Passive underwater acoustic target identification method based on semi-supervised learning
Through the semi-supervised learning method, the combination of labeled and labelless data sets is used to perform data amplification and pseudo-label extraction, and the model performance is optimized, which solves the problem of low accuracy of cross-sea water acoustic target recognition and achieves higher recognition Accuracy and robustness.
Patent Information
- Application Number
- CN202411006279.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-19
AI Technical Summary
In the field of water acoustic target recognition, it is difficult for the prior art to effectively utilize labeled and labelless data sets between different sea areas, resulting in a degradation in the performance of models in cross-sea applications, especially in the absence of sufficient labeled data. The identification accuracy rate is determined in the absence of sufficient labeled data. Low.
Using semi-supervised learning method, supervised pre-training of the labeled source domain data set, combined with the labeled target domain data set for data amplification and fine-tuning training, design the fusion loss function, extract the pseudo-labels and cluster it. Finally, through the unsupervised distillation optimization model, the identification of label-free water acoustic targets across sea areas is achieved.
It significantly improves the accuracy and robustness of water acoustic target recognition, effectively utilizes a small amount of label data and a large amount of labelless data, optimizes the model performance, and is suitable for labelless water acoustic target recognition in cross-sea environments.
Smart Images

Figure CN118981638A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of sound wave recognition, and in particular relates to a passive underwater acoustic target recognition method based on semi-supervised learning. Background Art
[0002] As the main carrier of underwater information collection, sound waves are used for information detection, positioning, observation, estimation, communication and other purposes of target objects through perception, identification and communication. Therefore, underwater acoustic target recognition technology is one of the most challenging research contents in today's marine information processing technology, and plays an extremely important role in marine scientific research, environmental investigation, rights protection and other fields.
[0003] For example, passive hydroacoustic target detection technology can be widely used in underwater defense, resource surveying, ship identification, and fish activity detection. As a breakthrough in realizing the intelligence of hydroacoustic equipment, it plays an important role in marine defense and economic construction. In addition, in the military field, the location detection of specific targets is conducive to tactical deployment and improves the early warning capability of coastal defense; in the civilian field, the detection of certain illegal ships is helpful for the monitoring and management of maritime transport.
[0004] When applying the underwater acoustic target recognition classifier in practice, it is often difficult to obtain labels. The labels of the data set obtained in a certain sea area are complete, but if the model is trained directly on this data set, the performance will drop rapidly when it is applied to the underwater acoustic target data set obtained in another sea area; even more difficult is that if the data set in another sea area has no labels at all, the obtained model cannot learn the data characteristics of this sea area at all. Obviously, there is an obvious domain mismatch problem between the two data sets, and the solution to this problem is almost blank in the field of underwater acoustic target recognition. How to make full use of the labeled data set in a certain sea area and the unlabeled data set in another sea area to obtain a more practical classifier model can effectively solve the model performance in such application contexts. Summary of the invention
[0005] The technical purpose of the present invention is to provide a semi-supervised learning passive underwater acoustic target recognition method to solve the problem of low recognition accuracy caused by lack of sufficient labeled data.
[0006] To solve the above problems, the technical solution of the present invention is:
[0007] A semi-supervised learning passive underwater acoustic target recognition method comprises the following steps:
[0008] Acquire a number of labeled known underwater acoustic target data as a source domain data set, and acquire a number of unlabeled unknown target data as a target domain data set;
[0009] Perform supervised pre-training on the source domain dataset to obtain a first training network;
[0010] The target domain data set is augmented in waveform domain and time-frequency domain, and fine-tuned in combination with the source domain data set, and a fusion loss function is designed to obtain a second training network; the fusion loss includes classification loss calculation for the source domain data set and contrast loss calculation for the target domain data set;
[0011] The target domain dataset is extracted using the second training network to obtain pseudo labels, and the contrastive center loss is introduced to fine-tune the second training network; the embedded codes of the target domain dataset are extracted using the fine-tuned second training network, and re-clustered to obtain pseudo-label data of the target domain with label noise;
[0012] Perform supervised classification training on the source domain dataset and the target domain pseudo-label data to obtain a third training network;
[0013] The model performance of the third training network is optimized through unsupervised distillation to obtain the fourth training model, which is used to realize the recognition of unlabeled underwater acoustic targets across sea areas.
[0014] The supervised pre-training of the source domain dataset is specifically as follows: using the labeled source domain dataset, utilizing standard classification loss, optimizing network parameters, and obtaining a first training network.
[0015] The training of the second training network specifically includes: extracting basic two-dimensional time-frequency acoustic features from the source domain data set;
[0016] The target domain dataset is augmented in waveform domain and time-frequency domain as the input of contrast loss;
[0017] The source domain dataset and the augmented target domain dataset are jointly supervised and trained; during training, the classification loss of the source domain dataset is combined with the contrast loss of the target domain dataset to fine-tune the first training network, thereby obtaining a second training network.
[0018] Among them, the target domain pseudo-label data with label noise is obtained as follows:
[0019] Using the second training network to extract the embedding codes of the target domain dataset and perform clustering;
[0020] The contrastive center loss is introduced to fine-tune the second training network. During the fine-tuning process, clustering updates are performed every preset training interval until convergence.
[0021] The fine-tuned second training network is used to extract the embedding codes of all audios in the target domain dataset and re-cluster them to obtain target domain pseudo-label data with label noise.
[0022] Among them, the supervised classification training of the source domain dataset and the target domain pseudo-label data is specifically as follows: the source domain dataset and the target domain pseudo-label data are fused for supervised classification training, and the total number is the sum of the number of target types contained in the source domain dataset and the number of target types contained in the unknown target data; during the fusion process, the proportion of erroneous labels in the target domain pseudo-label data is diluted to guide the correctness of the extraction network, and then the third training network is obtained.
[0023] Specifically, the model performance of the third training network is optimized through unsupervised distillation to obtain the fourth training model: the third training network is used as the teacher model, unlabeled data is input for training to obtain a new randomly initialized student model, and the student model is used to fit the output of the teacher model to improve the generalization ability and performance of the model, thereby obtaining the fourth training network; wherein, the distillation loss function in unsupervised distillation includes knowledge distillation loss and mutual learning loss.
[0024] An electronic device includes a processor and a memory, the memory stores computer instructions that can be run on the processor, and the processor is used to call the computer instructions in the memory to execute any of the above-mentioned semi-supervised learning passive underwater acoustic target recognition methods.
[0025] A storage medium is used to store computer instructions, wherein when the computer instructions are executed, the device where the storage medium is located is controlled to execute any of the semi-supervised learning passive underwater acoustic target recognition methods described above.
[0026] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:
[0027] By effectively utilizing a small amount of labeled data and a large amount of unlabeled data, combining the two data, and gradually optimizing the model, unlabeled underwater acoustic target recognition in a cross-sea environment can be achieved. This method can significantly improve the accuracy and robustness of underwater acoustic target recognition, and provides an effective solution for applications in the field of underwater acoustic signal processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.
[0029] Figure 1 The present invention is a flowchart of a semi-supervised learning method for passive underwater acoustic target recognition. DETAILED DESCRIPTION
[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and other implementation methods can be obtained based on these drawings without creative work.
[0031] In order to simplify the drawings, only the parts related to the present invention are schematically shown in each figure, and they do not represent the actual structure of the product. In addition, in order to simplify the drawings and facilitate understanding, in some figures, only one of the parts with the same structure or function is schematically drawn or marked. In this article, "one" not only means "only one", but also means "more than one".
[0032] The following is a further detailed description of a semi-supervised learning passive underwater acoustic target recognition method proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become more apparent from the following description and claims.
[0033] Example
[0034] See also Figure 1 This embodiment provides a passive underwater acoustic target recognition method based on semi-supervised learning. Different from the previous supervised classification model that limits the target type, this embodiment is a domain adaptation method that uses a known identified target in a certain sea area as a source domain dataset and another sea area without labels as a target domain dataset for joint training. This embodiment specifically includes the following steps:
[0035] First, relevant data is obtained, and labeled known hydroacoustic target data in any sea area is obtained as the source domain data set A. The source domain data set A contains M types of known identified targets, and M can be 200. Then, unlabeled unknown target data in another sea area is obtained as the target domain data set B. The target domain data set B contains N types of unknown identified targets. The specific number of N is unknown. In this embodiment, it is assumed that N=50.
[0036] Then, supervised pre-training is performed on the source domain dataset A to extract basic acoustic features and obtain the first training network. The specific steps are as follows: the source domain dataset A is input into the network and supervised classification training is performed to extract two-dimensional time-frequency acoustic features from each audio clip in the source domain dataset A, thereby obtaining the first training network f1, i.e., the pre-training network. The network input by the source domain dataset A is trained using a DenseNet or ResNet network structure and optimized using the AAMSoftmax loss function.
[0037] Next, data augmentation and joint training are performed on the target domain dataset B, specifically: data augmentation in the waveform domain and time-frequency domain is performed on the target domain dataset B. Waveform domain augmentation is to add noise, time scaling and translation to the target domain dataset B to generate multiple augmented data views; time-frequency domain augmentation is to perform frequency distortion, time masking and frequency masking on the target domain dataset B to generate multiple augmented data views. Then, fine-tuning training is performed in combination with the source domain dataset A, and a fusion loss function is designed. The fusion loss includes classification loss calculation for the source domain dataset A and contrast loss calculation for the target domain dataset B. The classification loss calculation ensures the classification ability of the model on known targets, and the contrast loss calculation enhances the representation ability of the model on unlabeled data. The source domain dataset A and the augmented target domain dataset B are jointly supervised and trained, that is, the joint input model is trained, and fine-tuning training is performed to obtain the second training network f2, that is, the joint training network. During the training process, the classification loss of the source domain dataset A is combined with the contrast loss of the target domain dataset B, so as to fine-tune the first training network f1.
[0038] Next, the second training network f2 is used to extract the features of the target domain data set B, that is, the second training network f2 is used to extract the embedded code of each audio clip in the target domain data set B, and then the K-means or DBSCAN clustering algorithm is used to cluster the extracted features. Including initializing the cluster center, that is, after extracting the embedded code, the cluster center is randomly initialized. These initial cluster centers are evenly distributed in the feature space, and then used as the starting point for iterative clustering. Then, the second training network f2 is fine-tuned by comparing the center loss so that the model can better cluster the data into their respective categories. During the fine-tuning process, the cluster center is updated once every certain training interval s, and the position of the cluster center is gradually optimized in an iterative manner. Through continuous fine-tuning, the fine-tuned second training network f2' is finally obtained. The update rule of the cluster center usually adopts K-means or its variant algorithm. Specifically, for each data point, its distance from the current cluster center is calculated, and it is divided into the category to which the nearest cluster center belongs. Then, the position of the cluster center is recalculated based on the data points in each category to minimize the distance from the data points in that category to the cluster center. The clustering update is iterated until the convergence condition is met. Usually, the iteration is stopped when the update change of the cluster center is less than a certain threshold, or the pre-set maximum number of iterations is reached. After the clustering iteration is completed, the target domain pseudo-label data B' with label noise is preliminarily generated based on the clustering results. Through the above iterative deep clustering learning method, the representation knowledge of unknown targets can be effectively extracted, and the cluster center is continuously optimized during the fine-tuning process, so as to obtain the target domain pseudo-label data B' with less label noise, providing more accurate pseudo-label information for subsequent supervised classification training.
[0039] Then, the source domain data set A and the target domain pseudo-label data B' are fused, and the total number of targets is M+N (250 in this example), so as to reduce the influence of the wrong labels in the pseudo-label data, and supervised classification training is performed after fusion. During the training process, different weights are assigned to the source domain data set A and the target domain pseudo-label data B'. The sample weight of the source domain data set A is large, while the sample weight of the target domain pseudo-label data B' is greatly reduced. In this way, when calculating the loss, the influence of the source domain data is greater, and the influence of the wrong labels of the target domain pseudo-label data is diluted. According to the results of each round of training, the sample weight of the target domain pseudo-label data B' is dynamically adjusted, so that as the model is improved, the credibility of the pseudo-label increases, and the weight is gradually increased, and finally the third training network f3, that is, the underwater acoustic target feature extraction network, is obtained. In a cross-sea environment, the above network can realize binary discrimination of the target without any underwater acoustic target labels in the target sea area.
[0040] Preferably, this implementation can further optimize the model performance of the third training network f3 through unsupervised distillation: the third training network f3 is used as a teacher model, and a large amount of unlabeled data is used to train a new randomly initialized student model. The distillation loss function includes knowledge distillation loss and mutual learning loss. At this stage, only unlabeled data is used, and the generalization ability and performance of the model are improved by fitting the output of the teacher model to the student model, thereby obtaining the fourth training network f4, i.e., the hydroacoustic target feature extraction network. Using this hydroacoustic target feature extraction network, cross-sea unlabeled hydroacoustic target recognition can be achieved.
[0041] Preferably, the present embodiment also provides an electronic device, which includes a processor and a memory, wherein the memory stores computer instructions that can be executed on the processor, and the processor is used to call the computer instructions in the memory to execute the above-mentioned semi-supervised learning passive underwater acoustic target recognition method.
[0042] Further preferably, the present embodiment also provides a storage medium, the storage medium is used to store computer instructions, wherein when the computer instructions are executed, the device where the storage medium is located is controlled to execute the passive underwater acoustic target recognition method based on semi-supervised learning as described above.
[0043] The embodiments of the present invention are described in detail above in conjunction with 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 still fall within the protection scope of the present invention.
Claims
1. A semi-supervised learning method for passive underwater acoustic target recognition, characterized in that: The steps include: Acquire a number of known underwater acoustic target data with labels as source domain data sets, and acquire a number of unknown target data without labels as target domain data sets; Performing supervised pre-training on the source domain dataset to obtain a first training network; Performing data amplification in waveform domain and time-frequency domain on the target domain data set, and performing fine-tuning training in combination with the source domain data set, designing a fusion loss function, and thereby obtaining a second training network; The fusion loss includes using classification loss calculation for the source domain dataset and using contrast loss calculation for the target domain dataset; Using the second training network to extract the target domain data set to obtain pseudo labels, and introducing contrastive center loss to fine-tune the second training network; Extracting the embedded codes of the target domain data set by using the fine-tuned second training network, and re-clustering the data to obtain target domain pseudo-label data with label noise; Performing supervised classification training on the source domain dataset and the target domain pseudo-label data to obtain a third training network; The model performance of the third training network is optimized through unsupervised distillation to obtain a fourth training model, and the fourth training model is used to realize the recognition of unlabeled hydroacoustic targets across sea areas.
2. The passive underwater acoustic target recognition method based on semi-supervised learning according to claim 1 is characterized in that: The supervised pre-training of the source domain dataset is specifically as follows: using the labeled source domain dataset, utilizing standard classification loss, optimizing network parameters, and obtaining the first training network.
3. The passive underwater acoustic target recognition method of semi-supervised learning according to claim 1 is characterized in that: Training the second training network specifically includes: extracting basic two-dimensional time-frequency acoustic features from the source domain data set; Performing data amplification in waveform domain and time-frequency domain on the target domain data set as input of contrast loss; Performing joint supervised training on the source domain dataset and the expanded target domain dataset; During training, the classification loss of the source domain data set is combined with the contrast loss of the target domain data set to fine-tune the first training network, thereby obtaining the second training network.
4. The passive underwater acoustic target recognition method of semi-supervised learning according to claim 1 is characterized in that: The target domain pseudo-label data with label noise is obtained as follows: Extracting the embedded codes of the target domain data set using the second training network and performing clustering; Introducing contrastive center loss to fine-tune the second training network, and performing clustering updates at preset training intervals during the fine-tuning process until convergence; The embedded codes of all audios in the target domain dataset are extracted using the fine-tuned second training network, and re-clustered to obtain the target domain pseudo-label data with label noise.
5. The passive underwater acoustic target recognition method of semi-supervised learning according to claim 1 is characterized in that: The supervised classification training of the source domain dataset and the target domain pseudo-label data is specifically as follows: the source domain dataset and the target domain pseudo-label data are fused for supervised classification training, wherein the total number of the target types included in the source domain dataset and the number of target types included in the unknown target data are combined; during the fusion process, the proportion of erroneous labels in the target domain pseudo-label data is diluted to guide the correctness of the extraction network, thereby obtaining the third training network.
6. The passive underwater acoustic target recognition method of semi-supervised learning according to claim 1 is characterized in that: The method of optimizing the model performance of the third training network through unsupervised distillation to obtain the fourth training model is specifically as follows: using the third training network as a teacher model, inputting unlabeled data for training to obtain a new randomly initialized student model, using the student model to fit the output of the teacher model, improving the generalization ability and performance of the model, and then obtaining the fourth training network; wherein the distillation loss function in the unsupervised distillation includes knowledge distillation loss and mutual learning loss.
7. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein the memory stores computer instructions that can be executed on the processor, and the processor is used to call the computer instructions in the memory to execute the semi-supervised learning passive underwater acoustic target recognition method according to any one of claims 1-6.
8. A storage medium, characterized in that: The storage medium is used to store computer instructions, wherein when the computer instructions are executed, the device where the storage medium is located is controlled to execute the semi-supervised learning passive underwater acoustic target recognition method according to any one of claims 1 to 6.
Citation Information
Patent Citations
Multi-stage unsupervised domain adaptive causal relationship identification method
CN114090770A
Systems and Methods for Implementing Smart Assistant Systems
US20230245654A1
Cited By
Multi-mode underwater acoustic target identification method and system, and storage medium
CN121142518A
