Single-channel underwater multi-target acoustic recognition method based on improved residual network

Abstract

The invention provides a single-channel underwater multi-target acoustic recognition method based on an improved residual network, and the method comprises the steps: obtaining an original acoustic signal of a to-be-recognized target, carrying out the preprocessing of the original acoustic signal to obtain a Mel-time spectrogram, inputting a multi-target recognition model which is trained in advance, and carrying out the recognition of the Mel-time spectrogram; a probability value of a [0, 1] interval is predicted for each possible target category, and the multi-target recognition model is composed of an improved ResNet network structure; a detection threshold value is preset for each possible target category, if the generation probability value is smaller than the preset detection threshold value, it is judged that the current scene does not have the corresponding target category, and if the generation probability value is larger than or equal to the preset detection threshold value, it is judged that the current scene has the corresponding target category. According to the method, the multi-target recognition capability of the model under a complex underwater sound scene and a multi-target signal aliasing condition is improved by performing deep mining on the feature information of the single Mel time-frequency spectrogram and combining multi-scale fusion and attention enhancement.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship radiation noise recognition, and particularly relates to a single-channel underwater multi-target acoustic recognition method based on an improved residual network. BACKGROUND

[0002] Underwater acoustic target recognition is a key technology in the field of ocean exploration and monitoring, and is of great significance for national defense security, ocean resource development and environmental protection. However, the complex and changeable marine environment and the non-stationary and diversity of ship target self-radiation noise make underwater acoustic target recognition face great challenges. In the underwater acoustic target recognition task, according to the number of targets in the scene, the recognition problem can be roughly divided into single-target recognition and multi-target recognition. In the ideal single-target scene, the acoustic sensor mainly receives the radiation noise signal from a single target, superimposed with a certain degree of marine environmental background noise. The signal source of this task is single, and the characteristics of the target signal are usually clear, and the recognition method can focus on constructing stable and discriminative feature representation.

[0003] However, in the actual application scene, there are usually multiple targets in the marine environment, thus forming a multi-target recognition problem. In this kind of scene, the signal received by the acoustic sensor is the superposition of the radiation noise of multiple different types of targets, accompanied by complex and changeable environmental noise. Multiple sound source signals inevitably overlap and interfere in time domain and frequency domain, resulting in overlapping, blurring and even mutual masking of the acoustic characteristics of different targets. The complexity of this multi-source signal overlap interference significantly increases the difficulty of feature extraction and multi-target classification, and becomes one of the important challenges in underwater target recognition research.

[0004] Single-channel hydrophones are widely used in underwater environment monitoring due to their simple structure and convenient deployment and recovery, but single-channel systems cannot utilize the spatial distribution difference of sound sources to separate and model the signals, so the radiation noise signals of all targets will be mixed and superimposed in the same channel. This superposition of multi-target signals in the time-frequency domain shows a high degree of overlap of multiple target characteristics, bringing great challenges to feature extraction and target differentiation under single-channel. In addition, in actual monitoring, the number and specific types of targets existing in the environment are usually unknown, which requires the recognition algorithm not only to recognize the target types, but also to judge which target combinations exist.

[0005] Therefore, developing a multi-label recognition method based on single-channel that can accurately recognize all these targets at the same time is of great importance for comprehensive understanding of underwater dynamics, improving monitoring and warning capabilities, and effectively managing marine activities. This not only provides more complete underwater acoustic environment situation awareness, but also lays a solid foundation for subsequent acoustic target state analysis. SUMMARY

[0006] The application aims to provide a single-channel underwater multi-target acoustic recognition method based on an improved residual network, aiming to solve the problem that the existing underwater multi-target recognition technology is limited in single feature extraction, and the traditional data enhancement method may introduce distorted data, thereby reducing the performance of the model.

[0007] To solve the above problems, the application provides a single-channel underwater multi-target acoustic recognition method based on an improved residual network, which comprises the following steps: Obtaining the original acoustic signal of the target to be recognized, and pre-processing the original acoustic signal to obtain a mel time-frequency spectrum; Inputting the mel time-frequency spectrum into a multi-target recognition model trained in advance to predict a probability value in the interval [0, 1] for each possible target class, wherein the multi-target recognition model is composed of an improved ResNet network structure; Presetting a detection threshold for each possible target class, if the generated probability value is less than the preset detection threshold, it is determined that the corresponding target class does not exist in the current scene, and if the generated probability value is greater than or equal to the preset detection threshold, it is determined that the corresponding target class exists in the current scene.

[0008] Further, the original acoustic signal is pre-processed, specifically including: Down-sampling the original acoustic signal to obtain a continuous audio stream; Dividing the continuous audio stream into non-overlapping short audio segments with a fixed time length; Normalizing each short audio segment, then performing short-time Fourier transform (STFT), and then calculating the mel spectrum through a mel filter bank.

[0009] Further, the normalized short audio segment is subjected to short-time Fourier transform, specifically including: Calculating the short-time Fourier transform and obtaining the power spectrum, the expression is:

[0010] wherein, is the power spectrum, is the original linear frequency, is the original audio segment, represents the result of the original signal after short-time Fourier transform processing, is the center position of the time window, i.e. the time point of short-time Fourier transform analysis, is the frequency variable in the short-time Fourier transform output result; Mapping the linear frequency f to the mel scale through a mel frequency conversion formula, the expression is:

[0011] The power spectrum of the audio segment is filtered through the mel filter bank to obtain a mel spectrogram.

[0012] Further, the ResNet network structure is improved, and specifically includes: The first 7x7 convolution kernel of the ResNet input layer is modified to a 3x3 convolution kernel, and the convolution step is adjusted to 1.

[0013] Further, the ResNet network structure is improved, and further includes: The improved ResNet network structure includes a backbone network, and the backbone network is composed of four convolution stages, and the four convolution stages include stage 1, stage 2, stage 3 and stage 4. Each convolution stage is sequentially stacked by a plurality of bottleneck residual blocks, and each bottleneck residual block internally extracts features through a plurality of convolution operations. The first bottleneck residual block of each convolution stage performs down-sampling on the second convolution layer and the residual branch.

[0014] Further, the stage 2, the stage 3 and the stage 4 of the convolution stage are used to extract acoustic feature maps of different levels. The acoustic feature maps output by the stage 2 and the stage 3 are subjected to an adaptive average pooling operation, and the size of the acoustic feature maps is adjusted to the size of the acoustic feature Figure One map output by the stage 4. The spatially aligned acoustic feature maps of the stage 2, the stage 3 and the stage 4 are spliced in the channel dimension to form a combined feature map containing multi-scale information.

[0015] Further, a channel attention module is applied to the spliced fusion combined feature map, and then a global adaptive average pooling layer is used to compress the spatial size to 1x1, and output a global feature map.

[0016] Further, the global feature map is flattened and converted into a one-dimensional feature vector, and then a fully connected layer with a Sigmoid layer at the back is used to generate a probability value in the interval [0, 1] for each predefined target class.

[0017] Further, a smoothness-induced regularization data enhancement strategy is used to enhance the training data of the multi-target recognition model.

[0018] Compared with the prior art, the beneficial effects of the present application are: The application provides a single-channel underwater multi-target acoustic recognition method based on an improved residual network. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of the provided drawings.

[0020] Figure 1 A flowchart of a single-channel underwater multi-target acoustic recognition method based on an improved residual network provided by the embodiment of the present application is shown in the figure. Figure 2 A pre-processing flowchart of a single-channel underwater multi-target acoustic recognition method based on an improved residual network provided by the embodiment of the present application is shown in the figure. Figure 3 An improved ResNet network structure diagram provided by the embodiment of the present application is shown in the figure. Figure 4 A macro average index comparison diagram of the improved ResNet network structure and ResNet18, ResNet34, ResNet50 and ResNet101 models is shown in the figure. Figure 5 A comparison diagram of overall accuracy of the improved ResNet network structure and ResNet18, ResNet34, ResNet50 and ResNet101 models is shown in the figure. Figure 6 An accuracy curve comparison diagram of the improved ResNet network structure and ResNet18, ResNet34, ResNet50 and ResNet101 models on a ShipEar verification set is shown in the figure. Figure 7 A verification set accuracy curve comparison diagram of the improved ResNet network structure in the case of using and not using smoothness-induced regularization data enhancement is shown in the figure. DETAILED DESCRIPTION

[0021] The present application will be further described in detail below in combination with the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present application, and not to limit the present application. In addition, it should be noted that, for the convenience of description, only the parts related to the present application are shown in the drawings, not all the structures.

[0022] The prior art solutions have made certain progress in the field of underwater acoustic target recognition. However, most of the current deep learning solutions are mainly designed and optimized for the recognition task when the environment only contains a single acoustic target. When multiple acoustic targets exist in the underwater environment at the same time, these single-target recognition frameworks are no longer applicable or the recognition performance (accuracy) is severely reduced.

[0023] Single-target recognition usually models the problem as a single-label classification problem (i.e., a sound sample only belongs to one class). Assuming that the time-domain acoustic signal received by the sensor at a certain moment is , it can be simply represented as:

[0024] where represents the target ( ∈ , is the set of all target types, and ) original radiation noise signal of the target, represents the ocean ambient noise, and this framework assumes that there are mutually exclusive target categories, and each input sample only corresponds to one determined target category. At this time, the goal of the recognition task is to determine the specific category of the target contained in according to the received signal .

[0025] Multi-target recognition is a typical multi-label classification problem (i.e., a sample can belong to multiple categories at the same time). The signal received by the sensor becomes a linear superposition of multiple target signals and ambient noise:

[0026] where is the set of actual target categories in the current scene, which is a subset of all possible target categories . Since multiple target signals are mixed at the sensor, the features between targets will be severely interfered and masked. At this time, the goal of the recognition task is to accurately identify the multiple target categories contained in , which belong to the set .

[0027] Currently, there are relatively few studies on underwater multi-target recognition based on single-channel acoustic features. Most existing studies use mel-spectrogram as input features because it can represent the local fine structure in the time domain and the global energy distribution in the frequency domain. However, the network architecture used in existing methods often fails to fully integrate and mine information from different levels and dimensions of input features, so there is still room for improvement in recognition ability when dealing with feature aliasing caused by multi-target signal superposition.

[0028] In addition, large-scale underwater acoustic multi-target data sets with real labels are very scarce, which greatly limits the training and generalization ability of deep learning models. In order to alleviate the lack of data, researchers usually use data enhancement methods to superimpose various noises on the data or perform time-frequency masking to improve the generalization performance of the model. However, data enhancement may introduce noise that does not match the real scene or change the key features of the original sample. If low-quality or distorted augmented data is directly used for model training, it is easy to mislead the model to learn false biased features, reducing the recognition performance and generalization ability of the model in real complex environments.

[0029] The purpose of the present application is to overcome the deficiencies of existing technologies, especially the classic residual network, in terms of multi-underwater acoustic target recognition accuracy and generalization performance, and to provide a single-channel underwater multi-target acoustic recognition method based on an improved residual network. The recognition model proposed in the present application focuses on optimizing the extraction and learning ability of a single acoustic feature (mel-frequency spectrogram).

[0030] The method aims to solve the problems of existing underwater multi-target recognition techniques in limited single feature extraction capability and traditional data enhancement methods that may introduce distorted data and thus reduce model performance. Specifically, the technical problems to be solved by the present application include: 1. By improving ResNet, different levels of information are extracted through branch setting and feature extraction and fusion, realizing fine feature fusion of the input feature map (mel-frequency spectrogram) at different scales and depths, and introducing a channel attention mechanism to adaptively enhance the most discriminative feature part. By fine multi-level analysis of a single acoustic feature map, the shortcomings of traditional single feature processing methods in terms of feature information utilization are overcome.

[0031] 2. To address the problem of scarcity of underwater acoustic multi-target data, a smoothness-induced regularization data enhancement strategy is used to improve the generalization ability of the model and reduce the impact of low-quality augmented data on the model.

[0032] To solve the above problems, with reference to Figure 1 The present embodiment provides a single-channel underwater multi-target acoustic recognition method based on an improved residual network, which comprises: An original acoustic signal of a target to be recognized is acquired, and the original acoustic signal is preprocessed to obtain a mel-frequency spectrogram.

[0033] The mel-frequency spectrogram is input into a multi-target recognition model trained in advance, and a probability value in the interval [0, 1] is predicted for each possible target category, the multi-target recognition model being composed of an improved ResNet network structure.

[0034] A detection threshold is preset for each possible target category, if the generated probability value is less than the preset detection threshold, it is determined that the corresponding target category does not exist in the current scene, and if the generated probability value is greater than or equal to the preset detection threshold, it is determined that the corresponding target category exists in the current scene.

[0035] In this embodiment, the original acoustic signal is preprocessed to obtain a mel-frequency spectrogram, an improved ResNet network structure is used for feature extraction and classification of the mel-frequency spectrogram, a multi-scale feature fusion mechanism is used to deeply mine time-frequency information, effectively integrate feature information from different levels of the network, and combine a channel attention module to adaptively enhance the expression of key features, and a data enhancement strategy of smoothness-induced regularization is introduced in the training process of the multi-target recognition model to improve the generalization ability of the multi-target recognition model.

[0036] As a preferred embodiment, the original acoustic signal is preprocessed, specifically including: The original acoustic signal is down-sampled to obtain a continuous audio stream.

[0037] The continuous audio stream is segmented into non-overlapping short audio segments of a fixed time length.

[0038] Each short audio segment is normalized, then subjected to a short-time Fourier transform (STFT), and then a mel-frequency spectrogram is calculated through a mel filter bank.

[0039] In this embodiment, the data processing flow is as shown in Figure 2 The original single-channel underwater acoustic signal (original acoustic signal) collected is preprocessed. The original acoustic signal is generally a time-domain waveform, a digital audio with a sampling frequency of 32 kHz or higher and a bit depth of 16 bits or higher. The data processing flow includes: down-sampling: the original audio signal is first down-sampled to 32 kHz; audio segmentation: the continuous audio stream after down-sampling is segmented into non-overlapping short audio segments of a fixed time length of 3 seconds; amplitude normalization: the sampling point values in each segment are divided by the maximum absolute value in the segment, so that the amplitude of each audio segment is scaled to the interval [-1, 1]; mel-frequency spectrogram generation: the normalized short audio segment is subjected to a short-time Fourier transform to obtain a power spectrum, and a mel-frequency spectrogram is calculated through a mel filter bank.

[0040] As a preferred embodiment, the normalized short audio segment is subjected to short-time Fourier transform, specifically including: The audio segment uses a Hanning window, a frame window size of 1024, and a step size of 754. The short-time Fourier transform is calculated to obtain a power spectrum, expressed as:

[0041] wherein, is the power spectrum, is the original linear frequency, is the original audio segment, represents the result of the original signal after short-time Fourier transform processing, is the center position of the time window, i.e., the time point of the short-time Fourier transform analysis, is the frequency variable in the short-time Fourier transform output result; A mel filter bank is designed. The linear frequency f is mapped to the mel scale by the mel frequency conversion formula, expressed as:

[0042] The mel filter bank is composed of a series of triangular filter banks, and the power spectrum of the audio segment is filtered by the mel filter bank to obtain a mel spectrogram.

[0043] As a preferred embodiment, the ResNet network structure is improved, specifically including: The first 7x7 convolution kernel of the ResNet input layer is modified to a 3x3 convolution kernel, and the convolution step size is adjusted to 1.

[0044] In this embodiment, the improved ResNet network structure is as shown in Figure 3 The improved ResNet network focuses on extracting multi-scale and discriminative features from single-channel mel time-frequency spectrograms. The input layer is adaptively modified: the first 7x7 convolution kernel of the standard ResNet-50 is modified to a 3x3 convolution kernel, the convolution step size is adjusted to 1, and the input channel number is modified to adapt to single-channel mel time-frequency spectrogram input.

[0045] As a preferred embodiment, the ResNet network structure is improved, and further includes: The improved ResNet network structure includes a backbone network, and the backbone network is composed of four convolution stages, including stage 1, stage 2, stage 3, and stage 4.

[0046] Each convolution stage is sequentially stacked by a plurality of bottleneck residual blocks, and each bottleneck residual block extracts features through multi-layer convolution operations.

[0047] The first bottleneck residual block of each convolution stage down-samples at its second convolution layer and residual branch.

[0048] In this embodiment, the improved ResNet network structure is as shown in Figure 3 The improved ResNet network backbone network mainly includes four convolution stages (stage 1 to stage 4). Each of the four convolution stages is sequentially stacked by a plurality of bottleneck residual blocks, and each bottleneck residual block extracts features through a plurality of convolution operations. The first bottleneck residual block of each of the four convolution stages down-samples at its second convolution layer and residual branch, so that the input feature map is reduced. With the deepening of the network, the spatial size of the feature map gradually decreases, and the channel dimension continuously increases.

[0049] As a preferred embodiment, stages 2, 3 and 4 of the convolution stage are used to extract acoustic feature maps of different levels.

[0050] The acoustic feature maps output by stages 2 and 3 are subjected to adaptive average pooling operation to adjust the feature size of the acoustic feature maps to be consistent with the acoustic feature maps output by stage 4.

[0051] The spatially aligned acoustic feature maps of stages 2, 3 and 4 are spliced in the channel dimension to form a combined feature map containing multi-scale information.

[0052] In this embodiment, the improved ResNet network structure is as shown in Figure 3 With the deepening of the network, the spatial size of the feature map gradually decreases, and the channel dimension continuously increases. The multi-target recognition model extracts acoustic feature maps of different levels from the last three main stages of the backbone network, i.e. stage 2 (outputting 512 channels), stage 3 (outputting 1024 channels) and stage 4 (outputting 2048 channels). To effectively integrate information of different scales, the acoustic feature maps output by stages 2 and 3 are subjected to adaptive average pooling operation to adjust their feature sizes to be consistent with the feature maps output by stage 4. Subsequently, the three groups of spatially aligned feature maps are spliced in the channel dimension to form a combined feature map containing rich multi-scale information, and the total number of channels is 3584.

[0053] As a preferred embodiment, a channel attention module is applied to the spliced combined feature map, and then a global adaptive average pooling layer is used to compress the spatial dimension to 1x1, and output a global feature map.

[0054] In this embodiment, the improved ResNet network structure is as shown in Figure 3As shown in the figure, a channel attention module is applied on the multi-scale combined feature map after splicing. The channel attention module learns the importance of each feature channel and adaptively weights these channels, thereby enhancing the feature information that contributes significantly to the classification task while suppressing the influence of redundant or noisy features. After the combined feature map is enhanced by the channel attention mechanism, it is compressed to 1x1 in spatial dimension through a global adaptive average pooling layer, while all 3584 feature channels are completely retained. The acoustic feature information extracted and enhanced in the previous step is efficiently processed into a compact global feature map for representation.

[0055] As a preferred embodiment, after the global feature map is flattened into a one-dimensional feature vector, a fully connected layer with a subsequent Sigmoid layer generates a probability value in the interval [0, 1] for each predefined target class.

[0056] In this embodiment, the improved ResNet network structure is as shown in the figure Figure 3 The global feature map with 3584 channels and spatial dimension of 1x1 is flattened and converted into a one-dimensional feature vector containing 3584 elements. The one-dimensional feature vector is then directly fed into a fully connected layer. The number of output nodes of the fully connected layer is set to the total number of predefined target classes. The output of the fully connected layer is then passed through a Sigmoid activation function to generate a probability value in the interval [0, 1] for each target class, representing the likelihood of the presence of the target class. A preset detection threshold is set for each possible target class. If the generated probability value is less than the preset detection threshold, it is determined that the corresponding target class does not exist in the current scene. If the generated probability value is greater than or equal to the preset detection threshold, it is determined that the corresponding target class exists in the current scene.

[0057] As a preferred embodiment, the training data of the multi-target recognition model is enhanced by the smoothness-induced regularization data enhancement strategy.

[0058] In this embodiment, to enhance the generalization ability of the multi-target recognition model in the case of scarce multi-target data and low data quality, the smoothness-induced regularization data enhancement strategy, which has been practiced in the field of underwater acoustic single-target recognition, is introduced to address the problem of scarce underwater acoustic multi-target data and the possible introduction of distorted data by traditional data enhancement methods. This strategy does not directly use enhanced samples to calculate classification loss, but rather constrains the consistency of the model's prediction output for the original sample and its enhanced version (through a symmetric KL divergence regularization term), prompting the multi-target recognition model to learn deeper and more essential features.

[0059] Specifically, for each original audio segment sample in the multi-target recognition model training set, an enhanced sample and its corresponding mel-frequency spectrum feature are generated by superimposing different signal-to-noise ratio Gaussian white noise. During model training, the original sample is used to calculate the binary cross-entropy loss (L) , and the KL divergence regularization loss function (L ) between the original sample and the enhanced sample is calculated. The total loss function is composed of two parts, L and L , and the expression of the total loss function is as follows:

[0060] wherein is the weight coefficient of the regularization term. The expression of the binary cross-entropy loss function is as follows:

[0061] wherein is the number of samples, is the label predicted by the model, is the true label. The symmetric KL divergence regularization term is as follows:

[0062] wherein is the output of the model for the original sample, is the output of the model for the enhanced sample.

[0063] To demonstrate the effect of the method described in the present application, a test was performed based on a public data set, and the specific steps are as follows: 1. Multi-target data generation: First, the underwater radiated noise signal of the target to be recognized is obtained. In this example, the specific implementation method is to use the ShipEar data set, which is a single-target data set, i.e., the audio segment contains only the radiated noise of one target, the sampling rate is 52734 Hz, and the three types of ships with the largest proportion are selected as the basic categories, plus natural noise, a total of 8 sub-categories can be generated, of which 4 are multi-target sub-categories. During data set production, the data is first down-sampled from the original 52734 Hz to 32000 Hz, then the audio is cut into segments with a duration of 3 seconds, and the corresponding audio segments are randomly selected from the corresponding basic categories according to the combination type of the multi-target; then the amplitude of the selected audio segments is normalized by the maximum value, and the audio of different categories is linearly superimposed to simulate a multi-target scene; finally, the amplitude of the superimposed audio is normalized by the number of targets.

[0064] 2、Multi-target data preprocessing: In the experiment, the synthesized multi-target data set is randomly divided into training set and validation set according to the ratio of 7:3. For all audio samples, the following processing flow is uniformly adopted for sample production: first, downsample all audio segments to 32kHz; then, randomly superimpose Gaussian white noise between -5dB and 5dB on the original data to obtain enhanced samples, which constitute a sample pair with the original samples. Then, perform short-time Fourier transform on the audio in the sample pair, with a window size of 1024 and a step size of 754, and use a Hanning window as the window function. Finally, process the transformed time-frequency graph through a Mel filter bank to generate a 128x128 resolution Mel time-frequency graph. To reduce training and inference costs, the Mel graph is compressed to 64x64 resolution when inputting the network.

[0065] 3、Model training parameter setting: The multi-target recognition model proposed in the application is trained using the Adam optimizer with an initial learning rate of 0.001. To balance training speed and final model effect, a dynamic learning rate adjustment strategy is used during training, with the learning rate decaying to 0.2 times the current learning rate every 30 epochs, and a total of 100 epochs.

[0066] 4、Validation results: To comprehensively evaluate the performance of the multi-label classification model proposed in the application, several residual models commonly used in target recognition schemes are selected as comparison models, and two sets of evaluation indicators are used: subset classification indicators and label-based indicators.

[0067] 5、Subset classification indicators: Each multi-label subset in the multi-target data set (including single-target and no-target cases, such as label [0, 0, 1] representing only the third type of target, and [0, 0, 0] representing no target) is regarded as an independent class. This allows the use of standard multi-class classification indicators to evaluate the model's ability to predict the combination of targets in the audio segment. Specifically, global accuracy (Global Accuracy) is used to evaluate the subset classification performance.

[0068] Global accuracy:

[0069] Where TP is the true positive, FP is the false positive, and FN is the false negative.

[0070] 6、Label-based indicators: To more accurately evaluate the model's recognition effect for each specific ship in a multi-target scenario, the precision (Precision), recall (Recall), and F1 score (F1-score) of each target label are calculated.

[0071] Label precision, recall, and F1 score:

[0072]

[0073]

[0074] where is the number of samples that are correctly predicted to belong to class , is the number of samples that are incorrectly predicted to belong to class , is the number of samples that belong to class but are not predicted, To obtain an overall performance measure across all labels, the metrics are averaged over the individual labels. Macro-averaging is used, where the metric is computed for each label independently and then averaged across labels, treating all labels equally:

[0075]

[0076]

[0077] where C is the total number of unique target classes.

[0078] Validation set results on the ShipEar multi-target dataset: 1. Performance metrics based on the base target label: As shown in Figure 4 , the improved ResNet model proposed in the present application achieved a macro-precision of 98.10%, a macro-recall of 97.28%, a macro-F1 score of 97.69%, and a macro-accuracy of 96.70% on the four macro indicators based on a single label, namely macro-averaged precision, macro-averaged recall, macro-averaged F1 score, and macro-averaged accuracy. Compared with the commonly used backbone models ResNet18, ResNet34, ResNet50, and ResNet101 in traditional target recognition, the improved ResNet model of the present application is higher than the aforementioned models in all indicators.

[0079] 2. Classification performance indicators based on target subset combinations: In the subset classification task, each unique target combination (including single target and no target) is considered as an independent class. As shown in Figure 5 , the improved ResNet model proposed in the present application achieved a global accuracy of 92.92% in subset classification.

[0080] 3. Accuracy curve: As shown in Figure 6As shown, the change curve of the model in the ShipEar multi-target data set training process verification set accuracy (with the training algebra), the improved ResNet model proposed in the application is higher than the accuracy of other commonly used models after the training is stable.

[0081] Smoothly induced regularization data enhancement result display: 1、Accuracy curve: Figure 7 The improved ResNet model using and not using smooth induction regularization data enhancement is compared when the accuracy of the verification set changes, when the smooth induction regularization data enhancement is not used, the verification set accuracy jumps sharply, and the improvement is slow with the increase of the training round, and the model is still unstable. The recognition of the target. When using smooth induction regularization data enhancement, the verification set accuracy also fluctuates, but the fluctuation is small, and with the training, the fluctuation amplitude gradually becomes small, and the verification set accuracy tends to be stable.

[0082] 3, inference time: the average inference time of the model is (6.0±0.074) ms, which meets the real-time requirement.

[0083] The above only describes the preferred embodiments of the present application, and is not intended to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A single-channel underwater multi-target acoustic identification method based on an improved residual network, characterized in that, The method comprises: obtaining an original acoustic signal of a target to be identified, and preprocessing the original acoustic signal to obtain a mel time-frequency spectrogram; inputting the mel time-frequency spectrogram into a multi-target identification model trained in advance to predict a probability value in the interval [0, 1] for each possible target category, wherein the multi-target identification model is composed of an improved ResNet network structure; presetting a detection threshold for each possible target category, and if the generated probability value is less than the preset detection threshold, it is determined that the corresponding target category does not exist in the current scene, and if the generated probability value is greater than or equal to the preset detection threshold, it is determined that the corresponding target category exists in the current scene.

2. The single-channel underwater multi-target acoustic identification method based on the improved residual network according to claim 1, wherein, The preprocessing of the original acoustic signal specifically comprises: performing down-sampling processing on the original acoustic signal to obtain a continuous audio stream; segmenting the continuous audio stream into non-overlapping short audio segments with a fixed time length; normalizing each short audio segment, then performing short-time Fourier transform (STFT), and then calculating a mel spectrogram through a mel filter bank.

3. The single-channel underwater multi-target acoustic identification method based on the improved residual network according to claim 2, characterized in that, The short-time Fourier transform of the normalized short audio segment specifically comprises: calculating the short-time Fourier transform to obtain a power spectrum, and the expression is: wherein, is the power spectrum, is the original linear frequency, is the original audio segment, represents the result of the original signal after short-time Fourier transform processing, indicates the center position of the time window, i.e. the time point of the short-time Fourier transform analysis, is the frequency variable in the short-time Fourier transform output result; mapping the linear frequency f to the mel scale through a mel frequency conversion formula, and the expression is: filtering the power spectrum of the audio segment through a mel filter bank composed of a series of triangular filters to obtain a mel spectrogram.

4. The single-channel underwater multi-target acoustic identification method based on the improved residual network of claim 1, wherein, The improvement of the ResNet network structure specifically comprises: modifying the first 7x7 convolution kernel of the ResNet input layer to a 3x3 convolution kernel, and adjusting the convolution step to 1.

5. The single-channel underwater multi-target acoustic identification method based on the improved residual network according to claim 1, characterized in that, The improvement of the ResNet network structure further comprises: the improved ResNet network structure comprises a backbone network, and the backbone network is composed of four convolution stages, and the four convolution stages comprise stage 1, stage 2, stage 3 and stage 4; each convolution stage is sequentially stacked by a plurality of bottleneck residual blocks, and each bottleneck residual block extracts features through a plurality of convolution operations; the first bottleneck residual block of each convolution stage performs down-sampling on its second convolution layer and residual branch.

6. The single-channel underwater multi-target acoustic identification method based on the improved residual network according to claim 5, characterized in that, Stage 2, stage 3 and stage 4 of the convolution stage are used to extract acoustic feature maps of different levels; the acoustic feature maps output by stage 2 and stage 3 are subjected to an adaptive average pooling operation to adjust the size of the acoustic feature maps to be consistent with the acoustic feature maps output by stage 4; the spatially aligned acoustic feature maps of stage 2, stage 3 and stage 4 are spliced in the channel dimension to form a fusion feature map containing multi-scale information.

7. The single-channel underwater multi-target acoustic identification method based on the improved residual network according to claim 6, characterized in that, A channel attention module is applied to the spliced fusion feature map, and then a global adaptive average pooling layer is used to compress the spatial size to 1x1, and output a global feature map.

8. The single-channel underwater multi-target acoustic identification method based on the improved residual network according to claim 7, characterized in that, The global feature map is flattened and converted into a one-dimensional feature vector, and then a fully connected layer with a Sigmoid layer at the back is used to generate a probability value in the interval [0, 1] for each predefined target category.

9. The single-channel underwater multi-target acoustic identification method based on the improved residual network according to claim 1, characterized in that, The training data of the multi-target identification model is enhanced through a smoothness-induced regularization data enhancement strategy.

Images