A situational reasoning method for unmanned surface vehicles

Abstract

The present invention discloses a situation reasoning method for surface unmanned boats, which includes the following steps: 1) modeling the situation elements of the surface unmanned boat's maritime mission based on the information of the surface unmanned boat, and modeling the cross-domain distributed mission situation elements; 2) situation encoding the modeled situation elements, extracting information related to situation prediction, and obtaining a situation matrix composed of task units; 3) inputting the situation-encoded data into an LSTM network, performing local action prediction and intention prediction on the task units, and then performing global intention prediction; 4) using a situation reasoning model to complete the situation information in the maritime environment. The present invention provides a situation reasoning method for surface unmanned boats, which provides strong support for unmanned boat mission planning and decision-making by modeling the situation elements of the unmanned boat, performing situation encoding, and training a situation reasoning model.

Description

Technical Field

[0001] The present invention relates to unmanned boat control technology, and in particular to a situation reasoning method for surface unmanned boats. Background Art

[0002] Situation analysis is the core of command and control for unmanned surface vehicles and the prerequisite for sound decision-making. How to predict and complete the current situation based on acquired situational information and historical data or prior knowledge, when information is uncertain and incomplete, is a key concern for scientists and engineers both domestically and internationally.

[0003] Due to the complex and changeable situation at sea, sensors are affected by the complex environment at sea, resulting in incomplete information. At the same time, there is less available training data on the marine environment. The existing situation reasoning technology is still insufficient for predicting the situation at sea. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a situation reasoning method for surface unmanned boats in view of the defects in the existing technology.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a situation reasoning method for an unmanned surface vehicle, characterized by comprising the following steps:

[0006] 1) Based on the information collected by the surface unmanned vehicle, a model of the surface unmanned vehicle maritime mission situation elements is established, and the cross-domain distributed mission situation elements are described by the model; the mission situation elements include environmental elements, static elements of the mission unit, and dynamic elements;

[0007] The environmental elements refer to the maritime environment information of the mission area; the static elements of the mission unit include the formation information of the unmanned boat and the type of mission performed by the unmanned boat; the dynamic elements of the mission unit are the position, speed and heading angle of the unmanned boat;

[0008] Among them, the grouping information of the unmanned boats is the grouping of the unmanned boats into subdivided task units;

[0009] 2) Situation coding is performed on the modeled situation elements to extract information related to situation prediction, and situation filtering is performed based on this information to obtain a situation matrix composed of task units;

[0010] Each row vector of the situation matrix represents the task unit A i ; Includes: (mission unit number, target identification, mission unit platform, position information, speed information, heading angle, loss status, sea state level);

[0011] 3) Input the situation-encoded data into the LSTM network to perform local action and intention prediction for the task unit, and then perform global intention prediction to train and obtain a situation reasoning model;

[0012] 4) Use situational reasoning models to complete situational information in the maritime environment.

[0013] According to the above scheme, step 3) is as follows:

[0014] 3.1) Predict local action of the task unit based on the situation coding data input;

[0015] The action prediction output of LSTM is two-dimensional discrete data [action, range]

[0016] Among them, the action prediction results include: acceleration, turning, moving forward at a constant speed, deceleration, and stopping;

[0017] The range of actions includes: urgent, rapid, smooth, slow;

[0018] 3.2) Predict the local intention of the task unit based on the situation coding data input;

[0019] The output of intention prediction is a one-dimensional discrete data tactic, which includes: attack, reconnaissance, escort, retreat, and standby;

[0020] 3.3) Define the input of the network at time t as n is the number of task units to be predicted, Use CNN network to extract features;

[0021] Among them, x1 is the task unit dynamic element corresponding to the action prediction of task unit i at time t, x1 = [x', y', z', v', α', β', θ'];

[0022] x2 is the intention prediction of task unit i, x3 is the inherent attribute of task unit i, x3 = [identification, platform, survival, sea state];

[0023] 3.4) Input the status of all task units at time t into the CNN and perform forward propagation to obtain the overall situation characteristics of the unmanned boat; the formula is as follows:

[0024] v t =conv2(W conv ,X t )+b conv

[0025] Among them, conv2 is the two-dimensional convolution layer function, W conv is the weight matrix of the convolutional layer, Xt is the input of the network at time t, b conv is the bias, v t It is the feature output by the convolutional network;

[0026] 3.5) Using recurrent neural network (RNN) units to memorize past states and integrate current state features to infer global intent;

[0027] Based on the overall situation characteristics of the unmanned boat t And the last output result of the recurrent neural network RNN ​​unit is used as input to obtain the output result of the RNN unit;

[0028] The output results are then classified through the output layer to obtain a global intention prediction. The global intention prediction is an overall situation judgment, including five possibilities: standby, reconnaissance, decision-making, strike, and evaluation;

[0029] 3.6) Define the global intention at time t as y t 1. Perform one-hot encoding on these five intentions. When y t 1=[1,0,0,0,0], it is in standby state; when y t 1=[0,1,0,0,0], it is in the reconnaissance state; when y t 1=[0,0,1,0,0], it is in the decision state; when y t 1=[0,0,0,1,0], it is in the attack state; when y t When 1=[0,0,0,0,1], it is in the evaluation state. The formula is as follows:

[0030] z1=f1(o t ×W1+b1)

[0031]

[0032] Among them, t is the output result of the RNN unit, W1 is the weight matrix of output layer 1, b1 is the bias of output layer 1, f1 is the activation function, and f1 is the Softmax function. is the global intention predicted at time t;

[0033] The loss function used in training is:

[0034]

[0035] Among them, p i is the probability that the sample belongs to the i-th category, p=[p1,p2,..,p c ]=z1,y i is the one-hot encoding representation of the sample, y t 1=[y1,y2,..,yc ], when the sample belongs to category i, y i =1, otherwise y i =0; C is the number of sample labels, C=5;

[0036] The result o output by the RNN unit t The winning rate is obtained through the output layer 2. The winning rate is a number representing probability. The winning rate at time t is defined as y t 2. The range is between 0 and 1. The closer it is to 0, the lower the winning rate is, and the closer it is to 1, the greater the probability of winning is. The formula is as follows:

[0037]

[0038] Among them, t is the output result of the RNN unit, W2 is the weight matrix of output layer 2, b2 is the bias of output layer 2, f2 is the activation function, is the predicted winning rate at time t;

[0039] The loss function used in training is:

[0040]

[0041] in, is the predicted winning rate at time t, y t 2 is the winning rate at time t.

[0042] The beneficial effects of the present invention are as follows: the present invention provides a situation reasoning method for surface unmanned boats under uncertain information conditions, by modeling the historical data and prior knowledge in the unmanned boat sea trials and performing situation encoding, the encoded situation information is used as the input of RNN and LSTM and a situation reasoning model is trained, and the trained situation reasoning model is used to complete the uncertain and incomplete situation information in the marine environment, providing strong support for the unmanned boat mission planning and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0044] Figure 1 It is a structural diagram of an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of task unit situation coding according to an embodiment of the present invention;

[0046] Figure 3 Schematic diagram of the action prediction and intention prediction network architecture according to an embodiment of the present invention;

[0047] Figure 4 2 is a schematic diagram of the recurrent neural network structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] like Figure 1 As shown, a situational reasoning method for a surface unmanned vehicle includes the following steps:

[0050] 1) Based on the modeling of the maritime mission situation elements of the surface unmanned vehicle, a mathematical model is used to describe the cross-domain distributed mission situation elements;

[0051] The data collected for modeling include historical data and prior knowledge from unmanned boat sea trials;

[0052] Mission situation elements include environmental elements, mission unit static elements and dynamic elements;

[0053] The environmental elements refer to the maritime environment information of the mission area; the static elements of the mission unit include the formation information of the unmanned boat and the type of mission performed by the unmanned boat; the dynamic elements of the mission unit are the position, speed and heading angle of the unmanned boat;

[0054] Among them, the grouping information of the unmanned boats is the grouping of the unmanned boats into subdivided task units;

[0055] 2) Situation coding is performed on the modeled situation elements to extract information related to situation prediction, and situation filtering is performed based on this information to obtain a situation matrix composed of task units;

[0056] like Figure 2 , each row of the situation matrix represents task unit A i ; Including mission unit number, target identification, mission unit platform, location information, speed information, heading angle, loss situation, and sea state level;

[0057] Loss refers to the energy and equipment loss of the unmanned boat;

[0058] Then the situation information is vectorized and customized normalized. The main work of situation coding is to process the original situation data. In addition, some static data (loss situation, sea state level) are input based on the set entity and scene parameters. Figure 2 The situation matrix shown.

[0059] When processing the original situation data to extract information related to situation prediction, the feature vector dimensionality reduction technology is used to project the high-dimensional vector after feature extraction into the low-dimensional space with the corresponding weight to obtain a new low-dimensional variable;

[0060] 3) Input the situation-encoded data into the LSTM network to perform local action and intention prediction for the task unit, and then perform global intention prediction to train and obtain a situation reasoning model;

[0061] The overall network structure is as follows Figure 3 , as follows:

[0062] 3.1) Predict local action of the task unit based on the situation coding data input;

[0063] The action prediction output of LSTM is two-dimensional discrete data [action, range]

[0064] Among them, the action prediction results include: acceleration, turning, moving forward at a constant speed, deceleration, and stopping;

[0065] The range of actions includes: urgent, rapid, smooth, slow;

[0066] 3.2) Predict the local intention of the task unit based on the situation coding data input;

[0067] The output of intention prediction is a one-dimensional discrete data tactic, which includes: attack, reconnaissance, escort, retreat, and standby;

[0068] The loss function of LSTM is:

[0069]

[0070] in, is the action and situation prediction result marked in the training sample, The action and situation results output by the LSTM network. i = 1, 2, and 3 correspond to the three-dimensional data output by the LSTM network.

[0071] 3.3) Define the input of the network at time t as n is the number of task units to be predicted, Use CNN network to extract features;

[0072] Among them, x1 is the task unit dynamic element corresponding to the action prediction of task unit i at time t, x1 = [x', y', z', v', α', β', θ'];

[0073] x2 is the predicted intent of task unit i, and x3 is the inherent attribute of task unit i, where x3 = [identification, platform, survival, sea state]. Therefore, the input can be viewed as a two-dimensional matrix, with each row representing the complete information of a task unit at the current moment. The states of all task units at time t are input into the CNN, and forward propagation is performed to obtain the overall situational characteristics; the formula is as follows:

[0074] v t =conv2(W conv ,X t )+b conv

[0075] Among them, conv2 is the two-dimensional convolution layer function, W conv is the weight matrix of the convolutional layer, X t is the input of the network at time t, b conv is the bias, v t It is the feature output by the convolutional network.

[0076] Global intent prediction depends not only on the current state of all task units but also on past action states. Therefore, the network must be able to comprehensively predict intent and determine success rate based on both past and current states. Recurrent neural networks are based on the idea of ​​a memory model, remembering previously presented features and inferring future outcomes based on these features. Therefore, recurrent neural network (RNN) units are used to memorize past states and integrate current state features to make global intent and success rate inferences.

[0077] Based on the overall situation characteristics of the unmanned boat t And the last output result of the recurrent neural network RNN ​​unit is used as input to obtain the output result of the RNN unit;

[0078] Recurrent neural network RNN ​​unit such as Figure 4 As shown, where U is the input v t The weight matrix, W is the last output value o t-1 As the input weight matrix, f is the activation function, H is the weight matrix of the RNN unit output layer, g is the activation function of the RNN unit output layer, s t is the value of the hidden layer, o t Output results for the RNN unit;

[0079] The output results are then classified through the output layer to obtain a global intention prediction, which includes five possibilities: standby, reconnaissance, decision-making, strike, and evaluation.

[0080] 3.4) Define the global intention at time t as y t 1. Perform one-hot encoding on these five intentions. When y t 1=[1,0,0,0,0], it is in standby state; when y t 1=[0,1,0,0,0], it is in the reconnaissance state; when y t 1=[0,0,1,0,0], it is in the decision state; when y t 1=[0,0,0,1,0], it is in the attack state; when yt When 1=[0,0,0,0,1], it is in the evaluation state. The formula is as follows:

[0081] z1=f1(o t ×W1+b1)

[0082]

[0083] Among them, t is the output result of the RNN unit, W1 is the weight matrix of output layer 1, b1 is the bias of output layer 1, and f1 is the activation function. Since this is a classification problem, f1 is the Softmax function. is the global intention predicted at time t. The goal during training is to make the network classify as correctly as possible, so its loss function is:

[0084]

[0085] where p i is the probability that the sample belongs to the i-th category, p=[p1,p2,..,p c ]=z1,y i is the one-hot encoding representation of the sample, y t 1=[y1,y2,..,y c ], when the sample belongs to category i, y i =1, otherwise y i = 0. C is the number of sample labels, and in this project, C = 5.

[0086] The result o output by the RNN unit t The winning rate is obtained through the output layer 2. The winning rate is a number representing probability. The winning rate at time t is defined as y t 2. The range is between 0 and 1. The closer it is to 1, the greater the probability of winning. The formula is as follows:

[0087]

[0088] Among them, t is the output result of the RNN unit, W2 is the weight matrix of output layer 2, b2 is the bias of output layer 2, f2 is the activation function, is the predicted winning rate at time t. The loss function is:

[0089]

[0090] in is the predicted winning rate at time t, y t 2 is the winning rate at time t.

[0091] After the LSTM network obtains the action prediction and intention judgment of each task unit, it adds the inherent attributes of the task unit as input, extracts features through CNN, passes through the RNN network unit, and obtains the global intention and winning rate through the output layer.

[0092] 4) Use the trained situational reasoning model to complete the uncertain and incomplete situational information in the maritime environment, providing strong support for unmanned boat mission planning and decision-making.

[0093] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A situation reasoning method for unmanned surface vehicles, characterized in that: The following steps are involved: 1) Model the maritime mission situation elements of the surface unmanned vehicle based on the information of the surface unmanned vehicle, and describe the cross-domain distributed mission situation elements; the mission situation elements include environmental elements, static elements of the mission unit, and dynamic elements; The environmental elements refer to the maritime environment information of the mission area; the static elements of the mission unit include the formation information of the unmanned boat and the type of mission performed by the unmanned boat; the dynamic elements of the mission unit are the position, speed and heading angle of the unmanned boat; Among them, the grouping information of the unmanned boats is the grouping of the unmanned boats into subdivided task units; 2) Situation coding is performed on the modeled situation elements to extract information related to situation prediction and obtain a situation matrix composed of task units; Each row vector of the situation matrix represents the task unit A i ; Including: mission unit number, target identification, mission unit platform, location information, speed information, heading angle, loss status, and sea state level; 3) Input the situation-encoded data into the LSTM network to perform local action and intention prediction for the task unit, and then perform global intention prediction to train and obtain a situation reasoning model; The step 3) is specifically as follows: 3.1) Predicting local actions of the task unit based on situation coding data input; The action prediction output of LSTM is two-dimensional discrete data [action, range] in, Action prediction results include: acceleration, turning, moving forward at a constant speed, deceleration, and stopping; The range of actions includes: urgent, rapid, smooth, slow; 3.2) Predict the local intention of the task unit based on the situation coding data input; Intent prediction output is one-dimensional discrete data , Including: attack, reconnaissance, escort, retreat, standby; 3.3) Definition The input of the network at this moment is , is the number of task units to be predicted, , use CNN network to extract features; in, for Time task unit The dynamic elements of the task unit corresponding to the action prediction, ; Task Unit Intention prediction, Task Unit The inherent properties of [identification, platform, survival, sea conditions]; 3.4) The states of all task units at the moment are input into CNN for forward propagation to obtain the overall situation characteristics of the unmanned boat; the formula is as follows: in, is the two-dimensional convolution layer function, is the weight matrix of the convolutional layer, for The input of the network at that moment, is the bias, It is the feature output by the convolutional network; 3.5) Using recurrent neural network (RNN) units to memorize past states and integrate current state features to infer global intent; Based on the overall situation characteristics of the unmanned boat And the last output result of the recurrent neural network RNN ​​unit is used as input to obtain the output result of the RNN unit; The output results are then classified through the output layer to obtain a global intention prediction. The global intention prediction is an overall situation judgment, including five possibilities: standby, reconnaissance, decision-making, strike, and evaluation; 3.6) Definition The global intention at all times is , perform one-hot encoding on these five intentions, then when When When When When When , it is in the evaluation state; the formula is as follows: in, Output result for RNN unit, is the weight matrix of output layer 1, is the bias of output layer 1, As the activation function, take is the Softmax function, for Global intentions to be predicted at all times; The loss function used in training is: in, The sample belongs to The probability of the class, , is the one-hot encoding representation of the sample, , when the sample belongs to the category hour, ,otherwise ; is the number of sample labels, ; The result of RNN unit output After the output layer 2, we get the winning rate, which is a number representing the probability. The winning rate at the moment is , the range is The closer it is to 0, the lower the winning rate is, and the closer it is to 1, the greater the probability of winning is. The formula is as follows: in, Output result for RNN unit, is the weight matrix of output layer 2, is the bias of output layer 2, is the activation function, for Winning rate predicted at each moment; The loss function used in training is: in, For prediction Time winning rate, for Time win rate; 4) Use situational reasoning models to complete situational information in the maritime environment.

Images