A collision avoidance optimization method for ships under conditions of uncertain obstacle motion information

Abstract

This invention discloses a ship collision avoidance optimization method under conditions of uncertain obstacle vessel motion information. It constructs a membership function to measure the collision hazard level of a ship, thereby building a mathematical model to obtain the current collision hazard level. Based on international maritime collision avoidance rules, it divides and obtains the collision hazard stages and the corresponding encounter situations and yield relationships of ships in the corresponding collision hazard stages according to a preset collision hazard range. By combining the DWA and VO collision avoidance algorithms, it provides support for the autonomous collision avoidance strategy formulation of intelligent ships under different collision avoidance situations. For situations where the ship's navigation status output is discontinuous, a CNN-GRU ship navigation status prediction network is proposed. Based on historical AIS data, it predicts the ship's navigation status and its uncertainty, obtaining the uncertain distribution of its position. This ensures that even when moving obstacles cannot be continuously perceived, the intelligent decision-making system can still provide effective collision avoidance decision support for the intelligent ship, fully considering the uncertainty of the ship's position and optimizing the ship domain.

Description

Technical Field

[0001] This invention relates to the field of intelligent collision avoidance technology for ships, and in particular to a collision avoidance optimization method for ships under conditions of uncertain obstacle vessel motion information. Background Technology

[0002] During autonomous navigation, intelligent ships can rely on various sensors for precise positioning and obstacle detection. However, they often encounter difficulties in continuously sensing obstacle signals during collision avoidance, making it challenging for them to accurately and in real-time understand the navigation environment and encounter situations. This limitation in perception poses a challenge to the effectiveness of collision avoidance algorithms based on the observable motion of obstacles. Summary of the Invention

[0003] This invention provides a collision avoidance optimization method for ships under conditions of uncertain obstacle vessel motion information, in order to overcome the above-mentioned technical problems.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A collision avoidance optimization method for ships under conditions of uncertain obstacle vessel motion information specifically includes the following steps:

[0006] S1: Detect and acquire the status information of obstacle ships and the target smart ship within the navigable waters perception area of ​​the target smart ship;

[0007] The status information includes at least the ship's position, ship's speed, and ship's attitude;

[0008] S2: Based on the state information of the obstacle ship and the target intelligent ship, construct a membership function to measure the collision risk of the ship, and construct a mathematical model of the collision risk of the ship based on the membership function to obtain the current collision risk of the ship.

[0009] S3: Based on international maritime collision avoidance rules, the collision avoidance hazard stage is divided according to the preset collision hazard level range, and the ship encounter situation and yield relationship in the corresponding collision avoidance hazard stage are obtained.

[0010] The collision avoidance phase of a ship includes the safe phase, the collision danger phase, the imminent situation phase, and the imminent danger phase.

[0011] S4: Based on the collision avoidance hazard phase, determine whether the target intelligent ship and the obstacle ship are in a safe phase according to the current collision hazard level;

[0012] If so, then control the target smart ship to sail along the originally set route;

[0013] Otherwise, proceed to step S5;

[0014] S5: Confirm whether the obstructing vessel cannot be stably tracked;

[0015] If so, the CNN+GRU network model, which is built by combining a convolutional neural network (CNN) and a gated recurrent neural network (GRU) after training, is used to obtain the status information of the obstacle ship in real time due to the loss of the detection signal.

[0016] Furthermore, based on the real-time acquired status information of the obstacle vessel, the real-time collision risk level with the target smart vessel is obtained, and step S6 is then executed.

[0017] If not, proceed directly to step S6;

[0018] S6: Based on the real-time ship collision risk level / current ship collision risk level, determine whether the target smart ship and the obstacle ship are in an imminent danger stage;

[0019] If so, the current reachable speed space of the target smart ship is obtained based on the DWA algorithm; and the optimal speed pair in the current reachable speed space is selected based on the objective function of the DWA algorithm.

[0020] If not, the DWA algorithm is improved based on the VO algorithm to obtain the current reachable speed space of the target smart ship, and an objective function for local trajectory planning of the smart ship is constructed to select the optimal speed pair in the current reachable speed space.

[0021] S7: Based on the selected optimal speed pair in the current achievable speed space, realize collision avoidance of ships under the condition of uncertain obstacle ship motion information.

[0022] Furthermore, the membership functions constructed in S2 include the nearest encounter distance membership function, the nearest encounter time membership function, the dangerous encounter area distance membership function, and the relative true azimuth membership function and the relative ship speed ratio membership function that affect the navigation safety of the intelligent ship.

[0023] The membership function D(DCPA) for the nearest encounter distance of the ship is:

[0024]

[0025] In the formula: d1 represents the distance of the area in which the collision avoidance maneuver can be performed, d2 represents the distance in the area in which the collision avoidance maneuver is performed last, and d2 = nd1 (n = 1, 2, ..., k);

[0026] The nearest encounter time membership function D(TCPA) is:

[0027]

[0028] In the formula: t1 and t2 represent the lower and upper bounds of the nearest encounter time TCPA, respectively; This represents the relative velocity modulus of the obstacle vessel with respect to the vessel itself, and

[0029]

[0030] The membership function D(d) of the danger encounter area is:

[0031]

[0032] The relative true azimuth membership function D(R) c The membership function D(K) for the relative ship speed ratio is:

[0033]

[0034] In the formula: R c Q represents the relative true bearing between the obstacle ship and the target intelligent ship; K represents the relative speed ratio between the obstacle ship and the target intelligent ship; Q represents the relative true bearing between the obstacle ship and the target intelligent ship. c This indicates the true bearing of the obstacle ship to the target intelligent ship;

[0035] The mathematical model for ship collision risk constructed based on the membership function is as follows:

[0036] CRI=ω1D(DCPA)+ω2D(TCPA)+ω3D(d)+ω4D(R c )+ω5D(K)

[0037] In the formula: CRI represents the ship collision risk level; ω1~ω5 represent the preset weight values.

[0038] Furthermore, S3 describes the process of dividing and obtaining the collision avoidance hazard stage Θ based on the International Maritime Collision Prevention Regulations and a preset range of ship collision hazard levels, expressed as follows:

[0039]

[0040] The ship encounter situation and yielding relationship during the collision avoidance phase are as follows:

[0041]

[0042] Furthermore, the training process of the CNN+GRU network model built in S5, which combines a convolutional neural network (CNN) and a gated recurrent neural network (GRU), is as follows:

[0043] S51: Obtain historical ship time-series trajectory sequence data, normalize it to obtain a standard sample set, and input the standard sample set into the CNN+GRU network model for training;

[0044] Furthermore, the CNN+GRU network model includes an encoder unit and a decoder unit;

[0045] The encoder unit includes a convolutional neural network layer, a two-layer GRU network layer, and a position dropout network layer connected in sequence.

[0046] The convolutional neural network layer is used to extract data features from the ship time-series trajectory sequence data in the standard sample set;

[0047] The data features include ship position, ship speed, and ship attitude;

[0048] The dual-layer GRU network layer is used to further extract data features of the ship time-series trajectory sequence from the output of the convolutional neural network layer, and to encode these features to obtain an intermediate vector.

[0049] The position Dropout network layer is used to proportionally ignore some hidden neurons during the training of the CNN+GRU network model in order to obtain the network model for estimating the uncertainty of the obstacle boat's motion information.

[0050] The decoder unit specifically adopts a recurrent network structure composed of a single-layer GRU and a fully connected layer;

[0051] The single-layer GRU is used to perform feature decoding operations on the intermediate vector;

[0052] The fully connected layer is used to extract the output of the single-layer GRU to obtain the obstacle ship state prediction information at the next moment;

[0053] S52: Use the mean squared error function as the loss function to determine whether the output of the trained CNN+GRU network model has converged, and use the trained CNN+GRU network model with converged output as the final CNN+GRU network model.

[0054] Furthermore, the DWA algorithm improved based on the VO algorithm described in S6, in order to obtain the current reachable velocity space of the target intelligent ship, specifically...

[0055] S61: Define the position of the obstacle ship at time t as P. t Its expression is

[0056]

[0057] In the formula: N represents the data acquisition period and N = (t - t0) / Δt; Δt represents the duration of the intelligent ship control cycle; t0 represents the start time of the obstacle ship prediction; P TS Indicates the predicted starting position of the obstacle vessel;

[0058] The speed range v that causes the target smart ship OS to collide with the obstacle ship TSOS , represented as

[0059]

[0060] In the formula: D(P,r) sum / (t-t0)) represents the overlapping area of ​​the target intelligent ship and the obstacle ship. Represents the Minkowski vector operator; r sum P represents the sum of the ship domains of the target intelligent ship and the obstacle ship; OS (t0) represents the position of the target smart ship at time t0; P TS (t0) represents the position of the obstacle ship at time t0;

[0061] S62: Due to the CNN+GRU network model predicting the ship speed v of the obstacle ship at the next moment... TS It follows a Gaussian distribution, that is...

[0062] The positional distribution of the obstacle ships is [P] TS (t)-(v e -1.96σ x )Δt,P TS (t)+(v e +1.96σ x )Δt]

[0063] Among them, v e σ represents the mean speed obtained from multiple predictions; σ represents the standard deviation of the speed.

[0064] And determine t based on the location distribution of the obstacle ships. n The constraints that the positions of the target smart ship and the obstacle ships satisfy at any given time are:

[0065] S63: Based on the aforementioned constraints, according to the velocity range v at which the target intelligent ship OS collides with the obstacle ship TS. OS The speed selectable interval VNLHO, considering the uncertainty of the obstacle ship's dynamic motion information, is obtained and used as the current achievable speed space of the target intelligent ship. Its expression is as follows:

[0066]

[0067] Furthermore, the objective function J for local trajectory planning of the intelligent ship in S6 is expressed as follows:

[0068] minJ=ω v J v +ω ψ J ψ

[0069] J v =|Δv OS |=|v OS (t+Δt)-v OS (t)|

[0070] J ψ =|Δψ OS |=|ψ OS (t+Δt)-ψ OS (t)|

[0071] In the formula: J v J ψ These represent the defined changes in speed and heading, respectively; v represents the achievable linear velocity; ψ represents the feasible angular velocity; ω v ,ω ψ Indicates the weighting parameter; Δv OS Δψ represents the change in reachable linear velocity. OS This represents the feasible change in angular velocity.

[0072] Beneficial Effects: This invention provides a ship collision avoidance optimization method under conditions of uncertain obstacle vessel motion information. It constructs a membership function to measure the collision hazard level of a ship, thereby building a mathematical model to obtain the current collision hazard level. Based on international maritime collision avoidance rules, it divides and obtains the collision hazard stages and the corresponding encounter situations and yield relationships of ships in the corresponding collision hazard stages according to a preset collision hazard range. By combining the DWA and VO collision avoidance algorithms, it provides support for the autonomous collision avoidance strategy formulation of intelligent ships under different collision avoidance situations. Furthermore, for situations where ship navigation status output is discontinuous, a CNN-GRU ship navigation status prediction network is proposed. Based on historical AIS data, it predicts the ship's navigation status and its uncertainty, obtaining the uncertain distribution of its position. This ensures that even when moving obstacles cannot be continuously perceived, the intelligent decision-making system can still provide effective collision avoidance decision support for the intelligent ship, fully considering the uncertainty of the ship's position and optimizing the ship domain. Attached Figure Description

[0073] 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.

[0074] Figure 1 This is a flowchart of the ship collision avoidance optimization method under conditions of uncertain obstacle vessel motion information according to the present invention.

[0075] Figure 2 This is a flowchart illustrating the technical route of the ship collision avoidance optimization method under the condition of uncertain obstacle vessel motion information in this embodiment.

[0076] Figure 3 This is a schematic diagram of the search space for the DWA algorithm in this embodiment;

[0077] Figure 4 This is a schematic diagram of the dynamic ship domain model in this embodiment;

[0078] Figure 5 This is a simplified schematic diagram of the collision avoidance waypoints in this embodiment;

[0079] Figure 6 This is a schematic diagram of the CNN-GRU prediction network structure for ship motion state considering trajectory uncertainty in this embodiment;

[0080] Figure 7 This is a schematic diagram of the process for estimating the uncertainty of ship trajectory prediction and motion state in this embodiment;

[0081] Figure 8 This is a schematic diagram showing the location distribution of the obstacle ships in this embodiment;

[0082] Figure 9 This is a schematic diagram of the intelligent ship collision avoidance algorithm under the condition of uncertain obstacle ship motion information in this embodiment;

[0083] Figure 10 This is a schematic diagram of the ship domain considering the extension of uncertainty of the obstacle ship in this embodiment;

[0084] Figure 11 This is a schematic diagram of the encounter situation when the navigation status information of the blue vessel is lost in this embodiment;

[0085] Figure 12 This is a schematic diagram of the encounter situation when the navigation status information of the red vessel is lost in this embodiment;

[0086] Figure 13 This is a schematic diagram illustrating the vessel's avoidance of the vessel in the blue area in this embodiment;

[0087] Figure 14 This is a schematic diagram of the vessel passing through the clear blue area in this embodiment;

[0088] Figure 15 This is a schematic diagram illustrating the vessel's avoidance of the vessel in the red zone in this embodiment;

[0089] Figure 16 This is a schematic diagram of the ship passing through the red area of ​​the Qing Dynasty in this embodiment. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] This embodiment provides a collision avoidance optimization method for ships under conditions of uncertain obstacle vessel motion information, such as... Figures 1 to 2 As shown, the specific steps include:

[0092] S1: Detect and acquire the status information of obstacle ships and the target intelligent ship within the navigable water area of ​​the target intelligent ship; the status information includes at least the ship's position, ship's speed, and ship's attitude.

[0093] S2: Based on the state information of the obstacle ship and the target intelligent ship, construct a membership function to measure the collision risk of the ship, and construct a mathematical model of the collision risk of the ship based on the membership function to obtain the current collision risk of the ship.

[0094] Specifically, although the DWA algorithm is widely used in robot path planning tasks, it still has certain limitations. The DWA algorithm mainly generates velocity commands based on dynamic windows and calculates the optimal motion strategy by considering the intelligent vessel's own dynamic constraints and obstacles in the environment. It typically has a significant advantage in rapidly responding to dynamic environments. However, intelligent vessels navigating at sea differ from land robots; their collision avoidance process is lengthy and must adhere to international maritime collision avoidance regulations (International Maritime Organization. (2002). International Regulations for Preventing Collisions at Sea (COLREGs). International Maritime Organization). This presents certain limitations for the DWA algorithm, which relies on dynamic constraints for path planning.

[0095] Therefore, to address the aforementioned issues, this embodiment integrates the DWA and VO algorithms. When the intelligent ship adheres to international maritime collision avoidance rules, a combined VO and DWA algorithm is used. However, in the "imminent danger zone" where there is a severe collision risk between the two ships and all available actions are required, the VO constraint is abandoned, and only the DWA algorithm, capable of rapid response to environmental dynamics, is employed. To achieve this objective, the collision risk level should first be reclassified, and the application timing of each sub-algorithm should be clarified (e.g., ...). Figure 2 (As shown).

[0096] First, the collision hazard level is constructed. The nearest encounter distance (DCPA) between the two ships is used as one of the factors to measure the collision hazard level, and the membership function of the nearest encounter distance is:

[0097]

[0098] In the formula: d1 represents the range distance at which collision avoidance maneuvers can be performed, d2 represents the range distance at which collision avoidance maneuvers can be performed last, and d2 = nd1 (n = 1, 2, ..., k). The setting of this range fully considers factors such as the size, speed, maneuverability and encounter situation of the ship.

[0099] In this embodiment, only the dynamic changes in collision avoidance dimensions in the bow and stern directions of the obstacle ship or the target intelligent ship are considered, and it is simplified into a ship domain model consisting of a semi-ellipse and a semi-circle.

[0100]

[0101] In the formula: R f ,R s Representing the elliptical and circular parameter radii of a ship domain model Represents the coordinates of the encircling ellipse; Indicates and Let R represent the horizontal and vertical coordinates of the ellipse, respectively; Q represents the set {R}. f ,R s};like Figure 4 As shown; sgn(x) represents the sign function. The radii of the ellipse and circle fully consider parameters such as ship size, relative speed between the ship and the target ship, the ship's speed, and the encounter situation. Its dynamic radius can be described as...

[0102]

[0103] In the formula: AD represents the ship's approach distance; V OS Let L represent the speed of the ship; L represent the length of the ship; and S(i) represent the piecewise function of the encounter situation.

[0104]

[0105] In the formula: R c Indicates relative true bearing; V r Indicates the relative speed of the obstacle vessel with respect to the vessel itself;

[0106] In addition, TCPA is another important indicator affecting collision hazard, and the nearest encounter time membership function is:

[0107]

[0108] Define t1 and t2 as the lower and upper bounds of the nearest encounter time TCPA, respectively. Let the relative velocity modulus of the obstacle vessel relative to the main vessel be given by:

[0109]

[0110] Let d be the distance between the obstacle ship and the target intelligent ship, and d2 be the distance within the territory where the collision avoidance operation is last executed. If the distance between them is within the danger encounter territory distance, it is considered an absolute danger; if it exceeds d2, it is considered no risk. That is, the danger encounter territory distance membership function is:

[0111]

[0112] In addition, relative true azimuth R c The ratio of ship speed to K also affects the navigation safety of intelligent ships, Q. c Assuming the obstacle vessel is positioned at the true bearing of the target intelligent vessel, the designed membership functions for relative true bearing and relative speed ratio are as follows:

[0113]

[0114] The membership functions of all the above indicators are D∈[0,1], and D(DCPA), D(TCPA), D(d), and D(R) are defined. c After assigning weights ω1 to ω5 to D(K) respectively, the collision hazard level can be mathematically expressed as:

[0115] CRI=ω1D(DCPA)+ω2D(TCPA)+ω3D(d)+ω4D(R c )+ω5D(K)

[0116] S3: Based on international maritime collision avoidance rules, the collision avoidance hazard stage is divided according to the preset collision hazard level range, and the ship encounter situation and yield relationship in the corresponding collision avoidance hazard stage are obtained.

[0117] The collision avoidance phase of a ship includes the safe phase, the collision danger phase, the imminent situation phase, and the imminent danger phase.

[0118] Specifically, after obtaining the Collision Risk Ratio (CRI), this study divides the different stages of collision avoidance risk into the following stages according to the International Maritime Collision Avoidance Code, as shown in Table 1.

[0119] Table 1. Criteria for Classifying Different Stages of Collision Hazard

[0120]

[0121] The expressions for the different stages of collision risk are as follows:

[0122]

[0123] The specific assessment of the encounter situation and the yielding relationship, i.e., the collision avoidance rules corresponding to the dangerous phase of a collision, are as follows:

[0124]

[0125] Other = {Q c R c Not included in the above situations} OS: Straight

[0126] Based on collision avoidance rules, the specific turning point location can be calculated according to the obstacle vessel and the dynamic vessel domain. Specifically, for ease of calculation, waypoints are selected as follows: Figure 5 As shown, when this vessel passes on the port or starboard side of the target vessel, the waypoints this vessel should choose are wt1 and wt2 respectively; CR in the figure s Let N be the radius of the passing point, N be the next target point, and P be the vector from the position coordinates P2 of the obstacle ship to the position coordinates P1 of the target intelligent ship. The formulas for calculating the positions of passing points wt1 and wt2 are shown below.

[0127]

[0128] S4: Based on the collision avoidance hazard phase, determine whether the target intelligent ship and the obstacle ship are in a safe phase according to the current collision hazard level, such as... Figure 9 As shown;

[0129] If so, then control the target smart ship to sail along the originally set route;

[0130] Otherwise, proceed to step S5;

[0131] S5: Confirm whether the obstructing vessel cannot be stably tracked;

[0132] If so, the trained CNN+GRU network model, which combines a convolutional neural network (CNN) and a gated recurrent neural network (GRU), is used to obtain information on the future time obstacle ship's motion status in real time due to the loss of detection signals.

[0133] Furthermore, based on the real-time acquired status information of the obstacle vessel, the real-time collision risk level with the target smart vessel is obtained, and step S6 is then executed.

[0134] If not, proceed directly to step S6;

[0135] In a specific embodiment, the training process of the CNN+GRU network model, which is constructed by combining a convolutional neural network (CNN) and a gated recurrent neural network (GRU), is as follows:

[0136] S51: Obtain historical ship time-series trajectory sequence data, normalize it to obtain a standard sample set, and input the standard sample set into the CNN+GRU network model for unsupervised learning training;

[0137] Furthermore, the CNN+GRU network model includes an encoder unit and a decoder unit;

[0138] The encoder unit includes a convolutional neural network layer, a two-layer GRU network layer, and a position dropout network layer connected in sequence.

[0139] The convolutional neural network layer is used to extract data features from the ship time-series trajectory sequence data in the standard sample set;

[0140] The data features include ship position, ship speed, and ship attitude;

[0141] The dual-layer GRU network layer is used to further extract data features of the ship time-series trajectory sequence from the output of the convolutional neural network layer, and to encode these features to obtain an intermediate vector.

[0142] The position Dropout network layer is used to proportionally ignore some hidden neurons during the training of the CNN+GRU network model in order to obtain the network model for estimating the uncertainty of the obstacle boat's motion information.

[0143] The decoder unit specifically adopts a recurrent network structure composed of a single-layer GRU and a fully connected layer;

[0144] The single-layer GRU is used to perform feature decoding operations on the intermediate vector;

[0145] The fully connected layer is used to extract the output of the single-layer GRU to obtain the obstacle ship state prediction information at the next moment;

[0146] S52: Use the mean squared error function as the loss function to determine whether the output of the trained CNN+GRU network model has converged, and use the trained CNN+GRU network model with converged output as the final CNN+GRU network model.

[0147] In this embodiment, the CNN-GRU prediction model mainly consists of two parts: an encoder and a decoder. In the encoder, feature codes {position, velocity, attitude} are extracted from the input time series data and converted into intermediate vectors, which are then used as the encoder's input. The decoder decodes these intermediate vectors, combines them with the current input data, and predicts the state sequence for the next time step. This embodiment utilizes CNN+GRU in the encoder structure to extract features of the ship's motion state. CNNs have parameter sharing characteristics and possess memory when processing time series data. In this application, the convolutional kernel layer has a kernel size of 3×3, 512 output channels, a stride of 1, an expansion rate of 2, and uses the ReLU activation function. In the encoder, this application employs a two-layer GRU network structure as follows: Figure 6 The first GRU layer has 128 hidden neurons, while the second GRU layer consists of 64 hidden neurons. The decoder uses a recurrent network structure composed of a single-layer GRU and a fully connected layer. The GRU network has 128 layers, and the fully connected layer has 128 neurons. Both use the tanh function as the activation function, the mean squared error function as the loss function, and Adam as the optimization function. The ship state prediction information obtained by the above method is affected by data noise, the accuracy of the prediction model itself, and the uncertainty of ship motion, inevitably resulting in some error. Currently, due to the lack of methods for estimating the uncertainty of ship motion state prediction, it is difficult to quantify this uncertainty. To ensure the navigation safety of intelligent ships at sea, the common technical approach is to leave a large safety threshold to avoid potential risks. However, this approach will increase the area of ​​unnavigable waters during collision avoidance, affecting collision avoidance decisions. This embodiment proposes to set a prior probability distribution on the model parameter weights to estimate the uncertainty of model output changes caused by variations in input data. Secondly, the Dropout algorithm (position Dropout network layer) is used to estimate the uncertainty of the prediction results, thereby obtaining the ship trajectory area (e.g., Figure 8 The uncertainty of the predicted values ​​of heading and speed (as shown) is calculated using the Dropout algorithm, which discards some neurons in the training network with a certain probability. This results in differences in the network structure during training, leading to variations in the training results. The complete network model for estimating the uncertainty of the prediction results is as follows: Figure 7 As shown, the Dropout algorithm in the network proportionally ignores some hidden neurons in the GRU network, which prevents the ship motion state information extracted by the convolutional layer from entering the same sub-network every time. This reduces the dependence on fixed nodes when updating parameters, thereby obtaining the uncertain distribution of prediction results from different sub-networks at future time points, in order to obtain the obstacle ship state prediction information at the next time step.

[0148] The specific expression for obtaining the uncertain distribution at future points in time is as follows:

[0149]

[0150] In the formula: E(x) t+1:t+k () represents the expected value of N predictions; x t+1:t+k Indicates the predicted location coordinates; σ t+1:k+1 This indicates the location of uncertainty in the prediction results, used to calculate the covariance matrix Q. t+i ;

[0151]

[0152] In the formula: σ t+i,x , σ t+i,y These represent the x-coordinate and y-coordinate of the position corresponding to the uncertain position, respectively.

[0153] Based on the covariance matrix Q t+i The elliptic confidence level of the ship position prediction results is obtained;

[0154] Assume the uncertainty follows a multivariate normal distribution, and the elliptic confidence rate is s. The center of the ellipse is obtained by averaging the N predictions. The elliptic confidence rate s of the ship position prediction result is then used to obtain the prediction result of the ship's motion state at future time.

[0155] Furthermore, the expression for the elliptic confidence rate s of the ship position prediction results is:

[0156]

[0157] In the formula: This represents the x-axis coordinate at time t+1; This represents the y-axis coordinate at time t+1;

[0158] During autonomous navigation, the intelligent ship in this implementation needs to continuously perceive its own and other ships' motion status using various sensors. However, during collision avoidance, there may be situations where the detection signal is lost, making it difficult for the intelligent ship to obtain the real-time motion status information of the obstacle ship. In this case, the obstacle ship's motion status has a high degree of uncertainty, and the intelligent ship collision avoidance algorithm based on known obstacle motion information will no longer be applicable. Therefore, this embodiment has predicted the uncertainty of the obstacle ship's trajectory and the ship's speed and course, which are highly related to the collision avoidance decision. This embodiment will fully utilize the prediction results to further optimize the collision avoidance decision algorithm studied above. Considering that the proposed method assumes "maintaining direction and speed" for the moving obstacle ship, however, the actual speed v of the obstacle ship... TS It is time-varying, and its speed can be sensed in real time by intelligent ships.

[0159] S6: Based on the real-time ship collision risk level / current ship collision risk level, and combined with the ship collision avoidance risk stage obtained according to the preset ship collision risk level range based on international maritime collision avoidance rules, determine whether the target smart ship and the obstacle ship are in an imminent danger stage.

[0160] If so, the current reachable speed space of the target smart ship is obtained based on the DWA algorithm; and the optimal speed pair in the current reachable speed space is selected based on the objective function of the DWA algorithm.

[0161] Specifically, the DWA algorithm in this embodiment considers the ship's acceleration and speed limitations. Its principle is to search for feasible solutions to the planned path within a velocity space comprised of planar and rotational velocities. Multiple sets of reachable velocity vector pairs (v, w) constitute the reachable velocity space of the intelligent ship. Each velocity pair contains the pose information of the intelligent ship within its motion cycle. The intelligent ship sets certain constraints to ensure that a unique feasible solution is found in the velocity space, thereby obtaining the corresponding motion commands and ensuring the safety of the intelligent ship's navigation. This application mainly imposes the following restrictions, limiting the sampling space to a specific window for ship avoidance under urgent and dangerous conditions, i.e., when CRI ≥ 0.7, such as... Figure 3 As shown, specifically including

[0162] (1) Intelligent ship speed sampling space

[0163] By sampling the reachable velocity vector pairs in the velocity space of the intelligent ship multiple times, its maximum reachable velocity constraint can be expressed as V s = (v,ω)∈{(v,ω)|v∈[v min ,v max ],ω∈[ψ min ,ω max In the formula, v min With v max These are the minimum and maximum achievable linear velocities of the intelligent ship, respectively. Similarly, ω min With ω max These represent the minimum and maximum feasible angular velocities of the intelligent ship. These metrics are directly determined by the maneuverability and decision-making requirements of the intelligent ship during modeling.

[0164] (2) Constraints on the speed vector of the intelligent ship

[0165] Due to environmental factors and its own maneuverability, the speed change of an intelligent ship should also be constrained within a certain space V. d Inside, its mathematical description is V d ∈{(v,w)|v∈[v o -Δv b t,v o +Δv ht],ω∈[ω o -Δω b t,v o +Δω h In the above expression, v]}, o With ω o Let Δv be the linear velocity and angular velocity of the current intelligent ship. b With Δv h These represent the maximum linear accelerations of the intelligent ship when it moves backward and forward, respectively, while Δω b With Δω h These represent the maximum bow angular accelerations of the intelligent ship when it is moving backward and forward, respectively. These constraints are determined by the ship's hull performance during modeling and are also subject to the constraints of the intelligent ship's navigation environment.

[0166] (3) Speed ​​limits allowed by safety regulations

[0167] In collision avoidance decision-making for intelligent ships, safety is always the top priority. When a ship encounters an obstacle, it should take timely measures to avoid danger. Therefore, the speed of the intelligent ship is constrained to ensure that it can always pass through obstacles at a safe distance. This requires the design of a permissible speed space V. t That is, it can collide with a static obstacle at a speed of [missing value]. In the formula, dist(v,ω) refers to the closest distance between the end of the simulated trajectory and the obstacle, and v smax and ω smax These refer to the maximum linear acceleration and maximum turning angular velocity of the intelligent ship, respectively. This velocity space is determined by multiple factors, including the hydrodynamic characteristics of the intelligent ship, the position and shape of obstacles, and the motion state of the intelligent ship. In path planning, this allowable velocity space needs to be treated as a constraint.

[0168] (4) Evaluation function

[0169] After setting the constraints, the DWA algorithm can calculate the ship's actual trajectory based on the current reachable velocity space (v,w), thus defining the intelligent ship's velocity search space V. m This can be expressed as the formula V m =V s ∩V d ∩V t .like Figure 3 As shown.

[0170] Once the search space is defined, the optimal velocity vector pair needs to be searched within that velocity space. In the DWA algorithm, the planned linear and angular velocities remain constant throughout the planning period. Therefore, within the planning period, the local trajectory of the intelligent ship consists of straight lines and circular arcs. m In space, the optimal velocity pair (v, w) of the classic DWA algorithm can be determined by the objective function of the corresponding local trajectory, and its expression is:

[0171]

[0172] Here, dist(v,w) is the evaluation factor for the distance between the local segment and the obstacle, α, β, and γ are weighting coefficients, vel(v,w) is the evaluation factor for (v,w) relative to the water speed, and head(v,w) is the deviation of the intelligent ship's heading angle from the critical target point's heading angle. θ is the angle between the intelligent ship and the target point, R is the intelligent ship's turning radius, and dist is the distance between the vehicle's current position and the target point. When there are no obstacles between the intelligent ship's current position and the local target point, n = 0; otherwise, n = 1. This function will be used for ship decision-making under conditions where collision avoidance rules fail.

[0173] If not, the DWA algorithm is improved based on the VO algorithm to obtain the current reachable velocity space of the target smart ship, specifically including the following steps.

[0174] S61: Define the position of the obstacle ship at time t as P. t Its expression is

[0175]

[0176] In the formula: N represents the data acquisition period and N = (t - t0) / Δt; Δt represents the duration of the intelligent ship control cycle; t0 represents the start time of the obstacle ship prediction; P TS Indicates the predicted starting position of the obstacle vessel;

[0177] The speed range v that causes the target smart ship OS to collide with the obstacle ship TS OS , represented as

[0178]

[0179] Where: D(P, r s um / (t-t0)) represents the overlapping area of ​​the target smart ship and the obstacle ship, and the two ships will collide within the time interval [t0,+∞). Represents the Minkowski vector operator; r sum P represents the sum of the ship domains of the target intelligent ship and the obstacle ship; OS (t0) represents the position of the target smart ship at time t0; P TS (t0) represents the position of the obstacle ship at time t0;

[0180] S62: Due to the CNN+GRU network model predicting the ship speed v of the obstacle ship at the next moment... TS It follows a Gaussian distribution, that is...

[0181] The positional distribution of the obstacle ships is [P] TS (t)-(v e -1.96σ x )Δt,P TS (t)+(v e +1.96σ x )Δt]

[0182] Among them, v e This represents the mean speed obtained from multiple predictions; σ represents the standard deviation of the speed; by setting the confidence level of the obstacle ship's speed prediction results to 95%, the speed distribution range at time t+1 can be obtained, such as... Figure 10 As shown, the red dashed line represents the range of uncertainty in the predicted distribution of obstacle vessel positions;

[0183] And determine t based on the location distribution of the obstacle ships. n The constraints that the positions of the target smart ship and the obstacle ships satisfy at any given time are:

[0184] S63: Based on the aforementioned constraints, according to the velocity range v at which the target intelligent ship OS collides with the obstacle ship TS. OS The speed selectable interval VNLHO, considering the uncertainty of the obstacle ship's dynamic motion information, is obtained and used as the current achievable speed space of the target intelligent ship. Its expression is as follows:

[0185]

[0186] An objective function for local trajectory planning of an intelligent ship is constructed to select the optimal velocity pair in the currently achievable velocity space. The objective function J for local trajectory planning of the intelligent ship is expressed as follows:

[0187] min J=ω v J v +ω ψ J ψ

[0188] J v =|Δv OS |=|v OS (t+Δt)-v OS (t)|

[0189] J ψ =|Δψ OS |=|ψ OS (t+Δt)-ψ OS (t)|

[0190] In the formula: J v J ψThese represent the defined changes in speed and heading, respectively; v represents the achievable linear velocity; ψ represents the feasible angular velocity; ω v ,ω ψ Indicates the weighting parameter; Δv OS Δψ represents the change in reachable linear velocity. OS This represents the feasible change in angular velocity.

[0191] S7: Based on the selected optimal velocity pair in the currently achievable velocity space, achieve collision avoidance for ships under conditions of uncertain obstacle vessel motion information. That is, based on the position uncertainty prediction results, numerically superimpose the position uncertainty ellipse and the ship's domain. For example... Figure 9 As shown, by inputting the updated ship domain into the improved DWA and VO hybrid collision avoidance decision method, an intelligent ship collision avoidance algorithm under the condition of uncertain obstacle ship motion information can be obtained. Based on the intelligent ship collision avoidance strategy, the ship collision avoidance process under the condition of uncertain obstacle ship motion information is realized. The basis for determining safe passage is passing outside the superimposed domain.

[0192] This embodiment combines the DWA and VO collision avoidance algorithms to support the formulation of autonomous collision avoidance strategies for intelligent ships under different collision avoidance scenarios. Furthermore, addressing the issue of discontinuous output of ship navigation status, a CNN-GRU ship navigation status prediction network is proposed. Based on historical AIS data, this network predicts the ship's navigation status and its uncertainties, obtaining the uncertainty distribution of its position. This ensures that even when moving obstacles cannot be continuously perceived, the intelligent decision-making system can still provide effective collision avoidance decision support for the intelligent ship, fully considering the uncertainty of the ship's position and optimizing the ship's domain.

[0193] The experimental results of this embodiment show that after the target intelligent ship passes through the yellow and green zones and passes the red turning point, it will encounter ships in the blue and red zones. At this point, assuming the ship loses the signal from the blue ship, the ship begins to predict the position uncertainty of the blue ship. The safe passage area, resulting from the superposition of the position uncertainty ellipse and the ship's own zone, is as follows: Figure 11 As shown in the diagram. Subsequently, the distance between our ship and the predicted blue-area vessel gradually decreases, the area of ​​the uncertainty ellipse for the blue vessel's position gradually increases, and at this time, the signal for the red vessel also begins to be lost. Our ship then begins to predict the position uncertainty of the red vessel, as shown in the diagram. Figure 12 As shown. As the vessel and the predicted blue vessel approached each other, until the vessel took a sharp right turn to avoid them, as... Figure 13 As shown. When this vessel passes through the clear blue area, such as... Figure 14As shown, after resuming the original route, a face-to-face situation arises with the vessel in the red area. At this point, the area of ​​the blue ellipse used to predict positional uncertainty is much larger than the vessel's own domain, and its domain is closer to an ellipse. It is noteworthy that during this process, the actual position of the vessel with the blue historical track remains within the predicted blue ellipse dashed line, proving the effectiveness of the algorithm. Figure 15 As shown, after our vessel completed the avoidance maneuver to avoid the vessel in the blue area, the area of ​​the uncertainty ellipse of the red vessel's position had exceeded its own inherent vessel territory, resulting in a relatively large safe area for avoidance. Despite this, our vessel still managed to avoid the red vessel by making a sharp right turn, and ultimately returned to its original course after clearing the way. Figure 16 As shown.

[0194] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A collision avoidance optimization method for ships under conditions of uncertain obstacle vessel motion information, characterized in that, Specifically, the following steps are included: S1: Detect and acquire the status information of obstacle ships and the target smart ship within the navigable waters perception area of ​​the target smart ship; The status information includes at least the ship's position, ship's speed, and ship's attitude; S2: Based on the state information of the obstacle ship and the target intelligent ship, construct a membership function to measure the collision risk of the ship, and construct a mathematical model of the collision risk of the ship based on the membership function to obtain the current collision risk of the ship. S3: Based on international maritime collision avoidance rules, the collision avoidance hazard stage is divided according to the preset collision hazard level range, and the ship encounter situation and yield relationship in the corresponding collision avoidance hazard stage are obtained. The collision avoidance phase of a ship includes the safe phase, the collision danger phase, the imminent situation phase, and the imminent danger phase. S4: Based on the collision avoidance hazard phase, determine whether the target intelligent ship and the obstacle ship are in a safe phase according to the current collision hazard level; If so, then control the target smart ship to sail along the originally set route; Otherwise, proceed to step S5; S5: Confirm whether the obstructing vessel cannot be stably tracked; If so, the CNN+GRU network model, which is built by combining a convolutional neural network (CNN) and a gated recurrent neural network (GRU) after training, is used to obtain the status information of the obstacle ship in real time due to the loss of the detection signal. Furthermore, based on the real-time acquired status information of the obstacle vessel, the real-time collision risk level with the target smart vessel is obtained, and step S6 is then executed. If not, proceed directly to step S6; The training process of the CNN+GRU network model, which combines a convolutional neural network (CNN) and a gated recurrent neural network (GRU), is as follows: S51: Obtain historical ship time-series trajectory sequence data, normalize it to obtain a standard sample set, and input the standard sample set into the CNN+GRU network model for training; Furthermore, the CNN+GRU network model includes an encoder unit and a decoder unit; The encoder unit includes a convolutional neural network layer, a two-layer GRU network layer, and a position dropout network layer connected in sequence. The convolutional neural network layer is used to extract data features from the ship time-series trajectory sequence data in the standard sample set; The data features include ship position, ship speed, and ship attitude; The dual-layer GRU network layer is used to further extract data features of the ship time-series trajectory sequence from the output of the convolutional neural network layer, and to encode these features to obtain an intermediate vector. The position Dropout network layer is used to proportionally ignore some hidden neurons during the training of the CNN+GRU network model in order to obtain the network model for estimating the uncertainty of the obstacle boat's motion information. The decoder unit specifically adopts a recurrent network structure composed of a single-layer GRU and a fully connected layer; The single-layer GRU is used to perform feature decoding operations on the intermediate vector; The fully connected layer is used to extract the output of the single-layer GRU to obtain the obstacle ship state prediction information at the next moment; S52: Use the mean squared error function as the loss function to determine whether the output of the trained CNN+GRU network model converges, and use the trained CNN+GRU network model with converged output as the final CNN+GRU network model. S6: Based on the real-time ship collision risk level / current ship collision risk level, determine whether the target intelligent ship and the obstacle ship are in an imminent danger stage; If so, the current reachable velocity space of the target smart ship is obtained based on the DWA algorithm; And select the optimal velocity pair in the currently achievable velocity space based on the objective function of the DWA algorithm; If not, the DWA algorithm is improved based on the VO algorithm to obtain the current reachable speed space of the target smart ship, and an objective function for local trajectory planning of the smart ship is constructed to select the optimal speed pair in the current reachable speed space. The improved DWA algorithm based on the VO algorithm is used to obtain the current reachable velocity space of the target intelligent ship, specifically... S61: Define the current moment of the obstacle ship The position is Its expression is In the formula: Indicates the collection period and ; Indicates the duration of the intelligent ship control cycle; Indicates the predicted start time of the obstacle vessel; Indicates the predicted starting position of the obstacle vessel; The speed range in which the target smart ship OS collides with the obstacle ship TS , represented as In the formula: This indicates the overlapping area between the target intelligent ship and the obstacle ship. This represents the Minkowski vector operator; This represents the sum of the ship domains of the target intelligent ship and the obstacle ship; express The location of the target intelligent ship is constantly monitored; express The position of the obstacle ship at all times; S62: The ship speed at the next moment predicted by the CNN+GRU network model. It follows a Gaussian distribution, that is... ; The positions of the obstacle ships are distributed as follows: in, This represents the average sailing speed obtained from multiple predictions. The standard deviation of sailing speed; And determine based on the location distribution of the obstacle vessels. The constraints that the positions of the target smart ship and the obstacle ships satisfy at any given time are: ; S63: Based on the aforementioned constraints, and according to the speed range at which the target intelligent ship OS collides with the obstacle ship TS... The selectable speed range is obtained by considering the uncertainty of the dynamic motion information of the obstacle vessel. And this is taken as the current reachable velocity space of the target intelligent ship, its expression is: ； S7: Based on the selected optimal speed pair in the current achievable speed space, realize collision avoidance of ships under the condition of uncertain obstacle ship motion information.

2. The ship collision avoidance optimization method under conditions of uncertain obstacle vessel motion information as described in claim 1, characterized in that, The membership functions constructed in S2 include the nearest encounter distance membership function, the nearest encounter time membership function, the dangerous encounter area distance membership function, and the relative true azimuth membership function and the relative ship speed ratio membership function that affect the navigation safety of the intelligent ship. The nearest encounter distance membership function of the ship for In the formula: This indicates the territorial distance at which a vessel can perform collision avoidance maneuvers. This indicates the distance in the area where the collision avoidance maneuver will be performed last, and ; The most recent encounter time membership function for In the formula: and These represent the lower and upper bounds of the most recent encounter time TCPA, respectively. This represents the relative velocity modulus of the obstacle vessel with respect to the vessel itself, and The danger encounter domain distance membership function for The relative true azimuth angle membership function Membership function relative to ship speed for In the formula: Indicates the relative true bearing between the obstacle ship and the target intelligent ship; This indicates the relative speed ratio between the obstacle ship and the target intelligent ship; This indicates the true bearing of the obstacle ship to the target intelligent ship; The mathematical model for ship collision risk constructed based on the membership function is as follows: In the formula: Indicates the degree of risk of a ship collision; This indicates the preset weight value.

3. The ship collision avoidance optimization method under conditions of uncertain obstacle vessel motion information as described in claim 2, characterized in that, S3 describes a system based on international maritime collision avoidance regulations, which classifies and obtains ship collision avoidance hazard stages according to a pre-defined range of ship collision hazard levels. Its expression is The ship encounter situation and yielding relationship during the collision avoidance phase are as follows: 。 4. The ship collision avoidance optimization method under conditions of uncertain obstacle vessel motion information as described in claim 3, characterized in that, The objective function for constructing local trajectory planning for intelligent ships in S6 Its expression is In the formula: , These represent the defined changes in speed and course, respectively. Indicates the achievable linear velocity; Indicates the feasible angular velocity; Indicates the weighting parameter; This represents the change in reachable linear velocity; This represents the feasible change in angular velocity.

Images