A ship autonomous navigation intention interpretation method and device

Abstract

The application relates to a ship autonomous navigation intention interpretation method and device, which comprises the following steps: obtaining a composite data set based on dynamic environment data of the ship and an autonomous future trajectory output by an autonomous navigation system; performing state decoding and motion mode identification on a state vector in the composite data set to obtain a segmentation point set; performing navigation scene type judgment on the state vector of the ship and state vectors of other ships in all dynamic environment data to obtain a segmentation point set; segmenting the state vector sequence based on the union set of the segmentation point set and the segmentation point set to obtain a minimum intention unit sequence; performing unit merging and intention recognition on each minimum intention unit to obtain an autonomous navigation intention interpretation result; the application combines the dynamic environment data with the autonomous future trajectory, and performs multi-level understanding on the autonomous future trajectory by determining the segmentation point at which the motion mode changes and the segmentation point at which the navigation scene changes, thereby improving the effective trust of the crew in the system.

Description

Technical Field

[0001] This invention relates to the field of intelligent ship technology, and in particular to a method and apparatus for interpreting the autonomous navigation intention of a ship. Background Technology

[0002] With the development of artificial intelligence technology, ships equipped with Autonomous Navigation Systems (ANS) and Autonomous Collision Avoidance Systems (CAS) have become a mainstream technological direction for improving maritime traffic safety and efficiency while reducing the workload of maritime personnel. Currently, ANS or CAS are typically based on complex algorithmic models, such as reinforcement learning, model predictive control, or expert systems. These models can integrate ship dynamic models, sensor data (such as radar, AIS, and cameras), and electronic chart information, and calculate optimal or near-optimal operational commands in real time based on maritime collision avoidance rules, such as steering angle adjustments, speed adjustments, or path planning. However, the core algorithms generally suffer from a "black box" characteristic of opaque decision-making processes, lacking semantic interpretation of the decision results. Before ships achieve full autonomy, their navigation will always be under the supervision, intervention, and control of human pilots. When machine-generated decisions contradict the crew's judgments based on experience and rules, the lack of a coherent and multi-layered understanding of the system's decision results makes it difficult for the crew to establish effective trust in the system. This cognitive gap not only hinders the improvement of human-machine collaboration efficiency but may also lead to misjudgments or operational conflicts in emergency situations, creating new safety hazards.

[0003] Therefore, there is an urgent need to propose a method and device for interpreting the autonomous navigation intentions of ships, in order to solve the technical problem in the existing technology that when the decision results generated by the machine contradict the judgment of the crew based on experience and rules, the crew has difficulty in establishing effective trust in the system due to the lack of a coherent and multi-level understanding of the system's decision results. Summary of the Invention

[0004] In view of this, it is necessary to provide a method and apparatus for interpreting the autonomous navigation intention of ships, in order to solve the technical problem in the prior art that when the decision results generated by the machine contradict the judgment of the crew based on experience and rules, the crew has difficulty in establishing effective trust in the system due to the lack of a coherent and multi-level understanding of the system's decision results.

[0005] To address the aforementioned problems, in a first aspect, the present invention provides a method for interpreting the autonomous navigation intention of a ship, comprising:

[0006] Based on the dynamic environmental data of the vessel in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system, a composite dataset is obtained; the composite dataset includes all dynamic environmental data and the state vector sequence in the autonomous future trajectory;

[0007] State decoding and motion pattern recognition are performed on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes.

[0008] Based on the state vector of the vessel at any given moment and the state vector of any other vessel in all the dynamic environmental data, the navigation scenario type is determined, and a set of cut-off points where the navigation scenario changes is obtained.

[0009] The state vector sequence is segmented based on the union of the segmentation point set and the cutting point set to obtain the minimum intent unit sequence;

[0010] Each minimum intent unit in the minimum intent unit sequence is merged and intent identified to obtain the autonomous navigation intent interpretation result.

[0011] In one possible implementation, the composite dataset obtained based on the ship's dynamic environmental data in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system includes:

[0012] Obtain the state vectors of other vessels at different times from AIS and radar data, and obtain the geographic information of multiple static obstacles from the electronic chart system;

[0013] By integrating the state vector of the other vessel at each moment with the geographical information, dynamic environmental data is obtained.

[0014] Based on the autonomous future trajectory, determine the state vector sequence;

[0015] All the dynamic environment data are time-aligned with the state vector sequence to obtain a composite dataset.

[0016] In one possible implementation, the step of performing state decoding and motion pattern recognition on each of the state vectors in the state vector sequence to obtain a set of segmentation points where the motion pattern changes includes:

[0017] Based on each of the state vectors in the state vector sequence, the observation sequence is obtained;

[0018] The observation sequence is input into a hidden Markov model for state decoding to obtain a set of hidden states.

[0019] The hidden state set is solved using the Viterbi algorithm to obtain the hidden state sequence;

[0020] Based on the hidden state sequence, determine the motion pattern recognition result at each time step in the state vector sequence;

[0021] When the motion pattern recognition results change at adjacent time points, a segmentation point is determined;

[0022] Based on all the split points, we obtain the set of split points.

[0023] In one possible implementation, obtaining the observation sequence based on each of the state vectors in the state vector sequence includes:

[0024] Determine the temporal features that characterize its dynamic behavior from the state vector sequence;

[0025] Based on the aforementioned temporal characteristics, determine the state vector at each time step;

[0026] Based on the state vector, determine the quantitative index reflecting its motion state at each moment;

[0027] Based on the quantification index, determine the observation vector at the corresponding time.

[0028] Based on all observation vectors, the observation sequence is obtained.

[0029] In one possible implementation, the step of determining the navigation scenario type based on the state vector of the vessel at any given time and the state vector of any other vessel in all dynamic environmental data, to obtain a set of cutoff points where the navigation scenario has changed, includes:

[0030] Based on the state vector of the vessel at any moment in the state vector sequence and the state vector of other vessels at any moment in the dynamic environment data, the nearest encounter distance and the nearest encounter time are obtained.

[0031] Whether to perform scene classification is determined based on the nearest encounter distance and the nearest encounter time;

[0032] If so, the navigation scenario type of the dynamic environment data at adjacent time points is compared, and the cut-off point is determined when a change occurs;

[0033] Determine the set of split points based on all split points.

[0034] In one possible implementation, the step of merging and identifying each minimum intent unit in the minimum intent unit sequence to obtain an autonomous navigation intent interpretation result includes:

[0035] Behavioral intent prediction is performed on each minimum intent unit in the minimum intent unit sequence to obtain an initial behavioral unit sequence;

[0036] The smallest adjacent units with the same behavioral intent in the initial behavioral unit sequence are merged to obtain multiple trajectory segments;

[0037] Based on the multiple trajectory segments, an operational layer behavioral intent sequence is obtained;

[0038] The operational layer behavioral intent sequence is mapped based on a preset finite state machine model to obtain the tactical layer intent and strategic layer intent for each trajectory segment.

[0039] Based on the operational layer behavioral intent sequence, the tactical layer intent, and the strategic layer intent, the autonomous navigation intent interpretation result is obtained.

[0040] In one possible implementation, the step of predicting the behavioral intent for each minimum intent unit in the minimum intent unit sequence to obtain an initial behavioral unit sequence includes:

[0041] By extracting feature vectors from each minimum intent unit in the minimum intent unit sequence using a preset set of behavioral intents, the actual feature data of each minimum intent unit is obtained.

[0042] The belief of multiple behavioral intentions of each minimum intention unit in the preset behavioral intention set is predicted to obtain the belief value of each minimum intention unit corresponding to each behavioral intention.

[0043] The behavioral intention with the highest belief value in the preset behavioral intention set of each minimum intention unit is determined as the final recognition result of the corresponding unit;

[0044] Based on all the final recognition results of the minimum intent unit sequence, the initial behavior unit sequence is obtained.

[0045] In one possible implementation, predicting the beliefs of multiple behavioral intentions for each minimum intention unit in the preset set of behavioral intentions to obtain a belief value for each minimum intention unit corresponding to each behavioral intention includes:

[0046] Construct observation and transfer models;

[0047] The prior probability of each behavioral intention is obtained by probabilistically predicting the behavioral intention of each smallest intention unit in the preset behavioral intention set through the transfer model.

[0048] The actual feature data is input into the observation model to obtain the observation likelihood of each behavioral intention;

[0049] The prior probability is corrected by the observed likelihood to obtain the belief value of each minimum intention unit corresponding to each behavioral intention.

[0050] In one possible implementation, after performing unit merging and intent recognition on each minimum intent unit in the minimum intent unit sequence to obtain the autonomous navigation intent interpretation result, the method further includes:

[0051] Based on the interpretation results of the autonomous navigation intent, each trajectory segment on the autonomous future trajectory is displayed in different colors, and the operational layer intention, the tactical layer intention, and the strategic layer intention are displayed in different colors respectively.

[0052] Secondly, the present invention also provides a device for interpreting the autonomous navigation intention of a ship, comprising:

[0053] The data acquisition module is used to obtain a composite dataset based on the dynamic environmental data of the vessel in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system; the composite dataset includes all dynamic environmental data and the state vector sequence in the autonomous future trajectory;

[0054] The pattern recognition module is used to perform state decoding and motion pattern recognition on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes.

[0055] The scenario judgment module is used to judge the navigation scenario type based on the state vector of the vessel at any time and the state vector of any other vessel in all the dynamic environment data, and to obtain the set of cut-off points where the navigation scenario has changed.

[0056] The sequence segmentation module is used to segment the state vector sequence based on the union of the segmentation point set and the splitting point set to obtain the minimum intent unit sequence;

[0057] The intent recognition module is used to merge and recognize the intent of each minimum intent unit in the minimum intent unit sequence to obtain the autonomous navigation intent interpretation result.

[0058] The beneficial effects of this invention are as follows: Based on the dynamic environmental data of the vessel in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system, a composite dataset is obtained; the composite dataset includes the state vector sequence in all dynamic environmental data and the autonomous future trajectory; state decoding and motion pattern recognition are performed on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes; the navigation scenario type is determined based on the state vector of the vessel at any time and the state vector of any other vessel in all dynamic environmental data to obtain a set of cutting points where the navigation scenario changes; the state vector sequence is segmented based on the union of the set of segmentation points to obtain a sequence of minimum intent units; each minimum intent unit in the sequence of minimum intent units is merged and intent is identified to obtain the autonomous navigation intent interpretation result; this application combines dynamic environmental data with the autonomous future trajectory output by the system, and performs multi-level understanding of the autonomous future trajectory by determining the segmentation points where the motion pattern changes and the segmentation points where the navigation scenario changes, thereby making the intent recognition of the minimum intent unit more consistent with the crew's judgment and improving the crew's effective trust in the system. Attached Figure Description

[0059] Figure 1 A schematic flowchart of an embodiment of the ship autonomous navigation intent interpretation method provided by the present invention;

[0060] Figure 2 For the present invention Figure 1 A schematic diagram of an embodiment of step S102;

[0061] Figure 3 For the present invention Figure 1 A schematic flowchart of an embodiment of step S105;

[0062] Figure 4 A schematic diagram of an embodiment of the ship autonomous navigation intention interpretation device provided by the present invention. Detailed Implementation

[0063] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0064] like Figure 1 As shown, a specific embodiment of the present invention discloses a method for interpreting the autonomous navigation intention of a ship, comprising:

[0065] S101. Based on the dynamic environmental data of the vessel in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system, a composite dataset is obtained; the composite dataset includes all dynamic environmental data and the state vector sequence in the autonomous future trajectory.

[0066] In this embodiment of the invention, the composite dataset is a structured, time-aligned dataset that integrates and formalizes the original decision results (i.e., autonomous future trajectories) output by ANS or CAS with environmental information obtained from various shipborne sensors and electronic chart systems, thereby providing complete and accurate input for subsequent trajectory segmentation and intent reasoning.

[0067] S102. Perform state decoding and motion pattern recognition on each state vector in the state vector sequence to obtain the set of segmentation points where the motion pattern changes.

[0068] In this embodiment of the invention, each state vector X in the state vector sequence is selected from... t Key motion features are extracted, including linear motion features: acceleration magnitude, acceleration direction, and rate of change of velocity; angular motion features: angular velocity, rate of change of heading, and turning radius; and higher-order motion features: jerk and rate of change of curvature. These are then used for state decoding and motion pattern recognition to obtain the set of segmentation points where motion patterns change.

[0069] S103. Based on the state vector of this vessel at any time and the state vector of any other vessel in all dynamic environmental data, determine the navigation scenario type and obtain the set of cut-off points where the navigation scenario changes.

[0070] In this embodiment of the invention, the navigation scenario type is determined by comparing the state vector of the vessel at any moment in the state vector sequence with the state vector of any other vessel in all dynamic environmental data. Specifically, the encounter characteristics of the vessel and all relevant other vessels are calculated, and then the scenario type is determined. When the navigation scenario type changes, the segmentation point is obtained, thereby obtaining the set of segmentation points where the navigation scenario has changed.

[0071] S104. The state vector sequence is segmented based on the union of the segmentation point set and the splitting point set to obtain the smallest intention unit sequence.

[0072] The embodiments of the present invention can calculate the union of the set of segmentation points and the set of cutting points to obtain the union set, and then use the union set to segment the state vector sequence to obtain the minimum intention unit sequence.

[0073] S105. Perform unit merging and intent identification on each minimum intent unit in the minimum intent unit sequence to obtain the autonomous navigation intent interpretation result.

[0074] In this embodiment of the invention, the smallest intent units that are temporally continuous and have similar intents are merged, and then the intent of the merged units is identified to obtain the autonomous navigation intent interpretation result.

[0075] The ship autonomous navigation intention interpretation method provided in this embodiment of the invention can be applied to the ship autonomous navigation intention interpretation system. The ship autonomous navigation intention interpretation system can be a software system running on a terminal device. The terminal device can be a server, tablet computer, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), mobile phone, etc. This application embodiment does not impose any restrictions on the specific type of terminal device.

[0076] Compared with existing technologies, this embodiment provides a composite dataset based on the dynamic environmental data of the vessel in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system. The composite dataset includes state vector sequences from all dynamic environmental data and the autonomous future trajectory. State decoding and motion pattern recognition are performed on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes. The navigation scenario type is determined based on the vessel's state vector at any given time and the state vector of any other vessel in all dynamic environmental data to obtain a set of cutting points where the navigation scenario changes. The state vector sequence is segmented based on the union of the segmentation point set and the cutting point set to obtain a sequence of minimum intent units. Each minimum intent unit in the minimum intent unit sequence is merged and intent is identified to obtain the autonomous navigation intent interpretation result. This application combines dynamic environmental data with the autonomous future trajectory output by the system, and performs multi-level understanding of the autonomous future trajectory by determining the segmentation points where the motion pattern changes and the cutting points where the navigation scenario changes. This makes the intent recognition of the minimum intent units more consistent with the crew's judgment, improving the crew's effective trust in the system.

[0077] In some embodiments of the present invention, step S101 includes:

[0078] Obtain the state vectors of other vessels at different times from AIS and radar data, and obtain the geographic information of multiple static obstacles from the electronic chart system;

[0079] By integrating the state vectors of other ships at each moment with geographical information, dynamic environmental data is obtained.

[0080] Determine the state vector sequence based on the autonomous future trajectory;

[0081] All dynamic environment data and state vector sequences are time-aligned to obtain a composite dataset.

[0082] This invention embodiment plans the controlled vessel, as intended by the autonomous system, for a future period of time (prediction time domain). The behavioral scheme within a given range is formalized as an ordered sequence of timestamped state vectors, denoted as . This sequence can characterize the future trajectory of the controlled vessel, as shown in formula (1):

[0083] (1)

[0084] In the formula, Represents the predicted endpoint; These are discrete points in time; It is the carrier at all times The state vector. This vector contains at least the core motion components necessary for resolving the intent, as shown in Equation (2):

[0085] (2)

[0086] In the formula, and These represent the latitude and longitude positions at that moment; Ground speed (SOG); For heading to ground (COG); For ships at time tangential acceleration; For ships at time angular velocity of turning.

[0087] This invention aligns environmental data (including the position, speed, heading, obstacle information, hydrological and meteorological data, etc.) from different sensors and data sources with the predicted trajectory state vector of the vessel using timestamps, ensuring that all data are on the same time reference. The state vector at each moment contains the vessel's navigation state (position, speed, heading, acceleration, etc.) and environmental state (distance to the nearest other vessel, relative speed, encounter angle, and other key features), yielding the state vector at each time point. Corresponding dynamic environment data . It is a set containing information about other ships and static obstacles, as shown in formula (3):

[0088] (3)

[0089] In the formula, This indicates that the information obtained from AIS and radar data about other ships at a given time... The state vector, , Indicates all ships, This represents all obstacles. Represents the first obtained from the electronic chart system Geographic information of static navigational obstructions (such as islands, reefs, shoals, no-navigation zones, etc.).

[0090] The final output is a complete composite dataset. : This provides the necessary information input for subsequent segmentation of ship trajectories based on two dimensions: motion characteristics and scene changes.

[0091] In some embodiments of the present invention, such as Figure 2 As shown, step S102 includes:

[0092] S201. Based on each state vector in the state vector sequence, the observation sequence is obtained.

[0093] In some embodiments of the present invention, the observation sequence is obtained based on each state vector in the state vector sequence. Specifically, in some embodiments of the present invention, step S201 includes:

[0094] Determine the temporal features that characterize the dynamic behavior from the state vector sequence.

[0095] Embodiments of the present invention start from the state vector sequence In this process, temporal features that characterize dynamic behavior are extracted. Temporal features can include position features, velocity features, heading features, and timestamps, which are used to provide feature representations at appropriate levels of abstraction for different tasks, from low-level signal features to high-level semantic features.

[0096] Based on the temporal characteristics, determine the state vector at each time step.

[0097] In this embodiment of the invention, after obtaining the temporal features, the state vector at each moment can be determined based on the temporal features. For example, based on features such as timestamps, the state vector corresponding to each moment in the state vector sequence can be determined. .

[0098] Based on the state vector, determine the quantitative index that reflects the motion state at each moment.

[0099] This invention allows for the extraction of quantitative indicators that directly reflect the motion state at each moment based on the state vector. These quantitative indicators are standardized metrics that can be precisely measured and expressed numerically, used to objectively describe, compare, and evaluate the state or behavioral characteristics of a system. In the context of autonomous ship navigation and trajectory analysis, quantitative indicators specifically refer to those that can transform continuous, complex dynamic behaviors into discrete, computable numerical features, such as ship speed and turning angular velocity.

[0100] Based on the quantitative indicators, the observation vector at the corresponding time point is determined.

[0101] The embodiments of the present invention can construct the observation vector at that moment using quantifiable indicators (i.e., ship speed and turning angular velocity). Among them, tangential acceleration It directly reflects the trend of ship speed change and can be approximated at discrete time points as follows: . Indicates acceleration. Indicates deceleration. Indicates uniform velocity; angular velocity of turning. It directly reflects the trend of changes in the ship's course and can be approximated at discrete time points as follows: . It is usually defined as turning right. To turn left, Indicates direct flights.

[0102] Based on all observation vectors, the observation sequence is obtained.

[0103] This invention provides a complete observation sequence by calculating the entire predicted trajectory. .

[0104] S202. Input the observation sequence into the Hidden Markov Model for state decoding to obtain the set of hidden states.

[0105] The Hidden Markov Model (HMM) in this invention is a statistical model used to describe a Markov process with hidden unknown parameters. The challenge lies in determining the hidden parameters of the process from the observable parameters. These parameters are then used for further analysis, such as pattern recognition. The observed sequence can be used as input to the HMM, which can be a pre-trained HMM, allowing for state decoding of the observed sequence. The Hidden Markov Model is formally defined as a quintuple: Among them, the set of hidden states , This indicates uniform straight-line flight, and its corresponding observation characteristics are: and ; This indicates accelerated straight-line flight, and its corresponding observation characteristics are: and ; This indicates decelerated straight-line flight, and its corresponding observation feature is: and ; This indicates a right turn, and its corresponding observation feature is: ; This indicates a left turn, and the corresponding observation feature is: Observation set It is a vector space consisting of all possible observation vectors. The state transition probability matrix. As shown in formula (4):

[0106] (4)

[0107] In the formula, Indicates the ship's motion mode from state Transition to state The probability of.

[0108] Observational probability distribution As shown in formula (5):

[0109] (5)

[0110] In the formula, This indicates that the ship is in a concealed state. When the eigenvector is observed The probability density.

[0111] Initial state probability distribution As shown in formula (6):

[0112] (6)

[0113] In the formula, This indicates that the ship is in motion mode at the initial moment. The probability of.

[0114] HMM parameters This was achieved through offline training using a large amount of labeled ship trajectory data and the Baum-Welch algorithm. The observation sequence was obtained after this process. After training the HMM, the observation sequence By inputting the data into the trained HMM, we can obtain the set of hidden states. .

[0115] S203. Solve the hidden state set using the Viterbi algorithm to obtain the hidden state sequence.

[0116] This invention embodiment can use the Viterbi algorithm to process the hidden state set. The solution is performed to generate the hidden state sequence most likely to produce the observed sequence. .

[0117] S204. Based on the hidden state sequence, determine the motion pattern recognition result at each time step in the state vector sequence.

[0118] Hidden state sequence in the embodiments of the present invention This can be represented as an autonomous future trajectory. The motion pattern recognition result is obtained by considering the hidden state at each time step in the hidden state sequence. This result is denoted as... .

[0119] S205. When the motion pattern recognition results change at adjacent time points, determine the segmentation point.

[0120] This invention determines the segmentation point by comparing the identification results of adjacent time moments. A moment is marked as a segmentation point if and only if the motion pattern changes in adjacent time moments.

[0121] S206. Based on all the split points, obtain the set of split points.

[0122] The set of all motion segmentation points in this embodiment of the invention is as follows: As shown in formula (7):

[0123] (7)

[0124] In the formula, Represents the time of controlled vessels The movement pattern.

[0125] In some embodiments of the present invention, step S103 includes:

[0126] Based on the state vector of this vessel at any moment in the state vector sequence and the state vector of other vessels at any moment in the dynamic environment data, the nearest encounter distance and the nearest encounter time are obtained.

[0127] The embodiments of the present invention construct an environmental dataset For any moment in the prediction time domain and any other ship It will be based on the ship's state vector and the state vector of his ship Calculate a set of key collision hazard parameters: nearest distance (DCPA) and nearest time to nearest (TCPA), as shown in Equation (8):

[0128]

[0129]

[0130]

[0131] (8)

[0132] In the formula, and These are the initial position vectors of our ship and the other ship, respectively; and These are the velocity vectors of our ship and the other ship, respectively.

[0133] Determine whether to classify the scene based on the nearest encounter distance and the nearest encounter time;

[0134] If so, compare the navigation scenario types of the dynamic environment data at adjacent time points, and determine the cut-off point when a change occurs;

[0135] Determine the set of split points based on all split points.

[0136] This invention embodiment determines whether to perform scene classification based on the nearest encounter distance and the nearest encounter time. This indicates that the two ships are approaching each other. A safety threshold can be preset. and ,like ,and If a collision is deemed to be at risk, scenario classification is required. Then, based on COLREGs and maritime practice, a basic set of navigation scenarios is defined, covering typical encounter situations: .in Indicates an encounter; Indicates a meeting at an intersection (giving way to another boat); Indicates a meeting at an intersection (for ships traveling in a straight line); It indicates to pursue or overtake; This indicates safety. A classification function is used to determine the specific scenario type.

[0137] When the type or number of elements in this basic scene set changes, it is considered that a fundamental change has occurred in the scene, and this moment is identified as a dividing point. All such dividing points constitute... As shown in formula (9):

[0138] (9)

[0139] In the formula, Representative moment The navigation scenario type.

[0140] Furthermore, step S104 includes: by analyzing the set and set Taking the union of the sets yields the final set of segmentation moments. ,Right now Using sets All time points in the original sequence of ship behavior. The segmentation process yields a sequence composed of the smallest units of intent. Within each unit, the controlled vessel's motion pattern is relatively stable, and the environment is singular.

[0141] In some embodiments of the present invention, such as Figure 3 As shown, step S105 includes:

[0142] S301. Perform behavioral intent prediction on each minimum intent unit in the minimum intent unit sequence to obtain the initial behavioral unit sequence.

[0143] After segmenting the behavior sequence, this embodiment of the invention enters the intent reasoning stage. The goal of this stage is to assign a basic operational intent that best represents the behavior of each smallest intent unit generated in the preceding steps and conforms to the cognition of a human driver. This allows for behavioral intent prediction of each smallest intent unit in the sequence, resulting in an initial behavior unit sequence.

[0144] In some embodiments of the present invention, step S301 includes:

[0145] By extracting the feature vector of each minimum intent unit in the minimum intent unit sequence through a preset set of behavioral intents, the actual feature data of each minimum intent unit is obtained.

[0146] This invention employs a probabilistic reasoning framework—the Recursive Bayesian Intention Estimation (RBII) model—to achieve this goal. This model effectively handles the uncertainty of time-series data and updates beliefs about intentions incrementally, aligning with the characteristics of interpreting continuous decision-making processes. Before performing reasoning, a discrete, finite set of predefined behavioral intentions at the operational level needs to be defined based on maritime practice, denoted as […]. ,For example For the output of the previous step... Minimum Intent Unit The vector describing the features of the trajectory segment is extracted as the actual feature data of the unit (including trajectory geometry features, kinematic features, navigation scene information, etc.), denoted as . .

[0147] In some embodiments of the present invention, the belief in multiple behavioral intentions of each minimum intention unit in a preset set of behavioral intentions is predicted to obtain a belief value for each behavioral intention corresponding to each minimum intention unit, including:

[0148] Construct observation and transfer models.

[0149] The present invention establishes an observation model This indicates an assumption of the true intention. In the case where the observed feature vector is The probability. This model can be built through supervised learning on a large amount of labeled ship trajectory data. Establishing a transfer model. This indicates that the ship's behavioral intention is different from the intention of the previous unit. Intent to transfer to the current unit The probability of.

[0150] The prior probability of each behavioral intention is obtained by probabilistically predicting the behavioral intention of each smallest intention unit in the preset behavioral intention set through a transfer model.

[0151] The actual feature data is input into the observation model to obtain the observation likelihood of each behavioral intention;

[0152] By correcting the prior probability through observational likelihood, the belief value corresponding to each behavioral intention for each smallest intention unit is obtained.

[0153] The embodiments of the present invention apply to sequences composed of minimal intent units. The system starts with the first unit and calculates the posterior probability of each unit belonging to a specific operational intent. For the first unit... Each unit is intended to Belief value The calculation follows the recursive Bayesian update rule as shown in formula (10):

[0154] (10)

[0155] In the formula: : indicates that after processing the first After one unit, the system believes the true intention is The belief or posterior probability; : Indicates that the first step has been completed After each unit, the intention The belief, as the prior knowledge for this update; This is the prediction step, predicting the intent of the current unit. The prior probability; This is the observation likelihood used in the update step, utilizing the actual feature data of the current cell. To correct the prediction, we obtain the belief value for each behavioral intention corresponding to each smallest intention unit. .

[0156] The behavioral intent with the highest belief value in the preset behavioral intent set of each smallest intent unit is determined as the final recognition result of the corresponding unit.

[0157] The embodiments of the present invention complete the first After updating the beliefs of each unit, the intention with the highest probability is selected as the final recognition result of that unit, as shown in formula (11):

[0158] (11)

[0159] in, It is the first Each unit is assigned an optimal operational layer intent label. Using this framework, the system can calculate the posterior probability distribution of all possible operational layer intents for each smallest intent unit.

[0160] The initial behavior unit sequence is obtained based on all the final recognition results of the minimum intent unit sequence.

[0161] In this embodiment of the invention, the intent with the highest probability is selected as the recognition result for that unit. The final output is an initial sequence of behavioral units with a clear operational intent label. .

[0162] S302. Merge the smallest adjacent units with the same behavioral intention in the initial behavioral unit sequence to obtain multiple trajectory segments.

[0163] After completing the serialization reasoning of basic operational intentions, this embodiment of the invention enters the interpretation generation stage. The core task of this stage is to systematically aggregate and enhance the discrete, machine-like operational intention sequences output from previous steps, ultimately generating a semantic interpretation for human drivers that includes three cognitive levels: operational, tactical, and strategic. This involves processing the initial behavioral unit sequences labeled with operational intentions. Conduct an inspection. Merge adjacent smallest units with the same operational-level behavioral intent to form longer, more meaningful units. Figure 1 This involves dividing the data into multiple trajectory segments. This step aims to eliminate minor fluctuations that may arise from overly fine segmentation, resulting in a more consistent and stable temporal representation of intent.

[0164] S303. Based on multiple trajectory segments, obtain the sequence of operational layer behavioral intentions.

[0165] Embodiments of the present invention can be based on an initial behavioral unit sequence The operational layer behavioral intent of each trajectory segment after merging is determined, thereby obtaining the operational layer behavioral intent sequence.

[0166] S304. Based on a preset finite state machine model, the sequence of operational layer behavioral intentions is mapped to obtain the tactical layer intention and strategic layer intention for each trajectory segment.

[0167] This step in the embodiments of the present invention is key to achieving semantic enhancement. Based on a preset rule base containing a large amount of knowledge from maritime experts or a preset finite state machine model, the operational layer behavioral intent sequence of multiple combined trajectory segments, combined with its context, is mapped to higher-level tactical and strategic intents.

[0168] Tactical intents describe a series of combined actions taken to accomplish a specific local task, typically directly related to collision avoidance and the execution of navigation rules. Mapping rules can be formally represented as: "IF [Scenario='Intersecting Encounter (Give Way Vessel)'] AND [Action Sequence='Slow Down' FOLLOWEDBY' Turn Right'] THEN [Tactical Intent='Execute Intersecting Encounter Avoidance (Turn Right and Slow Down)']". Other tactical intents include "overtaking maneuvers," "encounter avoidance," and "maintaining a safe encounter distance."

[0169] Strategic intent is linked to the highest-level planning of the entire journey, explaining which longer-term goal the current tactical action is intended to achieve. For example, when the system detects that a tactical intent to "perform an avoidance maneuver" has been completed, and the subsequent sequence of actions is "restore the original course," the system can interpret its strategic intent as: "Return and continue along the planned route to the next waypoint."

[0170] S305. Based on the sequence of operational-level behavioral intentions, tactical-level intentions, and strategic-level intentions, the interpretation results of autonomous navigation intentions are obtained.

[0171] After obtaining the operational layer behavioral intent, tactical layer intent, and strategic layer intent for each trajectory segment, the embodiments of the present invention can merge these three intents into the autonomous navigation intent interpretation result for the corresponding trajectory segment.

[0172] In some embodiments of the present invention, after step S105, the method further includes:

[0173] Based on the interpretation results of autonomous navigation intent, each trajectory segment on the autonomous future trajectory is displayed in different colors, and the operational layer intent, tactical layer intent, and strategic layer intent are displayed in different colors respectively.

[0174] The ultimate result of this invention is to efficiently and intuitively convey the multi-layered decision interpretations generated in the preceding steps to the navigator through a human-machine interface (HMI), thereby establishing trust and reaching consensus. The future navigation trajectory planned by the ship's autonomous system can be highlighted or color-coded on the ECDIS, with different colors representing different tactical intentions (e.g., green for "follow the planned route," orange for "maneuver and avoidance"). When the navigator hovers the mouse over or clicks on a segment of the trajectory, the system will display a clear text label explaining the operational, tactical, and strategic intentions corresponding to that segment of the trajectory.

[0175] Furthermore, to further enhance the naturalness and depth of human-computer interaction, this invention introduces an intelligent question-answering module based on Large Language Model (LLM), elevating human-computer interaction from traditional point-and-click queries to a new level of conversational understanding. Drivers can directly ask questions to the system using natural language via voice or text, eliminating the need for complex menu operations, thus enabling them to probe and explore the system's decision-making intentions.

[0176] This invention enhances decision-making transparency, addresses the "black box" problem, and improves human-machine trust. By transforming the predicted trajectory generated by the autonomous navigation system into a multi-layered interpretation of intent (including operational, tactical, and strategic layers) that can be intuitively understood by human drivers, this invention effectively solves the "black box" problem of opaque decision-making processes in existing autonomous systems. It establishes trust in the system and lays the foundation for efficient human-machine collaboration. It promotes consensus on human-machine intent, reduces operational conflicts, and improves collaborative efficiency. By interpreting machine decision results into behavioral intents that humans can understand, it significantly reduces the possibility of misunderstandings and operational conflicts caused by information asymmetry. When system decisions differ from the driver's experience-based judgment, the clear interpretation provided by this method acts as a translator, helping the driver quickly understand the system's decision-making basis, avoiding blind takeover or adversarial operations, and ensuring that both humans and machines can work collaboratively around a common navigation goal, thereby improving safety and efficiency. It strengthens situational awareness, enables early intervention, and improves maritime traffic safety. By interpreting the future behavioral intent of the controlled vessel, this invention allows the navigator to understand the vessel's future decisions at the initial stage of decision execution. This predictive explanatory information enhances the operator's ability to perceive and anticipate future situations. It enables the operator to intervene in the controlled vessel's behavior early and with evidence, thereby significantly improving the safety level of intelligent ship navigation.

[0177] To better implement the ship autonomous navigation intent interpretation method in the embodiments of the present invention, correspondingly, the embodiments of the present invention also provide a ship autonomous navigation intent interpretation device, such as... Figure 4 As shown, the ship autonomous navigation intention interpretation device 400 includes:

[0178] The data acquisition module 401 is used to obtain a composite dataset based on the dynamic environmental data of the ship in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system; the composite dataset includes all dynamic environmental data and the state vector sequence in the autonomous future trajectory.

[0179] The pattern recognition module 402 is used to perform state decoding and motion pattern recognition on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes.

[0180] The scenario judgment module 403 is used to judge the navigation scenario type based on the state vector of the ship at any time and the state vector of any other ship in all dynamic environment data, and obtain the set of cut-off points where the navigation scenario has changed.

[0181] The sequence segmentation module 404 is used to segment the state vector sequence based on the union of the segmentation point set and the split point set to obtain the minimum intent unit sequence;

[0182] The intent recognition module 405 is used to merge and recognize each minimum intent unit in the minimum intent unit sequence to obtain the autonomous navigation intent interpretation result.

[0183] The ship autonomous navigation intent interpretation device 400 provided in the above embodiments can realize the technical solutions described in the above embodiments of the ship autonomous navigation intent interpretation method. The specific implementation principles of each module or unit can be found in the corresponding content in the above embodiments of the ship autonomous navigation intent interpretation method, which will not be repeated here.

[0184] The above provides a detailed description of the method and apparatus for interpreting the autonomous navigation intent of ships provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for interpreting the autonomous navigation intention of a ship, characterized in that, include: Based on the dynamic environmental data of the vessel in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system, a composite dataset is obtained; the composite dataset includes all dynamic environmental data and the state vector sequence in the autonomous future trajectory; State decoding and motion pattern recognition are performed on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes. Based on the state vector of the vessel at any given moment and the state vector of any other vessel in all the dynamic environmental data, the navigation scenario type is determined, and a set of cut-off points where the navigation scenario changes is obtained. The state vector sequence is segmented based on the union of the segmentation point set and the cutting point set to obtain the minimum intent unit sequence; For each minimum intent unit in the minimum intent unit sequence, unit merging and intent identification are performed to obtain the autonomous navigation intent interpretation result; The step of performing state decoding and motion pattern recognition on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes includes: Based on each of the state vectors in the state vector sequence, the observation sequence is obtained; The observation sequence is input into a hidden Markov model for state decoding to obtain a set of hidden states. The hidden state set is solved using the Viterbi algorithm to obtain the hidden state sequence; Based on the hidden state sequence, determine the motion pattern recognition result at each time step in the state vector sequence; When the motion pattern recognition results change at adjacent time points, a segmentation point is determined; Based on all the split points, we obtain the set of split points.

2. The method for interpreting a ship's autonomous navigation intent according to claim 1, characterized in that, Based on the dynamic environmental data of the vessel in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system, a composite dataset is obtained, including: Obtain the state vectors of other vessels at different times from AIS and radar data, and obtain the geographic information of multiple static obstacles from the electronic chart system; By integrating the state vector of the other vessel at each moment with the geographical information, dynamic environmental data is obtained. Based on the autonomous future trajectory, determine the state vector sequence; All the dynamic environment data are time-aligned with the state vector sequence to obtain a composite dataset.

3. The method for interpreting a ship's autonomous navigation intention according to claim 1, characterized in that, The step of obtaining the observation sequence based on each of the state vectors in the state vector sequence includes: Determine the temporal features that characterize its dynamic behavior from the state vector sequence; Based on the aforementioned temporal characteristics, determine the state vector at each time step; Based on the state vector, determine the quantitative index reflecting its motion state at each moment; Based on the quantification index, determine the observation vector at the corresponding time. Based on all observation vectors, the observation sequence is obtained.

4. The method for interpreting a ship's autonomous navigation intention according to claim 1, characterized in that, The step of determining the navigation scenario type based on the state vector of the vessel at any given time and the state vector of any other vessel in all dynamic environmental data, and obtaining a set of cutoff points where the navigation scenario changes, includes: Based on the state vector of the vessel at any time in the state vector sequence and the state vector of other vessels at any time in the dynamic environment data, the nearest encounter distance and the nearest encounter time are obtained. Whether to perform scene classification is determined based on the nearest encounter distance and the nearest encounter time; If so, the navigation scenario type of the dynamic environment data at adjacent time points is compared, and the cut-off point is determined when a change occurs; Determine the set of split points based on all split points.

5. The method for interpreting a ship's autonomous navigation intent according to claim 1, characterized in that, The step of merging and identifying the intent of each minimum intent unit in the minimum intent unit sequence to obtain the autonomous navigation intent interpretation result includes: Behavioral intent prediction is performed on each minimum intent unit in the minimum intent unit sequence to obtain an initial behavioral unit sequence; The smallest adjacent units with the same behavioral intent in the initial behavioral unit sequence are merged to obtain multiple trajectory segments; Based on the multiple trajectory segments, an operational layer behavioral intent sequence is obtained; The operational layer behavioral intent sequence is mapped based on a preset finite state machine model to obtain the tactical layer intent and strategic layer intent for each trajectory segment. Based on the operational layer behavioral intent sequence, the tactical layer intent, and the strategic layer intent, the autonomous navigation intent interpretation result is obtained.

6. The method for interpreting a ship's autonomous navigation intention according to claim 5, characterized in that, The step of predicting the behavioral intent for each minimum intent unit in the minimum intent unit sequence to obtain an initial behavioral unit sequence includes: By extracting feature vectors from each minimum intent unit in the minimum intent unit sequence using a preset set of behavioral intents, the actual feature data of each minimum intent unit is obtained. The belief of multiple behavioral intentions of each minimum intention unit in the preset behavioral intention set is predicted to obtain the belief value of each behavioral intention corresponding to each minimum intention unit. The behavioral intention with the highest belief value in the preset behavioral intention set of each minimum intention unit is determined as the final recognition result of the corresponding unit; Based on all the final recognition results of the minimum intent unit sequence, the initial behavior unit sequence is obtained.

7. The method for interpreting a ship's autonomous navigation intention according to claim 6, characterized in that, The step of predicting the beliefs of multiple behavioral intentions for each minimum intention unit in the preset behavioral intention set to obtain the belief value for each behavioral intention corresponding to each minimum intention unit includes: Construct observation and transfer models; The prior probability of each behavioral intention is obtained by probabilistically predicting multiple behavioral intentions of each smallest intention unit in the preset behavioral intention set through the transfer model. The actual feature data is input into the observation model to obtain the observation likelihood of each behavioral intention; The prior probability is corrected by the observed likelihood to obtain the belief value of each minimum intention unit corresponding to each behavioral intention.

8. The method for interpreting a ship's autonomous navigation intention according to claim 5, characterized in that, After performing unit merging and intent recognition on each minimum intent unit in the minimum intent unit sequence to obtain the autonomous navigation intent interpretation result, the method further includes: Based on the interpretation results of the autonomous navigation intent, each trajectory segment on the autonomous future trajectory is displayed in different colors, and the operational layer intention, the tactical layer intention, and the strategic layer intention are displayed in different colors respectively.

9. A device for interpreting the autonomous navigation intention of a ship, characterized in that, include: The data acquisition module is used to obtain a composite dataset based on the dynamic environmental data of the vessel in the prediction time domain and the autonomous future trajectory output by the autonomous navigation system; the composite dataset includes all dynamic environmental data and the state vector sequence in the autonomous future trajectory; The pattern recognition module is used to perform state decoding and motion pattern recognition on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes. The scenario judgment module is used to judge the navigation scenario type based on the state vector of the vessel at any time and the state vector of any other vessel in all the dynamic environment data, and to obtain the set of cut-off points where the navigation scenario has changed. The sequence segmentation module is used to segment the state vector sequence based on the union of the segmentation point set and the splitting point set to obtain the minimum intent unit sequence; The intent recognition module is used to merge and recognize the intent of each minimum intent unit in the minimum intent unit sequence to obtain the autonomous navigation intent interpretation result; The step of performing state decoding and motion pattern recognition on each state vector in the state vector sequence to obtain a set of segmentation points where the motion pattern changes includes: Based on each of the state vectors in the state vector sequence, the observation sequence is obtained; The observation sequence is input into a hidden Markov model for state decoding to obtain a set of hidden states. The hidden state set is solved using the Viterbi algorithm to obtain the hidden state sequence; Based on the hidden state sequence, determine the motion pattern recognition result at each time step in the state vector sequence; When the motion pattern recognition results change at adjacent time points, a segmentation point is determined; Based on all the split points, we obtain the set of split points.

Images