A method for recognizing passive acoustic signals in the ocean based on a lightweight convolutional neural network

Abstract

This invention discloses a method for identifying marine passive acoustic signals based on a lightweight convolutional neural network. The method includes the following steps: constructing a marine passive acoustic signal dataset containing time-frequency spectral features and a low-frequency acoustic signal dataset containing low-frequency band time-frequency features; training an attention-optimized lightweight first-order fully convolutional detection network to obtain a full-band signal recognition model and a low-frequency signal recognition model; inputting the acoustic signal to be identified into the full-band and low-frequency signal recognition models respectively to obtain full-band and low-frequency signal recognition results, thereby obtaining the final acoustic signal recognition result. Compared with existing deep learning-based recognition algorithms, this invention is applicable to multi-target, wide-band scenarios, and has stronger recognition capabilities for low-frequency signals and small target signals; moreover, the model is lightweight, deployment-friendly, and suitable for long-term operation with limited computing power in underwater equipment, enabling real-time recognition.

Description

Technical Field

[0001] This invention relates to a signal recognition method in the field of underwater acoustic signal processing, and more particularly to a method for recognizing marine passive acoustic signals based on a lightweight convolutional neural network. Background Technology

[0002] As human exploration of the ocean deepens, collecting and analyzing vast amounts of marine data in complex marine environments, and transmitting it rapidly and reliably, has become an essential means of ocean development. Acoustic signals, as the primary carrier of marine information, exist in various forms, including marine background noise, active sonar signals, underwater acoustic communication signals, ship radiated noise, and marine animal signals. Their detection and identification are of great significance for ensuring maritime safety and protecting marine animals.

[0003] Underwater acoustic signal recognition is divided into two modules: feature extraction and signal recognition. For the feature extraction module, features should be extractable and distinguishable. Fourier transform is the most commonly used technique for extracting frequency domain features and serves as a bridge between the time and frequency domain features of a signal. However, many underwater acoustic signals are non-stationary, and extracting only frequency domain features is insufficient for their identification. Therefore, by combining time and frequency domain features and transforming them to a two-dimensional time-frequency space, a time-frequency-energy function triplet structure is obtained, which provides both time and frequency domain information and reflects the signal's local variation characteristics. The most commonly used time-frequency feature extraction method is the short-time Fourier transform.

[0004] The signal recognition module relies on traditional methods such as decision trees, Naive Bayes, and support vector machines. However, these methods are subject to numerous limitations and have low accuracy in complex underwater acoustic channels. Therefore, current underwater target recognition methods primarily rely on manual observation and listening to audio and image signals received by underwater acoustic signal equipment, combined with accumulated experience to complete manual identification. However, training skilled personnel is costly and difficult, and the identification work depends heavily on subjective judgment, lacking real-time discernment. Deep learning methods based on convolutional neural networks can better extract deep features and are widely used in target detection and recognition. Therefore, there is a need to develop an intelligent recognition method for passive marine acoustic signals that fully considers the characteristics of various underwater acoustic signals and is suitable for complex underwater acoustic channels. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a method for identifying marine passive acoustic signals based on a lightweight convolutional neural network. It employs an attention-optimized lightweight first-order fully convolutional network to extract key features with high discriminative power and good generalization ability from the target samples to be identified, and amplifies low-frequency local features to enhance the low-frequency feature extraction capability, thereby realizing the detection and identification of underwater acoustic signals.

[0006] The objective of this invention is achieved through the following technical solution.

[0007] I. A method for recognizing marine passive acoustic signals based on lightweight convolutional neural networks

[0008] Step 1: Construct a dataset of ocean passive acoustic signals containing the temporal spectral characteristics of the signals using sonar signal data;

[0009] Step 2: Construct a lightweight first-order fully convolutional detection network based on attention optimization. After pre-training the lightweight first-order fully convolutional detection network based on attention optimization using an image recognition dataset, retrain the pre-trained lightweight first-order fully convolutional detection network using an ocean passive acoustic signal dataset to obtain a full-band signal recognition model.

[0010] Step 3: Construct a low-frequency acoustic signal dataset containing low-frequency band time-frequency features using sonar signal data. After training the full-band signal recognition model using the low-frequency acoustic signal dataset, the low-frequency signal recognition model is obtained.

[0011] Step 4: Input the sound signal to be identified into the full-band signal recognition model and the low-frequency signal recognition model respectively. The models will output the full-band signal recognition result and the low-frequency signal recognition result respectively. Generate the final sound signal recognition result based on the full-band signal recognition result and the low-frequency signal recognition result.

[0012] The attention-optimized lightweight first-order fully convolutional detection network includes a multi-scale feature extraction module, a feature fusion module, and a detection head connected in sequence.

[0013] The attention-based multi-scale feature extraction module includes a connected attention module and a multi-scale feature extraction backbone network. The multi-scale feature extraction backbone network is connected to the feature fusion module. The attention module includes a connected spatial attention layer and an unsupervised 3D attention layer. The unsupervised 3D attention layer is connected to the multi-scale feature extraction backbone network.

[0014] The feature fusion module has a feature pyramid structure.

[0015] In step four, the signal category and start-end frequency information of the sound signal to be identified are calculated based on the full-band signal identification result, and the signal category and start-end frequency information of the sound signal to be identified are calculated based on the low-frequency band signal identification result. The signal category and start-end frequency information corresponding to the two identification results are fused to obtain the final sound signal identification result.

[0016] In step three, the steps for extracting low-frequency band time-frequency features from sonar signal data include signal windowing and framing, Fourier transform, low-frequency bandwidth extraction, sound pressure level conversion, and data normalization.

[0017] II. A computer device

[0018] The device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method.

[0019] III. A computer-readable storage medium

[0020] The medium stores a computer program that, when executed by a processor, implements the steps of the method.

[0021] IV. A computer program product

[0022] The product includes a computer program / instructions that, when executed by a processor, implement the steps of the method.

[0023] The beneficial effects of this invention are:

[0024] 1. By using a first-order fully convolutional detection network, this invention can not only identify the category of target signals and obtain signal duration and frequency information, but also is applicable to multi-target, wide-bandwidth scenarios and has a high accuracy rate in identifying small target signals.

[0025] 2. This invention improves the network's ability to extract frequency information by introducing an optimized spatial attention mechanism module, thereby enhancing the identification of underwater acoustic signals in different frequency bands.

[0026] 3. This invention effectively improves the recognition capability of various low-frequency underwater acoustic signals by locally amplifying the time-frequency characteristics of the low-frequency band and performing post-feature fusion. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] Figure 1 This is a flowchart illustrating the marine passive acoustic signal recognition method based on a lightweight convolutional neural network proposed in this invention.

[0029] Figure 2 This is a schematic diagram of the lightweight first-order fully convolutional network structure used in this invention;

[0030] Figure 3 This is a schematic diagram of the information fusion strategy between the low-frequency signal recognition model and the full-band signal recognition model proposed in this invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings.

[0032] This invention provides a method for recognizing marine passive acoustic signals based on a lightweight convolutional neural network, such as... Figure 1 As shown, the present invention includes the following steps:

[0033] Step 1: Construct a marine passive acoustic signal dataset containing the temporal spectrum characteristics of the signals using sonar signal data. In this embodiment, all collected sonar signal data, publicly available acoustic datasets, and some simulation data are acquired. The sonar signal data includes various types of underwater acoustic signals such as communication signals, active sonar signals, marine animal signals, and ship signals. The temporal spectrum characteristics of all the above signals are calculated, and then the marine passive acoustic signal dataset is generated.

[0034] Step one is as follows:

[0035] 1.1) Acquire data collected by underwater sonar equipment, including underwater acoustic communication signals, active sonar signals and ship radiated noise; acquire public dataset data, including marine animal signals and ship radiated noise; and simulate and generate supplementary datasets of some underwater acoustic communication signals and active sonar signals to balance the distribution of various signal samples.

[0036] 1.2) After preprocessing the above data, feature extraction is performed. The extraction of time-spectrum features mainly consists of four steps: signal windowing and framing, Fourier transform, sound pressure level conversion, and data normalization. Finally, a tensor suitable for network model input is obtained. The window function, frame length, number of Fourier points, and step size parameters are obtained by comparison based on actual conditions.

[0037] Step 2: Construct a lightweight first-order fully convolutional detection network based on attention optimization. After pre-training the lightweight first-order fully convolutional detection network based on attention optimization using a public image recognition dataset, the network is then retrained using a marine passive acoustic signal dataset to obtain a full-band signal recognition model.

[0038] like Figure 2 As shown, the attention-optimized lightweight first-order fully convolutional detection network includes a multi-scale feature extraction module based on attention, a feature fusion module, and a detection head connected in sequence.

[0039] The attention-based multi-scale feature extraction module includes a connected attention module and a multi-scale feature extraction backbone network. The multi-scale feature extraction backbone network is connected to the feature fusion module. The attention module includes a connected spatial attention layer and an unsupervised 3D attention layer. The unsupervised 3D attention layer is connected to the multi-scale feature extraction backbone network.

[0040] The attention module consists of a spatial attention layer (SA-Module) and an unsupervised 3D attention layer (SimAM). The computation process of the spatial attention layer for feature map F is as follows:

[0041]

[0042] Among them, M space (F) represents the weights of the feature maps at the spatial scale, M space (F)∈R H×W σ() represents the activation function sigmoid, f 7×7 () indicates a convolution with a kernel size of 7×7. This represents the feature map after average pooling along the channel dimension. This represents the feature map after max pooling along the channel dimension. H and W represent the width and height of the feature map, respectively. AvgPool() represents the average pooling operation, and MaxPool() represents the max pooling operation.

[0043] The multi-scale feature extraction backbone network selects ShuttleNetV2 as the backbone network. In order to increase the number of channels without significantly increasing the computational cost, ShuttleNetV2 adopts two techniques: pointwise group convolution and a bottleneck-like structure. In addition, ShuttleNetV2 introduces channel shuffling operation, which enables information exchange between channels of different groups and improves the network's feature extraction capability.

[0044] The feature fusion module employs a lightweight feature pyramid structure, which can propagate features both top-down and bottom-up. For feature maps at different levels, higher-level feature maps are upsampled using nearest-neighbor interpolation to fuse adjacent feature maps, thus completing the top-down feature propagation. Lower-level feature maps are dimensionality-reduced through convolution to match higher-level features.

[0045] The detection head selects each layer of features and uses a set of 5×5 depthwise separable convolutions for computation to obtain the results of bounding box regression and signal classification.

[0046] In this embodiment, a lightweight first-order fully convolutional detection network based on attention optimization is trained using a teacher-student model framework. Both the teacher and student models are specifically structured as lightweight first-order fully convolutional detection networks based on attention optimization. The teacher model, acting as a training aid module, differs from the student model in that its detection head consists of four 3×3×256 convolutions, meaning the teacher model has more parameters than the student model, and these parameters are shared across different feature layers. During training, the inference results of the teacher model are used as soft labels to calculate the loss function, providing guidance for the training of the student model and improving training efficiency. Furthermore, the training aid module is discarded after training, without affecting the model's inference speed. The trained student model is then used as the signal recognition model.

[0047] Step 3: Due to the abundance of low-frequency signals and interference, a low-frequency auxiliary identification module was added to amplify local low-frequency features for better identification of low-frequency underwater acoustic signals. A low-frequency acoustic signal dataset containing low-frequency time-frequency features was constructed using sonar signal data. This dataset was then used to train a full-band signal recognition model, resulting in a low-frequency signal recognition model, which serves as the auxiliary identification module. The low-frequency band refers to frequencies below 2kHz. The steps for extracting low-frequency time-frequency features from the sonar signal data include signal windowing and framing, Fourier transform, low-frequency bandwidth extraction, sound pressure level conversion, and data normalization. Larger frame length and Fourier point count parameters were selected to improve frequency resolution, resulting in a low-frequency feature dataset for passive acoustic signals.

[0048] Step 4: Input the sound signal to be identified into the full-band signal recognition model and the low-frequency signal recognition model respectively. The models will output the full-band signal recognition result and the low-frequency signal recognition result respectively. Generate the final sound signal recognition result based on the full-band signal recognition result and the low-frequency signal recognition result.

[0049] Specifically, the final acoustic signal recognition result is generated based on the full-band signal recognition result and the low-frequency band signal recognition result, as follows:

[0050] The signal category and start / end frequency information of the acoustic signal to be identified are calculated based on the full-band signal recognition results, and the signal category and start / end frequency information of the acoustic signal to be identified are calculated based on the low-frequency band signal recognition results. The signal category and start / end frequency information corresponding to the two recognition results are fused to obtain the final acoustic signal recognition result. Figure 3 As shown, the result of fusion is the union of the two. If the target signal is a low-frequency signal, the inference result of the latter shall prevail.

[0051] The format of the signal recognition result is [class,x c ,y c[,width,height,score], where class represents the signal category, x c ,y c Let x and y represent the normalized x and y coordinates of the detection box center, respectively. width and height represent the normalized width and height of the detection box, respectively. score represents the signal confidence score. The detection box coordinates are transformed to obtain the signal's initial frequency f. start and termination frequency f end .

[0052] f start =(y c -height / 2)*f s / 2

[0053] f end =(y c +height / 2)*f s / 2

[0054] The steps above describe in detail how to start by acquiring a dataset of marine passive acoustic signals, construct and train a lightweight first-order detection network with attention optimization, and add an auxiliary recognition module to ultimately achieve the recognition of underwater acoustic signals. This process involves multiple stages, including data processing, feature extraction, network optimization, model training, and information fusion, each aimed at improving the accuracy and efficiency of underwater acoustic signal recognition.

[0055] Finally, it should be noted that the above embodiments and descriptions are only used to illustrate the technical solutions of the present invention and not to limit it. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the disclosure of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for recognizing marine passive acoustic signals based on a lightweight convolutional neural network, characterized in that, The steps include the following: Step 1: Construct a dataset of ocean passive acoustic signals containing the temporal spectral characteristics of the signals using sonar signal data; Step 2: Construct a lightweight first-order fully convolutional detection network based on attention optimization. After pre-training the lightweight first-order fully convolutional detection network based on attention optimization using an image recognition dataset, retrain the pre-trained lightweight first-order fully convolutional detection network using an ocean passive acoustic signal dataset to obtain a full-band signal recognition model. Step 3: Construct a low-frequency acoustic signal dataset containing low-frequency band time-frequency features using sonar signal data. After training the full-band signal recognition model using the low-frequency acoustic signal dataset, the low-frequency signal recognition model is obtained. Step 4: Input the sound signal to be identified into the full-band signal recognition model and the low-frequency signal recognition model respectively. The models will output the full-band signal recognition result and the low-frequency signal recognition result respectively. Generate the final sound signal recognition result based on the full-band signal recognition result and the low-frequency signal recognition result.

2. The method for marine passive acoustic signal recognition based on a lightweight convolutional neural network according to claim 1, characterized in that, The attention-optimized lightweight first-order fully convolutional detection network includes a multi-scale feature extraction module, a feature fusion module, and a detection head connected in sequence.

3. The method for marine passive acoustic signal recognition based on a lightweight convolutional neural network according to claim 2, characterized in that, The attention-based multi-scale feature extraction module includes a connected attention module and a multi-scale feature extraction backbone network. The multi-scale feature extraction backbone network is connected to the feature fusion module. The attention module includes a connected spatial attention layer and an unsupervised 3D attention layer. The unsupervised 3D attention layer is connected to the multi-scale feature extraction backbone network.

4. The method for marine passive acoustic signal recognition based on a lightweight convolutional neural network according to claim 2, characterized in that, The feature fusion module has a feature pyramid structure.

5. The method for marine passive acoustic signal recognition based on a lightweight convolutional neural network according to claim 1, characterized in that, In step four, the signal category and start-end frequency information of the sound signal to be identified are calculated based on the full-band signal identification result, and the signal category and start-end frequency information of the sound signal to be identified are calculated based on the low-frequency band signal identification result. The signal category and start-end frequency information corresponding to the two identification results are fused to obtain the final sound signal identification result.

6. The method for marine passive acoustic signal recognition based on a lightweight convolutional neural network according to claim 1, characterized in that, In step three, the steps for extracting low-frequency band time-frequency features from sonar signal data include signal windowing and framing, Fourier transform, low-frequency bandwidth extraction, sound pressure level conversion, and data normalization.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 6.

Images