CN-122020382-A - Domain prompt self-adaption-based underwater sound signal detection method and system

CN122020382ACN 122020382 ACN122020382 ACN 122020382ACN-122020382-A

Abstract

The invention belongs to the field of underwater acoustic signal processing and deep learning, and discloses a domain prompt self-adaptive underwater acoustic signal detection method and system, wherein the method comprises the steps of collecting source domain underwater acoustic signals and preprocessing, and dividing the preprocessed source domain underwater acoustic signals into a plurality of related tasks according to working condition dimensions; the method comprises the steps of obtaining initialization parameters of a main network through meta learning pre-training based on a plurality of divided related tasks, carrying out domain prompt self-adaption and multi-granularity alignment training based on the initialization parameters of the main network, optimizing domain prompt vectors, splicing the optimized domain prompt vectors with target domain underwater sound signal characteristics to be detected, and then sending the spliced domain prompt vectors into the main network to finish underwater sound target classification.

Inventors

LAN BO
Xin Zhengang
SHI YANDA
ZHU YADONGYANG

Assignees

北京石油化工学院

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (8)

1. A domain hint-based adaptive underwater acoustic signal detection method, the method comprising: collecting a source domain underwater sound signal and preprocessing the source domain underwater sound signal, and dividing the preprocessed source domain underwater sound signal into a plurality of related tasks according to working condition dimensions; Based on a plurality of divided related tasks, obtaining initialization parameters of a backbone network through meta learning pre-training; Based on the initialization parameters of the backbone network, performing domain prompt self-adaption and multi-granularity alignment training, and optimizing domain prompt vectors; And (3) splicing the optimized domain prompt vector with the target domain underwater sound signal characteristics to be detected, and then sending the spliced domain prompt vector into a backbone network to finish underwater sound target classification.
2. The method of claim 1, wherein the method of dividing the preprocessed source domain acoustic signal into a plurality of related tasks in the operating mode dimension comprises: the source domain data is divided into tasks according to sea domain, season, sea state and signal to noise ratio dimensions, each task corresponds to a working condition combination, and each task comprises 16 samples of a support set and 64 samples of a query set.
3. The method of claim 2, wherein the pre-training by meta-learning to obtain the initialization parameters of the backbone network based on the partitioned plurality of related tasks comprises: for each sampling task First, an inner layer update is performed on a support set: ; Wherein, the For the learning rate of the inner layer, For the task Is a set of support for a (c) support, The meta-parameters of the backbone network are the targets of meta-learning optimization; For the task A loss function on; To be a pair of parameters Calculating an operator of the gradient; updating task adaptation parameters after one step on the support set; the updated model is then evaluated on the query set and an outer layer update is performed: ; Wherein, the For the learning rate of the outer layer, For the task Is a query set of (a).
4. The method of claim 1, wherein the domain hint adaptation comprises hint vector design, hint injection manner, and optimization strategy; wherein for hint vector design, a set of learnable hint vectors is defined for each target domain Wherein In order to suggest the number of vectors, For the number of characteristic channels, the prompt vector is randomly initialized by adopting normal distribution, the standard deviation is 0.02, the prompt vector does not carry any domain specific information during initialization, and the domain adaptation capability is obtained completely through training; for the prompt injection mode, the prompt vector and the input feature are spliced along the sequence dimension and then sent to the encoder: ; Wherein the method comprises the steps of The input characteristics are represented as such, Representing the output characteristics of the encoder, namely, the output characteristics obtained by processing the encoder after the prompt vector is spliced with the input characteristics, for subsequent classification detection, The hint vectors act as a learnable "domain context," interacting with the input features in a self-attention mechanism, directing the encoder to focus on key feature patterns of the target domain, the encoder's self-attention computation being: ; wherein queries Key and key Matrix of values The information of the prompt vector and the input feature are contained, and the prompt vector influences the final feature representation through attention weight distribution; For the optimization strategy, only the parameters of the prompt vector are updated in the training process, the parameters of the backbone network are kept fixed, and the optimization goal of the prompt vector is to minimize the classification loss and the alignment loss on the target domain: Wherein, the A label sample set is provided for the target domain; 、 respectively a source domain data set and a target domain data set; classifying the loss for cross entropy; loss for multi-granularity alignment; in order to freeze the parameters of the backbone network, The loss weights are aligned.
5. The method of claim 4, wherein the multi-granularity alignment training comprises performing signal level, feature level, semantic level three-layer alignment on the source domain data and the target domain data; The signal level alignment is performed in the original signal space, and the maximum mean difference MMD measurement distribution difference is adopted: ; Wherein, the 、 The number of samples of the source domain and the target domain respectively; 、 the underwater sound signal samples in the source domain and the target domain are respectively; is a Gaussian kernel mapping function; Feature level alignment performs hierarchical alignment in depth feature space, and calculates MMD on shallow, middle and deep features, respectively: ; Wherein, the 、 Sample of source domain and target domain respectively through backbone network Layer extracted features; is a hierarchical weight; Semantic level alignment is performed in a category condition distribution space, and condition distribution differences are calculated according to categories respectively: ; Wherein, the Is a hint vector for the source domain, Is a hint vector for the source domain, Is the total number of categories; is a category index; Is a category label; is a depth feature representation; is a category weight.
6. The method of claim 1, wherein the method for completing the classification of the underwater sound target comprises the steps of: The training optimized prompt vector is spliced with the characteristic of the underwater acoustic signal of the target domain to be detected, the encoder is guided to pay attention to the key characteristic mode of the target domain through a self-attention mechanism, the characteristic representation with cross-domain invariance is extracted, and finally the classification detection of the underwater acoustic signal of the target domain is realized.
7. A domain hint based adaptive underwater acoustic signal detection system for implementing the method of any of claims 1-6, wherein the system comprises a data preparation and task partitioning module, an initialization module, an optimization module, and a classification detection module; The data preparation and task division module is used for collecting source domain underwater sound signals and preprocessing the source domain underwater sound signals, and dividing the preprocessed source domain underwater sound signals into a plurality of related tasks according to working condition dimensions; The initialization module is used for obtaining initialization parameters of the backbone network through meta-learning pre-training based on a plurality of divided related tasks; the optimization module is used for performing domain prompt self-adaption and multi-granularity alignment training based on the initialization parameters of the backbone network, and optimizing domain prompt vectors; The classification detection module is used for splicing the optimized domain prompt vector with the target domain underwater sound signal characteristics to be detected and then sending the spliced domain prompt vector into a backbone network to finish underwater sound target classification.
8. The system of claim 7, wherein dividing the preprocessed source domain acoustic signal into a plurality of related tasks in a working dimension comprises: the source domain data is divided into tasks according to sea domain, season, sea state and signal to noise ratio dimensions, each task corresponds to a working condition combination, and each task comprises 16 samples of a support set and 64 samples of a query set.

Description

Domain prompt self-adaption-based underwater sound signal detection method and system Technical Field The invention belongs to the field of underwater acoustic signal processing and deep learning, and particularly relates to a field prompt self-adaptive underwater acoustic signal detection method and system. Background Underwater acoustic signal detection is a key technology for marine monitoring, underwater target identification and marine resource exploration. Due to the fact that ocean environments are complex and changeable, water temperatures, salinity and submarine topography of different sea areas are remarkably different, sea conditions and plankton distributions in different seasons are different, and the difference of hydrophone equipment models and installation positions is added, performance is greatly reduced when a model trained in a source domain (training environment with data) is moved to a target domain (new environment to be deployed), and the problem is called cross-domain distribution difference. The existing cross-domain detection method mainly focuses on feature space alignment, ignores distribution differences of a signal level and a semantic level, is insufficient in alignment, and meanwhile lacks of quick adaptability to a new target domain, and needs a large amount of retraining. More importantly, the traditional fine tuning mode needs to update the whole network parameters, has high calculation cost, and is easy to cause disastrous forgetting when adapting to a new domain, namely, the model loses the general representation capability of the learned source domain when learning the knowledge of the new domain, so that the model can not keep good performance on a plurality of domains at the same time. The invention aims to provide a domain prompt self-adaption-based underwater acoustic signal detection method and system, which overcome the defects and realize high-efficiency and accurate cross-domain underwater acoustic signal detection through the synergistic effect of multi-granularity layering domain alignment, meta-learning pre-training and domain prompt self-adaption. Disclosure of Invention Aiming at the requirements of multi-level distribution difference alignment and rapid adaptation in the cross-domain detection of the underwater acoustic signals, the invention provides a domain prompt self-adaptation-based underwater acoustic signal detection method and a domain prompt self-adaptation-based underwater acoustic signal detection system. In order to achieve the above object, the present invention provides the following solutions: a domain hint-based adaptive underwater acoustic signal detection method, the method comprising: collecting a source domain underwater sound signal and preprocessing the source domain underwater sound signal, and dividing the preprocessed source domain underwater sound signal into a plurality of related tasks according to working condition dimensions; Based on a plurality of divided related tasks, obtaining initialization parameters of a backbone network through meta learning pre-training; Based on the initialization parameters of the backbone network, performing domain prompt self-adaption and multi-granularity alignment training, and optimizing domain prompt vectors; And (3) splicing the optimized domain prompt vector with the target domain underwater sound signal characteristics to be detected, and then sending the spliced domain prompt vector into a backbone network to finish underwater sound target classification. Preferably, the method for dividing the preprocessed source domain underwater sound signal into a plurality of related tasks according to working condition dimensions comprises the following steps: the source domain data is divided into tasks according to sea domain, season, sea state and signal to noise ratio dimensions, each task corresponds to a working condition combination, and each task comprises 16 samples of a support set and 64 samples of a query set. Preferably, the method for obtaining the initialization parameters of the backbone network through meta-learning pre-training based on a plurality of divided related tasks comprises the following steps: for each sampling task First, an inner layer update is performed on a support set: ; Wherein, the For the learning rate of the inner layer,For the taskIs a set of support for a (c) support,The meta-parameters of the backbone network are the targets of meta-learning optimization; For the task A loss function on; To be a pair of parameters Calculating an operator of the gradient; updating task adaptation parameters after one step on the support set; the updated model is then evaluated on the query set and an outer layer update is performed: ; Wherein, the For the learning rate of the outer layer,For the taskIs a query set of (a). Preferably, the domain hint self-adaption comprises hint vector design, hint injection mode and optimization strategy; wherein for hint vector design, a