A ship maritime collision avoidance rule real scene training and evaluation method and system

Abstract

This invention belongs to the field of intelligent perception and simulation training technology for maritime navigation. It discloses a method and system for realistic training and evaluation of ship collision avoidance rules at sea. By constructing a hierarchical question bank, it generates scenario assembly lists and uniformly schedules ship motion simulation, 3D visual scenes, radar / AIS fusion display, and voice communication interaction. During training, it collects control inputs, motion states, target information, and communication records to form a traceable chain of evidence, and aligns sampling based on evaluation anchor points. An automatic scoring model is established based on result and process indicators, outputting an evaluation report with sub-scores, evidence location, a list of violations, and retraining suggestions. It is suitable for teaching, retraining, and assessment, improving the efficiency and achievement rate of collision avoidance decision-making training.

Description

Technical Field

[0001] This invention belongs to the field of intelligent perception and simulation training technology for maritime navigation, specifically relating to a method and system for real-world training and evaluation of ship collision avoidance rules at sea. Background Technology

[0002] During navigation, ships encounter complex situations and variable environmental conditions, making navigational safety highly dependent on crew members' understanding and execution of the International Regulations for Preventing Collisions at Sea (hereinafter referred to as the "Regulations"). Trainees must integrate multi-source information, including radar, AIS, visual lookout, and VHF communication, to perform situational awareness, risk identification, avoidance decisions, and rerouting maneuvers under different waterway traffic organization methods, visibility conditions, and sea states. Due to the randomness and non-repeatability of encounter situations, current maritime training generally suffers from high costs, high risks, difficulty in covering typical and boundary scenarios, and difficulty in achieving controllable progression of difficulty, thus hindering the systematic, large-scale, and standardized development of training.

[0003] Current collision avoidance teaching and training methods mainly rely on classroom lectures, case reviews, or partial simulations, which typically have the following shortcomings: 1) The lack of a consistent organization and constraints of "rules and clauses - triggering conditions - encounter situation - ship handling decision - result verification" makes training objectives difficult to trace and explain; 2) The training scenarios do not closely match the actual navigation process, making it difficult to stably reproduce verifiable encounter situations under conditions such as day and night, dawn and dusk, mutual visibility and poor visibility, which affects knowledge transfer; 3) The evaluation process relies heavily on the subjective judgment of teachers or manual review afterward. The evaluation criteria are inconsistent and the evidence chain is incomplete. It is difficult to quantify and locate "where the mistake was made, why it was made, and how to correct it". It is also difficult to distinguish the difference between "accidentally getting it right" and "the process is stable and controllable". 4) The lack of a unified scene generation and assembly scheduling mechanism makes it difficult to stably map abstract elements such as encounter types, environmental parameters and evaluation thresholds in training tasks into executable inputs for ship motion, 3D visual scene, voice interaction and evaluation algorithms, which can easily lead to a disconnect between "being able to present but difficult to generate according to the question and difficult to automatically evaluate according to the indicators".

[0004] Therefore, there is an urgent need for a real-world online training and evaluation method and system for the practical application of the "Rules". Summary of the Invention

[0005] To address the challenges of reproducing encounter situations, documenting training processes, and subjective scoring, this invention provides a method and system for real-world training and evaluation of maritime collision avoidance rules. It generates reproducible encounter situations through a hierarchical question bank and difficulty combination mechanism, synergistically couples a ship motion mathematical model, 3D visual visualization, and VHF voice interaction, and employs an automatic evaluation method combining outcome and process indicators to output quantifiable, traceable, and interpretable scoring and diagnostic results, supporting a closed-loop improvement process of "training-evaluation-diagnosis-retraining."

[0006] To achieve the above objectives, the present invention provides the following solution: A method for real-world training and assessment of ship collision avoidance rules at sea, the method comprising: Build a training question bank and configure scenario element parameters and evaluation index parameters for each question bank item; Based on the training question bank, a scenario assembly list is generated. The scenario assembly list is used to describe the ship set, initial relative situation, environmental and traffic organization constraints, and interaction and evaluation parameters required for real-world training of encounter situations. Based on the scenario assembly list, the ship motion simulation, 3D visual rendering, and radar / AIS fusion display are uniformly scheduled, and voice communication interaction is performed to form a training process; During the training process, the trainees' operational inputs, ship motion status, radar / AIS target information, and voice communication records are collected to form a traceable training evidence data chain. Based on a preset evaluation index system and combined with evaluation anchor points, the training evidence data chain is sampled and judged to achieve automatic scoring of training results and training process, and output evaluation report.

[0007] The present invention also provides a real-world training and evaluation system for ship collision avoidance rules at sea. The system is used to implement the aforementioned method and includes: a training question bank module, a ship motion simulation module, a three-dimensional visual simulation module, a human-computer voice interaction module, a collision avoidance real-world generation module, a real-world training terminal module, and an automatic evaluation module. The training question bank module is used to construct a training question bank and configure scene element parameters and evaluation index parameters. The collision avoidance real-scene generation module is used to map the abstract elements of the question bank to the scene parameters to be executed and to uniformly schedule the various subsystems; The ship motion simulation module is used to calculate the maneuvering response of the ship and the target ship and to provide state quantities for risk quantities and evaluation indicators. The three-dimensional visual simulation module is used to realize the realistic presentation of ships, waters and environment and provide interactive evidence objects; The human-computer voice interaction module is used to reproduce the VHF call process and form a traceable chain of voice evidence. The real-scene training terminal module is used to complete the presentation of multi-source information, manipulation input, and data collection during the training process. The automatic evaluation module is used to quantitatively score the correctness of training results and the standardization of the process, and output diagnostic and retraining suggestions, thereby forming a closed-loop improvement mechanism of "training-evaluation-diagnosis-retraining".

[0008] Preferably, the training question bank module follows the consistency principle of "rules and clauses - encounter situations - ship handling decisions - result verification," matching the question content with the clause connotations and applicable conditions of the preset rules, and adopting a "layered and graded + difficulty combination" structure to form a multi-level training set; wherein the difficulty levels include at least: time condition grading, visibility condition grading, water area and traffic organization method grading, comprehensive grading of ship quantity and ship type, and training purpose grading; and simultaneously defining result evaluation indicators and operation process evaluation indicators, threshold parameters, scoring weights and deduction criteria for each exercise to form a standardized question template.

[0009] Preferably, the ship motion simulation module adopts a layered modeling strategy: for the ship itself, a horizontal three-degree-of-freedom MMG-separated maneuvering motion model is used, decomposing external forces and torques into hull hydrodynamics, propeller thrust effect, and rudder force effect, and determining key coefficients by combining empirical parameter initial values ​​with maneuverability tests / system identification; for the target ship, a response model is used, preferably a first-order Nomoto model, to describe the steering response, and its parameters are calibrated through empirical estimation and data identification, realizing online simulation of multiple target ships and deployment of automatic heading / track control.

[0010] Preferably, the 3D visual simulation module aims for "reproducible, interactive, and evaluable real-world scenes," constructing a 3D model of the ship's hull and an interactive modular structure. It employs a physically-based rendering workflow to explicitly express training-sensitive elements. The water area and environment utilize GPGPU-based spectral / frequency domain wave simulation and rendering coupling, combining Fresnel reflection / refraction with environmental reflection to achieve sea surface optical consistency. Furthermore, it introduces wake / spray visualization cues coupled with flow field parameters to enhance the perceptibility of maneuvering effects. Simultaneously, it establishes an asset warehousing and version management mechanism, supporting automatic material assembly based on question bank configuration fields and performing LOD grading and instantiated rendering optimization.

[0011] Preferably, the human-machine voice interaction module is geared towards the collision avoidance and collaborative avoidance process, constructs a corpus based on the Standard Maritime Communication Protocol (SMCP), and provides ASR (Automatic Speech Recognition) and TTS (Text-to-Speech) capabilities for domain-specific terminology. On the ASR side, it combines a domain dictionary / language model to output recognition results with timestamps and confidence levels, and introduces a confidence-driven repeated confirmation mechanism and online adaptation. On the TTS side, it uses a standard answer library to drive standardized broadcasting, and optimizes the emphasis and pause rules for key information such as numbers, bearings, and headings. The entire voice interaction process is linked to the scene event timeline in a closed-loop manner of "call-response-confirmation-execution" to form a traceable evidence chain.

[0012] Preferably, the collision avoidance simulation generation module sets up a "scene assembly list" mechanism as a unified mapping layer from abstract elements of the question bank to executable inputs: an assembly list is generated based on the encounter type, clause triggering conditions, environmental level and evaluation threshold in the question bank, clarifying the ship assembly and initial relative situation, traffic organization constraints, sea state and visibility / light parameters, training evidence objects and sound effects and communication resources; and uniformly scheduling the ship motion simulation, 3D rendering, water environment, voice and sound effects subsystems to operate collaboratively, while solidifying the evidence anchor points and sampling points required for evaluation.

[0013] Preferably, the real-scene training terminal module includes a toolbar, visual view, nautical chart, radar / AIS and voice interaction function area, supports practice selection, start / pause / end, save playback and comments; and records the operation input, motion status, multi-source sensor information and voice log data during the training process to form a chain of evidence for playback.

[0014] Preferably, the automatic evaluation module takes result evaluation indicators and operation process evaluation indicators as inputs, and performs interpretable scoring around the two main lines of "accuracy of operation results" and "operation proficiency". During the simulation operation, a unified evaluation anchor point is set, and the indicator vector is calculated based on the state variables near the anchor point and the trainee input. The discrete indicators are judged by equivalent solution set according to the standard answer vector, and the continuous indicators are judged by threshold / segmentation / saturation soft scoring function. Finally, the total score, sub-item scores, key deduction points and their evidence, DCPA / TCPA evolution curve annotation, and targeted retraining suggestions are output.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Through the question bank mechanism of “layered and graded + difficulty combination” and the consistency constraint of clauses, a traceable and explainable mapping is established between the training tasks and the triggering conditions of the clauses of the “Rules”, the encounter situation and the ship handling decision, so as to realize the large-scale expansion of training content and the controllable progression of difficulty, and can cover typical and complex multi-objective scenarios.

[0016] (2) By using the “Scene Assembly List” to uniformly schedule the ship motion model, three-dimensional visual scene, voice and sound effects and other subsystems, a stable mapping from the abstract elements of the question bank to the executable scene parameters is achieved, avoiding system fragmentation and supporting multiple retraining comparisons of the same question and practical training under different disturbance conditions.

[0017] (3) By using the hierarchical modeling strategy of the ship's MMG model and the target ship's response model, the multi-objective concurrent computational load can be reduced while ensuring the expression and interpretability of maneuvering details, making it suitable for online deployment and real-time interaction.

[0018] (4) By enhancing the visual cues of sea surface optical consistency, the verifiable presentation of signal lights and the visualization of wake / flow field, trainees can obtain more practical situational awareness and manipulation feedback under conditions such as day and night, dawn and dusk and poor visibility, thereby improving the training transfer effect.

[0019] (5) By incorporating VHF voice communication into the training loop and performing structured recording and evaluation of the standardization of language, the completeness of information slots and the consistency of confirmation, the integrated training and evaluation of “cooperative avoidance ability” and “manipulation ability” can be achieved.

[0020] (6) By combining outcome indicators with process indicators and using the anchor evidence chain mechanism, we can achieve unified scoring criteria, explainable reasons for deductions, traceable evidence, and automatically generate diagnostic and retraining suggestions to support the closed-loop improvement of “training-assessment-diagnosis-retraining”. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall structure of the online training and evaluation system for real-world collision avoidance rules in this embodiment of the invention; Figure 2 This is a schematic diagram illustrating the hierarchical and combination relationships of training question bank elements in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating an example of training question bank element configuration in an embodiment of the present invention; Figure 4 This is a schematic diagram of a three-dimensional view of the navigation environment in an embodiment of the present invention; Figure 5 This is a schematic diagram of the user interface of the real-scene training terminal in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1 To address the shortcomings of existing collision avoidance rule teaching and training methods, such as the difficulty in reproducing scenarios stably, the inability to cover typical and boundary encounter situations, the uncontrollable progression of training difficulty, the unclear correspondence between training objectives and the triggering conditions of the International Maritime Collision Avoidance Regulations (IMDG), and the reliance on subjective human judgment in assessments, the lack of unified quantitative standards and traceable evidence chains, and the difficulty in outputting "error diagnosis and retraining suggestions," this invention proposes a realistic training and assessment method for ship maritime collision avoidance rules. This method aims to achieve: parameterized generation and assembly scheduling of encounter situations driven by a question bank; automatic collection and solidification of multi-source data during the training process; and interpretable automatic assessment based on a combination of outcome and process indicators. This supports a closed-loop improvement process of "training-assessment-diagnosis-retraining" and meets the requirements of online simulation deployment and concurrent operation. This includes: Construct a training question bank corresponding to the triggering conditions of the provisions of the International Regulations for Preventing Collisions at Sea (COCR), and configure scenario element parameters and evaluation index parameters for each question bank item; Based on the training question bank, a scenario assembly list is generated. The scenario assembly list is used to describe the ship set, initial relative situation, environmental and traffic organization constraints, and interaction and evaluation parameters required for real-world training of encounter situations. Based on the scenario assembly list, the ship motion simulation, 3D visual rendering, and radar / AIS fusion display are uniformly scheduled, and voice communication interaction is performed to form a training process; During the training process, the trainees' operational inputs, ship motion status, radar / AIS target information, and voice communication records are collected to form a traceable training evidence data chain. Based on a preset evaluation index system and combined with evaluation anchor points, the training evidence data chain is sampled and judged to achieve automatic scoring of training results and training process, and output evaluation report.

[0026] In this embodiment, the training question bank organizes questions according to the consistency principle of "rules and clauses - encounter situation - ship handling decision - result verification", and defines result evaluation indicators and operation process evaluation indicators for each question bank item simultaneously, and configures threshold parameters and weight parameters for the indicators.

[0027] In this embodiment, the training question bank adopts a hierarchical and difficulty combination mechanism, including at least: time condition classification, visibility condition classification, water area and traffic organization method classification, number of ships and ship type classification, and training purpose classification, and generates a set of training items of different difficulty by combining each level.

[0028] In this embodiment, the time condition classification includes daytime, dawn / dusk, and nighttime levels, and is used to achieve progressive training from relying primarily on visual information to relying on traffic light patterns and multi-source information for confirmation.

[0029] In this embodiment, the visibility condition classification includes mutual visibility and poor visibility levels, and under the poor visibility level, the training focus is switched to the fusion of multi-source information such as radar / AIS and the judgment of collision risk trends.

[0030] In this embodiment, the classification of water area and traffic organization methods includes open water areas, narrow waterways or separate navigation systems, and superimposed levels of narrow waterways and separate navigation systems, which are used to form a progressive training of spatial constraints and compliance boundaries from wide to narrow.

[0031] In this embodiment, the scenario assembly list includes at least: ship set and ship type parameters, ship quantity parameters; initial relative situation parameters, including relative bearing, relative distance, heading and speed; traffic organization constraint parameters, including channel, no-navigation zone, anchorage or lane separation constraints; water environment parameters, including sea state, wind field and current field parameters; visibility and illumination parameters; training evidence objects and interaction logic parameters; sound effects and voice communication link parameters; and evaluation anchor point set and sampling rule parameters.

[0032] In this embodiment, the ship motion simulation includes: performing horizontal three-degree-of-freedom motion calculations on the training ship using a separate maneuvering motion model, and implementing automatic heading control or track control on the target ship using a responsive maneuvering model.

[0033] In this embodiment, the separate maneuvering motion model is the MMG separate maneuvering model, the responsive maneuvering model is the Nomoto model, and the ship's position, heading, speed and turning rate are updated in real time based on the output of the model.

[0034] In this embodiment, the radar / AIS fusion display includes: simulating radar echoes and target tracking information, acquiring AIS static and dynamic information, and displaying radar targets and AIS targets in association and highlighting risk targets.

[0035] In this embodiment, the training evidence data chain includes at least: a sequence of manipulated input events, a sequence of key kinematic states, a sequence of target relative motion and CPA / TCPA, radar / AIS operation logs, text transcription of voice interaction and its confidence level and timestamps, and key event bookmarks.

[0036] In this embodiment, the evaluation anchor points include at least: the first target detection, the first encounter type determination, the first evasive action, the time of the minimum nearest encounter distance, the triggering or cancellation of the danger alarm, the start of the return flight, and the completion of the return flight. The training evidence data chain is then time-aligned based on the anchor points.

[0037] In this embodiment, the evaluation index system includes at least: The results evaluation indicators include the correctness of the encounter situation judgment, the assessment of collision risk, the achievement of the closest encounter distance or safety margin, and the rationality of resuming navigation; The evaluation indicators for the operational process include lookout effectiveness, sufficiency of multi-source information utilization, timing of avoidance actions, avoidance range and clarity of actions, standardization of voice command issuance, pre-resumption verification behavior, and process coherence or task completion time.

[0038] In this embodiment, the automatic scoring includes: calculating the accuracy score, the step proficiency score, and the time proficiency score respectively, and weighting them according to preset weights to obtain the final score; wherein when the accuracy does not reach the preset safety threshold, the weighting of the time score is reduced or canceled.

[0039] Example 2 (I) System Overall Architecture and Closed-Loop Process like Figure 1 As shown, the ship collision avoidance rules simulation training and evaluation system of the present invention is designed for online simulation training scenarios and includes at least a training question bank module, a collision avoidance simulation generation module, a ship motion simulation module, a three-dimensional visual simulation module, a human-computer voice interaction module, a simulation training terminal module, and an automatic evaluation module.

[0040] The training question bank module is used to organize training content and solidify evaluation indicators based on the provisions of the International Regulations for Preventing Collisions at Sea (hereinafter referred to as the "Regulations"). The collision avoidance scenario generation module is used to map the abstract elements of the question bank into executable scenario parameters and uniformly schedule the various subsystems. The ship motion simulation module is used to calculate the maneuvering response of the ship and the target ship and provide state quantities for risk quantities and evaluation indicators. The 3D visual simulation module is used to realize the realistic presentation of ships, waters and environment and provide interactive evidence objects. The human-computer voice interaction module is used to reproduce the VHF call process and form a traceable voice evidence chain. The real-scene training terminal module is used to complete the presentation of multi-source information, maneuver input and training process data collection. The automatic evaluation module is used to quantify the correctness of training results and process standardization and output diagnosis and retraining suggestions, thereby forming a closed-loop improvement mechanism of "training-evaluation-diagnosis-retraining".

[0041] (II) Implementation of the training question bank module To support the application of the "Online Simulation Platform for Multi-Task Ships" in the reconstruction of typical accidents, simulation correction, and realistic training and evaluation of the International Regulations for Preventing Collisions at Sea (hereinafter referred to as the "Regulations"), this invention adheres to the consistency principle of "rule clauses—encounter situations—ship handling decisions—result verification" during the question bank construction phase: the content of the questions strictly corresponds to the clause connotations and applicable conditions of the Regulations, ensuring that the training objectives are traceable and explainable; the questions are presented in a manner close to real-world maritime collision avoidance procedures, capable of forming reproducible encounter situations under different waters, different traffic organization methods, and different visibility conditions, thereby ensuring that the training not only covers the knowledge points of the Regulations but also tests the situational awareness and ship handling execution capabilities of the trainees. Based on this, this invention adopts a "layered and graded + difficulty combination" method to construct a multi-level training set, and simultaneously defines result indicators and process indicators for each exercise, forming a standardized question template that is quantifiable and automatically assessable.

[0042] 1. Standardized question templates and fixed fields To achieve scalable expansion of the question bank and uniformity of assessment criteria, this invention stores each real-life exercise in the database using structured fields, including at least: (1) Scene configuration fields: time conditions (daytime / dawn / dusk / nighttime), visibility conditions (mutual visibility / poor visibility), water area and traffic organization methods (open / narrow waterway / separate traffic system TSS and its combination), sea state / wind current level, number and type of ships, training purpose (identification / action / recovery of course), etc. (2) Initial situation field: initial position, course, speed, relative bearing / distance of each ship, as well as constraints such as waterway, no-navigation zone, and anchorage; (3) Clause mapping fields: corresponding to the clause number of the Rules, applicable boundary conditions, the logic for determining responsibility for giving way / direct flight, and the set of allowed equivalent avoidance solutions; (4) Threshold parameter fields: DCPA / TCPA threshold range, safe distance threshold, alarm trigger / cancellation threshold, recovery heading gating conditions, etc.; (5) Scoring and diagnostic fields: index weights, deduction criteria, evidence chain sampling points and mapping rules of “deduction reason - evidence - retraining suggestion”.

[0043] By solidifying the above fields, the question bank achieves "questions can be generated, results can be judged, processes can be traced, and deductions can be explained".

[0044] 2. The mechanism of hierarchical and difficulty combination in the "Rules" The difficulty design of the question bank is based on the cognitive and skill chains of the International Regulations for Preventing Collisions at Sea (hereinafter referred to as the "Regulations") in actual collision avoidance. It breaks down the Regulations learning and ship handling decision-making process into several controllable levels, allowing the difficulty of the questions to be progressively increased and flexibly assembled through "level selection + grade selection + element combination." In this invention, such as... Figure 2 As shown, the preferred hierarchical and graded mechanism includes five levels: time conditions, visibility conditions, water area and traffic organization methods, number and type of vessels, and training objectives. Each level is solidified in the question bank template in the form of enumeration level, parameter range and triggering conditions. Consistency constraints ensure that the selected elements match the applicable logic of the clauses, thereby avoiding the generation of questions that are "reasonable in combination of elements but distorted in triggering of clauses".

[0045] (1) Time condition classification The time condition is used to characterize the differences in target recognition load and information reliability of trainees under different lighting conditions, and to provide a stable "perceptual difficulty benchmark" for subsequent training in clause application judgment and ship handling decision-making. This invention divides the time condition into three levels: daytime, dawn / dusk, and nighttime, with the difficulty coefficient increasing sequentially.

[0046] Under daytime conditions, the outline of the target vessel, changes in relative bearing, and the surrounding traffic situation are more easily obtained visually, making it suitable for conducting rule-based introductory training and basic encounter situation recognition training, and for establishing a basic mapping relationship of "clause-situation-action". At this level, the system prioritizes the timing and clarity of standard avoidance maneuvers, enabling trainees to develop standardized operating habits in low-perceptual-load scenarios.

[0047] Under twilight conditions, the background brightness gradient is large and the target outline is significantly affected by the direction of light and the reflection from the sea surface. Trainees need to have stronger observation and information screening abilities, be able to continuously track key targets and maintain stable judgment of boundary situations. In this level, the question bank preferably introduces more complex backgrounds and interference conditions that are closer to reality, so that trainees can still complete the confirmation of encounter relationships and the expression of avoidance intentions even in the case of "unstable outline".

[0048] In dark conditions, identification relies heavily on signal lights and requires confirmation from multiple sources such as radar / AIS. This places higher demands on the system's accuracy, reaction timeliness, and clarity of action. The optimized question bank strengthens the identification of signal light combinations, ship type / state recognition, and the closed-loop process of "confirmation-action-verification," gradually transitioning training from a visually-driven approach with "sufficient visual information" to a comprehensive judgment mode that "relies on signal lights and confirmation from multiple sources."

[0049] (2) Visibility condition classification Visibility conditions are used to characterize the application boundaries of clauses and the shift in information source structure from "visual judgment-driven" to "risk control-driven". This invention divides visibility into two levels: mutual visibility and poor visibility, with increasing difficulty levels.

[0050] Under conditions of mutual visibility, the training focuses on visual lookout and encounter relationship identification, emphasizing the trainees' ability to delineate responsibilities for giving way / direct navigation and the compliance and clarity of avoidance actions. The question bank selects typical scenarios such as intersection, encounter, and overtaking to enable trainees to take timely actions that are "obvious, decisive, and easily understood by other vessels" while maintaining a good lookout, and to verify the avoidance effect through outcome indicators.

[0051] Under poor visibility conditions, visual information is weakened or even unreliable. The training focus shifts to the fusion of multi-source information such as radar / AIS, collision hazard identification and trend judgment, and conservative action strategies. The question bank is optimized to introduce urgency assessment and verification requirements based on DCPA / TCPA, emphasizing the trainees' ability to manage risk margin under uncertain conditions. For example, whether they can complete target acquisition / tracking within the risk window, whether they can make consistent judgments on relative motion trends, and whether they can complete necessary reviews before and after the action and avoid excessive risk-taking.

[0052] (3) Classification of water area and traffic organization methods The waterway and traffic organization methods are used to depict the progression of ship handling space and compliance boundaries from wide to narrow, and from single constraints to multiple constraints. This invention preferably sets up three levels: open water, narrow waterway or traffic separation scheme (TSS), and narrow waterway + traffic separation scheme (TSS).

[0053] In the open water level, where there is relatively ample space for avoidance, the question bank is selected to train the timing and range control of standard avoidance maneuvers, enabling trainees to form a standard of "early judgment, early action, and clear action" with a greater decision-making margin; in this level, more emphasis is placed on the correctness and interpretability of the application of clauses and the expression of actions.

[0054] In narrow waterways or TSS class, space constraints or traffic separation constraints introduce additional compliance boundaries. The question bank emphasizes compliance avoidance and risk control under constraints: trainees not only need to consider the encounter relationship with the target vessel, but also need to take into account the channel boundaries, recommended course and traffic order, and impose stricter requirements on the timing of actions and turning range.

[0055] In narrow waterways and TSS (Traffic Safety Situation) levels, the dual constraints of limited space and traffic organization result in less decision-making margin and a shorter time window. The question bank selection tests the trainees' ability to handle multiple constraints simultaneously and manage safety margins. For example, under the premise of maintaining navigation order and channel constraints, how to choose an avoidance plan that is more easily understood by other vessels and can reliably reduce risks, and how to reasonably restore the course after the danger has passed to avoid secondary approach or unnecessary deviation.

[0056] (4) Comprehensive classification of the number and type of ships The number and type of ships are used to characterize the progression of encounter situation complexity from "single object, typical ship type" to "multi-objective situation, superimposed special ship type / state". This invention preferably sets three levels: single-objective typical ship type, dual-objective mixed ship type, and multi-objective superimposed special ship type or state.

[0057] In the typical ship type level for single objectives, the situation is relatively clear. The question bank is preferred for establishing a basic mapping of clauses, situations, and actions, and emphasizes the trainees' correct understanding of the encounter relationship and the division of responsibilities. This level can serve as a typical example and introductory training for each clause.

[0058] In the dual-objective hybrid ship type class, the question bank introduces the issues of prioritizing primary and secondary objectives and action compatibility: trainees need to complete information screening, prioritization, and verification to ensure that actions on primary risk objectives do not introduce new risks to secondary objectives; this class emphasizes "robust decision-making under multiple objectives".

[0059] In situations involving multiple objectives and superimposed special ship types or status levels, the traffic situation is complex and may involve multiple parallel constraints or conflicting competition. The question bank is designed to test trainees' comprehensive judgment, stable execution, and risk reassessment capabilities. Trainees need to maintain a consistent risk control strategy in a continuously changing situation and avoid risk transfer or amplification due to a single action through necessary review and reassessment.

[0060] (5) Training objectives are graded The training aims to build a progressive capability from "correct recognition" to "correct action and closed-loop regression." This invention preferably sets three levels: signal light and pattern recognition; signal light and pattern recognition + collision avoidance action; signal light and pattern recognition + collision avoidance action + course recovery.

[0061] In the "Light and Signal Recognition" level, the question bank focuses on training participants to correctly identify and confirm the types and signals of lights and the types / states of ships, emphasizing "accurate identification and correct confirmation." In the "Identification + Collision Avoidance Actions" level, the question bank further assesses trainees' ability to take compliant, clear, and easily understood avoidance actions based on identification. It focuses on evaluating whether the direction, timing, and extent of the action are reasonable, emphasizing "giving way correctly and clearly." In the "Identification + Action + Resume Course" level, the question bank further requires trainees to resume the original course (or recommended course) after confirming that the danger has been eliminated and completing the necessary checks, so as to avoid turning back too early and causing a second approach or turning back too late and causing unnecessary deviation, emphasizing "returning safely after making a clean turn".

[0062] Based on the aforementioned five-layer structure, the question bank of this invention does not generate fixed difficulty packages all at once. Instead, it combines different levels from different layers to form practice sets with varying difficulty levels, and avoids distorted questions due to "mismatch between element and clause triggering logic" through consistency constraints. For example, high perception difficulty (night / poor visibility) can be combined with medium traffic complexity (dual objectives) and closed-loop training objectives (identification + action + course recovery) to construct practical training; similarly, medium environmental constraints (narrow waterways or TSS) can be combined with multi-objective traffic situations to construct comprehensive situational training. Through this hierarchical and difficulty combination mechanism, the question bank can achieve large-scale expansion and personalized scheduling while ensuring authenticity and compliance, and provides a unified and reusable question generation and management framework for subsequent progressive difficulty training schemes and automatic evaluation models.

[0063] 3. Practice evaluation index system and quantitative methods The training question bank not only needs to "generate scenarios," but more importantly, it needs to "objectively score, explain the reasons for deductions, and be used for error correction and retraining." Therefore, this invention stipulates that each real-world exercise must be simultaneously bound to evaluation indicator points upon being added to the database. These indicator points are further broken down into two categories: result evaluation indicators and operational process evaluation indicators. Result evaluation indicators address whether the training output is correct, safe, and compliant, answering "whether the final result was correct and safe." Operational process evaluation indicators address whether the training process follows standardized procedures and whether it possesses good situational awareness and decision-making execution capabilities, answering "whether the process was performed correctly and whether it possesses operational habits transferable to real navigation." These two types of indicators together constitute the input to the automatic evaluation model, enabling the system to provide not only a total score but also diagnostic conclusions such as "where the error occurred, why it occurred, and how to correct it," and to generate targeted retraining content accordingly.

[0064] To ensure the repeatability and comparability of scores under different questions and disturbance conditions (wind flow, visibility, traffic density, etc.), this invention optimizes and solidifies the "evaluation anchor point" mechanism in the question bank template. During simulation, key events are uniformly set with locatable time points or time windows, including but not limited to the first target detection, the first determination of the encounter relationship, the first reaching of the risk threshold, the first execution of an effective avoidance maneuver, the minimum encounter distance, the triggering / clearing of a danger alarm, the start of a return to flight, the completion of a return to flight, and the completion of communication confirmation. Each evaluation index is calculated jointly by the state variables near the anchor point (heading, speed, turning rate, relative bearing / distance, etc.) and the trainee's input (rudder angle, engine telegraph, radar / AIS operation, voice commands, etc.), thereby avoiding score drift caused by fluctuations in sampling frequency, playback truncation, or rendering frame rate, and providing consistent evidence for post-replay review.

[0065] Regarding the scoring structure, this invention preferably adopts a total score model that weights and sums the results, processes, and time, as shown in the following expression: (1) in: This is the total score for the exercise; The results are scored to measure whether the collision avoidance performance and compliance standards are met. The process score is used to measure whether the "observation-judgment-decision-execution-review" chain is standardized, stable, and explainable. The time score is used to measure the efficiency and pace of completing training tasks while meeting safety and compliance requirements. , , These are three weighted coefficients. The question bank template can be configured according to the training objective, such as improving performance in the "assessment / evaluation" mode. To enhance process standardization and improve skills under the "introductory teaching" model. This is achieved by highlighting the key points of the rules.

[0066] 3.1 Practice Configuration and Indicator Point Generation by Figure 3 As shown, the configuration elements for this exercise may include: Time = Daytime, Visibility = Mutual Visibility, Water Area = Open, Number of Vessels = Single Target, Vessel Type = Motorized Vessel, Training Objective = Light and Formation Recognition + Collision Avoidance Actions + Resumption of Course. Based on this configuration, the question bank automatically generates a set of indicator points highly matched to the question during the input phase, and fixes the calculation caliber, threshold parameters, weights, and deduction reasons for each indicator point, ensuring that a structured evaluation report of "Total Score - Sub-items - Evidence - Clause Explanation - Retraining Recommendations" can be output after training.

[0067] 3.2 Result Evaluation Indicators The outcome evaluation indicators are used to assess whether trainees have ultimately achieved the training objectives of "no collision risk, compliance with rules, controllable risk, and reasonable course recovery," emphasizing the accuracy of judgments and safety margins. This invention preferably divides the outcome evaluation indicators into at least the following items, and provides a quantification method and evidence output rules for each item in the question bank template.

[0068] (1) Determining the type / state of a ship by its light and signal This indicator assesses the trainee's ability to correctly identify the target vessel's displayed lights / types and their corresponding vessel type and operational status. It is a core fundamental skill under nighttime and twilight conditions, and a prerequisite for determining the applicability of subsequent clauses. This invention preferably uses a tiered scoring method to quantify this item: full marks are awarded for completely correct identification; partial marks are awarded for only partially correct elements (e.g., identifying a motorized vessel but misjudging its operational status, or identifying the main lights but omitting auxiliary lights); and low or zero marks are awarded when key light combinations or type identifications are incorrect, leading to misjudgments of type or status. To support "explainable deductions," the system simultaneously records the error category and key evidence, such as "type misjudgment," "status misjudgment," "light combination misjudgment," and "missing confirmation process," and outputs corresponding timestamps, visual screenshots, or light display status snapshots in the evaluation report, facilitating review and retraining for trainees.

[0069] (2) Encounter situation assessment and division of responsibility for giving way / direct flight This indicator is used to assess whether trainees correctly judge encounter relationships (crossing, meeting, overtaking, etc.) and yielding / direct flight attributes, directly corresponding to the applicable premise and action obligations of the "Rules". This invention preferably uses the situation labels in the system reference solution as the standard answer and compares the trainees' judgment results with the standard labels. For boundary situations (e.g., relative azimuth angles approaching the threshold, situations where the situation may evolve from crossing to overtaking in a short time), the question bank template can set a tolerance range or an "equivalent judgment set" to avoid scoring instability due to small initial differences. Evidence output for this item may include: a snapshot of the relative azimuth / distance at the time of judgment, the target's relative motion vector, and an explanation text of the responsibility division determined by the system, enabling the verification of "why it belongs to a certain situation and why yielding / direct flight should be performed".

[0070] (3) Collision Hazard Identification and Urgency Assessment (DCPA / TCPA) This indicator is used to assess whether trainees can correctly identify the existence and urgency of collision risks and take timely and compliant avoidance measures accordingly. The system calculates the DCPA (Distance to Nearest Encounter) and TCPA (Time to Nearest Encounter) in real time based on the relative motion of the two vessels in the background, and classifies the risk into "dangerous / controllable / safe" levels using thresholds or piecewise functions. To ensure consistent calculation methods, this invention preferably adopts the following definitions: (2) in: The position vector of the target ship relative to the ship itself; It is a relative velocity vector; This represents the time from time t to the nearest meeting point; This represents the shortest distance to the nearest point where the two points will meet.

[0071] The question bank template includes fixed safety distance and emergency time window thresholds: when the DCPA (Discretionary Distance Approach) is below the safety threshold and the TCPA (Temporary Risk Assessment) falls within the emergency window, it is considered "dangerous." If the trainee fails to take effective avoidance action within the danger window or their risk assessment deviates significantly from the system benchmark, points will be deducted accordingly. The evaluation report preferably outputs DCPA and TCPA curves over time, along with key anchor point markers (avoidance start point, minimum DCPA point, and rerouting point), to visually present the risk evolution and action effectiveness.

[0072] (4) Final CPA results and safety margin This indicator is used to assess whether the minimum safe distance between the two vessels after a collision avoidance maneuver meets training requirements and whether there are risky behaviors such as "grazing the edge" or "excessive approach". This invention preferably uses the final minimum encounter distance (i.e., (or its discrete sampling approximation) is compared with the safe distance threshold in the question bank template to form a continuous score: the greater the distance, the higher the score, but an upper limit is set to avoid excessive deviation from the route to obtain an unreasonably high score; at the same time, a minimum safety red line is set, and when the minimum distance is lower than the red line, a strong deduction or judgment failure is triggered to ensure that the safety bottom line is not offset by factors such as time score.

[0073] (5) Timing of course restoration and closed-loop confirmation This indicator is used to assess whether trainees resume their original (or recommended) course after confirming the danger has been eliminated and completing necessary verifications, avoiding "premature turning leading to a secondary approach" or "delayed turning causing unnecessary deviations." This invention preferably links the timing of course restoration with danger elimination criteria, relative bearing change rate, and closest distance trends: if course restoration occurs before the risk has been eliminated or necessary verifications have been completed, points are deducted; if course restoration is too late, resulting in significant deviations or disrupting traffic flow, points are also deducted. Evidence output for this item may include: DCPA / TCPA status at the time of course restoration, verification action records (radar verification / AIS verification / visual confirmation), and risk trends after restoration, used to explain "why the timing was reasonable / unreasonable."

[0074] 3.3 Evaluation Indicators for Operation Process The operational process evaluation indicators are used to assess whether trainees follow the "observation-judgment-decision-execution-verification" chain during training, emphasizing the compliance, timeliness, and explainability of the behavioral process. Unlike judging "right / wrong" solely based on the final result, this invention uses process indicators to constrain trainees to form standard operating habits that can be transferred to real-world navigation: even if no collision occurs, if there are non-standard behaviors such as insufficient observation, unclear actions, repeated probing, or lack of necessary verification during the process, the system can still deduct points and provide corrective suggestions, thereby achieving a closed-loop improvement of "training-assessment-diagnosis-retraining".

[0075] To ensure the objectivity and traceability of the process evaluation, this invention preferably links process indicators to the training event timeline: during simulation, the system continuously records the trainee's operational inputs (rudder angle, engine telegraph / target speed commands, mode switching), sensor and tool usage behaviors (telescope, radar range / gain / clutter adjustment, target acquisition and plotting, AIS query, VHF call and confirmation, etc.), and key state quantities (heading, speed, rate of turn, relative bearing / distance, DCPA / TCPA trends, etc.). Each process indicator is fixed in the question bank template in the manner of "triggering condition - evidence sampling - discrimination rule - deduction reason text" so as to automatically generate interpretable process diagnostics after training.

[0076] (1) Lookout effectiveness and multi-source information verification This indicator assesses whether trainees have conducted continuous and effective lookout and whether they have appropriately used tools such as binoculars, radar, AIS, and VHF at key points to complete situational awareness and cross-verification. Specifically, at key points such as "first target detection," "risk escalation," "pre-action verification," and "pre-resumption verification," the system checks whether trainees have completed the necessary observation and verification actions. For example, whether they have stably tracked risky targets and determined their relative motion trends, whether they have verified the target's heading, speed, and CPA / TCPA information using radar or AIS, and whether they have increased their reliance on radar / AIS and reduced their reliance on single-point visual observation under poor visibility conditions. The question bank template can set minimum usage requirements (such as at least one target acquisition and one trend verification) and bonus items (such as completing plotting and maintaining continuous tracking). If behaviors such as "failure to use key tools," "only a single brief look," or "verification occurring after the risk window" occur, the lookout is deemed insufficient and points are deducted.

[0077] (2) Timing of taking evasive action This indicator assesses whether trainees make evasive decisions and execute actions within a time window that still allows for sufficient leeway, avoiding delays that could lead to emergency evasive maneuvers or reactive responses. The system preferentially uses the "first time the risk threshold is reached" as a reference point to calculate the lead or lag of the trainee's first effective action input relative to that moment: if a clear action is taken shortly after the risk threshold is triggered, the timing is considered reasonable; if action is taken only within the emergency window or if there is still no effective action input during a period of escalating risk, the evasive maneuver is deemed too late and points are deducted. For different training difficulties (e.g., nighttime, poor visibility, narrow waterways / TSS, etc.), different reasonable lead intervals can be configured in the question bank template to ensure that the scoring aligns with practical understanding and risk margin requirements.

[0078] (3) Avoidance range and clarity of action This indicator is used to assess whether the trainee's actions are "obvious, decisive, and easily understood by other vessels," avoiding small, repeated swaying, ambiguous actions, or frequent speed / course changes that lead to unclear intentions. The system records the amount of course change, speed change, number of actions, and duration of actions, and uses the "minimum effective action threshold" from the question bank template for judgment: if multiple actions are below the threshold and the overall effect is not significant, it is considered ambiguous; if there are repeated attempts of "avoidance-correction-again avoidance," or frequent changes in rudder angle / engine telegraph within a short period of time causing wavering intentions, points are deducted. To provide interpretable quantification for "repeated attempts," this invention can introduce an action complexity indicator: (3) Where: C is the motion complexity; N is the number of actions identified as valid manipulations within the statistical time window; This represents the effective speed change for the kth time. Let be the effective change vector for the kth time. Upper thresholds for C and N can be set in the question bank template: when the threshold is exceeded, it is judged as "fragmented action / repeated trial and error" and points are deducted; when a single change of direction or speed meets the explicitness threshold and is consistent with the rule logic, points can be added to the action explicitness item.

[0079] (4) Consistency between the performance and the verbal command when giving the avoidance command. When the training system includes voice / command interaction, this metric assesses whether trainees can issue commands clearly, loudly, and decisively at key points, whether their expression conforms to the standard format, and whether the command content matches the actual operating actions. The system can perform keyword matching and format verification on the speech-recognition transcribed text, such as checking whether it contains core slots like "right / left," "xx degrees," or "decelerate to xx knots," and verifying whether the steering / telephone input near the time the command was issued matches the command. If key fields are missing from the command, the expression is ambiguous, the timing of the issuance is significantly delayed, or the "command and action are inconsistent," points are deducted and corresponding evidence (speech transcription, confidence score, timestamp, and operating input fragment) is output.

[0080] (5) Resumption of heading orders and verification procedures This indicator assesses whether trainees have completed necessary checks before resuming course (e.g., reconfirming CPA trends using radar or verifying target motion status via AIS), and whether the course resumption command is standardized and matches the environmental conditions. The question bank template preferably sets "checking actions" as a prerequisite for course resumption: if course resumption is initiated without checking, points should be deducted even if no collision occurs, to reinforce the closed-loop process of "hazard clearance confirmation—review—resumption"; if the check is thorough and the timing of the resumption is appropriate, a positive evaluation is given in the process score. Evidence output may include: check tool call logs, DCPA / TCPA status and trends at the time of course resumption, and whether the risk increased again after resumption.

[0081] (6) Total task completion time and process coherence This indicator is used to assess the efficiency of trainees in completing training tasks while ensuring safety and compliance, as well as the continuity and smoothness of the process. The system records the total time from the start of the exercise to the resolution of the risk and the completion of the course recovery, and can be further broken down into stages such as "discovery-judgment time," "initiation of avoidance time," "avoidance execution time," "review and return-to-course time," and "VHF communication time" for fine-grained diagnosis. To avoid risky operations due to a pursuit of speed alone, this invention preferably sets gating conditions: when key safety indicators (such as minimum DCPA, timing of action, etc.) do not reach the minimum qualified threshold, the time score is not increased or its weight is reduced, ensuring that the scoring orientation always prioritizes safety and compliance.

[0082] In summary, the operational process evaluation indicators and the outcome evaluation indicators complement each other: the outcome indicators ensure that "avoidance effectiveness meets standards," while the process indicators ensure that "operational procedures are standardized and transferable." After the thresholds, weights, and evidence output rules of the above indicators are solidified in the question bank template, the system can automatically generate a process diagnostic checklist after training, clearly pointing out the trainees' weaknesses in each stage of lookout, judgment, decision-making, execution, and review, and providing retraining recommendations linked to the modular content of the question bank (such as special training on signal light patterns, radar / AIS verification, action clarity, or course closure), thereby achieving sustainable capability improvement.

[0083] 3.4 Indicator solidification, scoring rules and evaluation report generation To support the scalable expansion of the question bank and the implementation of automated evaluation, this invention preferably embeds the aforementioned result evaluation indicators and operational process evaluation indicators in a structured manner within the question bank template. This ensures that each exercise possesses the inherent attribute of being "consistently evaluable" while simultaneously "generating a scenario." Specifically, in addition to including exercise scenario configuration, clause mapping, and initial state, the question bank template also embeds at least the following key fields directly related to evaluation: (1) Discrimination threshold parameter and tolerance rule The discrimination threshold parameter is used to provide an executable discrimination boundary for continuous indicators (such as DCPA, TCPA, yaw distance, etc.) and to suppress scoring instability in boundary situations through tolerance rules. Preferably, the question bank solidifies the safety distance threshold, urgent time window threshold, alarm trigger / cancellation threshold, heading recovery gating condition, minimum effective action threshold, etc.; it provides a tolerance range or "equivalent discrimination set" for encounter situation boundary discrimination (such as relative bearing approach threshold); and it provides an allowed equivalent avoidance solution set for cases where the rules allow multiple solutions, avoiding incorrect point deductions due to "multiple solutions to one question".

[0084] (2) Scoring weights and deduction criteria The scoring weights are used to reflect differences in training objectives and to allow for adjustable parameters: for example, increasing the weight of indicators related to identification and clause understanding in introductory training, increasing the weight of risk control and process standardization in practical training, and increasing the weight of process compliance and closed-loop review in assessment and evaluation. The deduction criteria are fixed in the form of "deduction reason—triggering condition—evidence output," for example: — "Late start avoidance": This condition is met when TCPA falls into the time window and the trainee does not provide any valid manipulation input within the specified time window; — "Unclear actions / repeated attempts": The complexity C of the action or the number of actions N exceeds the threshold, and the risk reduction is not significant or there are multiple attempts to correct the situation and then avoid it again; — "No verification before restoration": There is a lack of radar / AIS verification or visual confirmation logs before restoring the course anchor point; — “VHF information slot missing”: The voice command does not contain the necessary elements (object, location / distance, intent, confirmation statement, etc.) or the command and action are inconsistent.

[0085] By linking deduction criteria with evidence output, the evaluation results can be directly reviewed and understood by trainees.

[0086] (3) Anchor points of the chain of evidence and sampling strategies To ensure the traceability of scoring, this invention optimizes and solidifies the definition of evidence chain anchor points and sampling strategies in the question bank template. These include snapshots of key moments (relative situation, heading and speed, radar / AIS target status), curve sampling (DCPA / TCPA evolution over time, changes in heading / turn rate, and changes in yaw distance), and event logs (tool calls, alarm triggering / clearing, voice interaction rounds and confirmation results, etc.). The anchor point mechanism ensures stable evidence location even under different terminal frame rates and sampling frequencies, and supports consistent citation of the same evidence fragment through both automatic and manual review.

[0087] (4) Mapping between assessment report structure and retraining recommendations Based on the aforementioned fixed fields, the system automatically generates an evaluation report after training. The evaluation report preferably includes: total score and sub-scores (result / process / time), a list of key deduction points and corresponding evidence (timestamps, situational screenshots, radar / AIS playback bookmarks, speech transcription and confidence levels), explanations of corresponding clauses (why a clause was triggered, why the action was compliant / non-compliant), and targeted retraining suggestions. These retraining suggestions are linked to the modular content organization of the question bank: for example, when "misjudgment of signal lights / signages" occurs, a special study session on signal lights / signages is pushed; when "insufficient lookout / failure to verify" occurs, a special study session on radar / AIS verification is pushed; when "unclear actions / repeated probing" occurs, a special study session on action clarity is pushed; and when "resuming course too early or without verification" occurs, a special study session on closed-loop course recovery is pushed.

[0088] Through the structured and automated generation mechanism of "indicators-thresholds-weights-evidence-recommendations" mentioned above, the question bank is upgraded from a "set of questions" to a training system that is "configurable, assessable, diagnosable, and retrainable": the same question type can be repeatedly trained under different perturbation conditions while maintaining consistent scoring criteria; different trainees can make horizontal comparisons under unified criteria; and the same trainee can make vertical retraining comparisons based on the chain of evidence, thereby achieving continuous improvement in practical skills.

[0089] 4. Question bank content organization and learning path design After completing the hierarchical and graded design of the question bank and the indicator system, to ensure that trainees can form a systematic knowledge structure at a low cost and achieve a closed-loop improvement of "learning-practice-evaluation-correction" in real-world training, this invention further modularizes the question bank content and designs a path-based learning approach. Overall, the learning organization is based on the knowledge structure of the "Rules" and the practical procedures for ship collision avoidance, constructing a three-tiered learning system of "line-by-line learning—specialized learning—comprehensive application." It also allows trainees to select different content and training intensities based on their job position (captain / driver / watchkeeping personnel / trainee) and training objectives (learning, retraining, assessment, evaluation), thus balancing coverage, relevance, and operability.

[0090] (1) The organizational logic of the three-level learning system The three-level learning system of this invention is used to connect the knowledge of the rules and regulations with ship handling decision-making skills from simple to complex, and to form a corresponding relationship with the question bank difficulty combination mechanism.

[0091] In the "Article-by-Article Learning" level, a complete knowledge framework is established using the articles as an index, focusing on "knowing what the article is, when it applies, and what the core requirements are." The system ideally organizes each rule into "Article Card + Scenario Example + Common Mistakes Hints + Self-Test Questions." The Article Card provides the key points and applicable conditions of the article; the Scenario Example offers typical encounter situations and visual elements matching the article; the Common Mistakes Hints summarize common misjudgment patterns and their risk consequences; and the Self-Test Questions assess the trainee's understanding of the article's triggering conditions and responsibility allocation. The questions at this level are designed with low perceptual load and typical situational configurations to help trainees establish a stable "Article-Situation-Action" mapping.

[0092] In the "Specialized Learning" level, training is intensively conducted based on key capabilities, focusing on addressing the problem of "knowing how to do something, but being unstable or inaccurate in specific capability areas." This invention preferably divides specialized capabilities into several areas, including signal light and pattern recognition, sound and light signals, radar / AIS verification and trend judgment, clarity of avoidance maneuvers, course recovery and verification, and VHF communication standardization. In this level, the question bank preferably uses parameter perturbations and boundary conditions (such as twilight backlight, complex backgrounds, light fog, and multi-target interference) to enable trainees to maintain stable judgment and standardized procedures even under "imperfect information" conditions.

[0093] At the "Comprehensive Application" level, realistic drills are conducted using typical accident chains and complex traffic situations as indexes, focusing on addressing the problem of "being able to perform individual tasks correctly, but easily encountering conflicts or omissions when multiple constraints are in parallel." This invention preferably employs combinations of conditions such as multi-objective traffic situations, narrow waterways / TSS constraints, poor visibility, and wind disturbances to test trainees' ability to handle multiple parallel constraints, their continuous management of safety margins, and their stable execution of the "judgment-action-review-retake" closed loop. This level is deeply integrated with the automatic evaluation system, outputting a comprehensive report after training that emphasizes the evidence chain and retraining path.

[0094] (2) Content composition of the four core modules In accordance with the platform's training objectives, this invention preferably organizes the question bank learning content into the following four core modules, which can be expanded as needed: 1) Learning modules one by one The learning module is indexed by the clauses of the "Rules," organizing each rule into "clause card + scenario example + common mistake tips + self-test questions." The clause card clearly defines the applicable conditions and key points for assigning responsibility; the scenario example provides encounter situation materials corresponding to each clause, and marks key identification clues (relative bearing changes, target lights / types, basis for determining responsible vessel, etc.); the common mistake tips summarize common misjudgments (such as misjudging an intersection as an overtaking maneuver, ignoring special vessel type conditions leading to incorrect responsibility judgments, etc.) and provide correction points; the self-test questions provide immediate feedback and record the types of incorrect answers, providing a basis for subsequent specialized retraining.

[0095] 2) Light and Signal Type Learning Module The navigation light and shape learning module adopts a progressive structure: from single elements to combined elements, from typical ship types to special ship types, and from static recognition to dynamic confirmation. Lower-level training focuses on recognizing single light / shape elements to establish a basic mapping. Intermediate-level training introduces multi-light combinations, distance changes, and relative bearing changes, emphasizing the "observation-confirmation-reconfirmation" process. Higher-level training introduces complex backgrounds, dawn / dusk and nighttime conditions, and multi-target interference, enabling trainees to stably identify target ship types and states under conditions closer to real-world scenarios, thus addressing the challenges of nighttime and dawn / dusk recognition.

[0096] 3) Sound and light signal learning module The sound and light signal learning module divides signal learning into three levels: "signal recognition—signal semantics—scenario application." First, it trains participants to accurately identify signal patterns; second, it trains them to understand the semantics of the signals and the corresponding action requirements; and finally, it trains them to develop correct response and cooperation strategies in scenarios with limited line of sight, restricted communication, or cooperative avoidance. This module can be linked with the voice interaction module, enabling participants to complete the "recognition—understanding—action" loop in a near-realistic conversation and noisy environment.

[0097] 4) Navigation and Collision Avoidance Learning Module The navigation avoidance learning module adopts an organizational logic of "situation type - action strategy - closed-loop recovery," forming a training path that progresses from simple to complex. Low-level training focuses on standard avoidance maneuvers and the clarity of actions in typical situations; intermediate-level training introduces channel / narrow waterway / TSS constraints, emphasizing compliance boundaries and risk control; high-level training introduces multi-target situations and poor visibility, emphasizing multi-source verification, action compatibility, and timing of return navigation. This module primarily addresses the problems of "knowing when to yield but not yielding, yielding unclearly, and not knowing how to return after yielding," and outputs interpretable evidence and retraining recommendations through automatic evaluation.

[0098] (3) Linkage with automatic evaluation and differentiated push notifications To achieve a closed-loop improvement process of "learning-practice-evaluation-correction," this invention preferably links the question bank learning path with the automatic evaluation results: based on the reasons for deductions and the chain of evidence in the evaluation report, the system automatically identifies the trainee's weak skills and selects matching retraining content from the question bank module for differentiated delivery. For example, if "misjudgment of signal lights and patterns" occurs, training on similar element combinations in the signal light and pattern specialty is prioritized; if "insufficient lookout / failure to verify" occurs, radar / AIS verification and trend judgment specialty is delivered; if "unclear actions / repeated probing" occurs, action clarity and timing control training is delivered; if "resuming course too early or without verification" occurs, closed-loop recovery and review training is delivered. Through the above linkage, the question bank is upgraded from a "set of questions" to a "configurable teaching and training system," significantly improving learning efficiency and ability transfer effects while ensuring system coverage.

[0099] (III) Implementation of the Ship Motion Simulation Module like Figure 1 As shown, the ship motion simulation module of the present invention is used to perform real-time simulation of the maneuvering motion of the ship and the target ship in an online simulation training scenario. It provides the three-dimensional visual simulation module with common state variables such as heading, speed, position, and turning rate, and provides the collision avoidance risk calculation and automatic assessment module with the basic data required for key assessment indicators such as DCPA, TCPA, CPA results, and yaw distance, so that the "scenario presentation - maneuvering interaction - risk judgment - result evaluation" are closed and consistent under the same dynamics.

[0100] When selecting a model, this invention comprehensively considers dynamic expression capability, parameter interpretability, parameter acquisition difficulty, and online computational load: for the ship that needs to reflect maneuvering details and support interpretable evaluation, the MMG separation model is preferred; for target ships that are numerous and diverse and require automatic heading / track control to reduce the concurrent computational burden on the server, the responsive (Nomoto type) model is preferred as its kinematic / dynamic approximation.

[0101] 1. Model of the ship's motion The preferred motion model for this ship is established within a three-degree-of-freedom (sway-bow-roll) framework in the horizontal plane. The external forces and moments acting on the ship are decomposed into three parts: hull hydrodynamics, propeller thrust effect, and rudder force effect. This ensures a clear path for propulsion and steering, facilitates the extraction of training and evaluation indicators such as steering rate, lateral drift, and TCPA / DCPA evolution, and makes it easier to integrate "steering / engine telegraph input - force decomposition - motion response" into the control and evaluation logic.

[0102] Preferably, the MMG motion equation can be expressed as: (4) Where: u, v, and r are the longitudinal velocity, lateral velocity, and bow angular velocity of the ship, respectively; m is the mass of the ship; , These are the additional masses; Let z be the moment of inertia about the z-axis; To add mass moment of inertia; The coordinates of the centroid; For the hydrodynamics and torques of the ship's hull; For propeller force and torque; This refers to the rudder force and torque.

[0103] In engineering implementation, the ship motion simulation module performs numerical integration on equation (4) with a fixed simulation step size, updates the ship's state vector (position, heading, speed, turning rate, etc.), and outputs the state to: (1) Three-dimensional visual simulation module, used to drive the ship's attitude, wake / spray effect and instrument display; (2) Radar / AIS simulation and risk calculation module, used to generate target relative motion and DCPA / TCPA evolution; (3) Automatic evaluation module, used to sample at evaluation anchor points and form a chain of evidence.

[0104] To ensure the interpretability and feasibility of MMG parameters, this invention preferably adopts a parameter determination method that combines "mechanism modeling and identification modeling": First, operational initial parameter values ​​are given based on the ship's main dimensions, rudder / propeller geometry, and empirical formulas; then, key coefficients are identified and calibrated using maneuvering data such as circling tests, Z-shaped tests, and course maintenance, and physical constraints such as sign / range are applied to the parameters to ensure stability and interpretability.

[0105] 2. Target ship motion model The variety and number of target ships directly impact the concurrent computing load on the server side. Furthermore, target ships in training scenarios typically require automatic heading / track control to stably generate encounter situations and maintain the reproducibility of the problem set conditions. To reduce online computational complexity and improve scenario configuration efficiency, this invention preferentially employs a responsive (Nomoto-type) model to describe the steering response of target ships. This model summarizes the main dynamic characteristics of "rudder angle input—heading response / steering rate response" with a small number of parameters, thereby enabling real-time simulations that can be deployed in batches.

[0106] (1) First-order Nomoto (K–T) model Preferably, the steering response of the target ship is represented by a first-order Nomoto model, and its transfer function is as follows: (5) Where: s is the Laplace operator; K is the gyration exponent, representing the gain of the target ship's steady-state steering capability; T is the follower exponent, representing the time constant of the target ship's steering response speed. Equation (5) is suitable for rapid deployment of online simulation and automatic control; when the target ship exhibits obvious second-order dynamics or requires higher precision, it can be extended to a second-order Nomoto form and configured according to ship type in the question bank template.

[0107] To facilitate discrete simulation, equation (5) can be written in the time domain as a first-order differential form: (6) Where: r(t) is the target ship's turning rate; δ(t) is the target ship's rudder angle command. As can be seen from equation (6), K determines the magnitude of the steady-state turning rate under the same rudder angle, and T determines how quickly the turning rate reaches the steady state. Therefore, the two parameters can directly affect the speed of the encounter situation evolution and the length of the avoidance window, and have clear training interpretability.

[0108] (2) Target ship's heading and track status update The target ship's heading at simulation time t It can be obtained by integrating the steering ratio: (7) Given the target ship's speed (Given either a fixed question bank or a simplified car telegraph-speed model), its planar position Updates can be calculated based on kinematic relationships: (8) in: Let be the position of the target ship in the scene coordinate system. Through the coupling of equations (6) to (8), the target ship motion simulation module can generate heading and trajectory changes consistent with the rudder angle input under a low computational load, and maintain consistency with the three-dimensional visual presentation, radar / AIS display, and risk calculation.

[0109] (3) Determination and calibration of parameters K and T The key to modeling a response model lies in determining K and T. This invention preferably employs a two-stage method: "empirical / experimental initial values ​​+ data identification and calibration." 1) Initial value determination: The initial values ​​of K and T can be estimated based on the ship's static parameters and maneuvering test results (such as Z-shaped test and circling test), and a parameter range library can be established according to ship type, so that different target ships can be quickly configured even without a large amount of measured data; 2) Calibration identification: When the platform has target ship maneuverability test data or historical high-reliability trajectory data, the least squares or recursive identification of equation (6) can be performed to obtain K and T that are more in line with the dynamic characteristics of the ship type. During the calibration process, physical constraints such as K>0 and T>0 are preferred to be constrained to ensure that the response characteristics are stable and interpretable.

[0110] Using the above methods, the target ship motion model can not only meet the requirements of online real-time simulation and concurrent deployment, but also improve the degree of closeness to the dynamic characteristics of real ship types through parameter calibration when necessary, thereby improving the reproducibility of training scenarios and the consistency of evaluation criteria.

[0111] (iv) Implementation of the 3D Visual Simulation Module The 3D visual simulation module of this invention is used to present the ship, water, and environmental elements configured in the question bank in a realistic manner, and to link them with the state variables output by the ship motion simulation module, thereby achieving an immersive scene display that is "reproducible, interactive, and evaluable" in online simulation training. This module must meet the real-time rendering performance constraints of the web / online deployment, and ensure that the model semantics and component structure serve the training closed loop of "encounter situation generation—manipulation interaction—evidence chain retention," so that the scene presentation is consistent with the applicable conditions of the rules, encounter relationship identification, and the interpretability of avoidance actions.

[0112] 1. Construction of 3D ship hull model and interactive componentization The hull modeling is based on the main dimensions and typical structural layout of the actual ship, and the preferred strategy is "precise geometric constraints on the shape + controllable restoration of detailed features": First, a standardized coordinate system and scale benchmark are established based on the actual ship's external dimensions data, and the main mesh topology planning and surface continuity control of the hull are completed to ensure that the bow and stern lines, bilge transitions and superstructure outlines have stable shape recognition at different viewing distances; then, combined with multi-view detailed images and key structural information, the areas such as the bridge, mast, hatch coaming, railings, and bulwark openings are partitioned and refined, and the high-poly details are baked to a low-poly mesh that can be rendered in real time through normal mapping / height mapping, etc., to achieve a balance between "detail appearance and polygon count budget".

[0113] At the material and texture level, to ensure the reliability of recognition under daytime, dawn and dusk, nighttime, and complex background lighting conditions, the ship's surface preferably adopts a physically based rendering (PBR) workflow. The texture channels such as background color, metallicity / roughness, normal and ambient occlusion (AO) are uniformly organized, and training sensitive elements such as ship paint, draft markings, ship name / identification number, and reflective coating are explicitly expressed to avoid distortion of the contrast between the navigation lights, outline and background at night due to over-idealization of materials.

[0114] To address the frequent state changes and equipment operation requirements during training, this invention loads functional components and interactive objects onto the ship's hull using a "componentized + semantic" approach. These include anchor winches and chains, navigation lights and signal lights, deck lighting, radar antennas, communication equipment, and deck work equipment. Each component is bound to a unified naming convention and interface parameters (position, orientation, on / off status, flashing period, luminous intensity, visible distance threshold, etc.) at the model level. This allows the system to dynamically enable / disable and adjust lighting logic and visibility performance based on the question bank configuration and environmental conditions. This supports real-world generation and verifiable training under combinations of factors such as "time condition grading, visibility grading, and training objective grading."

[0115] 2. Water area and environment modeling and sea surface optical consistency The water area and environment modeling follows the principles of "real-time performance + physical consistency + controllable parameterization". It preferably adopts a GPGPU-based spectral / frequency domain wave simulation and rendering coupling scheme to ensure that the sea state level defined in the question bank can be consistent in three aspects: "geometric shape (sea surface undulation) - optical performance (reflection / refraction / scattering) - motion influence (wind flow superposition)", thereby avoiding the situation where "sea state parameters are configured but visual and risk assessment are disconnected".

[0116] (1) Generation and online updating of the spectral model of sea surface height field On the computational side, the system preferably uses CUDA, OpenCL, or DirectX 11 computational shaders to rapidly update the sea surface height field. Based on the required sea state level for training, it selects spectral models such as JONSWAP, Pierson–Moskowitz, and Phillips to generate different wind and wave energy distributions, achieving continuous adjustability from calm to complex sea waves. For ease of engineering implementation, this invention can adopt a height field generation framework of "spectral energy—random phase—inverse transform," the discrete form of which can be expressed as: (9) in: The sea surface in planar position Height relative to time t; It is a wavenumber vector; The energy spectral density given for the selected spectral model; ω is the angular frequency corresponding to the dispersion relation; The phase is random. Through equation (9), the "sea state level" in the question bank can be mapped to a set of spectral parameters (such as significant wave height, main wave direction, peak frequency, etc.), and a consistent sea surface undulation pattern is generated in real time by the GPU.

[0117] (2) A unified expression of the high reflectivity characteristics of sea surface reflection, refraction and grazing angle. On the rendering side, to enhance visual credibility under low solar altitude angle, backlighting, and nighttime light mapping conditions, this invention preferably utilizes the Fresnel effect and environment map reflection to realize the angle-dependent characteristics of sea surface reflection / refraction. This is combined with screen-space reflection, reflection probes, and a fogging scattering model to ensure that the "visibility boundary" under different visibility and lighting conditions is consistent with the training rules. The Fresnel reflection coefficient can be approximated using the Schlick approximation to reduce computational overhead; its expression is: (10) in: The angle of incidence is Reflectance coefficient at time; The reference reflectivity at normal incidence (determined by the refractive index of the medium); It can be calculated from the angle between the line of sight and the normal to the sea surface. Through Equation (10), the system can express the characteristic of enhanced reflection of the sea surface at the grazing angle, thus making it more consistent with the real sea surface "bright band", "glare" and light reflection rules under dawn / dusk / night conditions, reducing the deviation of target outline and signal light recognition caused by unrealistic rendering.

[0118] (3) Controllable consistency between visibility and atmospheric extinction The visibility conditions (mutual visibility / poor visibility) in the question bank should not only manifest as fog / rain effects in a 3D visual scene, but also form controllable extinction and scattering, thus conforming to the boundary of "should be visible / invisible according to rules". Therefore, this invention preferably uses an exponential extinction model to parameterize the visibility distance, the basic form of which is: (11) in: The result of visible brightness (or irradiance) attenuation at a propagation distance of d; Initial brightness; This is the extinction coefficient. The question bank template can map the "visibility level" to a range of values ​​for β, so that the visual visibility is consistent with the radar / AIS usage weights, evaluation gating conditions, etc.

[0119] Through the above-mentioned geometric and optical consistency modeling of the sea surface, this invention can provide a stable, reliable and parameterizable representation of water areas and environment in online training: it not only meets the performance constraints of real-time rendering, but also ensures that sea conditions, visibility and lighting conditions have a consistent impact on the entire process of "identification-judgment-action-verification", providing an environmental basis that is consistent with vision for subsequent risk calculation and automatic assessment.

[0120] 3. Wake wave, spray, and flow field parameterization and their linkage with motion state To enhance the perceptibility of "track evidence" and "maneuvering effects" in encounter situations, and to enable trainees to understand speed changes, rate of turn changes, and maneuvering execution effects through visual cues, this invention preferably introduces parametric representations of wake waves, spray, and flow fields into the three-dimensional visual simulation module, and integrates them with the ship speed output from the ship motion simulation module. Turning rate ,course The same state variables are linked together: on the one hand, wakes and sprays can enhance the intuitive feedback of "whether the action is clear and decisive" as visual evidence; on the other hand, the superposition of flow fields can be used to construct a more realistic scenario of drift error and maneuver compensation, so that the training not only stays at "rule judgment", but also covers the practical ability transfer of "maintaining safety margin in complex environment".

[0121] (1) Multi-scale visualization of Kelvin wakes and spray This invention preferably incorporates a 3D Kelvin wake simulation model to parametrically construct multi-scale fluid phenomena such as ship wake, propeller backflow / turbulent wake, bow pressure wave, and particle spray on both sides of the hull, and updates the wake intensity and morphology in real time according to the ship's motion state. For the stability and controllability of engineering implementation, the wake intensity can preferably be scaled proportionally to the square of the speed or power, for example: (12) in: For wake visualization intensity factor (used to drive wake texture contrast, foam density, or particle emissivity, etc.); For ship speed; The scaling factor related to ship type, displacement and rendering scale can be specified in the asset metadata according to ship type. Through Equation (12), when trainees perform deceleration or acceleration actions, the change of wake / spray can provide intuitive feedback, help them understand the correspondence between the change of the engine clock and the motion response, and provide visual evidence of "action occurs - effect appears" for the evaluation module.

[0122] (2) The influence of steering and lateral velocity on wake morphology To make the effect of steering action on the wake more intuitive, this invention preferably correlates the degree of lateral expansion or curvature of the wake with the steering rate. Establish a linkage relationship and perform parametric control on the wake deflection during turning. For example, the wake curvature intensity factor can be defined as: (13) in: This is the wake deflection / bending intensity factor; This refers to the heading angular velocity (steering rate). is the proportionality coefficient. Equation (13) is used to drive the curvature of the wake trajectory, the lateral component of the particle velocity field, or the deflection angle of the foam band, so that trainees can more intuitively perceive the strength and duration of the turn through the changes in the wake when performing right / left turn avoidance, thereby enhancing the training feedback of "action clarity".

[0123] (3) Superposition of parameter fields of tidal current and coastal current The question bank can be configured with hydrodynamic elements such as tidal currents and coastal currents as parameter fields, enabling training to cover flow deviation compensation and control error management. This invention preferably represents the flow field as a spatiotemporally varying velocity vector field. The system provides weak visual cues in the visual field using drift textures, streamlines, or localized foam bands (the cues can be turned off in assessment mode). At the same time, the flow field parameters are shared with the motion simulation module for ground velocity correction and consistency of risk quantity calculation.

[0124] When the question bank only needs to express uniform flow, it can be simplified to constant flow: (14) in: The magnitude of the flow velocity; The azimuth angle is the direction of flow. Equation (14) can be used to generate the streamline direction of the scene, as well as for motion simulation and radar / AIS relative motion calculation, to ensure that "there are flow deviation clues in the picture" and "flow deviation is considered in risk calculation" are consistent.

[0125] By parametrically and dynamically linking wake waves, spray, and flow fields, this invention enhances the interpretable feedback of maneuvering actions at the visual level, provides more realistic sources of environmental disturbances and deviations at the training level, and provides traceable auxiliary evidence for process indicators such as "action clarity, timing rationality, and sufficiency of verification behavior" at the evaluation level, thereby further improving the realism of realistic training and the ability transfer effect.

[0126] 4. Asset entry and real-scene generation material management To ensure that the 3D visual simulation module can stably generate realistic training scenes under different question types, environmental disturbances, and terminal performance conditions, this invention preferably manages ship, water, and environmental assets in a database according to the method of "reusable materials - configurable parameters - traceable versions," and establishes a mapping relationship with the scene configuration fields of the training question bank. This enables the system to automatically assemble materials, load corresponding rendering parameters, and interactive logic based on the combination of question bank elements, thereby achieving "reproducible for the same question, scalable for the same type, and deployable across terminals."

[0127] (1) Mapping of asset metadata structure and question bank fields This invention preferably establishes a unified metadata description for each type of asset. This metadata includes at least: asset category (ship type / water area / environment), scale range, applicable sea state level, available visibility range, lighting configuration set, list of interactive components, performance level and LOD information, and version number. When generating a scene assembly order, the question bank queries the metadata and selects a matching set of materials based on combinations of conditions such as "time / visibility / water area / traffic organization / ship type and quantity / training purpose".

[0128] To ensure that the display of lights / signals is consistent with the requirements of the "Rules", this invention preferably stores the "light configuration set" as an independent reusable resource unit in the database and records the following in the metadata: light type, mounting point number, visible distance threshold, luminous intensity, flashing period, opening and closing conditions (navigation / anchoring / restricted operation, etc.) and related fields with AIS static information. This allows the question bank to drive the lighting logic with "ship status labels" instead of temporarily splicing them in the scene script.

[0129] (2) Cross-platform export format and resource / instance splitting To balance cross-platform compatibility and web loading efficiency, this invention preferably exports 3D assets in a WebGL-friendly format (preferably .gltf / .glb, compatible with .obj, etc.), and separates geometry, materials, textures, animations, and particle / lighting presets into a "resource layer + instance layer": 1) Resource layer: Resources that can be downloaded repeatedly, such as general ship model, PBR texture, environment texture, wake / spray preset, and sound effect pack; 2) Instance layer: Initial situational parameters, light on / off and flashing cycle, sea state / fog parameters, flow field parameters, target highlighting and interaction on / off, etc., which are instanced parameters bound to specific training sessions.

[0130] By splitting resources into resources and instances, the same ship type can reuse resource layer files in different exercises, and only the instance layer parameters need to be loaded to complete the rapid switching, thereby significantly reducing the overhead of repeated downloads and loading, and improving the engineering feasibility of "multiple retraining sessions for the same question type and rapid assembly under different disturbance conditions".

[0131] (3) Performance grading and consistency maintenance To address the real-time rendering and concurrent access requirements of online simulation platforms, this invention preferably incorporates LOD grading, mesh batching, and instantiated rendering optimization on the asset side, and selects different rendering strategies based on terminal performance: key identification details (contour features, light sources, key markers, etc.) are preserved for close-range targets, while low-polygon meshes and simplified materials are used for distant targets to ensure stable frame rates.

[0132] Meanwhile, to avoid performance optimization compromising the consistency of training evidence, this invention sets "consistency constraints" on key training evidence objects: the opening and closing logic of the signal lights and shapes, brightness contrast, and visibility distance thresholds remain consistent under different LODs; the target ship outline features remain identifiable at both near and far viewing distances; the mapping relationship between wake intensity and changes in engine telegraph / ship speed remains stable; and the fogging extinction parameters are strictly consistent with the visibility levels in the question bank. Through these constraints, the training objectives of "real-world visibility—interactive behavior—evaluable results" can still be achieved under different terminal performance and network conditions.

[0133] (4) Version management and traceable updates To support long-term iteration and the comparability of training results, this invention preferably implements versioned management of asset resources and their metadata: when model geometry, texture materials, lighting presets, or wake / fog parameters are updated, a new version is generated while the old version is retained. This allows the question bank to lock training scenes by version, and the evaluation report to record the material version number for review and traceability. Through asset version management, the system can gradually improve the realism of the image and rendering performance without disrupting existing training standards, thus ensuring consistency between engineering evolution and training evaluation.

[0134] (v) Implementation of the human-computer voice interaction module The human-machine voice interaction module of this invention is used to reproduce the VHF communication process in ship collision avoidance practice, and transforms requirements such as call standardization, completeness of key information, and consistency of confirmation into trainable and quantifiable interactive tasks. This module generates interactive scripts and scoring criteria based on communication scenarios and evaluation elements set in the question bank, and links with the timeline of scenario events and ship motion states to achieve a closed-loop voice operation of "call—response—confirmation—execution," forming a traceable data evidence chain for the automatic evaluation module to determine the consistency of instructions and the standardization of communication.

[0135] 1. Construction of a corpus of maritime communication terminology for training tasks To enable voice interaction to directly serve the "call-response-confirmation-execution" process in collision avoidance training, this invention preferably uses real ship communication scenarios as the data collection object, collects call recordings covering different sea areas, different equipment links (VHF, shipborne intercom, shore-based relay, etc.) and different speaker characteristics (speech rate, accent, background noise intensity), and aligns them with the Standard Maritime Communication Terms (SMCP) text system.

[0136] The corpus construction follows the principle of "standard text as the skeleton and real speech as evidence": on the one hand, standard short sentences / phrases are established using SMCP entries as indexes, covering elements such as call format, ship name / call sign, bearing and heading, intent statement, instructions and confirmations, emergency and safety information; on the other hand, real recordings are segmented by conversation rounds and labeled with key semantic slots, including object, location, heading, speed, maneuvering intent, request / permission / warning, etc., thereby forming structured samples that can be used for training and evaluation.

[0137] Meanwhile, to avoid the training system from overfitting to "standard sentence patterns" and thus failing to generalize to real conversations, this invention introduces variants such as synonyms, colloquial omissions, and error recovery at the corpus level (e.g., differences in number pronunciation, differences in location expression, repeated confirmation and error correction statements), and retains the original waveforms and noise-reduced versions under different noise conditions, forming three switchable data views: "clean speech - real speech - enhanced speech," providing a data foundation for subsequent robust modeling and online adaptation.

[0138] 2. Domain Adaptation and Online Learning of Speech Recognition Models In human-computer voice interaction, speech recognition (ASR) plays a crucial role in "converting the VHF expressions of trainees into computable text," and its output directly affects instruction confirmation, slot integrity verification, and the retention of evaluation evidence chains. To reduce homophonic misjudgments and out-of-domain substitution errors commonly found in general spoken language models in the maritime field, this invention preferably constructs a domain-adaptive recognition model oriented towards SMCP terminology, and introduces a continuous adaptation mechanism during training and operation to ensure stable and usable recognition results under different accents, different equipment links, and different noise conditions.

[0139] (1) Identifying model structure and domain constraint decoding The speech recognition model of the present invention preferably uses a recurrent neural network as the core acoustic modeling backbone (such as BiLSTM or GRU), and achieves end-to-end recognition through CTC or attention-based encoder-decoder framework; in the decoding stage, a dedicated language model of SMCP or a domain-constrained dictionary is introduced to explicitly improve the prior probability of segments such as maritime terminology, call sign / ship name spelling, bearing and number reading, thereby improving the reliability of "key field" recognition.

[0140] To ensure that the recognition results can be used for interpretable evaluation, in addition to the text sequence, the recognition output preferably retains word-level timestamps, confidence scores, and a candidate N-best list. Word-level timestamps are used to align the password with the manipulation input to determine "password-action consistency"; confidence scores are used to trigger a confirmation rollback strategy; and the N-best list is used for candidate completion and secondary confirmation when key slots are missing.

[0141] (2) Semantic slot extraction and integrity determination After obtaining the identified text, this invention preferably extracts slots from SMCP communication elements, including at least: object (other vessel / call sign), position (bearing / distance or relative relationship), motion elements (heading / speed), maneuvering intent (turn right / left, decelerate / stop, etc.), request / permission / warning, and confirmation statements. The system performs a completeness check on the extraction results: if a slot required by the question bank assembly list for the vessel's communication is not extracted, or if the extraction results are significantly inconsistent with the environmental conditions (e.g., bearing field is inconsistent with radar / AIS), it is determined as "incomplete or inconsistent information," and the missing slots and evidence fragments are output in the evaluation report.

[0142] To facilitate quantification, this invention can define a slot integrity rate index: (15) in: This refers to the slot integrity rate; This represents the number of slots required by the question bank in this round of calls; This represents the number of slots that were successfully extracted and passed the format / range validation. The "Information Completeness" item is used in process evaluation and can be combined with confidence gating: when When the data falls below the threshold or the confidence level of the key slot location is insufficient, a standardized follow-up questioning and reconfirmation mechanism is triggered.

[0143] (3) Online learning and continuous adaptation mechanism Considering the wide range of trainees and the significant differences in accents and speaking speeds, this invention preferably incorporates online learning and continuous adaptation into the training mechanism: with the authorization of the trainees, the system incrementally updates the interactive speech (using small-step fine-tuning, playback buffering, and regularization constraints to avoid catastrophic forgetting), and maintains lightweight self-adaptive parameters or speaker embeddings according to "sea area / accent / device type", so that the model gradually approaches the actual call distribution without sacrificing overall stability.

[0144] In addition to the conventional word / phrase error rate, task-oriented metrics such as SMCP key phrase recall rate, numerical and directional field accuracy, and instruction-level semantic consistency rate are preferred for evaluation to ensure that the model optimization direction is consistent with the training objective.

[0145] (4) Confidence-driven confirmation loop and evidence chain retention To ensure interactive security under complex noise conditions, this invention preferably employs a "confidence-driven confirmation mechanism": when the recognition confidence is insufficient or key slots are missing, the system triggers standardized follow-up questions and repeated confirmations, and records the entire chain of evidence as "original audio—recognized text—confidence—confirmation rounds—final instruction." This chain of evidence is used to evaluate "communication standardization and consistency" and also provides high-quality samples for subsequent model learning, thereby achieving a collaborative closed loop between training and model improvement.

[0146] 3. Speech synthesis and standard answer library-driven broadcast output To ensure consistency, standardization, and ease of understanding in system-side responses, repetition confirmations, and error correction prompts during training, this invention preferably incorporates a standard answer library within the voice interaction module. This library drives text-to-speech (TTS) output, using "standardized text-to-speech" as the unified broadcast interface for the training system. This approach allows the platform to output consistent, standardized language across different question types, ambient noise levels, and interaction rounds, reducing comprehension biases caused by variations in system expression and providing verifiable reference text for automatic evaluation.

[0147] (1) Structured organization of the standard answer database The standard response library is preferably indexed by SMCP entries and should at least cover: call format, response format, bearing and distance notification, heading and speed notification, intent declaration (intention to turn right / left, decelerate / stop, etc.), request / permission / warning statements, repeated confirmation and correction statements, and emergency and safety information. To adapt to scenario-based configuration driven by the question bank, this invention preferably templates the response library text, using variable fields (such as the other vessel's name / call sign, bearing, distance, heading, speed, target intent, etc.) as slot parameters, which are then filled in by the scenario status to generate the final text before broadcasting.

[0148] To ensure traceability, the system prioritizes recording each broadcast: the item number used, the slot value filled, the generated text, the broadcast timestamp, and the scene anchor (e.g., risk increase, pre-action confirmation, pre-recovery confirmation, etc.), so that subsequent playback and evaluation can pinpoint "what the system said, when it said it, and why it said it".

[0149] (2) Optimization of speech synthesis model and maritime communication terminology At the speech synthesis level, this invention preferably employs a neural vocoder architecture (such as WaveNet) to map text (or phoneme sequences) into natural speech waveforms, and specifically optimizes the expression style of maritime communication terms to be "clear, concise, and unambiguous": 1) Intonation and stress control: Strengthen the rules for stressing and pausing key information such as numbers, directions, directions, and distances to reduce the risk of key information being swallowed or read in a connected manner, which could cause ambiguity. 2) Pronunciation consistency: Establish domain-specific pronunciation dictionaries and standardized abbreviation pronunciations (call signs, ship name spellings, heading pronunciations, etc.) to avoid pronunciation drift of the same term in different contexts; 3) Comprehensibility priority: Appropriately control the speech rate, pauses between sentences and clarity of the final sound to make the synthesized speech closer to the requirements of "clear, concise and unambiguous" VHF calls.

[0150] (3) Controllable matching of link timbre and ambient noise To enhance training immersion and scenario consistency, this invention preferably selects different preset "communication link timbre / bandwidth" based on the question bank and environmental parameters (wind and wave levels, cockpit noise levels, and intercom equipment bandwidth), allowing for a controllable trade-off between fidelity and communication quality in the broadcast. Preferably, bandpass filtering and compression can be applied to the broadcast signal to simulate a VHF link, and environmental noise can be superimposed with a controllable signal-to-noise ratio, enabling trainees to complete listening and confirmation under near-realistic conditions.

[0151] To quantify the intensity of noise superposition and align it with the difficulty level of the question bank, this invention can define the signal-to-noise ratio (SNR) and configure it according to levels: (16) in: For voice signal power; This refers to noise power. The question bank can map different sea conditions / equipment links to different SNR ranges, ensuring that the "listening difficulty" and "environmental difficulty" are linked in a consistent manner; at the same time, the system should ensure that key terms and fields can still be identified under this SNR, to avoid training becoming inoperable due to inaudibility.

[0152] (4) Binding of broadcast output with evaluation comparison To support interpretable evaluation, this invention preferably binds the TTS broadcast text and ASR recognition text in the same evidence chain: after each round of interaction, the system performs slot consistency and key phrase matching on the "standard text broadcast by the system" and the "recognition text repeated / confirmed by the trainee," and uses the differences as one of the deduction criteria for "communication standardization and confirmation consistency." Through broadcast output driven by a standard answer library, the system can form a stable benchmark during training and evaluation, improving the reproducibility and consistency of criteria in the voice interaction process.

[0153] 4. Robust processing and interactive closed loop in complex noise environments Maritime communications face interference from multiple sources, including wind noise, wave noise, engine and ventilation noise, and equipment electrical noise. Furthermore, the signal-to-noise ratio fluctuates rapidly depending on deck position, cabin structure, and equipment links. To improve the stability and security of voice interaction during online training, this invention preferably introduces noise-robust processing technology at the voice recognition front-end, combined with confidence gating, standardized follow-up questioning, and repeated confirmation mechanisms, forming a traceable interactive closed loop. This allows the system to "understand and speak clearly" while controlling the rollback of misidentifications and preserving a complete chain of evidence.

[0154] (1) Robust processing link at the front end The present invention preferably constructs a processing chain of "voice activity detection - noise reduction - enhancement - feature extraction" at the ASR front end: 1) Voice Activity Detection (VAD): Used to remove long silences and non-speech segments, reduce the interference of background noise on decoding, and provide start and end timestamps for each speech segment to facilitate alignment with the training timeline. 2) Noise reduction: For steady-state noise, methods such as spectral subtraction and Wiener filtering can be used, while for non-steady-state noise, a deep noise reduction model can be used to suppress the masking effect of wind and engine noise. 3) Data augmentation and robust training: During the model training phase, augmentation strategies such as mixed noise, reverberant convolution, and frequency band occlusion are used to improve the generalization ability to different noise distributions, so that the model has stable performance across the three views of "clean speech - real speech - augmented speech".

[0155] In addition to being used for identification, the above-mentioned processing link outputs preferably include noise intensity estimates and effective voice duration, which are used for the quantification of subsequent process indicators such as "communication clarity" and "call effectiveness".

[0156] (2) Confidence-driven inquiry and confirmation mechanism Under complex noise conditions, a single identification result may result in missing key slots or incorrect fields. To avoid erroneous instructions directly driving manipulation behavior, this invention preferably adopts a confidence-driven confirmation mechanism: when the identification confidence is lower than the threshold or the key slot completeness rate is lower than the requirement, the system triggers standardized follow-up questions and repeated confirmations, and incorporates the confirmation rounds into the process evaluation.

[0157] To facilitate implementation and quantification, the system can define the "instruction acceptability" of a round of interaction as a gating function jointly determined by confidence level and slot integrity rate, for example: (17) in: The average confidence level of this round of identification (which can be the average of key fields or key phrases); The confidence threshold; The slot integrity rate (see formula (15)); The completeness threshold; This is an instruction function. When A=0, the system does not accept the instruction and instead enters the follow-up / confirmation process, thereby ensuring the security of the interaction and the controllability of training.

[0158] (3) Verification of consistency between password and operation action To achieve a closed loop of "call-response-confirmation-execution," this invention preferably aligns the trainee's verbal commands (ASR-recognized text) with the operational inputs (rudder angle, engine telegraph, mode switching, etc.) in time to verify the consistency between the commands and actions. For example, after recognizing an intention statement such as "turn xx degrees to starboard / full starboard," the system checks whether a matching operational input occurs within a specified time window; if the command and action are inconsistent, it is recorded as a deduction point and an evidence chain (command fragment, timestamp, operational input curve, ship response curve) is output. This consistency check can effectively constrain trainees' non-standard behaviors such as "verbal confirmation but no execution" or "execution but no proper reporting."

[0159] (4) Structured retention and replayability of interactive evidence chains This invention preferably archives key data from each round of interaction in a structured format, including at least: the original audio, the denoised audio, the recognized text, word-level timestamps, confidence scores, N-best candidates, slot extraction results, confirmation rounds, system-broadcast text (TTS entries and filled slots), and association information with scene anchors. During training playback, speech segments and situational / manipulation evidence can be displayed in a timeline, allowing instructors and trainees to review "why it was judged as non-standard, where the error occurred, and how to improve it."

[0160] Through the robust processing and interactive closed-loop design described above, the human-machine voice interaction module of this invention can remain stable and usable under complex noise and diverse accent conditions, and form a consistent linkage with the question bank-driven scenario events and evaluation criteria, thereby supporting standardized call training and interpretable quantitative evaluation oriented towards practical skills.

[0161] (vi) Implementation of the collision avoidance real-scene generation module The collision avoidance scenario generation module of this invention undertakes the core functions of system integration and parameter scheduling. It is used to structure the abstract elements defined in the training question bank in the form of encounter type, rule trigger conditions, visibility / sea state level, traffic density and evaluation index threshold into an executable set of scenario parameters, and uniformly schedule the collaborative operation of subsystems such as ship motion simulation module, 3D visual simulation module, radar / AIS and voice interaction. At the same time, it solidifies the evidence anchor points and sampling caliber required for automatic evaluation, thereby ensuring consistency and full traceability of the three processes of "scenario presentation - maneuvering behavior - index judgment".

[0162] 1. Modular disassembly and scenario assembly order mechanism To achieve "question bank-driven, indicator-constrained, and automatically generated" real-world training, this invention preferably breaks down the collision avoidance real-world generation module into four basic units: ship, water area, environment, and sound effects. Based on this, a scene assembly list is set up as an intermediate layer: when a trainee selects a certain exercise content and its evaluation indicator points, the system does not directly load the scene in a static level manner. Instead, the rule engine / configuration engine generates an assembly list based on the question bank elements. The assembly list is used to specify the assets to be loaded, the situation parameters to be initialized, the interaction logic to be enabled, and the evidence anchor points to be collected.

[0163] Preferably, the assembly order includes at least the following set of fields.

[0164] (1) Vessel and Initial Situation Fields The assembly order is used to specify the assembly, quantity and type of the vessel and the target vessel, and to provide reproducible initial relative situation parameters of the encounter situation, including relative bearing, relative distance, initial heading, initial speed, and the direction and density of traffic flow generated by the target. At the same time, it provides traffic organization constraints (channels / restricted areas / anchorages / TSS lane dividers, etc.) and permissible maneuvering boundaries, so that the triggering logic of the clauses remains consistent under different disturbance conditions for the same type of problem.

[0165] (2) Sea state, visibility and illumination fields The assembly form is used to map the sea state levels in the question bank to wave spectrum model types and parameters (significant wave height, main wave direction, peak frequency, etc.), the visibility level to fog extinction parameters and precipitation particle parameters, and the time conditions to day / night / twilight illumination and environmental textures, thereby ensuring that visual visibility is consistent with the applicable boundary of the rules and that it is of the same origin as the wind flow superposition in the motion simulation.

[0166] (3) Training evidence objects and prompting policy fields The assembly sheet is used to specify the set of training evidence objects, such as navigation light / signal light status, fog signal triggering logic, wake wave intensity visualization switch, key target highlighting strategy, clickable / trackable / annotable objects, etc., and provides prompting strategies (annotation / weak prompts / no prompts) and their gating conditions, so that the evidence collection and interaction under different training purposes are consistent and controllable.

[0167] (4) Sound effects and communication fields The assembly sheet is used to configure the ambient noise level, VHF link tone / bandwidth preset, alarm prompt tone and communication script trigger conditions, etc., so that voice interaction and scene events are linked offline at the same time, and provide a traceable voice evidence chain for the evaluation module.

[0168] 2. 3D Asset Import, Format Export, and Performance Classification Once the scene assembly list is finalized, the collision avoidance real-scene generation module prioritizes the ship type set, water area type, sea state, and visibility level fields in the assembly list, automatically calls the matching ship, water area, and environment models in the 3D asset library, and completes resource loading, instantiation parameter distribution, and performance grading configuration, thereby ensuring the loading efficiency and rendering frame rate stability of the scene generated online.

[0169] (1) Asset matching and allocation strategy The system prioritizes allocating assets using a "mandatory resources + optional enhancement resources" approach. 1) Required resources include: models of the ship and the target ship, corresponding light / signal pattern presets, basic water surface materials and environment textures, and basic fog / precipitation particle presets; 2) Optional enhancement resources include: wake / spray particle presets, local streamline visualization, special vessel type operation status attachments (such as towing, restricted operation markings, etc.), and shoreline / navigation mark resources for specific waters.

[0170] When the vessel type configured in the question bank is a special vessel type or status (such as a towed vessel), the system prefers to read its "status component list" from the asset metadata and attach the corresponding signal light type and visibility parameters in the scene to ensure that the visual evidence is consistent with the triggering conditions of the rule clauses.

[0171] (2) Separate loading of resource layer and instance layer To reduce the overhead of repeated downloads and loading, the collision avoidance real-world generation module preferably loads and distributes data in a "resource layer + instance layer" manner: 1) The resource layer is only downloaded upon first use or version update, including model geometry, texture materials, general lighting / particle presets, and sound effect packs; 2) Instance layers are generated and distributed with each training session, including initial situational parameters, light on / off and flashing cycle, sea state / fog parameters, flow field parameters, and prompting strategies and interaction switches for this training session.

[0172] This splitting mechanism allows the system to quickly reuse the same ship type in multiple retraining sessions, meeting the online training requirements for "quick start and quick topic selection".

[0173] (3) Format export and cross-platform compatibility To ensure cross-platform compatibility and efficient web loading, assets are preferably exported to a WebGL-friendly format (preferably .gltf / .glb, compatible with .obj, etc.), while maintaining consistent reference relationships between geometry, materials, textures, animations, and lighting / particle presets. For scenarios requiring deployment on different terminals (PC / mobile), the system preferably provides multiple resource packages (high / medium / low), and records the selected resource package level in the assembly list to trace the "visual performance standard" during evaluation.

[0174] (4) LOD classification, instantiation and performance constraints The collision avoidance real-world scene generation module prioritizes LOD (Level of Detail) grading and instantiation rendering configuration based on terminal performance and scene scale: 1) For close-range targets, retain key identification details (outline features, light sources, key markings, etc.) to ensure that the siren and ship outline are identifiable within the training visual range; 2) Low-polygon meshes and simplified materials are used for distant targets, and instantiation rendering is used to reduce draw calls; 3) When there are many targets, the system can reduce the material complexity or particle density of non-critical targets to ensure stable frame rate.

[0175] Meanwhile, to avoid performance grading from disrupting the consistency of training criteria, the system sets consistency constraints on the on / off logic of traffic lights, brightness comparison, and visible distance thresholds, so that key training evidence under different LODs remains stable and reproducible.

[0176] 3. Consistency constraints between web-based loading and rendering and the actual scene On the web, the collision avoidance real-world generation module preferably calls the WebGL rendering engine to complete model loading, scene tree construction and real-time rendering, and solidifies the consistency constraints required for collision avoidance training into common boundary conditions for rendering and logic, so that what trainees see, manipulate and the system evaluates are all under the same caliber, avoiding inconsistencies such as "can be presented but does not meet the conditions of the question bank" and "can be interactive but the evidence is not traceable".

[0177] (1) Scene loading and timeline are unified During the loading phase, the system preferably completes the following steps in the order listed in the assembly list: resource layer download and verification, instance layer parameter distribution, scene node attachment, light / particle preset initialization, and alignment with the motion simulation timeline. After the simulation begins, rendering frame updates and kinematic stepping are preferably driven by a unified simulation clock, and the state freeze and breakpoint recording are kept consistent during pause / resumption to ensure stable playback and evaluation evidence positioning.

[0178] (2) Visual consistency constraints To ensure that visibility, illumination, and sea surface reflection / refraction parameters are strictly consistent with the conditions in the question bank, the system preferentially sets visual consistency constraints: 1) Visibility constraints: The fog extinction, precipitation particle density and visibility distance thresholds must be consistent with the visibility levels in the question bank to avoid the deviation of "the rules should be invisible but the picture is clearly visible"; 2) Illumination constraints: The ambient light and skybox during day and night / dawn and dusk must be consistent with the time conditions in the question bank, and the luminous intensity and visible distance of the signal lights switch with the time conditions; 3) Sea surface optical constraints: The Fresnel reflection and environmental reflection intensity remain stable with the change of viewing angle, making the nighttime light reflection and backlight glare performance reliable and avoiding systematic misleading of contour recognition.

[0179] (3) Rule consistency constraints The system preferably links the navigation light / signal display logic to the ship's status (navigation / anchoring / restricted operations, etc.), and aligns it with the training objectives of the question bank and the evaluation evidence chain: 1) When the question bank requires identification of signal lights and shapes, it must be ensured that the combination of lights / shapes, flashing period and visible direction meet the corresponding ship status; 2) When the question bank is under poor visibility conditions, the fog signal triggering logic and alarm prompts should be consistent with the rule requirements and coordinated with the triggering conditions of the voice interaction script; 3) When the question bank involves multiple targets or special ship types, the system must ensure that the status display of each target is consistent with the static information of AIS, so as to avoid the "visual and AIS contradiction" affecting the training judgment.

[0180] (4) Dynamic consistency constraint To enable trainees to understand the manipulation effects through "situational changes + visual cues," the system preferentially sets dynamic consistency constraints: 1) The target ship's motion, tidal current superposition, wake / spray effect, and state variables such as speed and turning rate are linked; 2) The target location calculated and visualized by radar / AIS relative motion, CPA / TCPA, and other methods are consistent and originate from the same source; 3) Key event points (nearest encounter, avoidance start / end, return to flight start / completion, etc.) are defined using the same timestamp and the same anchor point in both the visual and assessment.

[0181] (5) Interactive consistency constraints To support the training process of "observation-judgment-decision-execution" and to serve evidence retention, the system prioritizes providing clickable / trackable / annotable capabilities for key objects: trainees' actions such as selecting targets, plotting, observing with telescopes, adjusting radar parameters, querying AIS and making VHF calls should all be recorded in the same event bus and can be replayed; at the same time, interactive events need to be linked to evaluation indicators so that the evaluation report can pinpoint "when what was done and why points were deducted or added".

[0182] 4. From "Generating Scenes" to "Generating Evaluable Training" The result generated by the collision avoidance real-scene generation module of the present invention is as follows: Figure 4As shown, this is not just a visual representation, but also a set of training examples that can be evaluated. To ensure that the training results are quantifiable, the deductions are interpretable, and the retraining is comparable, the system preferably generates evaluation connection points and evidence sampling schemes simultaneously with the scenario generation, and archives them along with the assembly list, forming a complete data loop of "scenario-process-result". This supports comparisons among multiple trainees on the same topic, multiple retraining comparisons among the same trainees, and a controllable progression of training difficulty.

[0183] (1) Simultaneous solidification of assessment anchor points and evidence objects The collision avoidance scenario generation module preferably uses a unified set of evaluation anchor points fixed in the assembly order. These anchor points are used for repeatable sampling and alignment of the training process. The anchor points should include at least: first target detection, first encounter type determination, first execution of avoidance maneuvers, minimum DCPA moment, hazard alarm triggering / clearing, start of return-to-go, and return-to-go completion. In addition to recording timestamps, each anchor point is preferably bound to replayable evidence objects, including situational screenshots, radar / AIS target status snapshots, manipulation input segments, and voice interaction segments, to ensure that the evaluation report points to verifiable evidence.

[0184] (2) Sampling and labeling of risk quantity curves To support the quantitative assessment of "whether hazard identification is timely, whether avoidance is effective, and whether rerouting is reasonable," the system preferentially samples key risk quantities over time and associates them with anchor points, including DCPA, TCPA, CPA results, target relative bearing change rate, and yaw distance (XTD). The sampling frequency and sampling window are preferably provided by a question bank template and recorded in the assembly order to ensure consistent sampling across different terminals and network conditions.

[0185] (3) Communication and alarm event timeline When the question bank requires training on VHF cooperative avoidance or fog / alarm handling, the system preferentially generates a timeline of communication and alarm events: recording interaction nodes such as calls, responses, confirmations, follow-ups, and final instruction adoption, as well as CPA / TCPA alarm triggering, deactivation, and handling behaviors (e.g., adjusting radar range, target acquisition, and plotting). This timeline, together with motion status and visual snapshots, constitutes an evaluation evidence chain, enabling the evaluation module to determine whether "communication is standardized, information is complete, and instructions are consistent with actions."

[0186] (4) Comparability of training instance archives and retraining The collision avoidance scenario generation module prioritizes the unified archiving of assembly orders, asset version numbers, evaluation anchor point configurations, sampling criteria, and process data generated during training runs. This ensures that subsequent playback and retraining can reproduce the same encounter situation under identical conditions. For instantiation training of the same question type under different perturbation conditions, the system prioritizes maintaining a strategy of "unchanged rule constraints and variable parameter range": that is, the clause triggering logic and evaluation index definition remain consistent, while environmental perturbation parameters vary within the given range of the question bank. This encourages trainees to develop transferable skills while ensuring the fairness and interpretability of training comparisons.

[0187] (vii) Implementation of the real-scene training terminal module The real-scene training terminal module of this invention not only undertakes the presentation and manipulation input of three-dimensional scenes, but also serves to connect the key links of "task-driven - process interaction - data retention - intelligent evaluation": On the one hand, the terminal needs to unify the capabilities of ship motion simulation, visual rendering, water environment, radar / AIS and voice interaction onto the same timeline to ensure that multi-source information is consistent, interpretable and traceable during training; on the other hand, the terminal needs to build a low-learning-cost human-machine interface for ship bridge operation habits, enabling trainees to complete observation, judgment, decision-making and manipulation under multi-source information loads close to the real duty environment, thereby providing a stable interactive foundation for safety awareness cultivation and emergency response capability improvement.

[0188] like Figure 5 As shown, the present invention preferably divides the terminal UI into five functional areas that are decoupled but coordinated, and ensures the training effect and evaluation usability through three mechanisms: "partitioned full screen", "event linkage" and "evidence collection".

[0189] 1. Function area division and partitioned full-screen mechanism The terminal UI preferably includes a toolbar function area, a visual function area, a nautical chart function area, a radar and AIS function area, and a voice interaction function area; each function area can be displayed in full screen independently to meet the information presentation focus switching of different usage scenarios such as teaching demonstrations, personal training, and assessment.

[0190] To ensure that the switching between partitions does not disrupt the evaluation criteria, the terminal preferably maintains the continuity of the simulation timeline and data acquisition when switching between full-screen and split-screen modes, and records interface state change events (switching timestamp, currently selected target, current display mode, etc.) as part of the evidence chain, thereby supporting the reproduction of the trainee's observation path and information load state during playback.

[0191] 2. Implementation of the toolbar function area The toolbar serves as the control center for the training session, and is ideally used to complete training process control, scene invocation, evidence retention, and result archiving. It works in conjunction with the collision avoidance real-world generation module and the automatic evaluation module on the same timeline, thereby ensuring that the entire training process from "entering practice" to "generating an evaluation report" is controllable, replayable, and traceable.

[0192] (1) Practice selection and condition screening The toolbar preferably provides an entry point for selecting exercises, loading training tasks according to the question bank index. Each task should at least include the type of encounter situation, environmental conditions (time, visibility, sea state), traffic density level, and evaluation indicators. To reduce search costs and support differentiated training, the toolbar preferably supports filtering by job position (captain / driver / watch officer / trainee), training objective (learning / retraining / assessment / evaluation), and difficulty level. During loading, a summary of the key elements of the exercise (such as the number of objectives, whether it includes narrow waterways or lane separation systems, whether it includes voice coordination requirements, etc.) is displayed simultaneously, allowing trainees to clearly understand the training objectives and evaluation criteria before entering the scenario.

[0193] Once the trainee confirms their selection, the terminal writes the selected practice number and the current filtering conditions into the training session record, and triggers the collision avoidance real-world generation module to generate the corresponding scenario assembly list and evidence collection configuration, ensuring that the training input and evaluation output correspond.

[0194] (2) Start, Pause and End The toolbar preferably provides start, pause, and end controls for uniformly driving the simulation timeline: 1) Start: After loading is complete, start simulation timing and data acquisition, record the training start timestamp, and initialize evaluation anchor point monitoring; 2) Pause: When paused, the motion state, radar / AIS echo and scene rendering state are frozen, and the breakpoint is recorded; at the same time, input events during the pause are masked or marked separately to avoid interference with the evaluation; 3) End: When the simulation ends, the simulation stops and the training summary is generated. The summary includes at least a list of key events, violation points, indicator achievement and preliminary retraining suggestions. All process data is submitted to the automatic evaluation module for scoring and report generation.

[0195] (3) Saving, playback and commenting The toolbar preferably provides save, replay, and comment functions to solidify the evidence chain of the training process and support post-training review: 1) Preservation: Preserve the complete chain of evidence, including situation snapshots, control inputs (rudder angle, engine telegraph, mode switching, etc.), radar / AIS target status, voice interaction logs, and key event timestamps; 2) Playback: Playback is performed along the timeline and supports speed adjustment, pause, bookmarking key moments, and jumping; during playback, the display of each function area is synchronized with the original training process, making it easy to locate "when to judge, when to act, and when to review"; 3) Feedback: Supports the overlay of automatic and manual feedback. Automatic feedback is generated by the assessment module, which identifies deduction points and evidence locations. Manual feedback allows instructors to add text or voice explanations at key bookmarks, which are saved along with the evidence chain for later review and comparison.

[0196] (4) Manipulating cards and key input records The toolbar preferably provides control cards to display and record key control commands and their trigger times, including at least changes in engine telegraph, rudder angle / ratio, and switching between heading hold and track hold modes. Control cards are linked to evaluation anchor points: when the system detects key nodes such as the first valid control input, the avoidance start point, and the go-around start point, it automatically generates a marker on the card and writes it into the evidence chain, providing structured input for subsequent judgments on "whether the timing of the action was reasonable, whether the action was clear, and whether the go-around was standardized."

[0197] Meanwhile, the control card preferably supports quick review of the most recent input and current status (heading, speed, rate of turn, yaw distance, etc.) during training, so as to reduce the operating cost of trainees under multi-source information load and reduce misoperation.

[0198] 3. Implementation of the visual function area The visual function area serves as the primary visual channel for trainees' situational awareness and maneuvering decisions. It is ideally designed to provide a 3D scene display that closely resembles the observation habits of a real ship's bridge, maintaining consistency with the output of ship motion simulation. To meet the diverse needs of teaching, training, and assessment, this invention emphasizes "realistic visibility" and "interpretable feedback" within the visual function area: ensuring that target presentation under day / night, dawn / dusk, fog / rain, and sea state conditions conforms to the question bank settings and rule boundaries; and providing trainees with understandable feedback on key maneuvering effects (turning, deceleration, wake changes, relative situation evolution, etc.) and forming replayable evidence.

[0199] (1) Multi-view switching and observation constraints The visual function area offers multiple perspectives, including bridge view and external follow view, to support different training modes: the bridge view is used to closely resemble practical observation and judgment habits, emphasizing signal light and shape recognition, outline recognition, and relative orientation judgment; the external follow view is used for teaching demonstrations and understanding of action effects, making it easier to observe the clarity of avoidance actions and changes in encounter distance.

[0200] To ensure fairness in training and consistency in assessment, some auxiliary perspectives can be restricted or the intensity of prompts can be reduced in the assessment mode. Perspective switching events (switching timestamp, perspective type, duration) can be recorded as process evidence to evaluate trainees’ observation strategies and information use behaviors.

[0201] (2) Scene parameter linkage and visibility consistency The optimal selection of visual function areas is linked to the time conditions, visibility conditions, and sea state levels in the question bank: 1) Time-based linkage: The ambient light and contrast strategies differ under daytime, dawn / dusk and nighttime conditions. In nighttime mode, target recognition prioritizes the use of hazard lights and auxiliary information. 2) Visibility linkage: Under conditions of mutual visibility and poor visibility, the fogging intensity and visibility distance are controlled according to the parameters in the question bank to ensure that the boundary between "visible / invisible" and the applicable rules are consistent; 3) Sea state linkage: The sea surface undulation, reflection and wave intensity change with the sea state level, and generate consistent visual feedback with the ship's speed and turning status.

[0202] The above-mentioned linkage results are written into the scene instance record at the beginning of training and reconstructed with the same parameters during playback to ensure that the retraining comparison can be reproduced.

[0203] (3) Key objective prompting strategies and evidence objects To balance learning and assessment, this invention preferably provides a configurable target prompting strategy: in learning mode, key targets can be marked or given weak prompts to guide trainees in establishing an "observation-identification-judgment" path; in assessment mode, prompts can be turned off, retaining only necessary interactive evidence objects. Regardless of whether prompts are enabled or disabled, the system should retain traceable evidence objects, including the target ship's hull number light status, outline feature display, key moment situation snapshots, and relative bearing change records, so that the assessment report can locate evidence and explain the reasons for deductions.

[0204] (4) Ship dynamic instruments and status markings The visual function area preferably integrates ship dynamic instruments, displaying at least heading, speed, turning rate, engine telegraph, rudder angle, and wind and current information, and ensuring that they share the same driving source and consistent values ​​with the motion simulation module. To improve playback and review efficiency, the system preferably automatically inserts status markers at key event points, such as approaching the closest encounter, avoidance begins, avoidance ends, danger cleared, and resumption of navigation begins / completed, and these are simultaneously marked on the instrument curves, providing intuitive evidence for subsequent analysis of "when to start the action, whether the action was sufficient, and whether the resumption of navigation was appropriate."

[0205] (5) Telescope and lookout behavior records The visual function area preferably provides a telescope for localized magnification and stable observation, supporting signal light and shape recognition, hull outline and intent assessment. The number of times the telescope is used, its duration, and the objects observed are preferably recorded as evidence of "lookout behavior," and can be used in conjunction with radar / AIS queries, target selection, and other events for process evaluation to determine whether trainees have conducted continuous and effective lookout and whether cross-validation was performed at key points.

[0206] 4. Implementation of the nautical chart function area The chart function area provides trainees with navigational environment and route constraints information closely related to collision avoidance training. Its design principle emphasizes "collision avoidance guidance, information convergence, and operational controllability" within the constraints of terminal performance and learning costs, avoiding excessive layers and complex functions that could distract or interfere with identification. The chart function area maintains consistency with ship motion simulation, radar / AIS, and assessment sampling, supporting trainees in considering both route execution and navigational environment constraints when making encounter judgments and avoidance decisions, preventing training from merely focusing on "correct avoidance" while neglecting "route maintenance and safety margins."

[0207] (1) Display mode and cognitive load control The nautical chart function area preferentially provides north-facing and relative motion display modes, and the display strategy can be locked according to the training mode to reduce cognitive burden: in the initial training, north-facing and relative motion are preferred to reduce the extra burden on trainees in orientation conversion and understanding relative motion; in advanced training, more display switching can be gradually opened to improve the ability to adapt to different display habits. The system should record display mode switching events and be reproducible in playback so that instructors can analyze the trainees' information usage paths.

[0208] (2) Information hierarchy and layer convergence The nautical chart information display is preferably locked at the basic display level, focusing on elements related to collision avoidance height, including channels, anchorages, no-navigation zones, traffic separation system boundaries, recommended routes, and key turning points. Layers irrelevant to the training objectives or that are highly distracting for beginners are preferably turned off by default, and can be enabled by instructors in teaching mode when necessary. By converging information levels, the risk of "not being able to see the key points clearly or misinterpreting the environment" is reduced, thereby improving training efficiency.

[0209] (3) Interactive control and target information viewing The chart function area preferentially supports zooming in / out, target point selection, and key waterway feature prompts. When trainees select a target, the system can display key fields such as the target's relative bearing, distance, heading, and speed, consistent with radar / AIS targets. For constraint elements such as channel boundaries and restricted area boundaries, the system can provide prompts when the risk of a target approaching or crossing the boundary occurs, and record the prompt triggering event as part of the evidence chain to assess "whether the waterway constraints were recognized and the deviation was reasonably controlled".

[0210] (4) Route and Collision Avoidance Assistance Information To integrate collision avoidance decision-making with route execution, the chart function area prioritizes automatically loading practice routes and provides core navigation parameters related to collision avoidance training: 1) Yaw Distance (XTD): Used to evaluate track maintenance and yaw control, and to avoid excessive yaw in pursuit of large avoidance. 2) Next turning point bearing (BTW), next turning point distance (DTW), and time to the next turning point (TTG): These are used to support "avoidance decision-making under route constraints" and help trainees rationally choose the avoidance range and rerouting timing while meeting safety margins.

[0211] The above parameters are preferably calculated from the same source as the motion simulation, and are sampled and labeled at key event points (avoidance start, minimum encounter distance, re-entry start / completion, etc.) to provide a basis for the evaluation module to judge "avoidance effectiveness and route execution rationality".

[0212] (5) Linkage with assessment evidence Key interactive events in the nautical chart function area (zooming in / out, target selection, route viewing, constraint confirmation, etc.) should be recorded and aligned with assessment anchor points: for example, whether high-risk targets were selected during periods of heightened risk, and whether route deviations were checked before resuming operations. Through this linkage, the assessment report can explain "why the lookout was deemed insufficient or the review insufficient" and provide targeted retraining recommendations.

[0213] 5. Implementation of Radar and AIS Functional Areas The radar and AIS functional areas are designed around the core training chain of "target detection - tracking - relative motion analysis - collision risk assessment - avoidance verification". The principle is to retain the set of functions related to collision avoidance under the constraints of terminal performance, and to ensure that its display, calculation and alarm logic are consistent with the parameters of ship motion simulation and scenario assembly. This enables trainees to form transferable risk identification and verification habits under different conditions such as mutual visibility and poor visibility.

[0214] (1) Radar core functions and parameter control The radar section preferably retains key collision avoidance functions from the analog radar model, including range switching, echo gain, sea clutter / rain clutter suppression, target acquisition and tracking, relative / true motion display, CPA / TCPA calculation, and alarm threshold setting, while ensuring that echo performance is consistent with sea state and rain / fog conditions. Trainees' adjustments to radar parameters, target acquisition / tracking behaviors, and alarm handling behaviors should all be recorded as quantitative evidence of "situational awareness and risk identification capabilities."

[0215] In training under poor visibility conditions, the system can increase the weight of radar information in the assessment and set key node verification requirements (such as pre-operation verification and pre-return verification) to guide the formation of a standardized process "led by radar / multi-source information".

[0216] (2) AIS information fusion and target list management The AIS section preferably provides a fusion display of static and dynamic target information, including fields such as ship name or identification number, heading and speed, ship type, destination, and draft; it also supports association with radar targets, target list sorting (by CPA / TCPA / distance, etc.), and highlighting of high-risk targets. Through target sorting and highlighting, trainees can quickly locate and prioritize high-risk targets in multi-target situations, thereby supporting decision-making in complex scenarios.

[0217] (3) Risk quantity calculation, display and alarm linkage The radar and AIS functional areas should ideally display risk parameters such as CPA and TCPA using a unified caliber, triggering alarms when risk parameters reach thresholds. Simultaneously, risk trends (e.g., decreasing TCPA, approaching CPA) should be marked on the interface to help trainees understand risk evolution rather than relying solely on one-off numerical values. Alarm triggering, alarm clearing, and trainee actions (e.g., adjusting range, reacquisition, replotting, confirming high-risk targets) should be recorded and aligned with evaluation anchors to assess whether "hazards were identified in a timely manner, necessary checks were performed, and avoidance was initiated within a reasonable timeframe."

[0218] (4) Data collection for training and evaluation linkage To support interpretable evaluation, the system preferentially selects key radar and AIS operations and states as process data acquisition items, including... This includes: radar power-on and operating mode, range switching frequency and timing, gain and clutter parameter adjustment trajectory, target acquisition / tracking list changes, alarm threshold settings, target selection and information viewing records, etc. The above data, along with operational inputs, visual observation behaviors, and voice interaction records, constitute a chain of evidence, enabling the assessment report to pinpoint "when radar verification was lacking, when alarms were not handled correctly, or when high-risk targets were not locked," and to generate targeted retraining recommendations accordingly.

[0219] 6. Implementation of the voice interaction function area The voice interaction function area is used to replicate the VHF communication process in practice, enabling training to simultaneously cover "communication standardization" and "cooperative avoidance capabilities." This function area works in tandem with the human-machine voice interaction module: on the one hand, it provides trainees with interactive entry points such as calling, answering, confirming, and correcting; on the other hand, it aligns and records voice content with key maneuvering behaviors and risk evolution processes, forming a traceable chain of evidence to support post-event review and automatic evaluation.

[0220] (1) VHF call flow and interactive controls The voice interaction function area preferably provides a voice coordination and avoidance channel between ships, supports basic operations such as calling, answering, confirming and correcting, and can set the communication style and response strategy of the other ship according to the question bank assembly list (such as standard language, simplified language, delayed response, etc.) to construct communication environments of different difficulty.

[0221] To ensure the training process is controllable, the system can trigger communication tasks at key scenario nodes, such as when the risk rises to a threshold, before preparing to perform an avoidance maneuver, or before resuming flight, and prompt trainees to complete the corresponding communication steps on the interface. In assessment mode, the intensity of prompts can be reduced or prompts can be given only when key steps are missed, in order to test the trainees' mastery of the standardized procedures.

[0222] (2) Voice input / output and peripheral device adaptation The voice interaction function area preferably supports peripherals such as headsets and speakers, adapting to the input and output capabilities of different terminals; voice input and output are linked to environmental noise parameters, enabling trainees to conduct listening and expression training under conditions of wind and wave noise, cabin noise, or limited link bandwidth. The system should save the types of input and output devices used and the noise level in the training session records for playback and evaluation of the interpretation of "sources of difficulty in identification" and "clarity of expression".

[0223] (3) Alignment of evidence chain records and assessments The voice interaction function area preferably archives voice content in a structured manner, including at least the recognized text, confidence level, timestamp, key field extraction results, confirmation rounds, and the final adopted command; at the same time, it records voice events in alignment with manipulation inputs, radar / AIS verification events, and key evaluation anchor points. For example, after "password confirmation," it can check whether a manipulation action consistent with the password has occurred within a specified time window, and use the consistency result as a basis for process evaluation.

[0224] During playback, the system can simultaneously display voice clips and situation snapshots, risk curves and manipulation curves on the timeline, enabling instructors and trainees to review "when the communication took place, whether the communication was standardized, whether confirmation was completed, and whether the commands and actions were consistent," thereby supporting the closed-loop improvement of "training-review-retraining."

[0225] (viii) Implementation of the automatic evaluation module The automatic evaluation module of this invention serves as the core of the online training loop. It provides interpretable and quantifiable scores for trainees' operational results and processes during collision avoidance training, and outputs verifiable evidence, reasons for deductions, and targeted retraining suggestions. Unlike simply determining whether a collision occurred, this invention's automatic evaluation simultaneously covers two main aspects: "whether correct and effective avoidance methods were adopted" and "whether the judgment-avoidance-return-to-goal maneuver was completed efficiently with minimal trial and error, in accordance with practical procedures, and with standardized VHF communication skills." This supports online evaluation and comparative retraining for accident case correction training and scenario-based training of the rules.

[0226] 1. Evaluation Data Collection and Evaluation Anchor Point Definition The automatic evaluation module preferably uses multi-source process data available on the platform as input, including maneuver inputs (rudder angle, engine telegraph, control mode switching, etc.), ship motion status (heading, speed, turning rate, position, yaw distance, etc.), radar / AIS target information and risk quantities, voice communication records, and key event timestamps. To ensure the repeatability and cross-subject comparability of the indicators, this invention preferably sets a unified evaluation anchor point during the simulation operation to align sampling of key moments and solidify evidence location.

[0227] (1) Evaluation of the set of anchor points The assessment anchor points include at least: initial target detection, initial encounter type determination, initial effective avoidance maneuver execution, risk alarm triggering, minimum encounter distance time, risk resolution, commencement of re-entry, and completion of re-entry. For questions involving voice collaboration, these further include: initial call, counterparty response, initial confirmation, follow-up / correction rounds, and final instruction adoption. Each anchor point is timestamped and linked to corresponding situation snapshots, radar / AIS snapshots, maneuver input segments, and voice segments for scoring interpretation and replay review.

[0228] (2) Structured extraction of index vectors For any training question or its instantiation scenario, the system extracts and structures a vector of student operation results from the aforementioned multi-source data, preferably represented as follows: (18) The first m components are used to characterize indicators related to the accuracy of the operational results, including at least whether the encounter type identification is correct, whether the responsibility for giving way / direct flight is correctly assigned, whether clear avoidance actions are taken according to the rules, whether the avoidance initiation timing is within a reasonable window, whether DCPA / TCPA meets the threshold requirements, whether the final encounter distance reaches the safety margin, and whether the re-entry timing is reasonable. The last n components are used to characterize indicators related to the standardization and proficiency of the operational process, including at least whether the lookout and multi-source verification are sufficient, whether the control steps are complete, whether the control frequency is reasonable, whether there are repeated attempts or frequent corrections, whether VHF communication is standardized, whether the commands and actions are consistent, and whether the necessary verifications are completed before the re-entry.

[0229] To avoid evaluation drift caused by differences in sampling frequency, playback truncation, or terminal performance, the calculation of each index in vector x is preferably based on sampling or statistics of state variables and event logs near the evaluation anchor point, thereby ensuring that "different people on the same topic, and different times by the same person" have a consistent calculation caliber.

[0230] 2. Generation and management of standard answers and grading criteria To achieve interpretable scoring and retrainable comparison, this invention preferably establishes a standard answer for each training question that corresponds one-to-one with its evaluation index, and manages the standard answers in a versioned manner according to question type and element combination, so that the same question type maintains consistent rule constraints under different environmental disturbances, while allowing the safety threshold to be adjusted within a reasonable range as the environment changes.

[0231] (1) Definition of standard answer vector For the same question, based on the rules, standard operating procedures, and the platform's built-in reference solutions (which may come from expert annotations, rule derivation, or historical high-score sample statistics), a standard answer vector is defined: in: Let i represent the standard objective of the i-th evaluation indicator.

[0232] For discrete indicators (such as "whether the encounter type is a cross", "whether a right turn should be taken to avoid it", "whether a standard call was initiated and confirmed", etc.), Preferred values ​​are 0 / 1 or category labels; for continuous indicators (such as DCPA, TCPA corresponding to the start of the avoidance, maximum yaw distance XTD, risk margin at the start of the go-around, etc.). It is preferable to define it as a threshold range or target range rather than a single point value, in order to adapt to environmental disturbances and multiple solutions.

[0233] (2) Source of standard answer and handling of multiple solutions The standard answer may come from one or more of the following sources: 1) Expert annotation: Personnel with practical experience in collision avoidance provide recommended avoidance strategies, communication templates, and verification steps for typical situations, and determine the acceptable safety threshold range; 2) Rule derivation: Based on the applicable conditions of the clauses and the division of responsibility for giving way, derive the "set of actions to be taken" and clarify the constraints that the actions must meet (e.g., the actions should be obvious, decisive, and easy for other vessels to understand). 3) Statistics on excellent samples: Extract typical trajectories and call templates for the same question from the high-scoring samples accumulated on the platform to form a reusable reference solution.

[0234] When a question has a situation where "the rules allow multiple solutions", this invention preferably defines the standard answer as an equivalent solution set or an allowed set: that is, discrete decision gives an acceptable category set, and continuous index gives a threshold range, thereby avoiding wrong deduction of points due to multiple solutions to a question and improving the fairness of scoring.

[0235] (3) Version management based on element combination Standard answers are preferably indexed and managed according to "question type - environment - ship type combination", and their version number and scope of application are recorded in the question bank template. For the same question in instantiation scenarios under different disturbance conditions, the preferred management strategy is to maintain "unchanged rule constraints and variable parameter range": the clause triggering logic and indicator definition are consistent, and the continuous threshold is adjusted within a preset range according to visibility, sea state and traffic density, so as to balance the consistency of rules and the rationality of environmental differences.

[0236] At the same time, the standard answers, scoring weights, deduction criteria, and evidence anchor configurations are all optimized and solidified into a structured configuration, so that an evaluation report containing the total score, sub-scores, key evidence, and retraining suggestions can be directly generated after training.

[0237] 3. Accuracy score of operation results The accuracy of operational results is used to measure the trainees' achievement of key rules and safety objectives, answering the question of "whether the correct and effective avoidance methods were adopted". This invention preferably establishes the accuracy score based on a consistent comparison of x and z, and allows for a soft scoring method of "threshold-segmentation-saturation" for continuous indicators to improve discrimination and fairness.

[0238] (1) Calculation of total accuracy score Let the full score of the test be... Accuracy has a weight in the total score of . The weight of the i-th accuracy indicator is (satisfy Then the accuracy score can be defined as: (20) in: Let be the discriminant function for the i-th accuracy index, used to map "whether the standard objective is met" to a score in the interval [0,1].

[0239] (2) Discrete index discriminant function For discrete indicators such as encounter type judgment, yielding / direct flight responsibility allocation, avoidance direction selection, and whether key verification steps have been completed, this invention preferably adopts hard discrimination or permissive set discrimination: 1) Hard discrimination: when = Time to take =1, otherwise take =0; 2) Allowable set discrimination: When the standard answer provides an acceptable set. At that time, if Then take =1, otherwise take = 0.

[0240] By allowing set discrimination, the case where "the rule allows multiple solutions" can be handled, avoiding misjudging reasonable equivalent solutions as errors.

[0241] (3) Soft scoring of continuous indicators For continuous indicators such as DCPA, TCPA, XTD, and re-entry timing, using a hard threshold of 0 / 1 can easily lead to unstable boundaries and insufficient discrimination. This invention preferably employs a soft scoring method with segmented decreasing thresholds and upper and lower bounds. For example, for safety distance indicators that are "the larger the better but with an upper limit," the following discriminant function can be constructed: (twenty one) in: This is the minimum safety threshold; anything below this value can be considered seriously unsafe and will result in a significant deduction of points. To ensure safety, a saturation limit is reached, and scores will not be increased beyond this value to avoid unreasonably encouraging "excessive avoidance." For yaw distance indicators where "the smaller the better," a symmetrical piecewise function can also be used, with a reasonable lower limit set to avoid excessive penalties due to extreme values.

[0242] (4) Critical security gating To avoid situations where "the final score is high but key safety indicators fail to meet the standards," this invention preferably sets gating conditions for accuracy: when core indicators directly related to collision hazards (such as minimum DCPA, whether action is taken within the danger window, etc.) fail to meet the standards, the question can be judged as unqualified or additional penalties can be imposed on the total score, thereby ensuring that the score is consistent with the safety objectives.

[0243] 4. Score for standardization and proficiency of operation process The operational process score is used to evaluate whether the trainee's "observation-judgment-decision-execution-review" chain during the training process is standardized, answering the question of "whether the training was completed in accordance with practical procedures, with minimal trial and error, and with high stability." This invention preferably establishes the process score based on the alignment comparison of x and z, and introduces penalty items for behaviors such as high-frequency small-amplitude manipulation, repeated re-judgment, and repeated calls, to distinguish between "accidentally getting it right" and "stable and controllable process."

[0244] (1) Step proficiency score Let the weight of process (proficiency) in the total score be... Process indicator weights can be reused Or set additional weights (satisfy Then, the proficiency score for each step can be defined as: (twenty two) in: Let be the discriminant function for the j-th process index.

[0245] (2) Process discriminant function and penalty term Process indicators include, but are not limited to: 1) Lookout effectiveness: Whether continuous and effective lookout is conducted at key points such as initial detection, risk escalation, pre-operation verification, and pre-resumption verification, and whether radar, automatic identification system and voice communication are used appropriately for cross-verification; 2) Standardization of Escape Initiation and Execution: Whether the escape was initiated within a window of sufficient margin, and whether there was any delay that led to emergency escape; 3) Clarity of action: Whether the action is obvious, decisive and easy for other ships to understand, and whether there are small, repeated swaying, frequent speed changes or repeated steering. 4) Voice communication standardization: Whether the call format and key information slots are complete, whether the response and confirmation are completed, and whether the commands and actions are consistent; 5) Pre-resumption checks: Ensure that necessary checks are completed after the danger has passed before resuming operations to avoid premature turning back and causing a second approach.

[0246] For the aforementioned indicators, in addition to using hard discrimination, this invention preferably introduces an "invalid operation penalty." For example, a penalty item Q can be constructed for high-frequency operations within a unit of time and superimposed on the process score: when the rudder angle or engine telegraph changes frequently and slightly within a short period of time without significantly improving the risk trend, it is determined that there is a trial operation and points are deducted; when there is a repeated cycle of "avoidance-correction-avoidance again," the deduction is increased. This penalty item can be configured according to the threshold of the question bank to ensure consistency in the criteria for different question types.

[0247] 5. Operation time proficiency score Time proficiency is used to measure overall efficiency and pacing, emphasizing the efficiency of completing training tasks while ensuring safety and compliance, and avoiding prolonged stagnation or ineffective waiting. This invention preferably scores the total time or the time spent in stages, while setting upper and lower bounds and gating conditions to prevent "too fast but unsafe" attempts from receiving unreasonable bonuses.

[0248] (1) Definition of Time Score Let the weight of the time score be... Standard time is The time spent by the students this time was Then, the time proficiency score can be defined as: (twenty three) Among them: when When the time item generates a positive bonus; when At that time, points will be deducted from the time item.

[0249] (2) Upper and lower bounds of time score and gating To avoid abnormal extreme values ​​affecting fairness, this invention preferably uses... The segment is truncated to fall within a preset range; simultaneously, a safety gate is set: when accuracy... If the minimum qualification threshold is not met (e.g., core safety indicators are not up to standard), time-based bonus points may be prohibited or time may be reduced. To ensure "safety first".

[0250] For questions with phased timing, the present invention preferably uses... It is broken down into components such as judgment time, avoidance initiation time, return time, and communication time to support more granular diagnosis and retraining recommendations.

[0251] 6. Final Score and Evaluation Report Generation The final score is the weighted sum of the above three parts: (twenty four) Output the final score Simultaneously, the present invention preferably automatically generates an interpretable evaluation report, the report including at least: 1) Display of sub-scores and weights: Sub-scores and weights for accuracy, process standardization, and time; 2) List of key deduction points and evidence location: timestamps, situation screenshots, radar / AIS replay bookmarks, manipulation input curve segments, and speech recognition text segments corresponding to each deduction point; 3) Risk quantity curve and key point marking: DCPA / TCPA change curve over time, and marking the avoidance start point, minimum encounter distance point, risk resolution point and rerouting point; 4) Process diagnosis and retraining suggestions: Based on the deduction model, actionable retraining suggestions are given, such as "insufficient lookout", "late start of avoidance", "unclear actions", "failure to check before resumption of flight", "missing or insufficient communication slots", etc., and targeted practice is pushed in conjunction with the question bank.

[0252] Through the output format of "score-evidence-explanation-recommendation", the automatic evaluation module achieves traceability, verifiability, and reliability. Closed-loop evaluation of retraining supports the large-scale deployment and continuous improvement of online training.

[0253] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A ship's rule of the road at sea simulation training and evaluation method, characterized in that, The method includes: Build a training question bank and configure scenario element parameters and evaluation index parameters for each question bank item; Based on the training question bank, a scenario assembly list is generated. The scenario assembly list is used to describe the ship set, initial relative situation, environmental and traffic organization constraints, and interaction and evaluation parameters required for real-world training of encounter situations. Based on the scenario assembly list, the ship motion simulation, 3D visual rendering, and radar / AIS fusion display are uniformly scheduled, and voice communication interaction is performed to form a training process; During the training process, the trainees' operational inputs, ship motion status, radar / AIS target information, and voice communication records are collected to form a traceable training evidence data chain. Based on a preset evaluation index system and combined with evaluation anchor points, the training evidence data chain is sampled and judged to achieve automatic scoring of training results and training process, and output evaluation report.

2. A ship's real scene collision avoidance rule training and evaluation system, the system is used to realize the method of claim 1, characterized in that, The system includes: a training question bank module, a ship motion simulation module, a 3D visual simulation module, a human-computer voice interaction module, a collision avoidance scene generation module, a scene training terminal module, and an automatic evaluation module. The training question bank module is used to construct the training question bank and configure scene element parameters and evaluation index parameters; The collision avoidance real-scene generation module is used to map the abstract elements of the question bank to the scene parameters to be executed and to uniformly schedule the various subsystems; The ship motion simulation module is used to calculate the maneuvering response of the ship and the target ship and to provide state quantities for risk quantities and evaluation indicators. The three-dimensional visual simulation module is used to realize the realistic presentation of ships, waters and environment and provide interactive evidence objects; The human-computer voice interaction module is used to reproduce the VHF call process and form a traceable chain of voice evidence. The real-scene training terminal module is used to complete the presentation of multi-source information, manipulation input, and data collection during the training process. The automatic evaluation module is used to quantitatively score the correctness of training results and the standardization of the process, and output diagnostic and retraining suggestions, thereby forming a closed-loop improvement mechanism of "training-evaluation-diagnosis-retraining".

3. The system according to claim 2, characterized in that, The training question bank module follows the consistency principle of "rules and clauses - encounter situation - ship handling decision - result verification", and corresponds the question content to the clause connotation and applicable conditions of the preset rules, and adopts a "layered and graded + difficulty combination" structure to form a multi-level training set. The difficulty levels include at least: time conditions, visibility conditions, water area and traffic organization methods, number and type of vessels, and training objectives. For each exercise, result evaluation indicators, operation process evaluation indicators, threshold parameters, scoring weights and deduction criteria are defined simultaneously to form a standardized question template.

4. The system according to claim 2, characterized in that, The ship motion simulation module adopts a layered modeling strategy: for the ship itself, a horizontal three-degree-of-freedom MMG-separated maneuvering motion model is used, which decomposes external forces and moments into hull hydrodynamics, propeller thrust effect and rudder force effect, and determines key coefficients by combining empirical parameter initial values ​​with maneuverability tests / system identification; for the target ship, a response model is used, preferably a first-order Nomoto model to describe the steering response, and its parameters are calibrated through empirical estimation and data identification, so as to realize online simulation of multiple target ships and deployment of automatic heading / track control.

5. The system according to claim 2, characterized in that, The three-dimensional visual simulation module aims to make the real scene "reproducible, interactive, and evaluable". It constructs a three-dimensional model of the ship's hull and an interactive component structure, and uses a physically based rendering workflow to explicitly express training sensitive elements. The water area and environment are simulated and rendered using GPGPU-based spectral / frequency domain wave simulation and rendering coupling. Fresnel reflection / refraction and environmental reflection are combined to achieve optical consistency of the sea surface. The visualization cues of wake / spray are introduced and coupled with flow field parameters to enhance the perceptibility of manipulation effects. At the same time, an asset storage and version management mechanism is established to support automatic material assembly according to the question bank configuration fields and to perform LOD grading and instantiated rendering optimization.

6. The system according to claim 2, characterized in that, The human-machine voice interaction module is designed for collision avoidance and collaborative avoidance processes. It constructs a corpus based on the Standard Maritime Communication Protocol (SMCP) and provides ASR (Automatic Speech Recognition) and TTS (Text-to-Speech) capabilities for domain-specific terminology. On the ASR side, it combines a domain dictionary / language model to output recognition results with timestamps and confidence levels, and introduces a confidence-driven repeat confirmation mechanism and online adaptation. On the TTS side, it uses a standard answer library to drive standardized broadcasting and optimizes the emphasis and pause rules for key information such as numbers, bearings, and headings. The entire voice interaction process is linked to the scene event timeline in a closed-loop manner of "call-response-confirmation-execution" to form a traceable evidence chain.

7. The system according to claim 2, characterized in that, The collision avoidance simulation generation module sets up a "scene assembly list" mechanism, which serves as a unified mapping layer from abstract elements of the question bank to executable inputs. It generates an assembly list based on the encounter type, clause triggering conditions, environmental level, and evaluation threshold in the question bank, clarifying the ship assembly and initial relative situation, traffic organization constraints, sea state and visibility / lighting parameters, training evidence objects, and sound effects and communication resources. It also uniformly schedules the collaborative operation of ship motion simulation, 3D rendering, water environment, and voice and sound effects subsystems, while solidifying the evidence anchor points and sampling points required for evaluation.

8. The system according to claim 2, characterized in that, The real-world training terminal module includes a toolbar, visual view, nautical chart, radar / AIS, and voice interaction function area. It supports practice selection, start / pause / end, saving playback, and comments. During training, it records manipulation input, motion status, multi-source sensor information, and voice log data to form a chain of evidence for playback.

9. The system according to claim 2, characterized in that, The automatic evaluation module takes result evaluation indicators and operation process evaluation indicators as inputs, and performs interpretable scoring around the two main lines of "accuracy of operation results" and "operation proficiency". During the simulation, a unified evaluation anchor point is set, and the indicator vector is calculated based on the state variables near the anchor point and the trainee input. The discrete indicators are judged by equivalent solution set according to the standard answer vector, and the continuous indicators are judged by threshold / segmentation / saturation soft scoring function. Finally, the module outputs the total score, the scores of each item, the key deduction points and their evidence, the DCPA / TCPA evolution curve annotation, and the targeted retraining suggestions.

Images