A feature-based adversarial enhanced underwater target recognition method based on acoustic embedding memory spatial encoder model

Abstract

The present invention belongs to the technical field of underwater target recognition, and specifically relates to a feature-adversarial-enhanced underwater target recognition method based on an acoustic embedding memory spatial encoder model. The present invention completes the spatial transformation reconstruction from FBank features to GBank features by constructing a self-supervised SAE model, and learns SAE Spec features with good voiceprint characteristics and anti-noise robustness; through a negative sample mining strategy based on the AEMU module, a dynamic queue dictionary is used to dynamically update and store negative samples, thereby improving the efficiency of negative sample learning; the AEMU module is combined with the SAE model using an improved CE Loss function to construct an AEMU‑SAE model, so that feature transformation reconstruction and feature negative sample learning are unified in the AEMU‑SAE model, ensuring that the AEMU‑SAE model can learn ASAE Spec features containing high-level semantic information during the feature-adversarial-enhancement process, and the ASAE Spec features are used as inputs of the MLP model and the MLR model respectively, and the recognition task is processed by a multi-classification target recognition model, and finally the underwater target recognition is completed.

Description

Technical Field

[0001] The invention belongs to the technical field of underwater target recognition, and in particular relates to a feature-adversarial enhanced underwater target recognition method based on an acoustic embedding memory space encoder model. Background Art

[0002] In recent years, underwater target recognition technology based on acoustic signals has developed rapidly and plays a very critical role in various fields. However, there are still many problems. On the one hand, the complexity of underwater environmental acoustic signals is much greater than that of land environments, which increases the labeling cost of high-quality marine target samples; on the other hand, the uneven fluctuations of water bodies and terrain cause underwater acoustic signals to undergo waveform distortion, random fluctuations, signal attenuation, and other phenomena during long-distance propagation in the ocean, making the generalization ability of single feature classes extracted for target samples weak, and many target recognition algorithms suitable for land acoustic signals cannot be applied to underwater environments. Therefore, domestic and foreign scholars have conducted in-depth analysis and research on underwater target recognition and achieved relevant breakthroughs. Among them, the most famous and effective underwater target recognition methods in the existing literature mainly include: 1. Research on underwater acoustic target recognition methods based on gamma frequency cepstral coefficients: 2018 Zhang W, Wu Y, Wang D, et al. Underwater target feature extraction and classification based on gammatone filter and machine learning. International Conference on Wavelet Analysis and Pattern Recognition, Chengdu, China. 2018, 95: 42-47. It is proposed to use gamma frequency cepstral coefficients as signal features for underwater target recognition, and comprehensively compare multiple features through 19 different machine learning classification models such as support vector machines and decision trees. Finally, it is proved that gamma frequency cepstral coefficients have more significant noise resistance and robustness. 2. Research on underwater target recognition method based on convolutional neural network and extreme learning machine: Hu G, Wang K, Peng Y, et al. Deep learning methods for underwater target feature extraction and recognition. Computational intelligence and neuroscience, 2018, 18(3): 1-10. A method for underwater acoustic data feature extraction and recognition based on convolutional neural network and extreme learning machine is proposed. This method effectively combines the depth and robustness of convolutional neural network feature extraction capabilities with the excellent classification ability of extreme learning machine, and finally achieves a recognition accuracy of 93.04%.3. Research on underwater target recognition method based on multi-dimensional fusion features and improved deep neural network: Wang X, Liu A, Zhang Y, et al. Underwater acoustic target recognition: acombination of multi-dimensional fusion features and modified deep neural network. Remote Sensing, 2019, 11(16): 1888-1890. It proposes to combine the multi-dimensional features of gamma frequency cepstral coefficients and improved empirical mode decomposition, and use Gaussian mixture model to improve deep neural network for underwater acoustic target recognition, which ultimately improves the accuracy of underwater target recognition tasks.

[0003] Self-supervised learning is a knowledge transfer mechanism. It mainly proposes an auxiliary task of automatic labeling. The auxiliary task is used to mine its own supervision information from large-scale unsupervised data. The network is trained through this constructed supervision information, so that it can learn representations that are valuable for downstream tasks. Self-supervised learning is widely used as a method to improve the feature learning ability of neural networks in scenarios where it is difficult to obtain high-quality samples. The most famous self-supervised learning methods in the existing literature mainly include: 1. Research on acoustic feature learning methods based on contrastive predictive coding algorithm: Oord A, Li Y, Vinyals O. Representation learning with contrastive predictive coding. arXiv, 2018, 7(3748): 2-3. It is proposed to use self-supervised methods to construct an auxiliary task of predicting future frames by randomly sampling past frames, and extract contrastive predictive coding through an autoregressive model. At the same time, the coding is used as information representation in speech, image, text and reinforcement learning, and finally achieves excellent performance. 2. Research on unrelated acoustic feature learning methods based on self-supervised multi-tasks: 2019 Pascual S, Ravanelli M, Serra J, et al. Learning problem-agnostic speech representations from multiple self-supervised tasks. arXiv, 2019, 4(3416): 1-12. It is proposed to use a convolutional coding network to learn acoustic features from the original acoustic signal, and use these features to build a self-supervised task training network model. Finally, the acoustic features learned through self-supervised tasks can improve the performance of speaker recognition, emotion classification and automatic speech recognition tasks. 3. Research on self-supervised acoustic feature learning methods based on bidirectional encoder networks: Wang W, Tang Q, Livescu K. Unsupervised pre-training of bidirectional speech encoders via masked reconstruction. IEEE International Conference on Acoustics, Speech and Signal Processing. Virtual Site. 2020, 156: 6889-6893. A method for pre-training speech representation through masked reconstruction loss was proposed, and a bidirectional encoder network was used for self-supervised reconstruction tasks. Finally, excellent speech recognition performance was demonstrated on the LibriSpeech and Wall Street Journal corpora. Summary of the invention

[0004] The purpose of the present invention is to provide a feature-adversarial enhanced underwater target recognition method based on an acoustic embedding memory space encoder model, which can better improve underwater target recognition accuracy, task adaptability and noise robustness.

[0005] A feature adversarial enhanced underwater target recognition method based on an acoustic embedding memory spatial encoder model comprises the following steps:

[0006] Step 1: Obtain the identified original underwater acoustic signal dataset, and extract FBank features and GBank features from the original underwater acoustic signals in the dataset;

[0007] After pre-emphasis, framing, windowing, and short-time Fourier transform of the original underwater acoustic signal, the FBank feature with good voiceprint distinguishing advantage and the GBank feature with anti-noise robustness advantage are extracted using the Mel filter bank and the gamma filter bank respectively.

[0008] FBank features are:

[0009]

[0010] Among them, s(n) is the frame information; N is the total number of frames; m is the feature order; L is the total feature order; M is the number of Mel filters;

[0011] GBank features:

[0012]

[0013] Where i is the frame number; E S (i) represents the frame energy spectrum information; I is the total number of frames; n is the feature order; Q is the number of gamma filters;

[0014] Step 2: Train the AEMU-SAE model using a dataset containing the FBank features and GBank features of the original underwater acoustic signal;

[0015] The AEMU-SAE model includes a random Gaussian noise embedding module, an encoder module, a decoder module, and an AEMU module. The specific training process is as follows:

[0016] Step 2.1: The random Gaussian noise embedding module generates random Gaussian noise during the training process. The FBank features of the original underwater acoustic signal are mixed with the random Gaussian noise and then input into the encoder module.

[0017] Step 2.2: The encoder module outputs the ASAE Spec features and inputs them into the decoder module;

[0018] Step 2.3: The decoder module reconstructs the output of the encoder module into pseudo GBank features through an upsampling process and inputs them into the AEMU module;

[0019] Step 2.4: The AEMU module executes the negative sample mining strategy and stores a large number of negative samples from the original data space in the form of a dynamic queue dictionary. During the training process, samples are directly taken from the dictionary by keyword query. In each iteration, a small batch of sample sets are updated in the dictionary by dequeuing and enqueuing, thereby improving the learning efficiency of negative samples.

[0020] The AEMU module calculates the spatial similarity loss between the pseudo-GBank features and the real GBank features to achieve the purpose of training the model;

[0021] The loss function adopts the improved CE Loss loss function, specifically:

[0022]

[0023] Among them, τ is the temperature hyperparameter coefficient; sim(·) represents the feature similarity function, that is, the CE Loss function; f(x i ) and f(x j ) represents a positive sample pair, i.e., the reconstructed pseudo-GBank feature and the homologous real GBank feature; f(x i ) and f(x k ) represents the negative sample pair, i.e., the reconstructed pseudo-GBank features and the non-homologous real GBank features; minimizing l means minimizing the mutual information distance of the positive sample pair and maximizing the mutual information distance of the negative sample pair, ensuring that the AEMU-SAE model learns the ASAE Spec features containing high-level semantic information;

[0024] Step 3: After FBank feature extraction, the original underwater acoustic signal data to be identified is input into the encoder module of the trained AEMU-SAE model to obtain the ASAE Spec features containing high-level semantic information; the ASAE Spec features are used as the input of the MLP model and the MLR model respectively, and the recognition task is processed by the multi-classification target recognition model to finally complete the underwater target recognition.

[0025] Furthermore, the random Gaussian noise generated by the random Gaussian noise embedding module in the step 2.1 during the training process is specifically:

[0026]

[0027] Among them, z is the noise information, μ is the noise expectation, and σ is the noise variance.

[0028] Furthermore, the feature similarity function sim(·) in step 2.4, i.e., the CE Loss function, is specifically:

[0029]

[0030] Among them, margin is the minimum threshold of feature similarity to prevent overfitting; x i and x j They represent the vector pair of the reconstructed pseudo-GBank features and the real GBank features after one-dimensional mapping expansion; y represents the automatically generated pseudo-labels of positive and negative samples. The calculated loss value can be regarded as the similarity between the reconstructed pseudo-GBank features and the real GBank features in the auditory space.

[0031] Furthermore, the training optimization algorithm of the AEMU-SAE model in step 2 adopts the Adam algorithm, which is specifically:

[0032]

[0033] in, is the bias-corrected first-order moment estimate; is the bias-corrected second-order moment estimate; ε is used to prevent division by zero errors during implementation; η is the learning rate, which is responsible for controlling the update rate of the weights.

[0034] The beneficial effects of the present invention are:

[0035] The present invention completes the spatial transformation reconstruction from FBank features to GBank features by constructing a self-supervised SAE model, learns SAE Spec features with good voiceprint characteristics and anti-noise robustness, and solves the problems of extremely high labeling cost of high-quality underwater target samples and weak generalization ability of extracted single feature classes in ocean scenes. The present invention proposes a negative sample mining strategy based on the AEMU module, and uses a dynamic queue dictionary to dynamically update and store negative samples, thereby improving the efficiency of negative sample learning. The present invention combines the AEMU module with the SAE model using an improved CE Loss function to construct an AEMU-SAE model, so that feature transformation reconstruction and feature negative sample learning are unified in the AEMU-SAE model, ensuring that the AEMU-SAE model can learn ASAE Spec features containing high-level semantic information in the feature adversarial enhancement process, and use the ASAE Spec features as inputs of the MLP model and the MLR model respectively, and process the recognition task through a multi-classification target recognition model, and finally complete underwater target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of the present invention.

[0037] Figure 2(a) to Figure 2(c) It is an example diagram of the spectrum of underwater mammals randomly selected from the actual collected underwater acoustic signal target data set. Figure 2(a) is an example diagram of the FBank feature spectrum, Figure 2(b) is an example diagram of the GBank feature spectrum, and Figure 2(c) is an example diagram of the original spectrum.

[0038] Figure 3 It is a structural diagram of the self-supervised SAE model proposed in the present invention.

[0039] Figure 4(a) to Figure 4(d) is a comparison diagram of ablation experiments on the random Gaussian noise embedding module of the SAE model, Figure 4(a) is a comparison diagram of ablation experiments on the MLP model for an actually collected underwater acoustic signal target dataset, Figure 4(b) is a comparison diagram of ablation experiments on the MLP model for the ShipsEar dataset, Figure 4(c) is a comparison diagram of ablation experiments on the MLR model for an actually collected underwater acoustic signal target dataset, and Figure 4(d) is a comparison diagram of ablation experiments on the MLR model for the ShipsEar dataset.

[0040] Figure 5(a) to Figure 5(d) Figure 5(a) is a comparison chart of ablation experiments on the AEMU module of the actually collected underwater acoustic signal target dataset on the MLP model. Figure 5(b) is a comparison chart of ablation experiments on the AEMU module of the ShipsEar dataset on the MLP model. Figure 5(c) is a comparison chart of ablation experiments on the AEMU module of the actually collected underwater acoustic signal target dataset on the MLR model. Figure 5(d) is a comparison chart of ablation experiments on the AEMU module of the ShipsEar dataset on the MLR model.

[0041] Figure 6 It is a structural diagram of the AEMU-SAE model proposed in the present invention.

[0042] Figure 7(a) to Figure 7(d) 7(a) is a comparison chart of the recognition accuracy of the actually collected underwater acoustic signal target dataset on the MLP model, FIG7(b) is a comparison chart of the recognition accuracy of the ShipsEar dataset on the MLP model, FIG7(c) is a comparison chart of the recognition accuracy of the actually collected underwater acoustic signal target dataset on the MLR model, and FIG7(d) is a comparison chart of the recognition accuracy of the ShipsEar dataset on the MLR model.

[0043] Figure 8(a) to Figure 8(d)When the maximum number of iterations is 500 Epochs, the ASAE Spec features learned by the AEMU-SAE model proposed in the present invention are compared with the training loss curves of the SAE Spec features, FBank features, GBank features, MFCC features and GFCC features on the MLP model and the MLR model respectively. FIG8(a) is a comparison diagram of the training loss curves of the actually collected underwater acoustic signal target dataset on the MLP model, FIG8(b) is a comparison diagram of the training loss curves of the ShipsEar dataset on the MLP model, FIG8(c) is a comparison diagram of the training loss curves of the actually collected underwater acoustic signal target dataset on the MLR model, and FIG8(d) is a comparison diagram of the training loss curves of the ShipsEar dataset on the MLR model.

[0044] Figure 9(a) to Figure 9(d) The ASAE Spec features learned by the AEMU-SAE model proposed in the present invention are compared with the noise robustness of SAE Spec features, FBank features, GBank features, MFCC features and GFCC features in the underwater target recognition task. FIG9(a) is a comparison diagram of the recognition accuracy of the actual collected underwater acoustic signal target dataset with background noise added on the MLP model, FIG9(b) is a comparison diagram of the recognition accuracy of the ShipsEar dataset with background noise added on the MLP model, FIG9(c) is a comparison diagram of the recognition accuracy of the actual collected underwater acoustic signal target dataset with background noise added on the MLR model, and FIG9(d) is a comparison diagram of the recognition accuracy of the ShipsEar dataset with background noise added on the MLR model. DETAILED DESCRIPTION

[0045] The present invention is further described below in conjunction with the accompanying drawings.

[0046] The present invention provides a feature-adversarial enhanced underwater target recognition method based on an acoustic embedded memory space encoder model. The method comprises the following steps: (1) extracting voiceprint features and auditory features from the original underwater acoustic signal; (2) proposing a self-supervised space encoder (SpaceAuto-Encoder, SAE) model; (3) proposing a negative sample mining strategy based on an acoustic embedded memory unit (Acoustic-Embedding Memory Unit, AEMU) module; (4) completing a feature-adversarial enhanced underwater target recognition method based on an acoustic embedded memory space encoder (Acoustic-Embedding Memory Unit Modified Space Auto-Encoder, AEMU-SAE) model. In order to achieve better underwater acoustic signal target recognition effect, the present invention proposes a feature-adversarial enhanced underwater target recognition method based on an acoustic embedded memory space encoder model.That is, the original underwater acoustic signal extracts the Mel filter bank spectrogram (FBank) features with voiceprint characteristics and the gamma filter bank spectrogram (GBank) features with auditory characteristics; in order to solve the problems of extremely high labeling cost of high-quality underwater target samples and weak generalization ability of the extracted single feature class in ocean scenes, it is proposed to build a self-supervised SAE model, and learn the Space Auto-Encoder Spectrogram (SAE Spec) features with good voiceprint characteristics and noise robustness by completing the spatial transformation and reconstruction from FBank features to GBank features; in view of the problem of insufficient negative sample learning of the SAE model, a negative sample mining strategy based on the AEMU module is proposed, and a dynamic queue dictionary is used to dynamically update and store negative samples to improve the efficiency of negative sample learning; on this basis, in order to solve the problem of the lack of high-level semantic information of the SAESpec feature, it is proposed to use an improved cosine embedding loss (CosineEmbedding Loss, CE The AEMU module is combined with the SAE model through a Loss function to construct the AEMU-SAE model, so that feature conversion reconstruction and feature negative sample learning are unified in the AEMU-SAE model, ensuring that the AEMU-SAE model can learn the acoustic embedding memory space encoder spectrogram (Acoustic-EmbeddingMemory Unit Modified Space Auto-Encoder Spectrogram, ASAE Spec) features containing high-level semantic information during the feature adversarial enhancement process, and use the ASAESpec features as the input of the Multi-Layer Perceptron (MLP) model and the Multinomial Logistic Regression (MLR) model respectively, and process the recognition task through the multi-classification target recognition model to finally complete the underwater target recognition. The feature adversarial enhancement underwater target recognition method based on the AEMU-SAE model proposed in the present invention has good performance in recognition accuracy, task adaptability and anti-noise robustness, and has certain effectiveness.

[0047] A feature adversarial enhanced underwater target recognition method based on an acoustic embedding memory spatial encoder model comprises the following steps:

[0048] Step 1: Obtain the identified original underwater acoustic signal dataset, and extract FBank features and GBank features from the original underwater acoustic signals in the dataset;

[0049] After pre-emphasis, framing, windowing, and short-time Fourier transform of the original underwater acoustic signal, the FBank feature with good voiceprint distinguishing advantage and the GBank feature with anti-noise robustness advantage are extracted using the Mel filter bank and the gamma filter bank respectively;

[0050] FBank features are:

[0051]

[0052] Among them, s(n) is the frame information; N is the total number of frames; m is the feature order; L is the total feature order; M is the number of Mel filters;

[0053] GBank features:

[0054]

[0055] Where i is the frame number; E S (i) represents the frame energy spectrum information; I is the total number of frames; n is the feature order; Q is the number of gamma filters;

[0056] Step 2: Train the AEMU-SAE model using a dataset containing the FBank features and GBank features of the original underwater acoustic signal;

[0057] The training optimization algorithm of the AEMU-SAE model adopts the Adam algorithm, which is as follows:

[0058]

[0059] in, is the bias-corrected first-order moment estimate; is the bias-corrected second-order moment estimate; ε is used to prevent division by zero errors during implementation; η is the learning rate, which is responsible for controlling the update rate of the weights.

[0060] The AEMU-SAE model includes a random Gaussian noise embedding module, an encoder module, a decoder module, and an AEMU module. The specific training process is as follows:

[0061] Step 2.1: The random Gaussian noise embedding module generates random Gaussian noise during the training process. The FBank features of the original underwater acoustic signal are mixed with the random Gaussian noise and then input into the encoder module.

[0062] The random Gaussian noise generated by the random Gaussian noise embedding module during training is specifically:

[0063]

[0064] Among them, z is the noise information, μ is the noise expectation, and σ is the noise variance.

[0065] Step 2.2: The encoder module outputs the ASAE Spec features and inputs them into the decoder module;

[0066] Step 2.3: The decoder module reconstructs the output of the encoder module into pseudo GBank features through an upsampling process and inputs them into the AEMU module;

[0067] Step 2.4: The AEMU module executes the negative sample mining strategy and stores a large number of negative samples from the original data space in the form of a dynamic queue dictionary. During the training process, samples are directly taken from the dictionary by keyword query. In each iteration, a small batch of sample sets are updated in the dictionary by dequeuing and enqueuing, thereby improving the learning efficiency of negative samples.

[0068] The AEMU module calculates the spatial similarity loss between the pseudo-GBank features and the real GBank features to achieve the purpose of training the model;

[0069] The loss function adopts the improved CE Loss loss function, specifically:

[0070]

[0071] Among them, τ is the temperature hyperparameter coefficient; sim(·) represents the feature similarity function, that is, the CE Loss function; f(x i ) and f(x j ) represents a positive sample pair, i.e., the reconstructed pseudo-GBank feature and the homologous real GBank feature; f(x i ) and f(x k ) represents the negative sample pair, i.e., the reconstructed pseudo-GBank features and the non-homologous real GBank features; minimizing e means minimizing the mutual information distance of the positive sample pair and maximizing the mutual information distance of the negative sample pair, ensuring that the AEMU-SAE model learns the ASAE Spec features containing high-level semantic information;

[0072] The feature similarity function sim(·), that is, the CE Loss function is specifically:

[0073]

[0074] Among them, margin is the minimum threshold of feature similarity to prevent overfitting; x i and x j Respectively represent the vector pair of the reconstructed pseudo-GBank features and the real GBank features after one-dimensional mapping expansion; y represents the automatically generated positive and negative sample pseudo labels, and the calculated loss value can be regarded as the similarity between the reconstructed pseudo-GBank features and the real GBank features in the auditory space;

[0075] Step 3: After FBank feature extraction, the original underwater acoustic signal data to be identified is input into the encoder module of the trained AEMU-SAE model to obtain the ASAE Spec features containing high-level semantic information; the ASAE Spec features are used as the input of the MLP model and the MLR model respectively, and the recognition task is processed by the multi-classification target recognition model to finally complete the underwater target recognition.

[0076] Embodiment 1:

[0077] The present invention comprises the following steps in the implementation process:

[0078] (1) Extract FBank features and GBank features from the original underwater acoustic signal: ① Use the Mel filter group to extract the FBank features with good voiceprint distinguishing advantages from the original underwater acoustic signal; ② Use the gamma filter group to extract the GBank features with the advantage of anti-noise robustness from the original underwater acoustic signal;

[0079] (2) Propose to construct a self-supervised SAE model: ① Combine the voiceprint distinguishing advantages of FBank features and the noise robustness advantages of GBank features in a way similar to the animal vocalization auditory system; ② Learn SAE Spec features with good voiceprint characteristics and noise robustness by completing the spatial transformation and reconstruction from FBank features to GBank features;

[0080] (3) Propose a negative sample mining strategy based on the AEMU module;

[0081] (4) Complete the feature adversarial enhancement underwater target recognition method based on the AEMU-SAE model: ① Use the improved CELoss function to combine the AEMU module with the SAE model to construct the AEMU-SAE model, so that the feature transformation reconstruction and feature negative sample learning are unified in the AEMU-SAE model, ensuring that the AEMU-SAE model can learn the ASAE Spec features containing high-level semantic information during the feature adversarial enhancement process; ② Use the ASAE Spec features as the input of the MLP model and the MLR model respectively, and process the recognition task through the multi-classification target recognition model to finally complete the underwater target recognition.

[0082] The present invention may also include:

[0083] 1. In step (1), the original underwater acoustic signal is pre-emphasized, framed, windowed, and short-time Fourier transformed to obtain basic spectrogram features.

[0084] 2. In step (1), the Mel filter is used to extract FBank features from the spectrogram features, specifically: Where n is the frame number, s(n) is the frame information, N is the total number of frames, m is the feature order, L is the total feature order, M is the number of Mel filters, and the gamma filter bank is used to extract the GBank feature from the spectrogram feature, specifically: where i is the frame number, E S (i) represents the frame energy spectrum information, I is the total number of frames, n is the feature order, and Q is the number of gamma filters.

[0085] 3. In step (2), according to the characteristics of reconstructive self-supervised learning, the proposed self-supervised SAE model includes a random Gaussian noise embedding module, an encoder module and a decoder module.

[0086] 4. The random Gaussian noise embedding module of the self-supervised SAE model in step (2) generates random Gaussian noise during the training process and mixes it with the FBank feature as the model input to improve the model's anti-noise robustness. Specifically, Where z is the noise information, μ is the noise expectation, and σ is the noise variance.

[0087] 5. In the step (2), the encoder module of the self-supervised SAE model obtains input data, performs operations including data dimensionality reduction and feature weight learning, and finally maps it to the acoustic embedding space to obtain a feature space with a smaller dimension than the input data.

[0088] 6. The decoder module of the self-supervised SAE model in step (2) uses the encoding result processed by the encoder module as input data, completes the task of reconstructing pseudo GBank features from low-dimensional encoding features through an upsampling process, and calculates the spatial similarity loss with the real GBank features to achieve the purpose of training the model.

[0089] 7. The loss function of the self-supervised SAE model in step (2) is the CE Loss function, specifically: Where margin is the minimum threshold of feature similarity to prevent overfitting, x i and x j They represent the vector pair of the reconstructed pseudo-GBank features and the real GBank features after one-dimensional mapping expansion, y represents the automatically generated pseudo-labels of positive and negative samples, and the calculated loss value can be regarded as the similarity between the reconstructed pseudo-GBank features and the real GBank features in the auditory space.

[0090] 8. The self-supervised SAE model in step (2) uses the Adam algorithm as the training optimization algorithm of the model, specifically: in is the bias-corrected first-order moment estimate, is the bias-corrected second-order moment estimate, ε is used to prevent division by zero errors during implementation, and η is called the learning rate or step factor, which is responsible for controlling the update rate of the weights.

[0091] 9. In step (3), a negative sample mining strategy based on the AEMU module is proposed, which stores a large number of negative samples from the original data space in the form of a dynamic queue dictionary. During the training process, samples are directly taken from the dictionary through keyword queries. In each iteration, a small batch of samples is updated by dequeuing and enqueuing.

[0092] 10. In step (4), since the model introduces a feature adversarial enhancement mechanism, an improved CE Loss function is used, specifically: where τ is the temperature hyperparameter coefficient, sim(·) represents the feature similarity function, and f(x i ) and f(x j ) represents a positive sample pair, f(x i ) and f(x k ) represents the negative sample pair, minimizing l i,j That is, the mutual information distance of positive sample pairs is minimized and the mutual information distance of negative sample pairs is maximized to ensure that the AEMU-SAE model learns the ASAE Spec features containing high-level semantic information.

[0093] 11. The AEMU-SAE model trained in step (4) extracts the ASAE Spec features as high-level semantic information features for underwater target recognition, and inputs them into the MLP model and the MLR model respectively. After being mapped and transformed by the Softmax function, the recognition results are finally obtained to complete the underwater target recognition task.

[0094] Compared with the prior art, the advantages of the present invention are as follows: a. The traditional deep learning method relies on a large number of high-quality annotated sample sets to train the model to complete the underwater target recognition task, but the cost of annotating high-quality underwater target samples is extremely high, which affects the effectiveness of recognition, while the self-supervised learning model automatically learns the target features by constructing auxiliary tasks. According to the advantage of the self-supervised learning model that automatically learns the target features without labels, combined with the characteristics of the underwater acoustic signal itself, the present invention proposes a feature-adversarial enhanced underwater target recognition method based on the acoustic embedding memory space encoder model; b. In order to solve the problems of extremely high cost of annotating high-quality underwater target samples and weak generalization ability of the extracted single feature class in ocean scenes, the present invention proposes to construct a self-supervised SAE model, and learns SAE Spec features with good voiceprint characteristics and anti-noise robustness by completing the spatial transformation reconstruction from FBank features to GBank features; c. In view of the problem of insufficient negative sample learning of the self-supervised SAE model, the present invention proposes a negative sample mining strategy based on the AEMU module, and uses a dynamic queue dictionary to dynamically update and store negative samples, thereby improving the efficiency of negative sample learning; d. In order to solve the problem of the lack of high-level semantic information of the SAE Spec feature, the present invention uses an improved CE The Loss function combines the AEMU module with the SAE model to construct the AEMU-SAE model, which unifies feature transformation reconstruction and feature negative sample learning in the AEMU-SAE model, ensuring that the AEMU-SAE model can learn the ASAE Spec features containing high-level semantic information during the feature adversarial enhancement process. The ASAE Spec features are used as the input of the MLP model and the MLR model respectively, and the recognition task is processed through the multi-classification target recognition model to finally complete the underwater target recognition.

[0095] The feature adversarial enhanced underwater target recognition method based on the acoustic embedding memory space encoder model proposed in the present invention has good performance in recognition accuracy, task adaptability and noise resistance robustness, and has certain effectiveness.

[0096] Embodiment 2:

[0097] Combination Figure 1 , the specific steps of the present invention are as follows:

[0098] (1) Extracting voiceprint features and auditory features from the original underwater sound signal

[0099] The present invention uses two data sets. One is the underwater acoustic signal target data set collected in practice. The data set is divided into five categories, including four types of ship and underwater mammal sound data sets, with a total size of nearly 20 hours. Each sound signal is divided into multiple short speech segments with a duration of 2 seconds. The other is the ShipsEar ship noise data set recorded in different areas of the Spanish coast between 2012 and 2013. The data set has a total of 90 records of 11 types of ships within 15s to 10min. According to the annotations in the original data set, they can be merged into four major categories, A, B, C, and D, according to the types of ships.

[0100] After pre-emphasis, framing, windowing, and short-time Fourier transform of the original underwater acoustic signal, the Mel filter bank and gamma filter bank are used to extract the FBank features with good voiceprint distinguishing advantages and the GBank features with anti-noise robustness advantages respectively.

[0101] The FBank features extracted using the Mel filter are:

[0102]

[0103] Where s(n) is the frame information, N is the total number of frames, m is the feature order, L is the total feature order, and M is the number of Mel filters.

[0104] The GBank features extracted using the gamma filter bank are:

[0105]

[0106] Where i is the frame number, E S (i) represents the frame energy spectrum information, I is the total number of frames, n is the feature order, and Q is the number of gamma filters.

[0107] Figure 2 is an example diagram of the spectrum of underwater mammals randomly selected from the actual collected underwater acoustic signal target data set, where Figure 2(a) is an example diagram of the FBank feature spectrum, Figure 2(b) is an example diagram of the GBank feature spectrum, and Figure 2(c) is an example diagram of the original spectrum.

[0108] (2) Propose a self-supervised SAE model

[0109] Since the labeling cost of high-quality underwater target samples is extremely high and the generalization ability of the extracted single feature class in marine scenes is weak, it is proposed to construct a self-supervised SAE model. Based on the self-supervised auxiliary task of the animal vocalization auditory system, the spatial transformation and reconstruction from FBank features to GBank features is completed, and the advantages of the voiceprint distinguishing of FBank features and the advantages of the noise robustness of GBank features are combined to learn SAE Spec features with good voiceprint characteristics and noise robustness. According to the characteristics of reconstructive self-supervised learning, the proposed self-supervised SAE model includes a random Gaussian noise embedding module, an encoder module and a decoder module, and uses the CE Loss function as the loss function of the model, and uses the Adam algorithm as the training optimization algorithm of the model. Figure 3 It is a structural diagram of the self-supervised SAE model proposed in the present invention.

[0110] ① Random Gaussian noise embedding module

[0111] In order to make the SAE model have stronger task processing capabilities and noise robustness, a parallel random Gaussian noise embedding module is added to the data input layer. During model training, the random Gaussian noise embedding module generates random Gaussian noise and inputs it into the model together with the FBank feature. By attaching noise to the original underwater sound signal, the auxiliary task design of the SAE model based on the simulated animal vocalization auditory system becomes meaningful. In addition, the training method of adding noise to the input information can better improve the noise robustness of the SAE model, thereby improving the noise resistance performance of the downstream underwater target recognition task.

[0112] Random Gaussian noise is specifically:

[0113]

[0114] In the formula, z is the noise information, μ is the noise expectation, and σ is the noise variance.

[0115] In order to verify the effect of random Gaussian noise embedding module on improving the noise robustness of acoustic features, 15 groups of ablation experiments were designed on the actual collected underwater acoustic signal target dataset and ShipsEar dataset. The SAE Spec features learned by the SAE model containing random Gaussian noise module and the Un-robust Space Auto-Encoder Spectrogram features (USAE Spectrogram, USAE Spectrogram) learned by the Space Encoder (Un-robust Space Auto-Encoder, USAE Spectrogram, USAE Spectrogram) model without random Gaussian noise module were compared. Spec) on the MLP model and the MLR model. Figure 4 is a comparison of the ablation experiment of the random Gaussian noise embedding module of the SAE model, where Figure 4(a) is a comparison of the ablation experiment of the actual collected underwater acoustic signal target dataset on the MLP model, Figure 4(b) is a comparison of the ablation experiment of the ShipsEar dataset on the MLP model, Figure 4(c) is a comparison of the ablation experiment of the actual collected underwater acoustic signal target dataset on the MLR model, and Figure 4(d) is a comparison of the ablation experiment of the ShipsEar dataset on the MLR model. As can be seen from Figure 4, although the recognition accuracy of the SAE Spec feature and the USAE Spec feature have slight fluctuations on the MLP model and the MLR model, the noise robustness advantage of the SAE Spec feature is still obvious. In Figure 4(a) and Figure 4(b), in the MLP model, the recognition accuracy of the SAE Spec feature on the actual collected underwater acoustic signal target dataset and the ShipsEar dataset is about 2.10% higher than that of the USAE Sepc feature. In Figure 4(c) and Figure 4(d), in the MLR model, the recognition accuracy of the SAE Spec feature on the actual collected underwater acoustic signal target dataset and the ShipsEar dataset is about 2.05% and 2.12% higher than that of the USAE Sepc feature, respectively. In summary, the recognition accuracy of the SAE Spe feature in 15 groups of ablation comparison experiments is better than that of the USAE Spec feature, which verifies that the random Gaussian noise embedding module has an effect on improving the noise robustness of acoustic features in underwater target recognition tasks.

[0116] ②Encoder module

[0117] The encoder module of the self-supervised SAE model obtains input data, performs operations including data dimensionality reduction and feature weight learning, and finally maps it to the acoustic embedding space to obtain a feature space with a smaller dimension than the input data. The encoder module is a convolutional neural network structure, which is similar to the common ResNet50 model. The encoder module has a total of 5 stages, and each stage contains a residual module with 3 convolutional layers.

[0118] The principles of the encoder module are:

[0119] z i =f(x i )

[0120] In the formula, x i represents the i-th input sample, z i represents the i-th sound embedding obtained by the encoder, and f(·) represents the encoder module.

[0121] ③Decoder module

[0122] The decoder module of the self-supervised SAE model takes the encoding result processed by the encoder module as input data. Through the upsampling process, it completes the task of reconstructing pseudo GBank features from low-dimensional encoding features, and calculates the spatial similarity loss with the real GBank features to achieve the purpose of training the model.

[0123] The principle of the decoder module is:

[0124] r i =g(z i )

[0125] In the formula, z i represents the i-th sound embedding output by the encoder, r i represents the i-th pseudo GBank feature obtained by the decoder, and g(·) represents the decoder module.

[0126] ④CE Loss function and Adam optimization algorithm

[0127] Since the data set collection work for underwater target recognition tasks is relatively difficult, most of the data sets show the problem of uneven category distribution. In addition, the task processed in the self-supervised SAE algorithm is a feature reconstruction auxiliary task. Using CE Loss as the loss function can better measure the similarity of feature vectors and avoid overfitting. The CE Loss loss function is defined as:

[0128]

[0129] In the formula, margin is the minimum difference in feature similarity to prevent overfitting, x i and x j They represent the vector pair of the reconstructed pseudo-GBank features and the real GBank features after one-dimensional mapping expansion, y represents the automatically generated pseudo-labels of positive and negative samples, and the calculated loss value can be regarded as the similarity between the reconstructed pseudo-GBank features and the real GBank features in the auditory space.

[0130] The self-supervised SAE model uses the Adam algorithm as the training optimization algorithm of the model. The main advantage of the Adam algorithm is that after bias correction, the learning rate of each iteration has a certain range, making the parameters relatively stable. Specifically:

[0131]

[0132] In the formula, is the bias-corrected first-order moment estimate, is the bias-corrected second-order moment estimate, ε is used to prevent division by zero errors during implementation, and η is called the learning rate or step factor, which is responsible for controlling the update rate of the weights.

[0133] (3) Propose a negative sample mining strategy based on the AEMU module

[0134] A negative sample mining strategy based on the AEMU module is proposed, which stores a large number of negative samples from the original data space in the form of a dynamic queue dictionary. During the training process, samples are directly taken from the dictionary through keyword query. In each iteration, a small batch of samples is used to update the dictionary by dequeuing and enqueuing, thereby improving the learning efficiency of negative samples.

[0135] In order to verify the improvement effect of the AEMU module on negative sample learning, 8 groups of ablation experiments were designed on the actually collected underwater acoustic signal target dataset and ShipsEar dataset, and the recognition accuracy of ASAE Spec features learned by AEMU modules of different sizes on the MLP model and MLR model were compared. Figure 5 is a comparison diagram of ablation experiments of the AEMU module, where Figure 5(a) is a comparison diagram of ablation experiments of the AEMU module on the MLP model of the actually collected underwater acoustic signal target dataset, Figure 5(b) is a comparison diagram of ablation experiments of the AEMU module on the MLP model of the ShipsEar dataset, Figure 5(c) is a comparison diagram of ablation experiments of the AEMU module on the MLR model of the actually collected underwater acoustic signal target dataset, and Figure 5(d) is a comparison diagram of ablation experiments of the AEMU module on the MLR model of the ShipsEar dataset. As can be seen from Figure 5, the recognition accuracy of the ASAE Spec feature in the MLP model and the MLR model varies with the size of the AEMU module. The recognition accuracy of the specific data set is positively correlated with the size of the AEMU module. When the AEMU module size exceeds 256, the recognition accuracy of the ASAE Spec feature gradually stabilizes. Based on the performance of the ASAE Spec feature in the underwater target recognition task of the MLP model and the MLR model, the AEMU module has been verified to improve the negative sample learning.

[0136] (4) Complete feature-based adversarial enhanced underwater target recognition based on the AEMU-SAE model

[0137] ①Build AEMU-SAE model

[0138] The AEMU module is combined with the SAE model using the improved CE Loss function to construct the AEMU-SAE model, which unifies feature conversion reconstruction and feature negative sample learning in the AEMU-SAE model, ensuring that the AEMU-SAE model can learn the ASAE Spec features containing high-level semantic information during the feature adversarial enhancement process. Figure 6 It is a structural diagram of the AEMU-SAE model proposed in the present invention.

[0139] The improved CE Loss function is as follows:

[0140]

[0141] Where: τ is the temperature hyperparameter coefficient, sim(·) represents the feature similarity function, f(x i ) and f(x j ) represents a positive sample pair, f(x i ) and f(x k ) represents the negative sample pair, minimizing l i,j That is, it means minimizing the mutual information distance of positive sample pairs and maximizing the mutual information distance of negative sample pairs.

[0142] ②Use ASAE Spec features to achieve more accurate underwater target recognition

[0143] The trained AEMU-SAE model extracts ASAE Spec features as high-level semantic information features for underwater target recognition, which are input into the MLP model and MLR model respectively. The probability distribution of recognition is obtained through Softmax function mapping and transformation, and the recognition result is finally output.

[0144] The three major indicators of recognition accuracy (Accuracy, Acc.), convergence speed (Recognition Loss Convergence Rate, CLCR) and anti-noise robustness (Anti-Noise Robust, ANR) are the performance evaluation criteria of ASAE Spec features to complete underwater target recognition tasks. Acc. is the most basic performance evaluation indicator, which is used to verify the recognition accuracy of the current feature in the downstream underwater target recognition task. CLCR is an important indicator to verify whether the current model can learn high-level speech features. It puts different feature samples of the same order of magnitude into the same recognition model for training. The number of epochs required when the model reaches convergence is CLCR. ANR is a special evaluation indicator for underwater mission scenarios, which is used to measure the anti-noise robustness of the current model in processing complex underwater acoustic signals. The special processing of ANR is that it is necessary to first fuse the background noise of the current sea area with the underwater acoustic signals of various targets, and then perform feature extraction and target recognition on this basis.

[0145] In order to verify the effectiveness of the feature-based adversarial enhancement underwater target recognition method of the AEMU-SAE model proposed in the present invention, experiments on the recognition accuracy, convergence speed and anti-noise robustness of the underwater acoustic signal target data set and the ShipsEar data set actually collected are given. Figure 7 is a comparison chart of the recognition accuracy of the ASAE Spec feature learned by the AEMU-SAE model proposed in the present invention, and the SAESpec feature, FBank feature, GBank feature, MFCC feature and GFCC feature, where Figure 7 (a) is a comparison chart of the recognition accuracy of the underwater acoustic signal target data set actually collected on the MLP model, Figure 7 (b) is a comparison chart of the recognition accuracy of the ShipsEar data set on the MLP model, Figure 7 (c) is a comparison chart of the recognition accuracy of the underwater acoustic signal target data set actually collected on the MLR model, and Figure 7 (d) is a comparison chart of the recognition accuracy of the ShipsEar data set on the MLR model. It can be concluded from Figure 7 that the recognition accuracy of the ASAE Spec feature learned by the AEMU-SAE model proposed in the present invention is relatively the highest on both the MLP model and the MLR model. Due to the lack of AEMU module, the SAE model has the problem of insufficient negative sample learning in the self-supervised learning process, resulting in the lack of high-level semantics of acoustic features. Therefore, the recognition accuracy of SAE Spec features is lower than that of ASAE Spec features in MLP model and MLR model. Although FBank features have the advantage of voiceprint distinguishing ability, their noise robustness is weak. GBank features have strong noise robustness but their voiceprint distinguishing ability is not strong enough. Therefore, the adaptability of these two features for underwater target recognition tasks is not strong enough. MFCC features and GFCC features are both shallow machine learning features, with complex feature engineering and are not suitable for deep target recognition models. Therefore, their recognition accuracy is lower than that of ASAE Spec features. Table 1 shows the underwater acoustic signal target dataset and ShipsEar dataset collected in real time with a training set and test set ratio of 7 to 3. When the training iterations of MLP model and MLR model are 80 Epochs, the target recognition accuracy comparison results of MFCC, GFCC, FBank, GBank, SAE Spec and ASAE Spec features are compared.

[0146] Table 1 Comparison results of various feature recognition accuracy

[0147]

[0148] It can be seen from Table 1 that the extracted ASAE Spec features have higher recognition accuracy than other features on the MLP model and MLR model in the actual collected underwater acoustic signal target dataset and ShipsEar dataset. Therefore, it can be proved that the AEMU-SAE model proposed in this invention has certain effectiveness for underwater target recognition.

[0149] To further verify the convergence speed, Figure 8 shows the training loss curve comparison of ASAE Spec features, SAE Spec features, FBank features, GBank features, MFCC features and GFCC features on the MLP model and MLR model respectively when the maximum number of iterations is 500 Epochs. Figure 8(a) is the training loss curve comparison of the actual collected underwater acoustic signal target dataset on the MLP model, Figure 8(b) is the training loss curve comparison of the ShipsEar dataset on the MLP model, Figure 8(c) is the training loss curve comparison of the actual collected underwater acoustic signal target dataset on the MLR model, and Figure 8(d) is the training loss curve comparison of the ShipsEar dataset on the MLR model. It can be concluded from Figure 8 that although the convergence performance of various features shows a rapid decline at the beginning of training, followed by slight jitter, and finally tends to be flat and reaches the local optimal performance, since the ASAE Spec feature contains more advanced semantics, the ASAE Spec feature has the best convergence speed performance. Table 2 gives the comparison results of the convergence speed of MFCC, GFCC, FBank, GBank, SAE Spec and ASAE Spec features for the actually collected underwater acoustic signal target dataset and ShipsEar dataset.

[0150] Table 2 Comparison results of convergence speed of various features

[0151]

[0152] It can be seen from Table 2 that in the actual collected underwater acoustic signal target dataset and ShipsEar dataset, the ASAESpec feature convergence speed performance is better than other features. Therefore, the ASAE Spec feature learned by the AEMU-SAE model proposed in the present invention contains more advanced semantics.

[0153] Since the noise robustness of the model is also an important evaluation index in the underwater target recognition task, the actual collected underwater background noise dataset is added to the actual collected underwater acoustic signal target dataset and ShipsEar dataset to verify the noise robustness of the ASAE Spec feature. Figure 9 shows the ASAE Spec feature learned by the AEMU-SAE model proposed in the present invention, and the noise robustness comparison diagram of the SAE Spec feature, FBank feature, GBank feature, MFCC feature and GFCC feature in the underwater target recognition task, where Figure 9(a) is a comparison diagram of the recognition accuracy of the actual collected underwater acoustic signal target dataset with background noise added on the MLP model, Figure 9(b) is a comparison diagram of the recognition accuracy of the ShipsEar dataset with background noise added on the MLP model, Figure 9(c) is a comparison diagram of the recognition accuracy of the actual collected underwater acoustic signal target dataset with background noise added on the MLR model, and Figure 9(d) is a comparison diagram of the recognition accuracy of the ShipsEar dataset with background noise added on the MLR model. As can be seen from Figure 9, although the recognition accuracy has more severe jitters than the original dataset and has decreased, the recognition accuracy of the ASAE Spec feature is still higher than that of other features. Therefore, the ASAE Spec feature has good noise robustness. Table 3 shows the noise robustness comparison results of MFCC, GFCC, FBank, GBank, SAE Spec and ASAE Spec features when the training set and test set ratio is 7 to 3, and the underwater background noise dataset actually collected is added to the underwater acoustic signal target dataset and ShipsEar dataset actually collected. When the MLP model and MLR model training iteration number is 80 Epochs.

[0154] Table 3 Comparison results of noise robustness of various features

[0155]

[0156] As can be seen from Table 3, due to the strongest noise robustness of the ASAE Spec feature, the recognition accuracy of the ASAE Spec feature in the actual collected underwater acoustic signal target dataset and ShipsEar dataset and the actual collected underwater background noise dataset added to the actual collected underwater acoustic signal target dataset and ShipsEar dataset are higher than other features, and its accuracy reduction rate is the lowest. Combining the recognition accuracy experimental comparison data in Table 3 and the recognition model convergence speed experimental comparison data in Table 4, it can be obtained that the feature adversarial enhancement underwater target recognition method based on the AEMU-SAE model proposed in the present invention can effectively improve the overall performance of underwater target recognition accuracy under weak labels and large environmental noise interference through self-supervised learning combined with feature adversarial enhancement methods when high-quality marine target samples cannot be obtained and the labeling cost is high.

[0157] The feature adversarial enhanced underwater target recognition method based on the AEMU-SAE model proposed in the present invention has good performance in recognition accuracy, task adaptability and anti-noise robustness, and has certain effectiveness.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A feature-enhanced underwater target recognition method based on an acoustic embedding memory spatial encoder model, characterized in that: The following steps are involved: Step 1: Obtain the identified original underwater acoustic signal dataset, and extract FBank features and GBank features from the original underwater acoustic signals in the dataset; After pre-emphasis, framing, windowing, and short-time Fourier transform of the original underwater acoustic signal, the FBank feature with good voiceprint distinguishing advantage and the GBank feature with anti-noise robustness advantage are extracted using the Mel filter bank and the gamma filter bank respectively; FBank features are: Among them, s(n) is the frame information; N is the total number of frames; m is the feature order; L is the total feature order; M is the number of Mel filters; GBank features: Where i is the frame number; E S (i) represents the frame energy spectrum information; I is the total number of frames; n is the feature order; Q is the number of gamma filters; Step 2: Train the AEMU-SAE model using a dataset containing the FBank features and GBank features of the original underwater acoustic signal; The AEMU-SAE model includes a random Gaussian noise embedding module, an encoder module, a decoder module, and an AEMU module. The specific training process is as follows: Step 2.1: The random Gaussian noise embedding module generates random Gaussian noise during the training process. The FBank features of the original underwater acoustic signal are mixed with the random Gaussian noise and then input into the encoder module. Step 2.2: The encoder module outputs the ASAE Spec features and inputs them into the decoder module; Step 2.3: The decoder module reconstructs the output of the encoder module into pseudo GBank features through an upsampling process and inputs them into the AEMU module; Step 2.4: The AEMU module executes the negative sample mining strategy and stores a large number of negative samples from the original data space in the form of a dynamic queue dictionary. During the training process, samples are directly taken from the dictionary by keyword query. In each iteration, a small batch of sample sets are updated in the dictionary by dequeuing and enqueuing, thereby improving the learning efficiency of negative samples. The AEMU module calculates the spatial similarity loss between the pseudo-GBank features and the real GBank features to achieve the purpose of training the model; The loss function adopts the improved CE Loss loss function, specifically: Among them, τ is the temperature hyperparameter coefficient; sim(·) represents the feature similarity function, that is, the CE Loss function; f(x i ) and f(x j ) represents a positive sample pair, i.e., the reconstructed pseudo-GBank feature and the homologous real GBank feature; f(x i ) and f(x k ) represents the negative sample pair, i.e., the reconstructed pseudo-GBank features and the non-homologous real GBank features; minimizing l means minimizing the mutual information distance of the positive sample pair and maximizing the mutual information distance of the negative sample pair, ensuring that the AEMU-SAE model learns the ASAE Spec features containing high-level semantic information; Step 3: After FBank feature extraction, the original underwater acoustic signal data to be identified is input into the encoder module of the trained AEMU-SAE model to obtain the ASAE Spec features containing high-level semantic information; the ASAE Spec features are used as the input of the MLP model and the MLR model respectively, and the recognition task is processed by the multi-classification target recognition model to finally complete the underwater target recognition.

2. The feature-enhanced underwater target recognition method based on the acoustic embedded memory spatial encoder model according to claim 1 is characterized in that: The random Gaussian noise generated by the random Gaussian noise embedding module in the step 2.1 during the training process is specifically: Among them, z is the noise information, μ is the noise expectation, and σ is the noise variance.

3. The feature-enhanced underwater target recognition method based on the acoustic embedded memory spatial encoder model according to claim 1 is characterized in that: The feature similarity function sim(·) in step 2.4, i.e., the CE Loss function, is specifically: Among them, margin is the minimum threshold of feature similarity to prevent overfitting; x i and x j They represent the vector pair of the reconstructed pseudo-GBank features and the real GBank features after one-dimensional mapping expansion; y represents the automatically generated pseudo-labels of positive and negative samples. The calculated loss value can be regarded as the similarity between the reconstructed pseudo-GBank features and the real GBank features in the auditory space.

4. The feature-enhanced underwater target recognition method based on acoustic embedded memory spatial encoder model according to claim 1 is characterized in that: The training optimization algorithm of the AEMU-SAE model in step 2 adopts the Adam algorithm, which is specifically: in, is the bias-corrected first-order moment estimate; is the bias-corrected second-order moment estimate; ε is used to prevent division by zero errors during implementation; η is the learning rate, which is responsible for controlling the update rate of the weights.

Images