A method and system for underwater target recognition based on beam perception

Abstract

The present application relates to the technical field of underwater acoustic target recognition, in particular to a kind of underwater acoustic target recognition method and system based on beam perception, first extract acoustic statistical initial feature from beam domain signal, and then propose beam perception feature screening algorithm, the central echo and other beam features are comprehensively analyzed, and the preferred feature subset with cross-beam discrimination ability is screened;Subsequently, the spatial prior information is embedded using the self-defined beam perception position coding, and the beam perception attention mechanism is designed to capture the differences between beams, thereby realizing the effective fusion of multi-beam features;Finally, the recognition result is output by the classifier.The present application significantly reduces the false alarm rate of the system while maintaining the target recognition recall rate, and improves the robustness and generalization ability of the recognition model in complex underwater acoustic environment.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic target recognition technology, and in particular to an underwater acoustic target recognition method and system based on beam sensing. Background Technology

[0002] Underwater acoustic target recognition technology is a key component of the marine science and engineering technology system. This technology aims to analyze underwater acoustic signals in complex underwater acoustic combat environments using advanced signal processing methods, thereby completing tasks such as target identification and parameter estimation.

[0003] In recent years, deep learning technology has been widely applied in the field of underwater acoustic target recognition, forming two mainstream methods: one is the recognition path based on acoustic features and shallow networks. This method uses feature extraction techniques such as empirical mode decomposition, sparse signal decomposition, multivariate variational mode decomposition, spectrum analysis, statistical feature analysis, and wavelet transform, combined with traditional classifiers to achieve target recognition. Its focus is usually on data augmentation or the artificial construction of signal features. The other is the end-to-end recognition framework based on deep networks. This method usually first uses short-time Fourier transform, Mel-frequency cepstral coefficients, and other means to extract low-order features of the signal, and then uses deep structures such as convolutional neural networks, residual networks, and Transformers to automatically learn more discriminative high-level feature representations. Furthermore, it introduces attention mechanisms, reinforcement learning, semi-supervised learning, and other strategies to continuously improve the model's recognition performance.

[0004] For example, existing deep learning-based underwater acoustic target recognition methods include the following:

[0005] (1) An SE_ResNet_17 method for underwater acoustic target recognition in dynamic environments (CN114298090A): This method extracts underwater acoustic signal features through a residual network and combines the SE attention mechanism to weight the feature channels to improve the recognition accuracy. Specifically, it includes framing, windowing, and normalization preprocessing of the time-domain signal, then using a 1×64 convolutional kernel and a residual module for feature extraction, and finally generating channel weights through global average pooling and fully connected layers to achieve adaptive recalibration of the feature channels. This method is based on single-beam signal processing and cannot utilize the spatial distribution differences between multiple beams, resulting in limited ability to distinguish between targets and interference sources (such as seabed topography and marine life) in complex marine environments. At the same time, the SE attention mechanism only considers the channel relationship of features within a single beam and ignores the beam spatial structure, resulting in insufficient feature discriminativeness and a high false alarm rate.

[0006] (2) An underwater acoustic target recognition method based on an adaptive multi-feature fusion model (CN115909040A): This method extracts deep temporal features, spatial features, and frequency domain features from Mel frequency cepstral coefficient features and time-spectrum maps, and then performs adaptive weighted fusion through a channel attention layer; specifically, it uses a long short-term memory network to extract temporal features, a one-dimensional convolutional neural network to extract spatial features, and a two-dimensional convolutional neural network to extract frequency domain features, and optimizes the feature representation through an SE module (including global average pooling, weight generation, and feature weighting), and finally completes target recognition through a fully connected layer. Although this method fuses multi-dimensional features, it lacks beam dimension information, making it difficult to effectively identify targets by utilizing the energy difference between the center beam and adjacent beams; in addition, feature fusion is only performed through simple weighting, without a cross-beam feature screening mechanism for underwater acoustic target characteristics, resulting in performance degradation and insufficient generalization ability in low signal-to-noise ratio environments.

[0007] (3) A method for underwater acoustic target recognition based on a CNN-Transformer cooperative network model (CN118051831A): This method combines the advantages of convolutional neural networks and Transformers. First, a convolutional neural network is used to extract local features. Then, a Swing Transformer is used to capture global dependencies. Specifically, the audio data is converted into a Mel spectrogram. Preliminary feature extraction and downsampling are performed through 7×7 convolutional layers and 3×3 pooling layers. Channel attention processing is performed using the ECA-Res module. Global feature extraction is then performed through the Swing Transformer network. Finally, the recognition result is output through a classification layer. This method has a complex network structure and high computational cost, which affects real-time processing capabilities. Moreover, it does not involve multi-beam cooperative processing and cannot capture the spatial relationship between beams, resulting in insufficient robustness to marine environmental interference. It is difficult to achieve a high recall rate while maintaining a low false alarm rate.

[0008] While existing methods have filled some technological gaps in the field of underwater acoustic target identification and provided feasible solutions for practical applications, they still have significant limitations. Most of these methods focus on the center beam signal of the target echo. Although they possess basic identification capabilities, their over-reliance on a single beam can lead the model to focus only on the apparent differences between targets and non-targets in the center echo, neglecting the discriminative structural features of the target between the center beam and adjacent beams. This limitation not only easily leads to model overfitting but also results in a high false alarm rate in practical applications, restricting the robustness and generalization ability of existing identification methods in complex marine environments. Summary of the Invention

[0009] Therefore, the technical problem to be solved by the present invention is to overcome the problem that the existing technology only focuses on the apparent differences between the target and non-target in the center echo, while ignoring the distinguishable structural features of the target between the center beam and adjacent beams, resulting in insufficient generalization ability of underwater acoustic target recognition and excessively high false alarm rate.

[0010] To address the aforementioned technical problems, this invention provides a method for underwater acoustic target recognition based on beam sensing, comprising:

[0011] Initial features are extracted for each beam in a multi-beam signal of a known category; the category includes target and non-target.

[0012] Based on the intra-class variance, inter-class variance, and inter-beam variance, calculate the comprehensive discriminant index value for each initial feature; based on the comprehensive discriminant index value, select the best features from the initial features;

[0013] Extract the preferred features of each beam in the multi-beam signal to be identified, and construct the acoustic features of the multi-beam signal;

[0014] The acoustic features of the multibeam signal are input into the underwater acoustic target recognition model to obtain a predicted classification, including:

[0015] Position encoding is performed on the embedded acoustic features to obtain encoded features;

[0016] The encoded features are processed through a multi-head self-attention mechanism module and a beam attention mechanism module, respectively, to obtain self-attention features and beam attention features;

[0017] Adaptively fuse the self-attention features and beam attention features to obtain beam fusion features; then weight and aggregate the beam fusion features along the beam dimension to obtain beam aggregation features.

[0018] The beamforming features are input into the classifier to obtain the predicted classification.

[0019] Preferably, initial features are extracted for each beam in a multi-beam signal of a known category. The method is as follows: initial low-order features are extracted for each beam in a multi-beam signal of a known category from the time domain, frequency domain, time-frequency domain and auditory perception dimension; multiple statistical features are calculated for each initial low-order feature using statistical functions to obtain the initial features.

[0020] Preferably, the formula for calculating the comprehensive discriminant index value of each initial feature based on the intra-class variance, inter-class variance, and inter-beam variance is as follows:

[0021] ;

[0022] in, To comprehensively judge the index values, For the inter-beam variance, For inter-class variance, For within-class variance, and These are the weighting coefficients.

[0023] Preferably, the method for calculating the intra-class variance, inter-class variance, and inter-beam variance includes: using a Gaussian weighted distribution to calculate the intra-class variance, inter-class variance, and inter-beam variance.

[0024] Preferably, the embedding of the acoustic features is positionally encoded to obtain encoded features, the steps of which include:

[0025] Define the index of the center beam of the multibeam signal as c, and the neighborhood window of the multibeam signal as [ck, c+k], where k is the neighborhood size parameter;

[0026] The location encoding of the acoustic feature embedding is calculated using the following formula:

[0027] ;

[0028] in, Encode the position of the m-th beam in the d-th dimension. d is the beam index, and d is the dimension index. The total dimension of the acoustic feature embedding. For spatial information weighting coefficients, The Gaussian decay coefficient is... This is the normalized offset of the m-th beam;

[0029] By fusing the embedding of acoustic features and their positional encoding, coded features are obtained.

[0030] Preferably, the multi-head self-attention mechanism module includes three sequentially connected multi-head self-attention layers; the beam attention mechanism module includes three sequentially connected beam attention layers.

[0031] Preferably, the data processing procedure of the beam attention layer includes:

[0032] Calculate queries, keys, and values ​​based on the input features of the beam attention layer;

[0033] Spatial prior information is calculated based on beam offset and Gaussian attenuation coefficient;

[0034] Spatial prior information is introduced into the beam attention mechanism, and initial attention features are calculated based on the query, key, and value.

[0035] After the initial attention features are processed by the Dropout mechanism, they are added to the input features of the beam attention layer using residual connections, and then normalized to obtain the output features of the beam attention layer.

[0036] Preferably, spatial prior information is introduced into the beam attention mechanism. Based on the query, key, and value, initial attention features are calculated using the following formula:

[0037] ;

[0038] in, These are the initial attention features. For beam attention mechanism, For the softmax function, For query, As key, For value, For feature dimension, For spatial prior weights, This refers to prior information about space.

[0039] Preferably, the self-attention features and beam attention features are adaptively fused to obtain the beam fused features, and the calculation formula is as follows:

[0040] ;

[0041] in, For beam fusion features, This is a self-attention feature. For beam attention features, These are learnable weights.

[0042] The present invention also provides a beam-sensing-based underwater acoustic target recognition system, comprising:

[0043] An initial feature extraction module is used to extract initial features for each beam in a multi-beam signal of a known category; the category includes targets and non-targets.

[0044] The beam-sensing feature selection module is used to calculate the comprehensive discrimination index value of each initial feature based on intra-class variance, inter-class variance, and inter-beam variance; and to select the preferred features from the initial features based on the comprehensive discrimination index value.

[0045] The preferred feature extraction module is used to extract the preferred features of each beam in the multi-beam signal to be identified, and to construct the acoustic features of the multi-beam signal.

[0046] The recognition module is used to input the acoustic features of the multibeam signal into the underwater acoustic target recognition model to obtain a predicted classification, including:

[0047] Position encoding is performed on the embedded acoustic features to obtain encoded features;

[0048] The encoded features are processed through a multi-head self-attention mechanism module and a beam attention mechanism module, respectively, to obtain self-attention features and beam attention features;

[0049] Adaptively fuse the self-attention features and beam attention features to obtain beam fusion features; then weight and aggregate the beam fusion features along the beam dimension to obtain beam aggregation features.

[0050] The beamforming features are input into the classifier to obtain the predicted classification.

[0051] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0052] The present invention discloses an underwater acoustic target recognition method based on beam sensing. First, it quantifies the discriminative differences between features of different beams using a three-dimensional feature selection method based on intra-class variance, inter-class variance, and inter-beam variance. This selects preferred features with cross-beam discrimination capability from the high-dimensional initial features, and constructs acoustic features of multi-beam signals based on these preferred features. Second, the invention constructs an underwater acoustic target recognition model. This model utilizes beam sensing position coding technology to model the spatial relationships between beams as learnable position codes, enabling the model to perceive the distribution characteristics of targets in the beam space and enhance its focus on target areas. Furthermore, considering the characteristics of multi-beam signals, the invention further introduces a beam attention mechanism on top of the multi-head self-attention mechanism to model the relationship between the central beam and neighboring beams. Finally, by fusing multi-head self-attention and beam attention, the invention achieves efficient mining and fusion of cross-beam discriminative features, improving the model's utilization efficiency of multi-beam information. This invention systematically introduces the feature differences between the center beam and neighboring beams into underwater acoustic target recognition, breaking through the technical limitations of the traditional single-beam recognition mode, realizing cross-beam feature collaborative recognition, effectively improving the accuracy of target recognition, reducing the false alarm rate, and improving robustness and generalization ability in complex underwater acoustic environments. Attached Figure Description

[0053] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0054] Figure 1 These are the time-domain waveforms of three typical echo signals after beamforming, among which... Figure 1 (a) in the image represents the echo signal from the underwater acoustic target. Figure 1 (b) in the image represents strong interference echo signals caused by seabed topography, passing ships, and large marine life. Figure 1 (c) in the figure represents marine environmental noise;

[0055] Figure 2This is a flowchart of an underwater acoustic target recognition method based on beam sensing according to the present invention;

[0056] Figure 3 This is a schematic diagram of initial feature extraction for multi-beam signals;

[0057] Figure 4 This is a time-frequency comparison analysis diagram of the target echo, in which... Figure 4 (a) in the figure is a time-domain comparison analysis diagram of the center beam of the target echo and its adjacent beams. Figure 4 (b) in the figure is a frequency domain comparison analysis diagram of the center beam of the target echo and its adjacent beams;

[0058] Figure 5 Time-frequency comparison analysis chart of non-target echoes. Figure 5 (a) in the figure is a time-domain comparison analysis diagram of the center beam and its adjacent beams of the non-target echo. Figure 5 (b) in the figure is a frequency domain comparison analysis diagram of the center beam and its adjacent beams of the non-target echo;

[0059] Figure 6 This is a schematic diagram of the beam sensing feature filtering algorithm;

[0060] Figure 7 This is a schematic diagram of the beam sensing feature filtering results, in which... Figure 7 (a) in the figure is a visualization analysis of the 8th coefficient in the Mel frequency cepstral coefficients. Figure 7 (b) in the figure is a visualization analysis of the 90% spectral roll-off point. Figure 7 (c) in the figure is a visualization analysis of the spectral harmonicity. Figure 7 (d) in the figure is a visualization analysis of the 75% spectral roll-off point. Figure 7 (e) in the figure is a visualization analysis chart of the zero-crossing rate. Figure 7 (f) in the figure is a visualization analysis of spectral sharpness;

[0061] Figure 8 This is a schematic diagram of the underwater acoustic target recognition model;

[0062] Figure 9 This is a schematic diagram of the beam attention layer structure;

[0063] Figure 10 This is a structural flowchart of an underwater acoustic target recognition method based on beam sensing according to the present invention;

[0064] Figure 11 This is a graph showing the trend of recognition performance under different numbers of beams;

[0065] Figure 12 This is a graph showing the comparison results of ablation test performance. Detailed Implementation

[0066] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0067] Figure 1 The time-domain waveforms of three typical echo signals after beamforming are shown, among which... Figure 1 (a) in the image represents the echo signal from the underwater acoustic target. Figure 1 (b) in the image represents strong interference echo signals caused by seabed topography, passing ships, and large marine life. Figure 1 In the diagram, (c) represents marine environmental noise, with red indicating the center beam and blue representing multiple adjacent beams within its neighborhood window. It can be seen that both the target and strong interference exhibit concentrated energy and significant amplitude on the center beam, making effective differentiation difficult using only the center beam. However, the energy of the target echo in the center beam is significantly higher than that of adjacent beams, forming a distinct peak; while the energy difference between the center and adjacent beams in the interference echo is smaller, with a relatively flat distribution. Therefore, by utilizing the characteristic differences between the center beam and adjacent beams, effective differentiation between the target and interference can be achieved.

[0068] This invention, starting from the overall characteristics of echo signals, addresses the limitations of existing methods in feature utilization by proposing a beam-sensing-based underwater acoustic target recognition method. This method utilizes feature fusion across multiple beams to deeply explore the inherent correlations and structural differences exhibited by targets in different beams, rather than focusing solely on distinguishing targets from non-targets in a single beam. This strategy can significantly reduce the false alarm rate of the system while maintaining the target recognition recall rate, thereby improving the robustness and generalization ability of the recognition model in complex underwater acoustic environments.

[0069] Reference Figure 2 As shown, the present invention proposes a method for underwater acoustic target recognition based on beam sensing, the steps of which include:

[0070] S1: Extract initial features for each beam in a multi-beam signal of a known category; the category includes targets and non-targets;

[0071] S2: Calculate the comprehensive discriminant index value for each initial feature based on the intra-class variance, inter-class variance, and inter-beam variance; select the preferred features from the initial features based on the comprehensive discriminant index value.

[0072] S3: Extract the preferred features of each beam in the multi-beam signal to be identified, and construct the acoustic features of the multi-beam signal;

[0073] S4: Input the acoustic features of the multibeam signal into the underwater acoustic target recognition model to obtain a predicted classification, including:

[0074] Position encoding is performed on the embedded acoustic features to obtain encoded features;

[0075] The encoded features are processed through a multi-head self-attention mechanism module and a beam attention mechanism module, respectively, to obtain self-attention features and beam attention features;

[0076] The self-attention features and beam attention features are adaptively applied to obtain beam fusion features; the beam fusion features are then weighted and aggregated along the beam dimension to obtain beam aggregation features.

[0077] The beamforming features are input into the classifier to obtain the predicted classification.

[0078] Because underwater acoustic target echo signals typically possess complex characteristics such as low signal-to-noise ratio, high redundancy, and strong non-stationarity, and are also massive in scale, directly using the raw signal as input to the recognition model would result in extremely high computational complexity and storage overhead. Therefore, it is necessary to perform effective feature extraction, filtering, and dimensionality reduction on the echo signal.

[0079] Reference Figure 3 As shown, preferably, in S1, in order to comprehensively characterize the echo signal characteristics, initial features are extracted for each beam in the multi-beam signal of known categories. The method is as follows: initial low-order features are extracted for each beam in the multi-beam signal of known categories from the time domain, frequency domain, time-frequency domain and auditory perception dimensions; multiple statistical features are calculated for each initial low-order feature using statistical functions to obtain the initial features.

[0080] Reference Figure 4 As shown, Figure 4 (a) in the figure is a time-domain comparison analysis diagram of the center beam of the target echo and its adjacent beams. Figure 4 Figure (b) shows a frequency domain comparison analysis of the center beam and its adjacent beams of the target echo. (Refer to...) Figure 5 As shown, Figure 5 (a) in the figure is a time-domain comparison analysis diagram of the center beam and its adjacent beams of the non-target echo. Figure 5 Figure (b) shows a frequency domain comparison analysis of the center beam and its adjacent beams of the non-target echo. It can be seen that there are significant differences between the target and non-target in both the time and frequency domains, between the center beam, between different beams, and between different beams.

[0081] In this embodiment, basic initial low-order features, including Mel-Frequency Cepstral Coefficients (MFCC), zero-crossing rate, spectral sharpness, spectral centroid, and temporal envelope, are extracted from the time domain, frequency domain, time-frequency domain, and auditory perception dimensions. Basic statistical functions such as mean, standard deviation, extreme values, percentiles, and regression slope are calculated for the original sequence and its first- and second-order difference sequences of each initial low-order feature. These are then fused to construct a complete 6373-dimensional hybrid statistical feature set as the initial features. Taking a multi-beam signal as a sample, which includes a center beam and adjacent beams, totaling M beams, the initial features extracted from this multi-beam signal are an M*6373 feature matrix.

[0082] The initial feature set of a multibeam signal of a known category can be represented as:

[0083] ;

[0084] in, For the initial feature set, Let i be the initial feature value of the j-th category in the m-th beam. For beam index, there is , The total number of beams, For the initial feature index, The initial total number of features, For category indexing, Indicates a positive sample (target). This is a negative sample (not the target). In this embodiment... .

[0085] In S2, based on the spatial structure characteristics of multi-beam signals, this invention proposes a Beam-Aware Feature Selection (BAFS) method: Based on intra-class variance, inter-class variance, and inter-beam variance, a comprehensive discriminant index value is calculated for each initial feature; based on the comprehensive discriminant index value, preferred features are selected from the initial features. (Refer to...) Figure 6 As shown, the core idea of ​​the BAFS algorithm is to select a subset of features with cross-beam discrimination capability by quantifying the differences in the discriminative power of features between different beams.

[0086] The BAFS algorithm uses a three-dimensional variance analysis model to perform feature selection. Based on the intra-class variance, inter-class variance, and inter-beam variance, the formula for calculating the comprehensive discriminant index value of each initial feature is as follows:

[0087] ;

[0088] in, To comprehensively judge the index values, For the inter-beam variance, For inter-class variance, For within-class variance, and This is the weighting coefficient. The comprehensive discrimination index value enhances cross-beam feature consistency and inter-class separability through the numerator, while suppressing intra-class fluctuation noise through the denominator. The larger the value, the stronger the ability of the initial feature to jointly discriminate and represent underwater acoustic targets.

[0089] Preferably, the method for calculating the intra-class variance, inter-class variance, and inter-beam variance includes: using a Gaussian weighted distribution to calculate the intra-class variance, inter-class variance, and inter-beam variance.

[0090] Intra-class variance measures the characteristic stability of similar samples within the same beam. It emphasizes the contribution of the center beam through a Gaussian weighted distribution and is defined as follows:

[0091] ;

[0092] ;

[0093] ;

[0094] in, Let be the variance of the j-th class in the m-th beam. Let m be the Gaussian weighting function for the m-th beam. Let be the initial feature mean of the j-th category in the m-th beam. Let be the variance of the j-th category. The variance is the within-class variance.

[0095] Inter-class variance is used to characterize the separability between samples of different classes within the same beam. It also uses a Gaussian weighted distribution and is defined as follows:

[0096] ;

[0097] ;

[0098] in, Let m be the inter-class variance of the m-th beam. Let be the initial mean of the features of the target in the m-th beam. Let be the initial mean of the non-target features in the m-th beam. The variance is between classes.

[0099] Inter-beam variance measures the difference in response of different beams to the same type of target. It is distributed using a Gaussian weighted distribution and is defined as follows:

[0100] ;

[0101] ;

[0102] in, Let be the inter-beam variance of the j-th category. Let be the global feature mean of the j-th category across all beams. The inter-beam variance of the target. For non-target inter-beam variance, This represents the inter-beam variance.

[0103] Gaussian weighting function of the m-th beam Defined as:

[0104] ;

[0105] Where c is the index of the middle beam, and in this embodiment the value is 3; The value is the Gaussian attenuation coefficient, which is set to 0.5 in this embodiment.

[0106] Since the mixed feature set may contain a large number of features that are weakly related to the characteristics of the underwater acoustic target, this embodiment uses the BAFS method to optimize and screen the initial features. Finally, the 10 initial features with the highest scores are selected as the most preferred features, including: 90% spectral roll-off point, 75% spectral roll-off point, zero-crossing rate, the 8th coefficient in the Mel frequency cepstral coefficient, the 13th coefficient in the Mel frequency cepstral coefficient, spectral sharpness, spectral entropy, spectral slope, root mean square energy, and spectral harmonicity.

[0107] Preferably, in S3, the preferred features of each beam in the multi-beam signal to be identified are extracted to construct the acoustic features of the multi-beam signal. The method is as follows: after extracting the preferred features of each beam in the multi-beam signal to be identified, statistical functions are used to perform statistical analysis on the preferred features to construct the acoustic features of the multi-beam signal.

[0108] In this embodiment, eight statistical functions—arithmetic mean, mean, standard deviation, skewness, kurtosis, first quartile, third quartile, and 99th percentile—are selected to perform statistical analysis on the 10 preferred features mentioned above. This results in the formation of an 80-dimensional acoustic feature of the multibeam signal, which is used to effectively characterize the underwater acoustic target characteristics.

[0109] To visually demonstrate the differences in feature distribution between targets and non-targets, between targets in different beams, and between non-targets in different beams, this embodiment uses BAFS to select six low-order features with high scores for visualization analysis. The results are referenced... Figure 7 As shown, where Figure 7 (a) in the figure is a visualization analysis of the 8th coefficient in the Mel frequency cepstral coefficients. Figure 7 (b) in the figure is a visualization analysis of the 90% spectral roll-off point. Figure 7 (c) in the figure is a visualization analysis of the spectral harmonicity. Figure 7 (d) in the figure is a visualization analysis of the 75% spectral roll-off point. Figure 7 (e) in the figure is a visualization analysis chart of the zero-crossing rate. Figure 7 Figure (f) shows a visualization of spectral sharpness. Red represents negative samples (non-target), and blue represents positive samples (target). It can be seen that positive and negative samples exhibit significant differences not only at the center echo location but also at other beams. Furthermore, within the samples, there are clear differences in the feature distribution between different beams: positive samples show more pronounced feature differences between beams, with significant feature values ​​at the center beam, gradually flattening outwards; in contrast, negative samples exhibit smaller feature fluctuations between beams and a more uniform overall distribution.

[0110] Compared to single-beam features, processing multi-beam features requires considering both target attribute information within the beam and the differences in feature distribution between beams. To fully exploit the effective discriminative information contained in multi-beam features, this invention proposes an underwater acoustic target recognition model that integrates beam-aware position coding and beam-aware attention mechanisms. The overall structure of the model is as follows: Figure 8 As shown.

[0111] After the acoustic features of the multi-beam signal are input into the underwater acoustic target recognition model, they first pass through a feature embedding layer to obtain the embedded acoustic features. Beam-Aware Positional Encoding (BAPE) is then used to positionally encode the embedded acoustic features, embedding spatial prior information to obtain coded features. Next, the coded features are input into a dual-path Beam-Aware Attention Mechanism (BAAM), which includes a parallel multi-head self-attention mechanism module and a beam attention mechanism module, obtaining self-attention features and beam attention features respectively. Then, the self-attention features and beam attention features are adaptively fused to obtain beam fusion features. The beam fusion features are then weighted and aggregated along the beam dimension to obtain beam aggregate features. Finally, the beam aggregate features are input into a classifier to obtain a predicted classification.

[0112] In this embodiment, the feature embedding layer is used to embed a feature of size 1. The acoustic features are mapped as High-dimensional representations of size are used to improve the expressive power of features and the flexibility of subsequent processing. This represents the total number of dimensions of the acoustic features; in this embodiment, the value is 80. This represents the total number of dimensions for embedding acoustic features; in this embodiment, the value is 256.

[0113] The embedding of the acoustic features is positionally encoded using beam-sensing position coding to obtain coded features. The steps include:

[0114] Define the index of the center beam of the multibeam signal as c, and the neighborhood window of the multibeam signal as [ck, c+k], where k is the neighborhood size parameter;

[0115] Considering both absolute position and relative beam relationship, the positional encoding of the acoustic feature embedding is calculated using the following formula:

[0116] ;

[0117] in, Encode the position of the m-th beam in the d-th dimension. d is the beam index, and d is the dimension index. The total dimension of the acoustic feature embedding. For spatial information weighting coefficients, The Gaussian decay coefficient is... This is the normalized offset of the m-th beam;

[0118] By fusing the embedding of acoustic features and their positional encoding, coded features are obtained.

[0119] To enable the recognition model to better perceive the spatial structure between beams, this invention will and Set as learnable parameters to obtain optimal beam weight allocation.

[0120] The beam-aware position coding designed in this invention can focus on modeling the spatial relationships between the central beam and its neighboring beams.

[0121] Building upon traditional multi-head attention mechanisms, this invention proposes a dual-attention-based BAAM mechanism specifically tailored to the characteristics of multi-beam data. This attention mechanism comprises two paths: multi-head attention and beam attention. Beam attention is used to model the relationship between the center beam and neighboring beams.

[0122] Preferably, the multi-head self-attention mechanism module includes three sequentially connected multi-head self-attention layers; the beam attention mechanism module includes three sequentially connected beam attention layers.

[0123] Reference Figure 9 As shown, the data processing procedure of the beam attention layer includes:

[0124] S41: Calculate queries, keys, and values ​​based on the input features of the beam attention layer.

[0125] S42: Calculate spatial prior information based on beam offset and Gaussian attenuation coefficient, using the following formula:

[0126] ;

[0127] in, For spatial prior information, (mc) represents the beam offset.

[0128] S43: Introduce spatial prior information into the beam attention mechanism. Based on the query, key, and value, calculate the initial attention features using the following formula:

[0129] ;

[0130] in, These are the initial attention features. For beam attention mechanism, For the softmax function, For query, As key, For value, For feature dimension, For spatial prior weights.

[0131] S44: After passing the initial attention features through the Dropout mechanism, the residual connections are added to the input features of the beam attention layer, and then normalized to obtain the output features of the beam attention layer. The formula is as follows:

[0132] ;

[0133] in, The output characteristics of the beam attention layer, For the input features of the beam attention layer, It uses the Dropout mechanism. This is a layer normalization operation.

[0134] Specifically, the self-attention features and beam attention features are adaptively fused to obtain the beam fused features, and the calculation formula is as follows:

[0135] ;

[0136] in, For beam fusion features, This is a self-attention feature. For beam attention features, These are learnable weights.

[0137] Overall, the Transformer encoding layer of the multi-head self-attention mechanism module can effectively capture the complex dependencies between beams, while the beam attention mechanism module can focus on the aggregation of multi-beam information. The two work together to significantly improve the model's target recognition ability in complex underwater acoustic environments. In particular, while maintaining attention to the core echo region, it can effectively utilize the auxiliary information of multiple beams.

[0138] Specifically, the beam fusion features are weighted and aggregated along the beam dimension to obtain the beam aggregation features, calculated using the following formula:

[0139] ;

[0140] in, Let m be the feature vector of the m-th beam in the beam fusion feature. This is a beamforming feature.

[0141] The complete structural flowchart of this invention is shown below. Figure 10 As shown.

[0142] To systematically verify the effectiveness of the underwater acoustic target recognition method based on beam sensing proposed in this invention, this embodiment conducts a comprehensive performance evaluation and analysis from four aspects: feature extraction method, feature selection strategy, limited number of beams, and module ablation experiment.

[0143] The experimental data were obtained from underwater acoustic target echo data collected in the sea area. Each detection cycle lasted 40–70 seconds, and the target type was fixed. The transmitted signal used an HFM waveform with a pulse width of 2 seconds, covering the 1375–1875 Hz and 1200–1700 Hz frequency bands. The dataset contained 496 valid echo samples with a balanced distribution of positive and negative samples. To comprehensively evaluate the model performance, this embodiment used accuracy, recall, F1 score, and G-mean as the main evaluation metrics.

[0144] (1) Performance comparison of different feature sets

[0145] To verify the applicability of the mixed statistical features of multi-beam signals extracted by the present invention as initial features in underwater acoustic target recognition, this embodiment compares its performance with four commonly used acoustic feature sets under single-beam conditions, and the results are shown in Table 1.

[0146] Table 1. Recognition performance (%) of different feature sets

[0147]

[0148] Experimental results show that the hybrid feature performs best across all evaluation metrics, achieving an accuracy of 84.65%, an F1 score of 84.34%, and a mean G of 84.65%. Compared to MFCC, PLP, STFT, and WT, the hybrid feature improves accuracy by 0.72, 1.61, 4.31, and 5.37 percentage points, respectively, and improves F1 scores by 0.97, 1.81, 4.66, and 5.78 percentage points, respectively. This demonstrates that the rich acoustic descriptors and statistical functionals contained in the hybrid feature can more effectively characterize the complex properties of underwater acoustic targets.

[0149] (2) Comparison of Feature Filtering Strategies

[0150] To further verify the effectiveness of the BAFS feature selection algorithm, this embodiment compares and analyzes it with the classic Fisher discriminant method. The Fisher method first selects the top 80 discriminative features based on the center beam features to form a Single-Fisher feature set, which is then expanded to adjacent beams to obtain a Multi-Fisher feature set. Based on this, the BAFS algorithm is applied to select the multi-beam features, obtaining the BAFS feature set. All methods are trained using the standard Transformer multi-head attention mechanism, and the specific results are shown in Table 2.

[0151] Table 2 Performance comparison of different feature selection methods (%)

[0152]

[0153] The results show that the BAFS algorithm outperforms the comparison methods on all evaluation metrics. Compared with unfiltered mixed features, BAFS improves accuracy by 7.08 percentage points and by 3.61 percentage points compared with the Single-Fisher method. Notably, Multi-Fisher only improves accuracy by 0.76 percentage points compared to Single-Fisher, indicating that BAFS effectively filters features with cross-beam consistency and strong inter-class discriminative power, significantly enhancing the model's ability to identify negative samples and its overall robustness.

[0154] (3) Selection of the optimal number of beams

[0155] To further explore the impact of the number of beams on the performance of the BAFS algorithm, this embodiment also systematically compares different configurations from single beam to nine beams, and the results are as follows: Figure 11 As shown.

[0156] Analysis shows that under single-beam conditions, due to the lack of inter-beam variance information, the BAFS algorithm degenerates into an equivalent Fisher method with a recognition rate of 87.08%. As the number of beams increases, the model performance initially increases and then decreases, reaching its optimal performance (92.10% accuracy) with a seven-beam configuration. This indicates that the key feature information of underwater acoustic targets is mainly contained in the center echo and its six adjacent echoes. When the number of beams is insufficient, incomplete feature information limits model performance; while when the number of beams is excessive, background noise introduced by edge beams leads to data contamination, causing a decline in recognition performance.

[0157] (4) Module ablation experiment analysis

[0158] To gain a deeper understanding of the respective mechanisms of beam-sensing position coding and beam attention layer, this embodiment designed six ablation experiments and systematically compared the combined effects of different position coding and attention mechanisms. The experimental configurations are shown in Table 3.

[0159] Table 3 Experimental Configuration

[0160]

[0161] All experiments used multi-beam features selected by the BAFS algorithm as input, maintaining consistent settings for 8-head and 3-layer attention network structures. Specific performance comparisons are as follows: Figure 12 As shown.

[0162] Analysis of the results from Experiments 1-3 reveals that beam-aware location coding, by incorporating prior information about the beam spatial structure, effectively guides the model to focus on the target region, significantly improving the recall rate. However, due to the continued use of the standard attention mechanism, the model's ability to suppress interference from non-target regions is limited, resulting in a false alarm rate of 7.26% in Experiment 3 (compared to 9.68% in Experiments 1 and 10.89% in Experiments 2). This indicates that while beam-aware location coding can effectively improve recall, its performance gain is constrained by the structure of the attention mechanism.

[0163] Comparing experiments 4-6 reveals that the overall model performance is significantly improved after introducing the beam-aware attention mechanism. This mechanism effectively captures cross-beam discriminative features, enhancing the model's target discrimination ability. Specifically, it effectively reduces false alarms while maintaining recall. Experiment 6, combining beam-aware position encoding and the beam-aware attention mechanism, achieved optimal performance across all evaluation metrics: an F1 score of 96.69% and a mean G-score of 96.71%, indicating that the model achieved an optimal balance between precision and recall. Simultaneously, the 97.58% specificity metric demonstrates excellent false alarm control, effectively reducing the probability of false alarms while maintaining a high recognition rate.

[0164] In summary, the beam-aware underwater acoustic target recognition method of this invention first quantifies the differences in discriminative power between features of different beams based on a three-dimensional feature selection method using intra-class variance, inter-class variance, and inter-beam variance. It then selects preferred features with cross-beam discriminative capabilities from high-dimensional initial features and constructs acoustic features of multi-beam signals based on these preferred features. Secondly, this invention constructs an underwater acoustic target recognition model that utilizes beam-aware position coding technology to model the spatial relationships between beams as learnable position codes. This enables the model to perceive the distribution characteristics of targets in the beam space, enhancing its focus on target areas. Furthermore, considering the characteristics of multi-beam signals, this invention further introduces a beam attention mechanism on top of the multi-head self-attention mechanism to model the relationship between the central beam and neighboring beams. Finally, by fusing multi-head self-attention and beam attention, it achieves efficient mining and fusion of cross-beam discriminative features, improving the model's utilization efficiency of multi-beam information. This invention systematically introduces the feature differences between the center beam and neighboring beams into underwater acoustic target recognition. Through feature fusion across multiple beams, it delves into the intrinsic correlations and structural differences exhibited by targets in different beams, rather than merely focusing on distinguishing targets from non-targets on a single beam. This overcomes the technical limitations of traditional single-beam recognition modes, achieving cross-beam feature collaborative recognition. This strategy can significantly reduce the false alarm rate of the system while maintaining the target recognition recall rate, effectively improving the accuracy of target recognition and enhancing the robustness and generalization ability of the recognition model in complex underwater acoustic environments.

[0165] Based on the above-mentioned beam-sensing-based underwater acoustic target recognition method, the present invention also provides a beam-sensing-based underwater acoustic target recognition system, comprising:

[0166] An initial feature extraction module is used to extract initial features for each beam in a multi-beam signal of a known category; the category includes targets and non-targets.

[0167] The beam-sensing feature selection module is used to calculate the comprehensive discrimination index value of each initial feature based on intra-class variance, inter-class variance, and inter-beam variance; and to select the preferred features from the initial features based on the comprehensive discrimination index value.

[0168] The preferred feature extraction module is used to extract the preferred features of each beam in the multi-beam signal to be identified, and to construct the acoustic features of the multi-beam signal.

[0169] The recognition module is used to input the acoustic features of the multibeam signal into the underwater acoustic target recognition model to obtain a predicted classification, including:

[0170] Position encoding is performed on the embedded acoustic features to obtain encoded features;

[0171] The encoded features are processed through a multi-head self-attention mechanism module and a beam attention mechanism module, respectively, to obtain self-attention features and beam attention features;

[0172] Adaptively fuse the self-attention features and beam attention features to obtain beam fusion features; then weight and aggregate the beam fusion features along the beam dimension to obtain beam aggregation features.

[0173] The beamforming features are input into the classifier to obtain the predicted classification.

[0174] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0175] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0176] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0177] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0178] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for underwater acoustic target recognition based on beam sensing, characterized in that, include: Initial features are extracted for each beam in a multi-beam signal of a known category; The categories include targets and non-targets; Based on intra-class variance, inter-class variance, and inter-beam variance, calculate the comprehensive discriminant index value for each initial feature; Based on the comprehensive discrimination index value, the preferred features are selected from the initial features; Extract the preferred features of each beam in the multi-beam signal to be identified, and construct the acoustic features of the multi-beam signal; The acoustic features of the multibeam signal are input into the underwater acoustic target recognition model to obtain a predicted classification, including: Position encoding is performed on the embedding of the acoustic features to obtain encoded features; wherein, the step of obtaining the encoded features includes: Define the index of the center beam of the multibeam signal as c, and the neighborhood window of the multibeam signal as [ck, c+k], where k is the neighborhood size parameter; The location encoding of the acoustic feature embedding is calculated using the following formula: ； in, Encode the position of the m-th beam in the d-th dimension. d is the beam index, and d is the dimension index. The total dimension of the acoustic feature embedding. For spatial information weighting coefficients, The Gaussian decay coefficient is... This is the normalized offset of the m-th beam; By fusing the embedding of acoustic features and their positional encoding, coded features are obtained; The encoded features are processed through a multi-head self-attention mechanism module and a beam attention mechanism module, respectively, to obtain self-attention features and beam attention features; Adaptively fuse the self-attention features and beam attention features to obtain beam fusion features; then weight and aggregate the beam fusion features along the beam dimension to obtain beam aggregation features. The beamforming features are input into the classifier to obtain the predicted classification.

2. The underwater acoustic target recognition method based on beam sensing according to claim 1, characterized in that, For each beam in a multi-beam signal of a known category, initial features are extracted separately. The method is as follows: initial low-order features are extracted for each beam in the multi-beam signal of a known category from the time domain, frequency domain, time-frequency domain and auditory perception dimension; multiple statistical features are calculated for each initial low-order feature using statistical functions to obtain the initial features.

3. The underwater acoustic target recognition method based on beam sensing according to claim 1, characterized in that, Based on the intra-class variance, inter-class variance, and inter-beam variance, the formula for calculating the comprehensive discriminant index value of each initial feature is as follows: ； in, To comprehensively judge the index values, For the inter-beam variance, For inter-class variance, For within-class variance, and These are the weighting coefficients.

4. The underwater acoustic target recognition method based on beam sensing according to claim 3, characterized in that, Methods for calculating intra-class variance, inter-class variance, and inter-beam variance include using Gaussian weighted distribution to calculate intra-class variance, inter-class variance, and inter-beam variance.

5. The underwater acoustic target recognition method based on beam sensing according to claim 1, characterized in that, The multi-head self-attention mechanism module includes three sequentially connected multi-head self-attention layers; the beam attention mechanism module includes three sequentially connected beam attention layers.

6. The underwater acoustic target recognition method based on beam sensing according to claim 5, characterized in that, The data processing procedure for the beam attention layer includes: Calculate queries, keys, and values ​​based on the input features of the beam attention layer; Spatial prior information is calculated based on beam offset and Gaussian attenuation coefficient; Spatial prior information is introduced into the beam attention mechanism, and initial attention features are calculated based on the query, key, and value. After the initial attention features are processed by the Dropout mechanism, they are added to the input features of the beam attention layer using residual connections, and then normalized to obtain the output features of the beam attention layer.

7. The underwater acoustic target recognition method based on beam sensing according to claim 6, characterized in that, In the beam attention mechanism, spatial prior information is introduced. Based on the query, key, and value, initial attention features are calculated using the following formula: ； in, These are the initial attention features. For beam attention mechanism, For the softmax function, For query, As key, For value, For feature dimension, For spatial prior weights, This refers to prior information about space.

8. The underwater acoustic target recognition method based on beam sensing according to claim 1, characterized in that, Adaptive fusion of self-attention features and beam attention features yields beam fused features, calculated using the following formula: ； in, For beam fusion features, This is a self-attention feature. For beam attention features, These are learnable weights.

9. A beam-sensing-based underwater acoustic target recognition system, used to implement the beam-sensing-based underwater acoustic target recognition method according to any one of claims 1 to 8, characterized in that, include: The initial feature extraction module is used to extract initial features for each beam in a multi-beam signal of known category; The categories include targets and non-targets; The beam-sensing feature filtering module is used to calculate the comprehensive discrimination index value of each initial feature based on intra-class variance, inter-class variance, and inter-beam variance. Based on the comprehensive discrimination index value, the preferred features are selected from the initial features; The preferred feature extraction module is used to extract the preferred features of each beam in the multi-beam signal to be identified, and to construct the acoustic features of the multi-beam signal. The recognition module is used to input the acoustic features of the multibeam signal into the underwater acoustic target recognition model to obtain a predicted classification, including: Position encoding is performed on the embedded acoustic features to obtain encoded features; The encoded features are processed through a multi-head self-attention mechanism module and a beam attention mechanism module, respectively, to obtain self-attention features and beam attention features; Adaptively fuse the self-attention features and beam attention features to obtain beam fusion features; then weight and aggregate the beam fusion features along the beam dimension to obtain beam aggregation features. The beamforming features are input into the classifier to obtain the predicted classification.

Images