranged circumferentially at the midpoint of the columnar surface on the side, and evenly arranged radially at 50cm intervals per group. Each measuring point should be equipped with a stress sensor, a triaxial 60° strain rosette, and an acceleration sensor. The pressure sensor should be arranged inside the deep-water explosion container, 10cm from the inner wall. One measuring point should be arranged at the midpoint of each of the two hemispheres of the capsule-shaped deep-water explosion container, and four measuring points should be evenly arranged circumferentially at the midpoint of the columnar surface on the side, and evenly arranged radially at 1m intervals per group. The uTeKL dynamic signal analysis system should be used to export the collected acceleration and strain signals, and the Genesis high-speed underwater shock wave acquisition and analysis system should be used to export the pressure signals. Real-time monitoring of various physical and mechanical parameters of the deep-water explosion container should be achieved through the above monitoring equipment and systems. Step 3: Establish a manual inspection mechanism. Conduct a manual inspection once a month or before and after equipment use, and record the inspection results, such as whether the sealing performance of the deep-water explosion container is intact, whether there is rust inside and outside the tank, and whether there are cracks in the tank. Step 4: Construct a multi-dimensional data acquisition and transmission network architecture. This multi-dimensional data acquisition and transmission network architecture is divided into three levels: access layer, aggregation layer, and core layer. Reliable and effective physical and mechanical parameters are collected through the access layer. The data collected by the access layer is transmitted to the core layer quickly and efficiently through the aggregation layer using wired transmission. The core layer processes and analyzes the massive amount of data transmitted from the aggregation layer. Step 5, as follows Figure 2 As shown, the data-driven mechanism that uses the obtained real-time monitoring data to build a digital twin model is used to dynamically correct and optimize the digital twin by collecting multi-data of the deep-sea explosion container in real time during the use phase, forming an updatable and iterative digital twin model. Step 6: Conduct laboratory accelerated degradation experiments. Apply rated loads to the deep-water explosion container sample until failure. Define the remaining lifetime (RUL) of the deep-water explosion container at the failure time as 0, and the remaining lifetime (RUL) before failure as t-th time as t. Collect time-series data such as acceleration, strain, temperature, and pressure from the laboratory accelerated degradation experiment, as well as time characteristic data such as usage data and running time. The frequency of the time-series data must satisfy Nyquist's theorem (at least twice the highest frequency of the signal). Perform preprocessing on the time-series data, including denoising, missing value processing, normalization, or standardization. Divide the original time-series data from the laboratory accelerated degradation experiment into three types of features: time domain, frequency domain, and time-frequency domain. Divide the time-domain and frequency-domain data into training sets, test sets, and prediction sets for training the LTSM model. 1. The training set is for the early stage of structural degradation (RUL=10). The time-frequency domain data is divided into training, testing, and prediction sets for training the CNN model. Specifically, the training set consists of samples from the early stage of structural degradation (RUL=100%-30%), accounting for 70%; the validation set consists of samples from the mid-degradation stage (RUL=30%-10%), used for hyperparameter tuning, accounting for 20%; and the test set consists of samples from the late stage of degradation (RUL=10%-0%), used to simulate the critical scenario of "approaching failure" in actual prediction, accounting for 10%. The dataset consists of three sets of data: 1) a training set (70%) of samples in the early stage of structural degradation (RUL=100%-30%); 2) a validation set (20%) of samples in the middle stage of degradation (RUL=30%-10%), used for hyperparameter tuning; and 3) a test set (10%) of samples in the late stage of degradation (RUL=10%-0%), simulating the critical scenario of "near failure" in actual predictions. The remaining lifetime of the deep-water explosion container was predicted using LSTM, CNN, and SVM models in the time domain, frequency domain, and time-frequency domain of the most recent monitoring data. Health criteria were established: a weighted fusion of lifetime prediction results from the LSTM, CNN, and SVM models yielding more than 100 results indicates normal health; more than 50 but less than 100 results indicate slight damage; more than 20 but less than 50 results indicate moderate damage; and less than 20 results indicate severe damage. Step 7: Build a rule engine, whose rules include: 1. Severe damage requires immediate replacement of parts; 2. Moderate damage requires immediate repair; 3. Minor damage requires immediate inspection. Build a Q-learning model, including: 1. Define a state space, including equipment health status, production plan, and spare parts inventory. Equipment health status is categorized into H1 normal, H2 minor damage, H3 moderate damage, and H4 severe damage based on the remaining lifespan of the deep-sea explosion vessel in step 6. Production plan includes P1 peak production (e.g., peak season, usage period) and P2 off-peak (e.g., off-season, planned maintenance window). Spare parts inventory includes S1 sufficient spare parts (required parts inventory ≥ 1) and S2 spare parts shortage (required parts inventory = 0). Combine these dimensions into a discrete state, s = (H, P, S). 2. Define the action space, including: A1 Immediate shutdown for comprehensive maintenance (replace all aging parts and restore health to H1); A2 Planned off-peak maintenance (perform comprehensive maintenance during P2 hours to avoid peak downtime losses); A3 Partial repair (replace only critical degraded parts to improve health to H2); A4 No maintenance for now (continuous monitoring, as health may deteriorate over time); A5 Emergency procurement + maintenance (procure parts first and then perform maintenance when spare parts are scarce, incurring additional procurement costs). 3. Design the reward function, the reward function is as follows: ,in To maintain direct costs, To avoid downtime losses, Costs related to potential failure risks; 4. Initialize and update the Q-table, including: 1. Q-table initialization: The Q-table is a matrix of [number of states × number of actions]. The initial value is set to 0 or a small random number, indicating that the value of "state-action" is unknown at the beginning. 2. Q-table update: Through interaction with the environment (or simulation), the Q-value is iteratively updated according to the following formula: Q(s,a)←Q(s,a)+α[r+γ⋅maxa'Q(s',a')-Q(s,a)]; where, s: current state; a: action chosen in state s; r: immediate reward obtained after performing action a; s': new state transitioned to after performing action a; maxa'Q(s',a'): maximum Q-value of all possible actions in the new state s'; α (learning rate, 0<α<1): controls the impact of "new information" on the old Q-value; γ (discount factor, 0<γ<1): balances "immediate reward" and "future reward". A maintenance decision model based on a rule engine and reinforcement learning is constructed to form a decision system. The method is to determine whether the decision method for the current state is included in the rule engine. If it is, the decision is made directly based on the rule engine; otherwise, the decision system makes a decision based on the Q table and returns the solution effect under the decision to the Q learning program to further update the Q table. Step 8: Establish a visual human-computer interaction interface. Important content is displayed intuitively in the form of graphs, tables, reports, etc., improving the operability of the system. Its main functions include: 1. Allowing users to view detailed information of different modules through clicks, zooms, etc. Provides filtering and search functions, allowing users to view data based on conditions such as time, sensor location, and the service life of deep-water explosion containers. 2. Integrating charts and maps into reports or providing them to users through a web interface. Specifically, for charts about physical quantity measurements, the X-axis represents time, and the Y-axis represents the specific physical quantity; for charts about predicting the remaining service life of deep-water explosion containers, the X-axis represents the number of times the container is used (nth time), and the Y-axis represents the predicted remaining service life at the nth use (nth time). 3. Providing a download function, allowing users to download charts and raw data; Step 9: During the usage phase, when an abnormal state of the equipment is detected or a potential fault is predicted, the system will automatically trigger an early warning mechanism (including audible and visual alarms, SMS alarms, etc.) and provide corresponding emergency handling solutions and maintenance suggestions. Step 10: During the scrapping stage, when the equipment is found to be unable to meet the test requirements, it is determined that the equipment is scrapped. After scrapping, digital twin technology is used to record the disassembly process of the container and the recycling status of the parts, so as to realize closed-loop management of the entire life cycle.

[0044] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for full life-cycle health management of deep-sea explosion containers based on digital twins, characterized in that, Includes the following steps: Step 1: Collect point cloud data of the deep-water explosion container based on 3D laser scanning technology, and construct a digital twin of the deep-water explosion container based on the point cloud data; Step 2: Stress and strain sensors and acceleration sensors are installed on the sides and ends of the deep-water explosion container, and pressure sensors are installed inside the deep-water explosion container to collect physical and mechanical parameters in real time and form an automated monitoring data source. Step 3: Establish a manual inspection mechanism that combines regular and event-triggered inspections to obtain information on the surface condition and macroscopic defects of containers that cannot be covered by sensors, which will complement and verify the automated monitoring data from Step 2. Step 4: Construct a multi-source data acquisition and transmission network architecture to upload the physical and mechanical parameters from Step 2 and the inspection results from Step 3 to the server. Step 5: The server dynamically corrects and optimizes the digital twin based on real-time multi-source data to maintain the consistency between the state of the digital twin and the physical entity. Step 6: Based on the updated digital twin from Step 5 and the corresponding historical and real-time data, establish a machine learning model that integrates LSTM, CNN, and SVM. Based on the input real-time multi-source data, assess the health status of the deep-sea explosion container and output a health status assessment result that includes container health status classification and remaining life prediction. Step 7: Based on the health status assessment results, and combining the preset rule engine and Q-learning reinforcement learning model, generate the optimal maintenance strategy while considering maintenance costs, downtime losses and failure risks.

2. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 1, characterized in that, Step 1 specifically includes the following sub-steps: Step 11: Remove rust, oil, and biological deposits from the surface of the deep-water explosion container; Step 12: Attach high-reflectivity circular targets to the surface of the container, with a target spacing of no more than 10cm; Step 13: Scan the container using a handheld 3D laser scanner, ensuring a surface point cloud density of no less than 50 points / cm². 2 The point cloud density in the welding area is not less than 100 points / cm². 2 ; Step 14: Use point cloud processing software to denoise, register and fuse the collected point cloud data, and use the Poisson reconstruction algorithm to reconstruct the triangular mesh to generate a surface mesh model. Step 15: Import the mesh model into CAD software, extract geometric features through feature recognition, and generate a parametric solid model; Step 16: Import the physical model into the 3D engine to build an interactive and updatable digital twin.

3. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 2, characterized in that, In step 2, stress-strain sensors and acceleration sensors are arranged on the outer surface of the container. One measuring point is arranged at the middle of each of the two hemispheres of the capsule-shaped container, and no less than 6 measuring points are evenly arranged circumferentially in the middle of the cylindrical surface on the side. The density of measuring points on the entire outer surface is no less than 4 per m. 2 ; At each measuring point on the outer surface, at least one stress sensor, one triaxial 60° strain rosette, and one accelerometer are arranged. Pressure sensors are placed inside the container, 5-15 cm from the inner wall. One measuring point is placed at the midpoint of each of the two hemispheres, and at least four measuring points are evenly distributed circumferentially at the midpoint of the cylindrical surface on the side. The density of measuring points on the entire inner surface is no less than 1 sensor / m². 2 .

4. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 3, characterized in that, The manual inspection mechanism in step 3 is specifically as follows: When the deep-water explosion container is not in use, a regular inspection shall be carried out once a month. When the deep-water explosive container is in use, an event-triggered inspection shall be carried out before and after each use. The inspection record should include at least the container's sealing performance, the condition of rust on the inner and outer surfaces, and any visible cracks or defects.

5. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 4, characterized in that, The multi-data acquisition and transmission network architecture constructed in step 4 is a three-level structure, including: Access layer: Consists of various sensors and manual inspection terminals, responsible for raw data collection; Aggregation layer: Using wired, Bluetooth, 4G or 5G wireless transmission methods, it is responsible for aggregating and uploading data from the access layer; Core layer: Deployed on the server, responsible for receiving, storing and integrating the data uploaded from the aggregation layer.

6. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 5, characterized in that, Step 6, which involves building a fusion machine learning model, includes the following sub-steps: Step 61: Obtain multi-sensor time-series data of the deep-water explosion container sample from its intact state to failure through laboratory accelerated degradation tests; Step 62: Preprocess the time series data and extract time-domain features, frequency-domain features, and time-frequency-domain features; Step 63: Use the time-domain and frequency-domain feature data to train the LSTM model, use the time-frequency domain feature data to train the CNN model, and use the structured feature data to train the SVM model. Step 64: Integrate the trained LSTM, CNN, and SVM models, and use a weighted fusion method to perform collaborative prediction of remaining lifetime on real-time monitoring data. Step 65: Classify the health status level according to the predicted remaining lifespan value: more than 100 times is normal, more than 50 times and less than or equal to 100 times is minor damage, more than 20 times and less than or equal to 50 times is moderate damage, and less than or equal to 20 times is severe damage.

7. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 6, characterized in that, Step 7 requires the preset rule engine to include at least the following rules: If the health status assessment result is severe damage, then maintenance action of immediately replacing parts is triggered; If the health status assessment result is moderate damage, then immediate repair action is triggered; If the health assessment indicates minor injury, an immediate examination will be initiated.

8. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 7, characterized in that, In step 7, the Q-learning reinforcement learning model is constructed and run as follows: Define the state space S=(H,P,K), where H represents the health status level, P represents the production plan status, and K represents the spare parts inventory status. Action space A is defined as including: immediate full maintenance, planned off-peak maintenance, partial repair, no maintenance for the time being, and maintenance after emergency procurement; Design reward function ,in To maintain direct costs, To avoid downtime losses, Costs related to potential failure risks; By iteratively updating the Q-table, the system learns a strategy to select the optimal maintenance action under different states.

9. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 8, characterized in that, Also includes: Step 8: Construct a visual human-computer interaction interface to display the digital twin model, real-time monitoring data, health status assessment results and maintenance decision suggestions, and provide data filtering, historical query and report generation functions.

10. The method for full life-cycle health management of deep-sea explosion containers based on digital twins according to claim 9, characterized in that, Also includes: Step 9: When the health status assessment results determine that the container cannot meet the usage requirements, the scrapping process is initiated, and the dismantling process, component recycling status and scrapping reasons are recorded using a digital twin to complete the full life cycle information loop.

Images