An underwater target recognition method based on dual-channel self-supervised acoustic feature learning

Abstract

In order to further study the performance of underwater target recognition task by contrastive self-supervised feature learning method, a kind of underwater target recognition method based on double-channel self-supervised acoustic feature learning is proposed. It includes the following steps: (1) a double-channel self-attention audio encoder model is proposed; (2) a double-channel self-attention audio encoder model with dynamic positive sample storage is proposed; (3) the underwater target recognition method based on the double-channel self-attention audio encoder model with dynamic positive sample storage is completed, the spectral feature of the double-channel self-attention audio encoder with dynamic positive sample storage is extracted, and the multi-layer perception machine model and the multi-classification logistic regression model are used to complete the underwater target recognition task. The underwater target recognition method based on double-channel self-supervised acoustic feature learning has good recognition accuracy and convergence speed, can effectively recognize underwater targets in noisy environment, and has strong robustness.

Description

Technical Field

[0001] This invention relates to the field of underwater target recognition technology, and in particular to an underwater target recognition method based on dual-channel self-supervised acoustic feature learning. Background Technology

[0002] Underwater target recognition is a crucial research direction and technical challenge in the field of marine exploration, with widespread applications in both military and civilian sectors. Developing high-precision, intelligent underwater target recognition technologies to provide accurate and efficient decision support for situational analysis is an important task that aligns with my country's strategic development needs and is vital for implementing the national maritime power strategy. With the development of deep learning and the increasing volume of related data, deep neural network-based underwater target recognition has attracted growing attention from researchers. In recent years, scholars at home and abroad have conducted in-depth analysis and research on acoustic event detection. Among them, the most famous and effective underwater target recognition methods in the existing literature mainly include: 1. Underwater target recognition based on hybrid convolutional neural networks: B Vikram Deep, Ratnakar Dash. Underwater fish species recognition using deep learning techniques[C] / / 2019 6th International Conference on Signal Processing and Integrated Networks (SPIN), Noida, India, 2019:665-669. This paper proposes a hybrid convolutional neural network framework to extract underwater marine biological target features and uses SVM and K-nearest neighbor algorithms for classification to achieve the recognition of 12 fish species. Its recognition performance is better than traditional machine learning methods. 2. Underwater acoustic signal recognition and classification based on deep learning: Jin Guanghao, Liu Fan, Wu Hao, Song Qingzeng. Deep learning-based framework for expansion, recognition and classification of underwater acoustic signal. Journal of Experimental & Theoretical Artificial Intelligence, 2020, 32(2):205-218. This paper proposes to use the LOFAR spectral preprocessing method to retain key features. In order to address the problems of scarce datasets and non-standard data formats, generative adversarial networks are used to expand the samples to improve classification performance, which significantly improves the classification accuracy of the model.3. Deep learning method for underwater target recognition based on channel attention mechanism: Xue Lingzhi, Zeng Xiangyang, Jin Anqi. A Novel Deep-Learning Method with Channel Attention Mechanism for Underwater Target Recognition. Sensors, 2022, 22(15): 5492-5511. This paper proposes a residual network structure that introduces channel attention mechanism to enhance the stable spectral feature energy of residual convolution, realize the deep abstract spectral feature extraction of the target, and use the feature classification method based on one-dimensional convolution to complete the underwater target recognition, with good recognition effect. 4. Self-supervised underwater target recognition method based on acoustic embedding spatial encoder: Wang X, Meng J, Liu Y, et al. Self-supervised acoustic representation learning via acoustic-embedding memory unit modified space autoencoder for underwater target recognition. The Journal of the Acoustical Society of America, 2022, 152(5):2905-2915. This paper proposes to unify FBank features and GBank features in generative self-supervised learning, extract new acoustic features for underwater target recognition, and significantly improve the recognition accuracy and robustness. 5. Underwater Target Recognition Model Based on Convolutional Neural Network and Attention Mechanism: Jin Anqi, Zeng Xiangyang. A Novel Deep Learning Method for Underwater Target Recognition Based on Res-Dense Convolutional Neural Network with Attention Mechanism. Journal of Marine Science and Engineering, 2023, 11(1): 69-89. This paper proposes an underwater target recognition model based on convolutional neural network and attention mechanism. It combines residual neural network and densely connected convolutional neural network and introduces channel and temporal attention mechanism. Experimental results show that the proposed model has good recognition effect of marine ship radiated noise. Summary of the Invention

[0003] The purpose of this invention is to provide an underwater target recognition method based on dual-channel self-supervised acoustic feature learning that focuses on abstract semantic information and possesses high accuracy and robustness.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] (1) A model for constructing a dual-channel self-attention audio encoder (DSAE) is proposed;

[0006] (1.1) The features of Mel filter-bank (FBank) with rich low-frequency information and the features of Gammatone filter-bank (GBank) that focus on high-frequency signals are unified in contrastive self-supervised feature learning.

[0007] (1.2) The underwater target feature extraction module utilizes a local self-attention mechanism to better extract local feature semantic information;

[0008] (2) A dual-channel self-attention audio encoder with dynamic positive sample memory module (DSAE-DMM) model is proposed to construct a dynamic positive sample memory module.

[0009] (2.1) Based on the DSAE model, a data augmentation strategy of time-domain-frequency-domain enhancement is used to increase the diversity of data samples and improve robustness;

[0010] (2.2) A positive and negative sample balancing strategy based on the Dynamic positive sample Memory Module (DMM) is adopted to store the historical spatiotemporal features of anchor samples as positive sample augmentation to balance the number of positive and negative samples.

[0011] (3) Complete the underwater target recognition method based on DSAE-DMM;

[0012] (3.1) The target encoder obtained by self-supervised training is used for downstream tasks to extract the features of the dual-channel self-attention audio encoder with dynamic positive sample memory module spectrogram (DSAE-DMM Spec) of underwater target dynamic positive sample storage.

[0013] (3.2) The DSAE-DMM Spec features are used as inputs to the Multi-Layer Perceptron (MLP) model and the Multinomial Logistic Regression (MLR) model, respectively. The underwater target recognition task is completed by processing the recognition task through the multi-class target recognition model.

[0014] The aforementioned step (1.1) specifically includes:

[0015] After pre-emphasis, framing, windowing, and fast Fourier transform of the original audio data, the data is filtered using Mel filter bank and gamma filter bank respectively, and the logarithmic energy is calculated to obtain FBank features and GBank features.

[0016] The aforementioned step (1.1) specifically includes:

[0017] Using FBank and GBank features as dual-channel audio features, a DSAE is constructed, which includes a target encoder module, a query encoder module, a projection head module, and a dynamic negative sample storage queue.

[0018] Step (1.2) specifically includes:

[0019] The target encoder and query encoder employ an underwater target feature extraction module with a local self-attention mechanism, focusing on local features in the time-frequency domain and increasing the network's information selection ability to effectively extract features.

[0020] Step (1.2) specifically includes:

[0021] The target encoder extracts features from the FBank, and the query encoder extracts features from the GBank. For input x... i Get the output r i =f(x) i ) and r i ′=f′(x i The target encoder is denoted as f(·), and the query encoder is denoted as f′(·). The Projection Head module is used to extract features r. i and r i Dimensionality reduction is performed to obtain the embedding vector h. i and h i ′, represented as h i =W (2) σ(W (1) r i ) and h i ′=W (2)σ(W (1) r i ′), W is the weight matrix, and σ is the ReLU activation function.

[0022] Step (1.2) specifically includes:

[0023] Negative samples are stored using a dynamic negative sample queue, which is implemented using a first-in-first-out queue and is dynamically updated as the model is trained.

[0024] Step (2.1) specifically includes:

[0025] Data augmentation strategies utilizing time-domain and frequency-domain enhancement increase data sample diversity and improve robustness. Time-domain enhancement is represented as... The signal at time t is represented as X(t), where t1 and t2 represent the start and end times of the data augmentation in the time domain, and the frequency domain augmentation is represented as... The signal in frequency band f is represented as X(f), and f1 and f2 represent the start and end frequency domains of data enhancement.

[0026] Step (2.2) specifically includes:

[0027] A positive-negative sample balancing strategy based on Direct Modeling (DMM) is employed to augment the number of positive samples and balance the ratio of positive to negative samples, enabling the model to learn more accurate feature representations. The number of input samples in an epoch is denoted as N, the length of the dynamic queue is i, and the size of the embedding vector obtained through the target encoder is j. Initially, the DMM is an empty N*i*j matrix. In the first epoch, the input sample x obtains an embedding vector h1 through the target encoder. Before each iteration, h1 is stored in the dynamic queue corresponding to the input sample x until i historical embedding vectors are stored, at which point the queue is full. In the (i+1)th epoch, the embedding vector h1 of sample x is... i+1 The history of the head of the queue is included at the tail, and the history of the head of the queue is embedded into the vector and dequeued to achieve iterative dynamic updates.

[0028] Step (3) specifically includes:

[0029] The trained target encoder is used for downstream recognition tasks to extract DSAE-DMM Spec features. The recognition model adopts MLP and MLR models, and is mapped and transformed through the normalized exponential (Softmax) function to finally obtain the recognition result and complete the underwater target recognition task.

[0030] The beneficial effects of this invention are as follows:

[0031] Due to the difficulty in labeling and recognizing underwater datasets, this invention builds upon the feature adversarial enhancement underwater target recognition method proposed by Wang Xingmei et al. in 2022 using acoustic embedding memory space encoders. It combines SwinTransformer and momentum contrastive unsupervised visual representation learning to propose an underwater target recognition method based on dual-channel self-supervised acoustic feature learning. This invention proposes constructing a DSAE contrastive self-supervised feature learning model, which, compared to generative self-supervised feature learning methods, focuses on more abstract latent semantic information. The spectrogram features of the Mel filter bank, rich in low-frequency information, and the spectrogram features of the gamma filter bank, focusing on high-frequency signals, are unified in the contrastive self-supervised feature learning, enabling the encoder to learn high-level semantic features that combine the advantages of different channel features. Furthermore, to enhance the network's information selection ability, an underwater target feature extraction module using a local self-attention mechanism is employed, focusing on local features to better extract semantic information. This invention proposes to construct a DSAE-DMM model, introduce data augmentation strategy and positive and negative sample balancing strategy. The data augmentation strategy of time domain-frequency domain augmentation is used to increase the diversity of data samples and improve robustness. The positive and negative sample balancing strategy based on DMM is adopted to store the historical spatiotemporal embedding vectors as positive samples to achieve dynamic updates and balance the ratio of positive and negative samples, so that the model learns more accurate feature representations and improves recognition accuracy. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the method of the present invention;

[0034] Figure 2 This is a structural diagram of the DSAE model proposed in this invention;

[0035] Figure 3 This is a schematic diagram of the window attention mechanism and the sliding window attention mechanism. Figure 3 (a) is the spectrum after block division. Figure 3 (b) is a window partitioning diagram of the window attention mechanism. Figure 3 (c) is the window distribution diagram before sliding. Figure 3 (d) is the window distribution diagram after sliding. Figure 3 (e) is a window distribution diagram after sliding and shifting;

[0036] Figure 4 This is a structural diagram of the DSAE-DMM model proposed in this invention;

[0037] Figure 5 This is a diagram illustrating data augmentation strategies. Figure 5 (a) is a schematic diagram of the time-domain enhancement of FBank features. Figure 5 (b) is a schematic diagram of the frequency domain enhancement of FBank features. Figure 5 (c) is a schematic diagram of FBank feature mixing in the time-frequency domain. Figure 5 (d) is a schematic diagram of GBank feature enhancement in the time domain. Figure 5 (e) is a schematic diagram of frequency domain enhancement for GBank features. Figure 5 (f) is a schematic diagram of GBank feature mixing time-frequency domain enhancement;

[0038] Figure 6 This is a DMM update strategy diagram;

[0039] Figure 7 This is a map showing the recognition accuracy of various features on the MLP model in the actual collected underwater acoustic signal target dataset and the ShipsEar dataset. Figure 7 (a) is a graph showing the recognition accuracy of various features on the MLP model in the actual collected underwater acoustic signal target dataset. Figure 7 (b) is a graph showing the recognition accuracy of various features on the MLP model in the ShipsEar dataset;

[0040] Figure 8 This is a map showing the recognition accuracy of various features on the MLR model in the actual collected underwater acoustic signal target dataset and the ShipsEar dataset. Figure 8 (a) is a graph showing the recognition accuracy of various features on the MLR model in the actual collected underwater acoustic signal target dataset. Figure 8 (b) is a graph showing the recognition accuracy of various features on the MLR model in the ShipsEar dataset;

[0041] Figure 9 This is a comparison chart of feature recognition accuracy on MLP models with noisy and clean datasets. Figure 9 (a) is a comparison of the target recognition accuracy of the MLP model on noisy / clean underwater acoustic signal target datasets. Figure 9 (b) is a comparison of the recognition accuracy of the ShipsEar dataset with and without noise on the MLP model;

[0042] Figure 10 This is a comparison chart of feature recognition accuracy on noisy and clean datasets using the MLR model. Figure 10 (a) is a comparison of the target recognition accuracy of the MLR model on noisy / clean underwater acoustic signal target datasets. Figure 10 (b) is a comparison of the recognition accuracy of the ShipsEar dataset with and without noise on the MLR model;

[0043] Figure 11 This is a diagram illustrating audio noise addition. Figure 11 (a) is a schematic diagram of the clean audio waveform in the dataset. Figure 11 (b) is a schematic diagram of a noisy frequency waveform with a signal-to-noise ratio of -10dB. Figure 11 (c) is a schematic diagram of a noisy frequency waveform with a signal-to-noise ratio of -5dB. Figure 11 (d) is a schematic diagram of a noisy frequency waveform with a signal-to-noise ratio of 0dB. Figure 11 (e) is a schematic diagram of a noisy frequency waveform with a signal-to-noise ratio of 10dB;

[0044] Figure 12 This is a line graph showing the feature recognition accuracy under different noise intensities. Figure 12 (a) is a line graph showing the feature recognition accuracy of the MLP model. Figure 12 (b) is a line graph showing the feature recognition accuracy of the MLR model. Detailed Implementation

[0045] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Combination Figure 1 The specific steps of this invention are as follows:

[0048] (1) Propose constructing the DSAE model

[0049] This paper unifies FBank features, which possess rich low-frequency information, and GBank features, which focus on high-frequency signals, in a contrastive self-supervised feature learning model (DSAE). The underwater target feature extraction module, employing a local self-attention mechanism, serves as the encoder. The model integrates FBank features with rich detail and low-frequency signal information with GBank features possessing strong robustness and high-frequency signal resolution capabilities within the contrastive self-supervised feature learning process. Positive and negative sample pairs are constructed using FBank and GBank features, and a contrastive learning approach enables the encoder to learn hybrid semantic features that combine the advantages of both FBank and GBank features. Figure 2 This is a structural diagram of the DSAE model proposed in this invention.

[0050] (1.1) Preprocessing the raw audio data

[0051] This invention uses a real-world collected underwater acoustic signal target dataset and the ShipsEar ship noise dataset for self-supervised auxiliary task training and supervised downstream task training. Two-thirds of the data samples from the noise-removed underwater acoustic signal target dataset are randomly selected for self-supervised auxiliary task training, and one-third are used for supervised downstream task training, with a training to testing data ratio of 7:3. The allocation method for the ShipsEar ship noise dataset is consistent with that of the underwater acoustic signal target dataset.

[0052] FBank and GBank features are extracted from the underwater acoustic signal. Pre-emphasis, framing, windowing, and short-time Fourier transform are performed. The signals are then filtered by Mel and gamma filters, respectively, and logarithmic filter banks are used for computation to obtain the FBank and GBank features of the original underwater acoustic signal.

[0053] (1.2) Underwater target feature extraction module with local self-attention mechanism

[0054] An underwater target feature extraction module employing a local self-attention mechanism serves as both the target encoder and query encoder to extract embedding vectors from FBank and GBank features. The FBank or GBank feature spectrograms of the underwater target are divided into blocks, and a high-dimensional vector representation is obtained through linear embedding, which is then input into the Swin-Transformer block. The Swin-Transformed block employs both Window-Multi-head Self-Attention (W-MSA) and Shifted Window-Mutialhead Self-Attention (SW-MSA) mechanisms. The computation of two consecutive Swin-Transformer blocks is as follows:

[0055]

[0056]

[0057]

[0058]

[0059] In the formula, This represents the output characteristic of the l-th Swing-Transformer block after W-MSA, z l express The output characteristics, Let z represent the (l+1)th SW-MSA output feature. l+1 express The output characteristics.

[0060] Figure 3 This is a schematic diagram of the window attention mechanism and the sliding window attention mechanism. Figure 3 (a) is the spectrum after block division. Figure 3 (b) is a window partitioning diagram of the window attention mechanism. Figure 3 (c) is the window distribution diagram before sliding. Figure 3 (d) is the window distribution diagram after sliding. Figure 3 (e) shows the window distribution after sliding and shifting. Sliding the window allows local attention to include global correlation information, and after shifting, a masking mechanism avoids information interaction between non-adjacent spectral blocks. Specifically, without changing the window size, the feature map is shifted M / 2 to the right and downwards respectively. This reduces computation through the window's local attention mechanism while introducing information interaction between non-adjacent windows, preventing information fragmentation between different windows and increasing the correlation between different regions. To avoid the increased computational load caused by increasing the window size, a cyclic shifting method is used. Figure 3 (e) Regions A, B, and C shown in the upper left corner are shifted to regions A, B, and C in the lower right corner, resulting in four new complete windows. During the self-attention mechanism computation within each window, the shifted spectral blocks are masked to avoid semantic errors that might arise from the shifting of non-adjacent windows. Furthermore, the underwater target feature extraction module of the local self-attention mechanism uses a similar hierarchical structure, setting different factors in the feature downsampling dimension to adjust the number of channels and achieve hierarchical feature extraction. At the end of each Swintransformer block, a patch merging layer is added to reduce the sequence size, resulting in the underwater target feature vector h.

[0061] (2) Propose constructing the DSAE-DMM model

[0062] To improve the recognition accuracy and robustness of the DSAE model in downstream tasks, a DSAE-DMM model is proposed. Based on the DSAE model, data augmentation strategies are introduced, along with a positive-negative sample balancing strategy. The time-domain to frequency-domain augmentation strategy increases data sample diversity and improves robustness. A DMM-based positive-negative sample balancing strategy amplifies historical spatiotemporal embedding vectors as positive samples and dynamically updates them, enabling the model to better learn long-term dependencies between data and balance the positive-negative sample ratio to improve recognition accuracy. Figure 4 This is a structural diagram of the DSAE-DMM model proposed in this invention.

[0063] (2.1) Data Augmentation Strategies

[0064] This invention introduces a time-domain-frequency domain augmentation strategy. By adding time-domain-frequency enhancement to FBank and GBank features, it simulates underwater noise interference and information distortion, increasing the learning difficulty of the model and helping the encoder learn more effective features. The time-domain-frequency domain augmentation strategy consists of time-domain masks and frequency-domain masks. The time-domain mask simulates acoustic signal loss caused by signal acquisition equipment factors, improving the model's robustness to noise interference and information loss, allowing the model to better adapt to the real underwater environment. The frequency-domain mask simulates common noise and data distortion conditions, increasing data diversity and improving the model's learning ability.

[0065] Time-domain masking simulates the loss or attenuation of sound signals by masking the time-domain waveform. Examples include signal loss due to limitations in signal acquisition equipment or information loss due to packet loss during transmission. Time-domain masking is implemented using a rectangular window function to mask a portion of the time domain. This is achieved by setting the sampling points of the signal to 0 within a continuous time step. If the signal's time domain is represented as X(t), then masking over time [t1, t2] is specifically expressed as follows:

[0066]

[0067] Due to the signal attenuation characteristics of underwater media and the high complexity of the underwater environment, frequency domain masks are used to mask or suppress signals in different frequency domains to simulate noise interference in certain frequency bands. Masking some high-frequency signals in the frequency domain can simulate the attenuation characteristics of underwater signals, while masking low-frequency signals can simulate the influence of background noise. A rectangular window function is used to mask the entire frequency band within a certain frequency range. Frequency domain masking is achieved by setting the amplitude of a certain interval on the frequency axis to 0. The frequency domain representation of the signal is X(f). Then, bandpass filtering is performed on the frequency [f1, f2], specifically as follows:

[0068]

[0069] A data augmentation strategy combining random time-domain and frequency-domain masks is applied to FBank and GBank features to obtain enhanced feature spectrograms, increasing the learning difficulty of the model and improving the performance and robustness of underwater target identification tasks. Figure 5 This is a diagram illustrating data augmentation strategies. Figure 5 (a) is a schematic diagram of the time-domain enhancement of FBank features. Figure 5 (b) is a schematic diagram of the frequency domain enhancement of FBank features. Figure 5 (c) is a schematic diagram of FBank feature mixing in the time-frequency domain. Figure 5 (d) is a schematic diagram of GBank feature enhancement in the time domain. Figure 5(e) is a schematic diagram of frequency domain enhancement for GBank features. Figure 5 (f) is a schematic diagram of GBank feature mixing time-domain-frequency domain enhancement.

[0070] (2.2) Positive and negative sample balancing strategy

[0071] This invention employs a positive-negative sample balancing strategy based on Dynamic Model (DMM). It stores historical embedding vectors to provide abundant positive sample pairs for anchor samples, balancing the number of positive and negative samples to further improve recognition accuracy. As the model trains and parameters are updated, the embedding vectors learned by anchor samples in historical iterations exhibit a slow shift, approximating the features extracted by the current encoder. To enable the model to focus more on positive sample features, a DMM is constructed to store historical embedding vectors as positive sample expansion. The DMM consists of several fixed-size dynamic positive sample queues. Each queue contains the historical embedding vector of a sample over the past i epochs; that is, the embedding vector obtained by the anchor sample in each epoch is stored in the corresponding dynamic positive sample queue. In a new epoch, the dynamic positive sample queue corresponding to the anchor sample contains its historical embedding vector, which is treated as a positive sample and used together with negative samples to train the model. This method allows the model to collect positive sample pairs across multiple historical time periods, balancing the number of positive and negative samples. The DMM stores the embedding vectors of the same anchor point in past time periods in the form of dynamic queues. In each iteration, the embedding vector obtained by the anchor sample is updated by dequeueing and enqueueing. Figure 6 This is a DMM update strategy graph. The target encoder is denoted as f(·), a given input sample as x, the number of input samples in one epoch as N, the length of the dynamic queue as i, and the size of the embedding vector obtained through the target encoder as j. Initially, the DMM is an empty N*i*j matrix. In the first epoch, the input sample x obtains an embedding vector h1 through the target encoder. h1 is stored in the dynamic queue corresponding to the input sample x until i historical embedding vectors are stored, at which point the queue is full. In the (i+1)th epoch, the embedding vector h1 of sample x is... i+1 The historical embedding vector at the head of the queue is dequeued to achieve dynamic updates of the DMM after each iteration. In a new epoch, the historical embedding vector stored in the dynamic queue corresponding to the sample is taken as the positive sample.

[0072] (2.3) Loss Function

[0073] During model training, the feature similarity between each anchor sample and both positive and negative samples is calculated. The similarity of negative sample pairs is maximized, while the similarity of positive sample pairs is minimized. This optimizes the encoder, enabling it to learn hybrid semantic features that combine the advantages of FBank and GBank features. This yields acoustic features with better robustness and accuracy in underwater target recognition tasks. Multiple positive and negative sample pairs are constructed for model loss calculation, allowing the model to learn the target's feature representation from unlabeled data. The loss function is expressed as:

[0074]

[0075] In the formula, x i ,x j y represents the embedding vector, and y represents the category label.

[0076] (3) Complete the underwater target recognition task based on DSAE-DMM

[0077] This invention completes the underwater target recognition task in the downstream stage, retaining the target encoder module of the model as the feature extraction part of the downstream underwater target recognition task. The FBank spectrogram obtained from underwater acoustic signal data is input into the downstream task to obtain DSAE-DMM Spec features. MLP and MLR models are selected as the recognition models. Classification accuracy (Acc), Classifier Loss Convergence Rate (CLCR), and Anti-Noise Robust (ANR) are used as evaluation metrics.

[0078] To verify the algorithm's recognition accuracy in underwater target identification tasks, Acc is analyzed and calculated as follows:

[0079]

[0080] Where f(·) represents the classification model, D_Test represents the test dataset, and label represents the sample label.

[0081] To verify that the DSAE-DMM Spec features learned in this invention possess good semantics, CLCR is introduced as one of the performance evaluation metrics. Training is performed on different feature samples of the same order of magnitude, and the number of epochs required for the classification model to converge on the training dataset D_Train is evaluated. This compares the modeling and learning performance of different features on the same training dataset, thus assessing the performance of each feature. CLCR is defined as:

[0082] CLCR(f,D_Train)=epochs(loss=loss_min)

[0083] An ANR (Advanced Subjective Recognition) metric is introduced to evaluate robustness. Noise is fused into the dataset as a test set, and various features are extracted and learned based on this. Underwater target recognition tasks are performed on MLP and MLR models. By comparing the noisy test set Acc, the ANR metric is analyzed, and the performance of the underwater target recognition algorithm in noisy environments is evaluated. For the recognition network f, and the noisy test sample set DN_Test (sample set size N) and sample labels, ANR is defined as:

[0084]

[0085] Table 1 presents a comparison of target recognition accuracy for the actual underwater acoustic signal target dataset and the ShipsEar dataset, using MFCC, GFCC, FBank, GBank, SAE Spec, ASAE Spec, DSAE Spec, and DSAE-DMMSpec features, with a training set to test set ratio of 7:3. In the actual underwater acoustic signal target dataset, the DSAE-DMMSpec feature achieved the best recognition accuracy on the MLP model, improving upon DSAE Spec by 0.69%. On the MLR recognition model, the accuracy improved from 90.58% to 91.51%, and it achieved the best recognition accuracy compared to features such as ASAE Spec. The recognition accuracy performance on the ShipsEar dataset is similar to that of the actual underwater acoustic signal target dataset. Under the MLP recognition model, the DSAE-DMMSpec feature achieved a recognition accuracy of 88.64%, an improvement of 0.83% compared to DSAE, from 87.81%. Under the MLR recognition model, the DSAE-DMMSpec feature achieved a recognition accuracy of 85.93%. Achieving optimal recognition accuracy. DSAE-DMM Spec features demonstrate higher recognition accuracy than other features on both the actual collected underwater acoustic signal target datasets and the ShipsEar dataset. The combined performance of these two models on underwater target recognition tasks proves the effectiveness of the DSAE-DMM Spec features extracted in this invention; that is, the DSAE-DMM-based underwater target recognition method has better recognition accuracy. Figure 7 This is a map showing the recognition accuracy of various features on the MLP model in the actual collected underwater acoustic signal target dataset and the ShipsEar dataset. Figure 7 (a) is a graph showing the recognition accuracy of various features on the MLP model in the actual collected underwater acoustic signal target dataset. Figure 7 (b) is a graph showing the recognition accuracy of various features on the ShipsEar dataset in the MLP model. Figure 8This is a map showing the recognition accuracy of various features on the MLR model in the actual collected underwater acoustic signal target dataset and the ShipsEar dataset. Figure 8 (a) is a graph showing the recognition accuracy of various features on the MLR model in the actual collected underwater acoustic signal target dataset. Figure 8 (b) is a graph showing the recognition accuracy of various features on the MLR model in the ShipsEar dataset.

[0086] Table 1 Comparison of Recognition Accuracy for Various Features

[0087]

[0088] To demonstrate that the DSAE-DMM Spec features obtained by the proposed DSAE-DMM algorithm contain higher-level semantic information than other comparative features, comparative experiments on the CLCR performance of all features were conducted on both MLP and MLR models. Table 2 shows the CLCR of different acoustic features at model convergence. The proposed DSAE-DMM Spec shows the best CLCR performance on both the MLP and MLR models. On the MLP model, DSAE-DMM Spec converges after 17 epochs on the actual collected underwater acoustic signal target dataset and after 25 epochs on the ShipsEar dataset, showing a certain improvement in convergence speed compared to mainstream features such as MFCC. It also achieves the best convergence speed compared to SAE Spec and ASAE Spec features obtained by self-supervised learning methods. In the MLR recognition model, the performance of each feature is similar to that on the MLP model, exhibiting a fast convergence speed, further demonstrating the effectiveness of DSAE-DMM Spec features in underwater target recognition tasks. This demonstrates that DSAE-DMM Spec features are more adaptable to underwater target recognition tasks, and are audio features that contain more comprehensive semantics and have greater recognition advantages, further proving the effectiveness of the method proposed in this invention.

[0089] Table 2. CLCR at model convergence for different acoustic features.

[0090]

[0091] To verify the robustness of DSAE-DMM Spec features, the features were validated on noisy test data, and the accuracy loss was calculated. To visually demonstrate the accuracy changes of each feature and analyze its noise resistance robustness, Figure 9 This is a comparison chart of feature recognition accuracy on MLP models with noisy and clean datasets. Figure 9 (a) is a comparison of the target recognition accuracy of the MLP model on noisy / clean underwater acoustic signal target datasets. Figure 9(b) is a comparison of the recognition accuracy of the MLP model on the noisy / clean ShipsEar dataset. Figure 10 This is a comparison chart of feature recognition accuracy on noisy and clean datasets using the MLR model. Figure 10 (a) is a comparison of the target recognition accuracy of the MLR model on noisy / clean underwater acoustic signal target datasets. Figure 10 (b) is a comparison of the recognition accuracy of the MLR model on the noisy / clean ShipsEar dataset. On the MLP model, the DSAE-DMM Spec proposed in this invention suffers a 3.11% accuracy loss under noise interference on the actual collected underwater acoustic signal target dataset, which is an improvement of 1.02% compared to ASAE Spec features and 0.71% compared to DSAE Spec features. On the ShipsEar dataset, the accuracy loss is 1.59%, which is an improvement of 0.63% compared to ASAE Spec features and 0.3% compared to DSAE Spec features. The accuracy loss trend on the MLR recognition model is similar to that of the MLP; in both the actual collected underwater acoustic signal target dataset and the ShipsEar dataset, the recognition accuracy of various features suffers to some extent. MFCC features suffer the greatest loss in noisy datasets. GFCC features, which have lower recognition accuracy, have a slightly better accuracy loss than MFCC features. FBank and GBank features show improved robustness against noise than MFCC and GFCC features. This demonstrates that the DSAE-DMM Spec, which incorporates data augmentation and positive / negative sample balancing strategies, is more robust and exhibits good noise resistance even with added noise.

[0092] To further verify the robustness of the extracted features in complex underwater environments, Gaussian white noise of varying intensities was added to the test samples in the actual collected underwater acoustic signal target dataset to simulate real underwater environments under different noise intensities. The stability and noise resistance of the extracted features in real underwater environments were then analyzed. Four different noise intensities (-10dB, -5dB, 0dB, and 10dB) were selected to contaminate the actual collected underwater acoustic signal target dataset, thus creating a noisy dataset.

[0093] Figure 11 This is a diagram illustrating audio noise addition. Figure 11 (a) is a schematic diagram of the clean audio waveform in the dataset. Figure 11 (b) is a schematic diagram of a noisy frequency waveform with a signal-to-noise ratio of -10dB. Figure 11 (c) is a schematic diagram of a noisy frequency waveform with a signal-to-noise ratio of -5dB. Figure 11 (d) is a schematic diagram of a noisy frequency waveform with a signal-to-noise ratio of 0dB. Figure 11(e) is a schematic diagram of the noisy frequency waveform with a signal-to-noise ratio of 10dB. The underwater target recognition models for each feature were tested using a noisy dataset to analyze the noise resistance of each model. Figure 12 This is a line graph showing the feature recognition accuracy under different noise intensities. Figure 12 (a) is a line graph showing the feature recognition accuracy of the MLP model. Figure 12 (b) is a line graph showing the recognition accuracy of each feature on the MLR model. The experimental results show that the DSAE-DMMSpec features exhibit better noise resistance in both high and low signal-to-noise ratio (SNR) noisy data. Although the model recognition accuracy decreases with increasing noise intensity, the loss is smaller compared to other features, further verifying that the underwater target recognition method based on DSAE-DMM proposed in this invention has strong robustness and applicability in underwater target recognition tasks.

[0094] In summary, the underwater target recognition method based on dual-channel self-supervised acoustic feature learning proposed in this invention can effectively complete the underwater target recognition task and has good recognition accuracy and robustness.

[0095] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An underwater target recognition method based on dual-channel self-supervised acoustic feature learning, characterized in that, include: The contrastive self-supervised feature learning refers to a self-supervised method that learns feature representations by comparing positive and negative sample pairs. (1) A dual-channel self-attention audio encoder, DSAE model, is proposed; (1.1) The features of Mel filter bank spectrograms with rich low-frequency information, FBank and Gammatone filter bank spectrograms with high-frequency signals are unified in contrastive self-supervised feature learning. (1.2) The underwater target feature extraction module utilizes a local self-attention mechanism to better extract local feature semantic information; (2) A dual-channel self-attention audio encoder with dynamic positive sample memory module, DSAE-DMM model, is proposed; (2.1) Based on the DSAE model, a data augmentation strategy of time-domain-frequency-domain enhancement is used to increase the diversity of data samples and improve robustness; (2.2) The positive and negative sample balancing strategy of DMM based on Dynamic positive sample Memory Module is adopted, which stores the historical spatiotemporal features of anchor samples as positive sample augmentation to balance the number of positive and negative samples. (3) Complete the underwater target recognition method based on DSAE-DMM; (3.1) The target encoder obtained by self-supervised training is used for downstream tasks to extract the spectrogram of the dual-channel self-attention audio encoder with dynamic positive sample memory module (DSAE-DMM) for underwater targets. (3.2) The DSAE-DMM Spec features are used as inputs to the Multilayer Perceptron (MLP) model and the Multiclass Logistic Regression (MLR) model, respectively. The underwater target recognition task is completed by processing the recognition task through the multiclass target recognition model.

2. The underwater target recognition method based on dual-channel self-supervised acoustic feature learning according to claim 1, characterized in that, The aforementioned step (1.1) specifically includes: After pre-emphasis, framing, windowing, and fast Fourier transform of the original audio data, the data is filtered using Mel filter bank and gamma filter bank respectively, and the logarithmic energy is calculated to obtain FBank features and GBank features.

3. The underwater target recognition method based on dual-channel self-supervised acoustic feature learning according to claim 1, characterized in that, The aforementioned step (1.1) specifically includes: Using FBank and GBank features as dual-channel audio features, a DSAE is constructed, which includes a target encoder module, a query encoder module, a projection head module, and a dynamic negative sample storage queue.

4. The underwater target recognition method based on dual-channel self-supervised acoustic feature learning according to claim 1, characterized in that, Step (1.2) specifically includes: The target encoder and query encoder employ an underwater target feature extraction module with a local self-attention mechanism, focusing on local features in the time-frequency domain and increasing the network's information selection ability to effectively extract features.

5. The underwater target recognition method based on dual-channel self-supervised acoustic feature learning according to claim 1, characterized in that, Step (1.2) specifically includes: The target encoder extracts features from the FBank, and the query encoder extracts features from the GBank for input x. i Get the output r i =f(x) i ) and r i ′=f′(x i The target encoder is denoted as f(·), and the query encoder is denoted as f′(·). The Projection Head module is used to extract features r. i and r i Dimensionality reduction is performed to obtain the embedding vector h. i and h i ′, represented as h i =W (2) σ(W (1) r i ) and h i ′=W (2) σ(W (1) r i ′), W is the weight matrix, and σ is the ReLU activation function.

6. The underwater target recognition method based on dual-channel self-supervised acoustic feature learning according to claim 1, characterized in that, Step (1.2) specifically includes: Negative samples are stored using a dynamic negative sample queue, which is implemented using a first-in-first-out queue and is dynamically updated as the model is trained.

7. The underwater target recognition method based on dual-channel self-supervised acoustic feature learning according to claim 1, characterized in that, Step (2.1) specifically includes: Data augmentation strategies utilizing time-domain and frequency-domain enhancement increase data sample diversity and improve robustness. Time-domain enhancement is represented as... The signal at time t is represented as X(t), where t1 and t2 represent the start and end times of the data augmentation in the time domain, and the frequency domain augmentation is represented as... The signal in frequency band f is represented as X(f), and f1 and f2 represent the start and end frequency domains of data enhancement.

8. The underwater target recognition method based on dual-channel self-supervised acoustic feature learning according to claim 1, characterized in that, Step (2.2) specifically includes: A positive-negative sample balancing strategy based on DMM is adopted to augment the number of positive samples and balance the ratio of positive to negative samples, enabling the model to learn more accurate feature representations. The number of input samples in an epoch is denoted as N, the length of the dynamic queue is i, and the size of the embedding vector obtained by the target encoder is j. Initially, the DMM is an empty N*i*j matrix. In the first epoch, the input sample x obtains an embedding vector h1 through the target encoder. Before each iteration, h1 is stored in the dynamic queue corresponding to the input sample x until i historical embedding vectors are stored, at which point the queue is full. In the (i+1)th epoch, the embedding vector h1 of sample x is... i+1 The history of the head of the queue is included at the tail, and the history of the head of the queue is embedded into the vector and dequeued to achieve iterative dynamic updates.

9. The underwater target recognition method based on dual-channel self-supervised acoustic feature learning according to claim 1, characterized in that, Step (3) specifically includes: The trained target encoder is used for downstream recognition tasks to extract DSAE-DMM Spec features. The recognition model adopts MLP and MLR models, and is mapped and transformed by the normalized exponential Softmax function to finally obtain the recognition results and complete the underwater target recognition task.

Images