A method for classifying a torpedo trajectory based on an LSTM network

Abstract

The application discloses a method for classifying a torpedo trajectory based on an LSTM network, and comprises the following steps: 1) establishing a pursuit scene model of a torpedo and a submarine, which is divided into three types, i.e., a tail pursuit method of zero lead angle guidance of an air-launched torpedo, a fixed lead angle method of a ship-launched torpedo, and a parallel approach method of a submarine-launched torpedo; 2) performing simulation and simulation based on mathematical model features of the trajectory; 3) establishing a data set; 4) establishing a trajectory classification model based on the LSTM network; and 5) training based on the LSTM network. The method has the following advantages: the LSTM is introduced to solve the problem of discriminating the type of the torpedo through the torpedo trajectory, and the problem of strong constraint and poor robustness of current torpedo trajectory classification is overcome, so that more time is obtained for intercepting the torpedo; the problem that a traditional mathematical model method does not have a self-correction ability for parameter errors and cannot quantitatively evaluate and judge a result in a clear probability standard is overcome; and the problem that a current deep learning-based method has poor effect on complex torpedo time sequence data and may have problems of gradient explosion and gradient disappearance is overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of torpedo detection, and particularly relates to a torpedo trajectory classification method based on an LSTM network. BACKGROUND

[0002] Torpedo has become one of the biggest threats to submarines in sea battles due to its high concealment, great destructive power and few countermeasures. In order to establish an effective torpedo defense system, relevant scholars have conducted a large amount of research. Among the many technical links of the torpedo defense system, classifying torpedoes according to trajectory clusters is a key link, so as to effectively counter. Although torpedo trajectory classification methods have emerged in recent years, it is still a challenging topic to design and identify torpedo trajectories in various scenarios according to mathematical models due to the high volatility and multiple variables of modern marine environments.

[0003] In the paper "Research on the Prediction Method of Incoming Torpedo Trajectory under Azimuth Condition Only" (Journal of Naval University of Engineering, 2016, 28(02): 108-112) by Yin Wenjin, Zhang Jingyuan, Zhang Yongping and Rao Zhe, a torpedo azimuth information detected by a sonar system is proposed for the prediction of an incoming torpedo trajectory under azimuth condition. The movement elements are reasonably discriminated and calculated by backstepping, that is, the estimated speed, the estimated and calculated movement heading set. However, this method contains multiple interval analyses, which can easily miss reasonable values while eliminating unreasonable torpedo movement routes. Traditional mathematical model-based methods, such as the likelihood trajectory prediction method, do not have the ability to self-correct parameter errors and cannot quantitatively evaluate the judgment results with a clear probability standard. The current deep learning-based methods, such as BP neural networks, are not very ideal for complex torpedo time series data and may have problems such as gradient explosion and gradient disappearance. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art, introduce LSTM to solve the problem of discriminating torpedo types through torpedo trajectories, utilize the powerful time series analysis and prediction ability of the neural network to realize the classification of torpedo trajectory clusters, and propose a torpedo trajectory classification method based on an LSTM network.

[0005] The technical solution adopted by the present application to solve the technical problem is: a torpedo trajectory classification method based on an LSTM network, characterized by comprising the following steps:

[0006] Step 1: Establish a pursuit scenario model of torpedoes and submarines, which is divided into three types: tail pursuit method of zero lead angle guidance of air-launched torpedoes; fixed lead angle method of ship-launched torpedoes; and parallel approach method of submarine-launched torpedoes;

[0007] Step 2: Simulation based on trajectory-based mathematical model features: set the torpedo launch point as the fixed origin, and the offset angle of the torpedo from the submarine position as the launch lead angle. Randomly generate target set trajectories under different launch speeds, target speeds, etc. that conform to a uniform distribution;

[0008] Step 3: Establish a data set: according to the simulated trajectory data, classify and label them according to the three types in step 1, then normalize the data in each category as the model's time series input; 70% of the total data is used as the training set for the neural network, and the remaining 30% of the total data is used as the test set;

[0009] Step 4: Establish a trajectory classification model based on LSTM network: composed of an input layer, two LSTM layers, a fully connected layer, a SoftMax layer, and an output layer; the initial lead angle of the torpedo, the angle change rate, the relative distance from the submarine, the distance change rate, the submarine speed, the torpedo speed, etc. are used as the input of the model, and the torpedo type is used as the output of the model;

[0010] Step 5: Train based on LSTM network: set the appropriate parameters, including optimizer, learning rate, and iteration number, and use the training set to train the model; finally, get the trained model and use the test set to judge the accuracy.

[0011] Preferably, the method for classifying the trajectory of a torpedo based on an LSTM network comprises the following steps:

[0012] The torpedo trajectory in step 1 has three models, and the relative motion equation when the torpedo is guided to move is:

[0013]

[0014] where r is the relative distance between the torpedo and the target position, V T is the submarine's sighting line; q is the angle between the submarine's sighting line and a certain baseline of the attack plane, referred to as the submarine's bearing angle; σ V , σ T are the angles between the torpedo and submarine speed vectors and the baseline of the attack plane, referred to as the torpedo trajectory angle and the submarine heading angle; η V , η T are the angles between the torpedo and submarine speed vectors and the submarine's sighting line, referred to as the torpedo speed lead angle and the submarine speed lead angle; η V = 0 is called the tail-chasing method; η V is a constant, it is called the fixed lead angle method; η V changes according to certain rules, but in any case, the line of sight is parallel to its own movement, which is called the parallel approach method.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention introduces LSTM to solve the problem of identifying torpedo types based on torpedo trajectories. It utilizes the powerful time series analysis and prediction capabilities of this neural network to classify torpedo trajectory clusters, proposing a new method for determining torpedo guidance patterns based on torpedo trajectories. This overcomes the problems of strong constraints and poor robustness in current torpedo trajectory classification, thus gaining more time for torpedo interception. It also overcomes the problems of traditional mathematical model-based methods lacking self-correction capabilities for parameter errors and being unable to quantitatively evaluate judgment results with clear probability standards. Furthermore, it overcomes the problems of current deep learning-based methods performing poorly on complex torpedo time series data and potentially suffering from gradient explosion and gradient vanishing. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps in Embodiment 1 of the present invention;

[0017] Figure 2 This is a network structure diagram of Embodiment 1 of the present invention;

[0018] Figure 3 This is a confusion matrix diagram comparing the predicted label and the actual label in Embodiment 1 of the present invention;

[0019] Figure 4 This is a schematic diagram showing the relative positions of the torpedo (V) and the submarine (T) in Embodiment 1 of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0021] Example 1

[0022] like Figures 1 to 4 As shown, the LSTM network is also known as the Long Short-Term Memory (LSTM) network.

[0023] The flowchart for constructing a torpedo trajectory classification and prediction model based on LSTM network is as follows: Figure 1 As shown, the specific implementation method includes the following steps:

[0024] (1) Establish a torpedo and submarine pursuit scenario model, divided into three categories: tail-chase method with zero lead angle guidance for air-dropped torpedoes; fixed lead angle method with ship-launched torpedoes; and parallel approach method with submarine-launched torpedoes. There are three torpedo trajectory models, and the relative motion equations during torpedo guidance are as follows:

[0025]

[0026] Where r is the relative distance between the torpedo and the target position, and V Tis the angle between the submarine's line of sight and a certain baseline of the attack plane, called the submarine's azimuth. q V , q T are the angles between the torpedo, submarine speed vector and the baseline of the attack plane, called the torpedo's trajectory angle and the submarine's heading angle. η V , η T are the angles between the torpedo, submarine speed vector and the submarine's line of sight, called the torpedo's speed lead angle and the submarine's speed lead angle. η V = 0 is called the tail-chase method; η V is a constant is called the fixed lead angle method; η v changes according to certain rules, but in any case, the line of sight is parallel to the self-moving guide method is called the parallel approach method.

[0027] (2) Simulation based on the mathematical model characteristics of the trajectory: According to the mathematical model characteristics of different motion trajectories, the pursuit scene model of the torpedo and the submarine is established, and simulation is carried out. A certain type of torpedo is used as the incoming target, and the launch method is submarine launch, ship launch and air launch. According to the launch platform, different initial positions of the incoming torpedo are set, the submarine position is randomly set, and the torpedo trajectory is generated under the design of different launch speed and fixed lead angle. The torpedo trajectory is guided according to the motion model with acceleration curve. Each group of different initial angle, speed and acceleration initial setting value of the torpedo; each group of different initial angle, speed and acceleration initial setting value of the submarine as the initial launch condition, the simulation of the incoming torpedo of different launch platforms.

[0028] (3) Establish a data set: At the initial moment of the relative motion of the projectile (torpedo and target), the distance between the enemy and ourselves is far apart, and the sampling frequency is set to 50Hz. The motion trajectory of the torpedo during the incoming process is sampled, so the incoming torpedo from the initial moment to the end of the guidance process is taken as the end, and the sampled data in the guidance process is output as the original data document, including 6 categories such as torpedo speed, target speed, projectile distance, projectile distance change, projectile azimuth angle, and projectile azimuth angle change. The data set file is integrated and processed, and three labeled data files of air drop, ship launch and submarine launch are made, of which 70% is the training set and the remaining 30% is the test set.

[0029] (4) Establish a trajectory classification model based on LSTM network: It is composed of input layer, two LSTM layers, one fully connected layer, one SoftMax layer and one output layer; the initial lead angle, angle change rate of the torpedo, relative distance, distance change rate, submarine speed, torpedo speed, etc. are taken as the input of the model, and the torpedo type is taken as the output of the model. The specific structure of the network is as follows: Figure 2As shown, it comprises an input layer, two LSTM layers, a full connection layer, a SoftMax layer and an output layer. The first LSTM layer contains 50 neurons and outputs 50 dimensions; the second LSTM layer contains 50 neurons and outputs 50 dimensions; the full connection layer contains 50 neurons and outputs 3 dimensions; and finally, a SoftMax layer is connected to output the classification trajectory results.

[0030] (5) Training of the torpedo trajectory classification model based on the LSTM network: the training round is set to 100, the initial learning rate is set to 0.001, the batch size is 256, the training is performed on the GPU, and finally the trained network is saved.

[0031] (6) The trained model is used to test the divided test set, and finally the confusion matrix of the predicted classification trajectory and the actual classification trajectory is obtained (the horizontal coordinate is the predicted trajectory label, and the vertical coordinate is the actual trajectory label).

[0032] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in other forms. Any skilled person in the art can modify or change the above disclosed technical content to obtain equivalent embodiments. However, any simple modification, equivalent change and modification of the above embodiments made without departing from the technical scheme of the present application, according to the technical essence of the present application, still belongs to the protection scope of the technical scheme of the present application.

Claims

1. A torpedo trajectory classification method based on LSTM network, characterized in that: Includes the following steps: Step 1: Establish a torpedo and submarine pursuit scenario model, divided into three categories: tail pursuit method with zero lead angle guidance for air-dropped torpedoes; fixed lead angle method with ship-launched torpedoes; and parallel approach method with submarine-launched torpedoes. Step 2: Simulation based on the mathematical model features of the trajectory: Set the torpedo launch point as a fixed origin, set the offset angle of the line connecting the torpedo and the submarine position as the launch advance angle, and randomly generate target trajectories under different launch velocities and target velocities that follow a uniform distribution. Step 3: Establish dataset: Based on the simulated trajectory data, classify and label it according to the three types in Step 1, and then normalize the data in each category as the time series input of the model; use 70% of the total data as the training set of the neural network, and the other 30% as the test set. Step 4: Establish a trajectory classification model based on an LSTM network: consisting of an input layer, two LSTM layers, a fully connected layer, a SoftMax layer, and an output layer; the initial lead angle of the torpedo, the rate of change of the angle, the relative distance with the submarine, the rate of change of the distance, the submarine speed, and the torpedo speed are used as the inputs to the model, and the torpedo type is used as the output of the model. Step 5: Training based on LSTM network: Set the corresponding parameters, including optimizer, learning rate, and number of iterations, and train the model using the training set mentioned above; finally, obtain the trained model and use the test set to judge the accuracy.

2. The torpedo trajectory classification method based on LSTM network according to claim 1, characterized in that: There are three models for the torpedo trajectory described in step 1, and the relative motion equations during torpedo guidance motion are as follows: Where r is the relative distance between the torpedo and the target position, V T σ is the submarine's aiming line; q is the angle between the submarine's aiming line and a baseline of the attack plane, called the submarine's azimuth angle; V σ T These are the angles between the torpedo's and submarine's velocity vectors and the baseline of the attack plane, respectively, called the torpedo trajectory angle and the submarine's heading angle; η V η T These are the angles between the torpedo's velocity vector, the submarine's velocity vector, and the line of sight to the submarine, respectively, and are called the torpedo velocity lead angle and the submarine velocity lead angle. η V The method when = 0 is called the tail-chasing method; η V When the angle is constant, it is called the fixed advance angle method; η V The guiding method that changes according to a certain pattern, but always keeps the line of sight parallel to the movement of itself, is called the parallel approach method.

Images