A Ship Wake Detection Method Based on Convolutional Neural Network

Abstract

The present invention relates to a method for detecting ship acoustic wake based on a convolutional neural network, belonging to the field of target detection. It includes collecting scattered echo signals with and without wake, preprocessing the collected scattered echo signals, performing wavelet transform on the preprocessed scattered echo signals to obtain a wavelet coefficient matrix, dividing the time-frequency feature image set into a test set and a training set according to a certain ratio, building and training a convolutional neural network, and using the trained model for testing to achieve ship wake target detection. The method of the present invention utilizes the convolutional neural network to mine deeper information and improves the detection accuracy under low signal-to-noise ratio.

Description

Technical Field

[0001] The present invention belongs to the field of target detection, and particularly relates to a method for ship target detection by processing the scattered echo signal of ship wake. Background Art

[0002] During the navigation of a ship, a large amount of turbulent flow containing bubbles will be generated at its tail. Acoustic detection of wake has become an effective way to detect ship wake due to its advantages such as high detection probability, strong anti-interference ability, and long guiding distance.

[0003] Scholars at home and abroad have carried out a large number of studies on wake acoustic detection and proposed relevant detection algorithms, mainly focusing on detection after feature extraction. However, manual feature extraction requires certain professional knowledge and inevitably loses some information during the feature extraction process, resulting in poor robustness and generalization of the detection method based on feature extraction.

[0004] Deep learning uses multiple layers to gradually extract higher-level features from the original input and integrates the traditional algorithm of classification and recognition after extraction into an end-to-end classification model. In image processing, deep learning methods have demonstrated the ability of generalization and robustness. Summary of the Invention

[0005] Technical Problems to be Solved

[0006] In order to overcome the deficiencies of the prior art, the present invention proposes a ship wake detection method based on a convolutional neural network, which uses the convolutional neural network to mine deeper features of ship wake targets, thereby improving the accuracy of wake target detection under low signal-to-noise ratio.

[0007] Technical Solution

[0008] A ship acoustic wake detection method based on a convolutional neural network, characterized by the following steps:

[0009] Step 1: Collect scattered echo signals with and without wake; respectively collect scattered echo signals with and without wake when the detection platform is at different depths, different speeds, different signal emission angles, and different signal frequencies and forms.

[0010] Step 2: Preprocess the collected scattered echo signals to filter out the noise outside the frequency response range.

[0011] Step 3: Perform wavelet transform on the preprocessed scattered echo signals to obtain a wavelet coefficient matrix, and then perform visualization to obtain a time-frequency feature image set.

[0012] Step 4: Divide the time-frequency feature image set into a test set and a training set according to a certain ratio, and label each time-frequency diagram with a corresponding label.

[0013] Step 5: Considering the detection accuracy, computational complexity, and memory occupancy of the network comprehensively, build a convolutional neural network; the convolutional neural network is the WakeNet detection network, which consists of three ordinary convolutional layers, four bottleneck structures, and a global average pooling layer. The time-frequency diagram input to the network first passes through a convolutional layer with a convolutional kernel size of 3×3 to capture macroscopic features, then passes through four bottleneck structures to obtain deeper features, then passes through a convolutional layer with a convolutional kernel size of 1×1 to improve the non-linear classification performance of the network, then uses the global average pooling layer to reduce the dimension of the feature matrix, and finally uses a convolutional layer with a convolutional kernel size of 1×1 as a fully connected layer to output the classification result;

[0014] Step 6: Use the training set to train and optimize the convolutional neural network;

[0015] Step 7: After preprocessing and time-frequency feature extraction of the data to be measured, give the detection result through the optimized neural network model to achieve ship wake target detection.

[0016] A further technical solution of the present invention: The bottleneck structure described in Step 5 is an inverted residual structure, which is improved from the residual structure in the Resnet network. The bottleneck structure first passes through a convolutional layer with a convolutional kernel size of 1×1 to increase the feature dimension, then passes through a 3×3 DW convolution to extract features, and finally passes through a convolutional layer with a convolutional kernel size of 1×1 to reduce the feature dimension.

[0017] A further technical solution of the present invention: In Step 6, the learning effect and performance of the network are improved by adjusting the hyperparameters of the network. The hyperparameters include: learning rate and batch size.

[0018] A computer system, characterized by comprising: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the above method.

[0019] A computer-readable storage medium, characterized by storing computer-executable instructions, which are used to implement the above method when executed.

[0020] Beneficial effects

[0021] A ship acoustic wake detection method based on a convolutional neural network provided by the present invention improves the detection accuracy of ship wakes under low signal-to-noise ratios. Compared with traditional detection algorithms, this method uses a convolutional neural network to mine deeper information and improves the detection accuracy under low signal-to-noise ratios. Description of the drawings

[0022] The accompanying drawings are only for the purpose of showing specific embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference signs denote the same components.

[0023] Figure 1 It is a flowchart of the ship wake detection algorithm based on convolutional neural network of the present invention;

[0024] Figure 2 It is a schematic diagram of the bottleneck structure. Detailed implementation manners

[0025] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] The technical solution adopted by the present invention to solve its technical problems is: first, collect the scattered echo signals and construct a time-frequency feature image set, then build a convolutional neural network, and finally train and optimize the convolutional neural network to achieve ship wake detection. The general block diagram of this method is shown in the accompanying Figure 1 , and the main steps are as follows:

[0027] Step 1: Collect the scattered echo signals with and without wake. In order to improve the generalization of ship wake detection, collect the scattered echo signals with and without wake when the detection platform is at different depths, different speeds, different signal emission angles, and different signal frequencies and forms.

[0028] Step 2: Preprocess the collected scattered echo signals to filter out the noise outside the frequency response range.

[0029] Step 3: Perform wavelet transform on the preprocessed scattered echo signals to obtain a wavelet coefficient matrix, and then perform visualization to obtain a time-frequency feature image set.

[0030] Step 4: Divide the time-frequency feature image set into a test set and a training set according to a certain ratio, and label each time-frequency diagram with the corresponding label.

[0031] Step 5: Considering the detection accuracy, computational complexity and memory occupancy of the network comprehensively, build a convolutional neural network.

[0032] Step 6: Train and optimize the convolutional neural network.

[0033] Step 7: Realize the detection of ship wake targets.

[0034] To enable those skilled in the art to better understand the present invention, the present invention will be described in detail below in conjunction with specific embodiments.

[0035] The basic idea of the present invention is to first construct a time-frequency feature image set, including collecting scattered echo signals with / without wake, preprocessing the echo signals, and extracting the time-frequency feature maps of the preprocessed signals. Then, a convolutional neural network is built. Finally, the convolutional neural network is trained and optimized to classify the time-frequency feature image set, thereby realizing ship wake detection. The specific implementation manners of the above steps are as follows:

[0036] Step 1: A hydrophone collects scattered echo signals x(t), including scattered echo signals with wake and without wake. To improve the generalization of ship wake detection, scattered echo signals with wake and without wake are respectively collected when the detection platform is at different depths, different speeds, different signal emission angles, and different signal emission frequencies and forms.

[0037]

[0038] Where s1(t) is the wake scattered echo, and n(t) is the interference including noise and reverberation.

[0039] Step 2: The collected scattered echo signals are preprocessed to filter out the noise outside the frequency response range. The impulse response function h(t) of the band-pass filter is calculated based on the center frequency of the transmitted signal and the Doppler frequency shift.

[0040] Then the scattered echo signal after passing through the band-pass filter is

[0041]

[0042] Step 3: Perform wavelet transform on the preprocessed scattered echo signal y(t):

[0043]

[0044] Where represents the scale a>0, τ represents the displacement, is the mother wavelet function. In the present invention, db2 is selected as the mother wavelet function. After the signal is wavelet-transformed, a wavelet coefficient matrix is obtained, and then visualization is performed to obtain a time-frequency feature image set.

[0045] Step 4: The time-frequency feature image set is divided into a test set and a training set according to a ratio of 1:5, and each time-frequency diagram is labeled with the corresponding label.

[0046] Step 5: Construct a convolutional neural network. Traditional convolutional neural networks have large memory requirements and large computational amounts, resulting in inability to run on embedded devices. In the present invention, considering the detection accuracy, computational amount, and memory occupancy of the convolutional network, the WakeNet detection network is designed.

[0047] WakeNet consists of three convolutional layers, four bottleneck structures, and a global average pooling layer. The first layer of WakeNet is a convolutional layer with a kernel size of 3×3 to capture macroscopic features. Then, four bottleneck structures are used to obtain deeper features. After that, a convolutional layer with a kernel size of 1×1 is used to improve the non-linear classification performance of the network. Subsequently, a global average pooling layer is adopted to reduce the dimension of the feature matrix. Finally, a convolutional layer with a kernel size of 1×1 is used as a fully connected layer to output the classification result. The network structure parameters are shown in Table 1.

[0048] In Table 1, conv is an ordinary convolutional layer, avgpool is an average pooling layer, and bottleneck is an inverted residual structure, as follows Figure 2 . The expansion factor is the expansion ratio of the depth of the output feature matrix after the first convolutional layer of the bottleneck structure relative to the depth of the input feature matrix.

[0049] Figure 2 The BN layer in it is similar to a data preprocessing operation, mainly used for accelerating the training and convergence of neural networks, and has now become an essential layer in convolutional neural networks. The activation function is Relu6.

[0050] Bottleneck

[0051] Step 6: Train and optimize the convolutional neural network. Adjust the hyperparameters of the network to improve the learning effect and performance of the network. The hyperparameters include: learning rate and batch size, and the values are shown in Table 2.

[0052] Step 7: Implement the detection of ship wake targets. After preprocessing and time-frequency feature extraction of the data to be measured, the optimized neural network model is used to give the detection result, realizing the detection of ship wake targets.

[0053] By verifying this detection algorithm on the simulation dataset, the results show that this algorithm can achieve a detection accuracy of more than 95% when the signal-to-noise ratio is above -24dB. The number of network parameters is 64,386, and the average time to process 1 time-frequency diagram is 24.3ms, indicating that this detection model has good real-time performance.

[0054] Table 1 WakeNet network structure parameters

[0055]

[0056] Table 2 WakeNet hyperparameter values

[0057]

[0058] As described above, it is only the specific implementation manner 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 all be covered within the protection scope of the present invention.

Claims

1. A ship acoustic wake detection method based on a convolutional neural network, characterized in that The steps are as follows: Step 1: Collect the scattered echo signals with and without wake; respectively collect the scattered echo signals with and without wake when the detection platform is at different depths, different speeds, different signal emission angles, and emits different signal frequencies and forms; Step 2: Preprocess the collected scattered echo signals to filter out the noise outside the frequency response range; Step 3: Perform wavelet transform on the preprocessed scattered echo signals to obtain a wavelet coefficient matrix, and then perform visualization to obtain a time-frequency feature image set; Step 4: Divide the time-frequency feature image set into a test set and a training set according to a certain ratio, and label each time-frequency graph with a corresponding label; Step 5: Considering the detection accuracy, computational complexity, and memory occupancy of the network comprehensively, build a convolutional neural network; the convolutional neural network is the WakeNet detection network, which is composed of three ordinary convolutional layers, four bottleneck structures, and a global average pooling layer. The time-frequency graph input to the network first passes through a convolutional layer with a kernel size of 3×3 to capture macroscopic features, then passes through four bottleneck structures to obtain deeper features, then passes through a convolutional layer with a kernel size of 1×1 to improve the non-linear classification performance of the network, then uses a global average pooling layer to reduce the dimension of the feature matrix, and finally uses a convolutional layer with a kernel size of 1×1 as a fully connected layer to output the classification result; Step 6: Use the training set to train and optimize the convolutional neural network; Step 7: After preprocessing and time-frequency feature extraction of the data to be measured, give the detection result through the optimized neural network model to achieve the detection of ship wake targets.

2. The method for detecting ship acoustic wake based on convolutional neural network according to claim 1, wherein: The bottleneck structure described in Step 5 is an inverted residual structure, which is improved from the residual structure in the Resnet network. The bottleneck structure first passes through a convolutional layer with a kernel size of 1×1 to increase the feature dimension, then passes through a 3×3 DW convolution to extract features, and finally passes through a convolutional layer with a kernel size of 1×1 to reduce the feature dimension.

3. The method for detecting ship acoustic wake based on convolutional neural network according to claim 1, characterized in that: In Step 6, the learning effect and performance of the network are improved by adjusting the hyperparameters of the network. The hyperparameters include: learning rate and batch size.

4. A computer system, characterized in that Including: One or more processors, a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in claim 1.

5. A computer-readable storage medium, characterized in that Stored with computer-executable instructions, the instructions are used to implement the method described in claim 1 when executed.

Images