A Complex Environment Echo Cognition Method Based on Invariant Scattering Convolutional Network and LSTM

Abstract

A complex environment echo recognition method based on an invariant scattering convolutional network and LSTM, comprising the following steps: S1, establishing geometric models and electromagnetic parameter models of several typical environments; S2, based on the geometric models and electromagnetic parameter models in S1, simulating echo signals under a pulsed Doppler radar system; S3, establishing an invariant scattering convolutional network to extract characteristic data of the echo signals in S2; S4, based on the characteristic data of the echo signals in S3, establishing a training data set and a test data set; S5, establishing an LSTM network model, and inputting the training data set in S4 to train the LSTM network model; S6, using the trained LSTM model in S5 to perform classification tests on the test data set in S3. The present invention can not only reduce the complexity of signals, but also obtain high-order semantic information of signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar information acquisition and processing, and particularly relates to a method for recognizing radar echoes in a complex environment. Background Art

[0002] With the rapid development of radar communication and electronic information technology, the types and quantities of signal systems have increased sharply, the signal waveform design has become more complex, and the types of ground objects in the radar detection environment are complex and diverse, the landform features are different, and the spatial heterogeneity is strong. As a result, traditional signal processing and analysis methods can no longer meet the current usage requirements. As a general technology that does not require professional domain knowledge, deep learning has the ability to integrate massive, multi-modal, and dynamic big data, and can quickly extract high-order semantic information from labeled data, which coincides with the development requirements and trends of recognizing radar echo signals in complex environments, and is expected to improve the processing ability and efficiency of radar systems for echo signals in complex environments.

[0003] Currently, the technology mainly uses time-frequency analysis methods to transform the problem of radio signal recognition into the field of image recognition, and uses deep neural networks to extract features from two-dimensional spectrograms or directly uses long short-term memory neural networks to mine the essence of signal sequences. Compared with traditional radar signal processing methods, the information processing efficiency and robustness have been improved, but there are mainly two problems: one is that relevant information in the signal layer will be lost through time-frequency transformation, the essential features of radar echo signals cannot be mined, and the two-dimensional convolution training has a high complexity, a large number of parameters, and a high network redundancy; the other is that the features extracted from different radar echo signals through traditional radar signal processing are almost indistinguishable, and directly inputting into LSTM (long short-term memory network) is likely to result in a poor classification effect.

[0004] Based on the above, the present invention proposes a method for recognizing complex environment echoes based on an invariant scattering convolutional network and LSTM, which can not only reduce the complexity of the signal, but also obtain high-order semantic information of the signal. Summary of the Invention

[0005] The purpose of the present invention is to propose an efficient method for recognizing radar echoes in a complex environment. It mainly uses an invariant scattering convolutional network to perform information dimensionality reduction and feature extraction on the input radar echo signal, reducing information redundancy, and then uses LSTM to mine the essential features of the signal sequence to achieve the recognition of radar echoes in a complex environment. At the same time, aiming at other noises existing in the signal, the attention mechanism is used to enhance the intrinsic features of the radar echo signal and improve the recognition rate of the signal. Compared with previous radar echo signal recognition methods, the present invention can not only reduce the complexity of the signal, but also obtain high-order semantic information of the signal, and is a radar echo recognition method with application prospects.

[0006] To achieve the above object, the present invention proposes a complex environment echo recognition method based on an invariant scattering convolutional network and LSTM, comprising the following steps:

[0007] S1. Establish geometric models and electromagnetic parameter models of several typical environments;

[0008] S2. Based on the geometric models and electromagnetic parameter models in S1, simulate the echo signals under the pulsed Doppler radar system;

[0009] S3. Establish an invariant scattering convolutional network and extract the characteristic data of the echo signals in S2;

[0010] S4. Based on the characteristic data of the echo signals in S3, establish a training data set and a test data set;

[0011] S5. Establish an LSTM network model and input the training data set in S4 to train the LSTM network model;

[0012] S6. Use the trained LSTM model in S5 to conduct classification tests on the test data set in S3.

[0013] Among them, the step S1 further includes the following steps:

[0014] S11. Select several typical environments and establish their geometric models;

[0015] S12. Establish the electromagnetic parameter models of the several typical environments described in S11.

[0016] Among them, under the pulsed Doppler radar system, a fuze antenna model is established. The fuze antenna emits waves to the environment model established in step S1, and the environment model scatters the waves. The scattered waves received by the fuze antenna are the echoes.

[0017] Among them, the environmental area that contributes to the echo signal is divided into M×N surface elements. Then, the echo signal described in step S2 is

[0018]

[0019] Among them, P t is the receiving power of the fuze antenna, λ is the wavelength of the wave emitted by the fuze antenna, k is the wave number, Δs is the area of each surface element in the irradiation area of the fuze antenna, and δ mn is the RCS of each surface element, and R is the distance from the fuze antenna to the surface element.

[0020] Among them, the value range of the said R is t1C / 2 < R < (t1 + T p +T v )C / 2, where T pis the transmission pulse width of the fuze antenna, t1 is the reception delay, and T v is the reception wave gate pulse width of the fuze antenna, and C is the speed of light.

[0021] Among them, when calculating the δ mn , according to different incident angles of the transmitted wave, for small incident angle cases, that is, when the incident angle is less than 20°, the Kirchhoff approximation method is used for calculation; for large incident angle cases, that is, when the incident angle is 20° - 90°, the perturbation method is used for calculation.

[0022] Among them, the invariant scattering convolutional network in the S3 includes multiple layers, and each layer generally includes three parts: wavelet convolution, non-linear processing, and average operation.

[0023] Among them, the step S4 is specifically to randomly scatter the echo signal feature data extracted in step S3, and establish a training data set and a test data set according to a certain ratio, where the proportion of the training data set is greater than that of the test data set.

[0024] Among them, the LSTM model includes an input layer, an LSTM layer, an attention mechanism layer, a fully connected layer, a softmax layer, and a classification result layer.

[0025] Among them, in the step S5, the following steps are further included:

[0026] S51: Input the feature data in the training data set in step S4 into the input layer in sequence;

[0027] S52: Calculate the echo data input from the input layer in the LSTM layer;

[0028] S53: Use the attention mechanism layer to enhance the output result of the LSTM layer;

[0029] S54: Classify the result using the fully connected layer and the softmax layer;

[0030] S55: Output the classification result in the classification result layer, and adjust the parameters in the attention mechanism layer, the fully connected layer, and the softmax layer according to the classification result;

[0031] S56: Repeat the above steps until all the feature data in the training data set are classified, and adjust the parameters of each layer to obtain a trained LSTM network model. Description of the Drawings

[0032] Figure 1 is a flowchart of a method for radar echo recognition in a complex environment based on an invariant scattering convolutional network and LSTM of the present invention;

[0033] Figure 2It is a schematic diagram of feature extraction by the invariant scattering convolutional network in the present invention.

[0034] Figure 3 It is a schematic diagram of the LSTM classification model in the present invention. Detailed implementation manners

[0035] The following further elaborates in detail on a complex environment echo recognition method based on an invariant scattering convolutional network and LSTM proposed by the present invention in conjunction with the accompanying drawings and a preferred specific embodiment. It should be specifically noted that the following embodiments are only used to illustrate the present invention, but do not limit the scope of the present invention. Similarly, the following embodiments are only partial embodiments of the present invention rather than all embodiments. Technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments should all be within the protection scope determined by the claims.

[0036] As Figure 1 shown, a complex environment echo recognition method based on an invariant scattering convolutional network and LSTM includes the following steps:

[0037] S1. Establish a geometric model and an electromagnetic parameter model of a typical environment;

[0038] Select four typical environments of soil, sand, cement, and wind-driven sea surface, and establish their geometric models and electromagnetic parameter models respectively.

[0039] S11. Establish geometric models of the four typical environments;

[0040] Generate digital geometric models of the four typical environments by frequency-domain conversion: for the wind-driven sea surface, mainly determine the simulated wave height spectrum by the PM sea spectrum, and use a Gaussian distribution centered on the wind direction to simulate the wave propagation direction, where the root mean square parameter of the Gaussian distribution is taken as 0.8; for the soil interface, use a negative power spectrum model to simulate its geometric model; for the sand interface, use a Gaussian correlation function with both a large variance and a large correlation length, that is, greater than 10 cm, to simulate its geometric model; for the cement interface, use a Gaussian correlation function with both a small variance and a small correlation length, that is, less than 0.1 cm, to simulate its geometric model.

[0041] S12. Establish electromagnetic parameter models of the four typical environments;

[0042] Establish electromagnetic parameter models of the four typical environments of soil, sand, cement, and wind-driven sea surface, that is, calculate the equivalent dielectric constant composed of the dielectric constants of various substances in the four environments.

[0043] The dielectric constant of the wind-driven sea surface, based on the dielectric constant of pure water, also depends on the salinity and temperature of seawater; and since soil, sand, and cement are all random mixtures composed of various substances such as gravel and water, their dielectric properties need to be represented by an equivalent average dielectric constant, so a semi-empirical model of volume mixing is used for calculation. And because the dielectric constant of water varies greatly with frequency in the microwave band, the equivalent dielectric constant of this type of ground such as soil, sand, and cement varies with frequency, and the variation is mainly determined by its water content.

[0044] Taking soil as an example, the empirical analytical calculation formula for its relative dielectric constant ε is given: Assume W T and β are parameters related to the structure of the soil itself, and W g represents the weight humidity of the soil. The calculation of the relative dielectric constant of the soil is divided into the following two cases:

[0045] When W g <W T ,

[0046] ε = W g ε x +(P - W g )ε a +(1 - P)ε r

[0047] ε x =ε i +(ε w - ε i )β

[0048] When W g >W T ,

[0049] ε = W g ε x +(W g - W T )ε w +(P - W g )ε a +(1 - P)ε r

[0050] ε x =ε i +(ε w - ε i )β

[0051] Among them, P is the product porosity, P = 1 - ρ b / ρ r , and ρ b is the dry soil density, ρ b ≈1.3g / cm 3 , ρ ris the density of the rock in the soil, and ρ r ≈2.6 g / cm 3 ; ε a 、ε r 、ε i are the dielectric constants of air, rock, and ice respectively. Generally, ε a = 1, ε r = 5 + 0.1i and ε i = 3.2 + 0.001i; ε w is the dielectric constant of free water; ε x is the dielectric constant of the water absorbed in the soil, and can be approximately taken as the average value of ε w and ε i .

[0052] Among them, the calculation expression of ε w is

[0053] ε w = 4.9 + (ε w0 (T) - 4.9) / (1 + j2πτ w (T)f)

[0054] ε w0 (T) = 88.045 - 0.4147T + 6.295×10 -4 T 2 + 1.075×10 -5 T 3

[0055] 2πτ w (T) = 1.1109×10 -10 - 3.824×10 -12 T

[0056] + 6.938×10 -14 T 2 - 5.096×10 -16 T 3

[0057] Among them, ε w0 (T) is the static dielectric constant of water at temperature T; τ w (T) is the relaxation time of pure water at temperature T; f is the electromagnetic wave frequency; T is the temperature of water.

[0058] The relative dielectric constants of the other three typical environments are calculated according to the existing calculation methods, and will not be elaborated here.

[0059] S2. Based on the four groups of geometric models and electromagnetic parameter models in step S1, simulate and calculate the echo signals under the pulsed Doppler radar system;

[0060] In the pulsed Doppler radar system, the wave emitted by the fuse antenna reaches the four groups of geometric models and electromagnetic parameter models established in step S1 respectively, and part of the waves scattered by each model are received by the fuse antenna again, and this received wave is the echo. Since the fuse antenna is a narrow-beam antenna with low side lobes, the fuse antenna model is described by a simplified two-dimensional antenna pattern. The fuse antenna model mainly combines spherical waves and the normalized gain distribution of the two-dimensional antenna pattern, and then through weighted calculation, the influence of the gains of the fuse transmitting and receiving antennas on the scattering contributions of various parts of the typical environmental surface is simulated.

[0061] In the pulsed Doppler radar system, assume that the width of the wave emitted by the fuse antenna is T p , the receiving delay of the fuse antenna is t1, and the receiving wave gate pulse width is T v , restricted by the region affected by the transmitted wave and the receiving wave gate, the region that contributes to the echo received by the fuse is the environmental region between two distance circles, that is, the part between two circles with radii of t1C / 2 and (t1 + T p +T v )C / 2 respectively, where C is the speed of light. It can be seen from this that the environmental region to be calculated in the pulsed Doppler radar system is the intersection of the antenna main lobe illumination region and the environmental region between two distance circles, and this part is divided into M×N surface elements.

[0062] Finally, the echoes of the soil, sand, cement ground and wind-driven sea surface received by the fuse antenna are uniformly recorded as echo V2

[0063]

[0064] t1C / 2 < R < (t1 + T p +T v )C / 2

[0065] Among them, P t is the receiving power of the fuse antenna, λ is the wavelength of the wave emitted by the fuse antenna, k is the wave number of the transmitted wave, Δs is the area of each surface element in the illumination region of the fuse antenna, and δ mn is the RCS (radar cross section) of each surface element, R is the distance from the fuse antenna to the surface element, and i is the imaginary unit.

[0066] Among them, the calculation of δ mn needs to be carried out separately according to different incident angles: the Kirchhoff (KA) approximation and the perturbation method are respectively used to calculate the scattering fields of each surface element, and then δ mn can be obtained. Substituting it into the echo formula, the echo V2 can be obtained. Next, the calculation of the scattering field of each surface element in two cases of small incident angle and large incident angle is given.

[0067] S21. For the case of small incident angles, i.e., when the incident angle is less than 20°, the initial electromagnetic current density of each surface element is obtained using the KA approximation, and then the scattered field vectors of each surface element are summed according to the Stratton-Chu integral equation. Here, the high-order scattering contributions between the sea surfaces are ignored during the calculation.

[0068] The initial scattered field E of a single sea surface element obtained using the KA approximation s has the following calculation formula:

[0069]

[0070] F(k i , k s ) = -(e i ·q i )(n·k i )(1 - R h )q i +(e i ·p i )(n×q i )(1 + R v ) +(e i ·q i )[k s ×(n×q i )](1 + R h )+(e i ·p i )(n·k i )[k s ×q i (1 - R v )]

[0071] Among them, k is the wave number of the transmitted wave, k i , k s are the unit wave vectors of the incident wave and the scattered wave respectively; r i , r s are the distances from the transmitting and receiving antennas to the surface element respectively; E0 is the amplitude of the incident field, I is the unit dyad, and ΔA is the area of the sea surface element calculated from the geometric model established in step S11; e i , n are the normal vectors of the surface element; p i = q i ×k i , q i and p i are the local horizontal and vertical polarization vectors respectively; R v , R h are the vertical and horizontal polarization reflection coefficients of the calm sea surface calculated from the electromagnetic parameter model established in step S12.

[0072] S22. For the case of large incident angles, that is, when the incident angle is 20° to 90°, the following calculation formula for the first-order backscattering coefficient of the ground-sea surface is given according to the perturbation method:

[0073]

[0074]

[0075] where k is the wave number of the transmitted wave; W(2ksinθ i ) is the wave height spectrum of the rough sea surface, and θ i is the incident angle. For a certain surface element of the rough ground-sea surface, the polarization scattering field can give the following statistical approximation:

[0076]

[0077] where or represents the first-order backscattering coefficient of vertical or horizontal polarization; is the random phase.

[0078] S3. Establish an invariant scattering convolutional network to automatically extract the characteristic data of the echo signal;

[0079] The invariant scattering convolutional network is a technology that can be used to automatically extract low-variance and compact features, which can minimize the differences within a class while retaining the distinguishability between classes. Using the invariant scattering convolutional network, low-variance and compact features are automatically extracted instead of providing the original representation to the LSTM so that the LSTM can quickly learn the patterns and then classify the signals.

[0080] The Fourier transform is a global transform that reflects the global energy of an image. Therefore, the energy spectrum of the Fourier transform cannot well describe the local features of the image; the wavelet transform is a localized transform that reflects the localized features of the image. Different choices of wavelet functions result in different extracted features. The wavelet transform can extract the frequency features of an image at multiple scales and in multiple directions, thereby obtaining the energy distribution features of the image at multiple scales and in multiple directions.

[0081] The invariant scattering convolutional network usually includes multiple layers. In each layer, the wavelet transform is first used to hierarchically extract the high-frequency information of the image, that is, the detailed information of the image, and then the average operator is used to convert the high-frequency into low-frequency, so as to maintain the stability of the high-frequency information. The high-frequency information lost in this hierarchical process can be recovered in the next layer of this invariant scattering convolutional network. Therefore, the invariant scattering convolutional network can better extract the high-frequency information of different levels of the image.

[0082] Such as Figure 2As shown, the invariant scattering convolutional network is a framework for automatically extracting relevant compact features. The invariant scattering convolutional network includes multiple layers, and each layer generally includes three parts: wavelet convolution, non-linear processing, and averaging operation.

[0083] The specific process in this embodiment is as follows:

[0084] S31. Use the software tool MATLAB to sequentially read the echo V2 obtained by simulation in S2;

[0085] S32. Perform feature extraction on the echo signal sequentially;

[0086] S321. Use the Wavelet Scattering function in MATLAB to create a wavelet time scattering decomposition framework with two filter banks;

[0087] S322. Use the Feature matrix in MATLAB to perform dimensionality reduction and feature extraction on the echo signal sequentially;

[0088] S323. Use the Scattering Transform and Scatter gram functions in MATLAB to visualize the features and observe the differences in radar echoes in different environments;

[0089] S33. Save the above-extracted feature data, re-enter the echo signal processed in steps S321 to S323 into step S321, and repeat the above process.

[0090] S34. Repeat step S33 and stop after 100 times, and save all feature data.

[0091] S4. Based on the feature data of the echo signal, establish a training data set and a test data set;

[0092] Randomly shuffle the feature data extracted in step S3, place the feature data in different folders and rename their file names. Establish a training data set and a test data set from the different folders according to a ratio of 7:3.

[0093] S5. Establish an LSTM network model and input the training data set into it to train the LSTM network model;

[0094] As Figure 3 shown, the entire LSTM architecture is constructed using MATLAB and has a total of six layers: the first layer is the input layer; the second layer is the LSTM layer with 100 hidden units; the third layer is the attention mechanism layer; the fourth layer is the fully connected layer; the fifth layer is the softmax layer; the sixth layer is the classification result layer. The specific steps are as follows:

[0095] S51. Input the feature data in the training dataset in step S4 into the input layer sequentially;

[0096] S52. The echo data input from the input layer enters the LSTM layer for calculation. The specific calculation formula of the LSTM layer is as follows:

[0097] i t = σ(W xi x t + W hi h t-1 + W ci c t-1 + b i )

[0098] f t = σ(W xf x t + W hf h t-1 + W cf c t-1 + b f )

[0099] g t = tanh(W xc x t + W hc h t-1 + W cc c t-1 + b c )

[0100] c t = i t g t + f t c t-1

[0101] o t = σ(W xo x t + W ho h t-1 + W co c t-1 + b o )

[0102] h t = o t tanh(c t )

[0103] Among them, W xi , W hi , W ci , W xf , etc. are the neuron weight parameters of different gates in the LSTM layer, and x tis the input at the current moment, x, h, c, and b are weight factors, and b i , b f , b c , b o are the biases of different gates, i t , f t , g t , c t and o t are the output results in the process, h t-1 , c t-1 are the results in the previous moment of the process, h t is the final result output by one layer of LSTM at the current moment, and it is denoted as H.

[0104] S53. Use the attention mechanism layer to enhance the output result H of the LSTM layer;

[0105] The formula of the attention mechanism layer is as follows:

[0106] M = tanh(H)

[0107] α = softmax(w T M)

[0108] r = Hα T

[0109] where w is the weight of the neuron, and M, α, and r are the model parameters in the attention mechanism layer.

[0110] S54. Use the fully connected layer and the softmax layer to classify the output result H enhanced by S53;

[0111] S55. Output the classification result in the classification result layer, and adjust the parameters in the attention mechanism layer, the fully connected layer, and the softmax layer according to the classification result;

[0112] S56. Repeat the above process until all the feature data in the training data set are classified, and at the same time adjust the parameters of each layer to obtain the trained LSTM network model.

[0113] S6. Use the trained LSTM model in S5 to classify and test the test data set;

[0114] Output the feature data in the test data set in step S4 into the LSTM model trained in S56. The feature data in the test data set pass through the input layer, the LSTM layer, the attention mechanism layer, the fully connected layer, and the softmax layer in the LSTM model in turn, and obtain the classification result from the classification result layer to complete the test.

[0115] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A complex environment echo recognition method based on an invariant scattering convolutional network and LSTM, characterized in that It includes the following steps: S1. Establish the geometric models and electromagnetic parameter models of four typical environments; the four typical environments include soil, sandy land, cement ground, and wind-driven sea surface; S2. Based on the geometric models and electromagnetic parameter models in S1, simulate the echo signals under the pulsed Doppler radar system; S3. Establish an invariant scattering convolutional network to extract the characteristic data of the echo signals in S2; S4. Based on the characteristic data of the echo signals in S3, establish a training data set and a test data set; S5. Establish an LSTM network model, and input the training data set in S4 to train the LSTM network model; S6. Use the trained LSTM model in S5 to conduct classification tests on the test data set in S3; Among them, step S1 further includes the following steps: S11. Select four typical environments and establish their geometric models; Generate digital geometric models of four typical environments by frequency-domain and time-domain conversion: for the wind-driven sea surface, determine the wave height spectrum generated by simulation from the PM sea spectrum, and use a Gaussian distribution centered on the wind direction to simulate the wave propagation direction, where the root mean square parameter of the Gaussian distribution is taken as 0.8; for the soil interface, use a negative power spectrum model to simulate its geometric model; for the sandy land interface, use a Gaussian correlation function with a variance and correlation length greater than 10 cm to simulate its geometric model; for the cement ground interface, use a Gaussian correlation function with a variance and correlation length less than 0.1 cm to simulate its geometric model; S12. Establish the electromagnetic parameter models of the four typical environments described in S11; Among them, under the pulsed Doppler radar system, establish a fuze antenna model. The fuze antenna emits waves to the environmental model established in step S1, and the environmental model scatters the emitted waves, and the scattered waves received by the fuze antenna are the echoes; Among them, the environmental area that contributes to the echo signals is divided into M×N surface elements, and then the echo signals described in step S2 are Among them, P t is the received power of the fuze antenna, λ is the wavelength of the wave emitted by the fuze antenna, k is the wave number, Δs is the area of each surface element in the illumination area of the fuze antenna, and δ mn is the RCS of each surface element, and R is the distance from the fuze antenna to the surface element; The value range of R is t1C / 2 < R < (t1 + T p + T v )C / 2, where T p is the transmission pulse width of the fuze antenna, t1 is the reception delay, T v is the reception wave gate pulse width of the fuze antenna, and C is the speed of light; when calculating the δ mn , according to different incident angles of the transmitted wave, for the case of small incident angles, that is, when the incident angle is less than 20°, the Kirchhoff approximation method is used for calculation; for the case of large incident angles, that is, when the incident angle is 20° to 90°, the perturbation method is used for calculation.

2. The complex environment echo recognition method based on the invariant scattering convolutional network and LSTM according to claim 1, characterized in that The invariant scattering convolutional network in S3 includes multiple layers, and each layer generally includes three parts: wavelet convolution, nonlinear processing, and average operation.

3. A complex environment echo recognition method based on an invariant scattering convolutional network and LSTM according to claim 1, characterized in that, Step S4 specifically randomizes the characteristic data of the echo signals extracted in step S3 and establishes a training data set and a test data set according to a certain ratio, where the proportion of the training data set is greater than that of the test data set.

4. A complex environment echo recognition method based on an invariant scattering convolutional network and LSTM according to claim 1, characterized in that The LSTM model includes an input layer, an LSTM layer, an attention mechanism layer, a fully connected layer, a softmax layer, and a classification result layer.

5. The complex environment echo recognition method based on the invariant scattering convolutional network and LSTM according to claim 4, characterized in that, In step S5, it further includes the following steps: S51. Sequentially input the characteristic data in the training data set in step S4 through the input layer; S52. Calculate the echo data input from the input layer in the LSTM layer; S53. Use the attention mechanism layer to enhance the output result of the LSTM layer; S54. Use the fully connected layer and the softmax layer to classify the result; S55. Output the classification result in the classification result layer, and adjust the parameters in the attention mechanism layer, the fully connected layer, and the softmax layer according to the classification result; S56. Repeat the above steps until all the feature data in the training dataset are classified and the parameters of each layer are adjusted to obtain a trained LSTM network model.

Images