An LSTM and transfer learning-based submarine intent prediction method

Abstract

The application discloses a kind of based on LSTM and transfer learning's submarine intention prediction method, it is related to submarine target intention prediction field, including the following steps: S1: establish virtual submarine intention dataset;S2: to the submarine intention dataset established does filtering processing;S3: using the virtual dataset that has done filtering processing trains LSTM neural network, obtains pre-training model LSTM1;S4: collection submarine real intention data, obtains experimental training set, by transferring learning LSTM1 model parameter moves to experimental application scene, using experimental training set is trained and obtains neural network model LSTM2.The application adopts the above-mentioned one based on LSTM and transfer learning's submarine intention prediction method, solves the problem that submarine is not high in the case of small amount of data prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of submarine target intention prediction, and particularly relates to a submarine intention prediction method based on LSTM and transfer learning. BACKGROUND

[0002] In the sea war, target intention prediction is an indispensable link. Target intention prediction is to obtain the intention of the next target by observing the behavior and data of the underwater target. If we can know the intention of the other party in advance from the data, it is very beneficial for the appropriate behavior of the attack and defense decision of the submarine. With the development of artificial intelligence technology, the combination of intelligent technology and modern sea battlefield has become the main core content of submarine combat. If the target intention prediction and artificial intelligence technology are combined together, a target intention prediction system is designed, which has important significance for maintaining national maritime security.

[0003] In the patent application No. 202010031249.4, the name is "a multi-target trajectory prediction method based on non-collision LSTM", which uses long short-term memory neural network (LSTM) as a trajectory prediction method. (Liu Zhaodong, Chen Mou, Wu Qingxian, Chen Shaodong. Target intention prediction of unmanned aerial vehicle air combat under incomplete information [J]. China science: information science, 2020, 50 (05): 704-717.) also uses LSTM to design an unmanned aerial vehicle target intention prediction system, but the above-mentioned target intention prediction method is to train the LSTM network by using a large amount of actual data set to obtain the prediction result. However, due to the limited and incomplete data of the underwater submarine cluster navigation experiment, it is difficult to train the intention prediction neural network by collecting a large amount of data, which leads to the difficulty of directly applying the existing intention prediction method to the underwater submarine intention prediction problem.

[0004] Therefore, it is necessary to provide a submarine intention prediction method based on LSTM and transfer learning to solve the above problems. SUMMARY

[0005] The purpose of the present application is to provide a submarine intention prediction method based on LSTM and transfer learning, which can quickly train and predict the target intention under the condition of a small amount of experimental data of the submarine, and improve the reliability of the target intention prediction.

[0006] To achieve the above purpose, the present application provides a submarine intention prediction method based on LSTM and transfer learning, comprising the following steps:

[0007] S1: establishing a virtual submarine intention data set;

[0008] S2: filtering the established submarine intention data set;

[0009] S3: training the LSTM neural network using the filtered virtual data set to obtain a pre-trained model LSTM1;

[0010] S4: collecting real submarine intent data to obtain an experimental training set, migrating the LSTM1 model parameters to an experimental application scenario through transfer learning, and training a neural network model LSTM2 using the experimental training set.

[0011] Preferably, in step S1, the specific method for generating the virtual submarine intent data set is:

[0012] S11: establishing a virtual intent data set Z all according to real submarine data attack , retreat , feint , patrol , scout ;

[0013] wherein Z attack = {Z attack_1 ,…,Z attack_r} is a data set with an attack intent and contains r samples;

[0014] Z retreat = {Z retreat_1 ,…,Z retreat_u} is a data set with a retreat intent and contains u samples;

[0015] Z feint = {Z feint_1 ,…,Z feint_o} is a data set with a feint intent and contains o samples;

[0016] Z patrol = {Z patrol_1 ,…,Z patrol_q} is a data set with a patrol intent and contains q samples;

[0017] Z scout = {Z scout_1 ,…,Z scout_w} is a data set with a scout intent and contains w samples;

[0018] one∈{attack,retreat,feint,patrol,scout},Z one_ik is the state information of the i-th sample of the intent represented by one at the k-th time, and Z one_ik = [P one_ik ,V one_ik ] T ;

[0019] wherein P one_ik = [P x_one_ik , P y_one_ik ] T one represents the projection of the i-th sample of the intention on the horizontal plane x-axis and y-axis at time k;

[0020] V one_ik = [V x_one_ik , V y_one_ik ] T one represents the projection of the i-th sample of the intention on the horizontal plane x-axis and y-axis at time k;

[0021] S12: Z one_ik The data generation method of the velocity and position is shown in the following formula:

[0022]

[0023] wherein k one_i = diag{k x_one_i , k y_one_i} is the given control gain of the i-th sample of the intention represented by one, k x_one_i is the control gain in the horizontal plane x-axis direction, and k y_one_i is the control gain in the horizontal plane y-axis direction; goal one_i = [goal x_one_i , goal y_one_i ] T , goal x_one_i and goal y_one_i are the target positions of the i-th sample of the intention represented by one on the horizontal plane x-axis and y-axis, respectively; P one_ik = [P x_one_ik , P y_one_ik ] T , P x_one_ik and P y_one_ik are the positions of the i-th sample of the intention represented by one on the horizontal plane x-axis and y-axis at time k; V one_ik = [V x_one_ik , V y_one_ik ] T , V x_one_ik and V y_one_ik are the velocities of the i-th sample of the intention represented by one on the horizontal plane x-axis and y-axis at time k; wherein P one_i(k+1) = [P x_one_i(k+1) , P y_one_i(k+1) ] T , P x_one_i(k+1) , P y_one_i(k+1)Let be the projections of the i-th sample in the intent represented by 'one' onto the horizontal x-axis and y-axis at time (k+1); Δt is the difference between the two times; noise = [noise] x_one_ik noise y_one_ik ] T noise x_one_ik and noise y_one_ik These represent the random noise along the x-axis and y-axis of the horizontal plane at time k for the i-th sample in the intent represented by 'one', with values ​​ranging from [-1.2m to 1.5m].

[0024] S13: Label the generated virtual intent dataset as follows: Z attack ={Z attack_1 ,…,Z attack_r In the dataset, r samples are manually labeled with a single '1'. retreat ={Z retreat_1 ,…,Z retreat_u In the data, u samples were manually labeled with a 2-fold withdrawal; Z feint ={Z feint_1 ,…,Z feint_o In the data, o samples were manually labeled with 3 tags; Z patrol ={Z patrol_1 ,…,Z patrol_q In the}, q samples are manually labeled with 4 patrols; Z scout ={Z scout_1 …,Z scout_w The data of w samples in the} were manually labeled with 5 tags.

[0025] Preferably, the filtering process performed on the target intent dataset established in S1 in step S2 specifically involves:

[0026] S21. Apply a Savitzky-Golay filter to the data in S1 for smoothing. The process is as follows:

[0027] Z-axis of the underwater vehicle intent dataset one_ik State information, construct an (n-1)th order polynomial to fit Z one_ik Status data:

[0028]

[0029] Where a0, a1, a2, ..., a n-1 It is a fit to P x_one_ik The polynomial constant terms, b0, b1, b2, ..., b n-1 It is a fit to P y_one_ik The polynomial constant terms, c0, c1, c2, ..., c n-1 It is a fit to Vx_one_ik the polynomial constant term, d0, d1, d2, …, d n-1 is the polynomial constant term of the fitting V y_one_ik k k k k is the residual of the fitting process, k, k 2 , …, k n-1 is the polynomial degree of the fitting;

[0030] Select the state information at {Z one_i(k-m) , Z one_i(k-m+1) , …, Z one_ik , …, Z one_i(k+m-1) , Z one_i(k+m)} 2m+1 time points as the filter window, calculate using the above formula, get 2m+1 formula, the matrix composed of formula is represented as:

[0031]

[0032] Simplify the above matrix to the following formula:

[0033] Z (2m+1)×4 = X (2m+1)×n ·R n×4 +E (2m+1)×4

[0034] Where:

[0035] In the formula: Z, X, R, E subscript is the parameter dimension, through the least square method, the least square solution of R is:

[0036]

[0037] The submarine data filtering value of Z is:

[0038]

[0039] Where H = X·(X T ·X) -1 ·X T is the relationship matrix between the filtered data and the original data;

[0040] The filter step is 2m+1, when the center point of the filter window is k time data filtering is completed, the center point is moved to k+2m+1 time and the above filtering process is repeated until the i th sample all time state information filtering is finished, the filtered Z one_sg_i is obtained;

[0041] ​​​​Finally all the samples in Z all = {Z attack , retreat , feint , patrol , scout} are filtered to obtain filtered data Z all_sg = {Z attack_sg , retreat_sg , feint_sg , patrol_sg , scout_sg};

[0042] Z one_sg_ik is the data of the i-th sample in the filtered intent represented by one at the k-th time point;

[0043] S22: convert the filtered data in S21 into the input data format required by LSTM1, specifically: Z one_sg_i = [Z one_sg_i1 ,..., Z one_sg_iN ], where N is the number of sampling points in the i-th sample, p is the number of LSTM1 long short-term memory units, and Z one_sg_i is cut into N-p+1 sub-sample data as follows: {Z one_sg_i_1 ,..., Z one_sg_i_j ,..., Z one_sg_i_(N-p+1)} = { [Z one_sg_i1 ,..., Z one_sg_ip ],..., [Z one_sg_ij ,..., Z one_sg_i(p+j-1) ],..., [Z one_sg_i(N-p+1) ,..., Z one_sg_iN},

[0044] Z one_sg_i_j is the data of the i-th sample in the filtered intent represented by one, which is cut into the j-th sub-sample, and Z all_sg = {Z attack_sg , Z retreat_sg , Z feint_sg , Z patrol_sg , Z scout_sg} all samples are cut to obtain all sub-sample data to be input to LSTM1,

[0045] Z all_sg_fin = {Z attack_sg_fin , Z retreat_sg_fin , Z feint_sg_fin , Z patrol_sg_fin , Z scout_sg_fin} ; wherein Z attack_sg_fin is the sub-sample data set of the attack intent after cutting; Z retreat_sg_finZ feint_sg_fin Z patrol_sg_fin Z scout_sg_fin Z

[0046] In step S3, preferably, an LSTM neural network composed of an input layer, a long short-term memory layer, an output layer, and a softmax layer is built, and the specific steps are as follows:

[0047] S31: The input layer is to input the sub-sample data in the submarine virtual data set in S22 to the long short-term memory layer;

[0048] The long short-term memory layer is composed of p long short-term memory units, and the calculation formula is as follows:

[0049] f k = σ(W f · [h k-1 , x k ] + b f )

[0050] i k = σ(W i · [h k-1 , x k ] + b i )

[0051] o k = σ(W o · [h k-1 , x k ] + b o )

[0052] c k = f k *c k-1 +i k *tanh(W c · [h k-1 , x k ] + b c )

[0053] h k = o k *tanh(c k )

[0054] x k is the state information of Z one_sg_ik in the input network; i k , f k and o kThese represent the input gate, forget gate, and output gate at time k in the long short-term memory layer, respectively; σ is the sigmoid function; W f W i W o W c b f b i b o b c These are the parameters obtained after the network is trained; tanh is the activation function; c k c represents the output of the candidate gate of the LSTM neural network at time k; k-1 This represents the output of the candidate gate of the LSTM neural network at time k-1; h k This represents the output of the LSTM neural network at time k; h k-1 This represents the output of the LSTM neural network at time k-1; * indicates the product of corresponding elements of two vectors;

[0055] The output layer is a fully connected layer;

[0056] The softmax layer outputs the information from the output layer in the form of probabilities, and the calculation formula is as follows:

[0057]

[0058] in, y i ∈{1,2,3,4,5} corresponds to the labels in S13 indicating intentions of retreat, attack, feint, patrol, and reconnaissance. Sum the exponential outputs for all predicted intentions;

[0059] S32: The pre-trained model LSTM1 is obtained as follows:

[0060] S321. Divide the filtered samples in S2 into a training set and a test set; the training set and the test set account for 80% and 20% of the total samples, respectively.

[0061] S322. Use the multi-class cross-entropy loss function (Cross Entropy Loss) to verify the model's accuracy. The calculation method is as follows:

[0062]

[0063] Where m represents the total number of subsample data of the underwater vehicle, and n represents the total number of intention types of the underwater vehicle. This represents the probability that the i-th real sample is j in the context of the underwater vehicle's intention to attack, retreat, feint, patrol, or reconnaissance. This represents the probability that the i-th predicted sample is j in the context of the submarine's intention to attack, retreat, feint, patrol, or reconnaissance.

[0064] S323: Put the labeled data in the training set in S322 into the network built in S31 for training, save the parameters with the highest accuracy rate using the labeled test set data, and obtain the LSTM1 model parameters.

[0065] Preferably, in step S4, the neural network model LSTM2 obtained by training using the experimental training set comprises the following steps:

[0066] S41: The real intention data is divided into a training set and a test set, which respectively account for 80% and 20% of the real intention data, and are labeled with corresponding attack, retreat, feint, patrol and reconnaissance labels. The labeled training set and test set data are subjected to the filtering operation in step S2, and then subjected to the format operation for sending into the network in S322, to obtain data to be input into the migration network;

[0067] S42: The parameters of the input layer and the long short-term memory layer in the LSTM1 model are frozen, only the parameters of the fully connected layer and the softmax layer in the LSTM1 model are changed, the LSTM1 model to be migrated is obtained, the labeled training set data in S41 is sent to the LSTM1 to be migrated for multiple training processing, the parameters of the fully connected layer and the softmax layer are changed, after 5 rounds of training, the test set data with labels is sent into the network to obtain the model parameters with the highest accuracy, and the obtained model parameters are LSTM2.

[0068] Therefore, the submarine intention prediction method based on LSTM and migration learning has the following beneficial effects:

[0069] (1) The long short-term memory neural network used in the present application can simultaneously learn the features of time series data with long and short time intervals, and is suitable for processing target data with time correlation.

[0070] (2) The migration learning algorithm used in the present application can achieve accurate prediction under the condition of a small sample data set, and solves the difficulty of intention prediction caused by insufficient experimental data of underwater vehicles.

[0071] (3) The Savitzky-Golay filtering algorithm used in the present application can remove noise points in the data and effectively improve the prediction accuracy.

[0072] The technical solutions of the present application will be further described in detail below with reference to the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1A flowchart of an embodiment of a method for predicting the intentions of a submarine target based on LSTM and transfer learning according to the present application;

[0074] Figure 2 A prediction accuracy value graph of an embodiment of a method for predicting the intentions of a submarine target based on LSTM and transfer learning according to the present application; DETAILED DESCRIPTION

[0075] The technical solutions of the present application are further described below through the accompanying drawings and embodiments.

[0076] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those with ordinary skills in the art to which the present application belongs.

[0077] The terms such as "comprise" or "include" and the like used in the present application mean that the elements before the terms encompass the elements listed after the terms, and do not exclude the possibility of also encompassing other elements. The orientations or positional relationships indicated by the terms "in", "on", "upper", "lower", etc. are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. In the present application, unless otherwise explicitly specified and limited, the term "attached" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be a connection or interaction relationship between two elements. For those of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0078] As shown in the accompanying drawings and embodiments, the present application provides a method for predicting the intentions of a submarine based on LSTM and transfer learning, comprising the following steps: Figure 1

[0079] S1: Integrating expert experience knowledge to establish a virtual submarine intention data set;

[0080] S11, according to the expert experience and the real data of the submarine to establish a virtual intention data set Z all ={Z attack ,Z retreat ,Z feint ,Z patrol ,Z scout}, wherein Z attack ={Z attack_1 ,…,Z attack_r} is a data set whose intention is attack, and contains 100 samples, Z​retreat = {Z retreat_1 ,…,Z retreat_u} is the dataset intended for retreat, including 100 samples; Z feint = {Z feint_1 ,…,Z feint_o} is the dataset intended for feint, including 100 samples; Z patrol = {Z patrol_1 ,…,Z patrol_q} is the dataset intended for patrol, including 75 samples, Z scout = {Z scout_1 ,…,Z scout_w} is the dataset intended for scout, including 75 samples.

[0081] Given one ∈ {attack, retreat, feint, patrol, scout}, Z one_ik is the state information of the i-th sample of the intention represented by one at the k-th time,

[0082] Z one_ik = [P one_ik ,V one_ik ] T , where P one_ik = [P x_one_ik ,P y_one_ik ] T : the projection of the position of the i-th attack sample of the intention represented by one at the k-th time on the horizontal x-axis and y-axis.

[0083] V one_ik = [V x_one_ik ,V y_one_ik ] T : the projection of the velocity of the i-th sample of the intention represented by one at the k-th time on the horizontal x-axis and y-axis.

[0084] S12, the specific method for generating data of velocity and position in Z one_ik is shown in the following formula:

[0085]

[0086] where k one_i = diag{k x_one_i ,k y_one_i} is the given control gain of the i-th sample of the intention represented by one, k x_one_i is the control gain in the horizontal x-axis direction, k y_one_i is the control gain in the horizontal y-axis direction, k x_one_i , k y_one_i ∈ [-0.8s -1 , 0.6s-1 ];

[0087] goal one_i =[goal x_one_i ,goal y_one_i ] T ,goal x_one_i ,goal y_one_i respectively represent the projection of the target position of the i-th sample in the intent of one on the horizontal plane x-axis and y-axis, set goal x_one_i ,goal y_one_i ∈[-5m,5m];P one_ik =[P x_one_ik ,P y_one_ik ] T ,P x_one_ik ,P y_one_ik represent the projection of the position of the i-th sample in the intent of one at time k on the horizontal plane x-axis and y-axis;

[0088] V one_ik =[V x_one_ik ,V y_one_ik ] T ,V x_one_ik ,V y_one_ik represent the projection of the velocity of the i-th sample in the intent of one at time k on the horizontal plane x-axis and y-axis; wherein P one_i(k+1) =[P x_one_i(k+1) ,P y_one_i(k+1) ] T ,P x_one_i(k+1) ,P y_one_i(k+1) respectively represent the projection of the position of the i-th sample in the intent of one at time (k+1) on the horizontal plane x-axis and y-axis; Δt is the difference between the two time values, set to 2s; noise=[noise x_one_ik ,noise y_one_ik ] T ,noise x_one_ik ,noise y_one_ik respectively represent the random noise of the i-th sample in the intent of one at time k on the horizontal plane x-axis and y-axis, set noise x_one_ik ,noise y_one_ik ∈[-1.2m,1.5m]。

[0089] S13, label the generated virtual intent data set, the specific processing is as follows: the submarine intent data set Z all ={Z attack ,Z retreat ,Z feint ,Z patrol ,Z scout} in, Z attack= {Z attack_1 ,…,Z attack_r} are manually labeled as 1 (attack) in 100 sample data; Z retreat = {Z retreat_1 ,…,Z retreat_u} are manually labeled as 2 (retreat) in 100 sample data; Z feint = {Z feint_1 ,…,Z feint_o} are manually labeled as 3 (decoy) in 100 sample data; Z patrol = {Z patrol_1 ,…,Z patrol_q} are manually labeled as 4 (patrol) in 75 sample data; Z scout = {Z scout_1 ,…,Z scout_w} are manually labeled as 5 (reconnaissance) in 75 sample data.

[0090] S2: Filtering is performed on the established submarine intention data set;

[0091] S21, Savitzky-Golay filter is used for smoothing filtering on the data in S1, and the process is as follows:

[0092] The Z one_ik state information in the submarine intention data set is constructed to fit a polynomial of order n-1 Z one_ik state data:

[0093]

[0094] Where a0, a1, a2, …, a n-1 are polynomial constant terms of the fitted P x_one_ik , b0, b1, b2, …, b n-1 are polynomial constant terms of the fitted P y_one_ik , c0, c1, c2, …, c n-1 are polynomial constant terms of the fitted V x_one_ik , d0, d1, d2, …, d n-1 are polynomial constant terms of the fitted V y_one_ik , ε k , β k , χ k , δ k are the residuals of the fitting process, k, k 2 , …, k n-1 are the polynomial degrees of the fitting;

[0095] {Z one_i(k-m) , Z one_i(k-m+1) , …, Z one_ik , …, Zone_i(k+m-1) Z one_i(k+m) Using 2m+1 time-state information points as the filtering window, and calculating using the above formula, we obtain 2m+1 equations. The matrix formed by these equations is represented as follows:

[0096]

[0097] The above matrix can be simplified to the following formula:

[0098] Z (2m+1)×4 =X (2m+1)×n ·R n×4 +E (2m+1)×4

[0099] in:

[0100] In the formula: Z, X, R, E, where the subscripts represent the parameter dimensions, and the least squares solution of R is obtained through the least squares method. for:

[0101]

[0102] Z's underwater vehicle data filter value for:

[0103]

[0104] Where H = X·(X) T ·X) -1 ·X T This is the relationship matrix between the filtered data and the original data;

[0105] With a filtering step size of 2m+1, following the above steps, when the center point of the filtering window is at time k, the data filtering is completed. The center point is then moved to time k+2m+1, and the filtering process is repeated until the state information of the i-th sample at all times is filtered, yielding the filtered Z. one_sg_i In this invention, m is set to 9 and n is set to 8.

[0106] Finally, Z all ={Z attack Z retreat Z feint Z patrol Z scout After the above filtering operation is performed on all samples in the dataset, the filtered data Z is obtained. all_sg ={Z attack_sg Z retreat_sg Z feint_sg Z patrol_sg Z scout_sg};

[0107] Zone_sg_ik is the data of the i-th sample in the filtered attack intention at the k-th time point;

[0108] S22: convert the filtered data in S21 into the required input data format of LSTM1, specifically: the filtered data of the i-th sample in the intention Z one_sg_i = [Z one_sg_i1 ,…,Z one_sg_iN ], where N is the number of sampling points in the i-th sample, p is the number of LSTM1 long short-term memory units, and Z one_sg_i is cut into N-p+1 sub-sample data as follows: {Z one_sg_i_1 ,…,Z one_sg_i_j ,…,Z one_sg_i_(N-p+1)} = { [Z one_sg_i1 ,…,Z one_sg_ip ],…,[Z one_sg_ij ,…,Z one_sg_i(p+j-1) ],…,[Z one_sg_i(N-p+1) ,…,Z one_sg_iN ]}.

[0109] Z one_sg_i_j is the data of the j-th sub-sample cut from the i-th sample in the filtered intention, and Z all_sg = {Z attack_sg ,Z retreat_sg ,Z feint_sg ,Z patrol_sg ,Z scout_sg} after cutting all samples to obtain all sub-sample data to be input to LSTM1.

[0110] Z all_sg_fin = {Z attack_sg_fin ,Z retreat_sg_fin ,Z feint_sg_fin ,Z patrol_sg_fin ,Z scout_sg_fin}; where Z attack_sg_fin is the sub-sample data set after cutting the attack intention; Z retreat_sg_fin is the sub-sample data set after cutting the retreat intention; Z feint_sg_fin is the sub-sample data set after cutting the feint intention; Z patrol_sg_fin is the sub-sample data set after cutting the patrol intention; and Z scout_sg_fin is the sub-sample data set after cutting the reconnaissance intention.

[0111] S3: train the LSTM neural network using the filtered virtual data set to obtain a pre-trained model LSTM1; build an LSTM neural network composed of an input layer, a long short-term memory layer, an output layer, and a softmax layer, and the specific steps are as follows:

[0112] S31: The input layer is to input the sub-sample data in the virtual data set of the submarine in S22 to the long short-term memory layer, and the number of neurons in the input layer is set to 5;

[0113] The long short-term memory layer is composed of 5 long short-term memory units, wherein the forget gate, the input gate and the output gate in the long short-term memory layer control the flow of internal information of the network. The calculation formula is as follows:

[0114] f k =σ(W f ·[h k-1 ,x k ]+b f )

[0115] i k =σ(W i ·[h k-1 ,x k ]+b i )

[0116] o k =σ(W o ·[h k-1 ,x k ]+b o )

[0117] c k =f k *c k-1 +i k *tanh(W c ·[h k-1 ,x k ]+b c )

[0118] h k =o k *tanh(c k )

[0119] x k is the state information of Z one_sg_ik in the input network; i k , f k and o k represent the input gate, the forget gate and the output gate at k time in the long short-term memory layer; σ is a sigmoid function; W f , W i , W o , W c , b f , b i , b o , b c are parameters obtained after training of the network; tanh is an activation function; and c krepresents the output of the LSTM neural network candidate gate at time k; c k-1 represents the output of the LSTM neural network candidate gate at time k-1; h k represents the output of the LSTM neural network at time k; h k-1 represents the output of the LSTM neural network at time k-1; * represents the product of corresponding elements of two vectors; the number of neuron nodes in the long short-term memory layer is set to 80.

[0120] The output layer is a fully connected layer, and the number of neurons thereof is set to 5;

[0121] The softmax layer is to output the information of the output layer in the form of probability, and the calculation formula is as follows:

[0122]

[0123] wherein, y i ∈{1,2,3,4,5} corresponds to the label of the intention of retreat, attack, feint, patrol, and reconnaissance in S13, is the sum of the index outputs of all predicted intentions;

[0124] S32: obtaining the pre-trained model LSTM1, specifically,

[0125] S321, dividing the samples after filtering in S2 into a training set and a test set; wherein the training set and the test set account for 80% and 20% of the total sample proportion, respectively;

[0126] S322, using a multi-class cross-entropy loss function Cross Entropy Loss to verify the accuracy of the model, and the calculation method is as follows:

[0127]

[0128] wherein, m represents the total number of submarine sub-sample data, m = 4800; n represents the total number of intention categories of the submarine, n = 5; represents the probability that the i-th real sample is j in the intentions of attack, retreat, feint, patrol, and reconnaissance of the submarine, represents the probability that the i-th predicted sample is j in the intentions of attack, retreat, feint, patrol, and reconnaissance of the submarine.

[0129] S323: putting the labeled data in the training set in S322 into the network built in S31 for training, saving the best parameters with the labeled test set data to obtain the LSTM1 model parameters.

[0130] S4: A small amount of submarine intention real data is collected to obtain an experimental training set, the LSTM1 model parameters in S3 are transferred to an experimental application scenario through transfer learning, and a neural network model LSTM2 is obtained by training the experimental training set.

[0131] In step S4, training the neural network model LSTM2 using the experimental training set includes the following steps:

[0132] S41: The real intention data is divided into a training set and a test set, which respectively account for 80% and 20% of the real intention data, and are labeled with corresponding attack, retreat, feint, patrol and reconnaissance labels. The labeled training set and test set data are subjected to filtering operation in step S2, and then subjected to format operation for adjusting the input of the network to obtain data to be input into the transfer network.

[0133] S42: The parameters of the input layer and the long short-term memory layer in the LSTM1 model are frozen, only the parameters of the fully connected layer and the softmax layer in the LSTM1 model are changed, the LSTM1 model to be transferred is obtained, the labeled training set data in S41 is sent to the LSTM1 to be transferred for multiple training processing, the parameters of the fully connected layer and the softmax layer are changed, after 5 rounds of training, the labeled test set data is sent into the network to obtain the model parameters with the highest accuracy, and the obtained model parameters are LSTM2.

[0134] Figure 2 The submarine prediction model LSTM training accuracy graph of the submarine target intention prediction method based on LSTM and transfer learning is shown in the figure. As shown in the figure, the horizontal axis is the sample number, and the vertical axis is the accuracy. It can be seen that the accuracy of the LSTM network after transfer learning reaches 93%, and a small amount of data reaches a high accuracy. The accuracy of the LSTM network without transfer learning is only 39%. It is shown that the submarine target intention prediction method based on LSTM and transfer learning has high accuracy with a small amount of data.

[0135] Therefore, the submarine intention prediction method based on LSTM and transfer learning can quickly identify transient interference signal characteristics and provide support for radio station communication effectiveness analysis.

[0136] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. An LSTM and transfer learning based method for predicting intentions of an underwater vehicle, the method comprising: The method comprises the following steps: S1: establishing a virtual submarine intention data set; In step S1, the virtual submarine intention data set is generated by the following method: S11: Establishing a virtual intent dataset Z from submarine real data all = {Z attack ,Z retreat ,Z feint ,Z patrol ,Z scout} where Z attack = {Z attack_1 ,…, Z attack_r} is the dataset intended for attack and contains r samples; Z retreat = {Z retreat_1 ,...,Z retreat_u} is the dataset intended for withdrawal and comprises u samples; Z feint = {Z feint_1 ,...,Z feint_o} is a set of data intended to be decoy, including o samples; Z patrol = {Z patr ol_1,...,Z patrol_q} is a data set intended for patrol, including q samples; Z scout = {Z sc out_1...,Z scout_w} is a set of data intended for reconnaissance, comprising w samples; one ∈ {attack, retreat, feint, patrol, scout}, Z one_ik is the state information of the i-th sample of the intention represented by one at the k-th time, Z one_ik = [P one_ik , V one_ik ] T ; where P one_ik = [P x_one_ik ,P y_one_ik ] T where one represents the projection of the i-th attacking sample k position in the horizontal plane x-axis, y-axis at the moment of intention. V one_ik = [V x_one_ik , V y_one_ik ] T where one represents the projection of the velocity of the i-th sample in the intent at time k on the horizontal plane x-axis, y-axis. S12: Z one_ik The data generation method of the middle speed and position is shown in the following formula: wherein k one_i = diag{k x_one_i ,k y_one_i} is the given control gain of the i-th sample in the intent represented by one, k x_one_i is the control gain in the x-axis direction of the horizontal plane, and k y_one_i is the control gain in the y-axis direction of the horizontal plane; goal one_i = [goal x_one_i , goal y_one_i ] T , and goal x_one_i and goal y_one_i are the projections of the target position of the i-th sample in the intent represented by one in the x-axis and y-axis directions of the horizontal plane, respectively; P one_ik = [P x_one_ik , P y_one_ik ] T , and P x_one_ik and P y_one_ik are the projections of the position of the i-th sample in the intent represented by one at time k in the x-axis and y-axis directions of the horizontal plane; V one_ik = [V x_one_ik , V y_one_ik ] T , and V x_one_ik and V y_one_ik are the projections of the speed of the i-th sample in the intent represented by one at time k in the x-axis and y-axis directions of the horizontal plane; wherein P one_i(k+1) = [P x_one_i(k+1) , P y_one_i(k+1) ] T , P x_one_i(k+1) and P y_one_i(k+1) are the projections of the position of the i-th sample in the intent represented by one at time (k+1) in the x-axis and y-axis directions of the horizontal plane, respectively; Δt is the difference between the two time points; noise = [noise x_one_ik , noise y_one_ik ] T , and noise x_one_ik and noise y_one_ik are the random noises of the i-th sample in the intent represented by one at time k in the x-axis and y-axis directions of the horizontal plane, respectively, with a value range of [-1.2 m, 1.5 m]. S13: label the generated virtual intention dataset, and the specific processing is as follows:Z attack ={Z attack_1 ,...,Z attack_r}r sample data in which the artificial label is 1 attack;Z retreat ={Z retreat_1 ,...,Z retreat_u}u sample data in which the artificial label is 2 retreat;Z feint ={Z feint_1 ,...,Z feint_o}o sample data in which the artificial label is 3 feint;Z patrol ={Z patrol_1 ,...,Z patrol_q}q sample data in which the artificial label is 4 patrol;Z scout ={Z scout_1 ,...,Z scout_w}w sample data in which the artificial label is 5 reconnaissance; S2: filtering the established submarine intention data set; S3: training an LSTM neural network using the filtered virtual data set to obtain a pre-trained model LSTM1; S4: collecting real submarine intention data to obtain an experimental training set, migrating the LSTM1 model parameters to an experimental application scenario, and training a neural network model LSTM2 using the experimental training set.

2. The method of claim 1, wherein the method is based on LSTM and transfer learning. The filtering of the target intention data set established in S1 in step S2 is as follows: S21: using a Savitzky-Golay filter to perform smoothing filtering on the data in S1, and the process is as follows: Z one_ik state information, construct a n-1 order polynomial fit one_ik state data: where a0, a1, a2,..., a n-1 are polynomial constant terms fitting P x_one_ik n-1 are polynomial constant terms fitting P y_one_ik n-1 are polynomial constant terms fitting V x_one_ik ​​​ d0, d1, d2,..., d n-1 are polynomial constants of the fit V y_one_ik , ε k , β k , χ k , δ k are residuals of the fitting process, k, k 2 ..., k n-1 is the polynomial degree of the fit; Select {Z one_i(k-m) ,Z one_i(k-m+1) ,...,Z one_ik ,...,Z one_i(k+m-1) ,Z one_i(k+m)} 2m+1 time state information as the filter window, use the formula to calculate 2m+1 formula, the formula is represented as a matrix: The above matrix is simplified as follows: Z (2m+1)×4 = X (2m+1)×n • R n×4 + E (2m+1)×4 wherein: where: Z, X, R, E subscripted by the parameter dimension, the least square solution of R is obtained by least square method is: Z's submersible data filtered value is: where H = X · (X T ·X) -1 ·X T is the relationship matrix between filtered data and original data; The filtering step is 2m+1. After the data filtering of the center point at time k is completed, the center point is moved to time k+2m+1 and the above filtering process is repeated until the filtering of the state information of the i-th sample at all times is completed, and the filtered Z one_sg_i ; Finally all samples in Z all = {Z attack ,Z retreat ,Z feint ,Z patrol ,Z scout} are filtered by the above filtering operation to obtain filtered data Z all_sg = {Z attack_sg ,Z retreat_sg ,Z feint_sg ,Z patrol_sg ,Z scout_sg}. Z one_sg_ik is the data of the i-th sample of the intention represented by one at the k-th time after filtering; S22: convert the filtered data in S21 into the required input data format in LSTM1, specifically: the i-th sample Z in the filtered intent represented by one one_sg_i = [Z one_sg_i1 ,...,Z one_sg_iN ], where N is the number of sampling points in the i-th sample, p is the number of long short-term memory units in LSTM1, and Z one_sg_i is cut into N-p+1 sub-sample data as follows: {Z one_sg_i_1 ,...,Z one_sg_i_j ,...,Z one_sg_i_(N-p+1)}={[Z one_sg_i1 ,...,Z one_sg_ip ],...,[Z one_sg_ij ,...,Z one_sg_i(p+j-1) ],...,[Z one_sg_i(N-p+1) ,...,Z one_sg_iN ]}, Z one_sg_i_j For the data cut into the jth sub-sample in the ith sample of the intent represented by the filtered one, Z all_sg = {Z attack_sg ,Z retreat_sg ,Z feint_sg ,Z patrol_sg ,Z scout_sg} all samples are cut to obtain all sub-sample data to be input to LSTM1: Z all_sg_fin = {Z attack_sg_fin , Z retreat_sg_fin , Z feint_sg_fin , Z patrol_sg_fin , Z scout_sg_fin}; where Z attack_sg_fin is the subset of data for attack intent after being cut; Z retreat_sg_fin is the subset of data for retreat intent after being cut; Z feint_sg_fin is the subset of data for feint intent after being cut; Z patrol_sg_fin is the subset of data for patrol intent after being cut; Z scout_sg_fin is the subset of data for reconnaissance intent after being cut.

3. The method of claim 2, wherein the method is based on LSTM and transfer learning. In step S3, an LSTM neural network composed of an input layer, a long short-term memory layer, an output layer, and a softmax layer is built, and the specific steps are as follows: S31: the input layer is to input the sub-sample data in the virtual submarine data set in S22 to the long short-term memory layer; The long short-term memory layer is composed of p long short-term memory units, and the calculation formula is as follows: f k = σ(W f · [h k-1 , x k ]+ b f ) i k = σ(W i · [h k-1 , x k ]+ b i ) o k = σ(W o · [h k-1 , x k ]+b o ) c k = f k * c k-1 + i k * tanh(W c * [h k-1 , x k ] + b c ) h k = o k * tanh(c k ) x k one_sg_ik is the state information of Z k in the input network; i k , f k and o f represent the input gate, the forget gate and the output gate of the LSTM layer at time k, respectively; σ is the sigmoid function; W i , W o , W c , W f , b i , b o , b c are the parameters obtained after training of the network; tanh is the activation function; c k represents the output of the LSTM neural network candidate gate at time k; c k-1 represents the output of the LSTM neural network candidate gate at time k-1; h k represents the output of the LSTM neural network at time k; h k-1 represents the output of the LSTM neural network at time k-1; * represents the product of the corresponding elements of two vectors;​ The output layer is a fully connected layer; The softmax layer outputs the information of the output layer in the form of probability, and the calculation formula is as follows: wherein, y i ∈ {1,2,3,4,5} corresponds to the labels in S13 for the intents retreat, attack, feint, patrol, reconnaissance, sum over all predicted intent indices. S32: obtaining the pre-trained model LSTM1 is as follows, S321: dividing the samples filtered in S2 into a training set and a test set, wherein the training set and the test set account for 80% and 20% of the total samples, respectively; S322: using a multi-class cross-entropy loss function Cross Entropy Loss to verify the accuracy of the model, and the calculation method is as follows: Wherein, m represents the total number of submarine subsample data, n represents the total number of submarine intention categories, Pi,j represents the probability that the i-th real sample is j in the submarine intention of attack, retreat, feint, patrol and reconnaissance, Pi,j represents the probability that the i-th predicted sample is j in the submarine intention of attack, retreat, feint, patrol and reconnaissance. S323: putting the labeled data in the training set in S322 into the network built in S31 for training, saving the best parameters with the labeled test set data, and obtaining the LSTM1 model parameters.

4. The method of claim 3, wherein the method is based on LSTM and transfer learning. In step S4, training a neural network model LSTM2 using an experimental training set comprises the following steps: S41: dividing the real intention data into a training set and a test set, which account for 80% and 20% of the real intention data, respectively, and labeling corresponding attack, retreat, feint, patrol, and reconnaissance, and then performing filtering operation in step S2 and format operation in S322 to obtain data to be input into the migration network; S42: freezing the parameters of the input layer and the long short-term memory layer in the LSTM1 model, only changing the parameters of the fully connected layer and the softmax layer in the LSTM1 model, obtaining the LSTM1 model to be migrated, sending the labeled training set data in S41 to the LSTM1 to be migrated for multiple training, changing the parameters of the fully connected layer and the softmax, and after 5 rounds of training, sending the labeled test set data into the network to obtain the model parameters with the highest accuracy, and obtaining the model parameters LSTM2.

Images