CN114296067A - Recognition method of low, slow and small target based on LSTM model of pulse Doppler radar - Google Patents

Abstract

The invention provides a method for identifying a low-slow small target of a pulse Doppler radar based on an LSTM model, which comprises the following steps: receiving a set of low-slow subclass target tracks of the collected pulse Doppler radar, and performing splitting and normalization pretreatment on track set information; initializing a data set based on initialization parameters and the number of current training single targets to execute selection of LSTM forward propagation, and calculating a loss function value corresponding to a current coefficient; if the loss function value is larger than a preset threshold value, updating the input gate coefficient, the output gate coefficient and the forgetting gate state coefficient; if the loss function value is smaller than the threshold value, performing next training single target number training based on the current neural network weight coefficient; after finishing the training of all the training single target numbers in the current period, comparing the loss function value in the current period with the preset threshold value of the loss function stopping iteration; and verifying the identification accuracy in the verification set data based on the final state neural network parameters and outputting.

Description

Recognition method of low, slow and small target based on LSTM model of pulse Doppler radar

technical field

The invention belongs to the technical field of pulse Doppler radar target recognition, in particular to a pulse Doppler radar low-slow and small target recognition method based on an LSTM model.

Background technique

Radar target recognition technology refers to the use of radar to detect the target and analyze the acquired echo information to determine the target attribute and type. That is, according to the features in the echo, the target type is identified. Its essence is the electromagnetic inverse scattering problem that requires the inversion of the target characteristics when the incident and scattered waves are known. The types of targets that radars need to identify cover all targets on the ground, sea, air, and sky, and even terrain, weather, interference and radiation sources. The degree of target recognition also has multi-level definitions. In addition to classification, identification, identification, and identification, it can also be extended to target friend or foe identification, threat assessment, etc. The use of radar to classify and identify targets is very important for military and civilian use. value.

Artificial intelligence is a science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. Radar target recognition technology is an important application of artificial intelligence in the field of equipment. With the development of artificial intelligence technology, radar recognition is also making continuous progress, from pattern recognition, machine learning to neural networks and migration learning, which have developed rapidly in recent years, in radar recognition. There are many research results. Although radar target recognition has a wide range of applications and has been successfully applied at some levels, radar target recognition technology has not yet formed a complete theoretical system, and some radar target recognition systems still have certain limitations in function. It is the diversity of target types and radar regimes and the extreme complexity of the environment.

Traditional radar recognition technology often uses statistical pattern recognition theory. Pattern recognition is a discipline that mainly uses tools such as statistics, probability theory, computational geometry, machine learning, signal processing, and algorithm design to reason from perceptible data. Its central task is to find out the essential attributes of certain things . For radar target recognition, firstly, according to the motion, echo and other information of the target tracked by the radar, the stable and iconic features of the target are extracted, which are called "recognition feature templates", and then the patterns to be recognized are divided into their respective patterns. in the class. Recognition/classification of a given pattern will face two types of tasks: supervised classification and unsupervised classification, where supervised classification classifies patterns into existing classes, and unsupervised classification classifies patterns into unknown classes.

Feature extraction is an important part of traditional radar recognition technology. Radar recognition features are strongly dependent on the prior knowledge and professional skills of employees. The design of radar target recognition algorithms requires a deep research background on target characteristics and feature extraction. The traditional radar target recognition is usually to receive the fixed information of the radar sensor and perform digital signal processing to extract the characteristics of the target to be identified, use the existing feature template to classify the extracted features, and identify the target according to the degree of membership. The main problem of traditional target recognition is that it works according to the preset recognition mode, and it does not have the ability to automatically change the recognition mode with the changes of the target and the environment. When the environment changes, it is difficult to obtain the ideal only by passive feature extraction and classification. effect, insufficient adaptability to the target and the environment.

Facing the challenges of increasingly complex battlefield environment and dense clutter and multi-target background, in order to meet the current and especially future operational needs, the recognition technology must be further innovated and developed to continuously improve the recognition mode and recognition performance in order to adapt to the increasingly complex combat environment. .

SUMMARY OF THE INVENTION

In order to solve the above-mentioned technical problems, the present invention proposes a method for identifying low, slow and small targets of pulsed Doppler radar based on the LSTM model, and the method includes the following steps:

Step 1: Receive a collection of pulse Doppler radar low and slow target tracks that have been collected, split and normalize the track set information for preprocessing; divide the normalized tracks into training tracks according to a predetermined proportion. set and validation set;

Step 2: Initialize the number of input nodes, the number of neural network layers, the number of training cycles, the number of targets in a single training session, the weight adjustment ratio for a single iteration, the threshold of the stop iteration loss function, and the type of output target; Output gate, forget gate state coefficient, cell coefficient, bias value; initialized hidden layer cell state value, cell hidden layer value;

Step 3, based on the initialization parameters in step 2 and the data set of the current training single target number, perform forward propagation of the selected LSTM, and calculate the loss function value corresponding to the current coefficient;

Step 4 compares the loss function value described in step 3 with the preset threshold of the stop iteration loss function,

If the loss function value is greater than the preset threshold, update the input gate coefficient, output gate coefficient and forget gate state coefficient; if the loss function value is less than the threshold, perform the next training single target number based on the current neural network weight coefficient number training;

Step 5, after completing the training of all training single target numbers in the current cycle through steps 3 and 4, compare the current cycle loss function value with the preset threshold of the stop iteration loss function;

If the current cycle loss function value is greater than the preset threshold, perform the next cycle training;

If the current period loss function value is less than the preset threshold, stop training and output the parameters at the current moment as the final neural network parameters;

Step 6, based on the final state neural network parameters in step 5, verify the recognition accuracy in the validation set data and output.

Further, step 1 includes the following sub-steps:

Step 1.1, set the collection of pulse Doppler radar low-slow and small-class target tracks that have been collected and expressed as

n=1,...,N,

l _n =1,...,L _n

表示航迹集合内第n条航迹中第l_n个点迹距离，

表示航迹集合内第n条航迹中第l_n个点迹方位角，

表示航迹集合内第n条航迹中第l_n个点迹俯仰角，

表示航迹集合内第n条航迹中第l_n个点迹雷达散射截面积RCS，N表示航迹个数，L_n表示第n条航迹内点迹个数；in

represents the distance of the _nth point track in the nth track in the track set,

represents the azimuth angle of the _nth point track in the nth track in the track set,

represents the pitch angle of the _nth point track in the nth track in the track set,

Represents the radar scattering cross-sectional area RCS of the _nth point track in the nth track in the track set, N represents the number of tracks, and _Ln represents the number of point tracks in the nth track;

Step 1.2, for the set of target tracks, add track labels according to the prior information of track collection types;

Step 1.3, the target track set normalizes the track information according to the following formula:

n=1,...,N,

l _n =1,...,L _n

where ∑ represents the summation operation;

Step 1.4, the normalized track is divided into a training set T _n and a validation set V _n according to a predetermined ratio.

Further, where N=12338; in the track label, the UAV track is marked as 1, and the non-UAV track is marked as 0; the training set proportion is 70%, and the verification set proportion is 30%.

Further, in step 2, the number of input nodes is 256, the initial value of the training cycle is 1000 times, the number of targets in a single training session is 500, the weight ratio of a single iteration is ρ=1%, and the threshold of the stop iteration loss function Tr=10 ^-6 .

Further, step 3 includes sub-steps:

Step 3.1, according to the normalized training set obtained in step 1 and the minibatch size of the number of targets in a single training session in step 2, the data set is divided into

N _b =N/minibatch batches, wherein batch is used as the basic operation unit of the following steps;

Step 3.2, take 1 batch as the operation unit to carry out the following operations: calculate the input gate, forget gate, and output gate output according to the initialization coefficient of step 2 in combination with the following formula;

where σ(x) represents the sigmoid activation function:

Step 3.3, update the cell state x and the hidden layer value h according to the result of step 3.2 combined with the following formula:

x=x ₀ □F _g +I _g □G

h=O _g tanh(x)

where □ represents the element-wise dot product, and tanh(x) represents the activation function:

Step 3.4, calculate the classification output value corresponding to the current weight coefficient according to the result of step 3.3 in combination with the following formula;

表示第l_n条航迹分别属于种类“0”与“1”的概率，当前权系数对应分类输出值取两概率较大值对应种类

in

Indicates the probability that the 1 _nth track belongs to the category "0" and "1" respectively, and the current weight coefficient corresponds to the classification output value to take the two larger probability values corresponding to the category

Step 3.5, calculate the loss function L _s corresponding to the coefficient according to the classification result of step 3.4;

The meaning of each variable is: input coefficient W _I at the initial time of initialization, input hidden layer coefficient W _h , input bias value B, input gate state coefficient W _Ig , cell coefficient W _Ic , bias value B _I , output gate state Coefficient W _Og , cell coefficient W _Oc , bias value _BO , forget gate state coefficient W _Fg , cell coefficient W _Fc , bias value _BF ; cell state value x ₀ , cell hidden layer value h ₀ at the initial time of initialization, The hidden layer output coefficient W _O , the bias value B _O ; the gate coefficients, bias values, cell states, and hidden layer values are all initialized to random values in the interval (0,1) at the beginning.

Further, step 4 includes updating each coefficient; specifically includes the following sub-steps:

输出系数

更新：Step 4.1, enter the coefficients

output coefficient

renew:

Among them, I represents an all-one vector;

Step 4.2, door state coefficient update:

表示输入门状态系数更新值，

表示遗忘门状态系数更新值，

表示输出门状态系数更新值；in,

represents the update value of the input gate state coefficient,

represents the update value of the forget gate state coefficient,

Indicates the update value of the output gate state coefficient;

Step 4.3, cell coefficient update:

表示输入门细胞系数更新值，

表示遗忘门细胞系数更新值，

表示输出门细胞系数更新值；in,

represents the updated value of the input gate cell coefficient,

represents the update value of the forget gate cell coefficient,

Indicates the updated value of the output gate cell coefficient;

Step 4.4, update the input hidden layer coefficients:

Among them, □ represents the element-wise dot product, and tanh(x) represents the activation function.

Further, the parameter set at the current moment is used as the final neural network parameter; the parameter set is:

Further, step 6 includes the following sub-steps:

Step 6.1, classify the validation set data based on the final state neural network parameter W _opt , input gate output, forget gate output, output gate output, cell state, hidden layer value and the corresponding classification output value of the current weight coefficient and output the classification result

Step 6.2, compare the output classification result with the track label, and count the recognition rate as follows:

The method of the present invention is used to learn and train the track features of the radar low-slow and small-class targets based on the long-short-term memory network LSTM model, and output corresponding network parameters, and based on the network parameters, a classification function is established to achieve real-time recognition of the radar system's low-slow and small class targets. with classification purposes. In order to achieve the above technical purpose, the present invention adopts the following technical solutions to achieve.

Description of drawings

Fig. 1 is the low-slow and small target identification technology realization general flow chart of the present invention;

Figure 2 is an iterative diagram of LSTM parameter transfer;

Fig. 3 is a training process diagram;

Figure 4 is a real-time recognition effect diagram.

Detailed ways

The invention discloses a pulse Doppler radar low-slow and small target identification method based on an LSTM model, which is suitable for pulse-Doppler radar. Based on the long-short-term memory network LSTM model, the low-slow and small-class target track characteristics of the radar are learned and trained. Corresponding network parameters, and establishing a classification function based on the network parameters, to achieve the purpose of real-time identification and classification of low-slow and small-class targets in radar systems.

A low-slow and small target recognition method based on LSTM model for pulse Doppler radar, including the following steps

1) Data preprocessing: The pulse Doppler radar low and slow target track set has been collected, and the track set information is split and normalized for preprocessing; the normalized track is divided into a certain proportion. are the training set and the validation set;

2) Initialization of LSTM model training parameters: initialize the number of input nodes, the number of neural network layers, the number of training cycles, the number of targets per training (minibatch), the weight adjustment ratio for a single iteration, the threshold of the stop iteration loss function, and the type of output target ; Initialize the input gate, output gate, forget gate state coefficient, cell coefficient, bias value at the initial time; initialize the hidden layer cell state value and cell hidden layer value at the initial time;

3) Based on 2) the initialization parameters and the current minibatch data set carry out LSTM forward propagation and calculate the loss function value corresponding to the current coefficient;

4) Based on 3) compare the loss function value with the stop iteration loss function threshold, if the loss function value is greater than the threshold, update each gate coefficient; if the loss function value is less than the threshold, perform the next minibatch training based on the current neural network weight coefficient;

5) After completing all minibatch training in the current cycle based on 3) and 4), compare the current cycle loss function value with the stop iteration loss function threshold, if the loss function value is greater than the threshold, proceed to the next cycle of training; if the loss function is less than the threshold, Stop training and output the current moment parameters as the final neural network parameters;

6) Based on 5) final state neural network parameters, verify the recognition accuracy in the validation set data and output.

The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

In view of the problems existing in the above-mentioned prior art, the purpose of the present invention is to propose a low-slow and small target identification method based on the LSTM model of the pulse Doppler radar. This method is based on the long short-term memory network LSTM model to learn and train the corresponding network parameters for the track features of the radar low-slow and small-class targets, and establish a classification function based on the network parameters to achieve real-time recognition and classification of low-slow and small-class targets in the radar system. Purpose. In order to achieve the above technical purpose, the present invention adopts the following technical solutions to achieve.

A low-slow and small target recognition method for pulsed Doppler radar based on LSTM model, comprising the following steps:

Step 1 Data preprocessing: The pulse Doppler radar low and slow target track set has been collected, and the track set information is split and normalized for preprocessing; the normalized track is divided into a certain proportion. are the training set and the validation set;

Step 2 LSTM model training parameter initialization: initialize the number of input nodes, the number of neural network layers, the number of training cycles, the number of targets per training (minibatch), the weight adjustment ratio for a single iteration, the threshold of the stop iteration loss function, and the type of output target ; Initialize the input gate, output gate, forget gate state coefficient, cell coefficient, bias value at the initial time; initialize the hidden layer cell state value and cell hidden layer value at the initial time;

Step 3 selects LSTM forward propagation based on the initialization parameters of step 2 and the current minibatch data set and calculates the loss function value corresponding to the current coefficient;

Step 4 compares the loss function value of step 3 with the stop iteration loss function threshold. If the loss function value is greater than the threshold, update each gate coefficient; if the loss function value is less than the threshold, perform the next minibatch training based on the current neural network weight coefficient;

Step 5, Step 3 and Step 4 After completing all minibatch training in the current cycle, compare the loss function value of the current cycle with the stop iteration loss function threshold. If the loss function value is greater than the threshold, perform the next cycle of training; if the loss function value is less than the threshold, Stop training and output the current moment parameters as the final neural network parameters;

Step 6: Based on the final state neural network parameters of step 5, the recognition accuracy is verified and output in the validation set data.

Referring to FIG. 1 , it is a general flow chart for realizing a method for recognizing low, slow and small targets of pulse Doppler radar based on the LSTM model of the present invention. The LSTM model-based pulse Doppler radar low-slow and small target recognition method includes the following steps:

Step 1 Data preprocessing: The pulse Doppler radar low and slow target track set has been collected, and the track set information is labeled and normalized for preprocessing; A ratio is divided into training set and validation set;

1a) A collection of low-slow and small-class target tracks collected by pulse Doppler radar

表示航迹集合内第n条航迹中第l_n个点迹距离，

表示航迹集合内第n条航迹中第l_n个点迹方位角，

表示航迹集合内第n条航迹中第l_n个点迹俯仰角，

表示航迹集合内第n条航迹中第l_n个点迹雷达散射截面积(RCS)，N表示航迹个数，L_n表示第n条航迹内点迹个数；in

represents the distance of the _nth point track in the nth track in the track set,

represents the azimuth angle of the _nth point track in the nth track in the track set,

represents the pitch angle of the _nth point track in the nth track in the track set,

Represents the radar cross-sectional area (RCS) of the 1 _nth point track in the nth track in the track set, N represents the number of tracks, and _Ln represents the number of point tracks in the nth track;

In this example, N=12338 is selected but not limited to.

1b) Based on the track set in step 1a), the track label is added according to the prior information of the track collection type;

In this example, the track label is selected but not limited to the UAV track (marked as 1) and the non-UAV track (marked as 0);

1c) According to step 1b), the track set is combined with the following formula to normalize the track information:

where ∑ represents the summation operation;

1d) according to step 1c) the normalized track is divided into training set T _n and verification set V _n according to a certain proportion;

In this example, the proportion of training set is selected but not limited to 70%, and the proportion of verification set is selected to be 30%.

Step 2: Initialize the training parameters of the LSTM model: initialize the number of neural network layers, the number of training cycles, the number of targets in a single training (minibatch), the weight adjustment ratio ρ for a single iteration, the threshold Tr of the stop iteration loss function, and the type of output target; initialization start time Input coefficient, hidden layer coefficient, bias value, input gate, output gate, forget gate state coefficient, cell coefficient, bias value; initialized cell state value, cell hidden layer value;

In this example, the number of input nodes is selected but not limited to 256 neural network layers, the initial value of the training cycle is 1000 times, the number of minibatches for a single training target is 500, the weight ratio of a single iteration is ρ=1%, and the threshold of the loss function to stop iteration Tr=10 ^-6 , output target type 2, in which 0 means non-UAV, 1 means UAV; input coefficient W _I at the initial time of initialization, input hidden layer coefficient W _h , input bias value B, input gate state Coefficient W _Ig , cell coefficient W _Ic , bias value B _I , output gate state coefficient W _Og , cell coefficient W _Oc , bias value _BO , forget gate state coefficient W _Fg , cell coefficient W _Fc , bias value B _F ; Initialize the cell state value x ₀ , the cell hidden layer value h ₀ , the hidden layer output coefficient W _O , the bias value B _O ; the gate coefficient, bias value, cell state, and hidden layer value are all initialized at the initial time is a random value in the interval (0,1).

Step 3 selects LSTM forward propagation based on the initialization parameters of step 2 and the current minibatch data set and calculates the loss function value corresponding to the current coefficient;

3a) according to the normalized training set obtained in step 1 and the size of step 2 minibatch, the data set is divided into N _b =N/minibatch batches, wherein batch is the following basic operation unit;

3b) take 1 batch in step 3a) as the operation unit and carry out the following operations: calculate input gate, forget gate, output gate output according to step 2 initialization coefficient in conjunction with following formula;

where σ(x) represents the sigmoid activation function:

3c) Update the cell state x and the hidden layer value h according to the result of step 3b) combined with the following formula:

where □ represents the element-wise dot product, and tanh(x) represents the activation function:

3d) According to the result of step 3c), the corresponding classification output value of the current weight coefficient is calculated in combination with the following formula;

表示第l_n条航迹分别属于种类“0”与“1”的概率，当前权系数对应分类输出值取两概率较大值对应种类

in

Indicates the probability that the 1 _nth track belongs to the category "0" and "1" respectively, and the current weight coefficient corresponds to the classification output value to take the two larger probability values corresponding to the category

3e) Calculate the loss function L _s according to the classification result of 3d);

The loss function includes cross entropy, Focal loss, etc. In the example of the present invention, since the proportion of samples is seriously unbalanced, Focal loss is selected to calculate the loss function:

Among them, log( ) is the logarithmic operation, α is the balance factor, γ is the ratio of adjusting the reduction of the simple sample coefficient, in this example α=0.25, γ=2;

Step 4 compares the loss function with the stop iteration loss function threshold based on step 3. If the loss function is greater than the threshold, update the coefficients of each gate; if the loss function is less than the threshold, perform the next minibatch training based on the current neural network coefficients;

4a) Update each coefficient in combination with the following formula;

输出系数

更新：input factor

output coefficient

renew:

where I represents an all-ones vector. Door state coefficient update:

表示输入门状态系数更新值，

表示遗忘门状态系数更新值，

表示输出门状态系数更新值；细胞系数更新：in

represents the update value of the input gate state coefficient,

represents the update value of the forget gate state coefficient,

Indicates the update value of the output gate state coefficient; the cell coefficient update:

表示输入门细胞系数更新值，

表示遗忘门细胞系数更新值，

表示输出门细胞系数更新值；输入隐层系数更新：in

represents the updated value of the input gate cell coefficient,

represents the update value of the forget gate cell coefficient,

Represents the updated value of the output gate cell coefficient; the input hidden layer coefficient is updated:

Step 5, Step 3 and Step 4 After completing all minibatch training in the current cycle, compare the current cycle loss function value with the stop iteration loss function threshold, if the loss function value is greater than the threshold, proceed to the next cycle of training; if the loss function value is less than the threshold value , stop training and output the parameter set at the current moment as the final neural network parameter;

The set of parameters described is:

Step 6: Based on the final state neural network parameters of step 5, the recognition accuracy is verified and output in the validation set data.

6a) Based on the final state neural network parameter W _opt in step 5 and formulas 3), 5), and 7), classify the validation set data and output the classification results

6b) Compare the output classification result with the track label in step 1b), and use the following formula to count the recognition rate:

The effect of the present invention is further illustrated by the following simulation comparison test:

1. Experimental scene:

In the ground clutter environment, radar equipment is used to collect UAV and clutter track data, a total of N=12338 tracks, and the method of the present invention is used to perform network training based on 0.7×N training set tracks to obtain the optimal coefficient W _opt , Based on this coefficient, the track recognition accuracy of the remaining validation set is verified.

2. Analysis of experimental results:

Table 1 shows the recognition accuracy rates with different neural network layers and different minibatches; in the parameters of this example, the number of neural network layers is 100 layers, and the recognition accuracy rate is the highest when the minibatch size is 400, which can reach 93.25%.

Table 1

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention and not to limit them. Although the embodiments of the present invention have been described in detail with reference to the above preferred embodiments, those of ordinary skill in the art should Modifications or equivalent substitutions to the technical solutions of the embodiments of the present invention should not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A method for identifying a low-slow small target of a pulse Doppler radar based on an LSTM model is characterized by comprising the following steps:

step 1, receiving a set of low-speed subclass target tracks of an acquired pulse Doppler radar, and performing splitting and normalization pretreatment on track set information; dividing the normalized flight path into a training set and a verification set according to a preset proportion;

step 2, initializing the number of input nodes, the number of neural network layers, the number of training cycles, the number of single training targets, the single iteration weight adjustment proportion, the threshold value of the stop iteration loss function and the output target type; initializing a starting moment input gate, an output gate, a forgetting gate state coefficient, a cell coefficient and a bias value; initializing a hidden layer cell state value and a hidden layer cell value at the starting moment;

step 3, selecting LSTM forward propagation based on the initialization parameters in the step 2 and the data set of the current training single target number, and calculating a loss function value corresponding to a current coefficient;

step 4 compares the loss function value in step 3 with a preset threshold value for stopping the iterative loss function,

if the loss function value is larger than a preset threshold value, updating the input gate coefficient, the output gate coefficient and the forgetting gate state coefficient; if the loss function value is smaller than the threshold value, performing next training single target number training based on the current neural network weight coefficient;

step 5, after finishing the training of the number of the single targets in the current period through the steps 3 and 4, comparing the loss function value of the current period with a preset threshold value of the iteration-stopping loss function;

if the current period loss function value is larger than the preset threshold value, executing the next period training;

if the loss function value of the current period is smaller than the preset threshold value, stopping training and outputting all parameters at the current moment as final neural network parameters;

and 6, verifying the data verification identification accuracy in the verification set based on the neural network parameters in the final state in the step 5 and outputting.

2. An identification method as claimed in claim 1, characterized in that step 1 comprises the following sub-steps:

step 1.1, setting a low-slow subclass target track set of the acquired pulse Doppler radar to be expressed as

n＝1,...,N,

l_n＝1,...,L_n

Wherein

Representing the l in the n track in the track set_nThe distance between the point traces is determined,

representing the l in the n track in the track set_nTrack of points orientationThe angle of the corner is such that,

representing the l in the n track in the track set_nThe pitch angle of the point trace is,

representing the l in the n track in the track set_nThe scattering sectional area RCS of the point trace radar, N represents the number of tracks, L_nRepresenting the number of trace points in the nth track;

step 1.2, adding a track label according to track acquisition type prior information aiming at the target track set;

step 1.3, the target track set normalizes the track information according to the following formula:

n＝1,...,N,

l_n＝1,...,L_n

where Σ · denotes a summation operation;

step 1.4, dividing the normalized flight path into training sets T according to a preset proportion_nAnd verification set V_n。

3. The identification method of claim 1, wherein N-12338; marking the unmanned aerial vehicle track as 1 and marking the non-unmanned aerial vehicle track as 0 in the track label; the training set proportion is 70%, and the validation set proportion is 30%.

4. The identification method of claim 1, wherein in step 2, the number of input nodes is 256, the initial value of the training period is 1000 times, the number of single training targets is 500, the weight ratio p of single iteration is 1%, and the threshold value Tr of the stop iteration loss function is 10 ═ 10^-6。

5. An identification method as claimed in claim 1, characterized in that step 3 comprises the sub-steps of:

step 3.1, dividing the data set into N according to the normalized training set obtained in step 1 and the size of the training single target number minipatch in step 2_bN/minipatch lots, wherein each lot is used as a basic operation unit in the following steps;

and 3.2, taking 1 batch as an operation unit to perform the following operations: calculating the output of the input gate, the forgetting gate and the output gate according to the initialization coefficient in the step 2 and the following formula;

where σ (x) denotes the sigmoid activation function:

step 3.3, updating the cell state x and the hidden layer value h according to the result of the step 3.2 by combining the following formula:

wherein

Represents the element dot product, tanh (x) represents the activation function:

step 3.4, calculating a classification output value corresponding to the current weight coefficient according to the result of the step 3.3 by combining the following formula;

wherein

Denotes the l_nThe flight paths respectively belong to the probabilities of the types of 0 and 1, and the classification output value corresponding to the current weight coefficient takes the larger value of the two probabilities to correspond to the type

Step 3.5, calculating coefficient corresponding loss function L according to the classification result of the step 3.4_s；

Wherein the meaning of each variable is: initial start time input coefficient W_IInput hidden layer coefficient W_hInputting the offset value B and the gate state coefficient W_IgCell coefficient W_IcOffset value B_IOutput gate state coefficient W_OgCell coefficient W_OcOffset value B_OForgetting the door state coefficient W_FgCell coefficient W_FcOffset value B_F(ii) a Initializing the initial cell state value x₀Value of cell envelope h₀Hidden layer output coefficient W_OOffset value B_O(ii) a At the starting time, the gate coefficients, the offset values, the cell states and the hidden layer values are initialized to random values in the interval (0, 1).

6. An identification method as claimed in claim 1, characterized in that step 4 comprises updating the coefficients; the method specifically comprises the following substeps:

step 4.1, input coefficients

Output coefficient

Updating:

wherein I represents a full 1 vector;

step 4.2, updating the door state coefficient:

wherein,

indicating that the gate state coefficient update value is entered,

indicating an update value of the forgetting gate state coefficient,

representing the output gate state coefficient update value;

step 4.3, updating cell coefficients:

wherein,

indicating the input of the gated cell coefficient update values,

indicating an update value of the forgetting gate cell coefficient,

representing the output gate cell coefficient update value;

step 4.4, input hidden layer coefficient updating:

wherein,

representing element dot product, and tanh (x) representing activation function.

7. The identification method according to claim 6, wherein the current time parameter set is used as a final neural network parameter; the parameter sets are:

8. an identification method as claimed in claim 1, characterized in that step 6 comprises the following sub-steps:

step 6.1, neural network parameters W based on final state_optClassifying the verification set data and outputting classification results corresponding to the classification output values of the input gate output, the forgetting gate output, the output gate output, the cell state, the hidden layer value and the current weight coefficient

Step 6.2, comparing the output classification result with the track label, and counting the recognition rate according to the following formula:

Images