Water surface unmanned system operation intelligent decision control method based on multi-source information

Abstract

The invention relates to the technical field of unmanned ships, in particular to a water surface unmanned system operation intelligent decision control method based on multi-source information. The method comprises the steps that information, unmanned ship pose information, environment information and obstacle pose information are collected; evaluating a navigation risk level according to the environment information; when the risk level exceeds a threshold value, controlling the avoidance operation; the avoidance operation is output by the CNN-RF hybrid model, and the input of the CNN-RF hybrid model is unmanned ship pose information, environment information and obstacle pose information. According to the unmanned ship operation strategy automatic decision-making method based on the CNN-RF hybrid model, a large amount of data can be rapidly processed, corresponding decisions can be made, and the working efficiency is greatly improved. And meanwhile, the method has relatively high adaptive capacity, and can continuously optimize the decision process according to new data and environment changes.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned ships, and in particular to an intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information. Background Art

[0002] Shipping plays an important role in inland waterway transportation due to its strong carrying capacity and environmental friendliness. However, in inland waters such as urban rivers and lakes, accidents such as ship collisions and groundings still occur frequently. Leakage accidents caused by collisions may cause large-scale water pollution, seriously threatening people's life safety, ecological environment and economic security.

[0003] Unmanned ships can effectively reduce the dependence of ships on humans and improve the safe operation of ships because of their ability to perceive, make decisions and execute autonomously. They also integrate the idea of ​​artificial intelligence into the collision avoidance decision-making process, and can autonomously evaluate and determine the optimal operation strategy to avoid collisions. At the same time, it can also reduce casualties and property losses caused by such accidents. In addition, it can effectively shorten the ship's voyage, save fuel, improve shipping efficiency, and thus reduce shipping costs. However, automatic decision-making during the operation process is the core difficulty in realizing the autonomous operation of unmanned ships. How to use multi-source information to realize the dynamic automatic selection of operation strategies needs to be solved urgently. Summary of the invention

[0004] The present invention discloses an intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information. The specific method is as follows:

[0005] Collect information, including the position information of unmanned ship, environment information, and obstacle position information;

[0006] Assess the level of navigation risk based on environmental information;

[0007] When the risk level exceeds the threshold, control avoidance operations;

[0008] The avoidance operation is output by the CNN-RF hybrid model, and the input of the CNN-RF hybrid model is the unmanned ship posture information, environmental information, and obstacle posture information.

[0009] Furthermore, the unmanned ship posture information includes: unmanned ship speed, unmanned ship roll angle, unmanned ship pitch angle, unmanned ship yaw angle, unmanned ship X-axis direction offset, unmanned ship Y-axis direction offset and unmanned ship Z-axis direction offset;

[0010] The environmental information includes: wind direction, wind speed and rainfall;

[0011] The obstacle position information includes: obstacle ship speed, obstacle yaw angle, obstacle X-axis direction offset, obstacle Y-axis direction offset and obstacle Z-axis direction offset.

[0012] Furthermore, the navigation risk level is assessed based on environmental information. The specific method is as follows:

[0013] The risk level is divided into low, medium and high by setting thresholds for wind speed and rainfall, as well as the distance to obstacles.

[0014] Furthermore, the avoidance maneuvers include: deceleration, constant speed, acceleration, left turn, constant heading, and right turn.

[0015] Furthermore, the CNN-RF hybrid model includes a CNN feature extraction module and an RF classification module.

[0016] Furthermore, the CNN feature extraction module includes a convolutional layer;

[0017] The convolution layer extracts features from the input. Each filter slides on the input data, calculates the dot product of the local area of ​​the filter and the input data, and generates a feature map. The convolution calculation formula is:

[0018]

[0019] f(x)=max(0,x)

[0020]

[0021] Where m and n represent the mth input neuron and the nth output neuron respectively; A is the set of input layer nodes of the hth layer of the feedforward neural network; h represents the hth convolutional layer; is the output of neuron n; Represents the input of the h-1th convolutional layer; is the convolution kernel; Represents the bias value of the nth neuron in layer h; It represents the output of the pooling layer h. is the input of the pooling window.

[0022] Furthermore, the RF classification module randomly extracts multiple subsets Y from the features extracted by CNN through self-service sampling. k To train each decision tree, at each decision tree node, a subset F is randomly selected from all features k To find the best split point;

[0023] For each bootstrap sample Y k and feature subset F k , build a decision tree, the splitting of decision tree nodes is done by selecting the best features and thresholds. Commonly used splitting criteria include Gini impurity and information gain, which are:

[0024]

[0025] Among them, Gini(Y) represents the Gini impurity of data Y, J is the number of classes, P i represents the probability of the i-th category; Y is the data set of the current node, B is the considered feature, and Y1 and Y2 represent the two subsets obtained by splitting the threshold of feature B;

[0026] For classification problems, random forest predicts the category by voting among all decision trees. The specific formula is:

[0027]

[0028] in, represents the predicted category, K is the total number of decisions, T k (x) represents the prediction of the kth decision tree for input x.

[0029] Furthermore, the CNN-RF hybrid model is used as follows:

[0030] The unmanned ship posture information, environmental information and obstacle posture information are divided, and the data of the first 5 moments are used to classify the speed and heading decisions of the data of the last moment, and the moving step length is 1;

[0031] First, perform feature extraction on the input data, and use two layers of 2*2 convolution kernels to perform convolution and pooling operations to extract local features;

[0032] The data after feature extraction is sent to the RF model for decision classification.

[0033] Furthermore, the CNN-RF hybrid model was tested, and the evaluation indicators included accuracy, precision, F1 score and recall rate. The specific formula is as follows:

[0034]

[0035] Among them, TP is the number of positive samples correctly identified by the model; TN is the number of negative samples correctly identified by the model; FP is the number of negative samples incorrectly predicted as positive samples by the model; FN is the number of positive samples incorrectly predicted as negative samples by the model.

[0036] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0037] 1. The present invention proposes an automatic decision-making method for the operation strategy of an unmanned ship based on a CNN-RF hybrid model, which can quickly process a large amount of data and make corresponding decisions, greatly improving work efficiency. At the same time, it has strong adaptive capabilities and can continuously optimize the decision-making process according to new data and environmental changes.

[0038] 2. The CNN model solves the dimensionality curse and nonlinearity problems of the input data. The RF model improves prediction accuracy by integrating multiple decision trees, reduces the overfitting risk of a single decision tree, and enhances tolerance to noise and outliers, thereby improving the reliability of autonomous decision-making of unmanned ships in complex environments.

[0039] 3. Although the traditional method takes into account the impact of wind and rain on the ship's navigation, it does not make corresponding decisions. The input data of the present invention takes into account the information of the unmanned ship itself, dynamic obstacles, static obstacles and weather information at the same time, and can make corresponding decisions according to different scenarios, so as to accurately determine the collision situation and make correct avoidance actions, thus achieving the purpose of efficient collision avoidance in complex environments.

[0040] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings of the present invention are as follows.

[0042] Figure 1 It is the overall flow chart of the present invention;

[0043] Figure 2 It is a schematic diagram of weather information wind direction of the present invention;

[0044] Figure 3 It is a network diagram of the multi-dimensional variable CNN-RF model proposed by the present invention;

[0045] Figure 4 It is the accuracy distribution diagram of the model proposed in the present invention in the same classification task;

[0046] Figure 5 It is the ROC curve diagram of the model proposed in the present invention;

[0047] Figure 6 is a confusion matrix diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] An intelligent decision-making control method for surface unmanned system operation based on multi-source information, such as Figure 1 As shown, the specific steps are as follows:

[0050] S1. Collect information, including the position information of the unmanned ship, the environment information, and the position information of obstacles.

[0051] In step S1, the radar, sensors and other equipment carried by the unmanned ship are used to efficiently and accurately obtain three types of key information, which can provide a more comprehensive understanding of the current environment and lay a solid foundation for subsequent navigation decision-making.

[0052] S11, the unmanned ship posture information includes: X-axis direction offset, Y-axis direction offset, Z-axis direction offset, roll angle, pitch angle, yaw angle, and ship speed as the state information of the unmanned ship itself, that is,

[0053] X1={ΔX O ΔY O ΔZ O α O β O γ O v O}

[0054] Among them, {ΔX O ΔY O ΔZ O} is the position information of the unmanned ship, which respectively represents the offset of the unmanned ship in the X-axis direction, the Y-axis direction and the Z-axis direction; {α O β O γ O} as the attitude angle of the unmanned ship, which respectively represents the roll angle, pitch angle and yaw angle information of the unmanned ship itself; v O is the speed of the unmanned ship itself.

[0055] S12, obstacle position information includes: X-axis offset, Y-axis offset, Z-axis offset, yaw angle, ship speed, and X-axis offset, Y-axis offset, and Z-axis offset of static obstacles, that is,

[0056] X2={ΔX T ΔY T ΔZ T γ T v T ΔX J ΔY J ΔZ J}

[0057] Among them, {ΔX T ΔY T ΔZ T γ T v T} is the information of dynamic obstacles, which respectively represent the offset of dynamic obstacles in the X-axis direction, Y-axis direction, Z-axis direction, yaw angle and ship speed; {ΔX J ΔY J ΔZ J} is the information of static obstacles, which respectively represent the offset of static obstacles in the X-axis direction, the Y-axis direction, and the Z-axis direction.

[0058] S13, environmental information includes wind speed, wind direction and rainfall, that is,

[0059] X3={v W ar}

[0060] Among them, {v W a} represent wind speed and wind direction respectively, and γ represents rainfall. The change of wind force has a great influence on the stability and speed of the ship, while the wind direction will adjust the direction of the ship, and rainfall will have a certain impact on visibility.

[0061] S2. Assess the level of navigation risk based on environmental information.

[0062] Wind direction classification Figure 2 As shown in the table below, eight different directions, such as east, south, west, and north, are represented by numbers from 1 to 8; and rainfall and wind speed are divided into three levels, as shown in the table below:

[0063]

[0064] First, we set the risk level based on the change in wind speed. If the wind speed is within a safe range, the risk level is low; when the wind speed reaches a certain threshold, the risk level will rise to medium, and if it increases further, it will be high risk. At the same time, changes in wind direction will also affect the stability of navigation, so we will adjust the navigation route based on wind direction data to reduce risks.

[0065] Secondly, the amount of rainfall also directly affects the decision-making of the unmanned ship. If the rainfall is heavy, visibility will decrease, which will directly affect the navigation ability of the ship. Therefore, when the rainfall increases, the risk level will be increased in a timely manner.

[0066] S3. When the risk level exceeds the threshold, control avoidance operations.

[0067] In step S3, in order to simplify the processing flow of the unmanned ship decision system and improve its efficiency, the avoidance operation is converted and mapped, so that the originally separate heading and speed decision labels are merged into a unified column of labels.

[0068] "-1", "0", and "1" represent the decisions of speed v: deceleration, unchanged, and acceleration; "1", "2", and "3" represent the decisions of heading a: turn left, unchanged, and turn right. There are 9 different decisions, which are classified using 0-8, as shown in the following table:

[0069]

[0070] Among them, "0" means decelerating and turning left; "1" means decelerating and keeping the heading unchanged; "2" means decelerating and turning right; "3" means turning left with the same speed; "4" means turning the heading unchanged with the same speed; "5" means turning right with the same speed; "6" means accelerating and turning left; "7" means accelerating and keeping the heading unchanged; "8" means accelerating and turning right. This classification method not only makes the decision-making process clearer and more orderly, but also facilitates the system to quickly identify and execute, thereby ensuring that the unmanned ship can quickly make the most appropriate response when facing different navigation risks, ensuring the safety of navigation and the smooth completion of the mission.

[0071] In step S3, the avoidance operation is output by the CNN-RF hybrid model, and the input of the CNN-RF hybrid model is the unmanned ship posture information, environmental information, and obstacle posture information.

[0072] The collected and processed multi-dimensional data, including weather information, heading and speed decision labels, are used to build a deep learning network model, namely the CNN-RF network model.

[0073] The CNN-RF model is an algorithm that combines convolutional neural network (CNN) and random forest (RF) to extract features from input data and make effective decision classification. CNN is a multi-layer network similar to artificial neural network, which includes input layer, convolution layer, pooling layer, fully connected layer, and output layer. The CNN-RF model consists of two parts: CNN feature extraction module and RF classification module, and uses RF model to replace CNN fully connected layer and output layer. Figure 3 shown.

[0074] The convolution layer extracts features from the input data. Each filter slides on the input data, calculates the dot product of the filter and the local area of ​​the input data, and generates a feature map. The basic formula for convolution calculation is as follows:

[0075]

[0076] f(x)=max(0,x) (2)

[0077]

[0078] Where m and n represent the mth input neuron and the nth output neuron respectively; A is the set of input layer nodes of the hth layer of the feedforward neural network; h represents the hth convolutional layer; is the output of neuron n; Represents the input of the h-1th convolutional layer; is the convolution kernel; Represents the bias value of the nth neuron in layer h; It represents the output of the pooling layer h. is the input of the pooling window.

[0079] RF randomly extracts multiple subsets Y from the features extracted by CNN through bootstrapping k To train each decision tree, at each decision tree node, a subset F is randomly selected from all features k To find the best split point. For each bootstrap sample Y k and feature subset F k , build a decision tree. The splitting of decision tree nodes is done by selecting the best features and thresholds. Commonly used splitting criteria include Gini impurity and information gain, which are:

[0080]

[0081]

[0082] Among them, Gini(Y) represents the Gini impurity of data Y, J is the number of classes, P i represents the probability of the i-th category. Y is the data set of the current node, B is the considered feature, and Y1 and Y2 represent the two subsets obtained by splitting the threshold of feature B.

[0083] For classification problems, random forest predicts the class by voting among all decision trees:

[0084]

[0085] in, represents the predicted category, K is the total number of decisions, T k (x) represents the prediction of the kth decision tree for input x.

[0086] The trained multi-dimensional variable CNN-RF model is used to make intelligent decisions on the data collected by the surface unmanned system to ensure its safe and efficient navigation in the complex and changing environment of rivers, lakes and reservoirs.

[0087] The data collected by the unmanned system is divided, and the data of the first 5 moments are used to classify the speed and heading decisions of the data of the last moment, with a moving step of 1. First, the input data is feature extracted, and convolution and pooling operations are performed using two layers of 2*2 convolution kernels to extract local features, further reduce the data dimension, retain important information, reduce computational complexity, and enhance the expression ability of features. The data after feature extraction is sent to the RF model for decision classification.

[0088] Finally, the classification effect is tested through classification indicators such as accuracy, precision, F1 score and recall rate. The formula is as follows:

[0089]

[0090] Among them, TP is the number of positive samples correctly identified by the model; TN is the number of negative samples correctly identified by the model; FP is the number of negative samples incorrectly predicted as positive samples by the model; FN is the number of positive samples incorrectly predicted as negative samples by the model.

[0091] simulation:

[0092] In order for the unmanned ship to be able to safely complete the shipping mission in the designated area, the driving process can be described as follows: the unmanned ship starts from the designated starting point and follows a designated route, using radar and other sensor equipment to detect the location of obstacles and the intensity of rainfall and storms. When obstacle information and weather information are obtained and security threats are perceived, the ship will make autonomous judgments and actively change the path to avoid them. When the danger is lifted, the previously established route will be restored. The unmanned ship collision avoidance system consists of a perception module, a decision module, and an execution module. The perception module can summarize information from multiple aspects such as dynamic and static obstacles, weather, and drone status. The decision module consists of a deep neural network and outputs the desired decision based on the status information. The execution module controls the speed and heading based on the actual navigation parameters and heading of the unmanned ship. The above three parts are performed in sequence, thereby achieving real-time dynamic collision avoidance of the unmanned ship.

[0093] We selected 2043 data as the data set for this experiment, with 1341 items in the training set and 702 items in the test set. The accuracy distribution of CNN-RF in the same classification task is shown in the figure below: Figure 4 shown.

[0094] Overall, CNN-RF has a generally high accuracy rate in all categories. This is because CNN can effectively capture local and global features in the data and generate feature vectors with rich information through multi-layer convolution and pooling operations, while RF can effectively process nonlinear data and reduce overfitting by integrating the results of multiple decision trees, thereby improving the stability of classification. The combined model can process data of more dimensions, and is more comprehensive in feature representation and classification decisions, with strong generalization ability and classification performance. The classification ROC curve of the model is shown in Figure 2. Figure 5 shown.

[0095] In the case of imbalanced categories, the ROC curve can more accurately reflect the performance of the model than the accuracy rate. As can be seen from the figure, CNN-RF is close to the upper left corner and the AUC value fluctuations on all categories are small, indicating that the performance of the model on different categories is more consistent and more stable. The confusion matrix provides an intuitive view of the correct and incorrect predictions of the model on each category. The classification results are shown in Figure 1. Figure 6 As shown, 0, 1, 2, 3, 4, 5, 6, 7, and 8 in the horizontal and vertical axes correspond to 9 types of decisions respectively.

[0096] In the confusion matrix, the darker the color, the more obvious the characteristics of the same class of data are and the lower the possibility of being confused with the characteristics of other classes of data. Figure 6 The results show that the classification effect of CNN-RF in different categories is relatively consistent. CNN and RF each have their own unique advantages, and combining them can further optimize the performance of the model. After the data is classified by the CNN-RF model feature extraction, the 9-category decision has a more accurate classification result. The classification result evaluation indicators of the data set are shown in the following table:

[0097]

[0098] The CNN-RF model performed well in the classification performance evaluation, achieving 95.49% accuracy, 95.95% precision, 94.44% recall, and 95.14% F1 score. These indicators not only reflect the high efficiency of the model in the classification task, but also demonstrate its stability when processing complex data. The accuracy of this method reached 95.95%, indicating that the model has a strong ability to avoid errors for non-normal samples. The outstanding recall rate shows that this method can effectively identify all positive class samples with a low missed detection rate. The significant F1 score further confirms that the model can maintain high accuracy and high recall while showing that CNN-RF can make good use of the features extracted by CNN and the classification ability of RF to achieve the best classification effect, providing solid technical support for the intelligent decision-making of safe navigation of unmanned ships.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information, characterized in that: The specific method is as follows: Collect information, including the position information of unmanned ship, environment information, and obstacle position information; Assess the level of navigation risk based on environmental information; When the risk level exceeds the threshold, control avoidance operations; The avoidance operation is output by the CNN-RF hybrid model, and the input of the CNN-RF hybrid model is the unmanned ship posture information, environmental information, and obstacle posture information.

2. The intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information according to claim 1 is characterized in that: The unmanned ship posture information includes: unmanned ship speed, unmanned ship roll angle, unmanned ship pitch angle, unmanned ship yaw angle, unmanned ship X-axis direction offset, unmanned ship Y-axis direction offset and unmanned ship Z-axis direction offset; The environmental information includes: wind direction, wind speed and rainfall; The obstacle position information includes: obstacle ship speed, obstacle yaw angle, obstacle X-axis direction offset, obstacle Y-axis direction offset and obstacle Z-axis direction offset.

3. The intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information according to claim 1 is characterized in that: The navigation risk level is assessed based on environmental information. The specific method is as follows: The risk level is divided into low, medium and high by setting thresholds for wind speed and rainfall, as well as the distance to obstacles.

4. The intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information according to claim 1, characterized in that: Evasion maneuvers include: deceleration, maintaining speed, acceleration, left turn, maintaining heading, and right turn.

5. The intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information according to claim 1 is characterized in that: The CNN-RF hybrid model includes a CNN feature extraction module and an RF classification module.

6. The intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information as claimed in claim 5, characterized in that: The CNN feature extraction module includes convolutional layers; The convolution layer extracts features from the input. Each filter slides on the input data, calculates the dot product of the local area of ​​the filter and the input data, and generates a feature map. The convolution calculation formula is: f(x)=max(0,x) Where m and n represent the mth input neuron and the nth output neuron respectively; A is the input layer node set of the hth layer of the feedforward neural network; h represents the hth convolutional layer; is the output of neuron n; Represents the input of the h-1th convolutional layer; is the convolution kernel; Represents the bias value of the nth neuron in layer h; It represents the output of the pooling layer h. is the input of the pooling window.

7. The intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information as claimed in claim 5, characterized in that: The RF classification module randomly extracts multiple subsets Y from the features extracted by CNN through self-service sampling. k To train each decision tree, at each decision tree node, a subset F is randomly selected from all features k To find the best split point; For each bootstrap sample Y k and feature subset F k , build a decision tree, the splitting of decision tree nodes is done by selecting the best features and thresholds. Commonly used splitting criteria include Gini impurity and information gain, which are: Among them, Gini(Y) represents the Gini impurity of data Y, J is the number of classes, P i represents the probability of the i-th category; Y is the data set of the current node, B is the considered feature, and Y1 and Y2 represent the two subsets obtained by splitting the threshold of feature B; For classification problems, random forest predicts the category by voting among all decision trees. The specific formula is: in, represents the predicted category, K is the total number of decisions, T k (x) represents the prediction of the kth decision tree for input x.

8. The intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information as claimed in claim 1, characterized in that: CNN-RF hybrid model, usage is as follows: The unmanned ship posture information, environmental information and obstacle posture information are divided, and the data of the first 5 moments are used to classify the speed and heading decisions of the data of the last moment, and the moving step length is 1; First, perform feature extraction on the input data, and use two layers of 2*2 convolution kernels to perform convolution and pooling operations to extract local features; The data after feature extraction is sent to the RF model for decision classification.

9. The intelligent decision-making control method for the operation of an unmanned surface system based on multi-source information as claimed in claim 1, characterized in that: The CNN-RF hybrid model was tested, and the evaluation indicators included accuracy, precision, F1 score and recall. The specific formula is as follows: Among them, TP is the number of positive samples correctly identified by the model; TN is the number of negative samples correctly identified by the model; FP is the number of negative samples incorrectly predicted as positive samples by the model; FN is the number of positive samples incorrectly predicted as negative samples by the model.

Images