Underwater target identification method based on Conformer model

Abstract

The invention relates to an underwater target recognition method based on a Conformer model, and belongs to the technical field of underwater target recognition. According to the collected and processed data, constructing a model based on a conformer, and carrying out training; and inputting the data processed in the step 1 into the trained model based on the conformer to obtain a denoised clean underwater sound audio. According to the method, the conformer is applied to underwater acoustic target recognition, the global dependence modeling capability of the Transformer and the local structure capturing capability of convolution are fused, the problem of long-time dependence and local structure coexistence of underwater acoustic signals can be well solved, and therefore accurate target classification is achieved.

Description

Technical Field

[0001] This invention belongs to the field of underwater target recognition technology, and in particular to an underwater target recognition method based on the Conformer model. Background Technology

[0002] With the development of signal processing theory, microprocessor technology, and information theory, underwater target recognition is becoming increasingly rapid and urgent. This has greatly stimulated the development of underwater target recognition technology, but also increased the difficulty of the task. In real-world environments, underwater sound recognition differs significantly from ordinary speech recognition. The sound of an underwater target is not produced by a single sound source, but rather is the complex physical result of the interaction between the sound source and the surrounding water medium. The characteristics of this sound signal are reflected in its extremely complex composition, wide frequency and power range, numerous noise components, enormous information content, and the difficulty in identifying the target signal. These characteristics are generally accompanied by nonlinearity and non-stationarity of the signal, making the research process particularly challenging. Due to these unique and complexities, underwater sound target recognition has always been considered a major challenge in the field of underwater acoustics. To overcome these difficulties, researchers need to deeply study the propagation laws of underwater acoustic signals, develop more advanced signal processing techniques, and design more accurate recognition algorithms. This requires not only profound theoretical knowledge but also rich practical experience and a deep understanding of the underwater acoustic environment. Therefore, despite the numerous challenges, underwater acoustic target identification remains an important research direction in the field of underwater acoustics and has significant practical application value.

[0003] When identifying underwater targets, classification is not solely based on acoustic features. The core of the identification process lies in revealing the potential correlation between sound features and different ship and submarine types. To uncover this correlation, many scholars have employed mechanistic analysis methods, hoping to find identifiable descriptions that can clearly distinguish these sound features. In exploring this correlation, scholars have adopted various approaches. Starting with the most intuitive and simple waveform structure analysis, they have gradually deepened their research, using Fourier transform spectral estimation to analyze the main components of various sound sources. In recent years, with rapid technological advancements, they have further employed time-frequency methods such as wavelet analysis and Hilbert-Huang transform, which provide more refined tools for analyzing the local features of signals. Furthermore, scholars have conducted in-depth research on sound features from multiple perspectives, including fractal theory, auditory models, and acoustic imaging, aiming to find more features related to ship and submarine types. These efforts have continuously propelled the development of underwater sound target identification, showcasing its broader prospects in practical applications.

[0004] Deep learning, which has emerged in recent years, is a non-parametric classification algorithm based on neurophysiology and cognitive science theories. It can, to a certain extent, achieve automatic knowledge acquisition and automatic generation of classification rules. Neural networks, as the cornerstone of deep learning, are simplified models simulating the operation of the brain's neural networks. Their significant advantages lie in their distributed information storage mechanism, large-scale adaptive parallel processing capabilities, excellent fault tolerance, and powerful nonlinear function approximation ability. These unique characteristics have made neural networks play a crucial role in the field of pattern recognition. In particular, the significant progress of deep learning in speech recognition and its outstanding autonomous learning capabilities have opened up entirely new horizons: enabling machines to autonomously learn complex underwater sound signals. With carefully designed deep neural network structures, it is possible to capture signal representations that are drastically different from those analyzed using traditional theories. Utilizing the large-scale parallel computing capabilities of neural networks, massive amounts of raw data can be input for training, thus overcoming the limitations of traditional pattern recognition methods, which can only extract low-dimensional features due to technical constraints, and achieving a more comprehensive and in-depth understanding of signals. Once these rich signal features are mastered, combined with a high-performance classifier, multi-classification tasks can be easily handled, providing new solutions for the processing and analysis of underwater sound signals. This method not only improves the accuracy of identification, but also provides more insights and understanding about underwater sound signals. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and propose an underwater target recognition method based on the Conformer model. By applying the Conformer to underwater acoustic target recognition, it integrates the global dependency modeling capability of the Transformer and the local structure capture capability of convolution, which can effectively solve the problem of long-term dependency and coexistence of local structure of underwater acoustic signals, thereby achieving accurate target classification.

[0006] The technical problem solved by this invention is achieved through the following technical solution: An underwater target identification method based on the Conformer model includes the following steps: Step 1: Collect and process data; Step 2: Based on the collected and processed data, construct a Conformer-based model and train it. Step 3: Input the data processed in Step 1 into the trained Conformer-based model to obtain clean underwater audio with denoised audio.

[0007] Moreover, the specific implementation method of step 1 is as follows: collect ship signals and a type of environmental signal, and finally divide the processed data into training set, validation set and test set.

[0008] Furthermore, step 2, based on the Conformer model, includes: a position encoder, n Conformer modules, and a classification head, wherein the position encoder, n Conformer modules, and classification head are connected sequentially. The position encoder is used to extract high-level representations of the data in the time and frequency domains, the Conformer modules are used to obtain the local and global features of the model, and the classification head is used to output the features passed through the Conformer modules as the final classification result through a linear layer.

[0009] Furthermore, the position encoder includes two 3x3 convolutional modules and one 9x9 convolutional module. The 3x3 convolutional module, the 9x9 convolutional module, and the other 3x3 convolutional module are connected in sequence. The two 3x3 convolutional modules and the 9x9 convolutional module progressively extract high-level representations in the time domain and frequency domain through progressive feature extraction and block compression, and transform them into low-dimensional patch sequences. Then, the high-level representations in the time domain and frequency domain are flattened to one dimension and input into the Conformer model.

[0010] Moreover, the two 3x3 convolutional modules have the same structure. The 3x3 convolutional module uses a 3x3 convolutional kernel and a padding value of 1. While keeping the time-frequency feature dimension unchanged, it extracts the local key features of the underwater acoustic signal. At the same time, it convolves the number of channels by a quarter of the embedding dim, and also adds a Layer Norm layer and a ReLU nonlinear activation function.

[0011] Moreover, the 9x9 convolution module divides the large matrix into multiple 9x9 patches to extract features from 9x9 regions. Layer Norm layers and ReLU non-linear activation functions are also added after convolution.

[0012] Furthermore, the n Conformer modules are stacked to process the high-level representation of the encoder input. Each Conformer module is based on residual connections and normalization, containing two feedforward modules, sandwiched between a multi-head self-attention module and a convolutional module. The feedforward module uses half-step residual weights and consists of two linear layers and a Swish activation function. By expanding the number of channels to a high-dimensional space and then compressing it back to a low-dimensional space, it achieves feature fusion and information filtering between channels to obtain key features. The multi-head self-attention mechanism decomposes the Query, Key, and Value in the self-attention mechanism into multiple low-dimensional subspaces. Each subspace can independently learn feature associations in different dimensions. The model will eventually fuse the attention results of all subspaces to comprehensively capture feature dependencies in different positions and dimensions, avoiding the information limitations of a single subspace. At the same time, the convolutional module in the Conformer model adopts a depthwise separable convolutional structure.

[0013] The advantages and positive effects of this invention are: This invention involves collecting and processing data; constructing and training a Conformer-based model based on the collected and processed data; and inputting the data processed in step 1 into the trained Conformer-based model to obtain clean, denoised underwater acoustic audio. By applying the Conformer to underwater target recognition, this invention integrates the global dependency modeling capability of the Transformer with the local structure capture capability of convolution, effectively solving the problem of long-term dependency and coexistence of local structure in underwater acoustic signals, thereby achieving accurate target classification. Attached Figure Description

[0014] Figure 1 This is a structural diagram of the Conformer model based on the present invention; Figure 2 This is a structural diagram of the position encoder of the present invention; Figure 3 This is a structural diagram of the Conformer module of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to the accompanying drawings.

[0016] An underwater target identification method based on the Conformer model includes the following steps: Step 1: Collect and process the data.

[0017] The specific implementation method of step 1 is as follows: collect ship signals and a type of environmental signals, and finally divide the processed signals into training set, validation set and test set.

[0018] In this embodiment, the dataset comes from the ShipsEar ship sonar database, which contains signals from passing ships collected by researchers in different areas along the northern Atlantic coast of Spain. The database records sound signals from more than 10 types of ships, such as dredgers, fishing boats, passenger ships, motorboats, tugboats, and sailboats, as well as ambient sounds. This invention selects four categories from this database as datasets, based on criteria such as the inclusion of effective signal length and signal quality. Ultimately, three categories of ship signals and one category of ambient signals were selected, with 70% used for training, 15% for validation, and 15% for testing.

[0019] Step 2: Based on the collected and processed data, construct a Conformer-based model and train it.

[0020] like Figure 1As shown, step 2, based on the Conformer model, includes: a position encoder, n Conformer modules, and a classification head. The position encoder, n Conformer modules, and classification head are connected sequentially. The position encoder is used to extract high-level representations of the data in the time and frequency domains. The Conformer modules are used to obtain the local and global features of the model. The classification head is used to output the features passed through the Conformer modules as the final classification result through a linear layer.

[0021] First, a short-time Fourier transform is performed on the signal to obtain its assigned spectrum and phase spectrum. The assigned spectrum is a real matrix, and the phase information is a complex-valued matrix. In this invention, the real and imaginary values ​​of the complex-valued matrix are separated into two channels, and these three matrices are combined with the real matrix to form three matrices as input to the position encoder. The encoder extracts high-level representations in the time and frequency domains step by step through two 3x3 convolutions and one 9x9 patch convolution. Then, the time and frequency domain information is viewed into one dimension, resulting in features that can be input to the conformer block. To preserve crucial positional information, this invention incorporates positional encoding into the features. Subsequently, this invention inputs these features into the conformer block, which obtains the model's local and global features. These features are then input into the classification head to obtain the final classification result.

[0022] like Figure 2 As shown, the position encoder includes two 3x3 convolutional modules and one 9x9 convolutional module. The 3x3 convolutional module, the 9x9 convolutional module and the other 3x3 convolutional module are connected in sequence. The two 3x3 convolutional modules and the 9x9 convolutional module gradually extract high-level representations in the temporal and frequency domains through progressive feature extraction and block compression, and transform them into low-dimensional patch sequences. Then, the high-level representations in the temporal and frequency domains are flattened to one dimension and input into the Conformer model.

[0023] The number of channels in a feature is a core dimension determining its expressive power; a larger number of channels better captures local time-frequency features. However, increasing the number of feature channels too much at once may lead to information gaps. Therefore, to further increase the number of model channels, this invention adds a 3x3 module before and after the 9x9 convolution module, gradually increasing the number of feature channels without changing the feature dimension. The 3x3 module uses a 3x3 convolution kernel with a padding value of 1. While maintaining the time-frequency feature dimension, it extracts local key features of the underwater acoustic signal and convolves the number of channels to a quarter of the embedding dim. Layer Norm layers and ReLU non-linear activation functions are then added.

[0024] The 9x9 convolution module has its kernel and stride set to (9, 9), which divides the large matrix into multiple 9x9 patches to extract features from 9x9 regions. Layer Norm and ReLU non-linear activation functions are also added after convolution.

[0025] The third convolutional module has the same structure as the first convolutional module, and it realizes the transformation of the embedding dimension from half of the embedding dim to the embedding dim.

[0026] After the three convolutional modules are completed, this invention flattens the time and frequency dimensions into one dimension to obtain the output of the Encoder layer.

[0027] This invention introduces a learnable positional encoding to capture the absolute positional information between sequence elements. The learnable positional encoding assigns a unique feature vector to each position in the sequence through a parameterized embedding layer, and optimizes the parameters through training.

[0028] like Figure 3 As shown, this invention selects the Conformer Block as the core feature processor because it combines the global dependency modeling capability of the Transformer with the local structure capture capability of convolution, effectively addressing the coexistence of long-term dependency and local structure in underwater acoustic signals. This model stacks N Conformer Blocks to process features, with the detailed structure as follows: Figure 3 The single Conformer block is based on a residual connection and normalization framework, containing two feedforward modules sandwiched between a multi-head self-attention module and a convolutional module. The feedforward modules employ half-step residual weights. The feedforward module consists of two linear layers and a Swish activation function, expanding the number of channels to a high-dimensional space and then compressing it back to a low-dimensional space to achieve feature fusion and information filtering between channels, obtaining key features. The multi-head self-attention mechanism decomposes the Query, Key, and Value in the self-attention mechanism into multiple low-dimensional subspaces. Each subspace can independently learn feature associations in different dimensions. The model ultimately fuses the attention results of all subspaces, comprehensively capturing feature dependencies at different locations and in different dimensions, avoiding the information limitations of a single subspace. The convolutional module in the Conformer block uses a depthwise separable convolutional structure, where PointwiseConv1d handles channel dimension transformation and DepthwiseConv1d handles local sequence association modeling. This convolutional structure effectively reduces the number of parameters required by the model while achieving the effect of standard convolution, accelerating model training and inference.

[0029] Step 3: Input the data processed in Step 1 into the trained Conformer-based model to obtain clean underwater audio with denoised audio.

[0030] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. An underwater target recognition method based on the Conformer model, characterized in that: Includes the following steps: Step 1: Collect and process data; Step 2: Based on the collected and processed data, construct a Conformer-based model and train it. Step 3: Input the data processed in Step 1 into the trained Conformer-based model to obtain clean underwater audio with denoised audio.

2. The underwater target recognition method based on the Conformer model according to claim 1, characterized in that: The specific implementation method of step 1 is as follows: collect ship signals and a type of environmental signal, and finally divide the processed data into training set, validation set and test set.

3. The underwater target recognition method based on the Conformer model according to claim 2, characterized in that: Step 2, based on the Conformer model, includes: a position encoder, n Conformer modules, and a classification head. The position encoder, n Conformer modules, and classification head are connected sequentially. The position encoder is used to extract high-level representations of the data in the time and frequency domains. The Conformer modules are used to obtain the local and global features of the model. The classification head is used to output the features passed through the Conformer modules as the final classification result through a linear layer.

4. The underwater target recognition method based on the Conformer model according to claim 3, characterized in that: The position encoder includes two 3x3 convolutional modules and one 9x9 convolutional module. The 3x3 convolutional module, the 9x9 convolutional module and the other 3x3 convolutional module are connected in sequence. The two 3x3 convolutional modules and the 9x9 convolutional module gradually extract high-level representations in the time domain and frequency domain through progressive feature extraction and block compression, and transform them into low-dimensional patch sequences. Then, the high-level representations in the time domain and frequency domain are flattened to one dimension and input into the Conformer model.

5. The underwater target recognition method based on the Conformer model according to claim 3, characterized in that: The two 3x3 convolutional modules have the same structure. The 3x3 convolutional modules use 3x3 convolutional kernels with a padding value of 1. While keeping the time-frequency feature dimension unchanged, they extract local key features of the underwater acoustic signal. At the same time, they convolve the number of channels by a quarter of the embedding dim, and also add a Layer Norm layer and a ReLU nonlinear activation function.

6. The underwater target recognition method based on the Conformer model according to claim 3, characterized in that: The 9x9 convolution module divides the large matrix into multiple 9x9 patches to extract features from 9x9 regions. After convolution, a Layer Norm layer and a ReLU non-linear activation function are also added.

7. The underwater target recognition method based on the Conformer model according to claim 3, characterized in that: The n Conformer modules are stacked to process the high-level representation of the encoder input. Each Conformer module is based on residual connections and normalization, containing two feedforward modules, sandwiched between a multi-head self-attention module and a convolutional module. The feedforward module uses half-step residual weights and consists of two linear layers and a Swish activation function. It achieves feature fusion and information filtering between channels by expanding the number of channels to a high-dimensional space and then compressing it back to a low-dimensional space to obtain key features. The multi-head self-attention mechanism decomposes the Query, Key, and Value in the self-attention mechanism into multiple low-dimensional subspaces. Each subspace can independently learn feature associations in different dimensions. The model will eventually fuse the attention results of all subspaces to comprehensively capture feature dependencies in different positions and dimensions, avoiding the information limitations of a single subspace. At the same time, the convolutional module in the Conformer model adopts a depthwise separable convolutional structure.

Images