Intelligent assistance method and system for port state control (PSC) inspection of liquefied natural gas (LNG) carrier

Abstract

This invention relates to the interdisciplinary field of maritime safety supervision in transportation and artificial intelligence, and discloses an intelligent auxiliary method and system for Port State Control (PSC) inspections of liquefied natural gas (LNG) vessels. The method includes structured processing and relational mapping of multi-source basic data, semantic storage of maritime regulatory texts in parent-child segmentation, and generation of a specialized knowledge base data with a five-dimensional relational mapping network. Responding to the identification information of the vessel to be inspected, the method calls upon the specialized knowledge base data and performs calculations using a risk analysis model combining a rule engine and machine learning classification to generate personalized inspection checklist data labeled with risk level, high-risk inspection areas, and inspection priorities. Through the synergy of the method and system, the method addresses pain points such as inefficient regulatory retrieval for LNG vessel PSC inspections, reliance on experience for defect determination, and cumbersome report generation, significantly improving inspection efficiency and accuracy, standardizing the inspection process, and enabling personalized training for professionals and continuous system optimization.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of maritime safety supervision and artificial intelligence, specifically to an intelligent auxiliary method and system for port state control (PSC) inspections of liquefied natural gas (LNG) vessels. Background Technology

[0002] The global liquefied natural gas (LNG) maritime transport industry continues to expand. Due to the inherent risks of LNG vessels—low temperature, high pressure, and flammability and explosiveness—their navigation safety and maritime regulation have become a key focus in the maritime transportation sector. Port State Control (PSC), as a crucial regulatory tool to ensure vessels comply with international maritime standards, is the last line of defense in preventing maritime safety accidents involving LNG vessels and protecting the marine environment. The accuracy, efficiency, and standardization of PSC inspections directly impact the safety of life and property at sea and the protection of the marine ecosystem, placing higher demands on the professionalism and intelligence of PSC inspections for LNG vessels.

[0003] The current LNG vessel PSC inspection process remains largely manual, failing to deeply integrate maritime regulatory knowledge, historical inspection data, and modern information technologies such as big data, artificial intelligence, and 3D visualization. The entire inspection process relies excessively on the personal experience, memory, and manual operation of port state control inspectors, resulting in industry pain points such as "fragmented and unconnected data, unsupported knowledge application, lack of intelligent inspection operations, and no feedback during process iterations." This core issue directly leads to the inability to quickly generate targeted, personalized inspection checklists before inspections, inefficient regulatory searches during inspections, a lack of scientific basis for defect determination without visual and intelligent assistance, cumbersome and low-standardized report generation after inspections, and the inability to effectively feedback data from the entire inspection process to optimize subsequent inspections. Ultimately, this severely restricts the regulatory effectiveness of LNG vessel PSC inspections and fails to meet the current regulatory needs of the LNG maritime transport industry. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes an intelligent auxiliary method and system for Port State Control (PSC) inspections of liquefied natural gas (LNG) vessels, which solves the aforementioned technical problems.

[0005] Firstly, it provides intelligent auxiliary methods for Port State Control (PSC) inspections of liquefied natural gas (LNG) vessels, including: Obtain multi-source basic data for port state control inspections of liquefied natural gas vessels; The multi-source basic data is structured and associated with each other. The maritime regulations text is stored in a parent-child segmented semantic manner. Based on the stored structured regulations data, a special knowledge base is constructed that includes a five-dimensional association mapping network of inspection locations, defect codes, legal provisions, action codes, and defect phenomenon descriptions. In response to the identification information of the vessel to be inspected, the specialized knowledge base data is invoked, and a risk analysis model combining rule engine and machine learning classification is used to calculate and generate personalized inspection list data marked with risk level, high-risk inspection areas and inspection priority. Based on the personalized inspection checklist data, the inspection items are spatially bound to the 3D model of the liquefied natural gas vessel to generate visual inspection navigation data; During the on-site inspection, in response to the query command triggered by the visual inspection navigation data, the large language model, which has been fine-tuned by the maritime professional corpus, is called. The legal knowledge engine and the historical case engine are simultaneously triggered to perform dual-engine retrieval. After fusion reasoning, the structured inspection auxiliary response data is output. The defect data and evidence materials collected during on-site inspections are structured and automatically associated with the five-dimensional correlation mapping network to generate bilingual (Chinese and English) inspection report data that conforms to the Tokyo Memorandum of Understanding standards. The system integrates structured data from the entire inspection process, including visual inspection navigation call records, structured inspection auxiliary response data, and standardized inspection report data, as feedback to update the specialized knowledge base data and optimize the parameters of the risk analysis model and the large language model.

[0006] Furthermore, the multi-source basic data undergoes structured processing and association mapping; the maritime regulatory text is stored semantically in parent-child segmentation; and a specialized knowledge base is constructed based on the stored structured regulatory data, comprising a five-dimensional association mapping network including inspection locations, defect codes, regulatory clauses, action codes, and defect phenomenon descriptions. Collect historical inspection records of members of the Tokyo Memorandum, maritime law texts, professional knowledge data, and training question bank data as multi-source basic data; A semantic segmentation algorithm is used to divide the legal text into independent semantic units, using the chapter-section-article-clause numbering and preset logical connectors as semantic boundaries. The segmented semantic units are converted into a structured markup language format, and a parent-child segmentation architecture is used to complete the semantic storage of maritime regulations text; Based on data mining technology, a five-dimensional relational mapping network is constructed, which includes inspection locations, defect codes, legal and regulatory basis, action codes, and natural language descriptions of typical defect phenomena. The five-dimensional relational mapping network is calibrated through a collaborative mechanism of large model formatting, regularization correction, and manual verification to generate specialized knowledge base data.

[0007] Furthermore, in response to the identification information of the vessel to be inspected, the specialized knowledge base data is invoked, and a risk analysis model combining a rule engine and machine learning classification is used to calculate and generate personalized inspection checklist data labeled with risk level, high-risk inspection areas, and inspection priorities, including: Based on the IMO number or ship name identification information of the vessel to be inspected, the historical inspection records of the corresponding vessel, the flag state to which it belongs, and the statistical defect data of the fleet are retrieved from the special knowledge base. Rigid risk assessment is performed through a preset rule engine. The rigid risk assessment includes directly determining high risk if there are residual records within a preset period, upgrading the risk if there are recurrence records of the same inspection defect, and determining high-risk inspection areas if there are cumulative high-frequency defects in the same area. A random forest machine learning classifier trained on historical inspection data is used to comprehensively calculate the ship risk score by weighting factors such as the frequency of defect recurrence in the same part, the severity of historically detained defects, the flag state of the ship and the statistical defect rate of the fleet, and output the corresponding risk level result. Based on the risk level results and high-risk inspection areas, personalized inspection checklist data is generated by prioritizing high-risk areas.

[0008] Furthermore, based on the personalized inspection checklist data, the inspection items are spatially linked to the 3D model of the liquefied natural gas vessel to generate visual inspection navigation data, including: Load a general 3D model of a liquefied natural gas vessel onto a mobile terminal and bind the inspection items in the personalized inspection checklist to their corresponding spatial locations in the 3D model one by one; Based on the bound 3D model, a visual inspection navigation path that supports touch zoom and rotation operations is generated; In response to trigger commands for the inspection area in the 3D model, the system outputs the basic information, inspection points, and relevant regulatory data of the corresponding inspection area in real time. High-risk area data and lightweight model parameters are cached locally on the mobile terminal, and the cached data is called to provide inspection and navigation services in weak network environments.

[0009] Furthermore, during on-site inspections, in response to query commands triggered by visual inspection navigation data, a large language model fine-tuned from a maritime professional corpus is invoked. Simultaneously, a dual-engine retrieval process is initiated using both the regulatory knowledge engine and the historical case engine. After fusion reasoning, structured inspection auxiliary response data is output, including: Perform intent recognition and entity extraction on received natural language or voice query commands triggered by the visual inspection navigation interface, and extract search keywords; The system utilizes a large language model finely tuned with professional maritime corpus and historical question-and-answer pairs to trigger a dual-engine search based on search keywords. Specifically, the regulatory knowledge engine performs precise matching of regulatory clauses based on a specialized knowledge base, while the historical case engine performs similarity matching of similar defect cases based on a historical inspection database. The system integrates and reasons the regulatory data obtained from dual-engine retrieval with historical case data, and outputs structured response data that includes defect codes, legal basis in Chinese and English, graded handling suggestions, and corresponding action codes.

[0010] Furthermore, the defect data and evidence materials collected during on-site inspections are structured and automatically associated with the five-dimensional relational mapping network to generate bilingual (Chinese and English) inspection report data conforming to the Tokyo Memorandum of Understanding standards, including: The defect classification information, text descriptions, images and video evidence materials collected during the on-site inspection based on the visual inspection navigation path are structured and organized to obtain a standardized defect dataset; Automatically associate and match the standardized defect dataset with the five-dimensional relationship mapping network in the specialized knowledge base to supplement and improve the corresponding data of defect codes, legal basis, and action codes; The matched full data will be automatically populated into the standard FORM A and FORM B report templates of the Tokyo Memorandum. The report content is automatically switched to Chinese and English based on a maritime bilingual terminology database, generating inspection report data in both Chinese and English.

[0011] Furthermore, structured data from the entire inspection process, including visual inspection navigation call records, structured inspection auxiliary response data, and standardized inspection report data, is integrated as feedback to update the specialized knowledge base data. Parameter optimization is also performed on the risk analysis model and the large language model, including: Collect structured feedback data throughout the entire inspection process. The feedback data includes defect information, handling results, differences between AI suggestions and official handling data, and correlation data of newly emerging defects. According to the preset first cycle, the specialized knowledge base and the five-dimensional relational mapping network are incrementally updated through a process of automatic formatting, regular expression correction, and manual verification. According to the preset second cycle, based on the collected feedback data, the weighted parameters of the risk analysis model and the large language model are fine-tuned through reinforcement learning.

[0012] Furthermore, it also includes training and assessment steps based on the specialized knowledge base data and the 3D model, the training and assessment steps including: Based on a personalized inspection checklist or a specified high-risk inspection scenario, the corresponding 3D model inspection parts and related inspection items are dynamically loaded to generate a 3D simulation inspection task, complete the full-process simulation training of the inspection, and generate a simulation inspection record. Based on the training question bank data of label classification, special assessment test papers are generated by adopting adaptive test paper generation or scenario-based targeted test paper generation strategies. Among them, adaptive test paper generation is based on matching questions with students' historical wrong questions and knowledge weaknesses, while scenario-based targeted test paper generation is based on matching questions with specific ship risk profiles or port high-frequency defect types. The assessment results are quantitatively scored, and a training effectiveness evaluation report and targeted reinforcement suggestions are generated, linking the corresponding learning materials in the specialized knowledge base.

[0013] Secondly, an intelligent auxiliary system for port state control (PSC) inspections of liquefied natural gas (LNG) vessels is provided, based on the intelligent auxiliary method for PSC inspections of LNG vessels described in any of the preceding paragraphs, including: A specialized knowledge base module is used to store and manage the structured semantic units, historical inspection cases, professional knowledge, and the multidimensional association mapping network; The intelligent judgment and service engine integrates the risk analysis model and the natural language processing model to provide computing and reasoning services for the online integrated management platform and mobile intelligent inspection terminal. An online integrated management platform is used for checking task management, data analysis, knowledge base maintenance, and training assessment. The mobile intelligent inspection terminal integrates 3D visualization inspection navigation, AI assistant, and report generation functions for on-site operations. The data closure and iteration module automatically collects data during the inspection process and the correlation data of newly emerging defects. It updates the specialized knowledge base and five-dimensional correlation mapping network monthly through an automatic formatting-regular expression correction-manual verification process. It also fine-tunes the risk analysis model parameters and natural language processing model quarterly based on the collected data. The online integrated management platform and the mobile intelligent inspection terminal collaborate through a data interface, sharing and calling the specialized knowledge base module and intelligent service engine.

[0014] Furthermore, the mobile intelligent inspection terminal has edge computing capabilities, and in weak network environments, it can provide basic inspection navigation and question-and-answer services based on locally cached high-risk area data and lightweight models.

[0015] The invention employing the above technical solution has the following advantages: This invention utilizes a risk profiling model based on massive historical inspection data (e.g., 9000+ vessel inspections) to automatically generate personalized inspection checklists focusing on high-risk areas of specific vessels. It also incorporates a 3D electronic map of the inspection areas for navigation between the checklist and the inspection locations, transforming the traditional inspection model that relies on manual memorization and review. Practical results show that this solution significantly reduces the overall time required for a single inspection, greatly improving efficiency in pre-inspection data retrieval and preparation, as well as post-inspection document processing and report generation. Specific practical data demonstrates that by following the professional procedures for LNG vessel PSC inspections, the average inspection time for a single LNG vessel PSC can be reduced by 1-2 hours, data retrieval efficiency is increased by over 30%, and document processing time is reduced by over 50%; knowledge retrieval response time is ≤1 second, far superior to traditional manual retrieval methods.

[0016] This invention employs a dual-engine AI-assisted real-time decision-making mechanism driven by "precise matching of regulatory clauses" and "matching of historical case similarity," and relies on a deeply structured five-dimensional relational mapping knowledge base to provide Port State Control Officers (PSCOs) with immediate and authoritative defect determination support. This mechanism effectively reduces the risk of misuse of defect codes and incomplete or inaccurate citation of regulatory basis due to differences in personal experience or memory bias, ensuring the objectivity, consistency, and authority of inspection conclusions.

[0017] 3. This invention pioneers a "parent-child segmentation" semantic storage method for maritime regulatory texts and constructs a multi-dimensional relational mapping network of "inspection location - defect code - regulatory clause - action code - defect phenomenon description". By combining a collaborative construction process of large language model, rule engine and expert verification, the originally unstructured and difficult-to-use complex regulatory texts and case data are transformed into a semantic knowledge system that can be deeply understood, retrieved and reasoned by machines, fundamentally solving the core problem of knowledge management and application in the maritime field.

[0018] 4. The system of this invention integrates a simulated inspection environment using a general LNG ship 3D model and supports the dynamic generation of training tasks and assessment content based on ship risk profiles or specific defect scenarios. Through an intelligent strategy combining adaptive test paper generation (targeting trainees' weaknesses) and scenario-based targeted test paper generation (targeting high-frequency risk points), it provides PSCOs with highly immersive, practical, and personalized skills training and assessment methods, effectively accelerating the cultivation and maintenance of professional talent. Practice shows that after training, PSCOs' mastery of professional knowledge and regulatory provisions improves by more than 40%, and their practical skills are significantly enhanced.

[0019] 5. This invention, through the design of a closed-loop feedback mechanism for inspection data, enables the system to automatically collect new defect phenomena, handling results, and discrepancies with AI suggestions generated during on-site inspections. This data is then used to periodically drive updates to the specialized knowledge base and optimization of the core analytical model's parameters. This allows the system to continuously learn from accumulated regulatory practices, and its intelligence level and application effectiveness can iteratively improve over time.

[0020] 6. The system architecture of this invention fully considers the complex operational environment of onboard inspections. The mobile terminal can be tailored to the actual inspection needs of PSCOs to ensure the operation of core functions during onboard inspections. Meanwhile, the entire system adopts a modular and loosely coupled design. Its core knowledge base construction method, risk analysis engine, and intelligent auxiliary framework are easily adapted and extended to other types of vessels' PSC inspections or shipping company safety management and other maritime regulatory scenarios, demonstrating good potential for widespread application. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the specific embodiments will be briefly described below. In all the drawings, the elements or parts are not necessarily drawn to scale.

[0022] Figure 1 The present invention provides a process for the intelligent auxiliary method of Port State Control (PSC) inspection for liquefied natural gas (LNG) vessels. Figure 1 ; Figure 2 This is a functional framework diagram of the intelligent auxiliary system for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to the present invention. Figure 3 This is a block diagram of the overall architecture of the intelligent auxiliary method and system for port state control (PSC) inspection of liquefied natural gas (LNG) ships of the present invention. Figure 4 This is a flowchart of the knowledge base update process in the intelligent auxiliary method and system for port state control (PSC) inspection of liquefied natural gas (LNG) vessels of the present invention. Figure 5 This is a schematic diagram of the entire process of generating a ship risk profile and conducting intelligent inspections in the intelligent auxiliary method and system for port state control (PSC) inspections of liquefied natural gas (LNG) ships, as described in this invention. Figure 6 This is a schematic diagram illustrating the working principle of the AI ​​assistant's dual-engine retrieval and reasoning in the intelligent auxiliary method and system for port state control (PSC) inspection of liquefied natural gas (LNG) vessels of the present invention. Figure 7 This invention provides a flowchart of a simulated inspection and training assessment process based on a three-dimensional model within the intelligent auxiliary method and system for port state control (PSC) inspections of liquefied natural gas (LNG) vessels. Figure 8This is a schematic diagram of the structured storage of the "parent-child segmentation" regulatory knowledge base in the intelligent auxiliary method and system for port state control (PSC) inspection of liquefied natural gas (LNG) ships of the present invention. Figure 9 This is a schematic diagram of the knowledge base management interface in the online integrated management platform of the intelligent auxiliary method and system for port state control (PSC) inspection of liquefied natural gas (LNG) vessels of the present invention. Figure 10 This is a schematic diagram of the three-dimensional panoramic map inspection interface in the mobile intelligent inspection terminal of the intelligent auxiliary method and system for port state control (PSC) inspection of liquefied natural gas (LNG) ships of the present invention. Figure 11 This is a flowchart illustrating the process from task creation to completion in the intelligent auxiliary method and system for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to the present invention. Figure 12 This is a schematic diagram of the inspection report generation interface and the Chinese-English bilingual switching effect in the intelligent auxiliary method and system for port state control (PSC) inspection of liquefied natural gas (LNG) vessels of the present invention. Figure 13 The present invention provides a process for the intelligent auxiliary method of Port State Control (PSC) inspection for liquefied natural gas (LNG) vessels. Figure 2 . Detailed Implementation

[0023] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0024] like Figures 1 to 13 As shown, the intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels of the present invention includes: Step S01: Obtain multi-source basic data for port state control inspections of liquefied natural gas vessels; Step S02: Perform structured processing and association mapping on multi-source basic data, perform parent-child segmented semantic storage on maritime regulations text, and construct a special knowledge base data based on the stored structured regulations data, which includes a five-dimensional association mapping network of inspection locations, defect codes, legal clauses, action codes, and defect phenomenon descriptions. Step S03: In response to the identification information of the vessel to be inspected, call the specialized knowledge base data, and calculate through a risk analysis model that combines rule engine and machine learning classification to generate personalized inspection list data marked with risk level, high-risk inspection areas and inspection priority. Step S04: Based on the personalized inspection checklist data, bind the inspection items to the spatial location of the liquefied natural gas vessel 3D model to generate visual inspection navigation data; Step S05: During the on-site inspection, in response to the query command triggered by the visual inspection navigation data, the large language model finely tuned by the maritime professional corpus is called, and the legal knowledge engine and historical case engine are simultaneously triggered to perform dual-engine retrieval. After fusion reasoning, the structured inspection auxiliary response data is output. Step S06: The defect data and evidence materials collected during the on-site inspection are structured and automatically associated with a five-dimensional relational mapping network to generate bilingual (Chinese and English) inspection report data that conforms to the Tokyo Memorandum of Understanding. Step S07: Integrate structured data from the entire inspection process, such as visual inspection navigation call records, structured inspection auxiliary response data, and standardized inspection report data, as feedback; update the specialized knowledge base data; and optimize the parameters of the risk analysis model and the large language model.

[0025] Specifically, a specialized knowledge base for LNG ship PSC inspections will be established.

[0026] Data collection: Collected historical records of PSC inspections of 9,000+ LNG vessels in the Asia-Pacific region that joined the Tokyo Memorandum of Understanding since 2014, the International Code for the Construction and Equipment of Liquefied Gas Ships in Bulk (IGC Code), the International Convention for the Safety of Life at Sea (1974) (SOLAS Convention) and its annual amendments, domestic maritime regulatory documents, and data from a question bank of 2,000+ questions for the LNG vessel PSC inspection knowledge competition.

[0027] In this embodiment, multi-source basic data undergoes structured processing and association mapping; maritime regulatory texts are stored using parent-child segmentation semantic storage; and a specialized knowledge base is constructed based on the stored structured regulatory data, comprising a five-dimensional association mapping network including inspection locations, defect codes, regulatory clauses, action codes, and defect descriptions. Collect historical inspection records of members of the Tokyo Memorandum, maritime law texts, professional knowledge data, and training question bank data as multi-source basic data; A semantic segmentation algorithm is used to divide the legal text into independent semantic units, using the chapter-section-article-clause numbering and preset logical connectors as semantic boundaries. The segmented semantic units are converted into a structured markup language format, and a parent-child segmentation architecture is used to complete the semantic storage of maritime regulations text; Based on data mining technology, a five-dimensional relational mapping network is constructed, which includes inspection locations, defect codes, legal and regulatory basis, action codes, and natural language descriptions of typical defect phenomena. The five-dimensional relational mapping network is calibrated through a collaborative mechanism of large model formatting, regularization correction, and manual verification to generate specialized knowledge base data.

[0028] Specifically, the structured processing involves: professionally adapting the collected data; using a semantic segmentation algorithm based on the hierarchical structure and terminology of maritime regulations for regulatory texts; dividing the regulatory text into semantically independent clause units using the numbering of "chapter-section-article-clause" and the logical connectors "shall" and "shall not" as semantic boundaries; and storing the text in a Markdown structured format. For the competition questions in the knowledge competition question bank, the assessment areas and assessment categories are marked.

[0029] Association mapping construction: Based on data mining technology and the data analysis results of the survey on inspectors' practical experience in LNG ship PSC inspections, a five-dimensional association mapping network is constructed, consisting of "inspection location - defect code - legal and regulatory basis - action code - natural language description of typical defect phenomena". The mapping relationship is calibrated through a collaborative processing mechanism of "large model formatting + regularization correction + manual verification". The defect code is extracted from the Tokyo Memorandum of Understanding Port State Control Inspection Defect Table and synchronized with the special knowledge base in real time.

[0030] In this embodiment, in response to the identification information of the vessel to be inspected, specialized knowledge base data is invoked, and a risk analysis model combining a rule engine and machine learning classification is used to calculate and generate personalized inspection checklist data labeled with risk level, high-risk inspection areas, and inspection priority, including: Based on the IMO number or ship name identification information of the vessel to be inspected, the historical inspection records of the corresponding vessel, the flag state to which it belongs, and the statistical defect data of the fleet are retrieved from the special knowledge base. Rigid risk assessment is performed through a preset rule engine. Rigid risk assessment includes directly judging a high-risk area if there are residual records within a preset period, escalating the risk if there are recurrence records of the same inspection defect, and judging a high-risk inspection area if there are cumulative high-frequency defects in the same area. A random forest machine learning classifier trained on historical inspection data is used to comprehensively calculate the ship risk score by weighting factors such as the frequency of defect recurrence in the same part, the severity of historically detained defects, the flag state of the ship and the statistical defect rate of the fleet, and output the corresponding risk level result. Based on the risk level results and high-risk inspection areas, personalized inspection checklist data is generated by prioritizing high-risk areas.

[0031] Specifically, this involves creating a risk profile of the vessel and developing a personalized inspection plan.

[0032] Data retrieval: Input the target vessel's identification information (such as IMO number) to retrieve its historical inspection records and statistical defect data of its flag state / company fleet.

[0033] Risk Level Determination: A risk analysis model combining a rule engine and a machine learning classifier is adopted. The pre-set rule set combines rigid rules such as "if there is a record of detention in the past year, it is directly judged as high risk" and algorithms trained based on historical data such as "if the same inspection defect is ≥2 times in historical inspection data, the risk is upgraded and identified as medium risk", "if the same part has more than 3 high-frequency defects, the part is judged as high risk", and "if there are more than 2 serious defects, it is judged as high risk". The machine learning classifier is trained with data from 9000+ ships and comprehensively evaluates based on "the frequency of defect recurrence in the same part (weight 40%), the severity of historical detention defects (weight 30%), and the statistical defect rate of the flag state (weight 30%)", and outputs a risk level of 0-30 (low risk), 31-60 (medium risk), and above 61 (high risk).

[0034] Checklist generation: Based on the risk level, a personalized checklist is generated, marking high-risk areas and their inspection priorities. High-risk areas are given priority in the inspection sequence.

[0035] In this embodiment, based on personalized checklist data, the inspection items are spatially bound to the 3D model of the liquefied natural gas vessel to generate visualized inspection navigation data, including: Load a general 3D model of a liquefied natural gas vessel onto a mobile terminal and bind the inspection items in the personalized inspection checklist to their corresponding spatial locations in the 3D model one by one; Based on the bound 3D model, a visual inspection navigation path that supports touch zoom and rotation operations is generated; In response to trigger commands for the inspection area in the 3D model, the system outputs the basic information, inspection points, and relevant regulatory data of the corresponding inspection area in real time. High-risk area data and lightweight model parameters are cached locally on the mobile terminal, and the cached data is called to provide inspection and navigation services in weak network environments.

[0036] In this embodiment, during the on-site inspection, in response to a query command triggered by visual inspection navigation data, a large language model finely tuned from a maritime professional corpus is invoked. Simultaneously, a dual-engine retrieval is performed using both the regulatory knowledge engine and the historical case engine. After fusion reasoning, structured inspection auxiliary response data is output, including: Perform intent recognition and entity extraction on received natural language or voice query commands triggered by the visual inspection navigation interface, and extract search keywords; The system utilizes a large language model finely tuned with professional maritime corpus and historical question-and-answer pairs to trigger a dual-engine search based on search keywords. Specifically, the regulatory knowledge engine performs precise matching of regulatory clauses based on a specialized knowledge base, while the historical case engine performs similarity matching of similar defect cases based on a historical inspection database. The system integrates and reasons the regulatory data obtained from dual-engine retrieval with historical case data, and outputs structured response data that includes defect codes, legal basis in Chinese and English, graded handling suggestions, and corresponding action codes.

[0037] In this embodiment, the defect data and evidence materials collected during on-site inspections are structured and automatically associated with a five-dimensional relational mapping network to generate bilingual (Chinese and English) inspection report data conforming to the Tokyo Memorandum of Understanding standards, including: The defect classification information, text descriptions, images and video evidence materials collected during the on-site inspection based on the visual inspection navigation path are structured and organized to obtain a standardized defect dataset; Automatically associate and match the standardized defect dataset with the five-dimensional relationship mapping network in the specialized knowledge base to supplement and improve the corresponding data of defect codes, legal basis, and action codes; The matched full data will be automatically populated into the standard FORM A and FORM B report templates of the Tokyo Memorandum. The report content is automatically switched to Chinese and English based on a maritime bilingual terminology database, generating inspection report data in both Chinese and English.

[0038] Specifically, it performs inspections that integrate navigation of the inspection site with real-time AI assistance.

[0039] Inspection Area Navigation: Load a general 3D model of an LNG vessel onto the mobile terminal, bind the items in the personalized inspection checklist to the spatial location of the model, provide a visual and interactive inspection area navigation, and support touch zoom and rotation operations; select an inspection area in the model to display the basic information (images, text descriptions), inspection points and corresponding legal basis of the area in real time.

[0040] AI Real-Time Assisted Question Answering: The mobile terminal integrates the deepseek v3 671b natural language processing model, finely tuned with maritime professional corpus, to form an AI assistant module, supporting PSCO queries using natural language and voice input; the AI ​​assistant simultaneously triggers a dual-engine search of "regulatory knowledge engine + historical case engine", accurately matching regulatory clauses and similar defect cases from a specialized knowledge base, and outputting a structured response in real time after fusion reasoning, consisting of "defect code + Chinese and English regulatory basis + graded handling suggestions (specific measures and handling behavior codes to be taken, 10 - defect has been corrected, 16 - corrected within 14 days, 17 - corrected before departure, 18 - corrected within 3 months, etc.)".

[0041] In this embodiment, structured data from the entire inspection process, including visual inspection navigation call records, structured inspection auxiliary response data, and standardized inspection report data, is integrated as feedback to update specialized knowledge base data. Furthermore, parameters of the risk analysis model and the large language model are optimized, including: Collect structured feedback data throughout the entire inspection process. The feedback data includes defect information, handling results, differences between AI suggestions and official handling data, and correlation data of newly emerging defects. According to the preset first cycle, the specialized knowledge base and the five-dimensional relational mapping network are incrementally updated through a process of automatic formatting, regular expression correction, and manual verification. According to the preset second cycle, based on the collected feedback data, the weighted parameters of the risk analysis model and the large language model are fine-tuned through reinforcement learning.

[0042] Specifically, it generates inspection reports and drives system intelligence.

[0043] Automatic report generation: The system automatically associates the defect information and evidence materials (text, images, and videos) recorded on-site with a five-dimensional correlation mapping network and fills them into templates that conform to the Tokyo Memorandum of Understanding standards FORM A / FORM B. Based on a maritime professional bilingual terminology database, it achieves automatic Chinese-English conversion and generates bilingual English-Chinese inspection reports. This addresses the practical needs and language usage habits of Chinese PSCO inspectors and solves the problems of low inspection efficiency and large errors caused by manual translation during the inspection process. It also supports export in PDF / Excel format.

[0044] System Iteration and Optimization: The structured data from the entire inspection process (including defect information, handling results, and newly emerging defect relationships) is used as feedback data. The specialized knowledge base and five-dimensional association mapping network are incrementally updated monthly through the process of "automatic formatting - regular expression correction - manual verification". The risk analysis model parameters and natural language processing model are fine-tuned quarterly based on the feedback data to continuously improve the system's intelligence level.

[0045] In this embodiment, a training and assessment step based on specialized knowledge base data and a 3D model is also included. The training and assessment step includes: Based on a personalized inspection checklist or a specified high-risk inspection scenario, the corresponding 3D model inspection parts and related inspection items are dynamically loaded to generate a 3D simulation inspection task, complete the full-process simulation training of the inspection, and generate a simulation inspection record. Based on the training question bank data of label classification, special assessment test papers are generated by adopting adaptive test paper generation or scenario-based targeted test paper generation strategies. Among them, adaptive test paper generation is based on matching questions with students' historical wrong questions and knowledge weaknesses, while scenario-based targeted test paper generation is based on matching questions with specific ship risk profiles or port high-frequency defect types. The assessment results are quantitatively scored, and a training effectiveness evaluation report and targeted reinforcement suggestions are generated, linking the corresponding learning materials in the specialized knowledge base.

[0046] Specifically, the training and assessment steps.

[0047] Simulated Inspection Mode: To meet the training needs of new maritime personnel, the system provides a simulated inspection function. The system supports dynamically loading corresponding inspection areas and related inspection items based on a personalized inspection checklist or designated high-risk scenarios (such as engine room anti-pollution equipment inspection and lifesaving equipment inspection). Trainees can simulate the entire PSC inspection process and generate simulated inspection records.

[0048] Specialized Assessment: Based on a knowledge competition question bank with simulated inspection scenarios and labeled classifications, specialized assessment papers for inspection parts are generated. The question types include single choice, multiple choice, and true / false. The accuracy of defect identification and the standardization of inspection procedures of trainees are quantitatively scored. The simulation training results and targeted reinforcement suggestions are output, and corresponding learning materials in the specialized knowledge base are used for learning.

[0049] In other embodiments, an intelligent assistance system for port state control (PSC) inspections of liquefied natural gas (LNG) vessels is provided, based on any of the preceding intelligent assistance methods for PSC inspections of LNG vessels, including: A specialized knowledge base module is used to store and manage structured semantic units, historical inspection cases, professional knowledge, and multidimensional relational mapping networks; The intelligent analysis and service engine integrates risk analysis models and natural language processing models to provide computing and reasoning services for online integrated management platforms and mobile intelligent inspection terminals. An online integrated management platform is used for checking task management, data analysis, knowledge base maintenance, and training assessment. The mobile intelligent inspection terminal integrates 3D visualization inspection navigation, AI assistant, and report generation functions for on-site operations. The data closure and iteration module automatically collects data during the inspection process and the correlation data of newly emerging defects. It updates the specialized knowledge base and five-dimensional correlation mapping network monthly through an automatic formatting-regular expression correction-manual verification process. It also fine-tunes the risk analysis model parameters and natural language processing model quarterly based on the collected data. The online integrated management platform and the mobile intelligent inspection terminal collaborate through a data interface, sharing and accessing specialized knowledge base modules and intelligent service engines.

[0050] In this embodiment, the mobile intelligent inspection terminal has edge computing capabilities, and can provide basic inspection navigation and question-and-answer services based on locally cached high-risk area data and lightweight models in weak network environments.

[0051] Specifically, its core consists of two parts: an online integrated management platform based on a B / S architecture and a mobile intelligent inspection terminal. The two achieve real-time data synchronization and functional collaboration through standardized API interfaces. They share a unified data support layer and a technical support layer at the bottom layer, and the overall design follows the principle of "high cohesion and low coupling".

[0052] Specifically, the core of this technical solution's device system consists of two parts: an online system based on a B / S architecture and a mobile tablet-based AI intelligent inspection terminal (AI Smart Eye Intelligent Inspection Assistant). The two parts achieve real-time data synchronization and functional collaboration through standardized API interfaces, sharing a unified data support layer and technical support layer at the underlying level. The overall design follows a modular principle of "high cohesion and low coupling." The following sections will detail the system composition, core components, and connection relationships in two parts.

[0053] In this embodiment, the online integrated management platform includes a specialized knowledge base module, a ship inspection electronic map module, a knowledge graph module, an intelligent inspection module, a training classroom module, and a system management module; The intelligent inspection module establishes bidirectional connections with the ship inspection electronic map module and the specialized knowledge base module. The ship inspection electronic map module transmits location data to the intelligent inspection module, the specialized knowledge base module transmits regulatory data to the intelligent inspection module, and the intelligent inspection module transmits inspection process data back to both. The training classroom module establishes a one-way connection with the specialized knowledge base module and the intelligent inspection module, respectively, to receive learning materials transmitted by the specialized knowledge base module and inspection case data transmitted by the intelligent inspection module; The system management module establishes bidirectional connections with the specialized knowledge base module, the ship inspection electronic map module, the knowledge graph module, the intelligent inspection module, and the training classroom module, respectively, for access control, data maintenance, and related data interaction.

[0054] In this embodiment, the specialized knowledge base module includes a regulatory database, a defect standard library, a historical case library, a professional knowledge encyclopedia, a knowledge retrieval unit, and an association mapping construction unit; The knowledge retrieval unit is connected to the legal database, defect standard database, historical case database, and professional knowledge encyclopedia. The association mapping construction unit establishes a five-dimensional association mapping network of "inspection location - defect code - legal clause basis - action code - natural language description of typical defect phenomena". The knowledge base module uses a hierarchical structure to store legal documents, cutting the regulations into semantically complete sub-blocks according to natural hierarchy and storing them.

[0055] Specifically, the online system adopts a B / S architecture, allowing users to access it through mainstream browsers without installing any additional client. It boasts advantages such as cross-platform compatibility, ease of maintenance, and strong scalability. Its core modules and components are as follows: LNG ship PSC inspection specialized knowledge base module: This module forms the intelligent foundation of the system. Its core function is to provide structured storage and intelligent retrieval services for regulatory clauses, defect standards, historical cases, and professional knowledge. It mainly includes a regulatory database (international conventions + domestic regulatory documents categorized by "international / domestic / safety / transportation / inspection basis"), a defect standard library (defect codes derived from the Tokyo Memorandum of Understanding's Port State Control Defect Table), a historical case library (PSC inspection records of over 9000 LNG vessels in the Asia-Pacific region from 2014 to the present), a professional knowledge encyclopedia, a bidirectional search engine, and a relational mapping construction unit. It employs a "parent-child segmentation" architecture to store professional books and reference materials, converting IGC rules and SOLAS international convention documents into Markdown structured format. Each regulation is divided into multiple semantically complete "sub-blocks" according to a natural hierarchy of "chapter-section-article-clause," and based on data mining technology, a five-dimensional relational mapping network is constructed: "inspection location - defect code - regulatory clause basis - action code - natural language description of typical defect phenomena." Embedding RAG technology and contextual semantic segmentation algorithms, it supports keyword retrieval, natural language query, and related knowledge recommendation, with a retrieval response time of ≤1 second. The consistency of the format is ensured through a three-level process of "large model formatting + regular expression correction + manual verification".

[0056] This module supports adding, deleting, modifying, querying, retrieving, and statistically analyzing knowledge, and can visualize the knowledge structure in the form of tree diagrams and other formats.

[0057] LNG ship inspection electronic map module: This module binds abstract inspection clauses to specific ship spatial locations. It provides a classic LNG ship visualization model and inspection location positioning. The inspection area covers thirteen key areas of LNG ships, such as the bridge, engine room, and cargo holds. Each inspection area is associated with its corresponding inspection location, inspection items, and related descriptions. Users can click on an area or equipment label on the map to directly jump to the page containing the inspection content, related descriptions, and images. Simultaneously, it displays a list of all items to be inspected at that location, inspection points, regulatory basis, and possible defect codes, achieving "what you see is what you check."

[0058] Knowledge graph module: The knowledge graph is used to link the inspection locations, key points, defects, inspection basis, and defect codes of LNG ship PSC inspections. Users can click on specific inspection locations to display relevant knowledge content.

[0059] Intelligent inspection module: This module is responsible for the lifecycle management of actual inspection tasks. Administrators can use this module to simulate specific inspection tasks for particular vessel locations and specify inspection areas and items. When PSCO executes an inspection task, the system first calls the integrated risk analysis model. This model combines a rule engine with a machine learning classifier. The rule engine's preset rule set includes rules such as "directly classifying a vessel as high-risk if it has a detention record within the past year" and "upgrading the risk if the same defect recurrences ≥2 times in historical inspection data." The machine learning classifier, based on a random forest algorithm trained on data from over 9000 vessels, comprehensively evaluates the vessel based on "frequency of defect recurrence in the same location (weight 40%), severity of historical detention defects (weight 30%), and statistical defect rate of the flag state / company fleet (weight 30%)," outputting a high / medium / low risk level. Based on the risk level and the analyzed high-frequency defect locations, the system automatically generates a personalized inspection list marking high-risk locations and inspection priorities. The report template strictly adheres to the Tokyo Memorandum of Understanding, supports one-click association of inspection data and evidence documents, and automatically fills in core fields such as defect codes and legal basis, reducing manual data entry workload. The system includes a checklist automatic generation unit, a defect data entry unit (including a multi-type defect classification dictionary), an evidence file upload and association unit (supporting text, image, and video formats), a FORM A / B report template unit, a Chinese-English bilingual conversion unit (based on a maritime professional bilingual terminology database), and a report export unit (supporting PDF / Excel formats).

[0060] Training Classroom Module: This module provides comprehensive training and assessment functions, including online learning, mock tests, question bank practice, exams, and training data statistical analysis.

[0061] ① Online learning: Learn about LNG ship PSC inspection and related knowledge in the knowledge base.

[0062] ②Simulation Training: Users can create simulated inspection tasks (t_simulation) to simulate the entire inspection process on electronic maps and 3D models. Finally, the system can automatically generate simulated FORM A and FORM B reports.

[0063] ③ Question Bank Training: The module integrates nearly 2,000 professional questions. Users can customize the number of questions to conduct corresponding question bank training, record their answers and wrong answers, and view the analysis of some questions.

[0064] ④ Online Examination: Supports users with administrative privileges to define the examination scope, supports two intelligent test paper generation strategies (adaptive test paper generation: based on students' historical wrong questions and knowledge weaknesses; scenario-based targeted test paper generation: based on specific ship risk profiles or port high-frequency defect types), supports users to participate in unified examinations, and allows for examination grading.

[0065] ⑤ Training statistics: Conduct multi-dimensional statistical analysis on users' training duration, quiz accuracy, and simulation check completion to form a personal training profile.

[0066] System Management Module: Responsible for the basic operation and maintenance of the entire platform, including core functions such as user management, role and permission management, department management, data maintenance, system parameter configuration, and operation log auditing. The user management unit has the following roles: PSCO, Administrator, and Training Personnel.

[0067] It supports user registration, login, and password reset, and assigns function access permissions according to roles. Administrators have data modification permissions, while PSCOs only have inspection, learning, and query permissions. It automatically backs up the database regularly, supports manual data update, and records all user operation logs for auditing purposes.

[0068] In this embodiment, the mobile intelligent inspection terminal includes a homepage / task center module, an AI assistant module, an inspection map module, a panoramic map module, an inspection profile module, a mobile knowledge base module, and an inspection training module; The AI ​​assistant module establishes bidirectional connections with the mobile knowledge base module and the profile inspection module, respectively, to receive cached data transmitted by the mobile knowledge base module and profile data transmitted by the profile inspection module, and to send back question-and-answer interaction data to both of them. The panoramic map module is connected to the inspection map module. By clicking, you can jump to the details page of the corresponding inspection part in the inspection map module. The homepage / task center module establishes one-way connections with the AI ​​assistant module, inspection map module, panoramic map module, inspection profile module, mobile knowledge base module, and inspection training module, respectively.

[0069] Specifically, this terminal is a dedicated tool for PCSO to conduct on-site inspections, focusing on portability, real-time and intelligent interaction, and core module online systems to form complementary functions and data synchronization.

[0070] Homepage / Task Center: It provides aggregated function entry points, to-do task reminders, message notifications, and quick navigation services, serving as a unified entry point for on-site work. Function entry units are displayed in convenient modules, allowing users to jump to core modules with a single click.

[0071] The system synchronizes inspection plans and training tasks with the online system, pushes important information such as regulatory updates and user feedback processing results in real time, and supports quick search function entry by module name.

[0072] Check the map module: It integrates the electronic map function of the online platform, supports touch operation on tablets, and allows for smooth browsing of 3D maps of various areas of the ship. It supports setting up convenient entry points for inspection areas on the homepage, allowing users to select and view inspection items; the specific functions are consistent with the online website. Users can also click within the inspection map module to view the knowledge graph display content for the corresponding inspection area.

[0073] Panoramic map module: Using the 3D ship model as the entry point for viewing inspection areas, users can select the area they want to view based on the 3D model and jump to the corresponding content in the inspection map module.

[0074] AI Assistant Module: It is used to provide natural language question answering, defect code matching, rapid query of legal basis and question answering interaction services, including natural language processing unit, AI model inference unit, speech recognition and synthesis unit and answer formatting output unit.

[0075] This module deeply integrates the Deepseek v3 671b localized fine-tuning model optimized with maritime terminology, supports natural language and voice input queries, and returns defect codes, Chinese and English descriptions, action codes and legal basis in real time. It has a fixed output format to improve on-site efficiency and has a question-and-answer accuracy rate of ≥95%.

[0076] The workflow is as follows: PSCO inputs a natural language question (such as "cargo hold isolation valve leakage defect") via voice or text. The AI ​​assistant first performs intent recognition and entity extraction, then simultaneously activates the "regulatory knowledge engine" to query the structured knowledge base and the "historical data engine" to query the historical defect database of LNG ship PSC inspections, performing fusion reasoning. Finally, it outputs results in a fixed, standard template, with a fixed format of "defect code, defect description in Chinese and English, corresponding convention clauses, defect severity assessment, and action code suggestion." This process helps replace manual retrieval of manuals and memorization of codes.

[0077] Knowledge base module: It provides online access to the knowledge base, enabling quick retrieval and viewing of regulations, defect standards, and case studies on mobile devices. As a supplement to the AI ​​assistant, it supports keyword-based searches for legal clauses. Its main components include a lightweight search engine and knowledge display units. It supports keyword and voice searches, and the knowledge content is displayed in a structured format adapted for mobile screens. It also supports downloading frequently used knowledge documents and automatically updates the knowledge base online.

[0078] Inspection training module: It provides lightweight 3D model training, question bank practice, exam and learning data query functions on mobile devices. Users can practice with the question bank, conduct some simulated training, and take exams on mobile devices, with the same content as the web version.

[0079] Check the portrait module: This system displays visualized data on vessel inspection profiles, detention profiles, and port inspection profiles, allowing for quick on-site viewing. It includes profile data visualization, data filtering, and detail viewing units. Core data such as annual vessel inspection frequency, high-frequency defects, and detention risk points are presented using pie charts and Pareto charts. Users can filter specific vessel profiles by vessel name and IMO number, and click on charts to view data details. PSCOs can quickly query historical inspection profiles of currently inspected vessels on-site, viewing high-frequency absences and detention records, providing historical data references for inspections.

[0080] In this embodiment, the data support layer includes a regulatory database, a historical inspection database, a ship basic information database, a training question bank database, and a user feedback database; The intelligent service layer includes artificial intelligence models, semantic segmentation algorithm units, risk analysis model units, 3D modeling units, big data analysis units, bilingual conversion units, and interface adaptation units; The data support layer transmits data to the intelligent service layer through the JDBC interface, and the intelligent service layer transmits processed data to the application layer through a standardized interface.

[0081] Specifically, the hierarchical connection is as follows: Data support layer → Intelligent service layer → Application layer (online system + mobile terminal). Data support layer: includes a regulatory database, a historical inspection database, a ship basic information database, a training question bank database, a user feedback database, and a risk assessment rule base. Structured data is stored through a MySQL 8.0 database management system, and data consistency is ensured between databases through foreign key relationships.

[0082] Intelligent Service Layer: Integrates deepseek v3 671b AI model, contextual semantic segmentation algorithm unit, 3D modeling unit, big data profiling and analysis unit, risk analysis model, Chinese-English bilingual conversion unit, and API interface adaptation unit, providing core technical capabilities for the application layer.

[0083] Connection method: The data support layer outputs structured data to the intelligent service layer through the JDBC interface. After processing the data, the intelligent service layer provides services to the online system and mobile terminal through the standardized API interface. The online system and mobile terminal achieve real-time data synchronization through the HTTP / HTTPS protocol. Specific Implementation Example 1: The following describes a complete PSC intelligent inspection process for an LNG vessel, illustrating the collaborative working process of the method and system of this invention. It simulates the entire process of a Port State Control Officer (PSCO) using the system described in this invention to perform a PSC inspection on the LNG vessel "Pengyuan" with IMO number "9412031".

[0085] Specifically, the core working principle of this technical solution is "data-driven + AI-enabled + full-process collaboration." It constructs a specialized knowledge base through data preprocessing, and relies on AI models and algorithms to achieve intelligent assistance across all scenarios of inspection, training, and decision-making. The online system and mobile terminal collaborate to complete a closed-loop process of "pre-event preparation - in-event execution - post-event analysis." Its complete workflow reflects deep collaboration between online and offline, and between front-end and back-end modules. The following is a detailed explanation categorized by core technical methods.

[0086] In this embodiment, the steps of inputting multi-source data related to port state control inspections and constructing a specialized knowledge base include inputting regulatory text data (international / domestic classification data such as IGC rules and SOLAS conventions), historical inspection record data (PSC inspection records of 9,000+ LNG vessels), and basic vessel information data. A collaborative processing mechanism of classification integration, formatting, regularization correction, and manual inspection is adopted to obtain structured regulatory data and associated historical data. Based on data mining technology, a five-dimensional relational mapping network of "inspection location - defect code - regulatory clause basis - action code - natural language description of typical defect phenomena" is constructed and stored in the corresponding database.

[0087] In this embodiment, the steps of performing inspection assistance and report generation operations include inputting ship identification information, using a risk analysis model for calculation, combining a preset rule set with a machine learning classifier, outputting the ship risk level and high-frequency defect locations, and generating a personalized inspection list accordingly. Input defect information and evidence materials during the inspection process, and use defect code matching, legal basis association and template filling actions to obtain a standard bilingual inspection report.

[0088] In this embodiment, the steps of performing knowledge retrieval include inputting a user query request; using semantic segmentation, keyword extraction, five-dimensional association mapping network association, and dual-engine retrieval actions, matching corresponding legal provisions, defect-related information, and historical case data from a specialized knowledge base to obtain structured retrieval results.

[0089] In this embodiment, the steps for performing professional training operations include inputting training scenario selection instructions and question bank configuration information, using 3D model loading, simulation inspection guidance, intelligent test paper generation (adaptive test paper generation or scenario-based targeted test paper generation), answer recording and statistical analysis actions to obtain simulation training results and training effect evaluation reports.

[0090] Specifically, data preprocessing and specialized knowledge base construction methods.

[0091] Multi-source data acquisition and integration: The system first integrates three types of core data: regulatory and document text data, mainly including standards related to administrative penalties for maritime violations, regulations on LNG vessel operating areas, international conventions and relevant domestic regulations, and LNG vessel operation data; a distributed version control system repository (Git repository) for regulatory data: converting the IGC / SOLAS convention into Markdown structured format and storing it in the maritime regulations Git repository; and historical inspection data: PSC inspection records of Asia-Pacific LNG vessels from 2014 to the present, data on 9,000+ vessels, etc., to support subsequent functions.

[0092] Integration Method: Data is stored in a three-dimensional classification of "data type-source-time". A three-level scheme of "large model format + regular expression correction + manual verification" is used to process regulatory data to ensure structured accuracy. Historical inspection data is deduplicated, completed, and correlated. Based on data mining technology, a five-dimensional correlation mapping network of "inspection location-defect code-legal clause basis-action code-natural language description of typical defect phenomena" is constructed.

[0093] Data standardization and storage: Standardization process: Convert IGC rules and SOLAS convention data into Markdown structured format to unify the coding rules for defect codes and action codes.

[0094] Structured data is stored using a MySQL 8.0 database, and incremental updates and version control of regulatory data are achieved through a Git library; mobile devices support offline caching of frequently used data, and encrypted storage is used to ensure data security.

[0095] Implementation methods of core technologies for online systems: The knowledge base intelligent retrieval method allows users to input query keywords or natural language, and the context semantic segmentation algorithm performs semantic segmentation and keyword extraction on the query content; RAG technology accurately matches relevant laws, defect standards and cases from the knowledge base; and the "five-dimensional association mapping network" is invoked to return related knowledge content.

[0096] Based on user search history and feedback data, the search algorithm is continuously optimized through A / B testing to improve the accuracy of related recommendations. Frequently queried knowledge is cached to shorten the search response time.

[0097] Intelligent inspection and report generation methods: Checklist Generation: After receiving information such as the vessel's IMO number and flag state, the system invokes a risk analysis model. This model employs a combination of a rule engine and a machine learning classifier for analysis. The rule engine's preset rule set includes rules such as "a vessel with a detention record within the past year is directly classified as high-risk" and "the risk is upgraded if the same defect recurs ≥2 times." The machine learning classifier, based on the random forest algorithm, comprehensively evaluates the risk level and frequently occurring defect locations based on factors such as "the frequency of defect recurrence in the same location (weight 40%), the severity of historical detention defects (weight 30%), and the statistical defect rate of the vessel's flag state / company fleet (weight 30%)." The system then generates a personalized checklist accordingly.

[0098] Automatic report generation: After inspectors enter defect information and upload evidence documents, the system automatically matches the defect code with the regulatory basis, calls the FORM A / B standardized template, and fills in core fields such as inspection data, vessel information, and defect details; the report can be switched between Chinese and English through a bilingual conversion unit and supports export in PDF / Excel format.

[0099] Training classroom interaction and statistical methods: 3D simulation training: After the user selects a training scenario, the system loads the corresponding 3D ship model and guides the user to complete the entire process of "locating parts - identifying defects - entering information - generating reports".

[0100] Question Bank and Exam Management: Administrators can set question bank categories, question types (single choice / multiple choice / true / false), and difficulty levels through the backend. It supports two intelligent test paper generation strategies: adaptive test paper generation and scenario-based targeted test paper generation. After the exam, the system automatically scores and calculates the correct answer rate and common mistakes, generating individual and group training effectiveness evaluation reports.

[0101] Core technology implementation methods for mobile AI intelligent inspection terminals: AI assistant question-answering and reasoning methods: In the interactive process, users input their query requirements via natural language input or voice input (such as "lifeboat release defect handling"). The voice recognition unit converts the speech into text. The deepseek v3 671b AI model performs logical reasoning on the query content, calls the knowledge base and inspection database, and matches the corresponding defect code, Chinese and English descriptions, action codes, and legal basis. The answer formatting unit outputs the results in the structure of "core conclusion - detailed description - legal reference".

[0102] Weak network adaptation: Using edge computing technology, frequently queried data and basic model parameters are cached locally on the mobile device. In weak network environments, local data is called first to respond, and data and model iteration results are updated synchronously after the network is connected.

[0103] Examining methods for visualizing profile data: Data processing: After receiving the raw profile data pushed by the online system, the mobile device performs local processing through a lightweight big data analysis unit to extract core indicators.

[0104] Visualization: Using the mobile-adapted version of ECharts, the profile data is displayed in the form of bar charts, pie charts, line charts, etc. It supports touch zoom to view details, and clicking on the chart can jump to the corresponding defect case or regulatory clause.

[0105] System iterative optimization method.

[0106] Data and knowledge base iteration: The system is regularly updated. It automatically retrieves the latest regulations and inspection data from the Tokyo Memorandum website and maritime authorities every month, and updates the database and knowledge base through an automatic formatting, regular expression correction, and manual verification process. It also performs incremental training on historical inspection data every quarter to optimize the defect identification and matching accuracy of the AI ​​model.

[0107] User feedback driven: Users submit feedback on the accuracy of knowledge and the usability of functions through the online system or mobile terminal. After verification, the administrator will classify and process the feedback, correct knowledge errors, incorporate function suggestions into the iteration plan, complete emergency updates within 1 working day, and summarize the functions monthly for optimization.

[0108] Model and algorithm optimization: AI Model Fine-tuning: Based on user question-and-answer logs and feedback data, the Deepseek v3 671b model is fine-tuned quarterly using a maritime terminology database and newly added defect cases to optimize reasoning logic and answer accuracy, ensuring question-and-answer accuracy.

[0109] Algorithm performance optimization: Distribute system computing resources through load balancing technology, optimize algorithm complexity for high-frequency access modules to reduce response time; conduct stress tests regularly, adjust system parameters based on test results, and ensure stability under concurrent access from multiple users.

[0110] Specific technical method phase description: For pre-inspection intelligent preparation and planning, after receiving an inspection assignment for an LNG vessel (e.g., IMO: 1234567), the Port State Control Officer (PSCO) inputs the target vessel's IMO number into the system. The system then invokes a risk analysis model for calculation. This model combines a rule engine with a machine learning classifier. The rule engine's preset rule set includes rules such as "a vessel with a detention record within the past year is directly classified as high-risk" and "if the same defect in the same location has recurred ≥2 times in historical inspection data, the risk is upgraded." The machine learning classifier, based on the random forest algorithm, comprehensively evaluates the vessel's risk level (high / medium / low) based on factors such as "the frequency of recurrence of defects in the same location (weight 40%), the severity of historical detention defects (weight 30%), and the statistical defect rate of the flag state / shipping company (weight 30%)."

[0111] Real-time intelligent interaction and decision support during inspection: PSCO boarded the ship carrying a tablet computer equipped with the "AI Smart Eye Assistant". During the inspection, its operation involved two parallel paths: Streamlined inspection process based on electronic maps: PSCOs open the "Inspection Map Module" on their mobile terminals and navigate to the corresponding area (e.g., engine bay) according to the checklist. By browsing the panoramic view and selecting the equipment to be inspected (e.g., "Emergency Generator"), the complete inspection items, regulatory basis, and common defects for that equipment are displayed on the side of the screen. PSCOs verify the inspection steps on-site and directly record the inspection results (pass / fail) and take photos as evidence via the terminal. This process spatializes and visualizes the inspection items, avoiding omissions.

[0112] AI-powered intelligent question answering and instant decision-making: When a PSCO discovers any anomalies or uncertainties outside the preset list (such as an unfamiliar valve status anomaly), they don't need to leave the site. They can directly activate the "AI assistant module" on their mobile terminal, describing the problem in natural language: "What should I do if the pressure relief valve setting in the liquid cargo tank is incorrect?" The AI ​​assistant module then responds in the background within milliseconds: its fine-tuned large model analyzes the problem's intent, accurately locating the specific clause regarding safety valves in Chapter 8 of the IGC Rules from the local structured Markdown knowledge base, and matching similar defect descriptions and codes (such as "15105") from the synchronized historical case library. Finally, the terminal interface immediately generates a standard answer card: "Defect Code: 15105; Description: Pressure relief valve setting incorrect; Basis: IGC 8.2.5; Action Code: 30 (Detention)." This provides timely and authoritative decision support for the PSCO's on-site assessment.

[0113] Automated report generation and data storage after inspection: At the end of the inspection, all defects, evidence photos, corresponding defect codes, and supporting documentation recorded on the mobile device are structurally saved in the task. The PSCO clicks the "Generate Report" button. The system will automatically invoke the report generation logic, accurately filling all defect information, vessel information, and inspector information into the FORM A and FORM B report templates that conform to the Tokyo Memorandum of Understanding international standards. The PSCO only needs to perform a final review before submission, reducing report writing time from several hours to just a few minutes.

[0114] Closed-loop feedback of data and system optimization: All structured data generated during this inspection (new defect associations and handling measures) will be automatically saved to the central database. The online platform's data analysis and profiling module will update the vessel's "inspection profile," the port's "port profile," and macro-level "defect statistics" in real time based on the collected data. More importantly, the novel "question-answer" pairs generated during this inspection will be added to the AI ​​model's reinforcement learning sample library for the next round of incremental fine-tuning of the AI ​​assistant module's underlying large model. This will enable the entire system to continuously learn from maritime regulatory expertise, constantly evolving its intelligence.

[0115] Specific Implementation Example 2: A Complete LNG Ship PSC Intelligent Inspection This embodiment simulates the entire process of a Port State Control Officer (PSCO) using the system described in this invention to conduct a PSC inspection on the LNG vessel "Pengyuan" with IMO number "9412031".

[0116] 1. Pre-boarding preparation: Intelligent task preparation and knowledge preparation In the office, PSCO accesses the online system on their PC by logging in with their username and password (via a browser). In the "Intelligent Inspection Module," they enter the target vessel's IMO number "9412031," and the system automatically retrieves historical inspection data and uses the risk analysis model for calculations. The system then automatically fills in the relevant information in the simulated inspection module's input field. In the AI ​​Q&A module, entering the corresponding vessel's IMO number allows users to query historical inspection information.

[0117] (1) Task Overview Area: Displays basic ship information (synchronized from the database).

[0118] (2) Intelligent Checklist Area: This is the system's first intelligent output. The risk analysis model runs automatically in the background, querying the database (h_check_error, h_stop_error tables) for all historical records of the "Pengyuan" vessel over the past five years.

[0119] The model employs a combination of a rule engine and a machine learning classifier for analysis. The rule engine makes judgments based on preset rules (such as "a ship with a detention record in the past year is directly classified as high-risk"); the machine learning classifier uses historical data to perform weighted evaluations of features such as "frequency of recurrence of defects in the same part," "severity of historical detention defects," and "statistical defect rate of the flag state." Ultimately, the model outputs that the ship's risk level is "high-risk."

[0120] Analysis revealed that the vessel had defects recorded twice during its previous three inspections, specifically in the "fixed gas detection system in cargo hold spaces" and the "pressure relief valve in liquid cargo tanks." Therefore, based on the risk level and high-frequency defect analysis results, the system automatically highlighted and pinned these two items to the top of the generated electronic inspection checklist, labeling them as "historical high-frequency defects," prompting the PSCO to pay close attention.

[0121] (3) Knowledge-related area: When Li hovers his mouse over any inspection item in the list (such as "emergency fire pump"), this area will automatically retrieve the detailed inspection points and legal basis (such as SOLAS II-2 / 10.2.3.3.1) corresponding to the item from the special knowledge base and present them in a clear Markdown format.

[0122] 2. On-site auxiliary inspection: Intelligent on-site operation in parallel with dual lines PSCOs boarded the ship with mobile intelligent inspection terminal tablets, and on-site inspections were carried out collaboratively using the following two modes: Mode 1: Streamlined and visualized inspection based on a 3D electronic map of the inspection area PSCO activates the "Inspection Map Module" on the terminal. The interface displays a 3D model of an LNG vessel. Following the personalized inspection checklist, he clicks to enter the "Engine Room" area. The screen switches to display the inspection details for specific parts of the engine room. Clicking on the target equipment displays the corresponding inspection points.

[0123] When the "Emergency Generator" icon is clicked, a detailed panel slides out from the side of the screen, listing all the inspection items for the equipment and the relevant regulatory basis. Each item is checked, and the results are recorded directly through the application terminal: "Normal" or "Defective." For defects, a photo is immediately taken using the tablet, and the photo is automatically associated with that inspection item.

[0124] This method of "three-dimensional inspection area electronic map navigation and following the map" ensures the standardization and completeness of the inspection process, leaving no omissions.

[0125] Mode 2: Real-time question answering and intelligent adjudication based on a dual-engine AI assistant During an inspection of the cargo hold area, PSCO discovered that the fire damper mechanism of a ventilation duct was malfunctioning, but he was unsure of the specific defect description and convention clauses. Without leaving the site to consult paper documents, he directly activated the "AI Assistant" module on his terminal and input the following in his voice: "Cargo hold ventilation duct fire damper stuck."

[0126] (1) Intelligent background processing ① Intent analysis: The locally fine-tuned DeepSeek V3 671b model instantly analyzed the core of the problem: the object to be checked is "fireproof baffle", and the problem is "operation lag".

[0127] ② Dual-engine parallel retrieval: The system starts simultaneously. Regulatory knowledge engine: Based on the parent-child segmentation structure of the local "specialized knowledge base", the "five-dimensional association mapping network" accurately retrieves Article 9.7.5 of the IGC Rules regarding the provision that "fireproof baffles should be flexible to operate".

[0128] Historical data engine: Sends anonymous queries to the server, matches similar cases in the historical defect database, and obtains the most frequently corresponding defect code, such as "11108".

[0129] ③ Result Fusion and Generation: The AI ​​merges the two types of information and generates answer cards on the tablet using a fixed template (AI question-and-answer accuracy ≥ 95%). Defect code: 11108 Defect Description: Fire damper not operable Convention basis: IGC 9.7.5 Action Code: 17 (Correction before departure) PSCO quickly completed the recording based on this information, with the entire process taking no more than 10 seconds. This solved the core pain points of "inaccurate application of regulations and difficulty in remembering defect codes".

[0130] 3. Post-inspection: Automated report generation and one-click archiving After the on-site inspection, PSCO returned to the office. In the "Inspection Module" of the online management platform, the recorded process was entered, and the "Generate Report" button was clicked. The automatic report generation module was triggered and performed the following operations: (1) Automatically extract the ship information, inspector information, inspection time and location of this mission, and automatically associate the defect information and evidence recorded on site with the "five-dimensional association mapping network".

[0131] ① The recorded defect data is categorized by defect code (e.g., 11108) and automatically filled into the corresponding column of the Tokyo Memorandum Standard FORM B (Defect Report).

[0132] ② Fill in the summarized information, such as the number of defects and whether they are held up, into FORM A (Inspection Report).

[0133] ③ The photos of the fireproof barrier taken by Li were automatically associated with the defect "11108" under the "evidence attachment".

[0134] The system invokes the bilingual conversion unit to generate a draft of a Chinese-English bilingual PDF report.

[0135] PSCO only needs to conduct a final review, and once it confirms that everything is correct, it can be submitted. The report work that originally required 1-2 hours of manual completion can be completed in minutes, significantly improving efficiency.

[0136] 4. Data Closed Loop: Knowledge Accumulation and Intelligent System Evolution After PSCO submits its report, the system automatically initiates a data closed-loop process. The precise correspondence between the newly generated natural language processing question, "Flame baffle jamming in ventilation ducts," and the code "11108," along with all structured data from this inspection, is used as a high-quality training sample. This data is then encrypted and anonymized before being added to the system's feedback data pool. According to the system's iteration mechanism, this data will be used for: monthly incremental updates to the "Specialized Knowledge Base" and the "Five-Dimensional Association Mapping Network"; and quarterly fine-tuning of the parameters of the "Risk Analysis Model" and the "Natural Language Processing Model" of the AI ​​assistant module through reinforcement learning. This will enable the system to answer similar questions more accurately and quickly in the future, and its risk assessment capabilities will continue to evolve.

[0137] Implementation 3: Training Process Operation Step 1: 3D simulation training 1. Log in to the "Training Classroom - 3D Simulation Training" module of the PC online system and select the "Cabin Anti-contamination Equipment Inspection" scenario.

[0138] The system loads a 3D model of the engine room and marks the inspection points (such as "oil-water separator, oil discharge monitoring device"). The trainees click on the "oil-water separator" tab with the mouse, and the system pops up the inspection requirements: "separation effect, pipeline sealing, certificate validity".

[0139] The simulation revealed a defect of "oil-water separator pipeline leakage". After clicking "defect report", the system prompted the corresponding defect code "05103" based on the "five-dimensional correlation mapping network". The trainees followed the prompts to complete the defect entry and report generation.

[0140] Step 2: Question Bank Practice and Exams On mobile devices, users can open the "Training Check - Question Bank Practice" module. The system offers two practice modes: adaptive practice (based on the student's historical incorrect answers) and scenario-based targeted practice (e.g., for the "Pollution Prevention Equipment" topic). Students select the "Adaptive Practice" mode and choose the number of practice questions, such as 20. The system generates 20 single-choice, multiple-choice, and true / false questions based on the student's knowledge weaknesses. After completing the questions, the system automatically scores the student (out of 85 points) and displays explanations for incorrect answers (e.g., "Question 12 is incorrect; correct answer is C; legal basis: Article 20, Chapter III of the SOLAS Convention").

[0141] The administrator created an "LNG ship PSC inspection and assessment" through the "Training Classroom - Examination Management" module on the PC, selected the "scenario-based targeted test paper" strategy, and specified the scenario of "high-frequency defects in life-saving equipment". The system automatically extracted the corresponding questions from the question bank, set the number of questions to 50, the examination time to 90 minutes, and designated 10 PSCOs to participate in the examination.

[0142] After the exam, the system generates a group training effectiveness report and individual reports, and the administrator adjusts the training plan based on the reports.

[0143] Specific Implementation Four: Data Analysis and Decision-Making Implementation PSCO logs into the "Inspection Profile" module, selects the "Shenzhen Port 2024 January-June" time range, and the system calls the big data profile analysis unit to generate a port inspection profile: Chinese vessels accounted for 34% of the inspected vessels, the high-frequency defect "07105 - Lifesaving Equipment Defect" accounted for 68%, and 3 vessels were detained (all due to "lifesaving equipment failure"). This report can provide data-driven decision support for maritime authorities to optimize the inspection focus of LNG vessels at this port.

[0144] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A smart auxiliary method for port state control (PSC) inspections of liquefied natural gas (LNG) vessels, characterized in that, include: Obtain multi-source basic data for port state control inspections of liquefied natural gas vessels; The multi-source basic data is structured and associated with each other. The maritime regulations text is stored in a parent-child segmented semantic manner. Based on the stored structured regulations data, a special knowledge base is constructed that includes a five-dimensional association mapping network of inspection locations, defect codes, legal provisions, action codes, and defect phenomenon descriptions. In response to the identification information of the vessel to be inspected, the specialized knowledge base data is invoked, and a risk analysis model combining rule engine and machine learning classification is used to calculate and generate personalized inspection list data marked with risk level, high-risk inspection areas and inspection priority. Based on the personalized inspection checklist data, the inspection items are spatially bound to the 3D model of the liquefied natural gas vessel to generate visual inspection navigation data; During the on-site inspection, in response to the query command triggered by the visual inspection navigation data, the large language model, which has been fine-tuned by the maritime professional corpus, is called. The legal knowledge engine and the historical case engine are simultaneously triggered to perform dual-engine retrieval. After fusion reasoning, the structured inspection auxiliary response data is output. The defect data and evidence materials collected during on-site inspections are structured and automatically associated with the five-dimensional correlation mapping network to generate bilingual (Chinese and English) inspection report data that conforms to the Tokyo Memorandum of Understanding standards. The system integrates structured data from the entire inspection process, including visual inspection navigation call records, structured inspection auxiliary response data, and standardized inspection report data, as feedback to update the specialized knowledge base data and optimize the parameters of the risk analysis model and the large language model.

2. The intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to claim 1, characterized in that, The multi-source basic data undergoes structured processing and association mapping. Maritime regulatory texts are stored using parent-child segmentation semantic storage. Based on the stored structured regulatory data, a specialized knowledge base is constructed, comprising a five-dimensional association mapping network of inspection locations, defect codes, regulatory clauses, action codes, and defect descriptions. Collect historical inspection records of members of the Tokyo Memorandum, maritime law texts, professional knowledge data, and training question bank data as multi-source basic data; A semantic segmentation algorithm is used to divide the legal text into independent semantic units, using the chapter-section-article-clause numbering and preset logical connectors as semantic boundaries. The segmented semantic units are converted into a structured markup language format, and a parent-child segmentation architecture is used to complete the semantic storage of maritime regulations text; Based on data mining technology, a five-dimensional relational mapping network is constructed, which includes inspection locations, defect codes, legal and regulatory basis, action codes, and natural language descriptions of typical defect phenomena. The five-dimensional relational mapping network is calibrated through a collaborative mechanism of large model formatting, regularization correction, and manual verification to generate specialized knowledge base data.

3. The intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to claim 1, characterized in that, In response to the identification information of the vessel to be inspected, the system invokes the specialized knowledge base data and performs calculations using a risk analysis model that combines a rule engine with machine learning classification. This generates a personalized inspection checklist data labeled with risk level, high-risk inspection areas, and inspection priorities, including: Based on the IMO number or ship name identification information of the vessel to be inspected, the historical inspection records of the corresponding vessel, the flag state to which it belongs, and the statistical defect data of the fleet are retrieved from the special knowledge base. Rigid risk assessment is performed through a preset rule engine. The rigid risk assessment includes directly determining high risk if there are residual records within a preset period, upgrading the risk if there are recurrence records of the same inspection defect, and determining high-risk inspection areas if there are cumulative high-frequency defects in the same area. A random forest machine learning classifier trained on historical inspection data is used to comprehensively calculate the ship risk score by weighting factors such as the frequency of defect recurrence in the same part, the severity of historically detained defects, the flag state of the ship and the statistical defect rate of the fleet, and output the corresponding risk level result. Based on the risk level results and high-risk inspection areas, personalized inspection checklist data is generated by prioritizing high-risk areas.

4. The intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to claim 1, characterized in that, Based on the personalized inspection checklist data, the inspection items are spatially linked to the 3D model of the liquefied natural gas vessel to generate visualized inspection navigation data, including: Load a general 3D model of a liquefied natural gas vessel onto a mobile terminal and bind the inspection items in the personalized inspection checklist to their corresponding spatial locations in the 3D model one by one; Based on the bound 3D model, a visual inspection navigation path that supports touch zoom and rotation operations is generated; In response to trigger commands for the inspection area in the 3D model, the system outputs the basic information, inspection points, and relevant regulatory data of the corresponding inspection area in real time. High-risk area data and lightweight model parameters are cached locally on the mobile terminal, and the cached data is called to provide inspection and navigation services in weak network environments.

5. The intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to claim 1, characterized in that, During on-site inspections, in response to query commands triggered by visual inspection navigation data, a large language model fine-tuned from a maritime professional corpus is invoked. Simultaneously, a dual-engine retrieval process is initiated using both the regulatory knowledge engine and the historical case engine. After fusion reasoning, structured inspection auxiliary response data is output, including: Perform intent recognition and entity extraction on received natural language or voice query commands triggered by the visual inspection navigation interface, and extract search keywords; The system utilizes a large language model finely tuned with professional maritime corpus and historical question-and-answer pairs to trigger a dual-engine search based on search keywords. Specifically, the regulatory knowledge engine performs precise matching of regulatory clauses based on a specialized knowledge base, while the historical case engine performs similarity matching of similar defect cases based on a historical inspection database. The system integrates and reasons the regulatory data obtained from dual-engine retrieval with historical case data, and outputs structured response data that includes defect codes, legal basis in Chinese and English, graded handling suggestions, and corresponding action codes.

6. The intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to claim 1, characterized in that, The defect data and evidence collected during on-site inspections are structured and automatically associated with the five-dimensional relational mapping network to generate bilingual (Chinese and English) inspection report data conforming to the Tokyo Memorandum of Understanding standards, including: The defect classification information, text descriptions, images and video evidence materials collected during the on-site inspection based on the visual inspection navigation path are structured and organized to obtain a standardized defect dataset; Automatically associate and match the standardized defect dataset with the five-dimensional relationship mapping network in the specialized knowledge base to supplement and improve the corresponding data of defect codes, legal basis, and action codes; The matched full data will be automatically populated into the standard FORM A and FORM B report templates of the Tokyo Memorandum. The report content is automatically switched to Chinese and English based on a maritime bilingual terminology database, generating inspection report data in both Chinese and English.

7. The intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to claim 1, characterized in that, The system integrates structured data from the entire inspection process, including visual inspection navigation call records, structured inspection auxiliary response data, and standardized inspection report data, as feedback to update the specialized knowledge base data. It also optimizes the parameters of the risk analysis model and the large language model, including: Collect structured feedback data throughout the entire inspection process. The feedback data includes defect information, handling results, differences between AI suggestions and official handling data, and correlation data of newly emerging defects. According to the preset first cycle, the specialized knowledge base and the five-dimensional relational mapping network are incrementally updated through a process of automatic formatting, regular expression correction, and manual verification. According to the preset second cycle, based on the collected feedback data, the weighted parameters of the risk analysis model and the large language model are fine-tuned through reinforcement learning.

8. The intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to claim 1, characterized in that, It also includes training and assessment steps based on the specialized knowledge base data and the 3D model, the training and assessment steps including: Based on a personalized inspection checklist or a specified high-risk inspection scenario, the corresponding 3D model inspection parts and related inspection items are dynamically loaded to generate a 3D simulation inspection task, complete the full-process simulation training of the inspection, and generate a simulation inspection record. Based on the training question bank data of label classification, special assessment test papers are generated by adopting adaptive test paper generation or scenario-based targeted test paper generation strategies. Among them, adaptive test paper generation is based on matching questions with students' historical wrong questions and knowledge weaknesses, while scenario-based targeted test paper generation is based on matching questions with specific ship risk profiles or port high-frequency defect types. The assessment results are quantitatively scored, and a training effectiveness evaluation report and targeted reinforcement suggestions are generated, linking the corresponding learning materials in the specialized knowledge base.

9. An intelligent auxiliary system for port state control (PSC) inspections of liquefied natural gas (LNG) vessels, characterized in that: The intelligent auxiliary method for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to any one of claims 1 to 8 includes: A specialized knowledge base module is used to store and manage the structured semantic units, historical inspection cases, professional knowledge, and the multidimensional association mapping network; The intelligent judgment and service engine integrates the risk analysis model and the natural language processing model to provide computing and reasoning services for the online integrated management platform and mobile intelligent inspection terminal. An online integrated management platform is used for checking task management, data analysis, knowledge base maintenance, and training assessment. The mobile intelligent inspection terminal integrates 3D visualization inspection navigation, AI assistant, and report generation functions for on-site operations. The data closure and iteration module automatically collects data during the inspection process and the correlation data of newly emerging defects. It updates the specialized knowledge base and five-dimensional correlation mapping network monthly through an automatic formatting-regular expression correction-manual verification process. It also fine-tunes the risk analysis model parameters and natural language processing model quarterly based on the collected data. The online integrated management platform and the mobile intelligent inspection terminal collaborate through a data interface, sharing and calling the specialized knowledge base module and intelligent service engine.

10. The intelligent auxiliary system for port state control (PSC) inspection of liquefied natural gas (LNG) vessels according to claim 9, characterized in that, The mobile intelligent inspection terminal has edge computing capabilities, and in weak network environments, it can provide basic inspection navigation and question-and-answer services based on locally cached high-risk area data and lightweight models.

Images