Underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2

Abstract

The invention belongs to the field of sonar signal processing, and provides an underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2. According to the technology, a micro-motion echo model of propeller blades of an underwater vehicle is established, and micro-Doppler characteristics of blades of different numbers and blades of different rotating speeds are obtained by adopting a multi-harmonic signal combined transmitting signal form and utilizing a time-frequency analysis method. In order to more accurately identify and perceive propellers of different types, the invention constructs a CAMobileNetV2 model which is combined with a CBAM (Convolutional Block Attention Module) and MobileNetV2 (Convolutional Block Attention Module), and the CAMobileNetV2 (Convolutional Block Attention Module) model is used for identifying and perceiving propellers of different types. The model can effectively analyze and extract the micro-Doppler characteristics of the propeller, thereby realizing identification of different underwater vehicles. The constructed data set is trained and tested, and the effectiveness and robustness of the technology in underwater sonar signal processing and underwater target identification are shown. The invention provides a micro-Doppler feature recognition technology in the field of underwater sonar signal processing, and is beneficial to improving the target recognition capability of underwater sonar.

Description

Technical Field

[0001] The present invention belongs to the field of sonar signal processing, and specifically relates to an underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2. Technical Background

[0002] Underwater sonar target recognition technology is an important research topic in the field of underwater acoustic signal processing. Most traditional underwater target recognition methods are based on Doppler information carried by echo signals or underwater sonar imaging. When underwater vehicles are in motion, the micro-Doppler effect caused by the small frequency shift caused by the rotation of their own propeller blades is also an important recognition feature. The echo signal of the blade is analyzed in time and frequency, considering the number of propeller blades, the rotation speed and the emission form of the multi-harmonic signal. The number of blades directly affects the spectral characteristics of the echo signal. Each blade will generate a specific micro-Doppler frequency component during the rotation process. The more blades there are, the more frequency components there are in the echo signal. By analyzing these frequency components, information about the number of blades can be obtained. The rotation speed determines the frequency at which the blade passes through a specific spatial point, that is, the frequency shift of the micro-Doppler effect. High rotation speed will result in a larger frequency shift, while low rotation speed will result in a smaller frequency shift. By accurately measuring these frequency shifts, the propeller rotation speed can be inferred. By transmitting multi-harmonic signals, echo signals containing multiple frequency components can be obtained at the receiving end. The distribution of different harmonic components in the time-frequency domain and their changing trends can reflect the motion state and geometric characteristics of the propeller blades. The complete micro-Doppler characteristics can be obtained based on the number of propeller blades, the rotation speed, and the combination of the transmitted multi-harmonic signals.

[0003] In the field of deep learning, attention mechanism has become one of the important technologies to improve model performance. It enables the model to better capture key information in the data by emphasizing important features and suppressing unimportant features. CBAM (Convolutional BlockAttention Module) is an effective attention mechanism that combines channel attention and spatial attention to adaptively adjust the importance of feature maps in channel and spatial dimensions to further enhance model performance. MobileNetV2 is an efficient convolutional neural network structure that uses deep separable convolution and inverted residual structure to reduce model parameters and computation. Based on MobileNetV2, the present invention introduces the CBAM module to construct a CAMobileNetV2 fusion model to improve the feature extraction capability and classification performance of the lightweight model. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and propose an underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 to improve the ability of underwater sonar target recognition in the prior art.

[0005] In order to achieve the above object, the present invention adopts the following technical methods to achieve it.

[0006] The underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 includes the following steps:

[0007] Step 1, using the multi-harmonic constant frequency-frequency modulation signal as the sonar transmission signal, and establishing a transmission signal model;

[0008] Step 2, model the underwater propeller blades, and obtain the echo model and micro-Doppler information of the blades according to the sonar transmission signal;

[0009] Step 3, performing time-frequency analysis on the echo signal to obtain a micro-Doppler time-frequency diagram;

[0010] Step 4, taking the time-frequency diagrams obtained by combining different numbers of blades, different blade rotation speeds, and different multi-harmonic signals as a data set;

[0011] Step 5, divide the data set into training set and test set in proportion;

[0012] Step 6: Combine MobileNetV2 and the attention mechanism to build a CAMobileNetV2 model to improve the performance of the lightweight neural network, extract micro-Doppler features from the input time-frequency map, and fuse the spatial and channel features to obtain a fused feature map for classification and recognition.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] (1) The micro-Doppler generated by the rotation of the propeller blades of the underwater vehicle is used as a feature for identification, which solves the problem of excessive reverberation covering the target features in complex underwater scenes and can effectively extract the micro-motion information of the target.

[0015] (2) The transmission signal adopts a combination of multiple harmonics to ensure that the echo combination of different frequencies can carry more echo information, enhance the anti-reverberation performance, and improve the feature capture capability.

[0016] (3) The CAMobileNetV2 model can adapt to different environmental conditions and target types, making it more flexible in dealing with various underwater environments and diverse targets. The model can also learn to extract high-resolution features and capture subtle details that traditional methods may overlook. The CBAM attention mechanism is introduced into the model to improve feature extraction capabilities and improve the performance of micro-Doppler recognition features. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Figure 1 This is a flowchart of underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2.

[0019] Figure 2 This is a model diagram of propeller blade micromotion.

[0020] Figure 3 It is the time-frequency diagram of the simulated transmitted signal.

[0021] Figure 4 This is the CAMobileNetV2 model diagram.

[0022] Figure 5(a) is a micro-Doppler characteristic diagram when the number of blades is 3.

[0023] FIG5( b ) is a micro-Doppler characteristic diagram when the number of blades is 4. ...

[0024] FIG5( c ) is a micro-Doppler characteristic diagram when the transmitted harmonics are the first harmonic ( S1 ) and the second harmonic ( S2 ).

[0025] FIG5( d ) is a micro-Doppler characteristic diagram when the transmitted harmonic is the first harmonic ( S1 ) and the third harmonic ( S3 ).

[0026] FIG5(e) is a micro-Doppler characteristic diagram when the transmitted harmonic is the first harmonic (S1) and the third harmonic (S4).

[0027] Figure 5(f) is a micro-Doppler characteristic diagram at a rotation speed of 5 m / s.

[0028] Figure 5(g) is a micro-Doppler characteristic diagram at a rotation speed of 6 m / s.

[0029] Figure 5(h) is a micro-Doppler characteristic diagram at a rotation speed of 7 m / s.

[0030] Figure 6 is the model classification accuracy curve. DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0032] like Figure 1 As shown in the figure, the underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 includes the following steps:

[0033] Step 1, using the multi-harmonic constant frequency-frequency modulation signal as the sonar transmission signal, and establishing a transmission signal model;

[0034] Specifically, step 1 includes the following sub-steps:

[0035] Sub-step 1.1, the sonar system transmits a pulse composed of 4 harmonics, each of which contains a long constant frequency (CF) component and a short frequency modulation (FM) component, which is divided into 8 subcarriers in total, namely CF1-CF4 and FM1-FM4:

[0036] The time domain expression of the CF component is:

[0037]

[0038] Where u i (t) represents the complex envelope of the signal, f cf represents the carrier frequency, T represents the signal duration, and i represents the number of harmonics.

[0039] The FM component is a hyperbolic frequency modulation signal. Assume that the bandwidth of the transmitted signal is B, and the upper and lower limit frequencies are and f fm Represents the center frequency of the transmitted signal, recorded as:

[0040]

[0041] Then the time domain expression of the FM component is:

[0042]

[0043] Where k is the frequency modulation parameter, that is:

[0044]

[0045] r(t) represents a standard rectangular pulse function and has the following form:

[0046]

[0047] It can be seen that the phase and instantaneous frequency of the signal can be expressed as:

[0048]

[0049]

[0050] Then the four-harmonic CF-FM signal emitted by the bionic sonar is:

[0051]

[0052] Step 2, model the underwater propeller blades, and obtain the echo model and micro-Doppler information of the blades according to the sonar transmission signal;

[0053] Sub-step 2.1, mathematical modeling of the propeller blades, the phase function is:

[0054]

[0055] It can be seen that the echo signal of the blade is

[0056]

[0057]

[0058] The echo is integrated into the blade length L, and the blade echo is:

[0059]

[0060]

[0061] The micro-Doppler of the echo signal is expressed as:

[0062]

[0063] Step 3, performing time-frequency analysis on the echo signal to obtain a micro-Doppler time-frequency diagram;

[0064] Specifically, step 3 includes the following sub-steps:

[0065] Sub-step 3.1, use short-time Fourier transform to perform time-frequency analysis on the echo signal of the propeller blade to obtain a time-frequency diagram containing micro-Doppler information. The obtained time-frequency diagram includes the motion characteristics of the propeller blade at different times and frequencies, which can not only capture the rotation speed and angular velocity changes of the propeller blade, but also serve as a basis for the classification and identification of different propeller blades.

[0066] Step 4, taking the time-frequency diagrams obtained by combining different numbers of blades, different blade rotation speeds, and different multi-harmonic signals as a data set;

[0067] Specifically, step 4 includes the following sub-steps:

[0068] Sub-step 4.1, by changing the number of blades, the rotation speed of the blades, and the combination form of the transmitted signal, each echo signal is subjected to time-frequency analysis to obtain a micro-Doppler time-frequency diagram;

[0069] Sub-step 4.2, integrate the micro-Doppler features under different variable conditions, and set the input size of the time-frequency map to 224*224. In order to increase the diversity of training data, random cropping, flipping, rotation and scaling methods are used to enhance the data to form an underwater propeller blade micro-Doppler dataset.

[0070] Step 5, divide the data set into training set and test set in proportion;

[0071] Specifically, step 5 includes the following sub-steps:

[0072] In sub-step 5.1, each type of echo in the propeller blade micro-Doppler feature data set is randomly extracted and combined into a training set and a test set in a ratio of 8:2.

[0073] Step 6: Combine MobileNetV2 and the attention mechanism to build a CAMobileNetV2 model to improve the performance of the lightweight neural network, extract micro-Doppler features from the input time-frequency map, and fuse the spatial and channel features to obtain a fused feature map for classification and recognition.

[0074] Specifically, step 6 includes the following sub-steps:

[0075] Sub-step 6.1, build a channel attention module and generate an attention weight for each channel. Given an input feature map Where C is the number of channels, H and W are the height and width of the feature map respectively, and the calculation process of the channel attention module is as follows:

[0076] Calculate the global average pooling and global maximum pooling to obtain two feature vectors that describe the global information of the channel:

[0077] F avg =AvgPool(F),F max =MaxPool(F)

[0078] in,

[0079] Pass each of these two feature vectors through a shared multi-layer perceptron (MLP) and then add them together:

[0080] M c =σ(MLP(F avg )+MLP(F max ))

[0081] Among them, σ represents the sigmoid function, is the channel attention map.

[0082] Weight the input feature map using the channel attention map:

[0083] F′=M c ·F

[0084] Sub-step 6.2, build a spatial attention module to generate an attention weight for each spatial position. Given an input feature map The calculation process of the spatial attention module is as follows:

[0085] Calculate the average and maximum pooling of the feature map in the channel dimension to obtain two feature maps that describe spatial information:

[0086] F′ avg =AvgPool c (F′), F′ max =MaxPool c (F′)

[0087] in,

[0088] Concatenate these two feature maps along the channel dimension and pass them through a convolutional layer:

[0089] M s =σ(Conv([F′ avg ; F′ max ]))

[0090] Among them, σ represents the sigmoid function is the spatial attention map.

[0091] Weight the input feature map using the spatial attention map:

[0092] F″=M s ·F′

[0093] At this point, the CBAM module is built.

[0094] Through the above two steps, the CBAM module can adaptively adjust the weights of the feature maps in both channel and spatial dimensions, thereby improving the representation ability of the model.

[0095] Sub-step 6.3, according to the above steps, the channel attention mechanism module and the spatial attention mechanism module are fused to obtain the CBAM module. MobileNetV2 is used as the basic network, which contains multiple inverted residual modules. A CBAM module is added after each inverted residual module to enhance its feature extraction capability and build a fusion model of CAMobileNetV2.

[0096] Sub-step 6.4, shuffling the time-frequency feature graphs in the divided training set to obtain a shuffled training set;

[0097] In substep 6.5, the initial learning rate of the model is set to 0.001. The optimizer uses the Adam optimizer to iteratively update the parameters of the fusion network. The loss function uses the cross entropy loss function, which is suitable for multi-classification tasks. The training batch size is 32 and the number of training rounds is 50. Using the learning rate scheduler, the learning rate is gradually reduced during the training process. When the validation loss no longer decreases, the learning rate is reduced by half. The weight initialization uses He initialization, which is suitable for the ReLU activation function. To prevent overfitting, L2 regularization is used in the convolutional layer and the fully connected layer. Use the early stopping technique to monitor the validation set loss. When the validation loss no longer decreases for 10 consecutive epochs, stop training early.

[0098] Sub-step 6.6, training is performed according to the preset model parameters. The test set data is input into the trained model, and the output is the classification results of different propeller blade types.

[0099] Table 1 Transmitting signal parameter settings

[0100]

[0101] The multi-harmonic emission signal is simulated according to the parameters in Table 1, and the time-frequency diagram is as follows: Figure 3 As shown. When the number of propeller blades is 3, the blade echo micro-Doppler time-frequency diagram is shown in Figure 5(a). The micro-Doppler time-frequency diagram when the number of blades is 4 is shown in Figure 5(b). The number of blades can be distinguished according to the micro-Doppler distribution in the time-frequency diagram. The echo Dopplers of different harmonic combinations are shown in Figures 5(c), 5(d) and 5(e). By utilizing the characteristic that the combination of harmonics of different frequencies can carry echo information of different frequencies, the micro-Doppler information of the blade echo can be better extracted. The micro-Doppler time-frequency diagrams generated by the change in blade speed are shown in Figures 5(f), 5(g) and 5(h). The size of the echo micro-Doppler caused by the change in speed is also reflected in the time-frequency diagram. Therefore, the above differences can be used for neural network recognition.

[0102] The training process of the CAMobileNetV2 model and the size of each layer are as follows Figure 3 As shown in the figure, the input feature map size is 224*224. The model uses MobileNetV2 as the basic network, which contains multiple inverted residual modules. A CBAM module is added after each inverted residual module to enhance the feature extraction capability.

[0103] The model implementation includes training and testing phases.

[0104] Table 2 Dataset parameter settings

[0105]

[0106] The recognition performance of the model was verified by computer simulation. The simulation data parameter settings are shown in Table 2, and the number of training times is 500. According to the data set settings in Table 2, it is divided into two categories: 3-blade propellers and 4-blade propellers. The rotation speed in each data set is divided into 0.01m / s. Combined with the emission signal form of three multi-harmonic combinations, a total of 603 feature maps are generated for each data set. The feature maps are divided into training sets and test sets in a ratio of 8:2 and sent to the CAMobileNetV2 model for recognition. The recognition accuracy curve is shown in Figure 6 To verify the stability of the model and prevent overfitting, 10 groups of random experiments were set up for identification. Figure 6 It can be seen that compared with the traditional model, the recognition accuracy and stability of the proposed model are greatly improved, and the overall recognition accuracy can reach 97.45%, which further shows that the model has a high recognition rate for multi-harmonic micro-Doppler features.

[0107] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. Underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2, the method technology includes the following steps: Step 1, using the multi-harmonic constant frequency-frequency modulation signal as the sonar transmission signal, and establishing a transmission signal model; Step 2, model the underwater propeller blades, and obtain the echo model and micro-Doppler information of the blades according to the sonar transmission signal; Step 3, performing time-frequency analysis on the echo signal to obtain a micro-Doppler time-frequency diagram; Step 4, taking the time-frequency diagrams obtained by combining different numbers of blades, different blade rotation speeds, and different multi-harmonic signals as a data set; Step 5, divide the data set into training set and test set in proportion; Step 6: Combine MobileNetV2 and the attention mechanism to build a CAMobileNetV2 model to improve the performance of the lightweight neural network, extract micro-Doppler features from the input time-frequency map, and fuse the spatial and channel features to obtain a fused feature map for classification and recognition.

2. The underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 according to claim 1 is characterized in that Combined with the micro-Doppler features under time-frequency analysis, the CAMobileNetV2 model is used to achieve target classification and recognition. Step 1 includes the following sub-steps: Sub-step 1.1, the sonar system transmits a pulse composed of 4 harmonics, each of which contains a long constant frequency (CF) component and a short frequency modulation (FM) component, which is divided into 8 subcarriers in total, namely CF1-CF4 and FM1-FM4. The time domain expression of the CF component is: Where u i (t) represents the complex envelope of the signal, f cf represents the carrier frequency, T represents the signal duration, and i represents the number of harmonics. The FM component is a hyperbolic frequency modulation signal. Assume that the bandwidth of the transmitted signal is B, and the upper and lower limit frequencies are and f fm Represents the center frequency of the transmitted signal, recorded as: Then the time domain expression of the FM component is: Where k is the frequency modulation parameter, that is: r(t) represents a standard rectangular pulse function and has the following form: It can be seen that the phase and instantaneous frequency of the signal can be expressed as: Then the four-harmonic CF-FM signal emitted by the bionic sonar is:

3. According to the underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 in claim 2, step 2 comprises the following sub-steps: Sub-step 2.1, mathematical modeling of the propeller blades is performed, as shown in Figure 1, and its phase function is: It can be seen that the echo signal of the blade is The echo is integrated into the blade length L, and the blade echo is: It can be seen that the micro-Doppler is:

4. According to the underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 in claim 3, step 3 comprises the following sub-steps: Sub-step 3.1, use short-time Fourier transform to perform time-frequency analysis on the echo signal of the propeller blade to obtain a time-frequency diagram containing micro-Doppler information. The obtained time-frequency diagram includes the motion characteristics of the propeller blade at different times and frequencies, which can not only capture the rotation speed and angular velocity changes of the propeller blade, but also serve as a basis for the classification and identification of different propeller blades.

5. According to the underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 in claim 4, step 4 comprises the following sub-steps: Sub-step 4.1, by changing the number of blades, the rotation speed of the blades, and the combination form of the transmitted signal, each echo signal is subjected to time-frequency analysis to obtain a micro-Doppler time-frequency diagram; In sub-step 4.2, the micro-Doppler features under different variable conditions are integrated, and the input size of the time-frequency map is set to 224*224. In order to increase the diversity of training data, random cropping, flipping, rotation and scaling are used to enhance the data to form an underwater propeller blade micro-Doppler dataset, as shown in Table 1.

6. According to the underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 in claim 5, step 5 comprises the following sub-steps: In sub-step 5.1, each type of echo in the propeller blade micro-Doppler feature data set is randomly extracted and combined into a training set and a test set in a ratio of 8:

2.

7. According to the underwater sonar multi-harmonic micro-Doppler feature recognition technology based on CAMobileNetV2 in claim 6, step 6 comprises the following sub-steps: Sub-step 6.1, build a channel attention module and generate an attention weight for each channel. Given an input feature map Where C is the number of channels, H and W are the height and width of the feature map respectively, and the calculation process of the channel attention module is as follows: Calculate the global average pooling and global maximum pooling to obtain two feature vectors that describe the global information of the channel: F avg =AvgPool(F),F max =MaxPool(F) in, Pass each of these two feature vectors through a shared multi-layer perceptron (MLP) and then add them together: M c =σ(MLP(F avg )+MLP(F max )) Among them, σ represents the sigmoid function, is the channel attention map. Weight the input feature map using the channel attention map: F′=M c ·F Sub-step 6.2, build a spatial attention module to generate an attention weight for each spatial position. Given an input feature map The calculation process of the spatial attention module is as follows: Calculate the average and maximum pooling of the feature map in the channel dimension to obtain two feature maps that describe spatial information: F′ avg =AvgPool c (F′),F′ max =MaxPool c (F′) in, Concatenate these two feature maps along the channel dimension and pass them through a convolutional layer: M s =σ(Conv([F′ avg ;F′ max ])) Among them, σ represents the sigmoid function is the spatial attention map. Weight the input feature map using the spatial attention map: F″=M s ·F′ At this point, the CBAM module is built. Through the above two steps, the CBAM module can adaptively adjust the weights of the feature maps in both channel and spatial dimensions, thereby improving the representation ability of the model. Sub-step 6.3, according to the above steps, the channel attention mechanism module and the spatial attention mechanism module are fused to obtain the CBAM module. MobileNetV2 is used as the basic network, which contains multiple inverted residual modules. A CBAM module is added after each inverted residual module to enhance its feature extraction capability and build a fusion model of CAMobileNetV2. Sub-step 6.4, shuffling the time-frequency feature graphs in the divided training set to obtain a shuffled training set; In substep 6.5, the initial learning rate of the model is set to 0.

001. The optimizer uses the Adam optimizer to iteratively update the parameters of the fusion network. The loss function uses the cross entropy loss function, which is suitable for multi-classification tasks. The training batch size is 32 and the number of training rounds is 50. Using the learning rate scheduler, the learning rate is gradually reduced during the training process. When the validation loss no longer decreases, the learning rate is reduced by half. The weight initialization uses He initialization, which is suitable for the ReLU activation function. To prevent overfitting, L2 regularization is used in the convolutional layer and the fully connected layer. Use the early stopping technique to monitor the validation set loss. When the validation loss no longer decreases for 10 consecutive epochs, stop training early. Sub-step 6.6, training is performed according to the preset model parameters. The test set data is input into the trained model, and the output is the classification results of different propeller blade types.

Images