AFS neural coding method and system for time series feature representation

Abstract

The application provides an AFS neural coding method and system for time sequence feature representation, and belongs to the field of intelligent shipping and ship encounter situation recognition. The method comprises the following steps: taking a ship as a reference object, extracting the position, speed and heading information of a target ship in a continuous time window from AIS, radar or multi-source fusion perception data, and constructing a ship encounter situation time sequence sample; calculating relative distance, DCPA, TCPA, relative speed, bearing rate and other navigation dynamic indexes to form an original navigation dynamic index sequence; based on the experience distribution of the training sample and the monotonic relationship of the index risk, the original index sequence is mapped to an AFS simple concept sequence; an AFS complex concept is constructed through axiomatic fuzzy set logic; the AFS simple concept sequence is input into a GRU neural time encoder to learn the time sequence dependence, and a fixed dimension feature representation is obtained; and the encounter situation category and AFS semantic explanation chain are output according to the representation. The application has both time sequence representation capability and navigation semantic interpretability.

Description

Technical Field

[0001] This invention relates to the fields of intelligent shipping, ship traffic information processing, navigation safety of unmanned surface vessels, and ship encounter situation recognition. Specifically, it relates to an AFS neural coding method and system oriented towards time-series feature representation. This method encodes ship encounter processes formed from AIS, radar, shipborne navigation sensors, or multi-source fusion sensing data into a nautical-meaningful AFS semantic time series, and extracts temporal feature representations that can be used for encounter situation recognition, collision hazard evolution analysis, and auxiliary identification of key targets of interest through a neural time encoder. Background Technology

[0002] With the development of intelligent shipping, unmanned surface vessels, and autonomous navigation technologies, the demand for dynamic identification, risk evolution analysis, and semantic interpretation of ship encounters is constantly increasing. In complex encounter scenarios in open waters, motion elements such as the relative distance, DCPA, TCPA, relative bearing, relative heading, rate of change of heading, and rate of change of speed between the target vessel and the ship evolve continuously over time. The collision hazard state cannot be fully characterized by a single static risk value or instantaneous risk level.

[0003] Existing ship collision risk assessment methods typically construct collision hazard or risk level based on indicators such as DCPA, TCPA, relative distance, relative bearing, and speed. These methods can reflect the magnitude of risk at a certain moment or within a local window, but they are insufficient in expressing the process semantics of risk changes over time. They are difficult to characterize dynamic encounter patterns such as increased collision hazard, decreased collision hazard, continuous approach, enhanced relative motion conflict, and significant fluctuations in heading or speed.

[0004] Traditional fuzzy evaluation methods typically require manual setting of membership functions, risk boundaries, and indicator weights, mapping indicators such as DCPA and TCPA to risk membership degrees through expert experience. While these methods offer some interpretability, they are highly parameter-dependent and lack stability across different waters, ship sizes, and encounter types, making it difficult to meet the needs for objective characterization and automatic identification in complex and dynamic encounter environments.

[0005] Deep learning methods can learn temporal latent features from AIS trajectories or relative motion data, but the results are usually presented as black-box latent vectors, which are difficult to directly correspond to semantic concepts that navigators can understand. For autonomous ship navigation and risk-assisted decision-making, simply providing classification or prediction results is insufficient; it is also necessary to explain the time stage of risk formation, the main semantic sources, and the corresponding evidence of maneuvering state fluctuations.

[0006] Axiomatic fuzzy set theory possesses the ability to express fuzzy concepts, membership relationships, and logical combinations. However, current applications tend to focus on static evaluation, semantic clustering, or single-moment risk identification, failing to adequately address the unified encoding problem among original features, simple AFS concepts, complex AFS concepts, and neural temporal representations in ship encounter time series. Static AFS clustering alone cannot accurately reflect the phased changes and temporal dependencies of encounter risks; conversely, relying solely on neural networks makes it difficult to generate clear nautical semantic interpretations. Therefore, there is an urgent need for a method that combines temporal representation capabilities with nautical semantic interpretation capabilities. Summary of the Invention

[0007] To address the aforementioned technical challenges in ship encounter situation recognition, such as insufficient semantic representation of original features, reliance on experience in constructing fuzzy concepts, insufficient interpretability of deep models, and unclear boundaries of similar encounter processes, this invention provides an AFS neural coding method and system for time-series feature representation. This invention can transform the original dynamic indicators during ship encounters into simple AFS concepts, and form stable temporal feature representations through a neural time encoder. Simultaneously, it utilizes complex AFS concepts to provide an explanatory chain for the evolution of collision hazards and fluctuations in maneuvering states.

[0008] The technical means employed in this invention are as follows: An AFS neural coding method for time series feature representation includes: S1. Using the ship itself or the ship to be analyzed as a reference, extract the position, speed and heading information of the target ship within a continuous time window from AIS, radar, shipborne navigation sensor or multi-source fusion perception data to form a time series sample of the ship encounter situation. S2. Based on the time series sample of the ship encounter situation, calculate the relative distance, DCPA, TCPA, relative speed, relative approach speed, relative bearing, bearing change rate, distance change rate, DCPA change rate, TCPA change rate, target ship heading change rate, and target ship speed change rate to form the original navigation dynamic index sequence. S3. Based on the empirical distribution of training samples, the monotonic relationship of indicator risk, the support relationship, the weight function and the measure function, the original navigation dynamic indicator sequence is mapped to the AFS simple concept sequence to express the spatial proximity, time urgency, collision risk change trend and maneuvering state fluctuation during the ship encounter process. S4. Based on the simple concept sequence of AFS, complex nautical semantic concepts are constructed through conjunction, disjunction or common constraints in axiomatic fuzzy set logic to describe continuous approach danger, short-term close-range danger, high collision danger accompanied by fluctuations in maneuvering state and the process of collision danger reduction. S5. Input the simple concept sequence of AFS into the neural temporal encoder to learn the temporal dependencies in the encounter process and obtain a fixed-dimensional temporal feature representation of the encounter situation; among them, the complex concepts of AFS are used as auxiliary semantic features to participate in the weighted fusion input. S6. Based on the temporal characteristics of the encounter situation, output the encounter situation category, collision hazard evolution identification result or key target of concern auxiliary discrimination result, and simultaneously output AFS simple concept activation information, AFS complex concept activation information and main semantic interpretation chain.

[0009] Further, step S1 includes: S11. Obtain the ship's current time. Location ,speed and heading and the The target ship at the moment Location ,speed and heading ; S12. Define the relative position vector of the target ship with respect to the ship itself, as follows:

[0010] in, Indicates the relative position of the target vessel with respect to this vessel; S13. Define the relative velocity vector of the target ship with respect to the ship itself, as follows:

[0011] in, Indicates the relative speed of the target ship with respect to the ship itself; S14. Extract the continuous encounter process between the target ship and the ship using a fixed time window, and construct a time series sample of the ship encounter situation, as follows:

[0012] in, Indicates the first A time series sample of ship encounter situations Indicates the first A sample of encounters at time The relative motion state.

[0013] Further, step S2 includes: S21. Calculate the relative distance between the target ship and the ship based on the relative position vector. The calculation formula is as follows:

[0014] in, Indicates the target ship At the same time as this ship The relative distance between them; S22. Calculate the closest encounter time between the target ship and the ship based on the relative position vector and the relative velocity vector. The calculation formula is as follows:

[0015] in, Indicates the target ship At the same time as this ship The corresponding most likely time to meet, To prevent extremely small positive numbers with a denominator of zero, Represents a relative position vector Transpose of; This indicates that the larger of 0 and the value within the parentheses is taken, ensuring that the result is non-negative; S23. Calculate the nearest encounter distance between the target ship and the ship based on the relative position vector, relative velocity vector, and nearest encounter time. The calculation formula is as follows:

[0016] in, Indicates the target ship At the same time as this ship The closest possible distance; S24. Calculate the relative velocity between the target ship and the ship based on the relative velocity vector, using the following formula:

[0017] in, Indicates the target ship At the same time as this ship The relative velocity; S25. Calculate the relative approach speed between the target ship and the ship based on the relative position vector and the relative velocity vector, using the following formula:

[0018] in, Indicates the target ship At the same time as this ship The relative approach speed is used to characterize how quickly the relative distance between the target ship and the ship decreases. S26. Calculate the relative azimuth angle between the target ship and the ship based on the relative position vector, using the following formula:

[0019] in, Indicates the target ship Compared to the ship's time The relative azimuth angle, Represents the arctangent function in the four quadrants. Represents a relative position vector of Axial components, Represents a relative position vector of Axial components; S27. Calculate the rate of change of relative azimuth based on the relative azimuth angle, using the following formula:

[0020] in, Indicates the target ship Compared to the ship's time The rate of change of azimuth, Indicates the target ship Compared to the ship's time The relative azimuth angle, This indicates that the angle difference is constrained to... Angle wrap function within range, Indicates the time interval between adjacent sampling times; S28. Calculate the first-order rates of change of the relative distance, the nearest meeting distance, and the nearest meeting time, respectively, using the following formulas:

[0021]

[0022]

[0023] In the above formula, Indicates the target ship At the same time as this ship The rate of change of distance, Indicates the target ship At the same time as this ship The relative distance between them; Indicates the target ship At the same time as this ship The rate of change of DCPA Indicates the target ship At the same time as this ship The nearest meeting distance between them; Indicates the target ship At the same time as this ship TCPA change rate, Indicates the target ship At the same time as this ship The most recent meeting time between them; Indicates the time interval between adjacent sampling times; S29. Calculate the first-order rate of change of the target ship's course based on the target ship's course, using the following formula:

[0024] in, Indicates the target ship At any moment rate of change of course, Indicates the target ship At any moment The heading angle, This indicates that the angle difference is constrained to... Angle wrap function within range, Indicates the time interval between adjacent sampling times; S210. Calculate the first-order rate of change of the target ship's speed based on the target ship's speed, using the following formula:

[0025] in, Indicates the target ship At any moment The rate of change of speed, Indicates the target ship At any moment speed, Indicates the time interval between adjacent sampling times; S211. Based on relative distance, nearest encounter distance, nearest encounter time, relative approach speed, relative speed, rate of change of bearing, rate of change of distance, rate of change of DCPA, rate of change of TCPA, rate of change of target vessel's course, and rate of change of target vessel's speed, construct the original navigation dynamic feature vector to form the original navigation dynamic index sequence, as follows:

[0026]

[0027] in, Indicates the first A sample of encounters at time The original nautical dynamic characteristics.

[0028] Further, step S3 includes: S31. During the training phase, let all temporal state points in the training samples constitute the training universe of discourse, as follows:

[0029] in, Indicates the first training samples at time 1 state point, Indicates the number of sampling points within the time window; S32. Define a simple set of AFS concepts as follows:

[0030] in, Represents a simple set of concepts in AFS. Indicates the first A simple concept of AFS; S33. Construct concept support strength based on the monotonic relationship between each indicator and the encountered hazard, and define the AFS support relationship between state points with respect to simple concepts, as follows:

[0031] in, State point and A set of AFS support relationships relating to simple AFS concepts. This illustrates the basic concept of AFS. Represents a simple set of concepts in AFS. Represents the training domain Any two state points in the, Represents the training domain, Representing simple concepts Corresponding support relationships; S34. Based on the empirical distribution and support relationships of the training samples, calculate the consistency membership degree of each simple concept at each time step, and convert the original numerical time series into an AFS simple concept sequence. S35. For test samples, concept encoding is performed using only the empirical distribution function and concept support mapping determined during the training phase. Test samples do not participate in parameter fitting.

[0032] Further, step S4 includes: S41. Use min-type common constraints to represent conjunction operations and max-type combination to represent disjunction operations. S42. Complex nautical semantic concepts are generated from multiple simple AFS concepts through conjunction, disjunction or common constraints. Complex concepts are used to express the encounter risk evolution pattern formed by the combined effect of multiple nautical dynamic indicators. S43. Complex concepts in AFS are mainly used to form semantic interpretations and reveal complex encounter risk patterns under the combined effect of multiple simple concepts. Complex concepts do not replace simple concepts in AFS as the input for all classifications.

[0033] Further, step S5 includes: S51, Regarding the first Let there be a sample of encounters, and let it be at time [time]. The simple concept vector of AFS is The input to the neural encoder is defined as follows:

[0034] in, Indicates the first A sample of encounters at time Input the semantic feature vector of the neural temporal encoder, Indicates the number of sampling points within the time window; S52. When it is necessary to introduce complex semantics of AFS, construct a weighted fusion input as follows:

[0035] in, Indicates the first A sample of encounters at time AFS complex semantic vectors, These are the weight coefficients for complex semantics, used to control the degree of influence of complex semantics on the neural encoded input; S53. Using a GRU neural temporal encoder, the temporal dependencies in the encounter process are learned through recursive calculations of gate update, gate reset, and candidate hidden states. The recursive calculation formula is as follows:

[0036]

[0037]

[0038]

[0039] In the above formula, Indicates an update to the door. This indicates that the door is being reset. Indicates the candidate hidden state. Indicates the current hidden state; , , , , , This is the network weight matrix. , , For bias terms; This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. Represents the Hadamard product; S54. The hidden state at the last sampling time is used as the time series representation vector of the encounter sample, as follows:

[0040] in, Indicates the first Fixed-dimensional temporal feature representation of each encounter sample obtained by AFS neural temporal encoding; S55. Input the temporal feature representation vector into the classification layer, and obtain the encounter situation category prediction probability vector through the softmax function, as follows:

[0041] in, The model represents the first The probability vector of class prediction for each encountered sample. and These are the classification layer weight matrix and the bias term, respectively; S56. Based on the predicted probability vectors and true labels of the categories, train the model using the cross-entropy loss function, where the cross-entropy loss function is as follows:

[0042] in, Indicates the number of training samples. This indicates the number of situation types encountered. Indicates the first The sample belongs to the first The true label of the class, The model predicts the first... The sample belongs to the first The probability of a class.

[0043] Further, step S6 includes: S61. Output the encounter situation category, collision hazard evolution identification result, or key target of concern auxiliary judgment result based on the encounter situation time sequence characteristics. S62, For samples Define simple concepts and important time periods as follows:

[0044] in, Indicates the first The time periods in which simple concepts are significantly activated in a sample This means taking the maximum value over all simple concepts. Indicates the first Each sample at time... For the The membership degree is a simple concept. The activation threshold represents the basic concept of AFS. This represents the number of simple concepts in AFS; S63. Define important time periods with complex semantics as follows:

[0045] in, Indicates the first The time periods in which complex semantics are significantly activated in a sample This indicates taking the maximum value across all complex semantics. Indicates the first Each sample at time... For the Activation values ​​for complex semantics The activation threshold representing the complex semantics of AFS. Indicates the number of complex semantic concepts in AFS; S64. Synchronously outputs the encounter situation identification results, key time periods, main activated concepts, sources of complex semantics, and corresponding original navigation dynamic indicators, forming a complete explanatory chain from the original navigation dynamic indicators to the simple AFS concept, the complex AFS concept, and the identification results.

[0046] This invention also provides an AFS neural coding system based on the aforementioned AFS neural coding method for time-series feature representation. This system constructs the ship encounter process as a time-series sample and forms an interpretable temporal feature representation through AFS simple concept encoding, AFS complex semantic construction, and a neural time encoder. The system includes: an encounter situation time-series sample construction layer, a navigation dynamics index calculation layer, an AFS simple concept encoding layer, an AFS complex concept interpretation layer, a neural time coding layer, and a recognition and interpretation output layer, wherein: The encounter situation time series sample construction layer is used to extract the position, speed and heading information of the target ship within a continuous time window from AIS, radar, shipborne navigation sensor or multi-source fusion perception data, taking the ship itself or the ship to be analyzed as the reference object, to form a ship encounter situation time series sample. The navigation dynamic index calculation layer is used to calculate relative distance, DCPA, TCPA, relative speed, relative approach speed, relative bearing, bearing change rate, distance change rate, DCPA change rate, TCPA change rate, target ship heading change rate, and target ship speed change rate based on the time series sample of the ship encounter situation, forming the original navigation dynamic index sequence. The AFS simple concept encoding layer is used to map the original maritime dynamic indicator sequence into an AFS simple concept sequence based on the empirical distribution of training samples, the monotonic relationship of indicator risk, the support relationship, the weight function and the measure function, so as to express the spatial proximity, time urgency, collision risk change trend and maneuvering state fluctuation during the ship encounter process. The AFS complex concept interpretation layer is used to construct complex nautical semantic concepts based on the AFS simple concept sequence through conjunction, disjunction or common constraints in axiomatic fuzzy set logic, in order to describe continuous approach danger, short-term close-range danger, high collision danger accompanied by fluctuations in maneuvering state and collision danger reduction process. The neural temporal coding layer is used to input simple AFS concept sequences into the neural temporal encoder to learn the temporal dependencies in the encounter process and obtain a fixed-dimensional temporal feature representation of the encounter situation; wherein, complex AFS concepts participate as auxiliary semantic features in the weighted fusion input. The identification and interpretation output layer is used to output the encounter situation category, collision hazard evolution identification result or key target of concern auxiliary discrimination result based on the encounter situation temporal characteristics, and simultaneously output AFS simple concept activation information, AFS complex concept activation information and main semantic interpretation chain.

[0047] Compared with the prior art, the present invention has the following advantages: 1. This invention expands the object of ship encounters from a single static risk point to a continuous time series segment, enabling the description of the dynamic process of a ship from approaching, closest encounter, moving away, to the reduction of collision risk or fluctuations in maneuvering status. This processing helps to characterize the phased changes in encounter risk and improves the temporal completeness of ship encounter situation analysis.

[0048] 2. This invention constructs an AFS (Advanced Frontier Function) simple concept encoding layer oriented towards the evolution of maritime risks, converting raw indicators such as DCPA (Displacement Directional Approach), TCPA (Transient Coordinated Aspect), relative distance, relative approach speed, relative bearing, rate of change of heading, and rate of change of speed into semantic concepts with maritime meaning. This process reduces the reliance on manually set membership functions and fixed weights, giving the temporal feature representation a clear semantic basis.

[0049] 3. This invention encodes simple concept sequences in AFS using a neural temporal encoder to obtain a fixed-dimensional temporal feature representation of the encounter situation. Compared with neural models that directly input the original indicators, this processing can enhance the semantic representation capability of difficult encounter samples and perturbation conditions, and provide stable input for subsequent encounter situation recognition, collision hazard evolution analysis, and auxiliary identification of key targets of interest.

[0050] 4. This invention further constructs a complex AFS concept to express high-level risk evolution patterns such as continuous approach hazard, short-term close-range hazard, unstable high-risk state, and collision hazard reduction process. This process can form an explanatory chain between the original nautical dynamic indicators, the simple AFS concept, the complex AFS concept, and the identification results, so that the identification results not only have numerical basis, but also explain the causes of risk formation.

[0051] 5. This invention can combine clean testing, disturbance testing, small sample testing, ablation experiments, and quantitative evaluation in the characterization space to verify the effectiveness of the method from four aspects: identification performance, robustness, characterization structure, and semantic interpretation. This method is applicable to tasks such as ship encounter situation classification, collision hazard evolution identification, navigation behavior change identification, critical target of concern auxiliary identification, and navigation risk early warning, and has good engineering application value.

[0052] Based on the above reasons, this invention can be widely applied in fields such as maritime traffic situational awareness, ship traffic monitoring, navigation risk early warning, navigation auxiliary decision-making, and autonomous navigation systems. Attached Figure Description

[0053] 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.

[0054] Figure 1 This is a technical roadmap for the method of the present invention. Detailed Implementation

[0055] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0056] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product or device.

[0057] like Figure 1 As shown, this invention provides an AFS neural coding method for time series feature representation, comprising: S1. Using the ship itself or the ship to be analyzed as a reference, extract the position, speed and heading information of the target ship within a continuous time window from AIS, radar, shipborne navigation sensor or multi-source fusion perception data to form a time series sample of the ship encounter situation. S2. Based on the time series sample of the ship encounter situation, calculate the relative distance, DCPA, TCPA, relative speed, relative approach speed, relative bearing, bearing change rate, distance change rate, DCPA change rate, TCPA change rate, target ship heading change rate, and target ship speed change rate to form the original navigation dynamic index sequence. S3. Based on the empirical distribution of training samples, the monotonic relationship of indicator risk, the support relationship, the weight function and the measure function, the original navigation dynamic indicator sequence is mapped to the AFS simple concept sequence to express the spatial proximity, time urgency, collision risk change trend and maneuvering state fluctuation during the ship encounter process. S4. Based on the simple concept sequence of AFS, complex nautical semantic concepts are constructed through conjunction, disjunction or common constraints in axiomatic fuzzy set logic to describe continuous approach danger, short-term close-range danger, high collision danger accompanied by fluctuations in maneuvering state and the process of collision danger reduction. S5. Input the simple concept sequence of AFS into the neural temporal encoder to learn the temporal dependencies in the encounter process and obtain a fixed-dimensional temporal feature representation of the encounter situation; among them, the complex concepts of AFS are used as auxiliary semantic features to participate in the weighted fusion input. S6. Based on the temporal characteristics of the encounter situation, output the encounter situation category, collision hazard evolution identification result or key target of concern auxiliary discrimination result, and simultaneously output AFS simple concept activation information, AFS complex concept activation information and main semantic interpretation chain.

[0058] In a specific implementation, as a preferred embodiment of the present invention, step S1 includes: S11. Obtain the ship's current time. Location ,speed and heading and the The target ship at the moment Location ,speed and heading In this embodiment, when the ship is navigating in open water, it obtains information on the position, speed, and course of surrounding target ships through AIS, radar, or other sensing systems.

[0059] S12. Define the relative position vector of the target ship with respect to the ship itself, as follows:

[0060] in, Indicates the relative position of the target vessel with respect to this vessel; S13. Define the relative velocity vector of the target ship with respect to the ship itself, as follows:

[0061] in, Indicates the relative speed of the target ship with respect to the ship itself; S14. Extract the continuous encounter process between the target ship and the ship using a fixed time window, and construct a time series sample of the ship encounter situation, as follows:

[0062] in, Indicates the first A time series sample of ship encounter situations Indicates the first A sample of encounters at time The relative motion state. In this embodiment, the ship encounter situation time series sample is not a single moment of risk point, but a continuous time series segment including the approach, closest encounter, departure, collision avoidance actions or maneuvering fluctuation process between the target ship and the ship.

[0063] Unlike static risk assessments of target ships at a single moment, this embodiment defines a sample as a complete encounter process. This process may include typical situations such as safe passage, stable approach, close crossing, maneuver mitigation, and maneuver instability. For boundary samples, similar encounter processes between safe passage and close crossing, stable approach and maneuver mitigation, and maneuver mitigation and maneuver instability can be further constructed to test the method's ability to represent time series of difficult samples.

[0064] In a specific implementation, as a preferred embodiment of the present invention, step S2 includes: S21. Calculate the relative distance between the target ship and the ship based on the relative position vector. The calculation formula is as follows:

[0065] in, Indicates the target ship At the same time as this ship The relative distance between them; S22. Calculate the closest encounter time between the target ship and the ship based on the relative position vector and the relative velocity vector. The calculation formula is as follows:

[0066] in, Indicates the target ship At the same time as this ship The corresponding most likely time to meet, To prevent extremely small positive numbers with a denominator of zero, Represents a relative position vector Transpose of; This indicates that the larger of 0 and the value within the parentheses is taken, ensuring that the result is non-negative; S23. Calculate the nearest encounter distance between the target ship and the ship based on the relative position vector, relative velocity vector, and nearest encounter time. The calculation formula is as follows:

[0067] in, Indicates the target ship At the same time as this ship The closest possible distance; S24. Calculate the relative velocity between the target ship and the ship based on the relative velocity vector, using the following formula:

[0068] in, Indicates the target ship At the same time as this ship The relative velocity; S25. Calculate the relative approach speed between the target ship and the ship based on the relative position vector and the relative velocity vector, using the following formula:

[0069] in, Indicates the target ship At the same time as this ship The relative approach speed is used to characterize how quickly the relative distance between the target ship and the ship decreases. S26. Calculate the relative azimuth angle between the target ship and the ship based on the relative position vector, using the following formula:

[0070] in, Indicates the target ship Compared to the ship's time The relative azimuth angle, Represents the arctangent function in the four quadrants. Represents a relative position vector of Axial components, Represents a relative position vector of Axial components; S27. Calculate the rate of change of relative azimuth based on the relative azimuth angle, using the following formula:

[0071] in, Indicates the target ship Compared to the ship's time The rate of change of azimuth, Indicates the target ship Compared to the ship's time The relative azimuth angle, This indicates that the angle difference is constrained to... Angle wrap function within range, Indicates the time interval between adjacent sampling times; S28. Calculate the first-order rates of change of the relative distance, the nearest meeting distance, and the nearest meeting time, respectively, using the following formulas:

[0072]

[0073]

[0074] In the above formula, Indicates the target ship At the same time as this ship The rate of change of distance, Indicates the target ship At the same time as this ship The relative distance between them; Indicates the target ship At the same time as this ship The rate of change of DCPA Indicates the target ship At the same time as this ship The nearest meeting distance between them; Indicates the target ship At the same time as this ship TCPA change rate, Indicates the target ship At the same time as this ship The most recent meeting time between them; Indicates the time interval between adjacent sampling times; S29. Calculate the first-order rate of change of the target ship's course based on the target ship's course, using the following formula:

[0075] in, Indicates the target ship At any moment rate of change of course, Indicates the target ship At any moment The heading angle, This indicates that the angle difference is constrained to... Angle wrap function within range, Indicates the time interval between adjacent sampling times; S210. Calculate the first-order rate of change of the target ship's speed based on the target ship's speed, using the following formula:

[0076] in, Indicates the target ship At any moment The rate of change of speed, Indicates the target ship At any moment speed, Indicates the time interval between adjacent sampling times; S211. Based on relative distance, nearest encounter distance, nearest encounter time, relative approach speed, relative speed, rate of change of bearing, rate of change of distance, rate of change of DCPA, rate of change of TCPA, rate of change of target vessel's course, and rate of change of target vessel's speed, construct the original navigation dynamic feature vector to form the original navigation dynamic index sequence, as follows:

[0077]

[0078] in, Indicates the first A sample of encounters at time The original nautical dynamic characteristics. In this embodiment, the above indicators respectively reflect spatial proximity, encounter urgency, relative motion changes, and sources of maneuvering fluctuations.

[0079] In a specific implementation, as a preferred embodiment of the present invention, step S3 includes: S31. During the training phase, let all temporal state points in the training samples constitute the training universe of discourse, as follows:

[0080] in, Indicates the first training samples at time 1 state point, Indicates the number of sampling points within the time window; S32. Define a simple set of AFS concepts as follows:

[0081] in, AFS represents a simple set of concepts, including close proximity, small DCPA, short TCPA, high relative speed, high relative speed, drastic changes in bearing, increased collision risk, decreased collision risk, stable motion, strong relative motion conflict, significant fluctuations in heading, significant fluctuations in speed, and significant fluctuations in control status. Indicates the first A simple AFS concept; in this embodiment, the AFS simple concept encoding is based on the empirical distribution and support relationship obtained from the training samples to determine the parameters, and the test samples are only encoded based on the mapping relationship determined in the training phase to avoid the test data from participating in parameter fitting.

[0082] S33. Construct concept support strength based on the monotonic relationship between each indicator and the encountered hazard, and define the AFS support relationship between state points with respect to simple concepts, as follows:

[0083] in, State point and A set of AFS support relationships relating to simple AFS concepts. This illustrates the basic concept of AFS. Represents a simple set of concepts in AFS. Represents the training domain Any two state points in the, Represents the training domain, Representing simple concepts Corresponding support relationships; S34. Based on the empirical distribution and support relationships of the training samples, calculate the consistency membership degree of each simple concept at each time step, and convert the original numerical time series into an AFS simple concept sequence. S35. For test samples, concept encoding is performed using only the empirical distribution function and concept support mapping determined during the training phase. Test samples do not participate in parameter fitting.

[0084] In a specific implementation, as a preferred embodiment of the present invention, step S4 includes: S41. Use min-type common constraints to represent conjunction operations and max-type combination to represent disjunction operations. S42. Complex nautical semantic concepts are generated from multiple simple AFS concepts through conjunction, disjunction or common constraints. Complex concepts are used to express the encounter risk evolution pattern formed by the combined effect of multiple nautical dynamic indicators. S43. The AFS complex concept is mainly used to form semantic interpretations, revealing complex encounter risk patterns under the combined effect of multiple simple concepts. The complex concept does not replace the simple AFS concepts as the sole classification input. In this embodiment, the AFS complex concept is generated by logically combining multiple simple AFS concepts to express the encounter risk evolution pattern formed by the combined effect of multiple maritime dynamic indicators. The AFS complex concept does not replace the simple AFS concepts as the sole classification input, but rather participates in neural temporal encoding as a high-level semantic interpretation or auxiliary feature.

[0085] In a specific implementation, as a preferred embodiment of the present invention, step S5 includes: S51, Regarding the first Let there be a sample of encounters, and let it be at time [time]. The simple concept vector of AFS is The input to the neural encoder is defined as follows:

[0086] in, Indicates the first A sample of encounters at time Input the semantic feature vector of the neural temporal encoder, Indicates the number of sampling points within the time window; S52. When it is necessary to introduce complex semantics of AFS, construct a weighted fusion input as follows:

[0087] in, Indicates the first A sample of encounters at time AFS complex semantic vectors, This is the weight coefficient for complex semantics, used to control the degree of influence of complex semantics on the neural encoding input; in this embodiment, when the value is small, the model mainly uses simple concept representation; When the value is large, the influence of complex semantics on the encoding result is amplified. This setting can prevent complex semantics from excessively interfering with fine-grained dynamic information in simple concepts.

[0088] S53. Using a GRU neural temporal encoder, the temporal dependencies in the encounter process are learned through recursive calculations of gate update, gate reset, and candidate hidden states. The recursive calculation formula is as follows:

[0089]

[0090]

[0091]

[0092] In the above formula, Indicates an update to the door. This indicates that the door is being reset. Indicates the candidate hidden state. Indicates the current hidden state; , , , , , This is the network weight matrix. , , For bias terms; This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. Represents the Hadamard product; S54. The hidden state at the last sampling time is used as the time series representation vector of the encounter sample, as follows:

[0093] in, Indicates the first A fixed-dimensional temporal feature representation is obtained by encoding an encounter sample using AFS neural time coding. In this embodiment, the representation can be used for encounter situation classification, collision hazard evolution identification, or key target discrimination.

[0094] S55. Input the temporal feature representation vector into the classification layer, and obtain the encounter situation category prediction probability vector through the softmax function, as follows:

[0095] in, The model represents the first The probability vector of class prediction for each encountered sample. and These are the classification layer weight matrix and the bias term, respectively; S56. Based on the predicted probability vectors and true labels of the categories, train the model using the cross-entropy loss function, where the cross-entropy loss function is as follows:

[0096] in, Indicates the number of training samples. This indicates the number of situation types encountered. Indicates the first The sample belongs to the first The true label of the class, The model predicts the first... The sample belongs to the first The probability of a class.

[0097] To verify the impact of different coding structures on the ability to represent temporal features, this embodiment sets up four control models: Raw-GRU, AFS-basic-GRU, AFS-compound-GRU, and FullAFS-GRU. Each of the four models uses a different input feature sequence and obtains the corresponding category prediction results through a GRU neural temporal encoder.

[0098] The Raw-GRU model uses raw nautical dynamics index sequences as input to test the direct time-series representation capability of raw nautical dynamics indexes. Its model definition is as follows:

[0099] The AFS-basic-GRU model takes AFS simple concept sequences as input and is used to test the enhancing effect of AFS simple concept encoding on temporal feature representation. Its model definition is as follows:

[0100] The AFS-compound-GRU model takes AFS complex semantic sequences as input and is used to test the representation effect of AFS complex semantics as input alone. Its model definition is as follows:

[0101] The Full AFS-GRU model takes the fusion features of simple concept sequences and weighted complex semantic sequences from the AFS as input to examine the impact of the fusion input of simple concepts and complex semantics on recognition performance and semantic interpretation ability. Its model definition is as follows:

[0102] In the above formula, Indicates the first A sample of encounters at time The original nautical dynamic index vector; This represents the corresponding AFS simple concept vector; This represents the corresponding AFS complex semantic vector; It represents a fusion input of simple concepts and weighted complex semantics; Indicates the weight coefficients for complex semantics; , , and These represent the trainable parameters for the four models.

[0103] By using the above comparative model, we can analyze the contributions of original indicators, simple concepts, complex semantics, and fused features to the ability to represent ship encounter situations over time series. Overall, AFS simple concept encoding mainly serves a discriminative representation role, while AFS complex semantics primarily provides high-level semantic explanations or auxiliary features.

[0104] In a specific implementation, as a preferred embodiment of the present invention, step S6 includes: S61. Output the encounter situation category, collision hazard evolution identification result, or key target of concern auxiliary judgment result based on the encounter situation time sequence characteristics. S62, For samples Define simple concepts and important time periods as follows:

[0105] in, Indicates the first The time periods in which simple concepts are significantly activated in a sample This means taking the maximum value over all simple concepts. Indicates the first Each sample at time... For the The membership degree is a simple concept. The activation threshold represents the basic concept of AFS. This represents the number of simple concepts in AFS; S63. Define important time periods with complex semantics as follows:

[0106] in, Indicates the first The time periods in which complex semantics are significantly activated in a sample This indicates taking the maximum value across all complex semantics. Indicates the first Each sample at time... For the Activation values ​​for complex semantics The activation threshold representing the complex semantics of AFS. Indicates the number of complex semantic concepts in AFS; S64. Synchronously outputs the encounter situation identification results, key time periods, main activated concepts, sources of complex semantics, and corresponding original navigation dynamic indicators, forming a complete explanatory chain from the original navigation dynamic indicators to the simple AFS concept, the complex AFS concept, and the identification results.

[0107] This invention, through the aforementioned steps, forms a complete technical solution encompassing the construction of ship encounter situation time-series samples, calculation of nautical dynamic indicators, encoding of simple AFS concepts, interpretation of complex AFS semantics, neural temporal encoding representation, and output of recognition results and semantic interpretation. This method maps continuous ship encounter processes into AFS semantic time sequences with nautical meaning and utilizes a neural temporal encoder to extract representation vectors that reflect the evolutionary characteristics of the encounter situation. This method improves the objectivity, stability, and interpretability of ship encounter situation classification, collision hazard evolution identification, critical target discrimination, and navigation risk warning, providing interpretable situational feature inputs for maritime traffic situation awareness, ship traffic monitoring, navigational decision support, and autonomous navigation systems.

[0108] This invention also provides an AFS neural coding system based on the aforementioned AFS neural coding method for time-series feature representation. This system constructs ship encounter processes as time-series samples and forms interpretable temporal feature representations through AFS simple concept coding, AFS complex semantic construction, and a neural time encoder. The system includes: an encounter situation time-series sample construction layer, a navigation dynamics index calculation layer, an AFS simple concept coding layer, an AFS complex concept interpretation layer, a neural time coding layer, and a recognition and interpretation output layer, wherein: The encounter situation time series sample construction layer is used to extract the position, speed and heading information of the target ship within a continuous time window from AIS, radar, shipborne navigation sensor or multi-source fusion perception data, taking the ship itself or the ship to be analyzed as the reference object, to form a ship encounter situation time series sample. The navigation dynamic index calculation layer is used to calculate relative distance, DCPA, TCPA, relative speed, relative approach speed, relative bearing, bearing change rate, distance change rate, DCPA change rate, TCPA change rate, target ship heading change rate, and target ship speed change rate based on the time series sample of the ship encounter situation, forming the original navigation dynamic index sequence. The AFS simple concept encoding layer is used to map the original maritime dynamic indicator sequence into an AFS simple concept sequence based on the empirical distribution of training samples, the monotonic relationship of indicator risk, the support relationship, the weight function and the measure function, so as to express the spatial proximity, time urgency, collision risk change trend and maneuvering state fluctuation during the ship encounter process. The AFS complex concept interpretation layer is used to construct complex nautical semantic concepts based on the AFS simple concept sequence through conjunction, disjunction or common constraints in axiomatic fuzzy set logic, in order to describe continuous approach danger, short-term close-range danger, high collision danger accompanied by fluctuations in maneuvering state and collision danger reduction process. The neural temporal encoding layer is used to input the AFS simple concept sequence into the neural temporal encoder to learn the temporal dependencies in the encounter process and obtain a fixed-dimensional temporal feature representation of the encounter situation. Among them, the AFS complex concept is used as an auxiliary semantic feature to participate in the weighted fusion input. In this embodiment, the neural temporal encoding layer encodes the AFS simple concept sequence into a fixed-dimensional feature vector, which is used for encounter situation classification, collision hazard evolution identification, or key target discrimination.

[0109] The identification and interpretation output layer is used to characterize and output the encounter situation category, collision hazard evolution identification result, or key target of interest auxiliary discrimination result based on the temporal characteristics of the encounter situation, and simultaneously outputs AFS simple concept activation information, AFS complex concept activation information, and main semantic interpretation chain. In this embodiment, while outputting the encounter situation identification result, the identification and interpretation output layer also outputs the key time period, main activated concepts, complex semantic sources, and corresponding original nautical dynamic indicators.

[0110] The embodiments of the present invention are described simply because they correspond to those in the embodiments above. For any similarities, please refer to the descriptions in the embodiments above, which will not be elaborated here.

[0111] In summary, this invention addresses the problems in ship encounter situation recognition, such as insufficient semantic expression of original features, reliance on experience in the construction of fuzzy concepts, insufficient interpretability of deep models, and unclear boundaries of similar encounter processes. It constructs a technical solution consisting of an encounter situation temporal sample construction layer, a navigation dynamic index calculation layer, an AFS simple concept encoding layer, an AFS complex concept interpretation layer, a neural time encoding layer, and an identification and interpretation output layer.

[0112] This invention first uses the ship itself or the ship to be analyzed as a reference object, and extracts the target ship's position, speed, and heading information within a continuous time window from AIS, radar, shipborne navigation sensors, or other sensing data to form a time series segment of the ship encounter situation. Each sample is no longer represented as a single moment of risk point, but rather as a time series including processes such as approach, closest encounter, departure, collision avoidance actions, or maneuvering fluctuations.

[0113] This invention further calculates nautical dynamic indicators such as relative distance, DCPA, TCPA, relative speed, relative approach speed, bearing change rate, distance change rate, DCPA change rate, TCPA change rate, target ship course change rate, target ship speed change rate, and evidence of maneuvering fluctuations to form an original sequence of nautical dynamic indicators, which is used to describe the spatial proximity, time urgency, risk change trend, and source of maneuvering fluctuations of the risks encountered by the ship.

[0114] To address the issue that traditional fuzzy evaluation methods require manual setting of membership functions, risk boundaries, and indicator weights, this invention introduces axiomatic fuzzy set theory. Based on the empirical distribution of training samples and the monotonic relationship of indicator risks, an AFS simple concept set is constructed, and the consistent membership degree of each simple concept at each time step is calculated, thereby transforming the original numerical time series into an AFS simple concept sequence with nautical implications.

[0115] Based on the simple concept encoding results of AFS, this invention constructs complex AFS concepts through conjunction, disjunction, or common constraints in axiomatic fuzzy set logic. These complex concepts are used to express high-level risk evolution patterns such as continuous approach hazard, short-term close-range hazard, unstable high-risk state, and collision hazard reduction process. The complex AFS concepts are mainly used to form semantic interpretations, revealing complex encounter risk patterns under the combined effect of multiple simple concepts.

[0116] This invention inputs a sequence of simple AFS concepts into a GRU or other neural temporal encoder to learn the temporal dependencies of the encounter process and output a fixed-dimensional temporal feature representation. If needed, complex AFS concepts can also be used as auxiliary input in a weighted manner, but complex AFS concepts do not replace simple AFS concepts. This layer implements the encoding from AFS semantic concept sequences to neural temporal representation vectors.

[0117] This invention classifies encounter situations, identifies collision hazard evolution, or assists in the discrimination of key targets of concern based on neural coding representations, and simultaneously outputs AFS simple concept activation, AFS complex concept activation, and the main explanatory chain. The output results include not only the identified category, but also the risk stage on which the category is based, evidence of manipulated fluctuations, and the main AFS semantic sources.

[0118] This invention, through the above processing, forms a complete technical solution from the original ship encounter time series, nautical dynamic indicators, simple AFS concept, complex AFS concept to neural temporal coding representation. This method can improve the interpretability of nautical semantics in temporal feature representation while maintaining recognition performance, avoiding reliance solely on a single risk value or black-box latent vector for risk assessment.

[0119] 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. An AFS neural coding method for time series feature representation, characterized in that, include: S1. Using the ship itself or the ship to be analyzed as a reference, extract the position, speed and heading information of the target ship within a continuous time window from AIS, radar, shipborne navigation sensor or multi-source fusion perception data to form a time series sample of the ship encounter situation. S2. Based on the time series sample of the ship encounter situation, calculate the relative distance, DCPA, TCPA, relative speed, relative approach speed, relative bearing, bearing change rate, distance change rate, DCPA change rate, TCPA change rate, target ship heading change rate, and target ship speed change rate to form the original navigation dynamic index sequence. S3. Based on the empirical distribution of training samples, the monotonic relationship of indicator risk, the support relationship, the weight function and the measure function, the original navigation dynamic indicator sequence is mapped to the AFS simple concept sequence to express the spatial proximity, time urgency, collision risk change trend and maneuvering state fluctuation during the ship encounter process. S4. Based on the simple concept sequence of AFS, complex nautical semantic concepts are constructed through conjunction, disjunction or common constraints in axiomatic fuzzy set logic to describe continuous approach danger, short-term close-range danger, high collision danger accompanied by fluctuations in maneuvering state and the process of collision danger reduction. S5. Input the simple concept sequence of AFS into the neural temporal encoder to learn the temporal dependencies in the encounter process and obtain a fixed-dimensional temporal feature representation of the encounter situation; among them, the complex concepts of AFS are used as auxiliary semantic features to participate in the weighted fusion input. S6. Based on the temporal characteristics of the encounter situation, output the encounter situation category, collision hazard evolution identification result or key target of concern auxiliary discrimination result, and simultaneously output AFS simple concept activation information, AFS complex concept activation information and main semantic interpretation chain.

2. The AFS neural coding method for time-series feature representation according to claim 1, characterized in that, Step S1 includes: S11. Obtain the ship's current time. Location ,speed and heading and the The target ship at the moment Location ,speed and heading ; S12. Define the relative position vector of the target ship with respect to the ship itself, as follows: in, Indicates the relative position of the target vessel with respect to this vessel; S13. Define the relative velocity vector of the target ship with respect to the ship itself, as follows: in, Indicates the relative speed of the target ship with respect to the ship itself; S14. Extract the continuous encounter process between the target ship and the ship using a fixed time window, and construct a time series sample of the ship encounter situation, as follows: in, Indicates the first A time series sample of ship encounter situations Indicates the first A sample of encounters at time The relative motion state.

3. The AFS neural coding method for time series feature representation according to claim 1, characterized in that, Step S2 includes: S21. Calculate the relative distance between the target ship and the ship based on the relative position vector. The calculation formula is as follows: in, Indicates the target ship At the same time as this ship The relative distance between them; S22. Calculate the closest encounter time between the target ship and the ship based on the relative position vector and the relative velocity vector. The calculation formula is as follows: in, Indicates the target ship At the same time as this ship The corresponding most likely time to meet, To prevent extremely small positive numbers with a denominator of zero, Represents a relative position vector Transpose of; This indicates that the larger of 0 and the value within the parentheses is taken, ensuring that the result is non-negative; S23. Calculate the nearest encounter distance between the target ship and the ship based on the relative position vector, relative velocity vector, and nearest encounter time. The calculation formula is as follows: in, Indicates the target ship At the same time as this ship The closest possible distance; S24. Calculate the relative velocity between the target ship and the ship based on the relative velocity vector, using the following formula: in, Indicates the target ship At the same time as this ship The relative velocity; S25. Calculate the relative approach speed between the target ship and the ship based on the relative position vector and the relative velocity vector, using the following formula: in, Indicates the target ship At the same time as this ship The relative approach speed is used to characterize how quickly the relative distance between the target ship and the ship decreases. S26. Calculate the relative azimuth angle between the target ship and the ship based on the relative position vector, using the following formula: in, Indicates the target ship Compared to the ship's time The relative azimuth angle, Represents the arctangent function in the four quadrants. Represents a relative position vector of Axial components, Represents a relative position vector of Axial components; S27. Calculate the rate of change of relative azimuth based on the relative azimuth angle, using the following formula: in, Indicates the target ship Compared to the ship's time The rate of change of azimuth, Indicates the target ship Compared to the ship's time The relative azimuth angle, This indicates that the angle difference is constrained to... Angle wrap function within range, Indicates the time interval between adjacent sampling times; S28. Calculate the first-order rates of change of the relative distance, the nearest meeting distance, and the nearest meeting time, respectively, using the following formulas: In the above formula, Indicates the target ship At the same time as this ship The rate of change of distance, Indicates the target ship At the same time as this ship The relative distance between them; Indicates the target ship At the same time as this ship The rate of change of DCPA Indicates the target ship At the same time as this ship The nearest meeting distance between them; Indicates the target ship At the same time as this ship TCPA change rate, Indicates the target ship At the same time as this ship The most recent meeting time between them; Indicates the time interval between adjacent sampling times; S29. Calculate the first-order rate of change of the target ship's course based on the target ship's course, using the following formula: in, Indicates the target ship At any moment rate of change of course, Indicates the target ship At any moment The heading angle, This indicates that the angle difference is constrained to... Angle wrap function within range, Indicates the time interval between adjacent sampling times; S210. Calculate the first-order rate of change of the target ship's speed based on the target ship's speed, using the following formula: in, Indicates the target ship At any moment The rate of change of speed, Indicates the target ship At any moment speed, Indicates the time interval between adjacent sampling times; S211. Based on relative distance, nearest encounter distance, nearest encounter time, relative approach speed, relative speed, rate of change of bearing, rate of change of distance, rate of change of DCPA, rate of change of TCPA, rate of change of target vessel's course, and rate of change of target vessel's speed, construct the original navigation dynamic feature vector to form the original navigation dynamic index sequence, as follows: in, Indicates the first A sample of encounters at time The original nautical dynamic characteristics.

4. The AFS neural coding method for time series feature representation according to claim 1, characterized in that, Step S3 includes: S31. During the training phase, let all temporal state points in the training samples constitute the training universe of discourse, as follows: in, Indicates the first training samples at time 1 state point, Indicates the number of sampling points within the time window; S32. Define a simple set of AFS concepts as follows: in, Represents a simple set of concepts in AFS. Indicates the first A simple concept of AFS; S33. Construct concept support strength based on the monotonic relationship between each indicator and the encountered hazard, and define the AFS support relationship between state points with respect to simple concepts, as follows: in, State point and A set of AFS support relationships relating to simple AFS concepts. This illustrates the basic concept of AFS. Represents a simple set of concepts in AFS. Represents the training domain Any two state points in the, Represents the training domain, Representing simple concepts Corresponding support relationships; S34. Based on the empirical distribution and support relationships of the training samples, calculate the consistency membership degree of each simple concept at each time step, and convert the original numerical time series into an AFS simple concept sequence. S35. For test samples, concept encoding is performed using only the empirical distribution function and concept support mapping determined during the training phase. Test samples do not participate in parameter fitting.

5. The AFS neural coding method for time series feature representation according to claim 1, characterized in that, Step S4 includes: S41. Use min-type common constraints to represent conjunction operations and max-type combination to represent disjunction operations. S42. Complex nautical semantic concepts are generated from multiple simple AFS concepts through conjunction, disjunction or common constraints. Complex concepts are used to express the encounter risk evolution pattern formed by the combined effect of multiple nautical dynamic indicators. S43. Complex concepts in AFS are mainly used to form semantic interpretations and reveal complex encounter risk patterns under the combined effect of multiple simple concepts. Complex concepts do not replace simple concepts in AFS as the input for all classifications.

6. The AFS neural coding method for time-series feature representation according to claim 1, characterized in that, Step S5 includes: S51, Regarding the first Let there be a sample of encounters, and let it be at time [time]. The simple concept vector of AFS is The input to the neural encoder is defined as follows: in, Indicates the first A sample of encounters at time Input the semantic feature vector of the neural temporal encoder, Indicates the number of sampling points within the time window; S52. When it is necessary to introduce complex semantics of AFS, construct a weighted fusion input as follows: in, Indicates the first A sample of encounters at time AFS complex semantic vectors, These are the weight coefficients for complex semantics, used to control the degree of influence of complex semantics on the neural encoded input; S53. Using a GRU neural temporal encoder, the temporal dependencies in the encounter process are learned through recursive calculations of gate update, gate reset, and candidate hidden states. The recursive calculation formula is as follows: In the above formula, Indicates an update to the door. This indicates that the door is being reset. Indicates the candidate hidden state. Indicates the current hidden state; , , , , , This is the network weight matrix. , , For bias terms; This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. Represents the Hadamard product; S54. The hidden state at the last sampling time is used as the time series representation vector of the encounter sample, as follows: in, Indicates the first Fixed-dimensional temporal feature representation of each encounter sample obtained by AFS neural temporal encoding; S55. Input the temporal feature representation vector into the classification layer, and obtain the encounter situation category prediction probability vector through the softmax function, as follows: in, The model represents the first The probability vector of class prediction for each encountered sample. and These are the classification layer weight matrix and the bias term, respectively; S56. Based on the predicted probability vectors and true labels of the categories, train the model using the cross-entropy loss function, where the cross-entropy loss function is as follows: in, Indicates the number of training samples. This indicates the number of situation types encountered. Indicates the first The sample belongs to the first The true label of the class, The model predicts the first... The sample belongs to the first The probability of a class.

7. The AFS neural coding method for time series feature representation according to claim 1, characterized in that, Step S6 includes: S61. Output the encounter situation category, collision hazard evolution identification result, or key target of concern auxiliary judgment result based on the encounter situation time sequence characteristics. S62, For samples Define simple concepts and important time periods as follows: in, Indicates the first The time periods in which simple concepts are significantly activated in a sample This means taking the maximum value over all simple concepts. Indicates the first Each sample at time... For the The membership degree is a simple concept. The activation threshold represents the basic concept of AFS. This represents the number of simple concepts in AFS; S63. Define important time periods with complex semantics as follows: in, Indicates the first The time periods in which complex semantics are significantly activated in a sample This indicates taking the maximum value across all complex semantics. Indicates the first Each sample at time... For the Activation values ​​for complex semantics The activation threshold representing the complex semantics of AFS. Indicates the number of complex semantic concepts in AFS; S64. Synchronously outputs the encounter situation identification results, key time periods, main activated concepts, sources of complex semantics, and corresponding original navigation dynamic indicators, forming a complete explanatory chain from the original navigation dynamic indicators to the simple AFS concept, the complex AFS concept, and the identification results.

8. An AFS neural coding system for time-series feature representation implemented based on the AFS neural coding method for time-series feature representation as described in any one of claims 1-7, characterized in that, The ship encounter process is constructed as a time series sample, and an interpretable temporal feature representation is formed through AFS simple concept encoding, AFS complex semantic construction, and neural temporal encoder. This includes: an encounter situation time series sample construction layer, a navigation dynamics index calculation layer, an AFS simple concept encoding layer, an AFS complex concept interpretation layer, a neural temporal encoding layer, and a recognition and interpretation output layer. The encounter situation time series sample construction layer is used to extract the position, speed and heading information of the target ship within a continuous time window from AIS, radar, shipborne navigation sensor or multi-source fusion perception data, taking the ship itself or the ship to be analyzed as the reference object, to form a ship encounter situation time series sample. The navigation dynamic index calculation layer is used to calculate relative distance, DCPA, TCPA, relative speed, relative approach speed, relative bearing, bearing change rate, distance change rate, DCPA change rate, TCPA change rate, target ship heading change rate, and target ship speed change rate based on the time series sample of the ship encounter situation, forming the original navigation dynamic index sequence. The AFS simple concept encoding layer is used to map the original maritime dynamic indicator sequence into an AFS simple concept sequence based on the empirical distribution of training samples, the monotonic relationship of indicator risk, the support relationship, the weight function and the measure function, so as to express the spatial proximity, time urgency, collision risk change trend and maneuvering state fluctuation during the ship encounter process. The AFS complex concept interpretation layer is used to construct complex nautical semantic concepts based on the AFS simple concept sequence through conjunction, disjunction or common constraints in axiomatic fuzzy set logic, in order to describe continuous approach danger, short-term close-range danger, high collision danger accompanied by fluctuations in maneuvering state and collision danger reduction process. The neural temporal coding layer is used to input simple AFS concept sequences into the neural temporal encoder to learn the temporal dependencies in the encounter process and obtain a fixed-dimensional temporal feature representation of the encounter situation; wherein, complex AFS concepts participate as auxiliary semantic features in the weighted fusion input. The identification and interpretation output layer is used to output the encounter situation category, collision hazard evolution identification result or key target of concern auxiliary discrimination result based on the encounter situation temporal characteristics, and simultaneously output AFS simple concept activation information, AFS complex concept activation information and main semantic interpretation chain.

Images