A method for constructing a target classification and recognition model for underwater acoustic target radiated noise

Abstract

This invention proposes a target classification and recognition model construction method for underwater acoustic target radiated noise, belonging to the field of underwater acoustic target recognition. Based on the idea of ​​extracting the Mel-frequency cepstral coefficients of target radiated noise features, this invention presents a one-dimensional convolutional neural network model construction method for underwater acoustic target time-domain waveform data. This method solves two technical problems: First, the input data of this model is the original time-domain signal of the target's radiated noise, eliminating the need for manually designed feature extraction steps and simplifying the complexity of the entire classification system; second, this design method uses the sampling frequency of the radiated noise signal as the network design parameter, allowing for flexible adjustment of the network structure parameters for different frequency signals collected by different devices.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic target recognition and relates to a method for constructing an underwater acoustic target recognition model based on a one-dimensional convolutional neural network. Background Technology

[0002] Underwater acoustic target recognition is a pattern recognition technology that uses the characteristics of target radiated noise to distinguish target attributes, and it has significant economic and military value.

[0003] Traditional underwater acoustic target identification techniques primarily utilize signal analysis and feature engineering theories to extract features of the radiated noise signals from underwater acoustic targets. Based on this feature extraction, identification relies on skilled sonar operators through manual observation or the design of appropriate classifiers. Traditional methods inevitably lose some target feature information when designing classification features for the radiated noise signals, and also require researchers to possess extensive knowledge in the field of signal processing.

[0004] With the rapid development of deep learning technology, deep neural networks, relying on their unique multi-layered structure, can automatically extract features. However, current deep learning-based underwater acoustic target recognition technologies still focus on extracting shallow features from the target's radiated noise signal before inputting them into a deep network for deeper feature extraction and target classification. Existing technical literature 1, "A Method for Underwater Acoustic Target Recognition Based on Joint Neural Networks," Ren Chenxi, Wang Liming, Han Xingcheng, Ye Zefu, Zhu Zhujun, Ship Science and Technology, 2022, 44(1): 136-141, extracts multiple features such as Mel-frequency cepstral coefficients (MFCC) and chromatograms from the radiated noise signal of underwater acoustic targets to form a joint feature vector. Based on this, a joint network composed of a one-dimensional convolutional neural network and an LSTM recurrent neural network is used to realize target classification and recognition. Existing technical literature 2, "A Deep Learning Classification Method for Passive Underwater Acoustic Targets Based on MFCC Features," Yang Lufei, Zhang Xinhua, Wu Bingkun, Li Lanrui, Ship Science and Technology, 2020, 42(10): 129-133, extracts MFCC features from the target signal and designs a corresponding neural network to realize the classification and recognition of underwater acoustic targets. The disadvantage of this type of method is that a manual feature extraction stage is still required before the neural network processes the data, which increases the complexity of the network and loses some original signal information. Summary of the Invention

[0005] The purpose of this invention is to provide a deep neural network model design method for feature extraction and classification of target radiated noise time-domain signals, and to provide the specific structure of the model. This design method solves two technical problems: First, the input data for the model is the original target radiated noise time-domain signal, eliminating the need for manually designed feature extraction steps and simplifying the complexity of the entire classification system; second, the design method uses the sampling frequency of the radiated noise signal as the network design parameter, allowing for flexible adjustment of the network structure parameters for different frequency signals collected by different devices.

[0006] The technical solution of this invention is:

[0007] A method for constructing a target classification and recognition model for underwater acoustic target radiated noise includes the following steps:

[0008] Step 1: Preprocess the target radiated noise signal collected by the sonar equipment, converting it into a fixed-length data format of 3 seconds per frame according to the sampling frequency. The specific method is as follows:

[0009] Assume the target radiated noise signal collected by the sonar equipment is a processed digital sampled signal, denoted as x(n), and the sampling frequency is denoted as f. s The target signal x(n) is divided into frames, with each frame containing 3×f. s With sampling points and a 50% overlap between frames, the target radiated noise dataset X and the corresponding label data Y are obtained:

[0010] X = [x1(n), x2(n), ..., xn] i (n), ..., x p (n)]

[0011] Y=[label1, label2,…, label i , ..., label p ]

[0012] In the formula, p is the total number of frames; x i (n) is the frame signal, in the form of a [3×f s A column vector of dimension 1; the label data Y contains labels. i Set the Arabic numerals corresponding to the target category. After shuffling the target radiation noise dataset X, divide it into a training set X_train and a test set X_test in a 3:1 ratio; ensuring that the order of labels and data remains unchanged, perform the same operation on the label data Y to obtain the training set label Y_train and the test set label Y_test.

[0013] Step 2: Construct a network model for underwater acoustic target classification and recognition based on radiated noise time-domain data. The underwater acoustic target classification and recognition network model (named MFCC1D network model) mainly consists of two parts: a one-dimensional convolutional neural network for feature extraction and a fully connected network for classification and recognition.

[0014] The feature extraction part is constructed using a one-dimensional convolutional neural network model. The network structure design method is inspired by the Mel-Frequency Cepstrum Coefficient (MFCC) features of the target. The core idea of ​​MFCC feature extraction is to use multiple sets of filters mimicking human hearing characteristics to filter the target radiated noise signal, extracting the time-frequency features as the basis for target classification and recognition. The one-dimensional convolutional neural network is a deep learning model that uses convolutional kernels to extract features and classify one-dimensional signals. For the input target radiated noise dataset X, X... (k) Let h represent the input feature vector of the dataset at the k-th convolutional layer, assuming the j-th convolutional kernel is h. j If (n), then the output of the convolution operation of this layer is as follows:

[0015]

[0016] In the formula, M represents the total dimension of the input vector of the network layer, J represents the total number of convolution kernels, and σ(·) represents the nonlinear activation function. When performing one-dimensional convolution on radiated noise signals, the multi-dimensional convolution kernel convolution process is actually a filtering operation on the signal in the frequency domain. Utilizing the correlation inherent in the two methods, and based on the MFCC feature extraction idea, the following one-dimensional convolutional neural network structure for feature extraction is constructed: The one-dimensional convolutional neural network has a total of 5 convolutional layers, and the initial convolution kernel size is set to [10ms×f]. s The size of each subsequent convolutional layer is halved to improve the temporal resolution of the model when processing temporal signals; the kernel stride is set to 2, and the padding method is set to 'same'; the kernel dimension of the first 4 convolutional layers is set to 40, and the dimension of the last convolutional layer is set to 12; after the second and third convolutional layers, a max pooling layer is added, with both the kernel size and stride set to 2, and the ReLU function is used as the activation function between each layer.

[0017] After features are extracted by a one-dimensional convolutional neural network, the resulting feature vectors are concatenated and fed into a fully connected layer for classification and recognition. This fully connected layer employs a two-layer network structure. The first layer's input dimension is the same as the output feature vector dimension of the one-dimensional convolutional neural network, and its output dimension is set to 1024. The second layer's input dimension is set to 1024, and its output dimension is set to the total number of target categories in the dataset.

[0018] Step 3: Input the training set X_train and corresponding label Y_train obtained in Step 1 into the MFCC1D network model constructed in Step 2, determine the batch size and optimizer to train the model, stop network training with recognition accuracy as the evaluation index, and obtain and save the MFCC1D network parameters.

[0019] Step 4: Adjust the MFCC1D network model from the training phase in Step 3 to test mode and load the obtained network parameters. Perform category prediction on the test set X_test obtained in Step 1 and compare it with its true label Y_test to obtain the recognition accuracy of the model on the test set, thus realizing the classification and recognition of underwater acoustic targets for radiated noise signals.

[0020] The beneficial effects of this invention are as follows: The one-dimensional convolutional neural network model designed in this invention can directly identify underwater acoustic targets using time-domain radiated noise data, simplifying the complexity of previous models; furthermore, the model construction process of this invention takes the signal sampling frequency as the starting point, which facilitates model adaptation for different datasets. Attached Figure Description

[0021] Figure 1 This represents a target classification and recognition model framework for underwater acoustic time-domain waveform signals, where the network model is represented by the MFCC1D network.

[0022] Figure 2 This is a schematic diagram of the MFCC1D network structure.

[0023] Figure 3 The graph shows the training of the model under different hyperparameter conditions in the ShipsEar dataset; (a) shows the impact of different batches on model training, and (b) shows the model training under different optimizer conditions. Detailed Implementation

[0024] The following describes the specific implementation process of this invention in detail with reference to the model design method and accompanying drawings.

[0025] A method for constructing a target classification and recognition model for time-domain waveform signals of underwater acoustic target radiated noise includes the following steps:

[0026] Step 1: Preprocess the target radiated noise signal collected by the sonar equipment and convert it into a data format of equal length, with each frame lasting 3 seconds, according to the sampling frequency.

[0027] Taking the ShipsEar dataset as an example, the noise data sampling frequency is 52734Hz, and each frame contains 158202 sampling points. Targets are divided into five categories based on tonnage and type: Category A includes dredgers, tugboats, mussel boats, fishing boats, and trawlers, totaling 1230 frames; Category B includes motorboats, pilot boats, and sailboats, totaling 1016 frames; Category C includes passenger ferries, totaling 2808 frames; Category D includes ocean-going passenger ships and roll-on / roll-off ships, totaling 1623 frames; and Category E represents marine environmental noise, totaling 748 frames. A total of 7425 samples are collected across the five categories, with 5569 samples used as the training set and 1856 samples used as the test set.

[0028] Step 2: Based on the target signal sampling frequency, construct the MFCC1D network model. Taking the ShipsEar dataset as an example, the sampling frequency is 52734Hz. The model's first convolutional layer has a kernel size of 527, a stride of 2, and a dimension of 40; the second convolutional layer has a kernel size of 256, a stride of 2, and a dimension of 40; the first max pooling layer has a kernel size of 2 and a stride of 2; the third convolutional layer has a kernel size of 128, a stride of 2, and a dimension of 40; the second max pooling layer has a kernel size of 2 and a stride of 2; the fourth convolutional layer has a kernel size of 64, a stride of 2, and a dimension of 40; the fifth convolutional layer has a kernel size of 32, a stride of 2, and a dimension of 12; the first fully connected layer has an input dimension of 14268 and an output dimension of 1024; the second fully connected layer has an input dimension of 1024 and an output dimension of 5.

[0029] Step 3: Input the training set and corresponding labels into the MFCC1D network model for network training. This invention uses the ShipsEar dataset for experiments. Figure 3 The training results of the model under different batches and different optimizers are shown. The batch size is determined to be 16, the optimizer is Adam optimizer, and the model parameters are saved after 20 training rounds.

[0030] Step 4: Adjust the model to recognition mode, load the model parameters obtained in step 3, and classify and recognize the test set data to achieve target classification and recognition for radiated noise signals.

Claims

1. A method for constructing a target classification and recognition model for underwater acoustic target radiated noise, characterized in that, The method includes the following steps: Step 1: Preprocess the target radiated noise signal collected by the sonar equipment and convert it into a data format of 3 seconds per frame according to the sampling frequency; Step 2: Construct a network model for underwater acoustic target classification and recognition based on radiated noise time-domain data, namely the MFCC1D network model; The underwater acoustic target classification and recognition network model consists of two main parts: a one-dimensional convolutional neural network for feature extraction and a fully connected network for classification and recognition. The structure of the one-dimensional convolutional neural network for feature extraction is as follows: The one-dimensional convolutional neural network has a total of 5 convolutional layers, and the initial convolutional kernel size is set to... The size of each subsequent convolutional layer is halved to improve the temporal resolution of the model when processing temporal signals; the kernel stride is set to 2, and the padding method is set to 'same'; the kernel dimension of the first 4 convolutional layers is set to 40, and the dimension of the last convolutional layer is set to 12; after the second and third convolutional layers, a max pooling layer is added, with the kernel size and stride both set to 2, and the ReLU function is used as the activation function between each layer; After features are extracted by a one-dimensional convolutional neural network, the resulting feature vectors are concatenated and fed into a fully connected layer for classification and recognition. The fully connected layer for classification and recognition adopts a two-layer network structure. The input dimension of the first layer is the same as the output feature vector dimension of the one-dimensional convolutional neural network, and the output dimension is set to 1024. The input dimension of the second layer is set to 1024, and the output dimension is set to the total number of target categories in the dataset. Step 3: Determine the batch size and optimizer to train the model and obtain the MFCC1D network parameters; Step 4: Use the trained MFCC1D network parameters to build a recognition model to achieve target classification and recognition for radiated noise signals.

2. The method for constructing a target classification and recognition model for underwater acoustic target radiated noise according to claim 1, characterized in that, The specific method for step 1 is as follows: Let the target radiated noise signal collected by the sonar equipment be the processed digital sampled signal, denoted as... The sampling frequency is denoted as ; for target signal Perform frame segmentation, each frame contains 3 With sampling points and a 50% overlap between frames, the target radiated noise dataset is obtained. X With corresponding tag data Y : ； ； In the formula, Total number of frames; It is a framed signal, in the form of a Dimensional column vectors; label data Y middle Set to the Arabic numeral corresponding to the target category; for the target radiated noise dataset X After shuffling the data, it is divided into a training set X_train and a test set X_test in a 3:1 ratio; ensuring that the order of labels and data remains unchanged, the label data... Y Perform the same operation to obtain the training set label Y_train and the test set label Y_test.

3. The method for constructing a target classification and recognition model for underwater acoustic target radiated noise according to claim 1, characterized in that, The specific process of step 3 is as follows: input the training set X_train and the corresponding label Y_train obtained in step 1 into the MFCC1D network model constructed in step 2, determine the batch size and optimizer to train the model, stop network training with recognition accuracy as the evaluation index, and obtain and save the MFCC1D network parameters.

4. The method for constructing a target classification and recognition model for underwater acoustic target radiated noise according to claim 3, characterized in that, The specific process of step 4 is as follows: adjust the MFCC1D network model in the training stage of step 3 to test mode and load the obtained network parameters, perform category prediction on the test set X_test obtained in step 1, and compare it with its true label Y_test to obtain the recognition accuracy of the model on the test set, thereby realizing the underwater acoustic target classification and recognition for radiated noise signals.

Images