An underwater acoustic target recognition method and system based on an expert hybrid model

Abstract

The application discloses a kind of based on expert mixed model's underwater acoustic target identification method and system, since the distribution law of different frequency bands of underwater acoustic signal exists difference, to the problem of insufficient low-frequency feature extraction and unknown frequency distribution of underwater acoustic signal, introduce unbalanced bandwidth mel spectrum diagram, to improve the learning ability to different frequency band characteristics, to realize the time-frequency feature enhancement for underwater acoustic target identification task.For the problem that the model has limited learning ability to different frequency band characteristics when identifying different sub-band characteristics, an expert mixed model for different sub-band characteristics is designed to improve the learning ability of UBM-MoER network to different frequency band characteristics. In order to further explore how to improve the feature extraction capability of the model under the expert mixed model, a mixed domain attention mechanism module is designed, which is expected to improve the sensitivity of each expert mixed model to key features through channel and spatial attention mechanisms, thereby improving the overall performance of the model.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic target recognition technology, and relates to an underwater acoustic target recognition method and system based on an expert hybrid model. Background Technology

[0002] Passive underwater acoustic target identification is a crucial component of underwater sound source management, protection of key marine areas, and future marine information technology development. However, underwater acoustic target identification is an extremely challenging task. Underwater acoustic targets exhibit pseudo-randomness and are affected by marine environmental noise. The differences between targets are not significant, and their motion states are diverse, leading to frequent misjudgments or conflicting results. Traditional identification methods cannot meet the increasingly complex and demanding requirements, while deep learning methods offer a possibility for achieving underwater acoustic target identification.

[0003] Traditional underwater acoustic target recognition combines signal processing with machine learning. Underwater acoustic signal characteristics are generally divided into time-domain features and spectral features. Time-domain features mainly include mean, zero-crossing rate, and peak value. These features do not require complex feature extraction. However, due to the complex composition of underwater acoustic target radiated noise, these features are poorly descriptive. Therefore, applying frequency-domain features is crucial in target recognition technology. Time-frequency features can characterize the change of signal frequency over time, providing a more comprehensive description of the signal's characteristics. Short-time Fourier transform and wavelet transform methods can be used to achieve time-frequency representation. However, underwater acoustic signals exhibit frequency-division absorptivity during propagation; high-frequency signals are gradually absorbed by seawater at close range, while low-frequency signals propagate over longer distances. Therefore, in underwater acoustic target signal recognition tasks, due to the lack of prior target knowledge, the target distance corresponding to the received signal is unknown, and the main frequency band of the signal is also unknown. Machine learning methods based on statistical classification can achieve good results in certain ideal environments, but machine learning can only roughly fit the distribution of samples and cannot uncover the deep features of the input data. Public research on feature extraction or recognition of underwater passive targets using deep learning networks has been less than ten years, but it has shown significant advantages over traditional methods. Because it can automatically explore the features required for the task from complex raw data, it has shone brightly in fields such as computer vision and natural language processing. Convolutional neural networks have undergone a series of improvements, such as AlexNet, VGGNet, ResNet, and DenseNet, and have achieved strong feature extraction capabilities. Due to their superiority in feature extraction, many studies have applied them to underwater acoustic target recognition. However, in actual recognition tasks, the accuracy has shown a downward trend. The main reasons for this phenomenon are as follows: (1) Due to the differences in the distribution patterns of underwater acoustic signals in different frequency bands, insufficient extraction of low-frequency features and unknown frequency distribution of underwater acoustic signals affect recognition efficiency; (2) Due to the complexity and diversity of underwater acoustic signals, network models are prone to insufficient extraction of their time-frequency features; (3) Ship noise with large intra-class differences and high inter-class similarity cannot be effectively identified, leading to recognition errors. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of low recognition rate in underwater acoustic recognition in the prior art, and to provide an underwater acoustic target recognition method and system based on an expert hybrid model.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] This invention proposes an underwater acoustic target recognition method based on an expert hybrid model, comprising:

[0007] Acquire underwater acoustic signal samples, preprocess the underwater acoustic signal samples, and obtain the unbalanced bandwidth Mel spectrum time-frequency diagram;

[0008] A front-end backbone network is used to convert the unbalanced bandwidth Mel-spectral time-frequency map into a fixed-dimensional time-frequency map to obtain deep features.

[0009] Deep features are input into a pre-built expert hybrid model to obtain underwater acoustic target recognition results.

[0010] Preferably, the step of acquiring underwater acoustic signal samples and preprocessing the underwater acoustic signal samples to obtain an unbalanced bandwidth Mel-spectrum time-frequency diagram specifically involves:

[0011] The underwater acoustic signal samples are divided into multiple frequency bands by frame segmentation. Each frame is windowed, and a Fourier transform is performed on each frame group. Then, the underwater acoustic signal samples are input into different filters to perform logarithmic operations to obtain subband Mel spectra. The subband Mel spectra are concatenated along the frequency direction to obtain Mel features. Based on the Mel features, the unbalanced bandwidth Mel spectra time-frequency diagram is obtained.

[0012] Preferably, the front-end backbone network structure is based on ResNet architecture, using LayerNorm as the normalization method and GELU as the activation function; specifically: the unbalanced bandwidth Mel spectrum time-frequency map is passed through convolutional layers and LN layers, and the subsequent four stages are implemented through downsampling modules and stacked ConvNeXt Blocks of different numbers to achieve feature extraction, and finally through a global average pooling layer to output deep features;

[0013] The downsampling module consists of LN layers and convolutional layers, while the ConvNeXt Block consists of convolutional layers, LN layers, GELU layers, and a hybrid domain attention mechanism module.

[0014] Preferably, the hybrid domain attention mechanism module combines the channel attention module and the spatial attention module;

[0015] The feature map input to the channel attention mechanism module is processed by width-based global max pooling and height-based global average pooling, and then processed by a shared fully connected layer. After summation and sigmoid activation, the channel attention feature map is obtained.

[0016] The spatial attention mechanism takes the output of the channel attention module as input, performs global max pooling and global average pooling operations on the input feature map, then concatenates the two results, reduces the channel dimension to 1 through convolution operation, and generates a spatial attention feature map through Sigmoid.

[0017] Preferably, the expert hybrid model consists of a routing layer and an expert layer. The routing layer is composed of a softmax function, and the expert layer is composed of a multilayer perceptron with two linear layers.

[0018] Preferably, the step of inputting deep features into a pre-constructed expert hybrid model to obtain underwater acoustic target recognition results specifically involves:

[0019] Set the batch size to n, and the input unbalanced bandwidth Mel-spectrum time-frequency plot to {x}. i}, and its corresponding label is set to {y i}, (i = 1, 2...n), the number of expert layers is set to m, and the number of expert layers is set to E1(·),...E m (·), Input unbalanced bandwidth Mel spectrum time-frequency plot x i After passing through the front-end backbone network, deep features r are obtained. i , to deep feature r i The input is fed into the routing layer S(·) to obtain the routing score. The route score is normalized to the route probability p using the softmax function. i Routing probability p i This represents the probability of assigning a sample to each expert; the deep feature r i The data is input into the expert layer with the highest probability to obtain the ship's category, which is the underwater acoustic target recognition result.

[0020] This invention proposes an underwater acoustic target recognition system based on an expert hybrid model, comprising:

[0021] The data preprocessing module is used to acquire underwater acoustic signal samples, preprocess the underwater acoustic signal samples, and obtain the unbalanced bandwidth Mel spectrum time-frequency diagram.

[0022] The feature extraction module is used to convert the unbalanced bandwidth Mel spectrum time-frequency map into a fixed-dimensional time-frequency map using the front-end backbone network to obtain deep features;

[0023] The target recognition module is used to input deep features into a pre-built expert hybrid model to obtain underwater acoustic target recognition results.

[0024] Preferably, the step of acquiring underwater acoustic signal samples and preprocessing the underwater acoustic signal samples to obtain an unbalanced bandwidth Mel-spectrum time-frequency diagram specifically involves:

[0025] The underwater acoustic signal samples are divided into multiple frequency bands by frame segmentation. Each frame is windowed, and a Fourier transform is performed on each frame group. Then, the underwater acoustic signal samples are input into different filters to perform logarithmic operations to obtain subband Mel spectra. The subband Mel spectra are concatenated along the frequency direction to obtain Mel features. Based on the Mel features, the unbalanced bandwidth Mel spectra time-frequency diagram is obtained.

[0026] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement a method for underwater acoustic target recognition based on an expert hybrid model.

[0027] A computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements the steps of an underwater acoustic target recognition method based on an expert hybrid model.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] This invention proposes an underwater acoustic target recognition method based on an expert hybrid model. Due to the differences in the distribution patterns of underwater acoustic signals across different frequency bands, and addressing the issues of insufficient low-frequency feature extraction and unknown frequency distribution, an unbalanced bandwidth Mel spectrum is introduced to improve the learning ability for features across different frequency bands, thereby achieving time-frequency feature enhancement for underwater acoustic target recognition tasks. To address the limited learning ability of the model for different frequency band features during sub-band feature recognition, an expert hybrid model is designed for different sub-band features, enhancing the learning ability of the UBM-MoER network for different frequency band features. To further explore how to improve the feature extraction capability of the model under the expert hybrid model, a hybrid domain attention mechanism module is designed, aiming to improve the sensitivity of each expert hybrid model to key features through channel and spatial attention mechanisms, thereby improving the overall performance of the model.

[0030] This invention proposes an underwater acoustic target recognition system based on an expert hybrid model. By dividing the system into a data preprocessing module, a feature extraction module, and a target recognition module, it obtains underwater acoustic target recognition results and achieves underwater acoustic target recognition. The modular approach ensures that each module is independent, facilitating unified management of all modules. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of the underwater acoustic target recognition method based on an expert hybrid model according to the present invention.

[0033] Figure 2 This is a diagram of the overall network architecture of the present invention.

[0034] Figure 3 This is a diagram illustrating the unbalanced bandwidth Mel spectrum feature extraction process of the present invention.

[0035] Figure 4 This is a structural diagram of the front-end backbone network module of the present invention.

[0036] Figure 5 This is a structural diagram of the hybrid domain attention mechanism module of the present invention.

[0037] Figure 6 This is a structural diagram of the channel attention mechanism module of the present invention.

[0038] Figure 7 This is a structural diagram of the spatial attention mechanism module of the present invention.

[0039] Figure 8 This is a structural diagram of the expert hybrid model of the present invention.

[0040] Figure 9 This is a diagram of the underwater acoustic target recognition system based on an expert hybrid model according to the present invention.

[0041] Figure 10 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0043] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0044] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0045] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0046] The present invention will now be described in further detail with reference to the accompanying drawings:

[0047] This invention proposes an underwater acoustic target recognition method based on an expert hybrid model, such as... Figure 1 As shown, it includes the following steps:

[0048] S1. Acquire underwater acoustic signal samples, preprocess the underwater acoustic signal samples, and obtain the unbalanced bandwidth Mel spectrum time-frequency diagram;

[0049] The acquisition of underwater acoustic signal samples, including preprocessing the samples to obtain the unbalanced bandwidth Mel-spectrum time-frequency diagram, specifically involves:

[0050] The underwater acoustic signal samples are divided into multiple frequency bands by frame segmentation. Each frame is windowed, and a Fourier transform is performed on each frame group. Then, the underwater acoustic signal samples are input into different filters to perform logarithmic operations to obtain subband Mel spectra. The subband Mel spectra are concatenated along the frequency direction to obtain Mel features. Based on the Mel features, the unbalanced bandwidth Mel spectra time-frequency diagram is obtained.

[0051] S2. The unbalanced bandwidth Mel-spectral time-frequency map is converted into a fixed-dimensional time-frequency map using a front-end backbone network to obtain deep features;

[0052] The front-end backbone network structure is based on ResNet, using LayerNorm as the normalization method and GELU as the activation function. Specifically, the unbalanced bandwidth Mel spectrum time-frequency map is passed through convolutional layers and LN layers. The subsequent four stages use downsampling modules and stacked ConvNeXt Blocks of different numbers to achieve feature extraction. Finally, it passes through a global average pooling layer to output deep features.

[0053] The downsampling module consists of LN layers and convolutional layers, while the ConvNeXt Block consists of convolutional layers, LN layers, GELU layers, and a hybrid domain attention mechanism module.

[0054] The hybrid domain attention mechanism module combines the channel attention module and the spatial attention module;

[0055] The feature map input to the channel attention mechanism module is processed by width-based global max pooling and height-based global average pooling, and then processed by a shared fully connected layer. After summation and sigmoid activation, the channel attention feature map is obtained.

[0056] The spatial attention mechanism takes the output of the channel attention module as input, performs global max pooling and global average pooling operations on the input feature map, then concatenates the two results, reduces the channel dimension to 1 through convolution operation, and generates a spatial attention feature map through Sigmoid.

[0057] S3. Input the deep features into the pre-built expert hybrid model to obtain the underwater acoustic target recognition results.

[0058] The expert hybrid model consists of a routing layer and an expert layer. The routing layer is composed of a softmax function, and the expert layer is composed of a multilayer perceptron with two linear layers.

[0059] The process of inputting deep features into a pre-constructed expert hybrid model to obtain underwater acoustic target recognition results specifically involves:

[0060] Set the batch size to n, and the input unbalanced bandwidth Mel-spectrum time-frequency plot to {x}. i}, and its corresponding label is set to {y i}, (i = 1, 2...n), the number of expert layers is set to m, and the number of expert layers is set to E1(·),...E m (·), Input unbalanced bandwidth Mel spectrum time-frequency plot x i After passing through the front-end backbone network, deep features r are obtained. i , to deep feature r i The input is fed into the routing layer S(·) to obtain the routing score. The route score is normalized to the route probability p using the softmax function. i Routing probability p i This represents the probability of assigning a sample to each expert; the deep feature r i The data is input into the expert layer with the highest probability to obtain the ship's category, which is the underwater acoustic target recognition result.

[0061] The following description is in conjunction with the accompanying drawings:

[0062] Figure 2This is the overall network architecture diagram proposed in this invention, which consists of three parts. First, the signal is processed through time-frequency representation and non-evaluation filtering to obtain the unbalanced bandwidth Mel spectrum time-frequency diagram. Then, it is input into the front-end backbone network for feature extraction. Finally, the MoE network is used for classification and recognition.

[0063] Figure 3 This diagram illustrates the unbalanced bandwidth Mel spectrum time-frequency feature extraction process of this invention. To achieve effective classification and recognition, the underwater acoustic signal is framed into multiple frequency bands. Each band typically overlaps to eliminate errors caused by the segmentation method. Then, each frame is windowed using a Hamming window to reduce the impact of edge effects on the signal's frequency components. A Fourier transform is then performed on each frame group. Since the main acoustic features of distant targets in underwater acoustics are concentrated in the low-frequency band, and ocean background noise is also concentrated in the low-frequency band, effective identification of low-frequency information is particularly important. The sampling rate is set to 16000Hz. When the signal undergoes a short-time Fourier transform (STFT), the STFT window length is set to 2048, and the frame shift is set to 512. The signal is then input into different sub-band Mel filters. After a logarithmic operation, the Mel spectrum is logarithmized, typically to match the nonlinear loudness perception of the human ear. The subbands are divided into 0-1000, 1000-4000, and 4000-8000 Hz. 128 Mel filters are applied in the 0-1000 Hz range, 64 Mel filters in the 1000-4000 Hz range, and 32 Mel filters in the 4000-8000 Hz range. The resulting Mel spectra of the three subbands are concatenated along the frequency direction to obtain a 224-dimensional Mel feature, ultimately yielding the unbalanced bandwidth Mel spectrum. The conversion formula between Mel scaling and frequency is:

[0064] Mel(f) = 2595lg(1+f / 700)

[0065] Where f represents frequency. The Mel-spectrum is divided into sub-bands, and a different number of independent Mel filter banks are used in each sub-band. The resulting sub-band features are then spliced ​​together.

[0066] Figure 4This is a structural diagram of the front-end backbone network module of the present invention. The front-end backbone network converts the input acoustic features into a fixed-dimensional representation. The structure of the front-end backbone network consists of two main parts: the ConvNeXt network and the HAM module. ConvNeXt is based on the ResNet architecture and borrows the idea of ​​Swin-Transformer in its design process. It uses LayerNorm(LN) as the normalization method and GELU as the activation function to simplify the network structure, reduce parameters and computational cost, and improve the generalization ability and performance of the model. The input time-frequency image is 3*224*224 in size. It first passes through convolutional layers and LN layers. The next four stages extract features through a downsampling module and stacked ConvNeXt Blocks of different numbers. The downsampling module consists of LN layers and convolutional layers. The ConvNeXt Block includes convolutional layers, LN layers, GELU layers, and a Hybrid Attention Mechanism (HAM) module. The HAM module is placed at the end of the ConvNeXt Block to enhance the network's ability to extract key features. Finally, it passes through a global average pooling layer to output deep features.

[0067] Figure 5 This is a structural diagram of the hybrid domain attention mechanism module of the present invention. The hybrid domain attention mechanism combines a channel attention module and a spatial attention module, simultaneously focusing on the correlation between channels and key information and the importance of feature maps in the spatial dimension. Compared with attention mechanisms that only focus on one direction, it can extract richer deep features.

[0068] Figure 6 This is a structural diagram of the Channel Attention Module of this invention. The input feature map F is processed by width-based global max pooling (AvgPool(F)) and height-based global average pooling (MaxPool(F)). Then, it is processed through a shared fully connected layer (Shared MLP), and after summation and sigmoid activation, the channel attention feature map is obtained. c (F) represents the output feature map of the channel attention mechanism.

[0069] M c (F)=σ(MLP(AvgPool(F)+MLP(MaxPool(F)))

[0070] Figure 7This is a structural diagram of the Spatial Attention Module of this invention. The Spatial Attention Module takes the output of the Channel Attention Module as its input. It performs global max pooling and global average pooling (AvgPool) operations on the input feature map, then concatenates the results, reduces the channel dimension to 1 through convolution (Conv layer), and generates a spatial attention feature map using a sigmoid function. s (F) represents the output feature map of the spatial attention mechanism.

[0071] M s (F)=σ(f 7×7 ([AvgPool(F);MaxPool(F)]))

[0072] Figure 8 This is a structural diagram of the expert hybrid model of the present invention. The expert hybrid model consists of a routing layer and expert layers. The routing layer is composed of a softmax function, and the expert layer is composed of a multilayer perceptron with two linear layers. Different expert layers focus on differentiated deep features, thereby enhancing the model's classification ability. Each expert has the same structure, but its parameters are independent, and each expert is only activated when it encounters suitable input. This structure significantly improves the performance of underwater acoustic target recognition by capturing the latent characteristics of high-level features and adaptively learning from diverse data.

[0073] Set the batch size to n and the input spectrogram to {x}. i}, and its corresponding label is set to {y i}, (i = 1, 2...n), the number of expert layers is set to m, and the number of expert layers is set to E1(·),…E m (·), input x i After passing through the front-end backbone network, deep features are obtained, and these deep features r i The input is fed into the routing layer S(·) to obtain the routing score. Then, the softmax function is used to normalize the route score into the route probability p. i p i This represents the probability of assigning a sample to each expert.

[0074] The model will use deep features r i The input is sent to the expert layer with the highest probability. The entire route allocation is as shown in the formula below, using a UBM-MOER model with 4 experts. When p iWhen r = (0.1, 0.1, 0.2, 0.6), the depth feature r i This will be input into E4(·), and the final result (logits) will be obtained, which identifies the type of ship.

[0075] p i =softmax(S(r) i ),i∈[1,n]

[0076]

[0077] Example 2

[0078] This invention proposes an underwater acoustic target recognition system based on an expert hybrid model, such as... Figure 9 The above includes:

[0079] The data preprocessing module is used to acquire underwater acoustic signal samples, preprocess the underwater acoustic signal samples, and obtain the unbalanced bandwidth Mel spectrum time-frequency diagram.

[0080] The acquisition of underwater acoustic signal samples, including preprocessing the samples to obtain the unbalanced bandwidth Mel-spectrum time-frequency diagram, specifically involves:

[0081] The underwater acoustic signal samples are divided into multiple frequency bands by frame segmentation. Each frame is windowed, and a Fourier transform is performed on each frame group. Then, the underwater acoustic signal samples are input into different filters to perform logarithmic operations to obtain subband Mel spectra. The subband Mel spectra are concatenated along the frequency direction to obtain Mel features. Based on the Mel features, the unbalanced bandwidth Mel spectra time-frequency diagram is obtained.

[0082] The feature extraction module is used to convert the unbalanced bandwidth Mel spectrum time-frequency map into a fixed-dimensional time-frequency map using the front-end backbone network to obtain deep features;

[0083] The front-end backbone network structure is based on ResNet, using LayerNorm as the normalization method and GELU as the activation function. Specifically, the unbalanced bandwidth Mel spectrum time-frequency map is passed through convolutional layers and LN layers. The subsequent four stages use downsampling modules and stacked ConvNeXt Blocks of different numbers to achieve feature extraction. Finally, it passes through a global average pooling layer to output deep features.

[0084] The downsampling module consists of LN layers and convolutional layers, while the ConvNeXt Block consists of convolutional layers, LN layers, GELU layers, and a hybrid domain attention mechanism module.

[0085] The hybrid domain attention mechanism module combines the channel attention module and the spatial attention module;

[0086] The feature map input to the channel attention mechanism module is processed by width-based global max pooling and height-based global average pooling, and then processed by a shared fully connected layer. After summation and sigmoid activation, the channel attention feature map is obtained.

[0087] The spatial attention mechanism takes the output of the channel attention module as input, performs global max pooling and global average pooling operations on the input feature map, then concatenates the two results, reduces the channel dimension to 1 through convolution operation, and generates a spatial attention feature map through Sigmoid.

[0088] The target recognition module is used to input deep features into a pre-built expert hybrid model to obtain underwater acoustic target recognition results.

[0089] The expert hybrid model consists of a routing layer and an expert layer. The routing layer is composed of a softmax function, and the expert layer is composed of a multilayer perceptron with two linear layers.

[0090] The process of inputting deep features into a pre-constructed expert hybrid model to obtain underwater acoustic target recognition results specifically involves:

[0091] Set the batch size to n, and the input unbalanced bandwidth Mel-spectrum time-frequency plot to {x}. i}, and its corresponding label is set to {y i}, (i = 1, 2...n), the number of expert layers is set to m, and the number of expert layers is set to E1(·),...E m (·), Input unbalanced bandwidth Mel spectrum time-frequency plot x i After passing through the front-end backbone network, deep features r are obtained. i , to deep feature r i The input is fed into the routing layer S(·) to obtain the routing score. The route score is normalized to the route probability p using the softmax function. i Routing probability p i This represents the probability of assigning a sample to each expert; the deep feature r i The data is input into the expert layer with the highest probability to obtain the ship's category, which is the underwater acoustic target recognition result.

[0092] Example 3

[0093] Please see Figure 10As shown, the present invention also provides an electronic device 100 for an underwater acoustic target recognition method based on an expert hybrid model; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0094] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the underwater acoustic target recognition method based on the expert hybrid model described in Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0095] The at least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor. The processor 102 is the control center of the electronic device 100, connecting various parts of the electronic device 100 via various interfaces and lines.

[0096] The memory 101 in the electronic device 100 stores multiple instructions to implement an underwater acoustic target recognition method based on an expert hybrid model, and the processor 102 can execute the multiple instructions to achieve the following:

[0097] Acquire underwater acoustic signal samples, preprocess the underwater acoustic signal samples, and obtain the unbalanced bandwidth Mel spectrum time-frequency diagram;

[0098] A front-end backbone network is used to convert the unbalanced bandwidth Mel-spectral time-frequency map into a fixed-dimensional time-frequency map to obtain deep features.

[0099] Deep features are input into a pre-built expert hybrid model to obtain underwater acoustic target recognition results.

[0100] Example 4

[0101] If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0102] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0103] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0104] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0105] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for underwater acoustic target recognition based on an expert hybrid model, characterized in that, include: Acquire underwater acoustic signal samples, preprocess the underwater acoustic signal samples, and obtain the unbalanced bandwidth Mel spectrum time-frequency diagram; A front-end backbone network is used to convert the unbalanced bandwidth Mel-spectral plot time-frequency plot into a fixed-dimensional time-frequency plot to obtain deep features. Deep features are input into a pre-built expert hybrid model to obtain underwater acoustic target recognition results; The underwater acoustic signal samples are divided into multiple frequency bands by frame, each frame is windowed, and each frame group is subjected to Fourier transform. Then, the underwater acoustic signal samples are input into different filters and logarithmic operations are performed to obtain the sub-band Mel spectra. The sub-band Mel spectra are concatenated along the frequency direction to obtain Mel features. Based on the Mel features, the unbalanced bandwidth Mel spectra time-frequency diagram is obtained. The front-end backbone network structure is based on ResNet, using LayerNorm as the normalization method and GELU as the activation function. Specifically, the imbalanced bandwidth Mel spectrum time-frequency map is passed through convolutional layers and LN layers. The subsequent four stages use downsampling modules and stacked ConvNeXt Blocks of different numbers to achieve feature extraction. Finally, it passes through a global average pooling layer to output deep features. The downsampling module consists of an LN layer and a convolutional layer, while the ConvNeXt Block consists of a convolutional layer, an LN layer, a GELU layer, and a hybrid domain attention mechanism module. The subbands are divided into 0-1000, 1000-4000, and 4000-8000 Hz, with 128 Mel filters set in 0-1000 Hz, 64 Mel filters set in 1000-4000 Hz, and 32 Mel filters set in 4000-8000 Hz. The Mel spectra of the three subbands are concatenated along the frequency direction to obtain 224-dimensional Mel features, and finally, the unbalanced bandwidth Mel spectra are obtained.

2. The underwater acoustic target recognition method based on an expert hybrid model according to claim 1, characterized in that, The hybrid domain attention mechanism module combines the channel attention module and the spatial attention module; The feature map input to the channel attention mechanism module is processed by width-based global max pooling and height-based global average pooling, and then processed by a shared fully connected layer. After summation and sigmoid activation, the channel attention feature map is obtained. The spatial attention mechanism takes the output of the channel attention module as input, performs global max pooling and global average pooling operations on the input feature map, then concatenates the two results, reduces the channel dimension to 1 through convolution operation, and generates a spatial attention feature map through Sigmoid.

3. The underwater acoustic target recognition method based on an expert hybrid model according to claim 1, characterized in that, The expert hybrid model consists of a routing layer and an expert layer, wherein the routing layer is a... The function is structured such that the expert layer consists of a multilayer perceptron with two linear layers.

4. The underwater acoustic target recognition method based on an expert hybrid model according to claim 3, characterized in that, The process of inputting deep features into a pre-constructed expert hybrid model to obtain underwater acoustic target recognition results specifically involves: Set the batch size to Set the input unbalanced bandwidth Mel spectrum time-frequency plot to Its corresponding tag is set to (i=1,2... The number of expert levels is set to The expert layer is set as follows: Input unbalanced bandwidth Mel spectrum time-frequency plot After passing through the front-end backbone network, deep features are obtained. Deep features Input to the routing layer In the middle, the routing score is obtained. The softmax function is used to normalize the route scores into route probabilities. Routing probability This represents the probability of assigning a sample to each expert; deep features The data is input into the expert layer with the highest probability to obtain the ship's category, which is the underwater acoustic target recognition result.

5. A hydroacoustic target recognition system based on an expert hybrid model, characterized in that, The underwater acoustic target recognition method based on an expert hybrid model according to any one of claims 1 to 4 includes: The data preprocessing module is used to acquire underwater acoustic signal samples, preprocess the underwater acoustic signal samples, and obtain the unbalanced bandwidth Mel spectrum time-frequency diagram. The feature extraction module is used to convert the unbalanced bandwidth Mel spectrum time-frequency map into a fixed-dimensional time-frequency map using the front-end backbone network to obtain deep features; The target recognition module is used to input deep features into a pre-built expert hybrid model to obtain underwater acoustic target recognition results.

6. The underwater acoustic target recognition system based on an expert hybrid model according to claim 5, characterized in that, The acquisition of underwater acoustic signal samples, including preprocessing the samples to obtain the unbalanced bandwidth Mel-spectrum time-frequency diagram, specifically involves: The underwater acoustic signal samples are divided into multiple frequency bands by frame segmentation. Each frame is windowed, and a Fourier transform is performed on each frame group. Then, the underwater acoustic signal samples are input into different filters to perform logarithmic operations to obtain subband Mel spectra. The subband Mel spectra are concatenated along the frequency direction to obtain Mel features. Based on the Mel features, the unbalanced bandwidth Mel spectra time-frequency diagram is obtained.

7. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the underwater acoustic target recognition method based on the expert hybrid model as described in any one of claims 1 to 4.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the underwater acoustic target recognition method based on the expert hybrid model as described in any one of claims 1 to 4.

Images