Underwater weak target intelligent detection method based on time-frequency feature fusion

Abstract

The invention discloses an underwater weak target intelligent detection method based on time-frequency feature fusion. The method comprises the following steps: firstly, obtaining underwater acoustic signal data, then obtaining modal decomposition signals of signal time domain information and frequency domain information corresponding to original sonar signal data through variational modal decomposition, and obtaining a time-frequency domain modal feature data set after combination; and then, a deep neural network based on a cross attention mechanism is constructed, the deep neural network comprises a convolution module, a bidirectional gating circulation unit module, a cross attention mechanism module, a full connection layer and a classifier which are connected in sequence, and after the deep neural network is trained, an underwater weak target detection network is obtained. Finally, the high-accuracy, high-efficiency and high-generalization detection effect on the input signal is realized. The method can effectively detect single-frequency signals and broadband signals with weak target intensity in the ocean, and has high detection probability and good generalization performance.

Description

Technical Field

[0001] The invention relates to an underwater target intelligent detection method in the field of underwater acoustic targets, and in particular to an underwater weak target intelligent detection method based on time-frequency feature fusion. Background Art

[0002] Sonar systems are widely used in the detection and identification of underwater targets. Sonar can be divided into active sonar and passive sonar according to the working mode. Among them, passive sonar receives the noise emitted by far-field targets against the background of the ship's noise, and the target noise becomes very weak after long-distance propagation, so passive sonar often works under low signal-to-noise ratio conditions. Improving the detection capability of weak underwater signals has important application value in marine resource development, environmental monitoring and military fields.

[0003] The complexity and variability of the underwater environment often makes it difficult to meet actual needs. It is urgent to study and develop new methods for underwater target detection of hydroacoustic targets. The research focus of weak signal detection is on how to extract the signal to be processed in an environment with strong background noise. Therefore, the characteristics of the measured signal are often studied to detect weak targets masked by background noise. The rapid development of feature extraction technology based on deep neural networks has provided a series of new perspectives for underwater weak target detection. It can learn and extract effective features from underwater signal data, thereby achieving more accurate weak target detection. Feature extraction plays a key role in underwater weak target detection, but it also faces some problems and challenges.

[0004] There are many methods for target feature extraction, which vary according to the classification object and are highly targeted. Underwater target noise is complex and changeable, and coupled with the interference of the ocean noise environment, it is a difficult problem to obtain better target features. Therefore, it is necessary to develop a weak target detection method that is adapted to the background information of the actual ocean sound field environment and can effectively extract useful information from the features. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention provides an underwater weak target intelligent detection method based on time-frequency feature fusion. The present invention adopts a variational mode decomposition method and a deep neural network optimized based on a cross-attention mechanism to mine key recognition features with high discrimination and good generalization in the target samples to be detected, thereby achieving high-performance and high-generalization underwater acoustic weak target detection.

[0006] The purpose of the present invention is achieved through the following technical solutions.

[0007] 1. An intelligent underwater weak target detection method based on time-frequency feature fusion

[0008] Step 1: Obtain historical data of underwater sonar signals, extract the time-frequency domain fusion modal features corresponding to all historical underwater sonar signals, and thus obtain a time-frequency domain fusion feature data set;

[0009] Step 2: Construct a deep neural network based on the cross-attention mechanism, and use the time-frequency domain fusion feature dataset to train the deep neural network based on the cross-attention mechanism to obtain an underwater weak target detection network;

[0010] Step 3: Extract the time-frequency domain fusion modal features of the underwater signal to be detected and input them into the underwater weak target detection network to obtain the underwater acoustic target detection results.

[0011] In the step 1, the extraction process of the time-frequency domain fusion modal features of each historical underwater sonar signal is specifically as follows:

[0012] After performing variational modal decomposition in the time domain and frequency domain on each historical underwater sonar signal, the time domain modal decomposition signal and the frequency domain modal decomposition signal are obtained. After feature fusion of the time domain modal decomposition signal and the frequency domain modal decomposition signal, the time-frequency domain fusion modal feature is obtained.

[0013] In the step 2, the deep neural network based on the cross-attention mechanism includes a convolution module, a bidirectional gated recurrent unit module, a cross-attention mechanism layer, a fully connected layer and a classifier connected in sequence.

[0014] The convolution module includes a plurality of sequentially connected convolution blocks, and each convolution block is composed of a sequentially connected convolution layer, an activation function layer and a pooling layer.

[0015] 2. A computer device

[0016] The device comprises a memory and a processor, wherein the memory stores a computer program, and the steps of the method are implemented when the processor executes the computer program.

[0017] 3. A computer-readable storage medium

[0018] The medium stores a computer program, which implements the steps of the method when executed by a processor.

[0019] IV. A COMPUTER PROGRAM PRODUCT

[0020] The product comprises a computer program / instructions which, when executed by a processor, implement the steps of the method.

[0021] The beneficial effects of the present invention are:

[0022] 1. The present invention introduces variational mode decomposition to fuse time domain information and frequency domain information, which can accurately realize the effective decomposition and reconstruction of the original signal and effectively extract the local features in the signal. Through the feature fusion of time domain and frequency domain information, it can better understand the inherent structure and laws of the data, better utilize the data set and improve the generalization performance of the model, thereby improving the ability of the deep model to extract key features in the underwater acoustic signal.

[0023] 2. The present invention uses a cross-attention mechanism layer to calculate the attention weights between output sequences and then perform modeling and association. It can deeply fuse the feature information between different modalities, improve the information processing capabilities of different modalities, and thus effectively improve the detection performance of the original underwater acoustic signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0025] Figure 1 This is a flow chart of the underwater weak target intelligent detection method based on time-frequency feature fusion proposed by the present invention.

[0026] Figure 2 It is a schematic diagram of the process of extracting features by variational mode decomposition of the present invention.

[0027] Figure 3 Schematic diagram of the network structure of the deep neural network based on the cross-attention mechanism proposed in the present invention.

[0028] Figure 4 It is a structural schematic diagram of the convolution module of the present invention.

[0029] Figure 5 It is a schematic diagram of the structure of the bidirectional gated cycle unit module of the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the implementation of the present invention clearer, the technical solution in the embodiment of the present invention will be described in more detail below with reference to the accompanying drawings in the embodiment of the present invention.

[0031] The present invention provides an underwater weak target intelligent detection method based on time-frequency feature fusion. Figure 1 As shown, the present invention comprises the following steps:

[0032] Step 1: Obtain historical data of underwater sonar signals, extract the time-frequency domain fusion modal features corresponding to all historical underwater sonar signals, and thus obtain a time-frequency domain fusion feature data set;

[0033] The underwater sonar signal historical data is a data set obtained by using the public ocean noise data set and the underwater audio data received by the sonar system, through data preprocessing methods such as standardization, bandpass filtering and resampling, and by adding fixed-frequency single-frequency signals, random-frequency single-frequency signals and broadband signals with different signal strengths. It includes the corresponding label file, that is, the necessary target annotation information;

[0034] The specific process of extracting the time-frequency domain fusion modal features of each historical underwater sonar signal is as follows:

[0035] After performing variational modal decomposition in the time domain and frequency domain on each historical underwater sonar signal, the time domain modal decomposition signal and the frequency domain modal decomposition signal are obtained. After feature fusion of the time domain modal decomposition signal and the frequency domain modal decomposition signal, the time-frequency domain fusion modal feature is obtained. Figure 2 The schematic diagram of the process of fusing time domain and frequency domain features of underwater acoustic signals through variational mode decomposition is shown.

[0036] In step 1, the constrained variational expression of variational mode decomposition is:

[0037]

[0038] The constraints are:

[0039]

[0040] Among them, u k (t)={u1,u 2, …,u k} are k intrinsic mode functions (IMFs) at time t. IMFs represent the oscillation of different frequency components in the signal. The frequency of each IMF may change over time; ω k is the center frequency of each mode; is the partial derivative of the function with respect to time t; δ(t) is the impulse function; is the estimated center frequency; represents the square of the L2 norm; f(t) is the sum of all modal components.

[0041] To facilitate the solution, it is converted into an unconstrained optimization expression through the following formula:

[0042]

[0043] Among them, L() is the augmented Lagrange expression, λ(t) is the Lagrange multiplication factor, and α is the balance parameter.

[0044]

[0045] The Lagrange multiplication factor is updated according to the above formula. is the n+1th iteration form of the Fourier transform of the Lagrange multiplication factor λ(t); is the nth iteration form of the Fourier transform of the Lagrange multiplication factor λ(t); τ() is the noise tolerance; is the Fourier transform of the sum of all modal components f(t); The current remaining amount Wiener filter; ω is the frequency; n is the number of iterations.

[0046] Finally, solve the Wiener filter By inverse Fourier transform of , we can get the various modal components obtained after variational mode decomposition.

[0047] 1.3) The extracted time-frequency domain fusion modal features and the corresponding label files are used together to generate a time-frequency domain fusion modal feature dataset.

[0048] Step 2: Construct a deep neural network based on the cross-attention mechanism, and use the time-frequency domain fusion feature dataset to train the deep neural network based on the cross-attention mechanism to obtain an underwater weak target detection network; Figure 3 As shown in the figure, the deep neural network based on the cross-attention mechanism includes a convolution module, a bidirectional gated recurrent unit module, a cross-attention mechanism layer, a fully connected layer and a classifier connected in sequence. The convolution module and the bidirectional gated recurrent unit module are used to learn the key feature information description of the underwater acoustic signal; the cross-attention mechanism calculates the attention weight and converts the vector of the high-dimensional space into the representation of the low-dimensional space; the fully connected layer and the classifier obtain the final classification detection result. Figure 4 As shown, the convolution module includes multiple sequentially connected convolution blocks, and each convolution block consists of a sequentially connected convolution layer, an activation function layer, and a pooling layer. Figure 5 The network structure of the bidirectional gated recurrent unit module is shown.

[0049] Among them, the cross attention mechanism module is based on the calculation of query Q, key K and value V. The query Q comes from one input sequence, the key K and value V come from another input sequence, and the calculation mechanism is as follows.

[0050] First, calculate the similarity between query Q and key K: First, do a dot product between query Q and key K to get the correlation score between the two input sequences. The specific formula is as follows:

[0051]

[0052] Among them, Attention(Q,K,V) represents the output representation after cross attention adjustment; SoftMax() represents the SoftMax classifier; QK T is the dot product of the query Q and the key K, indicating the similarity of the two sequences at different positions; d k is the dimension of the key, and is a scaling factor used to scale the dot product to avoid excessive values.

[0053] Secondly, the attention weight is calculated: these similarities are converted into probability distributions through the SoftMax function, which represents the attention weight of the query for each key.

[0054] Finally, a weighted sum is performed: these attention weights are applied to the values, and the output vector is finally obtained. This is equivalent to extracting the attention information from the value sequence and inputting it into the next network layer.

[0055] Step 3: Extract the time-frequency domain fusion modal features of the underwater signal to be detected and input them into the underwater weak target detection network to obtain the underwater acoustic target detection results.

[0056] The above steps describe in detail how to start from acquiring underwater acoustic signal data, establish a feature data set that fuses time domain information and frequency domain information through variational mode decomposition, and then build and train a deep neural network based on the cross-attention mechanism, and finally realize the detection of underwater acoustic targets. This process involves multiple links such as data processing, feature extraction, network design optimization and model training, and each link is to improve the detection probability and efficiency of underwater acoustic targets.

[0057] Finally, it should be noted that the above embodiments and explanations are only used to illustrate the technical solution of the present invention rather than to limit it. Those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope disclosed in the technical solution of the present invention, which should be included in the scope of protection of the claims of the present invention.

Claims

1. An intelligent underwater weak target detection method based on time-frequency feature fusion, characterized in that: The steps include: Step 1: Obtain historical data of underwater sonar signals, extract the time-frequency domain fusion modal features corresponding to all historical underwater sonar signals, and thus obtain a time-frequency domain fusion feature data set; Step 2: Construct a deep neural network based on the cross-attention mechanism, and use the time-frequency domain fusion feature dataset to train the deep neural network based on the cross-attention mechanism to obtain an underwater weak target detection network; Step 3: Extract the time-frequency domain fusion modal features of the underwater signal to be detected and input them into the underwater weak target detection network to obtain the underwater acoustic target detection results.

2. According to claim 1, the method for intelligent detection of underwater weak targets based on time-frequency feature fusion is characterized in that: In the step 1, the extraction process of the time-frequency domain fusion modal features of each historical underwater sonar signal is specifically as follows: After performing variational modal decomposition in the time domain and frequency domain on each historical underwater sonar signal, the time domain modal decomposition signal and the frequency domain modal decomposition signal are obtained. After feature fusion of the time domain modal decomposition signal and the frequency domain modal decomposition signal, the time-frequency domain fusion modal feature is obtained.

3. According to claim 1, the method for intelligent underwater weak target detection based on time-frequency feature fusion is characterized in that: In the step 2, the deep neural network based on the cross-attention mechanism includes a convolution module, a bidirectional gated recurrent unit module, a cross-attention mechanism layer, a fully connected layer and a classifier connected in sequence.

4. According to claim 3, the method for intelligent detection of underwater weak targets based on time-frequency feature fusion is characterized in that: The convolution module includes a plurality of sequentially connected convolution blocks, and each convolution block is composed of a sequentially connected convolution layer, an activation function layer and a pooling layer.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

7. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images