Ship collision avoidance method
By combining deep learning models and ship collision avoidance mechanism models, and using discriminators to select the optimal collision avoidance behavior, the problem of poor conformity between mechanism models and navigation practices in the existing technology is solved, and the accuracy and reliability of collision avoidance decisions are improved.
Patent Information
- Application Number
- CN202411446941.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2025-02-11
AI Technical Summary
In the prior art, the mechanism model has poor compliance with navigation practices, and the data-driven model lacks the constraints of knowledge in the field of collision avoidance, resulting in insufficient accuracy and reliability of collision avoidance decisions.
Using a method combining deep learning model and ship collision avoidance mechanism model, the first collision avoidance behavior and the second collision avoidance behavior are determined by training the deep learning model and constructing the mechanism model, and the optimal collision avoidance behavior is selected from it through a discriminator.
The compliance of collision avoidance decisions and navigation practices has been improved, the accuracy and reliability of collision avoidance behavior models have been enhanced, and a ship collision avoidance behavior model has been established with data-mechanism fusion.
Smart Images
Figure CN119418555A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of ship navigation technology, and in particular to a ship collision avoidance method. Background Art
[0002] The study of ship collision avoidance behavior is the basis for safe collision avoidance decision-making.
[0003] In the prior art, traditional collision avoidance behavior models include mechanism models and data-driven models. From the perspective of mechanism models, by analyzing the relative motion relationship between ships, under certain assumptions, a mathematical method is used to describe the mapping relationship between input and output and various parameters. This is generally the optimal solution under the constraints of domain knowledge and many assumptions. This type of method has clear mathematical and physical meanings, and the decision results are relatively credible. However, in complex encounter scenarios, under the coupling of multiple factors such as man-ship-environment, there is actually a certain gap with navigation practice. From the perspective of data-driven, big data integrates navigation practice into the collision avoidance behavior model, but because of the unclear internal mechanism and the lack of constraints from the collision avoidance domain knowledge, the quality of the model is uneven. Summary of the invention
[0004] In view of this, it is necessary to provide a ship collision avoidance method to solve the problems in the existing technology that the mechanism model is poorly consistent with navigation practice and the data-driven model lacks constraints of collision avoidance domain knowledge.
[0005] In order to solve the above problems, the present application provides a ship collision avoidance method, comprising: Determine the first collision avoidance action based on the fully trained deep learning model; Determine the second collision avoidance behavior based on the constructed ship collision avoidance mechanism model; Based on the degree of conformity between the collision avoidance behavior and the crew's avoidance behavior preference and the degree of change in collision risk, a discriminator is constructed; An optimal collision avoidance behavior is selected from the first collision avoidance behavior and the second collision avoidance behavior based on the discriminator.
[0006] In some possible implementations, the process of determining the fully trained deep learning model includes: Obtain target historical traffic environment data; Extracting target feature data based on the target historical traffic environment data; An initial deep learning model is trained based on the target feature data to obtain a fully trained deep learning model.
[0007] In some possible implementations, obtaining target historical traffic environment data includes: Acquiring initial historical traffic environment data of the target area, wherein the initial historical traffic environment data includes ship AIS data, radar data, electronic chart data, hydrological data and meteorological data; Eliminating abnormal data and repairing the track of the ship in the ship AIS data, and cleaning and repairing the radar data, the electronic chart data, the hydrological data and the meteorological data to obtain intermediate historical traffic environment data; The intermediate historical traffic environment data are temporally and spatially matched and fused to obtain target historical traffic environment data.
[0008] In some possible implementations, the TSC-TIP algorithm is used to perform spatiotemporal matching on the intermediate historical traffic environment data, and the weighted and splicing processing method is used to fuse the intermediate historical traffic environment data.
[0009] In some possible implementations, extracting target feature data based on the target historical traffic environment data includes: Extracting first characteristic data from the target historical traffic environment data according to the correlation characteristics between human, ship and environmental parameters and the ship collision avoidance behavior; Performing feature processing on the first feature data to establish a dynamic mapping mechanism between data and features to obtain second feature data; The second feature data are unified to the same scale to obtain intermediate feature data, and target feature data is determined based on the intermediate feature data.
[0010] In some possible implementations, the first feature data includes collision risk, and the collision risk is extracted by the following method: By combining the controller and the ship motion model, a ship motion prediction model is constructed; Based on the ship AIS data in the target historical traffic environment data, a target ship motion prediction model is constructed; Based on the own ship motion prediction model, the target ship motion prediction model, the target historical traffic environment data and the probabilistic speed barrier method, the initial collision risk between the own ship and the target ship is calculated.
[0011] In some possible implementations, feature processing is performed on the first feature data to establish a dynamic mapping mechanism between data and features to obtain second feature data, including: Performing statistics and fitting on the random characteristic data in the first characteristic data to obtain distribution characteristic data; constructing a matrix for the variable-dimensional feature data in the first feature data, calculating matrix characteristic data of each matrix, and determining a target collision risk based on the matrix characteristic data and the initial collision risk; Processing the hierarchical feature data in the first feature data by feature coding to obtain coded data; Building a grid map based on the static obstacle feature data in the first feature data to obtain map data; Second characteristic data is formed based on the distribution characteristic data, the object collision risk, the coding data and the map data.
[0012] In some possible implementations, determining target feature data based on the intermediate feature data includes: Based on the nearest component analysis method, feature selection is performed on the intermediate feature data to obtain target feature data.
[0013] In some possible implementations, the fully trained deep learning model adopts a TimesNet network structure.
[0014] In some possible implementations, determining the second collision avoidance behavior based on the constructed ship collision avoidance mechanism model includes: Obtain the ship's encounter scenario and target's historical traffic environment data; Based on several basic ship collision avoidance mechanism models, a set of ship collision avoidance behaviors in the encounter scenario of the own ship is generated; Based on the clustering of the encounter scene features, extracting the collision avoidance behaviors of the similar encounter scenes similar to the encounter scene of the own ship from the target historical traffic environment data; Based on the degree of conformity between the collision avoidance behavior and the crew's avoidance behavior preference and the degree of change in collision risk, the optimal second collision avoidance behavior is selected from the ship collision avoidance behavior set and the similar encounter scenario collision avoidance behavior.
[0015] The beneficial effect of the present application is that the ship collision avoidance method provided by the present application, by introducing deep learning and discrimination mechanisms, explores the data and mechanism adversarial learning fusion mechanism, takes advantage of each other's strengths and overcomes each other's weaknesses, establishes a data-mechanism fusion ship collision avoidance behavior model, and improves the conformity of collision avoidance decisions with navigation practice. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 A schematic diagram of an embodiment of the ship collision avoidance method provided in this application; Figure 2 A flowchart of an embodiment of a process for determining a fully trained deep learning model provided in the present application; Figure 3 For this application Figure 2 A schematic flow chart of an embodiment of step S201; Figure 4 For this application Figure 2A schematic flow chart of an embodiment of step S202; Figure 5 A schematic diagram of an embodiment of a collision risk determination process provided by the present application; Figure 6 For this application Figure 3 A schematic flow chart of an embodiment of step S402; Figure 7 For this application Figure 1 A flowchart of an embodiment of step S102 in FIG. DETAILED DESCRIPTION
[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.
[0018] It should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowcharts may be implemented out of order, and steps without logical contextual relationships may be reversed in order or implemented simultaneously. In addition, those skilled in the art may add one or more other operations to the flowcharts, or may remove one or more operations from the flowcharts, under the guidance of the content of this application. Some of the block diagrams shown in the accompanying drawings are functional entities, and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor systems and / or microcontroller systems.
[0019] The descriptions of "first", "second", etc. involved in the embodiments of the present application are only used to describe the implicit purpose, and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the technical features defined as "first" and "second" may explicitly or implicitly include at least one of the features. "And / or" describes the association relationship of the associated objects, indicating that three relationships may exist, for example: A and / or B, which can represent: A exists alone, A and B exist at the same time, and B exists alone.
[0020] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.
[0021] The present application provides a ship collision avoidance method, which is described below.
[0022] Figure 1 A schematic diagram of an embodiment of the ship collision avoidance method provided in this application is shown in FIG. Figure 1 As shown, the ship collision avoidance method includes: S101. Determine a first collision avoidance behavior based on a well-trained deep learning model; S102, determining a second collision avoidance behavior based on the constructed ship collision avoidance mechanism model; S103, constructing a discriminator based on the degree of conformity between the collision avoidance behavior and the crew member's avoidance behavior preference and the degree of change in collision risk; S104: Selecting the optimal collision avoidance behavior from the first collision avoidance behavior and the second collision avoidance behavior based on the discriminator.
[0023] Compared with the existing technology, this application introduces deep learning and discrimination mechanisms, explores the data and mechanism adversarial learning fusion mechanism, draws on each other's strengths and overcomes each other's weaknesses, and establishes a data-mechanism fusion ship collision avoidance behavior model, thereby improving the conformity of collision avoidance decisions with navigation practice.
[0024] In order to better train the deep learning model, in some embodiments, such as Figure 2 As shown, the process of determining the fully trained deep learning model in step S101 includes: S201, obtaining target historical traffic environment data; S202, extracting target feature data based on target historical traffic environment data; S203: Train an initial deep learning model based on the target feature data to obtain a fully trained deep learning model.
[0025] In order to obtain better target historical traffic environment data, in some embodiments, such as Figure 3 As shown, step S201 includes: S301, obtaining initial historical traffic environment data of the target area, the initial historical traffic environment data including ship AIS data, radar data, electronic chart data, hydrological and meteorological data; It should be noted that the initial historical traffic environment data is multi-dimensional heterogeneous data ("multi-dimensional heterogeneous data" usually refers to multiple data types from different sources, with different characteristics, but related to the same phenomenon or process); among them, ship AIS data: the automatic identification system (AIS) provides information such as ship position, speed, and heading. This is spatiotemporal data, usually stored in the form of time series, reflecting the dynamic information of the ship; radar data: radar can detect the position and movement trajectory of ships and other objects, and supplement AIS data, especially in areas where AIS signals fail or coverage is insufficient; electronic chart data: provides geographical and channel information of the sea area, including static or semi-static spatial data such as navigation marks, shoals, and routes; hydrological data: including spatiotemporal data related to ocean dynamics such as tides, water speed, and flow direction, which affect the actual navigation status of the ship; meteorological data: such as meteorological information such as wind speed, wind direction, and rainfall, which will affect the environment.
[0026] S302, removing abnormal data and repairing the trajectory of the ship in the ship AIS data, and cleaning and repairing the radar data, electronic chart data, hydrological data and meteorological data to obtain intermediate historical traffic environment data; It should be noted that the identification and elimination of abnormal data adopts clustering algorithm, and the trajectory repair adopts dynamic interpolation method. S303: Perform spatiotemporal matching and fusion on the intermediate historical traffic environment data to obtain target historical traffic environment data.
[0027] It should be noted that the time-space matching adopts the TSC-TIP (Time-Space Convolution and Temporal Interpolation Processing) algorithm. The TSC-TIP algorithm is a time-space convolution and interpolation processing method that focuses on processing multi-dimensional data with inconsistent time and space. The algorithm generally includes the following steps: Time-space alignment: First, align the data according to the timestamp, normalize the data with different sampling frequencies, so that they are comparable in time; Time-space convolution: Map data from different sources to a common time-space scale through time-space convolution operations; Convolution operations help find the correlation between different data in time and space, ensuring that the data can jointly describe the traffic environment at the same time and location; Time interpolation: Missing data at certain time points can be interpolated through dynamic models or based on known data points to generate continuous trajectories or continuous meteorological and hydrological data, which is crucial to improving data integrity. It should also be noted that data fusion uses weighted and splicing processing methods. Weighted and splicing processing is mainly used to fuse data from different sources to ensure that the fused data is more reliable and accurate. The specific practices include: Weighted fusion: For the fusion of multi-source data, different weights can be set according to the reliability of the data, the authority of the data source or its importance to the research problem. For example, AIS data may be more reliable under certain conditions, while radar data may be more comprehensive in certain areas. Weighted fusion can be achieved through methods such as Bayesian estimation or Kalman filtering; Splicing processing: Sometimes, data may be inconsistent in time and space, such as different data sources covering different times or spaces. Through splicing processing, data can be combined into a more complete spatiotemporal sequence in different time periods or regions. This usually relies on splicing rules based on trajectory similarity or spatial position similarity.
[0028] In order to obtain better target feature data, in some embodiments, such as Figure 4 As shown, step S202 includes: In order to provide effective data representation for subsequent collision avoidance behavior modeling, in some embodiments, feature data is extracted from the encounter data to obtain encounter scene feature data, including: S401, extracting first feature data from target historical traffic environment data according to the correlation characteristics between human, ship and environmental parameters and ship collision avoidance behavior; It should be noted that based on the ship encounter data, guided by the knowledge of the "Collision Prevention Regulations" and good seamanship, different levels of multi-source high-dimensional feature data are mined from the three dimensions of people, ships and environment, that is, the first feature data, including the first feature data including crew avoidance preference feature data, ship encounter situation feature data and environmental feature data; among them, the human feature data, that is, the crew avoidance preference characteristics, are analyzed from the two perspectives of avoidance timing and operation preference, and are obtained by counting a large number of ship avoidance timing and avoidance actions; the ship feature data, that is, the ship encounter situation feature data, can be analyzed from different levels such as the ship's own state characteristics, ship geometric relationship characteristics, encounter situation and encounter urgency, and the ship's own state (captain) L , Ship width B , Draft d , Type of Ship S etc.) can be directly extracted through AIS data, and the geometric relationship characteristics of ships (ship encounter situations, relative distances ( dis ), relative position ( b )、Relative speed( v ) etc.) are extracted through AIS data after calculation, and the urgency characteristics are used to adopt collision risk ( risk ) to characterize; environmental characteristic data refers to natural environmental characteristic data, such as wind ( wind ),wave( wave ),flow( current ),visibility( visibility ), static obstacles ( s ) etc., obtained through environmental sensors, electronic nautical charts, etc.
[0029] It should also be noted that according to the Collision Regulations and good seamanship, ship encounter situations are divided into head-on encounter situations, cross encounter situations, overtaking situations and multi-ship encounter situations. The collision avoidance behaviors of ships in different encounter scenarios vary greatly. The encounter situation classification standards are quantified based on domain knowledge, and the encounter situations are identified based on the comparison of mutual side angles.
[0030] S402, performing feature processing on the first feature data to establish a dynamic mapping mechanism between data and features to obtain second feature data; It should be noted that the individual feature data in the first feature data come from different sources and are feature data with different modes, types, and structures. They should be matched with corresponding feature processing methods to establish a dynamic mapping mechanism between data and features.
[0031] S403: Unify the second feature data to the same scale to obtain intermediate feature data, and determine target feature data based on the intermediate feature data.
[0032] The first characteristic data includes an initial collision risk. In order to obtain better first characteristic data, in some embodiments, such as Figure 5 As shown, the process of determining the initial collision risk includes: S501, constructing a ship motion prediction model by combining the controller and the ship motion model; It should be noted that, for the prediction of the ship's motion, the controller is combined with the ship motion model to design the rudder operation model, which continuously updates the real-time motion status of the ship after the execution decision. Through negative feedback data, the control strategy is adjusted to achieve accurate prediction of the ship's motion.
[0033] S502, constructing a target ship motion prediction model based on the ship AIS data in the target historical traffic environment data; It should be noted that for the prediction of target ship motion, based on the basic idea of data-driven, the historical AIS data of ships in the waters are used to mine the route model and rules containing the traffic pattern. The Gaussian process probability is combined to characterize the uncertainty of the target ship's motion such as changing direction and speed, and a target ship motion prediction model is established.
[0034] S503, calculating the initial collision risk between the own ship and the target ship based on the own ship motion prediction model, the target ship motion prediction model, the target historical traffic environment data and the probabilistic speed barrier method.
[0035] It should be noted that the probability obstacle speed set (PVO) and feasible speed set caused by the target ship in the own ship's speed space are calculated according to the probability speed obstacle method, and the ship maneuvering avoidance space is used as an indicator (collision risk = probability speed obstacle set / feasible speed set) to achieve collision risk quantification; the speed obstacle set corresponding to each trajectory is calculated as shown in the following formula:
[0036] In the formula, Representative k The speed obstacle area of the target ship, represent t Moment k The position of the target ship track, Represents a safe area for both; The probability speed obstacle set is as follows:
[0037] In the formula, represents the set of probabilistic speed obstacles, represent k The probability of a trajectory, the sum of all possible trajectory probabilities is 1; The multi-ship speed barrier set is based on two ships, and the speed barrier sets between different ships are probabilized and calculated.
[0038] Since the individual feature data in the first feature data come from different sources and are feature data with different modes, types, and structures, corresponding feature processing means should be matched to establish a dynamic mapping mechanism between data and features. Therefore, in order to obtain better second feature data, in some embodiments, Figure 6 As shown, step S402 includes: S601, performing statistics and fitting on the random feature data in the first feature data to obtain distribution feature data; It should be noted that the distribution characteristic data (including timing distribution and action distribution) are obtained by statistics and fitting of random feature data (including avoidance timing and operation preference, etc.). Since the random feature data has the characteristics of high dimension and high correlation, on the basis of maintaining the function characteristics, before calculating the distribution characteristic data, it is necessary to perform corresponding dimensionality reduction and decorrelation processing on the random feature data, which can be done by using methods such as KL separable sorting method and wavelet.
[0039] S602, constructing a matrix for the variable-dimensional feature data in the first feature data, calculating matrix characteristic data of each matrix, and determining the target collision risk based on the matrix characteristic data and the initial collision risk; It should be noted that as the number of ships increases, the data used to describe the geometric relationship characteristics of the ships will also increase. The collision avoidance relationship between multiple ships is manifested as a complex network in time and space. Therefore, this application introduces spectral graph theory to perform feature processing on variable dimensional feature data such as ship geometric relationship characteristics. The matrix includes relative distance matrix, relative orientation matrix, relative orientation matrix and relative speed matrix. The matrix characteristic data includes matrix eigenvalue, spectral radius and matrix trace. Specifically, the relative distance matrix is constructed respectively. , relative orientation matrix , relative velocity matrix , find the eigenvalues of each matrix , , , using the spectral radius , the trace of the matrix Etc. are used to represent the characteristics of the matrix, and the geometric relationship characteristics of the encountering ships are reconstructed in combination with the collision risk.
[0040] It should also be noted that the collision risk is usually analyzed based on a specific threshold of the geometric relationship. For example, when the relative distance becomes smaller, and the relative speed and relative orientation indicate that the ships are approaching each other, the collision risk increases. Therefore, the key to combining geometric relationships with collision risk lies in how to extract features that are strongly correlated with the collision risk from the geometric elements: ① to construct a collision risk function based on relative distance, relative orientation and relative speed; ② based on the spectral radius or trace Reflecting the global geometric characteristics of the ship encounter system, these eigenvalues can be associated with the collision risk level by comparing with historical collision data or simulated collision processes. For example, a specific spectral radius value or trace value can indicate that there is a highly dangerous encounter mode in the system, indicating that the ship is approaching quickly and has a high risk of collision; ③ By analyzing the eigenvalues of the geometric matrix, select the features most relevant to the collision risk, for example, it may be found that the spectral radius of the relative distance matrix Some eigenvalues of the relative velocity matrix are closely related to the collision risk; ④ The geometric element features can be weighted to give higher weights to the elements that are more relevant to the collision risk; ⑤ By selecting the eigenvalues most relevant to the collision risk, a new geometric feature matrix can be reconstructed.
[0041] S603, processing the hierarchical feature data in the first feature data by feature coding to obtain coded data; It should be noted that the classifiable characteristic data include wind, waves, current, ship length, ship width, etc.
[0042] S604, constructing a grid map based on the static obstacle feature data in the first feature data to obtain map data; It should be noted that due to the irregular graphics of static obstacle feature data, the equivalent grid is used to improve the grid of the obstacle boundary area and construct a grid map to achieve accurate expression of the static obstacles.
[0043] S605: Form second characteristic data based on the distribution characteristic data, the target collision risk, the coding data and the map data.
[0044] In order to reduce data redundancy and improve algorithm learning efficiency, in some embodiments, in step S403, determining target feature data based on intermediate feature data includes: Based on the nearest component analysis method, feature selection is performed on the intermediate feature data to obtain the target feature data.
[0045] It should be noted that the projection matrix is automatically learned through Neighborhood Component Analysis (NCA) to obtain a low-dimensional space that retains global information as much as possible. The ReliefF algorithm is used to calculate the feature weights and analyze the feature contribution. The features are further analyzed in combination with local information to achieve feature selection and analysis and obtain the target feature data.
[0046] In order to build a better deep learning model, in some embodiments, the deep learning model adopts the TimesNet network structure, and the TimesNet network structure is as follows: In the deep learning model network layer, The one-dimensional time series of a layer is as follows:
[0047] In the formula, There are three modules.
[0048] First, the characteristic data is periodized, and the time length is set to T , the variable dimension is C , then the one-dimensional time series , the intensity of each frequency component A As shown below:
[0049] In the formula, FFT represents the Fourier transform of the variable. Amp Indicates the magnitude after transformation; For more significant intensities, their frequencies are extracted to form a new set, namely:
[0050] Each frequency has a corresponding period length, namely:
[0051] In order to prevent edge information loss, the one-dimensional time series needs to be padding Processing, so that the output size is consistent with the original size, and then padding The one-dimensional time series after periodization is reshaped according to the main characteristic law after periodization to obtain a two-dimensional time series, namely:
[0052] Then the typical Inception algorithm structure is used to extract features from the two-dimensional time series, namely:
[0053] Finally, the extracted feature information is transformed into one-dimensional space for information integration, and the
[0054] The deep learning model constructs loss functions based on trajectory similarity from three dimensions: time, space, and category:
[0055] In the formula, is the true trajectory, To predict the trajectory, is the temporal similarity loss metric, is the spatial similarity loss metric, is the category similarity loss metric.
[0056] The time dimension refers to the evolution of the ship's trajectory over time, which may include the ship's status at different points in time, including the timing of avoidance; the spatial dimension usually refers to the position coordinates of the ship at a certain point in time; the category dimension refers to the different types or behavior patterns to which the ship's trajectory belongs, and specific categories may include turning to avoid and slowing down to avoid.
[0057] In order to obtain a better ship collision avoidance mechanism model, in some embodiments, such as Figure 7 As shown, step S102 includes: S701, obtaining the encounter scenario of own ship and historical traffic environment data of target; S702, based on several basic ship collision avoidance mechanism models, generating a set of ship collision avoidance behaviors in a meeting scenario of own ship; It should be noted that the basic ship collision avoidance mechanism model includes a ship collision avoidance mechanism model constructed based on methods such as virtual potential field, maneuvering simulation and speed barrier.
[0058] S703, based on the clustering of the encounter scene features, extracting the collision avoidance behaviors of the similar encounter scenes as the own ship from the target historical traffic environment data; It should be noted that based on the clustering of encounter scene features, encounter scenes with collision risks are screened out from the target historical traffic environment data, the collision avoidance behaviors of ships in similar encounter scenes are analyzed, and their evolution laws are studied.
[0059] S704. Based on the degree of conformity between the collision avoidance behavior and the crew member's avoidance behavior preference and the degree of change in collision risk, select the optimal second collision avoidance behavior from the ship collision avoidance behavior set and the collision avoidance behavior in similar encounter scenarios.
[0060] It should be noted that a mechanism-driven iterative optimization model of ship behavior is constructed based on optimization theory from risk evolution, avoidance behavior preference, etc., and the ship collision avoidance behavior results generated by the optimal mechanism are selected.
[0061] The discriminator design of this application is as follows: the discriminant function is constructed by integrating the compliance of navigation practice and collision risk, and the discriminator is designed. The compliance of navigation practice refers to the compliance of the collision avoidance behavior with the crew's avoidance behavior preference. The avoidance timing and avoidance action are used as indicators, and the discriminant function is constructed by deeply integrating the avoidance timing, action, collision risk and other factors. ;in, , They represent the timing of collision avoidance and the probability of the crew's avoidance behavior preference corresponding to the avoidance behavior. The greater the probability, the more consistent it is with navigation practice. r Represents the change level of collision risk caused by the ship's collision avoidance behavior. By distinguishing the data-driven ship collision avoidance behavior and the mechanism-modeled ship collision avoidance behavior in the same encounter scenario, the better ship collision avoidance behavior is determined and fed back to the deep learning network loss function and the encounter scenario extension library to achieve data and mechanism confrontation fusion, and then realize the ship collision avoidance behavior modeling that conforms to navigation practice.
[0062] The above is a detailed introduction to a ship collision avoidance method provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the idea of the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.
[0063] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present application should be covered within the protection scope of the present application.
Claims
1. A ship collision avoidance method, characterized in that: include: Determine the first collision avoidance action based on the fully trained deep learning model; Determine the second collision avoidance behavior based on the constructed ship collision avoidance mechanism model; Based on the degree of conformity between the collision avoidance behavior and the crew's avoidance behavior preference and the degree of change in collision risk, a discriminator is constructed; An optimal collision avoidance behavior is selected from the first collision avoidance behavior and the second collision avoidance behavior based on the discriminator.
2. The ship collision avoidance method according to claim 1, characterized in that: The process of determining the fully trained deep learning model includes: Obtain target historical traffic environment data; Extracting target feature data based on the target historical traffic environment data; An initial deep learning model is trained based on the target feature data to obtain a fully trained deep learning model.
3. The ship collision avoidance method according to claim 2, characterized in that: Obtain the target's historical traffic environment data, including: Acquiring initial historical traffic environment data of the target area, wherein the initial historical traffic environment data includes ship AIS data, radar data, electronic chart data, hydrological data and meteorological data; Eliminating abnormal data and repairing the track of the ship in the ship AIS data, and cleaning and repairing the radar data, the electronic chart data, the hydrological data and the meteorological data to obtain intermediate historical traffic environment data; The intermediate historical traffic environment data are temporally and spatially matched and fused to obtain target historical traffic environment data.
4. The ship collision avoidance method according to claim 3, characterized in that: The TSC-TIP algorithm is used to perform spatiotemporal matching on the intermediate historical traffic environment data, and the weighted and splicing processing method is used to fuse the intermediate historical traffic environment data.
5. The ship collision avoidance method according to claim 2, characterized in that: Based on the target historical traffic environment data, target feature data is extracted, including: Extracting first characteristic data from the target historical traffic environment data according to the correlation characteristics between human, ship and environmental parameters and the ship collision avoidance behavior; Performing feature processing on the first feature data to establish a dynamic mapping mechanism between data and features to obtain second feature data; The second feature data are unified to the same scale to obtain intermediate feature data, and target feature data is determined based on the intermediate feature data.
6. The ship collision avoidance method according to claim 5, characterized in that: The first characteristic data includes an initial collision risk, and the initial collision risk is extracted by the following method: By combining the controller and the ship motion model, a ship motion prediction model is constructed; Based on the ship AIS data in the target historical traffic environment data, a target ship motion prediction model is constructed; Based on the own ship motion prediction model, the target ship motion prediction model, the target historical traffic environment data and the probabilistic speed barrier method, the initial collision risk between the own ship and the target ship is calculated.
7. The ship collision avoidance method according to claim 6, characterized in that: Performing feature processing on the first feature data to establish a dynamic mapping mechanism between data and features to obtain second feature data, including Performing statistics and fitting on the random characteristic data in the first characteristic data to obtain distribution characteristic data; constructing a matrix for the variable-dimensional feature data in the first feature data, calculating matrix characteristic data of each matrix, and determining a target collision risk based on the matrix characteristic data and the initial collision risk; Processing the hierarchical feature data in the first feature data by feature coding to obtain coded data; Building a grid map based on the static obstacle feature data in the first feature data to obtain map data; Second characteristic data is formed based on the distribution characteristic data, the object collision risk, the coding data and the map data.
8. The ship collision avoidance method according to claim 5, characterized in that: Determining target feature data based on the intermediate feature data includes: Based on the nearest component analysis method, feature selection is performed on the intermediate feature data to obtain target feature data.
9. The ship collision avoidance method according to claim 1, characterized in that: The fully trained deep learning model adopts the TimesNet network structure.
10. The ship collision avoidance method according to claim 1, characterized in that: The second collision avoidance behavior is determined based on the constructed ship collision avoidance mechanism model, including: Obtain the ship's encounter scenario and target's historical traffic environment data; Based on several basic ship collision avoidance mechanism models, a set of ship collision avoidance behaviors in the encounter scenario of the own ship is generated; Based on the clustering of the encounter scene features, extracting the collision avoidance behaviors of the similar encounter scenes similar to the encounter scene of the own ship from the target historical traffic environment data; Based on the degree of conformity between the collision avoidance behavior and the crew's avoidance behavior preference and the degree of change in collision risk, the optimal second collision avoidance behavior is selected from the ship collision avoidance behavior set and the similar encounter scenario collision avoidance behavior.