Underwater acoustic signal detection method and system based on domain prompt self-adaption

Abstract

The invention belongs to the field of underwater acoustic signal processing and deep learning, and discloses an underwater acoustic signal detection method and system based on domain prompt self-adaption, and the method comprises the steps: collecting and preprocessing a source domain underwater acoustic signal, and dividing the preprocessed source domain underwater acoustic signal into a plurality of related tasks according to the working condition dimension; based on the plurality of divided related tasks, obtaining initialization parameters of the backbone network through meta-learning pre-training; carrying out domain prompt self-adaption and multi-granularity alignment training based on the initialization parameters of the backbone network, and optimizing a domain prompt vector; and splicing the optimized domain prompt vector and the to-be-detected target domain underwater acoustic signal features, and sending the spliced features to a backbone network to complete underwater acoustic target classification.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic signal processing and deep learning, and specifically relates to an underwater acoustic signal detection method and system based on domain cueing adaptation. Background Technology

[0002] Underwater acoustic signal detection is a key technology for marine monitoring, underwater target identification, and marine resource exploration. Due to the complex and variable marine environment, there are significant differences in water temperature, salinity, and seabed topography in different sea areas. Sea conditions and plankton distribution vary in different seasons. In addition, the differences in hydrophone equipment models and installation locations lead to a significant performance drop when models trained in the source domain (the training environment with existing labeled data) are transferred to the target domain (the new environment to be deployed). This problem is called cross-domain distribution difference.

[0003] Existing cross-domain detection methods primarily focus on feature space alignment, neglecting the distributional differences at the signal and semantic levels, resulting in insufficient alignment. Furthermore, they lack rapid adaptation to new target domains, requiring extensive retraining. More importantly, traditional fine-tuning methods necessitate updating the entire network parameters, incurring high computational costs and prone to catastrophic forgetting when adapting to new domains—that is, the model loses its learned source domain general representation capabilities while learning new domain knowledge, preventing it from maintaining good performance across multiple domains simultaneously. The purpose of this invention is to propose a domain-cue adaptive underwater acoustic signal detection method and system. Through the synergistic effect of multi-granularity hierarchical domain alignment, meta-learning pre-training, and domain-cue adaptation, this method overcomes the aforementioned shortcomings, achieving efficient and accurate cross-domain underwater acoustic signal detection. Summary of the Invention

[0004] To address the need for alignment and rapid adaptation of multi-level distribution differences in cross-domain underwater acoustic signal detection, this invention provides a method and system for underwater acoustic signal detection based on domain cueing adaptation. This method achieves efficient and accurate cross-domain underwater acoustic signal detection by constructing a processing framework of "data preparation and task partitioning → meta-learning pre-training → domain cueing adaptation → multi-granularity domain alignment → target detection output".

[0005] To achieve the above objectives, the present invention provides the following solution: A domain-cue adaptive underwater acoustic signal detection method, the method comprising: Acquire underwater acoustic signals from the source region and preprocess them. Divide the preprocessed underwater acoustic signals from the source region into multiple related tasks according to the working condition dimension. Based on multiple related tasks, the initial parameters of the backbone network are obtained through meta-learning pre-training. Based on the initialization parameters of the backbone network, domain cue adaptation and multi-granularity alignment training are performed to optimize the domain cue vector; The optimized domain cue vector is concatenated with the underwater acoustic signal features of the target domain to be detected and then fed into the backbone network to complete the underwater acoustic target classification.

[0006] Preferably, the method for dividing the preprocessed source domain underwater acoustic signal into multiple related tasks according to the operating condition dimension includes: The source domain data is divided into tasks according to the dimensions of sea area, season, sea state, and signal-to-noise ratio. Each task corresponds to a combination of working conditions, and each task contains 16 samples in the support set and 64 samples in the query set.

[0007] Preferred methods for obtaining the initial parameters of the backbone network through meta-learning pre-training based on multiple related tasks include: For each sampling task First, perform an inner update on the support set: ; in, The inner learning rate, For the task Support set, The meta-parameters of the backbone network are the target of meta-learning optimization. For the task loss function on; For parameters Operators for finding the gradient; To update the task adaptation parameters one step on the support set; Then, evaluate the updated model on the query set and perform an outer update: ; in, The outer learning rate, For the task The query set.

[0008] Preferably, adaptive domain hints include hint vector design, hint injection method, and optimization strategy; Specifically, for cue vector design, a set of learnable cue vectors is defined for each target domain. ,in To indicate the number of vectors, The number of feature channels is given. The cue vector is randomly initialized using a normal distribution with a standard deviation of 0.02. During initialization, the cue vector does not carry any domain-specific information and acquires domain adaptability entirely through training. For the cue injection method, the cue vector and input features are concatenated along the sequence dimension and then fed into the encoder: ; in Indicates input features, This represents the encoder output features, which are the output features obtained after concatenating the cue vector with the input features and processing them by the encoder. These features are used for subsequent classification and detection. This represents concatenation along the sequence dimension; the cue vector, as a learnable "domain context," interacts with the input features in the self-attention mechanism, guiding the encoder to focus on key feature patterns in the target domain. The encoder's self-attention is calculated as follows: ; Among the queries ,key Value matrix Both contain information about cue vectors and input features, and the cue vectors influence the final feature representation through attention weight allocation; For the optimization strategy, only the cue vector parameters are updated during training, while the backbone network parameters remain fixed. The optimization objective of the cue vector is to minimize the classification loss and alignment loss in the target domain. in, The target domain has a labeled sample set; , These are the source and target domain datasets, respectively. Cross-entropy is used for classification loss; This represents the multi-granularity alignment loss. For the frozen backbone network parameters, To align the loss weights.

[0009] Preferably, multi-granularity alignment training includes: performing signal-level, feature-level, and semantic-level alignment on source domain data and target domain data; Signal-level alignment is performed in the original signal space, and the maximum mean difference (MMD) is used to measure the distribution difference. ; in, , These represent the number of samples in the source domain and the target domain, respectively. , These are underwater acoustic signal samples from the source and target domains, respectively. It is a Gaussian kernel mapping function; Feature-level alignment performs hierarchical alignment in the deep feature space, calculating MMD on shallow, middle, and deep features respectively: ; in, , Samples from the source and target domains are respectively processed by the backbone network. Features extracted from layers; For hierarchical weights; Semantic alignment is performed in the category conditional distribution space, and the conditional distribution differences are calculated separately for each category: ; in, The cue vector is the source domain. The cue vector is the source domain. Total number of categories; Indexed by category; Category labels; It is a deep feature representation; For category weights.

[0010] Preferably, the method of concatenating the optimized domain cue vector with the underwater acoustic signal features of the target domain to be detected and then feeding it into the backbone network to complete the underwater acoustic target classification includes: The trained and optimized cue vector is concatenated with the underwater acoustic signal features of the target domain to be detected. The encoder is guided to focus on the key feature patterns of the target domain through a self-attention mechanism, and feature representations with cross-domain invariance are extracted, ultimately achieving the classification and detection of underwater acoustic signals in the target domain.

[0011] The present invention also provides a domain-cue adaptive underwater acoustic signal detection system, the system being used to implement the aforementioned method, the system comprising: a data preparation and task division module, an initialization module, an optimization module, and a classification and detection module; The data preparation and task division module is used to collect source domain underwater acoustic signals and preprocess them, and divide the preprocessed source domain underwater acoustic signals into multiple related tasks according to the working condition dimension. The initialization module is used to obtain the initialization parameters of the backbone network through meta-learning pre-training based on multiple related tasks. The optimization module is used to perform domain cue adaptation and multi-granularity alignment training based on the initialization parameters of the backbone network, and optimize the domain cue vector. The classification and detection module is used to concatenate the optimized domain cue vector with the underwater acoustic signal features of the target domain to be detected and then send them into the backbone network to complete the underwater acoustic target classification.

[0012] Preferably, the process of dividing the preprocessed source domain underwater acoustic signal into multiple related tasks according to the operating condition dimension includes: The source domain data is divided into tasks according to the dimensions of sea area, season, sea state, and signal-to-noise ratio. Each task corresponds to a combination of working conditions, and each task contains 16 samples in the support set and 64 samples in the query set.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: By leveraging the synergistic effects of meta-learning pre-training, domain cue adaptation, and multi-granularity hierarchical alignment, this invention effectively addresses the issues of insufficient distribution alignment, weak adaptability, and high fine-tuning overhead in cross-domain underwater acoustic signal detection. Meta-learning pre-training endows the model with the ability to quickly adapt to new domains; domain cue adaptation achieves efficient cross-domain parameter transfer using lightweight cue vectors while avoiding catastrophic forgetting; and multi-granularity hierarchical alignment fully bridges the distribution differences between the source and target domains at the signal, feature, and semantic levels. This invention enables the model to efficiently adapt to multiple target domains while preserving source domain knowledge, achieving accurate cross-domain underwater acoustic signal detection. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a block diagram of the underwater acoustic signal cross-domain detection system architecture in an embodiment of the present invention; Figure 2 This is a flowchart of the cross-domain detection method in an embodiment of the present invention; Figure 3 This is a schematic diagram of the meta-learning pre-training mechanism in an embodiment of the present invention; Figure 4 This is a schematic diagram of the domain hint adaptive mechanism in an embodiment of the present invention; Figure 5 This is a diagram of a multi-granularity hierarchical domain alignment framework in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 This invention provides a domain-cue adaptive underwater acoustic signal detection method, the method comprising: Acquire underwater acoustic signals from the source region and preprocess them. Divide the preprocessed underwater acoustic signals from the source region into multiple related tasks according to the working condition dimension. Based on multiple related tasks, the initial parameters of the backbone network are obtained through meta-learning pre-training. Based on the initialization parameters of the backbone network, domain cue adaptation and multi-granularity alignment training are performed to optimize the domain cue vector; The optimized domain cue vector is concatenated with the underwater acoustic signal features of the target domain to be detected and then fed into the backbone network to complete the underwater acoustic target classification.

[0019] Specific implementation process: This method constructs a processing framework of "data preparation and task partitioning → meta-learning pre-training → domain prompting adaptation → multi-granularity domain alignment → object detection output".

[0020] For source domain underwater acoustic signal datasets The data covers five target categories: submarines, cargo ships, fishing vessels, whales, and background noise. Preprocessing includes mean and trend removal, fourth-order Butterworth bandpass filtering (20Hz-8kHz), and standardization. The source domain data is divided into multiple related tasks based on sea area, season, and sea state. Each task contains a support set (a small number of samples for rapid model adaptation) and a query set (samples for evaluating adaptation effectiveness). Stratified sampling is used to ensure class balance.

[0021] Meta-learning pre-training employs a model-independent meta-learning algorithm to learn a good set of initialization parameters. This allows the model to quickly adapt to new tasks through minimal gradient updates. The optimization objective of meta-learning is: in This is the inner learning rate. For each sampling task... First, perform an inner update on the support set. Then, evaluate the updated model on the query set and perform an outer update: in The outer learning rate is set to 0.01, and the inner learning rate is set to 0.001. The meta-learning loss combines classification loss and domain alignment loss, with a weighting coefficient of 0.3. This two-layer optimization structure ensures that the learned initial parameters have rapid adaptability.

[0022] After meta-learning pre-training, the model acquires good initial parameters and has the potential to quickly adapt to new tasks. When deployed to a new target domain (such as a new sea area, a new season, or underwater acoustic signals collected by new equipment), the model needs domain adaptation to maintain good performance due to the distribution differences between the target domain and the source domain. To address the problems of high parameter overhead and catastrophic forgetting in traditional fine-tuning methods, this invention introduces a domain cueing adaptive mechanism to achieve efficient cross-domain transfer. Its core idea is to freeze the backbone network parameters obtained from meta-learning pre-training and learn only a set of lightweight domain cueing vectors. These cueing vectors modulate the feature extraction process, enabling the model to adapt to the distribution characteristics of the target domain while fully preserving the general feature representation capabilities learned from the source domain.

[0023] Specifically, domain cue adaptation comprises three key components: cue vector design, cue injection method, and optimization strategy. For cue vector design, a set of learnable cue vectors is defined for each target domain. ,in To indicate the number of vectors, denoted as the number of feature channels. The cue vector is randomly initialized using a normal distribution with a standard deviation of 0.02. During initialization, the cue vector does not carry any domain-specific information and acquires domain adaptability entirely through training.

[0024] For the cue injection method, the cue vector and input features are concatenated along the sequence dimension and then fed into the encoder: in This represents the input features, specifically the feature representation extracted from the preprocessed underwater acoustic signal using a shallow network. This represents the encoder output features, which are the output features obtained after concatenating the cue vector with the input features and processing them by the encoder. These features are used for subsequent classification and detection. This represents concatenation along the sequence dimension. The cue vector, acting as a learnable "domain context," interacts with the input features in the self-attention mechanism, guiding the encoder to focus on key feature patterns in the target domain. The encoder's self-attention is calculated as follows: Among the queries ,key Value matrix Both contain information about cue vectors and input features, and the cue vectors influence the final feature representation through attention weight allocation.

[0025] For the optimization strategy, only the cue vector parameters are updated during training, while the backbone network parameters remain fixed. The optimization objective of the cue vector is to minimize the classification loss and alignment loss over the target domain. in, The target domain has a labeled sample set; , These are the source and target domain datasets, respectively. Cross-entropy is used for classification loss; This represents the multi-granularity alignment loss. For the frozen backbone network parameters, To align the loss weights, the Adam optimizer is used.

[0026] Domain cue adaptation provides a parameter-efficient cross-domain adaptation framework, but cue vectors alone are insufficient to fully bridge the multi-layered distribution differences between the source and target domains at the signal, feature, and semantic levels. Therefore, multi-granularity hierarchical domain alignment is introduced within the domain cue framework as the alignment loss during cue vector training. Distribution alignment is performed at three granularities: signal, feature, and semantic levels, eliminating distribution differences between the source and target domains and enabling the model to migrate from the "old environment" to the "new environment" without performance loss. Signal-level alignment is performed in the original signal space, using maximum mean difference (MMD) to measure distribution differences. in, , These represent the number of samples in the source domain and the target domain, respectively. , These are underwater acoustic signal samples from the source and target domains, respectively. This is a Gaussian kernel mapping function. Signal-level alignment eliminates distribution offsets caused by hardware factors such as hydrophone frequency response differences and sampling clock deviations, laying the foundation for high-level feature alignment.

[0027] Feature-level alignment performs hierarchical alignment in the deep feature space, calculating MMD on shallow, middle, and deep features respectively: in, , Samples from the source and target domains are respectively processed by the backbone network. Features extracted from layers; layer weights set to , , Shallow features contain the local time-frequency structure of the signal, mid-level features contain the target's local patterns such as propeller modulation features, and deep features contain the target's global semantic information. Layered MMD ensures effective transfer of features at each level of abstraction.

[0028] Semantic alignment is performed in the category conditional distribution space, and the conditional distribution differences are calculated separately for each category: in, The cue vector is the source domain. The cue vector is the source domain. Total number of categories; Indexed by category; Category labels; Deep feature representation; class weights The model is determined based on the proportion of samples from each category to address the class imbalance problem. Semantic-level alignment ensures that the model's discriminative ability for each category remains consistent across domains, avoiding the "domain shift" problem caused by edge distribution alignment. Three granularity alignment losses are weighted and fused. in, , , These are alignment losses at the signal level, feature level, and semantic level, respectively. , , For the corresponding weight coefficients, satisfying .

[0029] In the early stages of training, higher signal-level weights are used to quickly eliminate underlying differences, while in the later stages, higher semantic-level weights are used to finely adjust the category distribution.

[0030] During the training phase, the multi-granularity alignment loss, as part of the optimization objective, guides the cue vector to learn the distribution characteristics of the target domain, enabling it to gradually acquire domain adaptability.

[0031] During the detection phase, the trained and optimized cue vector is concatenated with the target domain signal features to be detected. The encoder is guided to focus on the key feature patterns of the target domain through a self-attention mechanism, thereby extracting feature representations with cross-domain invariance, and finally achieving accurate classification and detection of underwater acoustic signals in the target domain.

[0032] Specifically, in the detection phase, after preprocessing the input target domain underwater acoustic signal, the domain cue vector is concatenated with the input features (which are the feature representations obtained after preprocessing and initial feature extraction of the target domain underwater acoustic signal). The feature vector is then extracted by the backbone network, the classifier outputs the probability of each category, and the category with the highest probability is selected as the detection result.

[0033] Example 2 like Figure 2 As shown, this invention provides a multi-granularity cross-domain detection method for underwater acoustic signals based on domain cueing adaptation, comprising the following steps: Step 1: Data Collection and Task Division. Source domain data acquisition uses a standard hydrophone with a sampling rate of 16kHz and a signal duration of 2 seconds. The source domain data covers multiple sea areas, four seasons (spring, summer, autumn, and winter), and three sea states (calm, moderate, and severe), including five target categories: submarines, cargo ships, fishing boats, whales, and background noise. Preprocessing includes mean and trend removal, bandpass filtering (20Hz-8kHz fourth-order Butterworth), standardization (mean 0, standard deviation 1), and 50% overlap segmentation. The source domain data is divided into tasks based on sea area, season, sea state, and signal-to-noise ratio, with each task corresponding to a specific combination of operating conditions. Each task contains 16 samples in the support set and 64 samples in the query set, using stratified sampling to ensure class balance. Target domain data consists of underwater acoustic signals from new sea areas, with acquisition parameters consistent with the source domain.

[0034] Specifically, the task division first enumerates and combines the working condition dimensions of the source domain data, including multiple sea areas, seasons (spring, summer, autumn, winter), sea states (calm, moderate, severe), and signal-to-noise ratio (low, medium, high). Each working condition combination is defined as an independent task.

[0035] For each task, all samples that meet the conditions of that task are selected from the source domain dataset. Then, a stratified sampling strategy is used to construct the support set and the query set. During stratified sampling, samples are drawn separately for five categories: submarines, cargo ships, fishing boats, whales, and background noise. This ensures that each category has 3 to 4 samples in the support set and 12 to 13 samples in the query set, thereby guaranteeing a balanced distribution of categories.

[0036] Finally, work combination combinations that could not meet the sampling requirements due to insufficient sample size were removed, and only valid tasks were retained for meta-learning pre-training.

[0037] Step 2: Meta-learning pre-training. For example... Figure 3 As shown, a model-independent meta-learning algorithm is used for pre-training on the source domain data to learn a good set of initialization parameters. This allows the model to quickly adapt to new tasks through minimal gradient updates. The optimization objective of meta-learning is: in, The meta-parameters of the backbone network are the target of meta-learning optimization. To distribute from tasks; For the first sample Each working condition task; For the task loss function on; For A detection model with parameters; The inner learning rate is set to 0.01, which controls the step size for rapid adaptation on the support set. For parameters Operators for finding gradients.

[0038] Sample multiple tasks from the task distribution to form a batch, for each sampled task First, perform an inner update on the support set: in, For the task Support set; The inner learning rate is used. The inner update simulates the model's rapid adaptation process on new tasks. The updated model is then evaluated on the query set, and the meta-parameters are optimized by aggregating the losses from all tasks through the outer update. in, For the task Query set; To update the task adaptation parameters one step on the support set; The outer learning rate is denoted as . The meta-learning loss combines the classification loss and the domain alignment loss, with a weight of 0.3. After sufficient iterations of meta-learning, initial parameters with rapid adaptability are obtained. This two-layer optimization structure ensures that the learned initial parameters have rapid adaptability.

[0039] Step 3: Domain Adaptation. After meta-learning pre-training, the model is deployed to the target domain. Since there are distributional differences between the target and source domains, domain adaptation is necessary. For example... Figure 4 As shown, the backbone network parameters obtained from frozen meta-learning pre-training are... Initialize the target domain hint vector This embodiment indicates the number of prompt vectors. Number of feature channels The number of hint parameters is only 1.5% of that of the backbone network. The hint vectors are randomly initialized using a normal distribution with a standard deviation of 0.02.

[0040] During feature extraction, the cue vector and input features are concatenated along the sequence dimension and then fed into the encoder. The cue vector, acting as a learnable "domain context," interacts with the input features in the self-attention mechanism, guiding the encoder to focus on key feature patterns in the target domain through attention weight allocation. Specifically, the self-attention computation in the encoder incorporates joint information from the cue vector and input features. The cue vector influences the aggregation method of features through key-value pair matching, thereby achieving adaptive modulation of the target domain distribution.

[0041] During training, only the cue vector parameters are updated, and the optimization objective is: The Adam optimizer was used with a learning rate of 0.001 for 500 training epochs. The advantages of domain cue adaptation are: high parameter efficiency, avoiding complete fine-tuning of the large-scale backbone network and significantly reducing computational overhead; preservation of source domain knowledge, with backbone parameters completely frozen, avoiding catastrophic forgetting (i.e., the model losing previously learned knowledge in the source domain when learning a new domain), allowing the model to maintain good performance in both the source and target domains; and ease of multi-domain expansion, as each new domain only needs to store its corresponding cue vector, and the cue vectors for each domain are trained independently without interference.

[0042] Step 4: Multi-granular hierarchical domain alignment. Domain cue adaptation provides a parameter-efficient adaptation framework, but cue vectors alone are insufficient to fully bridge the multi-level distribution differences between the source and target domains. For example... Figure 5 As shown, multi-granularity alignment training is performed within the domain cueing framework, with the alignment loss serving as part of the cue vector optimization objective.

[0043] Signal-level alignment uses Gaussian kernel MMD to measure the difference in signal distribution between the source and target domains: in Using a Gaussian kernel mapping function, signal-level alignment eliminates low-level offsets caused by hardware and environmental factors. Feature-level alignment calculates MMD on shallow, mid-level, and deep features respectively. Hierarchical weight , , This ensures effective feature transfer across abstract levels. Semantic-level alignment calculates conditional distribution differences by category: Category weight The method addresses class imbalance by determining the proportion of samples. A weighted fusion of three granularity alignment losses is used. In the early stages of training, the signal-level weights are 0.5; in the later stages, the semantic-level weights are 0.5; and the weight for the overall alignment loss is 0.5.

[0044] Step 5: Underwater Acoustic Signal Detection. After model training, underwater acoustic signals in the target domain are detected. The input signal, after preprocessing, is concatenated with the target domain cue vector and fed into the backbone network, outputting a 128-dimensional feature vector. The classifier uses a multi-layer fully connected network, with the last layer using Softmax activation to output the probabilities of each class. The class with the highest probability is selected as the detection result, and a confidence score is output. Post-processing includes confidence threshold filtering (threshold 0.5) and time series smoothing (to eliminate transient false detections).

[0045] The final model consists of two parts: a backbone network (encoder + classifier) ​​pre-trained and frozen by meta-learning, and a cue vector optimized for the target domain.

[0046] Detailed Backbone Network Architecture Description Input Representation Layer: The input is a preprocessed underwater acoustic signal with a sampling rate of 16kHz, a duration of 2 seconds, and a total of 32,000 sampling points. First, the original signal is divided into frames of length 256, shifted by 128 (50% overlap), resulting in approximately 249 frames. Each frame is then mapped to a 128-dimensional feature vector through a one-dimensional convolutional embedding layer (kernel size 7, stride 1, output channels 128), forming a feature sequence. ,in The sequence length is given.

[0047] The encoder employs a Transformer encoder architecture, comprising four self-attention layers. Each layer includes a multi-head self-attention sublayer (8 attention heads, each with a dimension of 16) and a feedforward neural network sublayer (hidden layer dimension 512). Each sublayer is followed by normalization and residual connections. The encoder input is a concatenation of the cue vector and the feature sequence. , where the hint vector The shallow layer is defined as the output of layer 1, the middle layer as the output of layer 2, and the deep layer as the output of layer 4, which are used for feature-level alignment. The encoder's final output is then subjected to global average pooling to obtain a 128-dimensional feature vector.

[0048] Classifier: A three-layer fully connected network is used, with a structure of 128→64→32→5. The hidden layers use the ReLU activation function and Dropout (dropout rate 0.3) to prevent overfitting. The output layer uses the Softmax activation function to output the probability distribution of 5 target classes (submarine, cargo ship, fishing boat, whale, and background noise).

[0049] Example 3 The present invention also provides a domain-cue adaptive underwater acoustic signal detection system, which is used to implement the method described in Embodiment 1. The system includes: a data preparation and task division module, an initialization module, an optimization module, and a classification and detection module. The data preparation and task division module is used to collect source domain underwater acoustic signals and preprocess them, and divide the preprocessed source domain underwater acoustic signals into multiple related tasks according to the working condition dimension. The initialization module is used to obtain the initialization parameters of the backbone network through meta-learning pre-training based on multiple related tasks. The optimization module is used to perform domain cue adaptation and multi-granularity alignment training based on the initialization parameters of the backbone network, and optimize the domain cue vector. The classification and detection module is used to concatenate the optimized domain cue vector with the underwater acoustic signal features of the target domain to be detected and then send them into the backbone network to complete the underwater acoustic target classification.

[0050] The process of dividing the preprocessed source domain underwater acoustic signal into multiple related tasks according to the operating condition dimension includes: The source domain data is divided into tasks according to the dimensions of sea area, season, sea state, and signal-to-noise ratio. Each task corresponds to a combination of working conditions, and each task contains 16 samples in the support set and 64 samples in the query set.

[0051] The specific implementation process is as follows: Figure 1 As shown, this invention provides a multi-granularity cross-domain detection system for underwater acoustic signals based on domain cue adaptation, comprising: a data acquisition module for receiving and preprocessing underwater acoustic signals acquired by a hydrophone; a task partitioning module for dividing source domain data into multiple related tasks according to operating conditions; a meta-learning pre-training module for learning well-defined initialization parameters using a model-independent meta-learning algorithm; a domain cue adaptation module for freezing the backbone network and learning target domain-specific cue vectors to achieve efficient domain adaptation of parameters; a multi-granularity alignment module for performing three-level alignment at the signal level, feature level, and semantic level; a classification detection module for outputting the underwater acoustic target category and confidence level; and a result output module for displaying and storing the detection results.

[0052] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for underwater acoustic signal detection based on domain cueing adaptation, characterized in that, The method includes: Acquire underwater acoustic signals from the source region and preprocess them. Divide the preprocessed underwater acoustic signals from the source region into multiple related tasks according to the working condition dimension. Based on multiple related tasks, the initial parameters of the backbone network are obtained through meta-learning pre-training. Based on the initialization parameters of the backbone network, domain cue adaptation and multi-granularity alignment training are performed to optimize the domain cue vector; The optimized domain cue vector is concatenated with the underwater acoustic signal features of the target domain to be detected and then fed into the backbone network to complete the underwater acoustic target classification.

2. The method according to claim 1, characterized in that, Methods for dividing preprocessed source domain underwater acoustic signals into multiple related tasks according to operating conditions include: The source domain data is divided into tasks according to the dimensions of sea area, season, sea state, and signal-to-noise ratio. Each task corresponds to a combination of working conditions, and each task contains 16 samples in the support set and 64 samples in the query set.

3. The method according to claim 2, characterized in that, Methods for obtaining the initial parameters of the backbone network through meta-learning pre-training based on multiple related tasks include: For each sampling task First, perform an inner update on the support set: ； in, The inner learning rate, For the task Support set, The meta-parameters of the backbone network are the target of meta-learning optimization. For the task loss function on; For parameters Operators for finding the gradient; To update the task adaptation parameters one step on the support set; Then, evaluate the updated model on the query set and perform an outer update: ； in, The outer learning rate, For the task The query set.

4. The method according to claim 1, characterized in that, Domain-adaptive hinting includes hint vector design, hint injection methods, and optimization strategies. Specifically, for cue vector design, a set of learnable cue vectors is defined for each target domain. ,in To indicate the number of vectors, The number of feature channels is given. The cue vector is randomly initialized using a normal distribution with a standard deviation of 0.

02. During initialization, the cue vector does not carry any domain-specific information and acquires domain adaptability entirely through training. For the cue injection method, the cue vector and input features are concatenated along the sequence dimension and then fed into the encoder: ； in Indicates input features, This represents the encoder output features, which are the output features obtained after concatenating the cue vector with the input features and processing them by the encoder. These features are used for subsequent classification and detection. This represents concatenation along the sequence dimension; the cue vector, as a learnable "domain context," interacts with the input features in the self-attention mechanism, guiding the encoder to focus on key feature patterns in the target domain. The encoder's self-attention is calculated as follows: ； Among the queries ,key Value matrix Both contain information about cue vectors and input features, and the cue vectors influence the final feature representation through attention weight allocation; For the optimization strategy, only the cue vector parameters are updated during training, while the backbone network parameters remain fixed. The optimization objective of the cue vector is to minimize the classification loss and alignment loss in the target domain. in, The target domain has a labeled sample set; , These are the source and target domain datasets, respectively. Cross-entropy is used for classification loss; This represents the multi-granularity alignment loss. For the frozen backbone network parameters, To align the loss weights.

5. The method according to claim 4, characterized in that, Multi-granularity alignment training includes: performing signal-level, feature-level, and semantic-level alignment on source domain data and target domain data; Signal-level alignment is performed in the original signal space, and the maximum mean difference (MMD) is used to measure the distribution difference. ； in, , These represent the number of samples in the source domain and the target domain, respectively. , These are underwater acoustic signal samples from the source and target domains, respectively. It is a Gaussian kernel mapping function; Feature-level alignment performs hierarchical alignment in the deep feature space, calculating MMD on shallow, middle, and deep features respectively: ； in, , Samples from the source and target domains are respectively processed by the backbone network. Features extracted from layers; For hierarchical weights; Semantic alignment is performed in the category conditional distribution space, and the conditional distribution differences are calculated separately for each category: ； in, The cue vector is the source domain. The cue vector is the source domain. Total number of categories; Indexed by category; Category labels; It is a deep feature representation; This represents the category weight.

6. The method according to claim 1, characterized in that, The optimized domain cue vector is concatenated with the underwater acoustic signal features of the target domain to be detected and then fed into the backbone network to complete underwater acoustic target classification. Methods include: The trained and optimized cue vector is concatenated with the underwater acoustic signal features of the target domain to be detected. The encoder is guided to focus on the key feature patterns of the target domain through a self-attention mechanism, and feature representations with cross-domain invariance are extracted, ultimately achieving the classification and detection of underwater acoustic signals in the target domain.

7. A domain-cue adaptive underwater acoustic signal detection system, the system being used to implement the method according to any one of claims 1-6, characterized in that, The system includes: a data preparation and task allocation module, an initialization module, an optimization module, and a classification and detection module; The data preparation and task division module is used to collect source domain underwater acoustic signals and preprocess them, and divide the preprocessed source domain underwater acoustic signals into multiple related tasks according to the working condition dimension. The initialization module is used to obtain the initialization parameters of the backbone network through meta-learning pre-training based on multiple related tasks. The optimization module is used to perform domain cue adaptation and multi-granularity alignment training based on the initialization parameters of the backbone network, and optimize the domain cue vector. The classification and detection module is used to concatenate the optimized domain cue vector with the underwater acoustic signal features of the target domain to be detected and then send them into the backbone network to complete the underwater acoustic target classification.

8. The system according to claim 7, characterized in that, The process of dividing the preprocessed source domain underwater acoustic signal into multiple related tasks according to the operating condition dimension includes: The source domain data is divided into tasks according to the dimensions of sea area, season, sea state, and signal-to-noise ratio. Each task corresponds to a combination of working conditions, and each task contains 16 samples in the support set and 64 samples in the query set.

Images