A method and system for intelligent identification of ship targets based on underwater acoustic signals

Abstract

The present invention relates to the field of hydroacoustic physics technology, and in particular to a method and system for intelligent identification of ship targets based on hydroacoustic signals. The method comprises: preprocessing the hydroacoustic time domain signal of the sea area to be measured received by the hydrophone to obtain a frequency domain signal; inputting the preprocessed frequency domain signal into a pre-established ship target recognition model, combining with a priori database, to obtain the probability value of the frequency domain signal corresponding to different ship targets; the priori database includes a number of ship targets in the sea area to be measured and the corresponding frequency domain signals, and the ship target recognition model is a neural network with an attention mechanism introduced; the maximum value is selected from the probability value to determine the ship target. The present invention mines effective information of the target from the input features, and outputs an attention distribution map in real time to visualize the features of interest, completes specific decisions such as target detection and target recognition, overcomes the defect of poor interpretability of the neural network to a certain extent, and effectively improves recognition accuracy and credibility.

Description

Technical Field

[0001] The present invention relates to the field of hydroacoustic physics technology, and in particular to a method and system for intelligently identifying ship targets based on hydroacoustic signals. Background Art

[0002] Due to the complexity of the marine environment, underwater acoustic target recognition based on ship radiated noise faces great challenges. However, the performance of traditional feature extraction and classification methods is often limited by unfavorable factors such as environmental complexity and target adversarial nature, making it difficult to achieve higher classification accuracy. Since 2006, deep neural network (DNN) technology has greatly promoted the development of pattern recognition. This method has been applied to underwater acoustic target recognition in recent years. The complex model structure of deep learning methods and their ability to process massive data give them stronger feature representation capabilities and prediction performance in underwater acoustic target recognition. However, its high prediction accuracy is often difficult to explain and is considered unreliable because these black box models prevent users from analyzing the internal workings and behaviors of the model. Since deep learning methods are likely to over-focus on noise that is not related to the target, thereby affecting the model's decision, a network model that can show the internal working mechanism of deep neural networks when identifying underwater acoustic targets has become an urgent need for underwater acoustic recognition research. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects of the prior art and propose a method and system for intelligently identifying ship targets based on hydroacoustic signals.

[0004] In order to achieve the above object, the present invention proposes a method for intelligently identifying ship targets based on water acoustic signals, which is implemented based on a hydrophone set at a fixed position underwater. The method comprises:

[0005] Preprocessing the time domain signal of the underwater sound in the sea area to be tested received by the hydrophone to obtain the frequency domain signal;

[0006] The preprocessed frequency domain signal is input into a pre-established ship target recognition model, and combined with a priori database, the probability value of the frequency domain signal corresponding to different ship targets is obtained; the priori database includes a number of ship targets in the sea area to be tested and the corresponding frequency domain signals, and the ship target recognition model is a neural network that introduces an attention mechanism;

[0007] The maximum value is selected from the probability values ​​to determine the ship target.

[0008] As an improvement of the above method, the preprocessing specifically includes:

[0009] Performing short-time Fourier transform on the time-domain signal of the underwater acoustics in the sea area to be tested received by the hydrophone to obtain a frequency-domain signal;

[0010] According to the significance of the low-frequency characteristics of the ship, frequency domain signals of several frequency points are selected from the preset frequency range.

[0011] As an improvement of the above method, the ship target recognition model is a DNN network with an added attention module, specifically an input layer, an attention module, a fully connected layer, a Drop-out layer, a bottom fully connected layer and an output layer connected in sequence. The input of the ship target recognition model is a preprocessed frequency domain signal, and the output is a classification result.

[0012] As an improvement of the above method, the attention module is used to focus on target characteristics, suppress multi-source interference, and also serve as a dedicated feature extractor; the attention module includes: an attention layer, a Gaussian convolution layer and a fusion layer; wherein,

[0013] The attention layer is used to filter out uninteresting features and retain features that are highly relevant to the target; the attention layer includes a dense layer and an activation layer;

[0014] The Gaussian convolution layer is used to reduce the impact of Doppler frequency shift, measurement error and other random changes in features; it is also used to output attention weight values ​​to achieve network visualization;

[0015] The fusion layer is used to perform a point multiplication operation on the Gaussian convolution layer and the input layer.

[0016] As an improvement of the above method, the specific implementation process of the attention module is as follows:

[0017] Let I s is the input feature vector [I1,I2,…,I S ], S is the total number of feature vectors, after inputting the attention layer, the output attention weight α ts for:

[0018]

[0019] Among them, f(I s ) is the Luong scoring function, which means I s The degree to which each element of is related to the object of attention:

[0020] f(I s )=W·I s

[0021] Among them, W is the attention score matrix, which is obtained through network training and is specifically expressed as:

[0022]

[0023] Then α tsConvolution with the Gaussian kernel function H is mapped to a probability distribution in the range [0,1]. The output G of the Gaussian convolution layer is:

[0024] G=W·α ts

[0025] Where H contains the element h ij , and also serves as the attention weight, processing the attention area image of the input feature in real time during the network operation:

[0026]

[0027] The output vector o of the attention module t for:

[0028]

[0029] As an improvement of the above method, the output of the ship target recognition model also includes real-time output of attention weight values ​​by the attention module to achieve visualization of the DNN network.

[0030] As an improvement of the above method, the method further includes the step of training the ship target recognition model by using supervised learning, specifically including:

[0031] The hydrophones arranged in the sea area to be tested respectively receive the time domain signals corresponding to different ships and the time domain signals of the ocean background noise;

[0032] Preprocess various time domain signals to obtain frequency domain signals, and store them in the prior database according to the source of the ship;

[0033] The frequency domain signal of a certain ship and the frequency domain signal of the ocean background noise are randomly extracted from the prior database to establish a training set;

[0034] Input the training sets into the ship target recognition model respectively;

[0035] The Adam optimization algorithm is used for network training. The initial learning rate is set to 0.001. The cross entropy function is used as the loss function. Dropout regularization is introduced. The probability of neuron activation in each iteration is 90%. The initial weight of the network is generated by the truncated Gaussian distribution model with a standard deviation of 0.1. In the benchmark neural network, the activation function of the fully connected layer uses the ReLU function. The number of network iterations is set to K times, and a trained ship target recognition model is obtained.

[0036] A ship target intelligent recognition system based on underwater acoustic signals, the system comprises: a hydrophone arranged at a fixed position underwater, a trained ship target recognition model, a preprocessing module, a recognition module and an output module; wherein,

[0037] The preprocessing module is used to preprocess the time domain signal of the underwater acoustic wave in the sea area to be measured received by the hydrophone to obtain a frequency domain signal;

[0038] The recognition module is used to input the preprocessed frequency domain signal into a pre-established ship target recognition model, and obtain the probability value of the frequency domain signal corresponding to different ship targets in combination with the prior database; the prior database includes a number of ship targets in the sea area to be tested and the corresponding frequency domain signals, and the ship target recognition model is a neural network that introduces an attention mechanism;

[0039] The output module is used to select the maximum value from the probability values ​​to determine the ship target.

[0040] Compared with the prior art, the advantages of the present invention are:

[0041] 1. The present invention proposes a method of introducing an attention mechanism into a neural network, mining effective information of a target from input features, and outputting an attention distribution map in real time to visualize the features of interest, and completing specific decisions such as target detection and target recognition, which overcomes the defect of poor interpretability of neural networks to a certain extent and can effectively improve recognition accuracy and credibility;

[0042] 2. The method of the present invention can intuitively demonstrate the process of gradually focusing on the spectral characteristics of the target ship and suppressing the noise of the marine environment and the interference ship, achieving higher recognition accuracy than the traditional method; and the method has the potential to achieve multi-target resolution when only single target data is used;

[0043] 3. The method proposed in the present invention focuses on target characteristics, suppresses multi-source interference, focuses on feature visualization, and can achieve multi-target resolution using only single target data; it can be used as a dedicated feature extractor. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is the network structure of the ship target recognition model of the present invention;

[0045] FIG2( a ) is a sea chart of the experimental area of ​​the simulation example of the present invention;

[0046] FIG2( b ) is a radar / AIS diagram of the experimental area of ​​the simulation example of the present invention;

[0047] FIG3( a) is a time-frequency diagram of a first type of signal of a simulation example of the present invention;

[0048] FIG3( b ) is a time-frequency diagram of the second type of signal of the simulation example of the present invention;

[0049] FIG3( c ) is a time-frequency diagram of the third type of signal of the simulation example of the present invention;

[0050] FIG3( d ) is a time-frequency diagram of the fourth type of signal of the simulation example of the present invention;

[0051] FIG4( a ) is a time-frequency diagram of the ABNN-1 radiation noise signal of the detection ship A in the simulation example of the present invention;

[0052] FIG4( b ) is a time-frequency diagram of the ABNN-2 radiation noise signal of the detection ship B according to the simulation example of the present invention;

[0053] FIG4( c ) is an attention distribution diagram of the ABNN-1 radiation noise signal of the detection ship A in the simulation example of the present invention at the 10th, 20th, 30th, and 50th iterations;

[0054] FIG4( d ) is an attention distribution diagram of the ABNN-2 radiation noise signal of the detection ship B in the simulation example of the present invention at the 10th, 20th, 30th, and 50th iterations;

[0055] FIG5( a ) is a signal spectrum of a simulation example of the present invention on the first and second types of data and an attention distribution diagram output by a binary classification ABNN at the 10th, 50th, 100th, and 20,000th iterations;

[0056] FIG5( b ) is a diagram showing the cross entropy loss in network training according to Embodiment 1 of the present invention;

[0057] FIG5(c) is the accuracy of network training in Example 1 of the present invention;

[0058] FIG6(a) is a diagram showing the distribution of the first two element values ​​of the output node of the third type of signal (including the radiation noise of ships A and B) in the simulation example of Embodiment 1 of the present invention;

[0059] FIG6( b ) is an attention distribution diagram of the third type of signal output of the simulation example of Embodiment 1 of the present invention;

[0060] Figure 7 This is a comparison of the prediction accuracy between this method and the traditional method. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0062] Example 1

[0063] Embodiment 1 of the present invention proposes a method for intelligently identifying ship targets based on underwater acoustic signals, which adopts the following technical solutions:

[0064] The process of the ship target intelligent recognition method based on hydroacoustic signals can be divided into three main steps: data preprocessing and labeling, network model construction and training, measured data preprocessing and neural network output testing.

[0065] Step 1: Preprocess the experimental data. Perform short-time Fourier transform on the acoustic time domain signal received by the sonar to obtain its time-frequency expression. According to the significance of the low-frequency characteristics of the ship, 256 frequency points in the frequency range of 10-100Hz are taken as the dimension of the neural network input feature. According to the experimental records, all samples are divided into different categories of signals, such as four categories: only ship A, only ship B, coexistence of ships A and B, and only ocean background noise.

[0066] Step 2: Network model construction and training. In target recognition, the signal spectrum is directly input into the neural network input layer for training. The neural network goes through a series of complex nonlinear mappings and finally outputs the classification results according to the specific task. Figure 1 In the DNN network structure shown in , the input features are first weighted by a fully connected layer and a Softmax activation function to control the sensitivity to different frequency components. Next, a convolution layer with a Gaussian convolution kernel is introduced after the attention layer. This is because the attention module may pay too much attention to certain feature points in the feature space, and the radiation noise spectrum of surface ships, underwater vehicles, etc. contains both discrete narrowband line spectra and continuous broadband spectra. At the same time, this layer reduces the impact of Doppler shift, measurement errors, and other random changes in features. The attention layer acts as a masking layer to filter out uninteresting features and retain features that are highly correlated with the target. Finally, the attention module is connected to the underlying fully connected network for classification, and the attention weight value is output to achieve network visualization.

[0067] To build the attention module, let I s is the input feature vector [I1,I2,…,I S ], S is the total number of feature vectors, that is, the number of sampled frequency points. The feature expression used for classification is composed of the weighted sum of these feature vectors, and the attention network needs to learn this set of attention weights. The attention mechanism can be expressed as:

[0068] f(I s )=W·I s

[0069]

[0070] Where W and are the attention score matrices, which need to be obtained through network training, and f is the Luong score function, representing h s The degree of relevance of each element of with the object of attention, the more relevant, the greater the score. ts is the attention weight. Then the above result is convolved with the Gaussian kernel function and mapped to a probability distribution in the range [0,1], which is expressed by matrix operation as:

[0071]

[0072]

[0073] G=W·α ts

[0074] where H contains the element h ij , W is the Gaussian weight matrix, G is the output of the Gaussian convolution layer, and also serves as the attention weight, which is processed in real time as the attention area image of the input feature during the network operation. The output vector of the attention module is:

[0075]

[0076] During the network training process, the network gradually focuses on the features that are closely related to the training target, and the training loss gradually decreases. The attention layer acts as a mask to filter out uninteresting features and retain features that are highly relevant to the target.

[0077] The neural network in this paper adopts some commonly used network optimization strategies: due to the multi-path interference of environmental noise, reverberation and underwater acoustic channels, the target cost function may be a non-convex function, and it is difficult to jump out of the local optimal value. Therefore, the neural network trains the network parameters through the Adam optimizer to prevent falling into the local optimal value and over-learning. In addition, due to the sensitivity of the model to environmental changes and the distortion of underwater acoustic signals, DNN is prone to overfitting problems. Therefore, Dropout regularization is used to make the network parameters more sparse.

[0078] Step 3: Network testing. The measured data is processed by FFT to obtain a frequency data set. After training, the test set data is input into the trained DNN model to output the predicted results. The output label is a sparse label, each element of which represents the probability of a certain class, and the value range is [0,1]. The maximum value of these elements is used to determine the category.

[0079] The effects of the present invention include the following aspects:

[0080] (1) The attention model can find the most effective information to optimize model performance from the cluttered features and suppress irrelevant environmental noise and interference targets; (2) The network can output feature probability maps during operation to explain the features that the neural network pays attention to and the basis for prediction, and can also serve as an extractor of target features; (3) Compared with traditional DNNs, it improves accuracy and interpretability.

[0081] Simulation example:

[0082] In September 2020, a ship radiated noise measurement experiment was carried out in the northern part of the South China Sea. The topographic map of the experimental sea area is shown in Figure 2(a). According to the experimental plan, the two experimental ships sailed or moored at endpoints A, B and C, D along trajectory 1 and trajectory 2 respectively. The vector hydrophone used to receive the signal is located at a depth of 80m underwater at point O, and the frequency response range of the hydrophone is 1-28kH. According to the radar AIS display in Figure 2(b), there are more than 17 fishing boats within 30km of the experimental sea area, which can verify whether the attention network can suppress multi-source interference and focus on the target itself.

[0083] The sampling frequency of the hydrophone is 12kHz, and a portion of the near-field data with a relatively high signal-to-noise ratio is subjected to short-time Fourier transform to obtain its time-frequency expression. The number of FFT points is 32768 points, and the time length of each frame signal is 1 second. The amplitudes of 16385 frequency points of 1637 frames of signals are obtained, and the frequency range is 1-6000Hz. Due to the significance of the low-frequency characteristics of the ship, the frequency range is reduced to 10-100Hz, and 256 frequency points are taken as the dimension of the neural network input feature.

[0084] According to the experimental records, all samples are divided into four types of signals, and the division basis is shown in Table 1. Among them, the first type of signal contains 343 samples, the second type of signal contains 372 samples, the third type of signal contains 113 samples, and the fourth type of signal contains 426 samples. The time-frequency diagrams of the four types of signals are shown in Figure 3 (a), (b), (c), and (d).

[0085] Figure 3 shows that the target is in a complex environment with multiple sources of interference and environmental noise, and the original signal contains a large number of electrical noise line spectra caused by problems with the hydrophone equipment. Ship A and ship B each have a relatively significant low-frequency line spectrum (f A =53.44Hz,f B =29.12Hz), such as Figure 3(a) , 3(b) shown.

[0086] Table 1 Data division and line spectrum components

[0087]

[0088]

[0089] The Adam optimization algorithm is introduced into the neural network for network training. The initial learning rate is 0.001, and the other parameters follow the values ​​provided above. The cross entropy function is used as the loss function. Dropout regularization is introduced, and the probability of neuron activation is 90% for each iteration. The initial weight of the network is generated by a truncated Gaussian distribution model with a standard deviation of 0.1. In the benchmark neural network, the activation function of the fully connected layer uses the ReLU function. The output layer contains 4 Softmax nodes, representing the unique hot codes of the four categories. The number of network iterations is set to 20,000. During network training, 80% of the samples in each category are randomly selected as the training set, and the rest are the test set.

[0090] 1. Object Detection

[0091] In order to use ABNN to extract features from ship noise, two target detection networks were trained to determine whether there is self-radiated noise from ship A and ship B. The output label is a sparse label, each element of which represents the probability of a certain class, and the value range is [0,1]. The maximum value of these elements is used to determine the category. After 20,000 iterations, the accuracy of detecting presence or absence is 98.04% and 97.44%.

[0092] During the network training process, the vector of the attention weight layer of the attention module is recorded every 500 iterations, and the network attention distribution diagrams for the 10th, 20th, 30th, and 50th iterations are shown in Figure 4. The results show that as the number of iterations increases, the neural network gradually focuses on the specific line spectrum features of ship A and ship B (f A =53.44Hz,f B =29.12Hz), while other spectral features listed in Table I are assigned lower weights.

[0093] The attention map in Figure 4 explains the frequency components that may be used by the algorithm for decision making. Figure 4(a) is the time-frequency diagram of the ABNN-1 radiation noise signal of the detection ship A; Figure 4(b) is the time-frequency diagram of the ABNN-2 radiation noise signal of the detection ship B; Figure 4(c) is the attention distribution diagram of the ABNN-1 radiation noise signal of the detection ship A at the 10th, 20th, 30th, and 50th iterations; Figure 4(d) is the attention distribution diagram of the ABNN-2 radiation noise signal of the detection ship B at the 10th, 20th, 30th, and 50th iterations; This shows that the attention module can be used specifically as a feature extractor.

[0094] 2. Target Classification

[0095] The binary classification ABNN for ship A and ship B is trained. Its output layer contains two Softmax nodes, corresponding to the sparse labels of class I and class II. The training data set is randomly shuffled during the training process. Samples with index values ​​less than 343 are samples of ship A, and samples greater than 343 are samples of ship B, which are divided by the dotted line in Figure 5(a). The accuracy of the binary classification of ABNN is 96.56%. From the test set loss curve in Figure 5(b), it can be seen that the neural network does not overfit.

[0096] Figure 5(a) shows the attention distribution diagram for each sample at the 10th, 50th, 100th, and 20,000th iterations. The attention distribution diagram shows in real time that the network is mainly focused on ship A (f A =54.57Hz) and B ship (f B =29.30Hz), while the other frequencies given in Table I have almost negligible impact on the decision. These frequencies are also consistent with the frequencies extracted by ABNN-1 and ABNN-2 in Part B, as shown in Figure 5. The results also show that the model converges quickly and focuses on the target features in only 100 iterations, while suppressing most of the interference. This process is synchronized with the reduction of network loss and the improvement of accuracy, as shown in Figure 5(b) and (c).

[0097] For Class III data (including the noise of ship A and ship B), we only use single target and environmental noise data to train the model. The results show that for 100% of the samples, the third element value of the output node (corresponding to the probability of the environmental noise class) is less than 1e-5, indicating that ABNN believes that there are targets in all samples. In addition, in 74.34% of the samples, the first two elements of the output node (corresponding to the probabilities of class A and class B ships) are greater than 0.8, indicating that ABNN believes that the two targets are likely to coexist in these samples. The distribution of the detection results of each sample is shown in Figure 6(a). All samples are concentrated in the upper triangular part of the figure, and the proportion of samples increases as the first two elements of the output node approach (1,1). The output of the attention module shown in Figure 6(b) shows that the DNN focuses on the union of class 1 and class 2 at 20,000 iterations. This result shows that although the network is trained using single-target data, it has the potential to achieve multi-target resolution in the current dataset.

[0098] 3. Performance comparison

[0099] To summarize the above results, the detection accuracy of ship A and ship B is 98.04% and 97.44% respectively. The binary classification accuracy of ship A and ship B is 96.56%, and the four-classification accuracy in Table 2 is 96.56%. In order to compare with the traditional DNN, we removed the attention module and retested with the traditional neural network. The results show that the detection accuracy of ship A and the classification accuracy of the two ships of the proposed model are slightly higher than those of the traditional DNN model (0.92% and 0.70%) as shown in Table 2 and Figure 7 This shows that the improvement of the attention module on accuracy is limited. However, the accuracy of ABNN on multi-object resolution is 74.34%, which is 15.99% higher than the 58.35% of the traditional DNN. Therefore, ABNN has better potential for multi-object resolution when there are only a few single-object training samples.

[0100] Table 2 Method comparison

[0101] This method model Traditional Model Ship A detection 98.04％ 97.12％ Ship B detection 97.44％ 97.64％ Second ship classification 96.56％ 96.86％ Multi-target resolution 74.34％ 58.35％

[0102] In addition, the attention weight map output by the attention module can be used to explain the prediction results of the neural network, which is conducive to auxiliary decision-making and credibility analysis. When the attention area of ​​the neural network shifts to a frequency range beyond the prior knowledge, the user can consider that the prediction result is unreliable. This is a feature that traditional DNN does not have.

[0103] Example 2

[0104] The system includes: a hydrophone set at a fixed position underwater, a trained ship target recognition model, a preprocessing module, a recognition module and an output module; wherein,

[0105] The preprocessing module is used to preprocess the time domain signal of the underwater acoustic wave in the sea area to be measured received by the hydrophone to obtain a frequency domain signal;

[0106] The recognition module is used to input the preprocessed frequency domain signal into a pre-established ship target recognition model, and obtain the probability value of the frequency domain signal corresponding to different ship targets in combination with the prior database; the prior database includes a number of ship targets in the sea area to be tested and the corresponding frequency domain signals, and the ship target recognition model is a neural network that introduces an attention mechanism;

[0107] The output module is used to select the maximum value from the probability values ​​to determine the ship target.

[0108] The specific processing method is the same as that in Example 1.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A method for intelligently identifying ship targets based on underwater acoustic signals, which is implemented based on a hydrophone set at a fixed position underwater, and the method comprises: Preprocessing the time domain signal of the underwater sound in the sea area to be tested received by the hydrophone to obtain the frequency domain signal; The preprocessed frequency domain signal is input into a pre-established ship target recognition model, and combined with a priori database, the probability value of the frequency domain signal corresponding to different ship targets is obtained; the priori database includes a number of ship targets in the sea area to be tested and the corresponding frequency domain signals, and the ship target recognition model is a neural network that introduces an attention mechanism; The maximum value is selected from the probability values ​​to determine the ship target; The ship target recognition model is a DNN network with an added attention module, specifically an input layer, an attention module, a fully connected layer, a Drop-out layer, a bottom fully connected layer and an output layer connected in sequence. The input of the ship target recognition model is a preprocessed frequency domain signal, and the output is a classification result; The attention module is used to focus on target characteristics, suppress multi-source interference, and is also used as a dedicated feature extractor; the attention module includes: an attention layer, a Gaussian convolution layer and a fusion layer; wherein, The attention layer is used to filter out uninteresting features and retain features that are highly relevant to the target; the attention layer includes a dense layer and an activation layer; The Gaussian convolution layer is used to reduce the impact of Doppler frequency shift, measurement error and other random changes in features; it is also used to output attention weight values ​​to achieve network visualization; The fusion layer is used to perform a point multiplication operation on the Gaussian convolution layer and the input layer; The specific implementation process of the attention module is: Let I s is the input feature vector [I1,I2,…,I S ], S is the total number of feature vectors, after inputting the attention layer, the output attention weight α ts for: Among them, f(I s ) is the Luong scoring function, which means I s The degree to which each element of is related to the object of attention: f(I s )=W·I s Among them, W is the attention score matrix, which is obtained through network training and is specifically expressed as: Then α ts Convolution with the Gaussian kernel function H is mapped to a probability distribution in the range [0,1]. The output G of the Gaussian convolution layer is: G=W·a ts Where H contains the element h ij , and also serves as the attention weight, processing the attention area image of the input feature in real time during the network operation: h ij =j-i,i,j∈[0,S] The output vector o of the attention module t for:

2. The method for intelligently identifying ship targets based on underwater acoustic signals according to claim 1 is characterized in that: The pre-processing specifically includes: Performing short-time Fourier transform on the time-domain signal of the underwater acoustics in the sea area to be tested received by the hydrophone to obtain a frequency-domain signal; According to the significance of the low-frequency characteristics of the ship, frequency domain signals of several frequency points are selected from the preset frequency range.

3. The method for intelligently identifying ship targets based on underwater acoustic signals according to claim 1 is characterized in that: The output of the ship target recognition model also includes real-time output of attention weight values ​​by the attention module, which is used to realize the visualization of the DNN network.

4. The method for intelligently identifying ship targets based on underwater acoustic signals according to claim 3 is characterized in that: The method further comprises the step of training the ship target recognition model by using supervised learning, specifically comprising: The hydrophones arranged in the sea area to be tested respectively receive the time domain signals corresponding to different ships and the time domain signals of the ocean background noise; Preprocess various time domain signals to obtain frequency domain signals, and store them in the prior database according to the source of the ship; The frequency domain signal of a certain ship and the frequency domain signal of the ocean background noise are randomly extracted from the prior database to establish a training set; Input the training sets into the ship target recognition model respectively; The Adam optimization algorithm is used for network training. The initial learning rate is set to 0.

001. The cross entropy function is used as the loss function. Dropout regularization is introduced. The probability of neuron activation in each iteration is 90%. The initial weight of the network is generated by the truncated Gaussian distribution model with a standard deviation of 0.

1. In the benchmark neural network, the activation function of the fully connected layer uses the ReLU function. The number of network iterations is set to K times to obtain a trained ship target recognition model.

5. A system based on the method for intelligent identification of ship targets based on hydroacoustic signals according to claim 1, characterized in that: The system comprises: a hydrophone arranged at a fixed position underwater, a trained ship target recognition model, a preprocessing module, a recognition module and an output module; wherein, The preprocessing module is used to preprocess the time domain signal of the underwater acoustic wave in the sea area to be measured received by the hydrophone to obtain a frequency domain signal; The recognition module is used to input the preprocessed frequency domain signal into a pre-established ship target recognition model, and obtain the probability value of the frequency domain signal corresponding to different ship targets in combination with the prior database; the prior database includes a number of ship targets in the sea area to be tested and the corresponding frequency domain signals, and the ship target recognition model is a neural network that introduces an attention mechanism; The output module is used to select the maximum value from the probability values ​​to determine the ship target.

Images