Acoustic signal recognition method, electronic device and storage medium

Abstract

This application discloses an acoustic signal recognition method, electronic device, and storage medium, including: acquiring a time-domain acoustic signal; inputting the time-domain acoustic signal into an SC-CQT feature extraction module to obtain an SC-CQT spectrogram; inputting the SC-CQT spectrogram into a DAAR feature enhancement module to obtain an enhanced feature map; and inputting the enhanced feature map into an ABN dual-focus network module to obtain the category of the acoustic signal. This application combines SC-CQT feature extraction, DAAR feature enhancement, and ABN hierarchical network classification to form a complete and efficient underwater acoustic signal recognition process. By leveraging CQT's advantages in low-frequency analysis, key information is enhanced through sub-band weighting; the DAAR module is used to achieve intelligent feature focus and multi-semantic understanding; and the ABN architecture is used to achieve efficient decision-making from a global overview to local, detailed analysis, thereby improving the recognition accuracy and robustness of complex ship acoustic signals.

Description

Technical Field

[0001] The present application relates to the field of underwater acoustic recognition, and in particular to an acoustic signal recognition method, electronic equipment, and storage medium. Background Art

[0002] Acoustic signals, as an important physical phenomenon, are widely used in various fields, including industrial monitoring and environmental assessment. In the field of underwater acoustic recognition, the effective identification of radiated noise from underwater targets, such as ships, is key to target detection, tracking, and classification. However, real acoustic signals often exhibit non-stationarity, multi-component aliasing, and low signal-to-noise ratio. Traditional signal processing methods, such as the short-time Fourier transform (STFT) and its derived Mel-frequency cepstral coefficients (MFCC), often perform poorly when processing such complex signals due to their fixed time-frequency resolution or insufficient capture of low-frequency information.

[0003] In recent years, deep learning technology has made significant progress in pattern recognition and has also been applied to acoustic signal recognition tasks. However, existing deep learning models still face several challenges when applied to acoustics: first, how to design front-end feature extraction methods that better reflect the physical characteristics of acoustic signals to fully exploit the effective information in the signals; second, how to build systems that can intelligently focus on key information areas and effectively suppress interference; and third, how to design network architectures that balance recognition accuracy and computational efficiency to meet the needs of different application scenarios.

[0004] Therefore, developing a new acoustic signal recognition method that can provide refined time-frequency representation, realize intelligent feature enhancement, and adopt an efficient hierarchical decision-making mechanism is of great significance for improving the recognition performance of complex underwater acoustic signals. Summary of the Invention

[0005] In order to achieve the above objectives, the present application provides an acoustic signal recognition method, an electronic device and a storage medium.

[0006] The technical solution of this application is as follows:

[0007] A method for recognizing an acoustic signal, comprising:

[0008] S1. Collect the original acoustic vibration signal and obtain the digital time domain acoustic signal through preprocessing;

[0009] S2. Input the time domain acoustic signal into the SC-CQT feature extraction module to obtain the SC-CQT spectrogram;

[0010] S3, inputting the SC-CQT spectrum into the DAAR feature enhancement module for feature enhancement to obtain an enhanced feature map;

[0011] S4. Inputting the enhanced feature map into the ABN dual-focus network module for recognition to obtain the category of the acoustic signal;

[0012] in:

[0013] The SC-CQT feature extraction module performs:

[0014] Performing a constant Q transform on the time-domain acoustic signal to obtain a complex CQT coefficient;

[0015] Calculating a logarithmic CQT spectrum based on the complex CQT coefficients;

[0016] Dividing the logarithmic CQT spectrum into N frequency sub-bands along the frequency axis to generate N sub-generation spectrograms, where N≥2;

[0017] Multiply all frequency points of each daughter spectrum by the corresponding weight coefficient;

[0018] The weighted N daughter spectra are cascaded from low to high along the frequency axis to obtain the SC-CQT spectrum;

[0019] The DAAR feature enhancement module performs:

[0020] The input SC-CQT spectrogram is adaptively averaged and pooled along the frequency axis and time axis respectively to generate a one-dimensional frequency feature sequence and a one-dimensional time feature sequence;

[0021] Performing multi-scale depth-wise separable convolution on the frequency feature sequence and the time feature sequence to obtain a first semantic feature and a second semantic feature;

[0022] Performing attention enhancement processing on the first semantic feature, the second semantic feature, and the input SC-CQT spectrogram to obtain an enhanced feature map;

[0023] The ABN dual focus network module includes:

[0024] A base network Base-Net is used to perform a convolution operation on the enhanced feature map to output mid-level features;

[0025] Overview-Net, which performs at least two convolutional downsamplings on the mid-level features output by the base network Base-Net to generate global contextual prior features. It also performs global average pooling on the output of the last downsampling and generates auxiliary classification results through a fully connected layer.

[0026] The focusing network Focus-Net is used to concatenate the mid-level features and the global context prior features to obtain the main classification result.

[0027] In some embodiments, N=4, and the four frequency sub-bands are: first sub-band 20-150 Hz, second sub-band 150-800 Hz, third sub-band 800-3000 Hz, and fourth sub-band 3000-8000 Hz; the weight coefficients are: first sub-band 1.3, second sub-band 1.0, third sub-band 1.0, and fourth sub-band 0.9.

[0028] In some embodiments, the attention enhancement processing includes: respectively passing the first semantic feature and the second semantic feature through the Sigmoid activation function to generate a frequency attention map and a time attention map; copying and expanding the frequency attention map along the time axis to match the size of the SC-CQT spectrum; copying and expanding the time attention map along the frequency axis to match the size of the SC-CQT spectrum; performing element-wise multiplication on the expanded frequency attention map and time attention map to obtain a combined attention map; performing element-wise multiplication on the combined attention map and the SC-CQT spectrum to obtain an enhanced feature map.

[0029] In some embodiments, the convolution kernel size of the multi-scale depthwise separable convolution is 3, 5, 7, or 9, and the multi-scale output features are cascaded along the channel dimension and fused by 1×1 convolution.

[0030] In some embodiments, the base network Base-Net is composed of 3 convolution blocks, wherein the 1st and 3rd convolution blocks contain convolution layers with a stride of 2 to achieve downsampling, and each convolution block includes a convolution layer, a batch normalization layer and a ReLU activation function; the overview network Overview-Net performs two convolution downsamplings with a stride of 2 on the middle-level features; the focusing network Focus-Net performs: bilinear upsampling of the global context prior features; applies 1×1 convolution to the middle-level features and the sampled prior features respectively, and adjusts the number of channels to the same value; and splices the middle-level features and the global context prior features after channel alignment along the channel dimension.

[0031] In some embodiments, when training the ABN dual focus network module: calculating the cross entropy loss of the main classification result output by the focus network Focus-Net ; Calculate the cross entropy loss of the auxiliary classification results output by the overview network Overview-Net ; According to total loss Backpropagation optimizes the network, where .

[0032] A computer device comprises a memory, a processor and a computer program stored in the memory, wherein the processor implements any of the above-mentioned acoustic signal recognition methods when executing the program.

[0033] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements any of the above-mentioned acoustic signal recognition methods.

[0034] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0035] 1. Improve low-frequency analysis and characterization capabilities: The SC-CQT feature extraction module uses CQT's logarithmic frequency axis characteristics and targeted sub-band weighted cascade to more accurately capture and enhance low-frequency information, providing high-quality feature input for subsequent processing.

[0036] 2. Intelligent feature focusing and enhancement: The DAAR feature enhancement module can adaptively highlight key time-frequency patterns in feature maps through decoupled dimensional processing, multi-scale semantic extraction, and a combined attention mechanism, effectively improving the signal-to-noise ratio and discriminability of features.

[0037] 3. Efficient and accurate hierarchical decision-making: The ABN dual-focus network module architecture draws on human cognitive mechanisms. Through "overview-focus" hierarchical processing and using contextual priors to guide the deep network, it achieves a good balance between computational efficiency and recognition accuracy, enhancing the practicality and robustness of the model.

[0038] The present application proposes a method that is particularly suitable for processing acoustic signals with significant low-frequency characteristics and complex time-varying characteristics, such as underwater target radiation noise, and is particularly suitable for underwater ship radiation noise identification scenarios with a signal-to-noise ratio below 10dB and significant non-stationarity. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the embodiments. The text descriptions of these drawings are as follows, and the actual drawing can be based on these descriptions:

[0040] Figure 1 This is a flowchart of the execution of the acoustic signal recognition method in an embodiment of the present application.

[0041] Figure 2 Schematic diagram of the execution process of the acoustic signal recognition method in an embodiment of the present application.

[0042] Figure 3 This is the feature extraction process of the SC-CQT feature extraction module in the embodiment of this application.

[0043] Figure 4 This is the feature enhancement process of the DAAR feature enhancement module in the embodiment of the present application.

[0044] Figure 5 Schematic diagram of the hierarchical structure of the ABN dual-focus network module in an embodiment of the present application.

[0045] Figure 6 Schematic diagram of the structure of the computer device in the embodiment of the present application.

[0046] Figure 7 t-SNE visualization results of different models on the DeepShip dataset and ShipsEar dataset.

[0047] Figure 8 Confusion matrix of different models on DeepShip dataset. DETAILED DESCRIPTION

[0048] It should be noted that the contents of the following detailed description are exemplary and are intended to provide an indicative description of the contents of this application. It should be noted that all technical and scientific terms used in this application have the same meaning as commonly understood by ordinary technicians in the technical field to which the application belongs.

[0049] The following will be combined with the accompanying drawings in the embodiments of this application to provide a clear and complete description of the system architecture in the embodiments of this application and the solutions in the prior art. It should be noted that the described embodiments are only for the purpose of explaining and illustrating this application, and are not the complete content. Based on the embodiments provided in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work are within the scope of protection of this application.

[0050] Example 1:

[0051] like Figure 1 and Figure 2 As shown, this application provides a technical solution:

[0052] A method for recognizing an acoustic signal comprises the following steps:

[0053] S1. Collect the original acoustic vibration signal and obtain the digital time domain acoustic signal through preprocessing.

[0054] S2. Input the time domain acoustic signal into the SC-CQT feature extraction module to obtain the SC-CQT spectrogram.

[0055] S3. Input the SC-CQT spectrum into the DAAR feature enhancement module for feature enhancement to obtain an enhanced feature map.

[0056] S4. Input the enhanced feature map into the ABN dual-focus network module for recognition to obtain the category of the acoustic signal.

[0057] In the acoustic signal recognition method of the present application, the SC-CQT feature extraction module extracts features from the time-domain acoustic signal through the logarithmic frequency axis characteristics of CQT and targeted sub-band weighted cascade, thereby more accurately capturing and enhancing low-frequency information, providing high-quality feature input for subsequent processing. The DAAR feature enhancement module adaptively highlights the key time-frequency patterns in the feature map through decoupled dimensional processing, multi-scale semantic extraction and combined attention mechanism, thereby effectively improving the signal-to-noise ratio and discriminability of the features. The ABN dual-focus network module draws on the human cognitive mechanism, through the hierarchical processing of "overview-focus", and uses contextual priors to guide the deep network, thereby achieving a good balance between computational efficiency and recognition accuracy, enhancing the practicality and robustness of the model.

[0058] In summary, the present embodiment combines SC-CQT feature extraction, DAAR feature enhancement, and ABN hierarchical network classification to form a complete and efficient underwater acoustic signal recognition process. This process fully leverages CQT's advantages in low-frequency analysis, enhancing key information through sub-band weighting; utilizes the DAAR module to achieve intelligent feature focus and multi-semantic understanding; and leverages the ABN architecture to achieve efficient decision-making from global overview to localized, detailed analysis, significantly improving the recognition accuracy and robustness of complex ship acoustic signals.

[0059] The process of obtaining the time domain acoustic signal in step S1 is as follows:

[0060] The original acoustic vibration signal (for example, containing the noise of the ship's main engine or propeller) is collected by acoustic sensors (such as sonar and radar) and preprocessed to obtain a time domain acoustic signal.

[0061] Preprocessing includes filtering, signal amplification, and analog-to-digital conversion (ADC) to obtain a digital time-domain acoustic signal sequence, which is recorded as .

[0062] Example 2:

[0063] like Figure 2 and Figure 3 As shown, based on Example 1, the SC-CQT feature extraction module (hereinafter referred to as the SC-CQT module) in step S2 extracts features from the time domain acoustic signal to obtain the SC-CQT spectrogram in the following specific process:

[0064] S21. Perform constant Q transform processing on the time domain acoustic signal to obtain complex CQT coefficients.

[0065] The digitized time domain acoustic signal sequence Apply constant Q transform (CQT, Constant Q Transform). Constant Q transform parameters, minimum analysis frequency (Ship noise can be set to 20Hz), the number of frequency bands per octave (set to 24 for adequate low-frequency resolution), and the frame shift or jump length , the output of the constant Q transform is the complex CQT coefficient matrix.

[0066] The first in the complex CQT coefficient matrix frequency channels, CQT complex coefficients for time frames Obtained by the following formula:

[0067] ;

[0068] in, is the frequency channel index, is the timeframe index, For discrete time domain signals (sampling point index ), For the The complex conjugate window function of the frequency channels, For the Adaptive hopping step size of the channel, is the time offset, is the translation position of the window function on the time axis.

[0069] S22. Calculate the logarithmic CQT spectrum based on the complex CQT coefficients.

[0070] Specifically, first calculate the complex CQT coefficient The amplitude of the calculated amplitude is then taken logarithmically. The above process is implemented as follows: .

[0071] S23. Divide the logarithmic CQT spectrum into four frequency sub-bands along the frequency axis to obtain four sub-generation spectrograms.

[0072] Optionally, the sub-band boundaries are set according to the frequency range of noise generated by key components such as the ship's main engine and propeller. The four frequency sub-bands are: the first sub-band 20-150Hz, the second sub-band 150-800Hz, the third sub-band 800-3000Hz, and the fourth sub-band 3000-8000Hz.

[0073] The first sub-band 20-150 Hz is an extremely low frequency band, corresponding to the fundamental frequency of host vibration.

[0074] The second sub-band 150-800 Hz is a medium-low frequency band, covering the propeller blade frequency.

[0075] The third sub-band 800-3000 Hz is a mid-high frequency band, which contains mechanical harmonics.

[0076] The fourth sub-band 3000-8000 Hz is an extremely high frequency band, which contains cavitation high frequency noise.

[0077] Of course, in other embodiments, the logarithmic CQT spectrum may be divided into three, five, or other numbers of frequency sub-bands along the frequency axis according to specific needs, and correspondingly, three, five, or other numbers of child spectra may be obtained.

[0078] S24. Multiply all frequency points of each child spectrum by the corresponding weight coefficient.

[0079] In order to further highlight the frequency band information that contributes more to the recognition task, the divided sub-band spectrogram Specifically, a weight coefficient is assigned to each sub-band sub-generation spectrum graph. Usually, for the subband containing key low-frequency information, its weight The weight can be set larger to enhance its influence in subsequent analysis. For frequency bands that are relatively less important or may be subject to greater noise interference, their weight can be set to 1 or slightly less than 1.

[0080] For this reason, the weight coefficients assigned to the above four sub-generation spectrograms are: 1.3 for the first sub-band, 1.0 for the second sub-band, 1.0 for the third sub-band, and 0.9 for the fourth sub-band.

[0081] S25. Concatenate the four weighted daughter spectra along the frequency axis from low to high to obtain an SC-CQT spectrum.

[0082] That is to say, all weighted sub-band spectrograms The final sub-band concatenated CQT (Sub-band Concatenated- Constant QTransform, referred to as SC-CQT) spectrogram is formed by splicing along the frequency axis from low to high frequency. The SC-CQT spectrogram will be used as the input of the subsequent feature enhancement and classification network. Its dimension is usually ,in is the number of channels (usually 1, representing the spectrum itself), is the total number of frequency channels after splicing, is the number of time frames.

[0083] Example 3:

[0084] Underwater acoustic signals have independent characteristics along the frequency axis (ship line spectrum distribution) and the time axis (pulse duration). The frequency axis (or frequency dimension) primarily represents the fundamental frequency of the main engine vibration (stable), while the time axis (or time dimension) primarily represents the propeller cavitation pulse (transient). Therefore, the present embodiment introduces a dual-axis (or dual-dimension) network structure, also known as a dual-dimensional network structure.

[0085] The dual-dimension attention-driven adaptive refinement (DAAR) module combines a dual-axis network structure with attention-driven adaptive refinement to address the three major pain points of underwater acoustic recognition: low signal-to-noise ratio, signal non-stationarity, and signal multi-target mixing.

[0086] like Figure 2 and Figure 4 As shown, based on Example 1, the DAAR feature enhancement module (hereinafter referred to as DAAR module) in step S3 performs feature enhancement on the SC-CQT spectrum to obtain the enhanced feature map as follows:

[0087] S31. Adaptively average pool the input SC-CQT spectrogram along the frequency axis and the time axis respectively to generate a one-dimensional frequency feature sequence and a one-dimensional time feature sequence.

[0088] The input SC-CQT spectrum (which can be regarded as feature maps of channels ) is first fed into the DAAR module.

[0089] Inside the DAAR module, the input features are first The global context information is encoded independently along the frequency axis and the time axis. Adaptive average pooling is applied along the time axis and frequency axis respectively to implement encoding, and finally a one-dimensional frequency feature sequence is obtained. and one-dimensional time feature series .

[0090] S32, frequency feature sequence and time series Multi-scale depth-wise separable convolution is performed to obtain the first semantic feature and the second semantic feature.

[0091] The convolution kernel sizes of the multi-scale depth-wise separable convolution are 3, 5, 7, and 9, and the multi-scale output features are cascaded along the channel dimension and fused by 1×1 convolution.

[0092] Using multi-scale depth-wise separable convolution helps reduce the amount of computation and the number of parameters. Convolution kernels of different sizes can capture local dependencies from different receptive field scales.

[0093] For each dimension (frequency and time ), the output features from different convolution kernel sizes (i.e., sub-features with different semantics) are concatenated along the channel dimension. Then, these multi-scale features are usually fused through a 1×1 point convolution and the number of channels is adjusted to the same as the number of channels of the input features. Afterwards, Group Normalization (GN) is applied to stabilize the training process.

[0094] S33. Perform attention enhancement processing on the first semantic feature, the second semantic feature, and the input SC-CQT spectrogram to obtain an enhanced feature map.

[0095] The specific process is as follows:

[0096] The first semantic feature and the second semantic feature are respectively activated by the Sigmoid function to generate a frequency attention map and temporal attention maps .in, .

[0097] The frequency attention map Copy and expand along the time axis to match the size of the SC-CQT spectrogram, that is, the frequency attention map After the expansion is implemented, .

[0098] The temporal attention map Copy and expand along the frequency axis to match the size of the SC-CQT spectrogram, that is, the frequency attention map After the expansion is implemented, .

[0099] For the expanded frequency attention map and temporal attention maps Perform element-wise multiplication to obtain the combined attention map ,in, .

[0100] Combine the attention maps Perform element-wise multiplication with the input SC-CQT spectrogram and output an enhanced feature map.

[0101] This operation enables the model to adaptively reweight each element in the original feature map based on the multi-scale global contextual information extracted and fused independently from the frequency and time axes, thereby enhancing important time-frequency features and suppressing noise or irrelevant information.

[0102] Example 4:

[0103] like Figure 2 and Figure 5 As shown, based on Example 1, the ABN (Attentional Bifocal Network) dual focus network module (abbreviated as ABN module) in step S4 includes:

[0104] The base network Base-Net is used to perform convolution operations on the enhanced feature map to output mid-level features;

[0105] Overview-Net, which performs at least two convolutional downsamplings on the mid-level features output by the base network Base-Net to generate global contextual prior features. It also performs global average pooling on the output of the last downsampling and generates auxiliary classification results through a fully connected layer.

[0106] The focusing network Focus-Net is used to concatenate mid-level features and global context prior features to obtain the main classification results.

[0107] Specifically, the base network Base-Net consists of three convolutional blocks, where the first and third convolutional blocks contain convolutional layers with a stride of 2 to achieve downsampling. Each convolutional block includes a convolutional layer, a batch normalization layer, and a ReLU activation function to output mid-level features. ,in, .

[0108] Overview-Net outputs mid-level features of the base network Base-Net Perform two convolution downsamplings with a step size of 2 to generate global context prior features ,in, .

[0109] Global average pooling is performed on the last downsampling output, and the auxiliary classification result is generated through the fully connected layer.

[0110] Overview-Net is a relatively lightweight network. Further convolution and downsampling are performed. Overview-Net mainly completes two tasks: one is to generate global context prior features , through the convolution operation, the feature map with smaller spatial dimension and more concentrated information is obtained. The second is to output the logistic loss of auxiliary classification through an additional branch (global average pooling of its intermediate features followed by a fully connected layer). .

[0111] The focus network Focus-Net first analyzes the global context prior features output by the overview network Overview-Net Perform bilinear upsampling so that its spatial dimension is equal to matches the spatial dimensions.

[0112] Then, 1×1 convolution is applied to the middle-level features output by the base network Base-Net and the prior features after upsampling, and the number of channels is adjusted to the same preset number of channels. .

[0113] Finally, the channel-aligned mid-level features and the global context prior features are concatenated (fused) along the channel dimension, and the main classification results are output through the convolutional layer.

[0114] In other words, the Overview-Net extracts global contextual features and performs granular analysis. The Focus-Net, guided by global context, performs local, detailed recognition and fine-grained analysis. This dual-focus mechanism of "global overview + local focus" is the core of the network, hence its name, "dual-focus network."

[0115] According to the main classification output from the Focus-Net By applying the Softmax function, the probability distribution of each category can be obtained, and the category with the largest probability is taken as the final recognition result of the acoustic signal. During the training phase, the model usually uses the cross entropy loss function and combines it with the main loss (based on ) and auxiliary losses (based on , balanced by a weight factor λ) for end-to-end optimization.

[0116] Specifically, the training phase of the ABN dual-focus network module includes:

[0117] Calculate the cross entropy loss of the main classification result output by the focusing network Focus-Net ;

[0118] Calculate the cross entropy loss of the auxiliary classification results output by the overview network Overview-Net ;

[0119] By total loss Backpropagation optimizes the network, where .

[0120] Example 5:

[0121] This application provides a technical solution:

[0122] like Figure 6 As shown, a computer device includes a memory, a processor, and a computer program stored in the memory, and when the processor executes the program, the acoustic signal recognition method of any of the above embodiments is implemented.

[0123] Example 6:

[0124] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the acoustic signal recognition method of any of the above embodiments.

[0125] Experiment and result analysis:

[0126] This experiment was conducted on two publicly available datasets related to ship radiated noise: the DeepShip dataset and the ShipsEar dataset. These two datasets have been widely used in underwater acoustic recognition. More detailed information about the two datasets is provided below.

[0127] The DeepShip dataset contains data from 265 different ships from four different categories, totaling 47 hours and 4 minutes. The radiated noise was obtained using a ciliate AF hydrophone at a sampling frequency of 32,000 Hz. It mainly includes four common ship types: cargo ships, tankers, passenger ships, and tugboats. In the experiment of this application, the sample length was cut into 5 seconds, the samples overlapped for 3 seconds, and the total number of samples was 55,536. The dataset was first randomly divided into training and test sets in a ratio of 7:3, and then cut. The training set and test set contain 39,528 and 16,940 samples, respectively.

[0128] The ShipsEar dataset is derived from several ships cruising along the Atlantic coast, using DigitalHyD SR-1 recorders to capture radiated noise at a sampling frequency of 52,734 Hz. The dataset includes 90 recordings from 11 different ship types, totaling nearly three hours. The 11 ship types are categorized into four sizes and ship-radiated noise categories. Consequently, the dataset is ultimately divided into the following five categories:

[0129] Level A: Background noise.

[0130] Category B: Fishing vessels, trawlers, mussel boats, tugboats and dredgers.

[0131] Category C: Motorboats, pilot boats and sailboats.

[0132] Class D: Passenger ship.

[0133] Category E: Ocean-going passenger ships and ro-ro ships.

[0134] The ShipsEar dataset contains 2,223 samples, each 5 seconds long and non-overlapping. To prevent information leakage and ensure that the audio clips in the training and test sets come from different audio files, this application randomly splits the dataset into training and test sets in a 7:3 ratio, followed by a split process. Ultimately, the training set contains 1,556 samples, and the test set contains 667 samples.

[0135] The experiments were repeated 10 times on the two datasets to ensure the reliability of the experiments.

[0136] Table 1 lists the performance of the ABNet model (the recognition model provided in this application) on two test datasets, describing the model's evaluation metrics in terms of precision, recall, and F1 score. Based on these three evaluation metrics, the ABNet model achieved weighted average scores of 85.25%, 85.24%, and 0.8524 on the DeepShip dataset, respectively, while the ABNet model achieved weighted average scores of 96.42%, 96.41%, and 0.9642 on the ShipsEar dataset, respectively.

[0137] The performance of the ABNet model on the DeepShip dataset shows that the tanker class achieves the best F1-score. The performance of the ABNet model on the ShipsEar dataset shows that Class A (background noise) achieves the highest F1-score.

[0138] Compared to the DeepShip dataset, the ShipsEar dataset has a larger difference in sample size between categories, but the ABNet model achieves precision, recall, and F1-score exceeding 90% for each category. Therefore, the ABNet model is more suitable for imbalanced sample categories.

[0139] Table 1 Performance results of ABNet model on two datasets

[0140]

[0141] Figure 7The t-SNE visualization reveals the discriminative power of the features learned by the ABNet model. On the DeepShip dataset, while the ABNet model effectively achieves intra-class clustering, there is significant overlap between similar categories such as "cargo ship" and "tanker," suggesting less than ideal recognition accuracy. On the ShipsEar dataset, however, feature separation is excellent, with all categories (especially Class A noise, which is completely separated from the rest) forming independent, compact clusters. This intuitively demonstrates the ABNet model's superior feature learning capabilities in this scenario.

[0142] This experiment also compares the recognition performance of the ABNet model with six other representative models (AGNet, MSRDN, UALF, Swin Transformer, ConvNeXt, and CMoE).

[0143] For the DeepShip dataset, Table 2 shows the comparison results between the ABNet model and other models.

[0144] As shown in Table 2, the ABNet model achieves an 85.25% recognition accuracy on the DeepShip dataset, outperforming AGNet (77.09%), MSRDN (82.94%), UALF (82.39%), Swin Transformer (80.22%), ConvNeXt (83.54%), and CMoE (79.62%) by 8.16%, 2.31%, 2.86%, 5.03%, 1.71%, and 5.63%, respectively. Furthermore, the ABNet model also uses significantly fewer parameters and FLOPs than the other models.

[0145] Table 2 Performance comparison of different models on the DeepShip dataset

[0146]

[0147] For the ShipsEar dataset, Table 3 shows the comparison results of the ABNet model with six other representative models (AGNet, CRNN, ResNet 18, DRW-AE, ConvNeXt, CMoE).

[0148] As shown in Table 3, the recognition accuracy of the ABNet model on the ShipsEar dataset is 10.94%, 1.82%, 2.12%, 1.93%, 4.52% and 10.21% higher than that of AGNet, CRNN, ResNet 18, DRW-AE, ConvNeXt and CMoE, respectively.

[0149] The DRW-AE network has the lowest parameters and FLOPs, but its recognition accuracy is lower than that of the ABNet model. The ABNet model and CRNN model have similar parameters and FLOPs, but the ABNet model's recognition accuracy is higher than that of the CRNN model. Although the ABNet model's recognition accuracy is lower than that of the UATR-Transformer model, its parameter count and FLOPs are significantly lower than those of the UATR-Transformer model.

[0150] Therefore, the ABNet model achieves a good trade-off between recognition accuracy and model size on the ShipsEar dataset.

[0151] Table 3 Performance comparison of different models on the ShipsEar dataset

[0152]

[0153] Ablation experiments,based on the DeepShip dataset, analyze and evaluate the impact of SC-CQT, DAAR and ABN modules on the UTAR system. Figure 8 The confusion matrix for the six models on the DeepShip dataset is shown in Table 4. The overall performance of the different models is shown in Table 4. These models are labeled Models 1-6. Accuracy and F1-score are used to measure recognition performance.

[0154] Table 4 Performance comparison of different models on the DeepShip dataset

[0155]

[0156] Compared to Model 1, Model 2 only replaces the input features from standard CQT with SC-CQT, and both accuracy and F1-score improve. This indicates that SC-CQT, by incorporating prior knowledge of underwater acoustic physics and adaptively reconstructing the original spectrogram, enhances the discriminative characteristics of the signal, thereby providing a higher-quality input representation for the subsequent deep network.

[0157] Model 3 further incorporates the DAAR module, building on the SC-CQT. Compared to Model 2, both accuracy and F1-score are further improved, demonstrating that the DAAR module, through its decoupled dual-axis attention refinement mechanism, is able to further filter and enhance the temporal-frequency patterns of information from high-quality feature maps.

[0158] Model 2 and Model 4 both use SC-CQT as input features, but Model 4 replaces the standard ResNet-18 backbone network in Model 2 with an ABNet architecture. The accuracy increases from 78.91% to 82.13%. This suggests that ResNet-18 is not optimal for underwater acoustic target recognition. It may consume a large amount of computing power in insignificant background areas, while insufficiently exploring key areas.

[0159] Comparing Model 4 and Model 6, when the DAAR module is missing, the accuracy is only 82.13%, while the complete model reaches 85.25%, an accuracy improvement of 3.12%. This shows that dynamic channel calibration by decoupling the time and frequency context can effectively purify features and provide higher quality input for the subsequent ABN dual-focus network.

[0160] Comparing Model 5 and Model 6, after removing the Overview-Net from ABNet, the model's accuracy dropped from 85.25% to 82.49%, a performance loss of 2.76%. This demonstrates that the macro-contextual information provided by the Overview-Net is effective and necessary for achieving accurate classification, confirming the importance of the "overview first, focus later" synergistic mechanism for model integrity.

[0161] Table 4 also shows that the complete model including the SC-CQT module, DAAR module, and ABNet (with Overview-Net) module (i.e., the ABNet model of this application) has an accuracy and F1-score that are far superior to the other five models, indicating that the modules introduced in this invention have a synergistic effect in improving model performance, and the resulting effects can enhance each other.

[0162] The present invention has been described above in sufficient detail with certain particularities. Those skilled in the art will appreciate that the descriptions in the embodiments are merely illustrative, and that all modifications that do not depart from the true spirit and scope of the invention are intended to be within the scope of protection of the present invention. The scope of protection claimed in the present invention is defined by the appended claims, not by the foregoing description of the embodiments.

Claims

1. A method for recognizing acoustic signals, characterized in that: The acoustic signal recognition method comprises: S1. Collect the original acoustic vibration signal and obtain the digital time domain acoustic signal through preprocessing; S2. Input the time domain acoustic signal into the SC-CQT feature extraction module to obtain the SC-CQT spectrogram; S3, inputting the SC-CQT spectrum into the DAAR feature enhancement module for feature enhancement to obtain an enhanced feature map; S4. Inputting the enhanced feature map into the ABN dual-focus network module for recognition to obtain the category of the acoustic signal; in: The SC-CQT feature extraction module performs: Performing a constant Q transform on the time-domain acoustic signal to obtain a complex CQT coefficient; Calculating a logarithmic CQT spectrum based on the complex CQT coefficients; Dividing the logarithmic CQT spectrum into N frequency sub-bands along the frequency axis to generate N sub-generation spectrograms, where N≥2; Multiply all frequency points of each daughter spectrum by the corresponding weight coefficient; The weighted N daughter spectra are cascaded from low to high along the frequency axis to obtain the SC-CQT spectrum; The DAAR feature enhancement module performs: The input SC-CQT spectrogram is adaptively averaged and pooled along the frequency axis and time axis respectively to generate a one-dimensional frequency feature sequence and a one-dimensional time feature sequence; Performing multi-scale depth-wise separable convolution on the frequency feature sequence and the time feature sequence to obtain a first semantic feature and a second semantic feature; Performing attention enhancement processing on the first semantic feature, the second semantic feature, and the input SC-CQT spectrogram to obtain an enhanced feature map; The ABN dual focus network module includes: A base network Base-Net is used to perform a convolution operation on the enhanced feature map to output mid-level features; Overview-Net, which performs at least two convolutional downsamplings on the mid-level features output by the base network Base-Net to generate global contextual prior features. It also performs global average pooling on the output of the last downsampling and generates auxiliary classification results through a fully connected layer. The focusing network Focus-Net is used to concatenate the mid-level features and the global context prior features to obtain the main classification result.

2. The acoustic signal recognition method according to claim 1, wherein: N=4, the four frequency sub-bands are: first sub-band 20-150 Hz, second sub-band 150-800 Hz, third sub-band 800-3000 Hz, fourth sub-band 3000-8000 Hz; The weight coefficients are: 1.3 for the first sub-band, 1.0 for the second sub-band, 1.0 for the third sub-band, and 0.9 for the fourth sub-band.

3. The acoustic signal recognition method according to claim 1, wherein: The attention enhancement process includes: Pass the first semantic feature and the second semantic feature through the Sigmoid activation function to generate a frequency attention map and a time attention map respectively; The frequency attention map is copied and expanded along the time axis to match the size of the SC-CQT spectrogram; The temporal attention map is replicated and expanded along the frequency axis to match the size of the SC-CQT spectrogram; Perform element-wise multiplication on the expanded frequency attention map and the time attention map to obtain a combined attention map; The combined attention map is element-wise multiplied with the SC-CQT spectrogram to obtain an enhanced feature map.

4. The acoustic signal recognition method according to claim 3, wherein: The convolution kernel sizes of the multi-scale depthwise separable convolution are 3, 5, 7, and 9, and the multi-scale output features are cascaded along the channel dimension and fused by 1×1 convolution.

5. The acoustic signal recognition method according to claim 1, wherein: The base network Base-Net consists of three convolutional blocks, where the first and third convolutional blocks contain convolutional layers with a stride of 2 to achieve downsampling. Each convolutional block includes a convolutional layer, a batch normalization layer, and a ReLU activation function. The overview network Overview-Net performs two convolution downsamplings with a step size of 2 on the middle layer features; The Focus-Net performs: Bilinear upsampling of global context prior features; Apply 1×1 convolution to the middle-level features and the sampled prior features respectively, and adjust the number of channels to the same value; The channel-aligned mid-level features and global context prior features are concatenated along the channel dimension.

6. The acoustic signal recognition method according to claim 5, characterized in that: When training the ABN dual-focus network module: Calculate the cross entropy loss of the main classification result output by the focusing network Focus-Net ; Calculate the cross entropy loss of the auxiliary classification results output by the overview network Overview-Net ; By total loss Backpropagation optimizes the network, where .

7. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: When the processor executes the program, the acoustic signal recognition method according to any one of claims 1 to 6 is implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the acoustic signal recognition method according to any one of claims 1 to 6 is implemented.

Images