A ship radiated noise recognition method and system based on a double low-rank adjustment network

Abstract

The application provides a ship radiation noise recognition method and system based on a double low-rank adjustment network. The ship radiation noise recognition method based on the double low-rank adjustment network comprises the following steps: performing Mel filter bank transformation on a ship radiation noise signal to obtain a Mel time-frequency spectrum diagram of the noise signal; inputting the Mel time-frequency spectrum diagram into a trained noise recognition model to output a target category of the ship radiation noise; and generating the noise recognition model based on a pre-trained benchmark model containing a plurality of residual blocks, performing low-rank adaptation on a weight space through a weight low-rank adaptive module, and performing low-rank adaptation on a feature space through a feature low-rank adaptive module. The application has the advantages that the classification performance can be significantly improved on ShipsEar and DeepShip data sets by increasing only about 0.55% parameters, and the application has low overhead and high effectiveness.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of ship radiated noise recognition, and particularly relates to a ship radiated noise recognition method and system based on a double low-rank adjustment network. BACKGROUND

[0002] Ship radiated noise is one of the main components of ocean ambient noise, and the passive ship recognition method based on ship acoustic radiation characteristics is an important technical means for tasks such as ocean environment monitoring and military target detection. In recent years, there are the following several kinds of data-driven deep learning methods:

[0003] The first kind is a data expansion method, which aims to expand the size and diversity of the training set by transforming existing data or by generating models. Data expansion can alleviate overfitting to some extent, but the diversity of synthetic samples is insufficient, and the training cost is increased.

[0004] The second kind is a domain adaptation method, which aims to reduce the difference in feature distribution between the training set (source domain) and the test set (target domain) by designing a feature distribution alignment constraint. Domain adaptation methods usually require unlabeled target domain data in the training phase, but it is difficult to meet this requirement in actual applications. At the same time, the domain transfer process based on confrontation may not be stable, and when the domain difference is too large, it may lead to the phenomenon of "negative transfer", which damages the performance of the model.

[0005] The third kind is a transfer learning method. Since most underwater acoustic recognition tasks such as ship radiated noise recognition use two-dimensional time-frequency spectrograms as input, borrowing pre-trained models from the image classification field as weight initialization and fine-tuning the model have become an effective and commonly used means to improve model performance and accelerate model convergence. For ship noise recognition tasks, the limited labeled data increases the risk of overfitting of the fully fine-tuned model, which reduces the model's generalization ability. On the other hand, ship radiated noise time-frequency spectrum data has a special spectral line structure, which is inherently different from natural images, and fully fine-tuning image pre-trained weights is difficult to achieve good results.

[0006] In addition, the Chinese invention patent with the publication number CN 118982074 A and the invention name of an industry large model training method and system based on efficient fine-tuning and federal learning discloses that parameter efficient updating is realized by low-rank matrix fine-tuning at each client, and different client models are aggregated by federal learning, thereby improving the adaptability and generalization ability of the model under the premise of ensuring data privacy. This method has certain effect in reducing centralized data dependence and alleviating training resource consumption, but in the domain offset scenario caused by complex ocean propagation environment and differences in receiving equipment, it is still difficult to completely solve the problem of insufficient generalization ability of underwater acoustic recognition models.

[0007] In summary, the prior art relies on a large amount of training data, and the traditional transfer learning method often fails to achieve ideal results; and the complex marine propagation environment and the difference between the receiving equipment result in domain shift between the test data and the training data, which reduces the generalization ability of the model.

[0008] Therefore, improving the generalization ability has important research significance for ship noise recognition. SUMMARY

[0009] The purpose of the present application is to overcome the defects of low generalization ability caused by the prior art relying on a large amount of training data and domain shift, and to propose a ship radiated noise recognition method and system based on a double low-rank adjustment network for ship radiated noise recognition.

[0010] In order to achieve the above-mentioned purpose, the present application proposes a ship radiated noise recognition method based on a double low-rank adjustment network, comprising:

[0011] Mel filter bank transformation is performed on the ship radiated noise signal to obtain a Mel time-frequency spectrum of the noise signal;

[0012] The Mel time-frequency spectrum is input into a trained noise recognition model to output a target category of the ship radiated noise;

[0013] The noise recognition model is generated based on a pre-trained reference model containing a plurality of residual blocks, and the weight space is adapted by a low-rank adaptive module for weights, and the feature space is adapted by a low-rank adaptive module for features.

[0014] The low-rank adaptive module for weights acts on the linear layer weights in the residual block and is used for weight fine-tuning.

[0015] The processing process of the low-rank adaptive module for weights is: the pre-training weights are frozen A trainable low-rank weight matrix is constructed beside the linear layer weights as a learnable low-rank adjustment increment to obtain fused weights wherein is expressed by the product of two matrices and , that is .

[0016] The low-rank adaptive module for features directly acts on the feature maps between the residual blocks and is used for constraining the model to learn low-rank features.

[0017] The processing process of the low-rank adaptive module for features is: a convolution is used on the input features of the residual block.Channel dimensionality reduction is performed, then amplitude adjustment is performed by instance regularization, then the number of channels is expanded by convolution Convolution , and finally the original features are added, and the calculation formula is:

[0018] ;

[0019] ;

[0020] wherein, represents the residual block mapping, is the output of the instance regularization, and respectively represent the mean and variance of the input channel; is the first is to prevent division by zero variables; and are the learnable scaling and bias parameters of the instance regularization, respectively, to enhance the model expression ability, is the feature low-rank adaptive module output.

[0021] The loss function of the noise recognition model is a cross-entropy loss function, and Mixup data enhancement is combined to improve the generalization ability of the model, wherein the loss function of the noise recognition model is:

[0022] ;

[0023] wherein, and are two pairs of samples randomly sampled from the data set, is a mixing coefficient, is a hyperparameter, represents the network, is the network prediction result, is a mixed sample, is a mixed label, is the total number of categories. is the label of the th category generated by mixing, is the network output probability of the th category.

[0024] The benchmark model is ResNet18, which includes an initial convolution-batch regularization block-RELU activation function block, a plurality of residual blocks, a global average pooling layer, and a fully connected classification layer.

[0025] The application also proposes a ship radiated noise recognition system based on a double low-rank adjustment network, which is realized based on the above method, comprising:

[0026] The data acquisition module is used to receive ship radiated noise signals;

[0027] The feature extraction module is a Mel filter bank used to obtain the Mel time-frequency spectrum of the ship's radiated noise signal;

[0028] The noise recognition module, equipped with a noise recognition model, is used to receive the Mel time-frequency spectrum and output the target category of ship radiated noise;

[0029] The low-rank adaptive weight module is used to perform low-rank adaptation on the linear layer weights in the residual block. It achieves weight fine-tuning by freezing the pre-trained weights and building a trainable low-rank weight matrix next to them.

[0030] The low-rank adaptive feature module is used to perform low-rank adaptation on feature maps between residual blocks. It compresses and reconstructs features by directly learning low-rank features as residuals of the original features, thereby constraining the model to learn low-rank features.

[0031] The training and optimization module is used to train the model based on the cross-entropy loss function and combine it with Mixup data augmentation.

[0032] The baseline network structure module, equipped with the baseline model, is used to provide a pre-trained baseline network structure. The baseline model is ResNet18, which includes an initial convolution-batch regularization block-ReLU activation function block, four residual blocks, a global average pooling layer, and a fully connected classification layer.

[0033] Compared with existing technologies, the advantages of this application are:

[0034] (1) The low-rank adaptive weight module of this application draws on the idea of ​​fine-tuning training of current large models and constructs a trainable low-rank weight matrix next to the weights of the linear layers. During training, only the low-rank matrix needs to be updated, thereby compressing the number of trainable parameters and reducing the risk of overfitting.

[0035] (2) In the testing phase, this application uses structural reparameterization technology to merge low-rank weights with pre-trained weights, thus eliminating the need for additional inference overhead. The feature low-rank adaptive module targets the low-rank characteristics of the ship radiated noise time-frequency spectrum in the feature domain. It directly constrains and optimizes the intermediate feature map by explicitly constructing a low-rank feature adjustment branch, guiding the model to learn the core feature patterns of the Mel time-frequency map.

[0036] (3) This invention addresses the problems of domain differences and insufficient generalization ability faced by deep learning models in ship radiated noise identification tasks. It proposes a transfer learning framework based on dual low-rank adaptation (weight low-rank adaptation and feature low-rank adaptation). The weight low-rank adaptation module freezes pre-trained weights through low-rank decomposition, training only small-parameter quantum matrices, thus reducing the risk of overfitting. The feature low-rank adaptation module first reveals the low-rank properties of ship radiated noise in the time-frequency domain and feature domain, and constrains feature learning by directly learning low-rank features as residuals of the original features. The proposed method significantly improves the accuracy of ship noise classification without increasing inference overhead, providing a new paradigm for transfer learning in data-scarce scenarios.

[0037] (4) This application fully considers the inherent low-rank property of the Mel time spectrum, deeply explores the potential of the pre-trained model, and effectively improves the transfer learning performance. Extensive experimental results on the ShipsEar and DeepShip datasets show that only 0.55% of additional parameters are introduced compared to the pre-trained model, demonstrating the superiority of the proposed method. Experimental results show that the proposed method significantly improves classification performance, highlighting its effectiveness in ship radiated noise identification. Attached Figure Description

[0038] Figure 1 The image shows a dual low-rank adjustment transfer learning framework.

[0039] Figure 2 The diagram shows the structure of the low-rank weight / feature adjustment module and the fully fine-tuned method.

[0040] Figure 3 The image shows the low-rank spectral analysis at Mel;

[0041] Figure 4 The image shows a low-rank analysis of the ResNet18 feature map;

[0042] Figure 5 The diagram shows the design of the weighted low-rank adaptive module and the feature low-rank adaptive module. Detailed Implementation

[0043] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0044] To address the insufficient generalization performance of deep learning models in ship radiated noise identification, this application proposes a weight-feature dual low-rank transfer learning framework. This framework achieves dual low-rank adjustment through two core modules: in the weight space, pre-trained weights are frozen, and a lightweight weight low-rank adaptive module constructs learnable low-rank adjustment increments, enabling fine-tuning of weights with fewer parameters; in the feature space, based on the inherent low-rank characteristics of the Mel-time spectrum of ship radiated noise, a feature low-rank adaptive module compresses and reconstructs features, thus explicitly constraining the model to learn low-rank features. This framework fully considers the inherent low-rank properties of the Mel-time spectrum, deeply explores the potential of pre-trained models, and effectively improves transfer learning performance. Experimental results on two public datasets, ShipsEar and DeepShip, show that compared to the pre-trained model, this framework introduces only 0.55% more parameters, achieving a state-of-the-art level.

[0045] like Figure 2 As shown, this application employs a dual low-rank adaptive transfer learning framework—combining low-rank weight adjustment and low-rank feature adjustment—to perform low-rank analysis of the Mel time-frequency spectrum in both the data and feature domains. The weight low-rank adaptive module constructs a trainable low-rank weight matrix alongside the linear layer weights. During training, only the low-rank matrix needs to be updated, thus compressing the number of trainable parameters and reducing the risk of overfitting. During testing, structural parameter re-parameterization is used to merge the low-rank weights with the pre-trained weights, eliminating the need for additional inference overhead. The feature low-rank adjustment addresses the low-rank characteristics of the ship radiated noise time-frequency spectrum in the feature domain by explicitly constructing a low-rank feature adjustment branch to directly constrain and optimize the intermediate feature map, guiding the model to learn the core feature patterns of the Mel time-frequency spectrum.

[0046] Ship radiated noise mainly includes three types: mechanical noise, propeller noise, and hydrodynamic noise. Among them, mechanical noise and propeller noise are mainly caused by the rotation of the main engine and propeller, and usually have strong periodicity. These periodic components are represented in matrix representation (such as the Mel time spectrum) as repeating patterns along the frequency or time dimension, thus leading to a reduction in matrix rank.

[0047] Example 1

[0048] To better illustrate this, consider the Mel time spectrum. Perform rank analysis.

[0049] Taking a 128-dimensional Mel feature as an example, the temporal dimension is usually larger than the Mel frequency dimension, therefore the full rank of a single sample is 128. Firstly, regarding... Perform singular value decomposition (SVD).

[0050]

[0051] In the formula, and These are left and right orthogonal matrices, For a singular value diagonal matrix, the elements on its diagonal are... These are called singular values. For matrix The rank of singular values. The ordering property of singular values ​​determines that the first few singular values ​​usually carry the main structural information, and the energy is accumulated by normalizing the singular values. The number of effective rank values ​​in the time spectrum can be determined. The normalized singular value cumulative energy is defined as...

[0052]

[0053] In the formula, For singular value indexes. When When the low-rank condition is met, only the first few rank members need to be selected. The original matrix can be approximately reconstructed using only a few singular values. The reconstruction process can be represented as follows:

[0054]

[0055] In the formula, , These are the left and right singular vectors, respectively.

[0056] like Figure 3 As shown, the spectral samples of radiated noise from different ships in the ShipsEar dataset at time Mel are... The normalized singular value cumulative energy curves are shown in the figure. It can be observed that different samples exhibit different harmonic structures or broadband noise components. However, their singular value cumulative distributions share similar characteristics: the singular value cumulative energy increases sharply with the index, exceeding 90% around dimension 20, indicating that only a 20-dimensional rank is needed to retain most of the information. The reconstruction results at different ranks in the experiment also illustrate this point: when the rank is 16, the reconstruction result is very close to the original Mel-time spectrum. This means that ship radiated noise has significant low-rank characteristics and information redundancy. Utilizing this low-rank characteristic can effectively reduce the network learning complexity and improve the model robustness.

[0057] The aforementioned rank analysis primarily focuses on the Mel-time spectrum. However, for deep learning models, feature space data also exhibits low-rank properties. There are two main reasons for this: First, the feature extraction process by the convolutional kernel through a sliding window operation is localized, allowing repetitive patterns from the original Mel-time spectrum to be preserved in the feature map, leading to feature map rank degradation. Second, CNN features inherently possess low rank, with strong linear correlations between different channels. For example... Figure 4As shown, the cumulative energy distribution of singular values ​​in the feature maps of different residual blocks of ResNet18 is displayed. It can be observed that different input categories (passenger ships, roll-on / roll-off ships, and sailboats) have similar rank distributions; for example, in residual block 1, the cumulative energy of the first 3-5 singular values ​​already accounts for 95%. This indicates that the requirements for ship radiated noise identification can be met without adjusting all channels.

[0058] Traditional transfer learning methods often fail to achieve ideal results in ship noise identification tasks due to data scarcity. Analysis of the low-rank characteristics of the Mel time-frequency spectrum and depth features reveals that model fine-tuning does not actually require optimization of all parameters. Based on this, this application proposes a transfer learning framework based on dual low-rank adaptation, the structure of which is as follows: Figure 1 As shown.

[0059] This paper uses ResNet18, pre-trained on ImageNet and performing well in underwater acoustic target classification, as the baseline model. The standard ResNet-18 consists of an initial convolutional-batch regularization block-ReLU activation block, four residual blocks, a global average pooling layer, and a fully connected classification layer. Building upon this, the paper proposes two low-rank modules: a weight low-rank adaptive module and a feature low-rank adaptive module. These modules impose low-rank constraints on both weights and features. The weight low-rank adaptive module operates on the linear layer weights within the residual blocks, while the feature low-rank adaptive module operates on the feature maps between residual blocks. The feature extraction weights of the baseline model are frozen during training; the entire training process only requires training the weight low-rank adaptive module, the feature low-rank adaptive module, and the classification layer.

[0060] The loss function employs cross-entropy loss, combined with Mixup data augmentation to improve the model's generalization ability. Mixup obtains virtual training samples through linear interpolation between samples, expanding the data distribution and prompting the model to learn smoother decision boundaries. The final loss function of the noise recognition model is... for:

[0061]

[0062]

[0063]

[0064]

[0065] in, and For two pairs of samples randomly sampled from the dataset, The mixing coefficient, For superparameters, Represents a network. For network prediction results, Mixed samples, Mixed tags, This represents the total number of categories. For the mixed generation of the first Labels for each category, For the first The network output probabilities for each category.

[0066] The core idea of ​​the weighted low-rank adaptive module is: it eliminates the need to train the model's original weights. Instead, it introduces a trainable low-rank weight. This is used to approximate weight updates. The fused weights of the model during training... It can be represented as

[0067]

[0068] in, Due to its low rank, it can be passed through two smaller matrices. and It is expressed as the product of, that is This decomposition significantly reduces the number of parameters that need to be trained. Taking a fully connected layer as an example, let the weights of a certain fully connected layer be... , Introduction , ,in , These represent the number of input and output neurons in the fully connected layer, respectively. The hyperparameter represents the rank of the weights of the low-rank adaptive module, which typically satisfies... .when When the value is small, compared to the original parameter size The number of parameters that need to be trained is only Smaller training parameters reduce the risk of overfitting, thereby improving the network's generalization ability.

[0069] The forward propagation process of the weighted low-rank adaptive module is as follows

[0070]

[0071] In the formula, For input, This is for output features. In actual implementation, Random initialization is used. Initializing to zero ensures that the output of the low-rank adaptive module is zero at the start of training, thus guaranteeing training stability. The forward propagation described above can be restated as follows:

[0072]

[0073] Therefore, during the inference phase, the weighted low-rank adaptive module reparameters can be transformed into new weights according to Equation 8. This avoids additional parameters and computational overhead. It's important to note that this reparameterization process only applies to linear layers; therefore, this paper only applies the weighted low-rank adaptive module to the convolutional and fully connected layers of the four residual blocks in ResNet18. For example... Figure 5 As shown, 'a' represents the detailed structure of the weighted low-rank adaptive module.

[0074] Weighted low-rank adaptive modules can be used in single-layer linear layers, but not in multi-layer structures (such as residual blocks). Due to the low-rank nature of the spectral feature map in Mel, the feature map of the network can be adjusted by designing a lightweight additional module, the weighted low-rank adaptive module.

[0075] The low-rank adaptive feature module is inserted as a separate layer into each layer of the baseline network, and its output directly affects the feature map during the forward propagation. Given a residual block S as input features... First through Layered convolution Perform channel dimensionality reduction

[0076]

[0077] In the formula, This is a channel compression feature. For the number of input features, The hyperparameter represents the number of features output by the convolution. , These represent the height and width of the feature map, respectively. Next, The amplitude is adjusted through an instance regularization layer (IN layer) to stabilize the training. Then, through... Layered convolution To expand the number of channels to the original number of channels Finally, it is added to the original features. This process can be represented as:

[0078]

[0079]

[0080] In the formula, Represents instance regularization operations. Represents residual block mapping, For IN layer output, and Representative input No. The mean and variance of each channel. Prevent division by zero. and The IN layer provides learnable scaling and bias parameters to enhance the model's expressive power. This is the output of the feature low-rank adaptive module. Since the feature low-rank adaptive module only contains two layers... Convolution, with the following number of parameters: ,when When the weights are small, the number of parameters in the feature-based low-rank adaptive module is negligible compared to the baseline network weights. For example... Figure 5 As shown, b is the detailed structure of the feature low-rank adaptive module.

[0081] This application also provides a ship radiated noise identification system based on a dual low-rank adjustment network, including:

[0082] The data acquisition module is used to receive ship radiated noise signals;

[0083] The feature extraction module is a Mel filter bank used to obtain the Mel time-frequency spectrum of the ship's radiated noise signal;

[0084] The noise recognition module, equipped with a noise recognition model, is used to receive the Mel time-frequency spectrum and output the target category of ship radiated noise;

[0085] The low-rank adaptive weight module is used to perform low-rank adaptation on the linear layer weights in the residual block of the noise recognition model. It achieves weight fine-tuning by freezing the pre-trained weights and building a trainable low-rank weight matrix next to them.

[0086] The low-rank adaptive feature module is used to perform low-rank adaptation on the feature maps between residual blocks of the noise recognition model. By directly learning low-rank features as residuals of the original features, the features are compressed and reconstructed to constrain the model to learn low-rank features.

[0087] The training and optimization module is used to train the model based on the cross-entropy loss function and combine it with Mixup data augmentation.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. A method for identifying ship radiated noise based on a dual low-rank adjustment network, comprising: The ship radiated noise signal is subjected to Mel filter bank transformation to obtain the Mel time-frequency spectrum of the noise signal; Input the Mel time-frequency spectrum into the trained noise recognition model to output the target category of ship radiated noise; The noise recognition model is generated based on a pre-trained benchmark model containing several residual blocks. The weight space is adapted to low rank through a weight low rank adaptive module, and the feature space is adapted to low rank through a feature low rank adaptive module. The low-rank adaptive feature module directly acts on the feature map between residual blocks to constrain the model to learn low-rank features; The processing procedure of the feature low-rank adaptive module is as follows: using convolution Input features of residual blocks Perform channel dimensionality reduction, then adjust the amplitude using instance regularization, and then... convolution To expand the number of channels, it is finally added to the original feature. The calculation formula is as follows: ； ； in, Represents residual block mapping, For the output of instance regularization, and Representing input respectively No. The mean and variance of each channel; To prevent division by zero; and These are the learnable scaling and bias parameters for instance regularization, used to enhance the model's expressive power. This is the output of the low-rank adaptive module.

2. The ship radiated noise identification method according to claim 1, characterized in that, The aforementioned low-rank adaptive weighting module operates on the linear layer weights in the residual block for weight fine-tuning; The processing procedure of the weighted low-rank adaptive module is as follows: freeze the pre-trained weights. Construct a trainable low-rank weight matrix next to the weights of the linear layer. As a learnable low-rank adjustment increment, the fusion weights are obtained. ,in Through two matrices and It is expressed as the product of, that is .

3. The ship radiated noise identification method according to claim 1, characterized in that, The noise recognition model uses a cross-entropy loss function, combined with Mixup data augmentation to improve its generalization ability. The loss function of the noise recognition model... for: ； in, and These are two pairs of samples randomly sampled from the dataset. The mixing coefficient, For superparameters, Represents a network. For network prediction results, For mixed samples, For mixed tags, Total number of categories; For the mixed generation of the first Labels for each category, For the first The network output probabilities for each category.

4. The ship radiated noise identification method according to claim 1, characterized in that, The baseline model is ResNet18, which includes: initial convolution-batch regularization block-ReLU activation function block, several residual blocks, global average pooling layer and fully connected classification layer.

5. A ship radiated noise identification system based on a dual low-rank adjustment network, implemented according to the method described in any one of claims 1-4, characterized in that, include: The data acquisition module is used to receive ship radiated noise signals; The feature extraction module is a Mel filter bank used to obtain the Mel time-frequency spectrum of the ship's radiated noise signal; The noise recognition module, equipped with a noise recognition model, is used to receive the Mel time-frequency spectrum and output the target category of ship radiated noise; The low-rank adaptive weight module is used to perform low-rank adaptation on the linear layer weights in the residual block of the noise recognition model. It achieves weight fine-tuning by freezing the pre-trained weights and building a trainable low-rank weight matrix next to them. The low-rank adaptive feature module is used to perform low-rank adaptation on the feature maps between residual blocks of the noise recognition model. By directly learning low-rank features as residuals of the original features, the features are compressed and reconstructed to constrain the model to learn low-rank features. and The training and optimization module is used to train the model based on the cross-entropy loss function and combine it with Mixup data augmentation.

Images