Method for quantifying quality of intelligent ship personnel takeover and inferring optimal takeover time

Abstract

The application discloses a kind of intelligent ship personnel takeover quality quantification and optimal takeover time inference method, the method includes by constructing the safety quantization model for evaluating the safety margin of intelligent ship takeover operation, and the output of safety quantization model is used as safety quantization evaluation index, and the risk value of collision between ships is integrated as multidimensional quantification index, and the comprehensive takeover quality index is obtained by the CRITIC-TOPSIS multidimensional comprehensive evaluation mechanism constructed, the quantitative analysis of takeover quality in the process of takeover is realized;Introduce hierarchical Bayesian inference method to construct Bayesian-based optimal takeover time inference strategy, by fusing prior knowledge and experimental observation data in the field, reliable inference of optimal takeover time is realized under small sample condition, and the uncertainty quantization interval is given.The present application solves the technical problems that the existing method cannot effectively realize the objective quantification of takeover quality and the scientific inference of optimal time.

Description

Technical Field

[0001] This invention relates to the field of intelligent ship technology, and in particular to a method for quantifying the quality of personnel takeover and inferring the optimal takeover timing in intelligent ships. Background Technology

[0002] With the rapid development of artificial intelligence, the Internet of Things and automation technologies, intelligent ships (Maritime Autonomous Surface Ships, MASS) have become a key way to promote the digital transformation and intelligent upgrading of the shipping industry. Intelligent ships aim to achieve less manned or even unmanned operation of ships through technological innovation, thereby improving shipping efficiency, reducing operating costs and effectively reducing maritime accidents caused by human error[1]. However, the current level of technology and the high complexity of the marine environment restrict the ability of intelligent ships to achieve fully autonomous navigation throughout the entire voyage. In response to this real challenge, the International Maritime Organization (IMO) proposed the concept of "Modes of Operation" (Mo O) in the draft MASS Code, replacing the previous static classification method based on the level of autonomy with dynamic and scenario-based modes. The draft divides the operating modes into Autonomous Mode (AM) and Remote Mode (RM), and particularly emphasizes the necessity of onboard personnel or remote operators to intervene in the operation as a safety redundancy mechanism[2]. Section 8.7 of Chapter 8 "Operational Context" of the MASS Code draft clearly states regarding Operator Control Modes (OCM): "The operational capabilities of autonomous surface vessels at sea are limited without any human intervention. As a last resort, when the need is anticipated, personnel should be able to exercise control over all functions directly or by supervision." This provision establishes the core position of personnel takeover in the safe operation system of intelligent ships. The clause further emphasizes: "To ensure the effective assumption of responsibility, remote or onboard personnel should have at least sufficient time to acquire traffic situational awareness and gradually gain control of the ship." [3] This requirement directly raises two key scientific issues in the field of human-machine collaboration in intelligent ships.

[0003] The first issue is how to scientifically quantify the quality of personnel takeover. Quantifying takeover quality is not only a key indicator for measuring the effectiveness of human-machine collaboration, but also an important basis for crew training and certification. However, takeover quality is a complex concept involving multiple dimensions such as collision risk, rule compliance, and the timing of takeover. Its quantification faces many challenges, requiring consideration of both safety at the outcome level and compliance and rationality at the process level.

[0004] The second question is when is it most appropriate to initiate a takeover request. The choice of takeover timing is essentially a multi-objective trade-off problem. Taking over too early will lead to a decrease in system trust and “takeover fatigue”; taking over too late will cause the operator to respond hastily with insufficient situational awareness, significantly increasing the risk of collision. The optimal takeover timing should ensure that the operator has sufficient time budget to complete the four stages of situational awareness, risk assessment, collision avoidance decision-making and ship handling, and ultimately achieve a high-quality takeover [4].

[0005] In response to the above problems, scholars at home and abroad have carried out a series of exploratory studies. It should be noted that as an emerging research field, the research on the takeover of personnel on intelligent ships is still in its infancy and the number of existing literatures is relatively limited. The relevant research has drawn on the relatively mature takeover theories and methods in the field of autonomous vehicles, but there are essential differences between the two fields in terms of operating environment, time scale and risk characteristics. Autonomous vehicles face a highly dynamic and rapidly changing land traffic environment, and the takeover response time is usually measured in seconds; while intelligent ships face a relatively low-speed and high-inertia maritime traffic environment, and the takeover process often lasts for several minutes or even longer, and involves more complex collision avoidance rules (Convention on the International Regulations for Preventing Collisions at Sea, COLREGs) constraints[1]. In addition, the handling characteristics of ships are significantly different from those of automobiles. Directly applying the quantitative indicators and takeover timing inference models in the field of autonomous driving may lead to inapplicable or misleading conclusions. Therefore, it is urgent to develop a method model suitable for the characteristics of intelligent ships[5].

[0006] In terms of quantifying the quality of takeover in autonomous vehicles, existing research mainly adopts three types of methods. The first type is the assessment method based on subjective questionnaires. These studies use psychological measurement tools such as the NASA-TLX workload scale[6], Situation Awareness Rating Technique (SART)[7], and Driver Behavior Questionnaire (DBQ) to quantify the cognitive load, situation awareness level and subjective experience of drivers during takeover[8]. These methods can capture changes in the operator's psychological state, but they are highly dependent on the self-report of the subjects. Different individuals may have significant differences in their perception of the same takeover scenario, making it difficult to objectively quantify the actual operation quality and safety level[9]. The second type is the quantitative method based on a single indicator. These studies evaluate the safety and smoothness of takeover by calculating single physical indicators such as Time to Collision (TC), Distance to Collision (DTC), lateral offset, and braking intensity

[10] . This type of method is simple to calculate and has a clear physical meaning. It has been widely used in the field of autonomous driving. However, its fundamental defect is that it ignores the multidimensional characteristics of takeover quality. That is, successfully avoiding a collision is not the same as high-quality takeover. The takeover personnel may violate traffic rules or cause unstable vehicle control during the operation. The third type is based on multi-index comprehensive analysis. Some studies have attempted to construct a comprehensive analysis system that includes multiple dimensions such as safety, comfort and efficiency

[11] and have used the Analytic Hierarchy Process (AHP) and fuzzy comprehensive evaluation methods for in-depth analysis. However, there are still problems such as strong subjectivity and insufficient theoretical basis in terms of the systematicness of index selection and the objectivity of weight allocation. However, if these model methods from the field of autonomous vehicles are directly applied to intelligent ships, they will face significant adaptability problems. In addition, the complexity of ship encounter scenarios is fundamentally different from typical scenarios such as car overtaking and lane changing. Existing studies lack in-depth analysis of the correlation between encounter scenario characteristics and takeover quality.

[0007] In terms of inferring the timing of takeover in autonomous vehicles, existing research mainly falls into two categories. The first category is inference methods based on physiological cognitive models. These studies, from the perspective of human factors engineering, estimate the minimum time budget required to complete a safe takeover by establishing cognitive stage models such as the driver's out-of-the-loop (OOTL), situational awareness recovery time, and motion reaction time

[12] . For example, some studies decompose the takeover process into four stages: perceiving alarms, understanding the situation, decision planning, and executing operations, and estimate the time required for each stage based on cognitive psychology experimental data, thereby determining the lead time for the takeover request. However, these methods are often based on idealized assumptions and are difficult to fully consider the complexity and uncertainty of the real environment

[13] . The second category is inference methods based on data-driven approaches. These studies use machine learning algorithms to learn the correlation between takeover timing and takeover results from driving simulators or road test data, attempting to predict whether the driver can successfully take over within a specific time in a given scenario

[14] . These methods can capture complex nonlinear relationships from real data, but the interpretability of the models is often insufficient. For the field of intelligent ships, both of these methods face significant challenges. On the one hand, the timescale of ship takeover is much longer than that of automobile takeover, and the cognitive model needs to be recalibrated; on the other hand, the high cost and small sample size of experimental data acquisition in the maritime field restrict the application of data-driven methods. More critically, whether it is the cognitive model method or the data-driven method, existing studies focus on the binary classification problem of "whether the operator can successfully complete the takeover within a given time", while ignoring the more essential quantitative relationship of "what differences exist in the takeover quality corresponding to different takeover times"

[15] . This research perspective makes it impossible to answer the core question of "when is the best time to take over", and also makes it impossible to give an optimal timing inference that takes into account both safety and reliability.

[0008] Furthermore, the issues of quantifying takeover quality and inferring takeover timing are often treated separately in existing research, lacking a unified theoretical framework and methodological system. Research on quantifying takeover quality focuses on post-takeover analysis, evaluating its quality after completion; research on inferring takeover timing focuses on pre-takeover planning, determining when to initiate a takeover request before it occurs. The inherent connection between these two types of research has not been fully explored. However, logically, the selection of takeover timing should aim to ensure high-quality takeover, and the takeover timing quantification model must be based on quantifying the differences in takeover quality under different timings. In summary, how to construct an integrated theoretical framework and methodological system that adapts to the characteristics of intelligent ship operation and achieves both objective quantification of takeover quality and scientific inference of optimal timing is a current focus of human-machine collaboration research in intelligent ships. Summary of the Invention

[0009] This invention provides a method for quantifying the quality of personnel takeover and inferring the optimal takeover timing on intelligent ships, in order to overcome the aforementioned technical problems.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows:

[0011] A method for quantifying the quality of personnel takeover and inferring the optimal takeover timing in intelligent ships includes the following steps:

[0012] S1: Acquire AIS data for intelligent ships;

[0013] Furthermore, the intelligent ship AIS data includes at least ship speed and ship dimensions data;

[0014] S2: Define the security boundaries of SD in the ship domain based on intelligent ship AIS data;

[0015] S3: Preset a ship collision scenario and obtain the basic parameters between the ship and the target ship under the corresponding ship collision scenario. The basic parameters include the nearest encounter distance and the nearest encounter time. By introducing an encounter situation factor, the collision risk value between ships is obtained based on the basic parameters and the safety boundary.

[0016] S4: Construct a safety quantification model for evaluating the safety margin of intelligent ship takeover operations, and use the output of the safety quantification model as a safety quantification evaluation index.

[0017] S5: Use safety quantitative evaluation indicators and ship-to-ship collision risk values ​​as multi-dimensional quantitative indicators, and construct a CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism based on multi-dimensional quantitative indicators to obtain comprehensive takeover quality indicators.

[0018] S6: Construct a Bayesian-based optimal takeover timing inference strategy, obtain the inference result of the optimal takeover timing based on the comprehensive takeover quality index, and then provide quantitative decision-making for the dynamic switching of intelligent ship operation mode based on the inference result.

[0019] Beneficial Effects: This invention provides a method for quantifying the quality of takeover operations and inferring the optimal takeover timing for intelligent ships. It constructs a safety quantification model to evaluate the safety margin of takeover operations on intelligent ships, using the model's output as a safety quantification evaluation index. It also integrates inter-ship collision risk values ​​as multi-dimensional quantification indicators and obtains a comprehensive takeover quality index through a constructed CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism, thus achieving quantitative analysis of takeover quality during the takeover process. Addressing the scarcity of intelligent ship takeover data, this invention introduces a hierarchical Bayesian inference method to construct a Bayesian-based optimal takeover timing inference strategy. By fusing domain prior knowledge with experimental observation data, it achieves reliable inference of the optimal takeover timing under small sample conditions and provides a quantitative range of uncertainty. This invention combines takeover quality quantification with optimal timing inference, overcoming the limitations of the separation between the two in traditional research. It can be directly applied to the design of intelligent ship human-computer interaction systems, the development of takeover decision-making systems, and the construction of crew training systems, providing a theoretical foundation and technical support for promoting the safe and standardized development of intelligent ship technology. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the intelligent ship personnel takeover quality quantification and optimal takeover timing inference method of the present invention;

[0022] Figure 2 This is a schematic diagram illustrating the parameters for calculating the collision hazard index in this embodiment;

[0023] Figure 3 This is a schematic diagram illustrating the encounters at different stages in this embodiment;

[0024] Figure 4 This is a schematic diagram of the scene setup in this embodiment;

[0025] Figure 5 This is a schematic diagram illustrating the weight settings for different indicators in this embodiment;

[0026] Figure 6 This is a correlation coefficient matrix diagram of the evaluation indicators in this embodiment;

[0027] Figure 7 This is a schematic diagram of the OS trajectory after different subjects were taken over in this embodiment;

[0028] Figure 8This is a scatter plot of the takeover time and comprehensive quality index values ​​in this embodiment;

[0029] Figure 9 This is a curve showing the optimal takeover timing in the Bayesian inference in this embodiment. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0031] This embodiment provides a method for quantifying the quality of personnel takeover and inferring the optimal takeover timing on intelligent ships. The method extracts ship trajectory and attribute data, combines them with features such as dynamic ship domain, ship collision risk index, and collision avoidance rule compliance, to quantify the takeover characteristics of the personnel. Then, the CRITIC method is used for objective weighting, and the ideal solution distance is calculated using the TOPSIS method to obtain the final comprehensive quality index value, evaluating the takeover quality of the personnel within the takeover time. Finally, a hierarchical Bayesian model is used to achieve point and interval estimation of the optimal takeover timing, providing quantitative decision support for the timing decision-making of personnel takeover on intelligent ships. Figure 1 As shown, the specific steps include:

[0032] S1: Acquire AIS data of the intelligent ship; and the intelligent ship AIS data includes at least ship speed, ship size data and ship attitude data;

[0033] S2: Define the security boundaries of SD in the ship domain based on intelligent ship AIS data;

[0034] Specifically, in this embodiment, collision risk assessment is the core dimension of takeover quality evaluation. Accurately quantifying the real-time collision risk during the ship takeover process is crucial for judging the safety of the takeover operation. This embodiment first introduces the definitions of the basic parameters required for collision risk calculation, and then explains the process of obtaining the improved collision risk index proposed in this embodiment. In ship collision risk calculation, a series of basic risk parameters need to be clearly defined, such as... Figure 2 The text identifies ship collision hazard indicators, which are also used in the calculations below. Taking a collision scenario between the Own Ship (OS) and the Target Ship (TS) as an example, both ships are in the world coordinate system. middle, , These represent the relative positions of TS and OS, respectively. , These represent the true orientations of TS relative to OS and OS relative to TS, respectively. The relative distance between OS and TS. This represents the relative motion vector of TS with respect to OS. This indicates the direction corresponding to the vector. Based on the above basic parameters, the DCPA and TCPA between the two ships can be obtained, as shown in equations (1) to (2):

[0035] (1)

[0036] (2)

[0037] In the formula: This indicates that you will encounter a distance in the near future; Indicates that the time will be met soon; Indicates the true bearing of the target ship TS relative to the ship's OS; This indicates the relative distance between the ship's OS and the target ship's TS; This represents the relative motion vector of the target vessel TS with respect to the vessel OS. Represents the relative motion vector The corresponding direction angle;

[0038] There are currently many methods to quantify the risk of collision between ships, such as the DCPA and TCPA parameters commonly used at sea. These parameters can calculate the relative distance between ships to obtain the collision risk between them. However, this calculation method is based on the assumption that ships maintain their course and speed, that is, it is assumed that TS and OS maintain their course and speed. Although this premise simplifies the collision risk calculation model, it ignores the most common speed change and course change behaviors at sea, which affects the accuracy and timeliness of ship collision risk. Some scholars have proposed an improved method that integrates DCPA and TCPA

[18] , which calculates the collision risk (CR) by assigning different weights to the two parameters, as shown in equation (3):

[0039] (3)

[0040] However, this calculation method can easily lead to an exponential increase in CR when TS approaches OS, resulting in oversensitivity. To address this issue, Mou et al. proposed an improved method that considers DCPA, TCPA, and the relative distance between ships in the CR calculation, and sets different weighting coefficients according to different encounter scenarios

[19] , as shown in Equation (4):

[0041] (4)

[0042] In the formula: This represents the coefficient of the relative heading angle between OS and TS, and different coefficients are used depending on the different encounter scenarios.

[0043] However, the quantitative calculation of CR based on mathematical models does not take into account the characteristics of the ship itself. Different ships require different avoidance spaces, so it does not have a certain universality. The concept of Ship Domain (SD) introduced by Fujii and Tanaka provides another perspective for calculating ship collision risk. SD defines the navigation safety domain that does not allow other ships to enter the ship from a geometric perspective. Different ships can set different SD sizes according to their characteristics

[20] . Compared with DCPA and TCPA, SD realizes the visualization of ship collision risk in the spatial dimension, but it only provides information on the relative safe distance between ships and cannot determine the specific time of the collision. Therefore, in the complex maritime traffic environment, it means that SD can only be used as a reference and cannot be used to make decisions based on the information it provides. To make up for the shortcomings of a single method, Jisang et al. proposed a CR calculation based on the ship domain. They used the ship 3-DOF equation to obtain the scale of each boundary of SD through regression analysis

[21] and combined it with DCPA and TCPA to calculate the CR index, as shown in equations (5) to (7).

[0044] (5)

[0045] (6)

[0046] (7)

[0047] This method effectively integrates the advantages of DCPA, TCPA, and SD, and can better reflect the dynamic changes in collision risk between ships. However, this method does not consider the influence of ship size and speed on the CR value when constructing the ship domain, and only relies on static parameters for calculation. In fact, the SD size should show significant differences at different speeds. Ships need a larger safety distance when sailing at high speeds, while it can be appropriately reduced when sailing at low speeds. This static approach cannot fully reflect the influence of ship motion characteristics on collision risk. To address the above shortcomings, this embodiment proposes an improved CR calculation method that integrates dynamic ship domain, TCPA, and DCPA. By parameterizing SD as a function of speed and ship length, the risk assessment standard can be adaptively adjusted. The dynamic ship domain adopts the potential collision risk ship domain proposed by Zou et al., and the shape of the ship domain is defined as an eccentric ellipse. The calculation of its safety boundary is shown in Equation (8):

[0048] (8)

[0049] In the formula: This represents the forward boundary of the SD domain in the shipbuilding field; Represents the backward boundary of the SD domain in the shipbuilding field; Represents any point in the field of shipping. p The angle between the line connecting the ship's center and the bow direction; Indicates the length of the ship; Parameters representing the lateral influence of SD in the shipbuilding sector; Parameters representing the longitudinal impact of SD in the shipbuilding sector; Indicates the potential collision risk index; express The ship's speed at any given moment; This indicates the boundary of the SD (Shipbuilding Scope) in the shipbuilding field.

[0050] S3: Preset a ship collision scenario and obtain the basic parameters between the ship and the target ship under the corresponding ship collision scenario. The basic parameters include the nearest encounter distance and the nearest encounter time. By introducing an encounter situation factor, the collision risk value between ships is obtained based on the basic parameters and the safety boundary.

[0051] The specific steps include:

[0052] S31: Under the preset world coordinate system, preset the ship collision scenario and obtain the basic parameters between the ship and the target ship under the corresponding ship collision scenario. The basic parameters include the nearest encounter distance and the nearest encounter time, and their expressions are Equations (1) to (2).

[0053] S32: Obtain the safety margin of the domain boundary between the target ship TS and the ship's OS based on the nearest encounter distance.

[0054] The expression for the security margin of the domain boundary is:

[0055] (9)

[0056] In the formula: Indicates the security margin of the domain boundary; This represents the radius of the target vessel's domain in a specified direction; based on the realization of dynamic vessel domain, this embodiment proposes an adjusted DCPA calculation method, as shown in equation (9); where This represents the radius of the target ship's domain in a specific direction. This applies when TS is located inside or on the boundary of the OS domain. A value of 0 indicates that the area has been entered; otherwise... A value greater than 0 indicates a safety margin at the boundary of the OS (Ship Safety) domain. This adjustment shifts risk assessment from the traditional ship geometry center to the ship's safety domain boundary, better aligning with the ship's actual collision avoidance needs.

[0057] S33: The adaptive adjustment dynamic coefficients for the ship domain are obtained based on the nearest encounter distance and the nearest encounter time. The expression for this coefficient is:

[0058] (10)

[0059] In the formula: This indicates the collision risk value at the SD boundary in the marine field; , They represent The corresponding nearest encounter distance With the most recent meeting time ; This represents the adaptive adjustment dynamic coefficient in the shipbuilding field. To achieve an organic integration of the dynamic shipbuilding field and the CPA method, this embodiment improves the CR calculation framework of Jisang et al., proposes an adaptive adjustment method for dynamic coefficients, and sets a standard risk value based on the derivation logic of equation (6) and the parameter values ​​in that framework. The standard distance is 0.3, which is twice the radius of the ship's territory. ,parameter It can be calculated from equation (10), this improvement makes the coefficient It is dynamically linked to the shipbuilding industry, enabling adaptive and dynamic adjustment of risk assessment standards under different ship speeds and lengths.

[0060] S34: Based on this, the comprehensive collision risk index is calculated by equations (11) and (12), that is, by introducing the encounter situation factor and adjusting the dynamic coefficient according to the ship domain based on the domain boundary safety margin to obtain the inter-ship collision risk value. for:

[0061] (11)

[0062] (12)

[0063] In the formula: The value is set to the maximum value of the encounter situation factor, 8.5, to ensure that CRI is normalized to the [0,1] interval. express The reciprocal of the maximum value. This method comprehensively considers the influence of DCPA, TCPA, encounter situation, and dynamic ship domain, and can more accurately reflect the evolution of collision risk between ships. This embodiment introduces an encounter situation factor. Risk weights are adjusted according to different encounter types: a value of 1 for a face-to-face encounter, 8.5 for a cross encounter, and 2.34 for an overtaking encounter. These parameter values ​​remain consistent with the original text. Furthermore, to avoid discrepancies in risk calculations between different situations... There are discontinuous jumps; linear interpolation is used for smoothing within the transition interval to ensure the smoothness of the risk index. The OS is calculated time-by-time. CRI In this embodiment, three statistical indicators were extracted for quantifying the quality of takeover: the average CRI reflects the overall risk level throughout the entire trajectory, the maximum CRI captures extreme risk states, and the 95th percentile CRI characterizes a high-risk level after excluding outliers. In addition, average domain risk, average DCPA risk, average TCPA risk, as well as minimum DCPA and minimum TCPA, were calculated to comprehensively evaluate the safety margin of the takeover process from both statistical distribution and critical value perspectives.

[0064] S4: Construct a safety quantification model for evaluating the safety margin of intelligent ship takeover operations, and obtain safety quantification evaluation indicators based on the relative distance between the ship and the target ship and the ship's attitude data according to the safety quantification model;

[0065] Specifically, COLREGs, as traffic regulations for maritime navigation, clearly stipulate the avoidance responsibilities and action requirements of each ship under different encounter situations. Existing research mainly focuses on whether ships violate the rules, while the compliance with the rules during operation is limited and its quantification is limited. Woerner et al. first proposed a comprehensive framework for quantifying COLREGs compliance, and it has been continuously improved in subsequent studies. This method is based on ship track data, considers Rules 8 and Rules 13-17, and quantifies whether the ship's maneuvering behavior in encounter events complies with the rule requirements. Hagen et al. further proposed a more comprehensive COLREGs compliance quantification model based on this, covering the provisions of Rules 8a, 8b and Rules 13-17, and made several improvements to the evaluation algorithm of Woerner et al., which significantly improved the accuracy and robustness of quantification

[17] . This embodiment, based on the research of Hagen et al., addresses the shortcomings of their method, which relies on static relative bearing angles for encounter situation determination and does not fully consider the dynamic characteristics of ship relative motion. It innovatively introduces dynamic ship domain theory to improve the encounter situation identification method and combines DCPA and TCPA information, making the encounter situation classification more consistent with the actual maritime navigation environment. Based on this, this embodiment constructs a complete system including encounter scenario classification, a safety quantification model, and a rule compliance quantification model, realizing a quantitative evaluation of compliance with takeover operation rules. The parameters in the safety quantification model and the rule compliance quantification model are consistent with the values ​​in the research results of Hagen et al. Hagen et al. used the relative bearing angle between OS and TS to determine the encounter situation, but this method did not consider the dynamic changes in ship relative motion. For example, the relative bearing angles of the two ships may be within the encounter range, but if the relative motion direction indicates that the two ships are increasing distance (negative TCPA or large DCPA), it should not be determined as an encounter situation requiring avoidance. This embodiment adds DCPA and TCPA constraints to the relative azimuth angle, and divides the encounter situation into five types: encounter, crossing and yielding, crossing and straight driving, overtaking and being overtaken, as shown in Table 1.

[0066] Table 1. Encounter Situation Classification Rules

[0067]

[0068] Hagen et al. implicitly defined four areas or stages of avoidance responsibility based on the distance between two ships. However, the parameter settings were static and did not take into account the ship's own attributes. Considering this, this paper improves upon them, redefining the four areas or stages of avoidance responsibility by referring to the dynamic ship domain defined in equation (8). When TS is in OS stage 1, it is considered that the COLREGs rules do not apply because the two ships are too far apart. When TS is in stage 2, the ship with the straight-ahead responsibility should maintain its course and speed. However, if the ship with the give-way responsibility does not take appropriate action, the two ships enter stage 3, where the ship with the straight-ahead responsibility can take action to avoid a collision. In addition, if the action of the give-way ship alone is clearly insufficient to avoid a collision, the two ships may enter stage 4, where the ship with the straight-ahead responsibility must take action to avoid a collision. Schematic diagrams of the different stages are shown below. Figure 3 As shown.

[0069] Safety quantification models are used to evaluate the safety margin of takeover operations. Hagen et al. constructed a safety quantification model based on distance and attitude factors. This embodiment introduces the dynamic ship domain on this basis, so that the safety quantification standard is dynamically adjusted with the ship's motion state.

[0070] Specifically, a safety quantification model is constructed to evaluate the safety margin of intelligent ship takeover operations, including a first quantification model for quantifying the relative distance and attitude between the target ship TS and the ship OS, a second quantification model for quantifying whether the yielding ship's avoidance action meets the requirements of the first rule, a third quantification model for quantifying whether the straight-going ship's straight-going action meets the requirements of the second rule, and a fourth quantification model for quantifying whether the navigation of the two ships meets the requirements of the third rule in a confrontation situation.

[0071] The expression for the first quantization model is:

[0072] (13)

[0073] (14)

[0074] (15)

[0075] (16)

[0076] (17)

[0077] In the formula: This represents the distance quantization function; The CPA distance represents the distance between the nearest points on the territory of two ships when they meet at sea. , , , Both indicate that the distance parameter threshold is set based on the potential collision risk indicator and These are the preferred distance, minimum safe distance, near-accident distance, and collision distance, respectively, as follows: Figure 3 As shown; This represents the contact angle quantization function used to reflect the relative headings between the two ships at the CPA moment. This reflects the relative headings between the two ships at the CPA time. This function reflects the relative heading angle of the OS with respect to the TS. The minimum value of 0 indicates the highest risk when the bow of the ship is directly facing TS; This indicates the relative heading angle of the ship's OS with respect to the target ship's TS; Indicates the preset cutoff angle; This represents the relative bearing quantization function used to reflect the relative bearing between the two ships at the CPA moment; This indicates the relative heading angle of the target vessel TS with respect to the vessel's OS; This represents the minimum and maximum relative heading angles of the target vessel TS with respect to the vessel OS. The attitude score is the weighted sum of the contact angle quantization function and the relative orientation quantization function. Indicates the weighting coefficient; This represents the final security score of the first quantization model, which is determined by the distance quantization function. and attitude quantization function Calculated;

[0078] In this embodiment, the safety score of the second quantification model is the score of Rule 16 in the 1972 Convention on the Regulations for Preventing Collisions at Sea (COLREGs). It is used to quantify whether the give-way vessel's avoidance action meets the requirements of the first rule, namely, whether it conforms to the requirement of "taking significant and obvious avoidance action as early as possible." In this embodiment, quantification is performed from three dimensions: timing of maneuver, magnitude of course change, and magnitude of speed change. The expression of the second quantification model is:

[0079] (18)

[0080] (19)

[0081] (20)

[0082] (twenty one)

[0083] (twenty two)

[0084] (twenty three)

[0085] (twenty four)

[0086] In the formula: This represents the collision avoidance delay penalty function, through which the timing of the ship's maneuvers is determined. This function is implemented based on the relative distance between the ship and its own vessel when the ship begins maneuvering. The penalty value is calculated based on the relationship between the different domain boundaries in Equation (13). The penalty for action is 0 when the ship takes action outside stage 3. The penalty is greater when the two ships start maneuvering closer to the CPA. This represents the distance to the ship's domain boundary as set according to the potential collision risk index, i.e. The distance in stage2 as specified in equation (13); This is the boundary distance of stage3; Indicates the relative distance between the target vessel and this vessel; This indicates the maximum change in course of a vessel during a maneuver; the course maneuver amplitude is defined as the maximum change in course of a vessel during a maneuver. This refers to the maximum course change from the time the ship enters OS stage 2 to the time it encounters another ship at sea, CPA (Closest Point of Approach). The penalty function representing heading maneuver; Indicates the minimum detectable change in heading; Indicates the heading change threshold; This indicates the amount of deceleration of the ship's relative speed; Indicates the moment of entering the pre-set warning zone; express Ship speed at any given time; express Ship speed at any given time; The penalty function representing the ship's maneuvering; A threshold parameter representing the amount of deceleration of a ship's relative speed; Indicates the score for vessel maneuvering; Indicates the weighting coefficient; This represents the safety score used to quantify whether the give-way vessel's avoidance actions meet the requirement of taking early, significant, and obvious avoidance actions; This indicates the time when TS enters OS stage 2 and the time when it enters CPA; express The ship's course at any given time; express The ship's course at any given time;

[0087] In this embodiment, the safety score of the third quantification model is the score of Rule 17 in the 1972 Convention on the Regulations for Preventing Collisions at Sea (COLREGs). The Rule 17 score is used to quantify whether a straight-running vessel meets the requirements of the second rule, namely, whether it complies with the requirement to "maintain course and speed, take action only when necessary, and not turn to port." The method for obtaining the third quantification model is as follows:

[0088] S100: Based on empirical values ​​and the safety boundary of the SD (Safety Safe Zone) in the maritime domain, four encounter zone stages {stage1, stage2, stage3, stage4} are defined according to the distance between the target vessel and the vessel to avoid collision.

[0089] Stage 1: The target vessel is more than the maximum distance threshold of the target vessel, which is the encounter zone stage that is not considered; Stage 2: The target vessel is more than the first distance threshold of the target vessel, which is the encounter zone stage; Stage 3: The target vessel is more than the second distance threshold of the target vessel, which is the encounter zone stage; Stage 4: The target vessel is more than the third distance threshold of the target vessel, which is the encounter zone stage.

[0090] And the first distance region > the second distance region > the third distance region;

[0091] When the target ship and the current ship are in Stage 2 or Stage 3, the corresponding stage penalty applies. for:

[0092] (25)

[0093] (26)

[0094] (27)

[0095] (28)

[0096] In the formula: This represents the penalty function for heading changes within Stage 2 or Stage 3; This indicates the amount of deceleration of the ship's relative speed; This indicates the rate of increase in the relative speed of a ship; This represents a penalty function for increasing speed; This represents the penalty function for reduced speed; Indicates the weighting coefficient;

[0097] S101: When a vessel traveling in a straight course has a responsibility to give way to other vessels, a yielding compensation factor is introduced. Optimization phase penalty This is combined with the defined left-turn penalty function when the target ship and the current ship are in Stage 4. The third quantization model is constructed as follows:

[0098] (29)

[0099] (30)

[0100] (31)

[0101] In the formula: This indicates the optimized stage penalty; This represents the safety score used to quantify whether a straight-going vessel complies with maintaining its course and speed, taking action only when necessary, and not turning left. In this embodiment, based on Rule 17(c) of the COLREGs Rules, which states that "when the give-way vessel is on the port side and the straight-going vessel has to take action, it should not turn left," the left-turn penalty function is defined as follows: ;

[0102] In this embodiment, the safety score of the fourth quantification model is the score of Rule 14 in the 1972 Convention on the Regulations for Preventing Collisions at Sea (COLREGs). The Rule 14 score is used to quantify whether the navigation of two ships in an encounter situation meets the requirement of Rule 3, namely whether they "both turn to starboard and pass each other on the port side". In this embodiment, quantification is performed from two dimensions: the starboard turn and the passing method. The expression of the fourth quantification model is:

[0103] (32)

[0104] (33)

[0105] (34)

[0106] (35)

[0107] In the formula: This represents the maximum right turn angle between the target vessel and the vessel itself, i.e., the right turn amount. This is the maximum right turn angle after entering the restricted area; Indicates the penalty function for non-right turns; This represents the minimum effective right turn angle; Indicates the heading change threshold; This function represents the penalty function when the target ship and this ship pass each other on the starboard side. A penalty is applied when both ships pass each other on the starboard side, such as when... and At this time This indicates that the two ships are passing each other on their port side. and At this time A value close to 1 indicates that both vessels passed through each other's starboard sides, constituting a serious violation; Indicates the weighting coefficient; This represents a comprehensive score used to quantify whether both ships turn to starboard and pass each other's port side in a head-on encounter situation. Indicates that the ship is t The course of time; Indicates the ship's course when entering the pre-set warning zone.

[0108] S5: Using safety quantitative evaluation indicators and inter-ship collision risk values ​​as multi-dimensional quantitative indicators, a CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism is constructed based on these indicators to obtain comprehensive takeover quality indicators. After establishing a multi-dimensional quantitative indicator system, the key issue in quantifying takeover quality is how to scientifically and rationally determine the weights of each indicator and conduct a comprehensive evaluation. Traditional subjective weighting methods rely on expert experience, and the judgments of different experts may differ significantly. Furthermore, the consistency verification process is cumbersome and difficult to guarantee objectivity and repeatability. To overcome this limitation, this embodiment constructs the CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism to achieve objective weighting based on data characteristics and multi-attribute comprehensive evaluation. The core idea of ​​the CRITIC method is to simultaneously consider the comparative strength and conflict between indicators. Comparative strength reflects the degree of variation of an indicator among different samples; the greater the variation, the greater the contribution of the indicator to distinguishing different samples. Conflict reflects the correlation between indicators; the lower the correlation with other indicators, the more independent information the indicator contains. The CRITIC method combines comparative strength and conflict to achieve objective determination of weights, specifically including the following steps:

[0109] S51: Using AIS data from intelligent ships as sample data, and combining safety quantitative evaluation indicators and inter-ship collision risk values ​​as multi-dimensional quantitative indicators, a decision matrix is ​​obtained. and ,in n Indicates the number of sample data; m This indicates the number of evaluation indicators for multi-dimensional quantitative indicators; Represents the elements in the decision matrix;

[0110] S52: The CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism is constructed based on multi-dimensional quantitative indicators as follows:

[0111] S521: Use Z-score standardization to eliminate the influence of dimensions in the decision matrix and obtain the optimal decision matrix. and ,and

[0112] (36)

[0113] In the formula: express The result after Z-score standardization; Indicates the first j The mean of each evaluation indicator ; Indicates the first j Standard deviation of each evaluation indicator , used to characterize the comparative strength of evaluation indicators among different samples;

[0114] S522: Obtain the optimal decision matrix Furthermore, the Pearson correlation coefficient matrix among the various evaluation indicators;

[0115] Specifically, no. j The comparative strength of each evaluation indicator is expressed by its standard deviation. This indicates the degree of variability of the indicator across different samples; a larger value indicates that the indicator contributes most to distinguishing different samples. The conflict between indicators is measured using a correlation coefficient matrix. j The first evaluation indicator and the first k Pearson correlation coefficient among the evaluation indicators for:

[0116] (37)

[0117] S523: Obtain conflict indicators for evaluation metrics based on the Pearson correlation coefficient matrix;

[0118] The formula for obtaining the conflict index is as follows:

[0119] (38)

[0120] In the formula: Indicates the first j Conflicting indicators among evaluation metrics; The Pearson correlation coefficient matrix represents the first... j The first evaluation indicator and the first k The Pearson correlation coefficient between the evaluation indicators; in this embodiment, the... j The conflict of an indicator is defined as shown in Equation (38). The higher the conflict, the less overlap there is between the information contained in the indicator and other indicators, and therefore it should be given a higher weight.

[0121] S524: Based on the conflict index and contrast strength, the CRITIC weight is obtained as follows:

[0122] (39)

[0123] (40)

[0124] In the formula: Indicates the first j The product of the conflict index and the contrast strength of each evaluation indicator; Indicates CRITIC weight; Indicates the first k The product of the conflict index and the contrast strength of each evaluation indicator;

[0125] In this embodiment, the TOPSIS method ranks the indicators by calculating their relative proximity to the ideal solution. The optimal solution should be close to the positive ideal solution (each indicator takes its optimal value) and far from the negative ideal solution (each indicator takes its worst value). By combining the weights obtained from the CRITIC method with the TOPSIS method, a comprehensive evaluation based on objective weighting is achieved, specifically including:

[0126] S53: Decision Matrix Vector normalization is performed to obtain the normalized decision matrix; the decision matrix is ​​then processed. The expression for vector normalization is:

[0127] (41)

[0128] In the formula: This represents an element in the normalized decision matrix;

[0129] S54: Obtain the weighted normalized value based on the normalized decision matrix and CRITIC weights. for:

[0130] (42)

[0131] S55: Obtain the corresponding weighted normalized value positive ideal solution With negative ideal solution That is, to find the ideal solution. and negative ideal solution Each evaluation index is composed of its optimal and worst values:

[0132] (43)

[0133] (44)

[0134] S56: Based on the ideal solution With negative ideal solution Combined with weighted normalized values Calculate the sum of each evaluation index and the positive ideal solution. With negative ideal solution European distance for:

[0135] (45)

[0136] (46)

[0137] S57: Based on Euclidean distance Get the i The comprehensive takeover quality index of each evaluation indicator for:

[0138] (47)

[0139] In the formula A higher value indicates a higher quality of the pipe fitting.

[0140] The takeover quality indicator system constructed in this embodiment is shown in Table 2, which includes 12 indicators across five dimensions: collision risk, domain risk, collision avoidance safety, COLREGs compliance, and encounter statistics. The indicators are divided into two categories: benefit-based and cost-based. For benefit-based indicators (such as minimum DCPA, minimum TCPA, and compliance score), higher values ​​indicate better takeover quality; for cost-based indicators (such as average CRI, maximum CRI, and number of violations), lower values ​​indicate better takeover quality.

[0141] Table 2. Quality Indicator System for Acceptance

[0142]

[0143] S6: Construct a Bayesian-based optimal takeover timing inference strategy. Based on comprehensive takeover quality indicators, obtain the inference result for the optimal takeover timing, and then provide quantitative decision-making support for the dynamic switching of intelligent ship operation modes based on the inference result. In this embodiment, the selection of the timing for manual takeover of intelligent ships is a key decision problem. Taking over too early will reduce autonomous navigation efficiency, increase operator workload, and may lead to "takeover fatigue" and decreased system trust in the long run; taking over too late will force operators to respond hastily with insufficient situational awareness, significantly increasing the risk of collision. The optimal takeover timing should ensure that operators have sufficient time budget to complete situational awareness, risk assessment, collision avoidance decision-making, and ship control, and ultimately achieve a high-quality takeover. This embodiment aims to infer the optimal takeover timing and its uncertainty range based on limited takeover experimental data, providing quantitative decision support for the dynamic switching of intelligent ship operation modes.

[0144] Specifically, the Bayesian-based optimal takeover timing inference strategy is as follows:

[0145] S61: Settings nThe system records the takeover scenarios for several intelligent ships, including the actual takeover time for each scenario. (seconds, i.e., TOT time), comprehensive takeover quality indicators And the recorded data corresponding to the indicator system in Table 2; the objective of this embodiment is to infer the optimal takeover timing. posterior distribution ;

[0146] S62: This embodiment employs a three-step inference strategy:

[0147] The first step involves using correlation analysis and grouped statistical analysis to analyze the relationship between takeover time and takeover quality, and then calculating the Pearson correlation coefficient between the two. To preliminarily determine the linear correlation between the two:

[0148] (48)

[0149] The second step is to establish a quadratic regression model to characterize the nonlinear relationship between takeover time and quality: Assuming that there is an optimal point for takeover quality, i.e., quality first increases and then decreases over time, exhibiting a downward-opening parabolic relationship, a regression model is established based on this assumption. Specifically, a quadratic regression model based on linear correlation is established to characterize the nonlinear relationship between takeover time and the comprehensive takeover quality index:

[0150] (49)

[0151] In the formula: This represents the output of the quadratic regression model; Indicate design parameters; Indicates the regression parameters of the model; Indicates the actual takeover time Standardized time, ; This is random error;

[0152] S63: The regression parameters of the quadratic regression model are estimated using the least squares method as follows:

[0153] (50)

[0154] In the formula: This represents the estimated values ​​of the model's regression parameters; Represents the design matrix and ;

[0155] S64: If This indicates that the connection quality and connection time follow a downward-opening parabola, suggesting the existence of an optimal connection time. That is, to obtain the optimal takeover time based on the estimated values ​​of the model regression parameters. for:

[0156] (51)

[0157] S65: Based on optimized takeover time Obtain the optimal takeover time Soon The optimal tube-unpacking time is obtained after converting to the original scale. for:

[0158] (52)

[0159] In the formula: Indicates the actual takeover time The mean; Indicates the actual takeover time The variance;

[0160] However, due to the optimal takeover timing based on point estimation Without considering the uncertainty of parameter estimation and the limitations of limited samples, the third step is to construct a hierarchical Bayesian model to quantify the uncertainty of the optimal timing. This embodiment assumes that the optimal takeover timing follows a normal distribution and constructs a hierarchical Bayesian model, where the prior layer is set as follows:

[0161] (53)

[0162] In the formula: This represents the prior mean, reflecting the domain experts' prior knowledge of the optimal timing; It represents the prior standard deviation, reflecting the degree of prior uncertainty;

[0163] S66: Based on the optimal takeover time Construct a hierarchical Bayesian model to obtain the data mean of optimal timing uncertainty. That is, the likelihood layer is estimated by the regression model:

[0164] (54)

[0165] In the formula: An estimate representing the uncertainty of the optimal timing; This represents the optimal point estimated by the regression model; This indicates the uncertainty of the estimate;

[0166] In this embodiment, weak information prior is used based on expert knowledge, and the parameter settings are shown in Table 3:

[0167] Table 3. Weak Information Prior Parameter Setting

[0168]

[0169] In this embodiment, the prior weights are defined as follows:

[0170] (55)

[0171] Based on the estimated values ​​of the model regression parameters The uncertainty of the optimal timing is estimated by combining the error propagation formula with the standard error of the model regression parameters. for:

[0172] (56)

[0173] (57)

[0174] In the formula: Indicates intermediate parameters;

[0175] S67: Estimation of uncertainty based on optimal timing Obtain the posterior precision value Then, using the conjugate property of the normal distribution, the variance is converted into precision based on the posterior distribution analysis to obtain the prior precision, data precision, and posterior precision:

[0176] (58)

[0177] (59)

[0178] (60)

[0179] In the formula: These represent the prior precision and data precision, respectively, indicating the uncertainty of the optimal timing. Indicates the prior standard deviation;

[0180] S68: Data mean based on S67 combined with the uncertainty of optimal timing Obtain the posterior mean That is, the posterior mean is the precision-weighted average of the prior mean and the data mean:

[0181] (61)

[0182] In the formula: This represents the prior mean;

[0183] S69: Based on the posterior precision value Obtain the posterior standard deviation for:

[0184] (62)

[0185] (63)

[0186] S70: Based on the posterior standard deviation with posterior mean Obtain the Bayesian confidence interval for the optimal takeover timing. for:

[0187] (64)

[0188] The Bayesian confidence interval is the inference result of the optimal takeover timing. In this embodiment, the 95% Bayesian confidence interval is set as shown in equation (64), which means that the true value of the optimal takeover timing has a 95% probability of falling within this range.

[0189] In this embodiment, to evaluate the reliability of Bayesian inference, including shrinkage... and data impact The calculation of the two diagnostic indicators is shown in equations (65) and (66) respectively:

[0190] (65)

[0191] In the formula: The closer a value is to 0, the closer the posterior data is to 1, indicating that the prior data is closer to 1. Therefore, ideally, This indicates that both prior and posterior data contribute appropriately.

[0192] (66)

[0193] Using the hierarchical Bayesian inference framework described above, this embodiment not only provides a point estimate of the optimal takeover timing but also an interval estimate (95% confidence interval) considering parameter uncertainties, offering quantitative support for intelligent ship takeover timing decisions that combines accuracy and robustness. Furthermore, this method fully utilizes domain prior knowledge, providing reliable inference results under limited experimental data conditions, overcoming the limitations of purely data-driven methods under small sample conditions.

[0194] To verify the effectiveness of the method described in this embodiment, a manual takeover experiment based on a ship maneuvering simulator was designed. The core objective of the experiment was to obtain realistic takeover process data in a controlled experimental environment, including the operator's takeover reaction time, maneuvering decision trajectory, and final collision avoidance effect, providing empirical data support for quantifying takeover quality and inferring optimal timing. The experimental scenario was selected as a multi-ship encounter situation in open water. This choice was mainly based on the following considerations: open water is the main application scenario for autonomous navigation of intelligent ships, and studying the manual takeover problem in this scenario has strong practical significance and application value.

[0195] The specific parameters of the experimental scenario were carefully designed to ensure the repeatability of the experiment. The OS was set as a 259-meter container ship sailing from north to south in the Bohai Sea at an initial speed of 10 knots. The scenario included 11 TS (Transport Ships) creating a complex traffic situation with various encounter types. The OS's initial position and course were calculated so that, without evasive action, the nearest collision distance (CPA) was between 0.5 and 1.5 nautical miles, and the nearest trip distance (TCPA) was between 8 and 15 minutes. This ensured a certain risk of collision to trigger takeover requirements while also allowing the participants some room for maneuver. Environmental conditions were set to good visibility and sea state 2 to 3 to eliminate the influence of adverse weather on the experimental results. A schematic diagram of the scenario setup is shown below. Figure 4 As shown. The triggering conditions for the intelligent navigation system takeover request are set as follows: when the TCPA of any TS is less than 10 minutes or the DCPA is less than 3 nautical miles, a TOR is triggered, indicating a high risk of collision, requiring manual intervention. After the TOR is triggered, the subject is notified through visual and auditory alarms, and then manual control of the ship to handle the encounter situation begins. The setting of this triggering condition refers to the commonly used warning distance and time threshold in maritime collision avoidance practice to ensure the authenticity of the experimental scenario. In this embodiment, subjects were recruited through Dalian Maritime University. The recruitment criteria required subjects to hold a valid seafarer's certificate of competency, have at least one year of actual shipboard navigation experience, be in good health, and have no health problems such as color blindness or color weakness that may affect the experiment. In the end, a total of 18 subjects completed the entire experimental process. All of them were male, aged between 25 and 48 years old, with an average age of 34.2 years. The detailed information of the subjects is shown in Table 4.

[0196] Table 4. Basic Information of Participants

[0197]

[0198] Table 4 shows that the 18 participants covered the complete officer rank sequence from third mate to captain, including 4 captains, 4 chief mates, 6 second mates, and 4 third mates, demonstrating a relatively balanced rank distribution. The participants' seafaring experience ranged from 3 to 20 years, with an average of 9.5 years. This sample composition is highly representative, including both experienced senior officers and relatively young junior officers, reflecting the impact of different experience levels on leadership performance. The experiment was conducted at the Ship Handling Simulator Center of Dalian Maritime University. This simulator features a 180-degree field of view system, and the control panel realistically replicates the layout of a ship's bridge at a 1:1 scale, equipped with a full set of nautical equipment, including vehicles, steering wheels, electronic chart displays and information systems, and radar. The simulator uses a six-degree-of-freedom ship motion mathematical model, with a system response delay of less than 50 milliseconds, effectively reproducing the ship's handling characteristics. The data acquisition system recorded parameters such as ship position, heading, and speed at a frequency of 1Hz, providing complete trajectory data for subsequent analysis. In the pre-experiment preparation phase, participants were first given a detailed explanation of the experiment's purpose, procedures, and precautions. After reading and signing the informed consent form, participants filled out a background information form, and their demographic characteristics, maritime experience, and competency certificates were collected. Following this, participants underwent approximately 15 minutes of familiarization training, including practice of basic simulator operation, interface recognition, and the TOR response process, ensuring they fully understood the experimental tasks and procedures before the formal experiment. In the formal experimental phase, each participant was required to complete... Figure 5 The mid-scenario takeover task first consists of a monitoring phase of approximately 3 minutes. During this period, the vessel navigates autonomously according to the pre-set route plan. Participants can perform other tasks without any intervention. When the autonomous navigation system meets the TOR trigger conditions, it issues a takeover request. Participants then observe the current vessel's encounter situation. After confirming that takeover is permissible, the manual control phase begins. At this point, participants must maneuver the vessel to avoid obstacles based on traffic conditions and COLREGs rules. The task ends when the vessel reaches the last waypoint along the planned route. Throughout the experiment, the system automatically records the first set of maneuvering commands and the complete vessel trajectory data.

[0199] The experiment collected 18 sets of complete takeover data from 18 participants under a set scenario. Each set of data included information on the entire process from the issuance of the TOR to the completion of the task, covering multi-dimensional data such as takeover time, ship trajectory, sequence of maneuvering commands, and encounter event records. Based on the takeover quality quantification method proposed in Chapter 3, this embodiment systematically analyzed the 18 sets of data.

[0200] The takeover quality quantification adopts the CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism, which comprehensively considers 12 evaluation indicators in dimensions such as collision risk (average CRI, maximum CRI, 95th percentile CRI), domain risk (average domain risk, average DCPA risk, average TCPA risk), collision avoidance safety (minimum DCPA, minimum TCPA), COLREGs compliance (compliance score, number of violations), and encounter statistics (number of domain intrusions, number of high-risk encounters). The score range is [0,1], and the higher the value, the better the takeover quality. Figure 5 The text provides the calculated weights for each indicator, where... and The calculated weights are the largest, at 0.1178 and 0.1101 respectively, indicating that these two indicators contain less information that overlaps with other indicators, and therefore are given higher weights. Furthermore, the COLREGs compliance scores are relatively high, reflecting the importance of collision avoidance rule compliance indicators, suggesting that COLREGs are one of the important indicators for distinguishing operational quality. The correlation coefficient matrix of each evaluation indicator is as follows: Figure 8 As shown.

[0201] Figure 6 The study revealed a complex linear relationship pattern among the 12 evaluation indicators, with the most significant strong positive correlation appearing in... , , and The fact that these three risk factors are numerically equivalent indicates that, due to their different calculation logics, they depict potential collision risks from different perspectives. The series of indicators shows strong consistency, for example... and The correlation between them reached 0.726, indicating that the risk levels calculated by different risk indicators have a certain inherent consistency, and the risk indicators in the time dimension... and some spatial dimensions index( , and There is a certain positive correlation, reflecting the characteristics of spatiotemporal risk coupling. Table 5 shows the results and main evaluation indicators of 18 sets of experimental data.

[0202] Table 5. Experimental Data Results and Main Evaluation Indicators

[0203]

[0204] Table 5 presents the comprehensive quality index values ​​and main evaluation indicators for 18 sets of experimental data. The table shows that the comprehensive quality index values ​​of the 18 sets of data exhibit significant dispersion, ranging from a minimum of 0.1604 (P02) to a maximum of 0.6511 (P13), with an average score of 0.2929. The three participants with the highest comprehensive quality index values ​​are P13, P18, and P17. These three participants share the characteristics of high COLREG compliance scores (0.30, 1.00, and 0.70, respectively), few violations (3, 0, and 0, respectively), and high average collision avoidance safety scores. In contrast, the three participants with the lowest comprehensive quality index values, P02, P05, and P16, all exhibited very low COLREG compliance scores (0.00, 0.01, and 0.00, respectively) and a high number of violations, indicating that although rule compliance has a relatively small weight, it is one of the key factors affecting takeover quality. From the various evaluation indicators, the average COLREGs compliance score was 0.3394, and the median was 0.35, showing significant individual differences. Of the 18 participants, only P18 achieved a perfect score of 1.00, P17 achieved a high score of 0.70, while P02, P05, and P16 scored 0 or close to 0, indicating that strictly adhering to the COLREGs rules in complex encounter situations is a challenging task. Combined with... Figure 8 Further analysis revealed a positive correlation between the number of violations and the number of encounters, indicating that the more complex the encounter situation and the more target vessels involved, the higher the likelihood of violations.

[0205] The average collision avoidance safety score in this embodiment reflects the overall safety level of the inter-ship distance and attitude during takeover. The average safety score of the 18 participants was 0.76, ranging from 0.50 to 1.00. Most participants were able to maintain a relatively safe avoidance distance, but P16's safety score was significantly low. Combined with its takeover time of only 42 seconds (the shortest among all participants), it can be inferred that this participant may have hastily taken over with insufficient situational awareness, resulting in an inefficient avoidance decision. Ultimately, the participant successfully avoided a collision by maintaining a relatively small distance.

[0206] In this embodiment, TOT reflects the time required for operator situational awareness reconstruction and decision preparation. The takeover times of the 18 participants ranged from 42 to 165 seconds, with an average of 82.9 seconds and a median of 75.5 seconds. This result is orders of magnitude different from the takeover times in the field of autonomous vehicles (typically 2-7 seconds), validating the assertion in the literature review in Chapter 2 that ship takeover times are much longer than those in vehicles.

[0207] The distribution of takeover times exhibited a right-skewed characteristic, with most participants (13 people, 72%) having takeover times concentrated in the 50-100 second range, 5 people (28%) having takeover times exceeding 100 seconds, and 1 person (5%) having a takeover time less than 50 seconds. The shortest takeover time was P16 (42 seconds), and the longest was P09 (165 seconds), a difference of nearly four times, showing significant individual differences. Further analysis revealed a correlation between takeover time and participants' rank and experience: the average takeover time for captains was 103.5 seconds, for chief officers 96.75 seconds, for second officers 65.67 seconds, and for third officers 74.25 seconds. This showed that the average takeover time for senior officers was significantly longer than that for junior officers. This difference may reflect that experienced officers tend to fully assess the situation before confirming takeover, while less experienced officers may take over prematurely due to a less accurate grasp of the timing. From the perspective of encounter situations, the actual number and types of encounters faced by different test subjects after takeover varied. This difference stemmed from the different timing of takeovers, which led to different relative positional relationships between the test vessel and the target vessel, resulting in different encounter situation assessments. In the 18 data sets, the number of encounters ranged from 0 to 4, with an average of 2 encounters. The most common encounter type was yielding, followed by overtaking and being overtaken; face-to-face encounters and direct cross-course maneuvers were less frequent. This distribution characteristic is related to the target vessel's position and direction of movement in the scenario design, and also reflects the prevalence of cross-encounter situations in open waters. Figure 7 The diagram shows the maneuvering trajectories of the vessel after 18 participants took over. The diagram clearly shows significant differences in the avoidance strategies employed by different participants. Some participants (e.g., P17, P18) adopted a relatively smooth, large starboard turn to avoid the obstacle, resulting in a regular arc trajectory. This maneuvering method meets the "clear avoidance" requirement of the COLREGs rules and provides stable handling and good passenger comfort. Other participants (e.g., P02, P09) exhibited trajectories with multiple turns and swaying, reflecting hesitation in decision-making or difficulty in situational assessment. This maneuvering method not only increases operational workload but may also lead to misinterpretation of the vessel's intentions by surrounding vessels. Still other participants (e.g., P06, P14) adopted a deceleration avoidance strategy, resulting in trajectories significantly shorter than those of other participants. While this is a reasonable avoidance method in overtaking or being overtaken situations, it may not meet the requirements for the right-of-way vessel's actions in cross-traffic situations.

[0208] This embodiment explores the quantitative relationship between takeover time and takeover quality by pairing the takeover time of 18 sets of data with the corresponding comprehensive quality index values. Figure 8A scatter plot of takeover time and comprehensive quality index value is presented. The plot shows that the comprehensive quality index value does not exhibit a simple linear relationship with takeover time, but rather a non-linear trend of first rising and then falling. This phenomenon verifies the hypothesis that premature takeover leads to insufficient situational awareness, while delayed takeover results in insufficient remaining decision-making time; there exists a time window for optimizing takeover quality. Based on the above analysis, this embodiment uses a quadratic regression model to analyze the functional relationship between takeover time and comprehensive quality index value. Least squares regression was performed on 18 sets of data to obtain parameter estimates. , , ,because The hypothesis that quality changes in an inverted U-shape over time was verified. According to equations (51) and (52), the point estimate of the optimal takeover time is approximately 112.4 seconds, at which point the inferred comprehensive quality index is 0.35. This result indicates that in this experimental scenario, the operator confirming takeover approximately 112.4 seconds after the TOR is issued achieves the best takeover quality performance. However, the optimal timing based on point estimation does not consider the uncertainty of parameter estimation and the limitations of finite samples. To quantify this uncertainty, this embodiment further employs a hierarchical Bayesian inference method. The prior distribution setting incorporates the prior knowledge of experts in the maritime field: experts believe that it typically takes 60 to 120 seconds from the system issuing the TOR to the operator completing the situation assessment and making the takeover decision, therefore the prior mean is set as follows. The value of 1 second, located in the early part of the middle of this interval, reflects a safety tendency towards "early and adequate preparation." The prior standard deviation is set at 1. The time interval ensures that 95% of the prior probability distribution is within a reasonable range, and the equivalent prior sample size is set to [value missing]. This approach preserves the influence of prior information on inference without overly dominating the data. Utilizing the conjugate property of the normal distribution, the posterior distribution parameters are calculated according to equations (58) to (63), and the Bayesian inference results are shown in Table 6. The posterior mean is 101.2 seconds, the posterior standard deviation is 26.624 seconds, and the 95% Bayesian confidence interval is [49.0, 153.4] seconds. This means that confirming takeover approximately 101 seconds after the TOR is issued yields the best takeover quality, and the true optimal timing has a 95% probability of falling within the 49-153 second interval.

[0209] Table 6. Bayesian inference results for the optimal takeover timing

[0210]

[0211] Table 6 shows that the posterior mean of 101.2 seconds is approximately 21 seconds later than the prior mean of 80 seconds. This difference reflects the dominant role of experimental data in inference. Although experts tend to "take over as early as possible" in their prior knowledge, experimental data shows that in this scenario, operators need more time for situational awareness and decision preparation; taking over too early (e.g., within 80 seconds) actually leads to a decrease in quality. (Data Influence) This indicates that posterior inference incorporates both prior and data influences, with a larger prior weight and greater shrinkage. The values ​​are within the ideal range [0.2, 0.8], indicating that the prior and posterior data have achieved a good balance. Figure 9 This shows the posterior distribution curve obtained from Bayesian inference, from Figure 9 As can be seen, the posterior distribution follows an inverted U-shaped curve, concentrated around 101.2 seconds, with a 95% confidence interval spanning approximately 104 seconds. This level of uncertainty is reasonable, reflecting the inherent uncertainty of estimation under limited sample conditions. Compared to the point estimate of 112.4 seconds, the posterior mean of the Bayesian inference is approximately 11.2 seconds earlier. Based on the above analysis, the experimental results of this embodiment indicate that in open water multi-ship encounter scenarios, the optimal time for onboard personnel to take over the intelligent vessel is approximately 101 seconds after the TOR is issued (95% confidence interval [49.0, 153.4] seconds). This time window ensures that operators have sufficient time to complete situational awareness reconstruction and risk assessment, while avoiding insufficient remaining maneuvering time due to excessive delay. Significant deviations from this optimal interval in the takeover time (too early or too late) will lead to a significant decrease in takeover quality.

[0212] This embodiment addresses the issue of manual takeover on intelligent ships, constructing a quantification method for takeover quality based on the CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism, and using hierarchical Bayesian inference to determine the optimal takeover timing. Experimental results show significant individual differences in the takeover time of the pilots, with an average takeover time of 82.9 seconds. The corresponding comprehensive quality index also exhibits significant dispersion, with an average comprehensive quality index value of 0.2929. This indicates that takeover quality is not only affected by traffic conditions but also closely related to the pilot's decision-making and timing selection. There is a certain correlation between takeover time and the participants' rank. The average takeover time for the captain is 103.5 seconds, for the chief mate 96.75 seconds, for the second mate 65.67 seconds, and for the third mate 74.25 seconds. Senior crew members have significantly longer average takeover times than junior crew members. This difference may reflect that experienced crew members tend to confirm takeover only after fully understanding the situation, while less experienced crew members may take over prematurely due to inaccurate timing. Further analysis reveals a non-linear relationship between takeover time and quality. Quadratic regression analysis shows that the comprehensive quality index initially increases and then decreases with increasing takeover time, indicating an optimal time point that maximizes the comprehensive quality index. The optimal takeover time inferred from the least squares regression curve is 112.4 seconds, at which point the inferred comprehensive quality index is 0.358. However, the optimal timing based on point estimation does not consider the uncertainty of parameter estimation and the limitations of limited samples. The advantage of the Bayesian method lies in its ability to naturally quantify estimation uncertainty. The width of the confidence interval reflects the accuracy of the inference given the current sample size. The prior distribution is set as a normal distribution with a mean of 80 seconds and a standard deviation of 30 seconds, reflecting the safety tendency of early takeover in the maritime field. The posterior mean shifting towards 101.2 seconds indicates that the experimental data supports a later takeover timing, with the data playing a dominant role in the inference result. The optimal takeover timing determined using hierarchical Bayesian inference has a posterior mean of 101.2 seconds, with a 95% Bayesian confidence interval of 49.0 to 153.4 seconds.

[0213] This embodiment addresses the quality quantification problem of manual takeover of intelligent ships in complex traffic situations. It constructs a comprehensive quantitative method based on the CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism and uses hierarchical Bayesian inference to determine the optimal takeover timing. The system integrates multi-dimensional indicators such as collision risk index, COLREGs compliance, and ship domain intrusion, achieving an objective quantitative evaluation of takeover quality. The effectiveness of the model was verified through a ship maneuvering simulator experiment involving 18 certified crew members, revealing the intrinsic correlation between takeover time and takeover quality. Experimental results show significant individual differences in the takeover time of the pilots, with an average takeover time of 82.9 seconds, ranging from 42 seconds to 165 seconds. The corresponding comprehensive quality index value also exhibits large dispersion, with an average comprehensive quality index value of 0.2929, ranging from 0.1604 to 0.6511. A non-linear inverted U-shaped relationship exists between takeover time and takeover quality. The optimal takeover time predicted by the least squares regression curve is 112.4 seconds, at which point the predicted comprehensive quality index value is 0.358. Based on hierarchical Bayesian inference, the posterior mean of the optimal takeover timing is 101.2 seconds, with a 95% confidence interval of 49.0 to 153.4 seconds. The difference between the two methods is that the point estimation method mainly relies on mathematical extreme points for inference, and the inference results have exceeded the range of actual data. In contrast, Bayesian inference integrates prior information and observational data, and the optimal takeover time given is located in the data center area, exhibiting more robust characteristics.

[0214] Compared with existing technologies, the method described in this embodiment first constructs an objective quantitative indicator system based on AIS trajectory data and ship attribute parameters, including 12 dimensions such as collision risk, domain risk, collision avoidance safety, COLREGs compliance, and encounter statistics. Then, it employs the CRITIC (Criteria Importance Through Intercriteria Correlation) method to achieve objective weighting based on data characteristics, and uses the TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) method to calculate the comprehensive quality index value. Based on this, statistical analysis reveals the quantitative relationship between takeover quality and encounter scenario characteristics, identifying key scenario factors affecting takeover quality. Finally, hierarchical Bayesian inference is used to integrate domain prior knowledge and limited experimental data to establish a probabilistic mapping relationship between takeover timing and takeover quality, achieving the inference of the optimal takeover timing and its uncertainty range.

[0215] The beneficial effects of the method described in this embodiment are as follows:

[0216] (1) The method described in this embodiment constructs a multi-dimensional takeover quality quantification index system applicable to intelligent ships. In terms of collision risk quantification, based on the Potential Risk Ship Domain (PRSD) model improved by Zou et al.

[16] , a dynamic collision risk index (CRI) is proposed by integrating the three elements of ship domain intrusion degree, TCPA (Time to Closest Point of Approach, TCPA) and DCPA (Distance to Closest Point of Approach, DCPA). The overall risk level and extreme risk state are characterized by statistical indicators such as average CRI, maximum CRI, and 95th percentile CRI. In terms of domain safety quantification, the average domain risk, average DCPA risk, average TCPA risk, minimum DCPA, and minimum TCPA are introduced to evaluate the safety margin from the perspective of statistical distribution and critical value. In terms of rule compliance quantification, based on the COLREGs research framework of Hagen et al.

[17] , the dynamic ship domain theory is innovatively introduced to improve the identification of encounter situations. The rule compliance degree is quantified by rule compliance and the number of violations.

[0217] (2) The method described in this embodiment proposes an objective weight allocation method based on the CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism. This method uses the CRITIC method to objectively calculate the weights based on the data comparison intensity and the correlation between indicators, avoiding the situation where the weighting method is too subjective; on this basis, the TOPSIS method is used to calculate the relative closeness of each quantitative indicator to the positive ideal solution and the negative ideal solution, and multi-attribute comprehensive evaluation is achieved through geometric distance, which significantly improves the objectivity and scientificity of the calculation process;

[0218] (3) The method described in this embodiment proposes an optimal takeover timing inference model based on hierarchical Bayesian theory. Addressing the practical constraint of scarce experimental data in the maritime field, this model integrates prior knowledge with limited observational data through a hierarchical Bayesian framework, establishing a probabilistic mapping relationship between takeover timing and takeover quality. It can not only predict the optimal time budget required to ensure high-quality takeover, but also provide a credible interval considering parameter uncertainties, offering both scientific and robust quantitative support for the timing decision-making of takeover by personnel on intelligent ships.

[0219] The following documents are involved in this embodiment:

[0220] [1]Jo M, Choi W, Lim M, et al. A study on improving the internationalregulations for preventing collisions at sea (COLREG) for the introduction ofmaritime autonomous surface ships (MASS)[J]. Journal of InternationalMaritime Safety, Environmental Affairs, and Shipping, 2024, 8(4):2428006.

[0221] [2]Yang X, Zhou T, Utne IB, et al. A framework for quantification ofcoupling risk in the transition between operational modes of MASS[J].Reliability Engineering System Safety, 2025, 264:111436.

[0222] [3]IMO. Maritime Safety Committee - 110th session (MSC 110)[R].London:IMO, 2025.

[0223] [4]Gold C, Damböck D, Lorenz L, et al. “Take over!” How long does ittake to get the driver back into the loop?[J]. Proceedings of the HumanFactors and Ergonomics Society Annual Meeting, 2013, 57(1):1938–1942.

[0224] [5]Shyshova O, Gadhavi P, Tenzer M, et al. Preparation Times: AnExperimental-Based Discussion About Limits for Takeover in Highly AutomatedSystems[J]. 2024 IEEE Conference on Cognitive and Computational Aspects ofSituation Management (CogSIMA), 2024:71–78.

[0225] [6]Zhu Y, Li C, Qiao Z, et al. From Distraction to Action: ElevatingSituation Awareness with Visual Assistance in Level 3 Autonomous Driving[J].International Journal of Human-Computer Interaction, 2025, 41(7):4271–4283.

[0226] [7]Kim YW, Yoon SH. Understanding drivers’ situation awareness inhighly automated driving using SAGAT, SART, and eye-tracking data[J].Transportation Research Part F-Traffic Psychology and Behaviour, 2025,109:1437–1450.

[0227] [8]Gold C, Körber M, Lechner D, et al. Taking Over Control FromHighly Automated Vehicles in Complex Traffic Situations: The Role of TrafficDensity[J]. Human Factors, 2016,58(4):642–652.

[0228] [9]Braunagel C, Rosenstiel W, Kasneci E. Ready for Take-Over A NewDriver Assistance System for an Automated Classification of Driver Take-OverReadiness[J]. IEEE Intelligent Transportation Systems Magazine 2017,9(4):10–22.

[0229]

[10] Wang T, Han Y, Li W, et al. A Takeover Risk Assessment ApproachBased on an Improved ANP-XGBoost Algorithm for Human-Machine Driven Vehicles[J]. IEEE Access, 2024,12:48379–48387.

[0230]

[11] Agrawal S, Peeta S. Evaluating the impacts of situationalawareness and mental stress on takeover performance under conditionalautomation[J]. Transportation Research Part F-Traffic Psychology andBehaviour, 2021,83:210–225.

[0231]

[12] Liang K, Calvert SC, Lint JWC van. Towards Safe and ComfortableVehicle Control Transitions: A Systematic Review of Takeover Time, TimeBudget, and Takeover Performance[DB / OL]. https: / / arxiv.org / abs / 2507.22262.

[0232]

[13] Radlmayr J, Gold C, Lorenz L, et al. How Traffic Situations andNon-Driving Related Tasks Affect the Take-Over Quality in Highly AutomatedDriving[C]. Proceedings of the Human Factors and Ergonomics Society AnnualMeeting, 2014,58(1):2063–2067.

[0233]

[14] Wang W, Wang Z, Li Q, et al. Task-specific versus free-choicenon-driving-related tasks: Enhancing automated driving takeover researchthrough varied task engagement[J]. iScience, 2025,28(6):112783.

[0234]

[15] Han Y, Wang T, Shi D, et al. The Effect of MultifactorInteraction on the Quality of Human-Machine Co-Driving Vehicle Take-Over[J].Sustainability, 2023,15(6):5131.

[0235]

[16] Zou Y, Zhang Y, Wang S, et al. Ship regulatory method formaritime mixed traffic scenarios based on key risk ship identification[J].Ocean Engineering, 2024,298:117105.

[0236]

[17] Hagen IB, Vassbotn O, Skogvold M, et al. Safety and COLREGevaluation for marine collision avoidance algorithms[J]. Ocean Engineering,2023,288(1):115991.

[0237]

[18] Kearon J. Computer program for collision avoidance and trackkeeping[C]. Conference on Mathematical Aspects on Marine Traffic, 1977: 229-242.

[0238]

[19] Mou JM, Tak C van der, Ligteringen H. Study on collisionavoidance in busy waterways by using AIS data[J]. Ocean Engineering, 2010,37(5-6):483–490.

[0239]

[20] Fujii Y, Tanaka K. Traffic Capacity[J]. Journal of Navigation,1971,24(4):543–552.

[0240]

[21] Ha J, Roh M-I, Lee H-W. Quantitative calculation method of thecollision risk for collision avoidance in ship navigation using the CPA andship domain[J]. Journal of Computational Design and Engineering, 2021,8(3):894–909.

[0241] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for quantifying the quality of personnel takeover and inferring the optimal takeover timing on intelligent ships, characterized in that, The specific steps include: S1: Acquire AIS data of the intelligent ship; and the intelligent ship AIS data includes at least ship speed, ship size data and ship attitude data; S2: Define the safety boundary of the ship domain SD based on the intelligent ship AIS data, and its expression is: In the formula: Represents the forward boundary of the SD in the shipbuilding domain; Represents the backward boundary of the SD domain in the shipbuilding field; Represents any point in the field of shipping. p The angle between the line connecting the ship's center and the bow direction; Indicates the length of the ship; Parameters representing the lateral influence of SD in the shipbuilding sector; Parameters representing the longitudinal impact of SD in the shipbuilding sector; Indicates the potential collision risk index; express The ship's speed at any given moment; Indicates the boundary of the SD in the shipbuilding field; S3: Preset a ship collision scenario and obtain the basic parameters between the ship and the target ship under the corresponding ship collision scenario. The basic parameters include the nearest encounter distance and the nearest encounter time. By introducing an encounter situation factor, the collision risk value between ships is obtained based on the basic parameters and the safety boundary. S4: Construct a safety quantification model for evaluating the safety margin of intelligent ship takeover operations, and obtain safety quantification evaluation indicators based on the relative distance between the ship and the target ship and the ship's attitude data according to the safety quantification model; S5: Use safety quantitative evaluation indicators and ship-to-ship collision risk values ​​as multi-dimensional quantitative indicators, and construct a CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism based on multi-dimensional quantitative indicators to obtain comprehensive takeover quality indicators. S6: Construct a Bayesian-based optimal takeover timing inference strategy, obtain the inference result of the optimal takeover timing based on the comprehensive takeover quality index, and then provide quantitative decision-making for the dynamic switching of intelligent ship operation mode based on the inference result.

2. The method for quantifying the quality of personnel takeover and inferring the optimal takeover timing in intelligent ships according to claim 1, characterized in that, S3 specifically includes the following steps: S31: Under a preset world coordinate system, a ship collision scenario is preset, and the basic parameters between the current ship and the target ship under the corresponding ship collision scenario are obtained. The basic parameters include the nearest encounter distance and the nearest encounter time, and their expressions are as follows: In the formula: This indicates that you will encounter a distance in the near future; Indicates the time that will be encountered soon; Indicates the true bearing of the target ship TS relative to the ship's OS; This indicates the relative distance between the ship's OS and the target ship's TS; This represents the relative motion vector of the target vessel TS with respect to the vessel OS. Represents the relative motion vector The corresponding direction angle; S32: Obtain the safety margin of the domain boundary between the target ship TS and the ship's OS based on the nearest encounter distance. The expression for the security margin of the domain boundary is: In the formula: Indicates the security margin of the domain boundary; Indicates the radius of the target vessel's territory in the specified direction; S33: The adaptive adjustment dynamic coefficients for the ship domain are obtained based on the nearest encounter distance and the nearest encounter time. The expression for this coefficient is: In the formula: This indicates the collision risk value at the SD boundary in the marine field; , They represent The corresponding nearest encounter distance With the most recent meeting time ; This represents the adaptive adjustment dynamic coefficient in the marine field; S34: By introducing an encounter situation factor and adaptively adjusting dynamic coefficients based on the ship domain, combined with the domain boundary safety margin, the inter-ship collision risk value is obtained. for: In the formula: This indicates the situational factors that will be encountered; express The reciprocal of the maximum value.

3. The method for quantifying the quality of personnel takeover and inferring the optimal takeover timing in intelligent ships according to claim 2, characterized in that, S4 constructs a safety quantification model for evaluating the safety margin of intelligent ship takeover operations, including a first quantification model for quantifying the relative distance and attitude between the target ship TS and the ship OS, a second quantification model for quantifying whether the yielding ship's avoidance action meets the requirements of the first rule, a third quantification model for quantifying whether the straight-going ship's straight-going action meets the requirements of the second rule, and a fourth quantification model for quantifying whether the navigation of the two ships in a meeting situation meets the requirements of the third rule. The expression for the first quantization model is: In the formula: This represents the distance quantization function; The CPA distance represents the distance between the nearest points on the territory of two ships when they meet at sea. , , , Both indicate that the distance parameter threshold is set based on the potential collision risk indicator and ; This represents the contact angle quantization function used to reflect the relative headings between the two ships at the CPA moment; This indicates the relative heading angle of the ship's OS with respect to the target ship's TS; Indicates the preset cutoff angle; This represents the relative bearing quantization function used to reflect the relative bearing between the two ships at the CPA moment; This indicates the relative heading angle of the target vessel TS with respect to the vessel's OS; This represents the minimum and maximum relative heading angles of the target vessel TS with respect to the vessel OS. The attitude score is the weighted sum of the contact angle quantization function and the relative orientation quantization function. Indicates the weighting coefficient; This represents the final security score of the first quantization model; The expression for the second quantization model is: In the formula: This represents the collision avoidance delay penalty function; This indicates the distance to the ship's domain boundary as set according to the potential risk collision index; Indicates the relative distance between the target vessel and this vessel; This indicates the maximum change in course of a vessel during maneuvering; The penalty function representing the heading maneuver; Indicates the minimum detectable change in heading; Indicates the heading change threshold; This indicates the amount of deceleration of the ship's relative speed; Indicates the moment of entering the pre-set warning zone; express Ship speed at any given time; express Ship speed at any given time; The penalty function representing the ship's maneuvering; A threshold parameter representing the amount of deceleration of a ship's relative speed; Indicates the score for vessel maneuvering; Indicates the weighting coefficient; This represents the safety score used to quantify whether the give-way vessel's avoidance actions meet the requirement of taking early, significant, and obvious avoidance actions; The method for obtaining the third quantization model is as follows: S100: Based on empirical values ​​and the safety boundary of the SD (Safety Safe Zone) in the maritime domain, four encounter zone stages {stage1, stage2, stage3, stage4} are defined according to the distance between the target vessel and the vessel to avoid collision. Stage 1: The target vessel is more than the maximum distance threshold of the target vessel, which is the encounter zone stage that is not considered; Stage 2: The target vessel is more than the first distance threshold of the target vessel, which is the encounter zone stage; Stage 3: The target vessel is more than the second distance threshold of the target vessel, which is the encounter zone stage; Stage 4: The target vessel is more than the third distance threshold of the target vessel, which is the encounter zone stage. And the first distance region > the second distance region > the third distance region; When the target ship and the current ship are in Stage 2 or Stage 3, the corresponding stage penalty applies. for: In the formula: This represents the penalty function for heading changes within Stage 2 or Stage 3; This indicates the amount of deceleration of the ship's relative speed; This indicates the rate of increase in the relative speed of a ship; This represents a penalty function for increasing speed; This represents the penalty function for reduced speed; Indicates the weighting coefficient; S101: By introducing a yield compensation coefficient Optimization phase penalty Then, the left-turn penalty function is defined when the target ship and the current ship are in Stage 4. The third quantization model is constructed as follows: In the formula: This indicates the optimized stage penalty; This represents a safety score used to quantify whether a straight-running vessel complies with maintaining course and speed, taking action only when necessary, and not turning left. The expression for the fourth quantization model is: In the formula: Indicates the maximum right turn angle between the target vessel and this vessel; Indicates the penalty function for non-right turns; This represents the minimum effective right turn angle; Indicates the heading change threshold; The penalty function represents the situation where the target ship and the ship pass each other on the starboard side; Indicates the weighting coefficient; This represents a comprehensive score used to quantify whether both ships turn to starboard and pass each other's port side in a head-on encounter situation. Indicates that the ship is t The course of time; Indicates the ship's course when entering the pre-set warning zone.

4. The method for quantifying the quality of personnel takeover and inferring the optimal takeover timing in intelligent ships according to claim 3, characterized in that, S5 specifically includes the following steps: S51: Using AIS data from intelligent ships as sample data, and combining safety quantitative evaluation indicators and inter-ship collision risk values ​​as multi-dimensional quantitative indicators, a decision matrix is ​​obtained. and ,in n Indicates the number of sample data; m This indicates the number of evaluation indicators for multi-dimensional quantitative indicators; Represents the elements in the decision matrix; S52: The CRITIC-TOPSIS multi-dimensional comprehensive evaluation mechanism is constructed based on multi-dimensional quantitative indicators as follows: S521: Use Z-score standardization to eliminate the influence of dimensions in the decision matrix and obtain the optimal decision matrix. and ,and In the formula: express The result after Z-score standardization; Indicates the first j The average of each evaluation indicator; Indicates the first j The standard deviation of each evaluation indicator is used to characterize the comparative strength of the evaluation indicator among different samples. S522: Obtain the optimal decision matrix Furthermore, the Pearson correlation coefficient matrix among the various evaluation indicators; S523: Obtain conflict indicators for evaluation metrics based on the Pearson correlation coefficient matrix; The formula for obtaining the conflict index is as follows: In the formula: Indicates the first j Conflicting indicators among evaluation metrics; The Pearson correlation coefficient matrix represents the first... j The first evaluation indicator and the first k Pearson correlation coefficients among the evaluation indicators; S524: Based on the conflict index and contrast strength, the CRITIC weight is obtained as follows: In the formula: Indicates the first j The product of the conflict index and the contrast strength of each evaluation indicator; Indicates CRITIC weight; Indicates the first k The product of the conflict index and the contrast strength of each evaluation indicator; S53: Decision Matrix Vector normalization is performed to obtain the normalized decision matrix; the decision matrix is ​​then processed. The expression for vector normalization is: In the formula: This represents an element in the normalized decision matrix; S54: Obtain the weighted normalized value based on the normalized decision matrix and CRITIC weights. for: S55: Obtain the corresponding weighted normalized value positive ideal solution With negative ideal solution for: S56: Based on the ideal solution With negative ideal solution Combined with weighted normalized values Calculate the sum of each evaluation index and the positive ideal solution. With negative ideal solution European distance for: S57: Based on Euclidean distance Get the i The comprehensive takeover quality index of each evaluation indicator for: 。 5. The method for quantifying the quality of personnel takeover and inferring the optimal takeover timing in intelligent ships according to claim 4, characterized in that, The optimal takeover timing inference strategy based on Bayesian methods constructed in S6 specifically includes the following steps: S61: Settings n The system records the takeover scenarios for several intelligent ships, including the actual takeover time for each scenario. Comprehensive takeover quality indicators The corresponding record data; S62: Obtain the actual takeover time Comprehensive takeover quality indicators The Pearson correlation coefficient was used to determine the linear correlation between the two, and a quadratic regression model was established based on the linear correlation to characterize the nonlinear relationship between takeover time and comprehensive takeover quality index. The expression for the quadratic regression model is: In the formula: This represents the output of the quadratic regression model; Indicate design parameters; Indicates the regression parameters of the model; Indicates the actual takeover time Standardized time; S63: The regression parameters of the quadratic regression model are estimated using the least squares method as follows: In the formula: This represents the estimated values ​​of the model's regression parameters; Represents the design matrix and ; S64: Obtain the optimized takeover time based on the estimated values ​​of the model regression parameters. for: S65: Based on optimized takeover time Obtain the optimal takeover time for: In the formula: Indicates the actual takeover time The mean; Indicates the actual takeover time The variance; S66: Based on the optimal takeover time Construct a hierarchical Bayesian model to obtain the data mean of optimal timing uncertainty. for: In the formula: An estimate representing the uncertainty of the optimal timing; The method for obtaining the estimate of the uncertainty of the optimal timing is as follows: based on the estimated values ​​of the model regression parameters. The uncertainty of the optimal timing is estimated by combining the error propagation formula with the standard error of the model regression parameters. for: In the formula: Indicates intermediate parameters; S67: Estimation of uncertainty based on optimal timing Obtain the posterior precision value for: In the formula: These represent the prior precision and data precision, respectively, indicating the uncertainty of the optimal timing. Indicates the prior standard deviation; S68: Data mean based on S67 combined with the uncertainty of optimal timing Obtain the posterior mean : In the formula: This represents the prior mean; S69: Based on the a posteriori precision value Obtain the posterior standard deviation for: S70: Based on the posterior standard deviation with posterior mean Obtain the Bayesian confidence interval for the optimal takeover timing. for: The Bayesian confidence interval is the inference result of the optimal takeover timing.

Images