An acoustic recognition method for unknown underwater surface

Abstract

This invention discloses an acoustic recognition method for unknown underwater surfaces. The method involves acquiring labeled known targets as a source domain dataset; acquiring unlabeled known targets and unknown targets as a target domain dataset; supervising the training of an embedding code extraction model on the source domain dataset; and calculating the average embedding code on the source domain dataset to establish a source domain embedding code template library. The embedding codes of known and unknown targets in the target domain dataset are calculated and compared with the source domain embedding code template library. If the embedding codes are less than a preset threshold, they are considered to belong to the same category as the source domain dataset. Otherwise, they are treated as unknown targets. The model parameters are retrained based on multiple loss to form pseudo-labels. The embedding code extraction model is then updated until its performance no longer improves, thereby classifying and recognizing unknown targets.

Description

Technical Field

[0001] This invention belongs to the field of acoustic wave recognition, and in particular relates to an acoustic recognition method for unknown underwater surfaces. Background Technology

[0002] As the primary carrier of underwater information acquisition, sound waves are used for information detection, positioning, observation, estimation, and communication of target objects through sensing, identification, and communication. Therefore, underwater acoustic target recognition technology is one of the most challenging research areas in marine information processing technology today, playing an extremely important role in marine scientific research, environmental surveys, and rights protection.

[0003] For example, underwater acoustic target detection technology can be widely used in underwater defense, resource exploration, ship identification, and fish activity detection. As a breakthrough in realizing the intelligentization of underwater acoustic equipment, it plays a vital role in marine defense and economic development. Furthermore, in the military field, detecting the location of specific targets facilitates tactical deployment and improves maritime early warning capabilities; in the civilian field, detecting certain illegal vessels contributes to the monitoring and management of maritime transport.

[0004] Early underwater acoustic target identification relied on human experience, requiring extensive training in various marine environments to develop the ability to recognize the timbre characteristics of different target signals. However, underwater acoustic signals face various interference problems, some of which are thousands of times stronger than the target signal. Furthermore, with the improvement of sonar systems and the advancement of ship stealth technology, the ability of human experience to make judgments under high-intensity operations has declined, and its disadvantages have gradually become apparent.

[0005] Of course, underwater acoustic recognition technology has also undergone a long period of development, including two working modes: active and passive, targeting the reflected echo and radiated noise of the target, respectively. The passive working mode has advantages such as high concealment and long detection range, and is therefore an important detection method for various underwater acoustic targets, with high research value.

[0006] Traditional acoustic signal-based recognition systems, employing manual feature extraction techniques and relatively simple classifiers, often fail in complex environments, exhibiting rapid performance degradation. In the past five years, the rapid development of artificial intelligence has led to the emergence of deep learning-based underwater acoustic target recognition methods. However, current mainstream deep learning algorithms utilize supervised classification models with a limited number of categories, forcing unknown targets into a class of known targets, or failing to further subdivide unknown targets into subclasses. In practice, they are ineffective at handling unknown types within open sets, exhibiting a cross-functional performance limitation. The application requirements of underwater acoustic target recognition necessitate not only determining whether a target is of a known or unknown type, but also, if unknown, the ability to further subdivide it into a specific unknown subtype—a capability currently severely lacking. Summary of the Invention

[0007] The technical objective of this invention is to provide an acoustic identification method for unknown underwater surfaces, in order to solve the problem that it is difficult to identify unknown types through underwater acoustic target identification in the prior art.

[0008] To solve the above problems, the technical solution of the present invention is as follows:

[0009] An acoustic identification method for unknown underwater surface includes the following steps:

[0010] Obtain labeled known targets as the source domain dataset;

[0011] Obtain known and unknown targets without labels as the target domain dataset;

[0012] The source domain dataset is trained under supervision to obtain the embedding code extraction model, and the average embedding code of the known recognition targets in the source domain dataset is calculated to establish the source domain embedding code template library.

[0013] Calculate the embedding codes of known and unknown targets in the target domain dataset and compare them with the source domain embedding code template library. If the embedding codes are less than a preset threshold, they are considered to be of the same category as the source domain dataset; otherwise, they are treated as unknown target datasets. The self-supervised model obtained is retrained based on multiple loss and pseudo-labels are generated.

[0014] The embedding code extraction model is updated based on the labels of known targets and the pseudo-labels of unknown targets in the source and target datasets until the performance of the embedding code extraction model no longer improves.

[0015] The specific steps for establishing the source domain embedded code template library are as follows:

[0016] The known target in the source domain dataset is segmented into segments of equal length L; the known and unknown target in the target domain dataset are segmented into segments of equal length 2L; the basic acoustic features of the above segments are extracted.

[0017] Supervised training is performed on the basic acoustic features corresponding to the segments obtained from the source domain dataset to obtain the embedding code extraction model. The average embedding code of the known recognition target in the source domain dataset is calculated using the embedding code extraction model, and a source domain embedding code template library is established.

[0018] Among them, the basic acoustic features are one or more, including gamma-pass frequency cepstral coefficients, Mel cepstral coefficients, Mel spectrum and differential Mel spectrum, logarithmic Mel spectrum, fbank features, and features obtained through a self-designed auditory filter bank.

[0019] Specifically, matching the target domain dataset is as follows:

[0020] The source domain embedding code template library is matched with the basic acoustic features corresponding to the speech segments obtained from the target domain dataset to obtain the embedding code of the corresponding feature. The embedding code is then compared with the embedding code in the source domain embedding code template library. If the embedding code is less than the preset threshold, it is considered a known recognition target; otherwise, it is considered an unknown recognition target.

[0021] If the target is known, it is judged based on similarity, the confidence level is determined, and a closed set pseudo-label is affixed;

[0022] If the target is unknown, then self-supervised training is performed on the unknown target and corresponding pseudo-labels are obtained.

[0023] Specifically, the steps for obtaining the corresponding pseudo-labels for unknown targets are as follows:

[0024] Using the embedded code extraction model as the initial model, the loss function is a combination of contrastive loss and clustering loss. A self-supervised model is trained based on the unknown target and the corresponding pseudo-labels are obtained.

[0025] The clustering loss is determined based on the optimal value of one or more clustering performance metrics, including clustering purity, mutual information index, CH index, and silhouette coefficient.

[0026] The specific steps for updating the embedding code extraction model until its performance no longer improves are as follows:

[0027] The known recognition targets with labels, the known recognition targets with closed-set pseudo-labels, and the unknown recognition targets with pseudo-labels are merged to obtain a new source domain dataset, and supervised training is performed to obtain a new embedding code extraction model.

[0028] Calculate the average embedding code in the new source domain dataset and establish a template library of new source domain embedding codes;

[0029] The target domain dataset is matched and compared with the new source domain embedding code template library. If the value is less than the preset threshold, it is considered a known target and belongs to the same category as the new source domain dataset. Otherwise, it is considered an unknown target dataset. The model parameters obtained by retraining based on contrast loss, center contrast loss and clustering loss are self-supervised to form pseudo-labels, and the new embedding code extraction model is updated again.

[0030] Ideally, if the performance of the embedding code extraction model no longer improves during iterative updates, training should be stopped, and the latest embedding code extraction model should be used as the final embedding code extraction model.

[0031] The present invention, by adopting the above technical solution, has the following advantages and positive effects compared with the prior art: The present invention provides an acoustic recognition method for unknown underwater surfaces. It uses known recognition targets with labels as the source domain dataset to construct an embedding code extraction model and a source domain embedding code template library. Then, it uses known recognition targets of the same type as the source domain dataset but without labels and unknown recognition targets of unknown types as the target domain dataset. Then, it performs self-supervised and supervised training iterations to update the embedding code extraction model, and finally realizes the recognition of all unknown recognition targets. Attached Figure Description

[0032] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0033] Figure 1 This is a flowchart of an acoustic recognition method for an unknown underwater surface according to an embodiment of the present invention;

[0034] Figure 2 This is an embodiment of an acoustic identification method for an unknown underwater surface according to an example of the present invention. Detailed Implementation

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0036] To keep the drawings concise, only the parts relevant to the invention are shown schematically in each figure, and they do not represent the actual structure of the product. Furthermore, for ease of understanding, in some figures, only one of components with the same structure or function is shown schematically, or only one is labeled. In this document, "one" can mean not only "only one" but also "more than one".

[0037] The acoustic identification method for unknown underwater surfaces proposed in this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of this invention will become clearer from the following description and claims.

[0038] Example

[0039] See Figure 1 and Figure 2This embodiment provides an acoustic recognition method for unknown underwater surfaces. Unlike previous supervised classification models that limit target types, this method uses an iterative self-supervised learning algorithm that jointly trains data with known targets as the source domain dataset and data that may contain unknown targets as the target domain dataset. The method includes the following steps:

[0040] First, in step S1, labeled known targets are obtained as the source domain dataset. In this embodiment, data containing M = 200 known targets is used as the source domain dataset, where M is the number of categories. Unlabeled known targets and unknown targets are then combined to form the target domain dataset. In this embodiment, data containing N1 = 56 unknown targets and N2 = 50 known targets are used as the target domain dataset, where N1 and N2 are both numbers.

[0041] Then, in step S2, the known targets in the source domain dataset are segmented into segments of equal length L, in this embodiment L = 5S. The known and unknown targets in the target domain dataset are segmented into segments of equal length 2L, in this embodiment 2L = 10S.

[0042] Then, basic acoustic features are extracted from all the above-mentioned segments. The basic acoustic features can be one or more, including gamma-pass frequency cepstral coefficients (GFCC), mel-frequency cepstral coefficients (MFCC), mel spectrum and differential mel spectrum, logarithmic mel spectrum, fbank features, and features obtained through a self-defined auditory filter bank. In this embodiment, fbank features are used as an example.

[0043] The source domain dataset is segmented into segments, and their corresponding basic acoustic features are used for supervised training to obtain the embedding code extraction model H1. The network structure used here is DenseNet or ResNet, and the loss function is AAMSOFTMAX. The embedding code extraction model H1 is used to calculate the average embedding code of all known target categories in the source domain dataset, with a total of M categories, and a source domain embedding code template library of M=200 targets is established.

[0044] Next, in step S3, the source domain embedding code template library is matched with the basic acoustic features corresponding to the segments obtained from the target domain dataset to obtain the embedding codes for the corresponding features. Specifically, the data in the target domain dataset, all of which are 2L in length, are divided into two segments and identified as the same target. Other segments in the same batch are identified as different targets. The extracted embedding codes are then used as the intra-class and inter-class parts of the contrast loss, respectively. These are then compared with the embedding codes in the source domain embedding code template library, i.e., a distance comparison, to determine whether it is a known or unknown type. If the distance is less than a preset threshold, the sample to be compared is considered to have the same category as the known recognition target in the source domain dataset, i.e., one of the 200 categories in the source domain data, and is therefore a known recognition target. Otherwise, the sample to be compared is identified as an unknown target dataset, i.e., an unknown recognition target.

[0045] If the target is known, it is judged based on similarity, the confidence level is determined, and a closed-set pseudo-label is assigned. If the target is unknown, self-supervised training is performed on the unknown target. After algorithm iteration, the self-supervised model is retrained based on contrastive loss and clustering loss, and pseudo-labels are formed.

[0046] In self-supervised training, the embedding code extraction model H1 is used as the initial model, and the loss function is the contrast loss and the clustering loss. The self-supervised model H2 is trained based on the unknown identification target, and the pseudo-labels corresponding to the unknown identification target are obtained.

[0047] The termination condition for self-supervised training is that one or more of the clustering performance metrics reach their optimal values. The clustering performance metrics include clustering purity, normalized mutual information (NMI), Calinski-Harabaz index, and silhouette coefficient. In this embodiment, clustering purity is used as the stopping criterion.

[0048] Finally, in step S4, the known identified targets with closed-set pseudo-labels and the unknown identified targets with pseudo-labels in the source domain dataset and the target domain dataset are merged to form a new source domain dataset. Supervised training is then performed again to update the embedding code extraction model H1, resulting in a new embedding code extraction model H1. The average embedding code of all categories in the new source domain dataset is recalculated to establish a new source domain embedding code template library. The embedding codes of all samples in the original target domain are recalculated and matched with the new source domain embedding code template library. Based on the calculated distance, the target is determined to be either known or unknown. If it is the latter, the model parameters obtained from retraining are self-supervised based on factors including but not limited to contrastive loss, center contrastive loss, and clustering loss, thereby updating the new embedding code extraction model and pseudo-labels until the performance of the embedding code extraction model no longer improves. If the performance of the embedding code extraction model no longer improves during iterative updates, training is stopped, and the latest embedding code extraction model is taken as the final embedding code extraction model.

[0049] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments. Even if various changes are made to the present invention, if these changes fall within the scope of the claims of the present invention and their equivalents, they shall still fall within the protection scope of the present invention.

Claims

1. An acoustic identification method for unknown underwater surface, characterized in that, Includes the following steps: Obtain labeled known targets as the source domain dataset; Obtain known and unknown targets without labels as the target domain dataset; The source domain dataset is trained under supervision to obtain an embedding code extraction model, and the average embedding code of the known recognition targets in the source domain dataset is calculated to establish a source domain embedding code template library. Calculate the embedding codes of known and unknown identification targets in the target domain dataset, and compare them with the source domain embedding code template library. If the distance is less than a preset threshold, it is determined that there is the same category as the source domain dataset. Otherwise, the dataset is treated as an unknown target dataset, and the self-supervised model is retrained based on multiple losses to form pseudo-labels; The embedding code extraction model is updated based on the closed set pseudo-labels of known targets and the pseudo-labels of unknown targets in the source domain dataset and the target domain dataset, until the performance of the embedding code extraction model no longer improves.

2. The acoustic identification method for unknown underwater surface according to claim 1, characterized in that, The specific steps for establishing the source domain embedded code template library are as follows: The known recognition targets in the source domain dataset are segmented into segments of equal length L; the known and unknown recognition targets in the target domain dataset are segmented into segments of equal length 2L; and the basic acoustic features of the above segments are extracted. The basic acoustic features corresponding to the segments obtained by segmenting the source domain dataset are subjected to supervised training to obtain the embedding code extraction model. The embedding code extraction model is used to calculate the average embedding code of the known recognition targets in the source domain dataset, and the source domain embedding code template library is established.

3. The acoustic identification method for unknown underwater surface according to claim 2, characterized in that, The basic acoustic features are one or more, including gamma-pass frequency cepstral coefficients, Mel cepstral coefficients, Mel spectrum and differential Mel spectrum, logarithmic Mel spectrum, fbank features, and features obtained through a self-defined auditory filter bank.

4. The acoustic identification method for unknown underwater surface according to claim 2, characterized in that, The specific steps for matching the target domain dataset are as follows: The source domain embedding code template library is matched with the basic acoustic features corresponding to the segments obtained based on the target domain dataset to obtain the embedding code of the corresponding feature. The distance is compared with the embedding code in the source domain embedding code template library. If the distance is less than a preset threshold, it is identified as a known recognition target; otherwise, it is an unknown recognition target. If the target is known, it is judged based on similarity, the confidence level is determined, and a closed set pseudo-label is affixed; If the target is unknown, then self-supervised training is performed on the unknown target and corresponding pseudo-labels are obtained.

5. The acoustic identification method for unknown underwater surface according to claim 4, characterized in that, The step of obtaining the corresponding pseudo-label for the unknown target is as follows: Using the embedded code extraction model as the initial model, the loss function is a combination of contrastive loss and clustering loss. The self-supervised model is trained based on the unknown identification target, and the pseudo-labels corresponding to the unknown identification target are obtained.

6. The acoustic identification method for unknown underwater surface according to claim 5, characterized in that, The clustering loss is determined based on the optimal value of one or more clustering performance metrics, including cluster purity, mutual information index, CH index, and silhouette coefficient.

7. The acoustic identification method for unknown underwater surface according to claim 4, characterized in that, The specific steps for updating the embedding code extraction model until its performance no longer improves are as follows: The known recognition targets with labels, the known recognition targets with closed-set pseudo-labels, and the unknown recognition targets with pseudo-labels are merged to obtain a new source domain dataset, and supervised training is performed to obtain a new embedding code extraction model. Calculate the average embedding code in the new source domain dataset and establish a new source domain embedding code template library; The embedding codes of the target domain dataset are matched with the average embedding codes of the new source domain embedding code template library and the distance is compared. If the distance is less than a preset threshold, it is identified as a known target and has the same category as the new source domain dataset. Otherwise, the model is treated as an unknown target dataset, and self-supervised training is conducted based on the model parameters obtained by retraining, including contrastive loss, center contrastive loss, and clustering loss, to form pseudo-labels, and the new embedding code extraction model is updated again.

8. The acoustic identification method for unknown underwater surface according to claim 7, characterized in that, Furthermore, if the performance of the embedded code extraction model no longer improves during iterative updates, training is stopped, and the latest embedded code extraction model is taken as the final embedded code extraction model.

Images