Ship user behavior self-learning recommendation system based on large model

Abstract

The invention relates to a ship user behavior self-learning recommendation system based on a large model, and relates to the technical field of ship informatization. According to the system, through collection and fusion of multi-source heterogeneous ship user behavior data, a large-scale pre-training language model (large model) is utilized to carry out deep understanding and semantic mining on massive ship field text information and user behavior sequences, and a ship field knowledge graph or semantic vector space is constructed. The large model can identify and predict potential demands, behavior patterns and preference changes of ship users, and generates highly personalized, accurate and prospective ship service, product, route or information recommendations in combination with real-time operation data and external environment factors. Besides, a user feedback self-learning mechanism is introduced into the system, recommendation strategies and model parameters are continuously optimized according to interaction behaviors and explicit evaluation of the users in modes of reinforcement learning or continuous learning and the like, and intelligent iteration of the system and continuous improvement of the recommendation effect are achieved. According to the method, the challenges of a traditional recommendation system in the aspects of data complexity, semantic gaps and dynamic demand adaptability in the ship field are effectively solved, and the ship operation efficiency and the user satisfaction degree are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship informatization, and particularly relates to a ship user behavior self-learning recommendation system based on a large model. BACKGROUND

[0002] With the rapid development and digital transformation of global shipping industry, the needs of ship users (including shipowners, ship management companies, crew, port operators, cargo owners, etc.) in ship services, spare parts procurement, route planning, crew training, insurance and finance are increasingly diversified and complex. Traditional ship service recommendation systems mainly rely on rule-based, collaborative filtering, content recommendation and other technologies. These methods have many limitations in handling massive heterogeneous data, understanding complex user intentions and adapting to rapidly changing market demands:

[0003] Data heterogeneity and fusion difficulty: The data sources in the ship field are extensive, including ship AIS (Automatic Identification System) data, port docking records, maintenance records, material procurement lists, crew resumes, maritime regulations, shipping news, weather forecasts, etc. The data formats are diverse, making it difficult to effectively integrate and utilize them.

[0004] Insufficient semantic understanding: Traditional methods are difficult to perform deep semantic understanding on unstructured text data related to ships (such as maintenance logs, maritime reports, user inquiries), resulting in an inability to accurately grasp the true intentions and potential needs of users.

[0005] Lack of recommendation accuracy and personalization: Existing systems often only provide coarse-grained recommendations, lack detailed portraits of individual ships or users, and are difficult to meet the growing demand for personalization.

[0006] Poor dynamic adaptability: The needs of ship users, ship status, market environment and other factors are dynamically changing, and traditional models are difficult to capture these changes in real time and adjust the recommendation strategy in a timely manner.

[0007] Cold start problem: For new users or new ships, traditional recommendation systems are difficult to provide effective recommendations due to the lack of historical behavior data.

[0008] Lack of self-learning ability: Most systems lack the ability to self-optimize and learn based on user feedback, and require manual intervention to update the recommendation logic. SUMMARY

[0009] The present application relates to the technical field of ship informatization, and particularly relates to a ship user behavior self-learning recommendation system based on a large model.

[0010] Technical solution: A ship user behavior self-learning recommendation system based on a large model, including five components: a multi-source data acquisition and fusion module, a ship domain large model module, a user behavior understanding and demand prediction module, a personalized recommendation generation module, and a user feedback and self-learning optimization module.

[0011] Among them, the multi-source data acquisition and fusion module is used to collect and integrate multi-source heterogeneous ship-related data, and to clean, standardize and feature extract the data;

[0012] The ship domain large model module deploys one or more large-scale pre-trained language models or multi-modal large models, which are pre-trained and / or fine-tuned on massive ship domain corpus, enabling semantic understanding and knowledge reasoning capabilities in the ship domain;

[0013] The user behavior understanding and demand prediction module is connected to the ship domain large model module, and is used to receive user input, historical interaction behavior, ship current state and external environment data, and use the ship domain large model to analyze the above data in depth to understand user intent, build user and ship portraits, and predict potential user needs;

[0014] The personalized recommendation generation module is connected to the user behavior understanding and demand prediction module, and generates personalized recommendation results based on the output of the user behavior understanding and demand prediction module, combined with the feature information of recommended items or services, using the generation or ranking capabilities of the ship domain large model;

[0015] The user feedback and self-learning optimization module is connected to the personalized recommendation generation module, and is used to collect user feedback data on the recommendation results, and input the feedback data into the ship domain large model, dynamically adjusting the model parameters and recommendation strategies through reinforcement learning or continuous learning mechanisms.

[0016] In further embodiments, the data collected by the multi-source data acquisition and fusion module includes ship static data, ship dynamic data, historical operation data, transaction data, user interaction data, maritime regulations and news, and weather and sea state data.

[0017] In further embodiments, the large model in the ship domain large model module has at least one of the following capabilities: semantic understanding capability, knowledge reasoning capability, behavior pattern recognition capability, and multi-modal understanding capability.

[0018] In further embodiments, the potential needs predicted by the user behavior understanding and demand prediction module include ship maintenance needs, required spare parts, crew skill training needs, port supply needs, or suitable insurance products.

[0019] In further embodiments, the recommendation results generated by the personalized recommendation generation module include ship services, ship products, optimized routes, crew training courses, maritime regulations interpretation, or industry information.

[0020] In further embodiments, the feedback data collected by the user feedback and self-learning optimization module includes clicks, browsing time, purchases, evaluations, collections, shares, or rejections.

[0021] In further embodiments, the user feedback and self-learning optimization module optimizes the model through reinforcement learning from human feedback (RLHF) or continuous learning.

[0022] In further embodiments, the recommendation result presentation and interaction module, connected to the personalized recommendation generation module, is used to present the recommendation results to the user and support user search, filtering, feedback interaction operations.

[0023] In further embodiments, the recommendation method of the large model-based ship user behavior self-learning recommendation system includes the following steps:

[0024] S1, multi-source data collection and preprocessing: collecting multi-source heterogeneous ship-related data, and cleaning and standardizing the data;

[0025] S2, ship domain large model construction and fine-tuning: selecting a general large-scale pre-trained language model as a base model, and pre-training and / or fine-tuning the base model using massive ship domain corpus to construct a ship domain large model;

[0026] S3, deep understanding of ship user behavior and demand: receiving user query information, historical interaction behavior, ship current state and external environment data, and using the ship domain large model to deeply analyze the information and data to understand user intent, construct user and ship portraits, and predict user's potential demand;

[0027] S4, personalized recommendation result generation: retrieving relevant content from the recommendation candidate set according to the understood user demand and prediction results, and using the ship domain large model to sort, filter and combine the retrieved candidate content to generate personalized recommendation results;

[0028] S5, user feedback collection and self-learning optimization: collecting user interaction feedback on the recommendation results, and using the feedback data as reward signals or training samples to continuously self-learn and optimize the ship domain large model and recommendation strategy through reinforcement learning or continuous learning mechanism.

[0029] In further embodiments, in step S3, the ship domain large model understands user intent and predicts potential needs by performing semantic analysis, intent recognition, and knowledge reasoning on the information and data. In step S4, the ship domain large model can also generate explanatory text for the recommended results. In step S5, the self-learning optimization includes adding new user interaction data to the training set and periodically or incrementally fine-tuning the ship domain large model. In step S6, the self-learning optimization also includes using reinforcement learning combined with user feedback to optimize the recommendation strategy of the large model.

[0030] In further preferred embodiments, the recommendation method of the large model-based ship user behavior self-learning recommendation system further includes a recommendation result presentation: presenting the personalized recommendation results to ship users through a user interface or an API interface.

[0031] Benefits: The present application relates to a large model-based ship user behavior self-learning recommendation system, and relates to the field of ship informatization technology. It has the following benefits:

[0032] 1. Greatly improves the accuracy and personalization level of recommendations: using the powerful semantic understanding and reasoning capabilities of large models, complex and variable ship user needs and potential intentions can be deeply mined, providing more accurate recommendations than traditional methods.

[0033] 2. Enhances the processing capability of heterogeneous data: large models can effectively fuse and understand ship data from different sources and formats, solving the data fusion problem of traditional methods.

[0034] 3. Achieves self-learning and dynamic adaptability of the recommendation system: through real-time learning mechanisms based on user feedback, the system can continuously optimize itself to adapt to dynamic changes in ship user needs, market environment, and regulatory policies.

[0035] 4. Supports forward-looking and preventive recommendations: large models can predict potential problems or future needs in ship operation, enabling a shift from passive response to proactive recommendation, such as early warning of equipment failure and recommendation of preventive maintenance.

[0036] 5. Improves ship operation efficiency and user experience: helps ship users quickly and accurately find the services and products they need, reduces information screening time, optimizes the decision-making process, and reduces operating costs.

[0037] 6. Solves the cold start problem: large models can effectively infer through the association of small amounts of new user / ship data with existing knowledge, combined with domain general knowledge, to alleviate the cold start effect.

[0038] 7. Explainability enhancement: Large models have the ability to generate text, which can provide explanations for recommended results, enhancing user trust in recommendations. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The system overall architecture schematic diagram of the whole application.

[0040] Figure 2 The multi-source data acquisition and processing flowchart of the application.

[0041] Figure 3 The ship field large model construction and fine-tuning flowchart of the application.

[0042] Figure 4 The user behavior understanding, recommendation generation and feedback self-learning flowchart of the application. DETAILED DESCRIPTION

[0043] In the following description, a large number of specific details are given in order to provide a more thorough understanding of the application. However, it is obvious to those skilled in the art that the application can be implemented without one or more of these details. In other examples, some technical features known in the art are not described in order not to obscure the application.

[0044] The application relates to a ship user behavior self-learning recommendation system based on a large model, comprising:

[0045] A multi-source data acquisition and fusion module is used for acquiring and integrating multi-source heterogeneous ship-related data, including but not limited to ship static data (ship type, ship age, load), dynamic data (AIS, speed, heading), historical operation data (maintenance record, fuel consumption, port of call), transaction data (purchase order, service agreement), user interaction data (search record, click, evaluation), maritime regulations and news, weather and sea state data, etc.; and the data is cleaned, standardized, structured and feature extracted.

[0046] A ship field large model module is used for deploying one or more large-scale pre-trained language models (LLM) or multi-modal large models, which are pre-trained and / or fine-tuned with massive ship field corpus (such as maritime documents, technical manuals, industry reports, shipping news, ship forum discussions, historical transaction texts, etc.), so as to have:

[0047] Strong semantic understanding ability: capable of understanding professional terms, complex sentence patterns and context in the ship field.

[0048] Knowledge reasoning ability: capable of extracting entities and relationships from text, constructing a ship field knowledge graph or performing logical reasoning.

[0049] Behavioral pattern recognition capability: It can analyze users' historical behavior sequences to identify potential behavioral patterns and demand trends.

[0050] Multimodal understanding capability (optional): Able to process non-textual data such as images and charts, improving the comprehensive understanding of ship status and environment.

[0051] The user behavior understanding and demand prediction module utilizes the aforementioned large-scale model in the shipping field to conduct in-depth analysis of user-input queries, historical interaction behaviors, current ship status, route information, and external environmental data. This allows the module to understand user intent, construct fine-grained user and ship profiles, and predict potential user needs (e.g., upcoming ship maintenance, required spare parts, crew skills training needs, resupply needs at the next port, suitable insurance products, etc.).

[0052] Personalized Recommendation Generation Module: Based on the output of the user behavior understanding and demand prediction module, and combined with the feature information of recommended items or services, this module leverages the generation or ranking capabilities of large models to generate highly personalized, diverse, and interpretable recommendation results. Recommended content may include ship services (such as repair, refueling, and port agency services), ship products (such as spare parts and navigation equipment), optimized routes, crew training courses, interpretations of maritime regulations, or industry information.

[0053] The user feedback and self-learning optimization module collects user feedback data on recommendation results, including implicit / explicit feedback such as clicks, browsing time, purchases, ratings, favorites, shares, or rejections. This feedback data is input into a large model, and through mechanisms such as reinforcement learning from human feedback (RLHF), continuous learning, or model fine-tuning, the model parameters and recommendation strategies are dynamically adjusted, enabling the system to self-optimize and iterate based on changes in actual user behavior and preferences.

[0054] Recommendation results presentation and interaction module: Provides a user-friendly interface (such as web, mobile app or API interface) to present recommendation results to users in an intuitive and easy-to-understand way, and supports users to perform interactive operations such as searching, filtering and feedback.

[0055] Based on the aforementioned destacking system, this invention proposes a self-learning recommendation method for ship user behavior based on a large model, comprising the following steps:

[0056] Step S1: Multi-source data acquisition and preprocessing:

[0057] Collecting ship static and dynamic data, historical operation and transaction data, user interaction data, maritime regulations and news, weather and sea conditions, and other heterogeneous data from various data sources.

[0058] Performing preprocessing operations such as data cleaning, deduplication, format conversion, and missing value filling, to unify data format and encoding.

[0059] Processing unstructured text data such as word segmentation, named entity recognition, and keyword extraction, to convert it into a form that large models can understand.

[0060] Step S2: Ship domain large model construction and fine-tuning:

[0061] Select a general large-scale pre-trained language model as the base model.

[0062] Collect massive ship-specific unstructured text data (such as maritime reports, ship technical specifications, port operation procedures, crew logs, etc.) and structured data (such as ship attribute tables, service lists, historical transaction records, etc.), and convert them into text form.

[0063] Use these ship domain data to pre-train the base large model (if starting from scratch) or domain-adaptive fine-tuning (such as parameter-efficient fine-tuning LoRA, Prompt Tuning, etc.), to enable it to have deep ship domain knowledge and semantic understanding capabilities.

[0064] Step S3: Deep understanding of ship user behavior and demand:

[0065] Receive user query information and context information.

[0066] Combine user historical behavior sequences, current ship state (obtained through AIS, sensor data, etc.), route information, and real-time external environment data (such as weather, port congestion), to build a comprehensive user-ship-environment context.

[0067] Use the above context information as input to the large model, and use the large model's semantic understanding and reasoning capabilities to analyze user intent, identify potential needs, and predict future possible needs, such as predicting maintenance needs for specific components by analyzing the ship's current location and historical failure records, or predicting fuel supply and port agent needs by analyzing the route and destination port.

[0068] Build or update fine-grained user portraits and ship portraits.

[0069] Step S4: Generation of personalized recommendation results:

[0070] Retrieving relevant ship services, products, routes, or information from the recommendation candidate set based on the user demand and prediction results understood in step S3.

[0071] Sorting, filtering, and combining the retrieved candidate content using the ship domain large model to generate a highly personalized, accurate, and diverse recommendation list.

[0072] The large model can generate explanatory text for the recommendation results based on the demand, enhancing the transparency and user trust of the recommendations.

[0073] Step S5: User feedback collection and self-learning optimization:

[0074] Collecting user interaction feedback on the recommendation results in real time or periodically, such as click-through rate, dwell time, purchase conversion rate, evaluation star rating, likes / dislikes, favorites, etc.

[0075] Using these user feedback data as reinforcement learning reward signals or training samples for model fine-tuning, continuously self-learning and optimizing the ship domain large model and recommendation strategy.

[0076] For example, through reinforcement learning, the model can learn how to generate recommendations that are more satisfying to users; through supervised learning, the model can adjust the ranking of recommended content based on explicit user evaluations.

[0077] Periodically or according to data volume thresholds, retrain or incrementally learn the model to ensure the recommendation system always maintains optimal performance.

[0078] Step S6: Presentation of recommendation results:

[0079] Presenting personalized recommendation results to ship users intuitively and clearly through user interfaces or API interfaces.

[0080] The specific embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings of the present application.

[0081] The present application provides a large model-based ship user behavior self-learning recommendation system and method. The system utilizes the powerful semantic understanding, knowledge reasoning, and generation capabilities of advanced large-scale pre-training models (LLM), combined with the unique complex data and business scenarios in the ship domain, to achieve deep mining and accurate prediction of ship user demand, and has the ability to continuously optimize itself through user feedback.

[0082] I. System architecture (refer to Figure 1 )

[0083] As shown in Figure 1 , the large model-based ship user behavior self-learning recommendation system described in the present application mainly includes the following modules:

[0084] 1. Data Collection Layer:

[0085] Vessel Static and Dynamic Data: AIS data (vessel position, speed, heading), GPS data, vessel registration information, vessel type, carrying capacity, year of construction, etc.

[0086] Vessel Operational Data: Historical maintenance records, fuel consumption data, equipment operational status data, sensor data, historical voyage reports, port docking records, etc.

[0087] Transaction and Service Data: Historical purchase orders, service contracts, voyage charter agreements, insurance records, etc.

[0088] User Interaction Data: Search queries, click behavior, browsing duration, purchase records, reviews, favorites, shares, etc. on the recommendation platform.

[0089] External Environment and Regulation Data: Real-time weather and sea state data, shipping market reports, maritime regulations, IMO (International Maritime Organization) conventions, port policies, etc.

[0090] Industry Knowledge Base: Vessel technical manuals, equipment manuals, industry standards, maritime accident reports, industry news, etc.

[0091] 2. Data Processing and Fusion Layer:

[0092] Data Cleaning and Standardization: Perform operations such as deduplication, error correction, format unification, and missing value handling on collected heterogeneous data.

[0093] Structured and Unstructured Data Fusion: Fuse different types of data such as text, numerical, and time series, for example, associate unstructured text descriptions in maintenance records (such as "diesel engine startup difficulty with abnormal noise") with structured maintenance component IDs, maintenance time, etc.

[0094] Feature Engineering: Extract useful features from raw data, such as vessel usage age, historical failure frequency, user activity level, route busy degree, etc.

[0095] Knowledge Graph Construction (Optional but Recommended): Based on vessel domain text and structured data, construct a knowledge graph of vessel entities (such as vessels, ports, equipment, service providers, crew) and relationships (such as "located in", "provided by", "faulty component is"), providing structured domain knowledge for large models.

[0096] 3. Vessel Domain Large Model Module:

[0097] Basic Large Model Selection: Can be a pre-trained language model based on the Transformer architecture, such as the GPT series, BERT series, Llama series, or open-source variants thereof.

[0098] Domain-adaptive training:

[0099] Continuous pre-training: Further pre-training on massive ship domain text corpus to make the general large model learn more deeply about the language patterns, terminologies, and knowledge of the ship industry.

[0100] Instruction fine-tuning: By constructing ship domain-specific instruction-response pair datasets (e.g., "Describe the recent maintenance needs of ship A" ≥ "Ship A's diesel engine may need to be checked, and its historical data shows that there is an increased risk of failure after XXX hours of operation"), the model's performance on specific tasks is improved.

[0101] Parameter-efficient fine-tuning (PEFT, such as LoRA, Prompt Tuning): By introducing a small number of trainable parameters or adjusting input prompts, the large model is efficiently adapted to the ship recommendation task while keeping most of the original model parameters unchanged.

[0102] 4、User behavior understanding and demand prediction module:

[0103] Multi-dimensional context construction: Receive user queries (such as "My cargo ship needs to be replenished"), historical interaction records, real-time state data of the ship (such as current speed, destination port), external environment data (such as congestion at the target port, recent fluctuations in oil prices), etc.

[0104] Large model inference analysis: Provide the above context as input to the ship domain large model. The large model uses its powerful semantic understanding and reasoning capabilities:

[0105] Intention recognition: Accurately identify the real intention behind the user's query (for example, "replenishment" may refer to fuel, water, food, or spare parts).

[0106] Demand prediction: Combine historical data of the ship and domain knowledge to predict the user's potential or future needs (for example, by analyzing the port the ship will soon arrive at, historical replenishment patterns, and fuel consumption rate, predict the amount of fuel replenishment and the best replenishment time).

[0107] Risk assessment: Combine ship equipment operation data and historical failure patterns to predict potential equipment failure risks and recommend preventive maintenance.

[0108] User and ship portrait update: Continuously update the user's preferences, service habits, and the ship's health status, operation characteristics, and other portrait information.

[0109] 5、Personalized recommendation generation module:

[0110] Candidate Set Generation: Based on the outputs of the user behavior understanding and demand prediction modules, a preliminary set of recommendation candidates is generated from the vast repository of ship services, products, routes, and information knowledge. This is achieved through semantic similarity matching, rule filtering, or vector retrieval.

[0111] Large Model Re-ranking and Generation: The recommendation items in the candidate set, along with their features and the user's current context information, are input again into the ship domain large model. The large model performs:

[0112] Fine-tuned Ranking: Based on user preferences, ship requirements, service provider reputation, price, distance, and other factors, the candidate recommendation items are ranked again.

[0113] Recommended Content Generation: Not only does it provide a list of recommendations, but it also generates detailed reasons for the recommendations, service introductions, comparative analyses, or operational suggestions, enhancing the transparency and practicality of the recommendations.

[0114] Diversity and Novelty: By introducing diversity penalties or exploration mechanisms, the homogeneity of recommended content is avoided, ensuring that users are exposed to new but still relevant services or products.

[0115] 6、User Feedback and Self-learning Optimization Module:

[0116] Feedback Data Collection: Real-time capture of all user interactions with the recommended results, such as clicks, browsing time, favorites, shares, purchases, explicit evaluations (ratings, reviews), complaints, etc.

[0117] Reward Signal Design: These behaviors are converted into reward signals for model learning. For example, purchase behavior receives high rewards, clicks receive moderate rewards, and neglect or negative evaluations receive low or negative rewards.

[0118] Model Optimization Mechanism:

[0119] Reinforcement Learning (RLHF): Use user feedback as a reward for reinforcement learning to train a reward model to evaluate the quality of recommendations and further guide the large model (such as through the PPO algorithm) to optimize its recommendation strategy, making it generate recommendations that better meet user expectations.

[0120] Continuous Fine-tuning / Incremental Learning: Add new user interaction data to the training set and regularly or incrementally fine-tune the ship domain large model to enable it to quickly adapt to user preferences and market changes.

[0121] A / B Testing and Multi-armed Bandit: Used to evaluate the performance of different recommendation strategies or model versions and dynamically allocate traffic to find the best practices.

[0122] 7、Recommendation Result Presentation and Interaction Module:

[0123] Provide multi-platform user interfaces such as web applications, mobile apps, and API interfaces to ensure that recommendation results can easily reach users.

[0124] Recommendation results are displayed in intuitive lists, cards, charts, etc., and provide search, filtering, comparison, one-click ordering / reservation, online consultation, and other interactive functions.

[0125] II. Method flow (refer to Figure 2 、 Figure 3 、 Figure 4 )

[0126] The main steps of the ship user behavior self-learning recommendation method based on large models are as follows:

[0127] Step S1: Multi-source data collection and preprocessing (refer to Figure 2 )

[0128] S101: Data source access: Access various ship-related data sources through API interfaces, database connections, crawlers, etc.

[0129] S102: Data collection: Collect data in real time or in batches, including ship static data (such as ship type, registration location), dynamic data (such as AIS signals), operational data (such as fuel consumption, maintenance records), transaction data (such as service orders), user behavior data (such as search terms, click streams), maritime regulations and news texts, and external environmental data such as weather and sea conditions.

[0130] S103: Data cleaning and standardization: Denoise, de-duplicate, handle outliers, and fill in missing values for the collected raw data. Unify different formats of data into structured formats (such as JSON, CSV), and encode and unify text data (such as UTF-8).

[0131] S104: Feature extraction and representation: Extract key features from raw data, such as converting time series data into statistical features (mean, maximum, trend), performing word segmentation on text data, word embedding (such as Word2Vec, BERT Embedding), or constructing named entity recognition (NER) results.

[0132] S105: Data fusion and storage: Fuse processed data by user ID, ship ID, timestamp, etc. and store it in a distributed database or data lake suitable for big data analysis. Optionally, build a ship domain knowledge graph to explicitly store entities and relationships.

[0133] Step S2: Ship domain large model construction and fine-tuning (refer to Figure 3 )

[0134] S201: Base Model Selection and Loading: Select a pre-trained general-purpose large model (such as LLaMA2, GPT series, etc.) as the foundation for the recommendation system.

[0135] S202: Ship Domain Corpus Construction: Collect and organize large-scale ship industry-related text corpus (such as maritime laws and regulations, ship technical manuals, shipping news, maritime reports, industry journals, ship management documents, historical consultations and fault records, etc.), and convert structured data (such as ship attributes, service catalog) into text descriptions.

[0136] S203: Domain Adaptation Pre-training (if needed): Continue pre-training the base large model on the above ship domain corpus to enhance its understanding of ship professional knowledge and language patterns.

[0137] S204: Task-oriented Fine-tuning (Instruction Fine-tuning / SFT): Construct a series of instruction-response pair datasets related to ship recommendation tasks, such as: "Please recommend the nearest fuel supply port for an oil tanker" -> "Recommend port X and port Y, which are closest to your current location and have moderate fuel prices." Use this dataset to supervise the fine-tuning of the large model, so that it better follows the ship recommendation instructions.

[0138] S205: Reward Model Training (RLHF Preposition): If RLHF is used, an additional reward model needs to be trained. By artificial labeling or crowdsourcing, the quality of the recommendation results generated by the large model is sorted, and a reward model that can evaluate the quality of the recommendation is trained.

[0139] S206: Reinforcement Learning Fine-tuning (RLHF, optional): Use reinforcement learning algorithms (such as PPO) combined with the reward model to further fine-tune the large model, so that it generates more human-preferred and higher-quality recommendation results.

[0140] Step S3: Deep understanding of ship user behavior and demand (refer to Figure 4 )

[0141] S301: Receive user input and context: The system receives queries from the user interface (such as "My container ship needs maintenance"), the current ship state (such as engine running hours, historical fault records), ship location, destination, and real-time weather, port congestion, and other external environment information.

[0142] S302: Context encoding and large model input: Integrate the above multi-modal, multi-source context information into an input sequence that the large model can process. This may include converting structured data into descriptive text, using specific Prompt templates to guide the large model to understand the intent.

[0143] S303: Large model intent recognition and semantic analysis: The ship domain large model performs deep semantic analysis on the input context, identifying the user's core needs and potential intent. For example, from "container ship maintenance", identify "maintenance type" (regular maintenance, fault maintenance), "maintenance components" (engine, hull), "time urgency", etc.

[0144] S304: User and ship portrait update: The large model continuously updates the user's interest preferences, consumption habits, and the ship's health status, operating route characteristics, etc. based on current interactions and historical behavior.

[0145] S305: Demand prediction and reasoning: Based on the large model's understanding of historical data and real-time context, predict future user needs. For example, based on ship route and international regulation changes, predict that crew members need new training certificates; based on fuel consumption trends, predict the next replenishment point and time.

[0146] Step S4: Personalized recommendation result generation (refer to Figure 4 )

[0147] S401: Candidate set recall: Based on the user understanding module's demand prediction, from a pre-established service / product / route / information knowledge base, through keyword matching, vector retrieval (such as similarity search through embedding vectors generated by the large model), or rule matching, etc., recall a batch of relevant recommendation candidates.

[0148] S402: Large model sorting and refinement: Input the recalled candidate set together with the user context into the ship domain large model. The large model sorts the candidate set according to its learned user preferences, service / product features, timeliness, price, reputation, etc., and generates the final recommendation list. The large model can recommend different types or levels of services for different users based on user portraits and ship portraits.

[0149] S403: Recommendation explanation and diversity generation: The large model can generate explanatory text to explain the recommendation reasons, such as "We recommend this maintenance service for you because it is closest to your current port and has good reputation on similar ship types." At the same time, by adjusting the generation strategy, ensure the diversity of the recommendation list, avoid the recommendation items too single.

[0150] Step S5: User feedback collection and self-learning optimization (refer to Figure 4 )

[0151] S501: Real-time feedback collection: Monitor all user interaction behaviors in the recommendation system: clicks, browsing time, favorites, shares, purchases / reservations, user-submitted ratings, comments, error correction feedback, etc.

[0152] S502: Reward signal calculation: Convert the collected user behavior into a reward signal for model learning. For example, successful service booking gets a high positive reward, and negative comments from users on recommendations get a negative reward.

[0153] S503: Model iteration and optimization:

[0154] Continuous fine-tuning: Incrementally fine-tune the ship domain large model using new, cleaned user feedback data (such as high-quality interaction sequences, error correction samples) to enable it to quickly adapt to new user preferences and market dynamics.

[0155] Reinforcement learning (RLHF): Use reward models and RL algorithms such as PPO to further optimize the large model's performance on the recommendation task, so that the generated recommendations can maximize user satisfaction.

[0156] Offline evaluation and online A / B testing: Regularly conduct offline evaluations to monitor model performance and verify the actual effects of new models or new strategies through online A / B testing.

[0157] S504: Model update and deployment: The verified and optimized model version is deployed to the production environment, replacing the old model, and realizing continuous self-learning and iterative optimization of the recommendation system.

[0158] Step S6: Presentation of recommendation results

[0159] Through various forms such as web, mobile app, API interface, etc., the personalized recommendation results are intuitively and conveniently presented to users.

[0160] Embodiment:

[0161] Suppose a bulk carrier is sailing from Brazil to China, and its crew queries "recent hull maintenance services" through the system.

[0162] 1. Data collection and preprocessing: The system collects static information of the bulk carrier (ship type, age), dynamic information (current position, estimated arrival time), historical maintenance records (last hull maintenance time), service provider information in the port of call, requirements for hull cleaning in maritime regulations, and historical user evaluations of hull maintenance services, etc.

[0163] 2. Large model fine-tuning: The ship domain large model has been fully fine-tuned through massive maritime reports, hull anti-pollution technical specifications, port service agreements, and other corpus.

[0164] 3. User behavior understanding and demand prediction:

[0165] The user inputs "hull maintenance service", and the large model identifies that the intention is to seek hull cleaning and anti-pollution services.

[0166] In combination with the current location of the ship, the estimated arrival at the port (such as Shanghai), the age of the ship, the last maintenance time, and the regulations on anti-pollution coatings in specific sea areas, the large model predicts that the ship needs to be cleaned immediately after arriving at the Chinese port and may need to be repainted with a specific type of anti-pollution paint.

[0167] The large model also infers from historical data that the ship is sensitive to service prices and tends to choose suppliers that provide Chinese services.

[0168] 4. Personalized recommendation generation:

[0169] The system recalls a list of suppliers that provide hull maintenance services near the Shanghai port.

[0170] The large model sorts the recalled suppliers according to the predicted demand (hull cleaning + anti-pollution paint), the user profile (price sensitivity, preference for Chinese services), and the ship profile (bulk carrier, age, historical maintenance records).

[0171] The recommendation results may include the top three hull maintenance service providers, with specific content of their services, estimated cost range, user reviews, and special mention of Chinese service support. The large model can also generate an explanation: "Based on your ship type, route, and the port you will soon arrive at, we recommend the following hull maintenance service providers. Company A has rich experience in cleaning and anti-pollution technology, and Company B offers competitive prices and provides Chinese services."

[0172] 5. User feedback and self-learning optimization:

[0173] If the user clicks and finally chooses Company B and gives a good review, the system will record this positive feedback.

[0174] This feedback is used to continuously fine-tune the large model, so that in future recommendations, it will be more inclined to recommend service providers with good reputation and user preferences (such as Chinese services, price sensitivity) for similar users and ships.

[0175] If the user chooses a company C that was not recommended and notes the reason as "provides integrated hull inspection services", the system will capture this information and use it to optimize the large model so that it can identify and recommend more comprehensive and complete service packages in the future.

[0176] Through the above examples, the present application can provide more accurate, personalized, and forward-looking service recommendations for ship users, greatly improving the efficiency of ship operation and user experience.

[0177] While the application has been described and illustrated with reference to specific preferred embodiments, it is not intended that it be limited to these particulars. Various changes in form and detail can be made without departing from the spirit and scope of the application as defined by the appended claims.

Claims

1. A large model-based ship user behavior self-learning recommendation system, characterized by The system comprises: a multi-source data collection and fusion module for collecting and integrating multi-source heterogeneous ship-related data, and cleaning, standardizing and feature extracting the data; a ship domain large model module deploying one or more large-scale pre-trained language models or multi-modal large models, which are pre-trained and / or fine-tuned on massive ship domain corpus, and have semantic understanding and knowledge reasoning capabilities in the ship domain; a user behavior understanding and demand prediction module connected to the ship domain large model module, for receiving user input, historical interaction behavior, current ship state and external environment data, and using the ship domain large model to analyze the data in depth to understand user intent, build user and ship profiles, and predict potential user demand; a personalized recommendation generation module connected to the user behavior understanding and demand prediction module, which generates personalized recommendation results based on the output of the user behavior understanding and demand prediction module, combines feature information of recommended items or services, and uses the generation or ranking capabilities of the ship domain large model; a user feedback and self-learning optimization module connected to the personalized recommendation generation module, for collecting user feedback data on the recommendation results, and inputting the feedback data into the ship domain large model to dynamically adjust the model parameters and recommendation strategies through reinforcement learning or continuous learning mechanism.

2. The ship user behavior self-learning recommendation system based on large models according to claim 1, wherein: the data collected by the multi-source data collection and fusion module includes ship static data, ship dynamic data, historical operation data, transaction data, user interaction data, maritime regulations and news, and weather and sea condition data.

3. The ship user behavior self-learning recommendation system based on large models according to claim 2, wherein: the large model in the ship domain large model module has semantic understanding capability, knowledge reasoning capability, behavior pattern recognition capability, and multi-modal understanding capability.

4. The ship user behavior self-learning recommendation system based on large models according to claim 3, wherein: the potential demand predicted by the user behavior understanding and demand prediction module includes ship maintenance demand, required spare parts, crew skill training demand, port supply demand, or suitable insurance products.

5. The ship user behavior self-learning recommendation system based on large models according to claim 3, wherein: the recommendation results generated by the personalized recommendation generation module include ship services, ship products, optimized routes, crew training courses, maritime regulations interpretation or industry information.

6. The ship user behavior self-learning recommendation system based on large models according to claim 1, wherein: the feedback data collected by the user feedback and self-learning optimization module includes clicks, browsing time, purchases, evaluations, collections, shares or rejections.

7. The ship user behavior self-learning recommendation system based on large models according to claim 1, wherein: The user feedback and self-learning optimization module continuously learns from human feedback through reinforcement learning to optimize the model.

8. The large model-based ship user behavior self-learning recommendation system according to claim 1, characterized in that, Further comprising: A recommendation result presentation and interaction module connected with the personalized recommendation generation module, for presenting the recommendation results to the user and supporting the user to perform search, screening, feedback interaction operations.

9. The large model-based ship user behavior self-learning recommendation system according to claim 8, characterized in that: The recommendation method of the large model-based ship user behavior self-learning recommendation system comprises the following steps: S1, multi-source data acquisition and preprocessing: collecting multi-source heterogeneous ship-related data, and performing cleaning and standardization processing on the data; S2, ship domain large model construction and fine-tuning: selecting a general large-scale pre-trained language model as a base model, and pre-training and / or fine-tuning the base model using massive ship domain corpus to construct a ship domain large model; S3, deep understanding of ship user behavior and demand: receiving user query information, historical interaction behavior, ship current state and external environment data, and utilizing the ship domain large model to perform deep analysis on the information and data to understand user intent, construct user portrait and ship portrait, and predict user's potential demand; S4, personalized recommendation result generation: retrieving relevant content from a recommendation candidate set according to the understood user demand and prediction result, and utilizing the ship domain large model to sort, screen and combine the retrieved candidate content to generate personalized recommendation results; S5, user feedback collection and self-learning optimization: collecting user interaction feedback on the recommendation results, and taking the feedback data as reward signals or training samples to continuously self-learn and optimize the ship domain large model and recommendation strategy through reinforcement learning or continuous learning mechanism.

10. The large model-based ship user behavior self-learning recommendation system according to claim 9, characterized in that: In step S3, the ship domain large model understands user intent and predicts potential demand by performing semantic analysis, intent recognition and knowledge reasoning on the information and data; In step S4, the ship domain large model can also generate explanatory text for the recommendation results; In step S5, the self-learning optimization includes adding new user interaction data to the training set and periodically or incrementally fine-tuning the ship domain large model; In step S6, the self-learning optimization further includes utilizing reinforcement learning combined with user feedback to optimize the recommendation strategy of the large model.

Images