An underwater acoustic target recognition method and system based on general acoustic knowledge transfer of a voice large model, and a storage medium

Abstract

The present application relates to the technical field of underwater acoustic target recognition (UATR), and proposes an underwater acoustic target recognition method and system based on general acoustic knowledge transfer of a large speech model and a storage medium, to solve the problems of data scarcity and insufficient cross-domain generalization performance in underwater acoustic target recognition for a long time. A large speech model pre-trained on large-scale human speech dialogue data is selected as the acoustic encoder backbone, and its original feature extractor is retained to fully utilize the general acoustic representation ability learned by the model. The original decoder is replaced by a lightweight linear classifier to adapt to the underwater acoustic target classification task. A small amount of underwater acoustic samples are used to fine-tune the system end-to-end, so that the model learns the mapping relationship from general speech features to underwater acoustic target acoustic fingerprints. The method of the present application not only can realize high-precision underwater acoustic target recognition with extremely low training cost, but also shows excellent robustness and excellent cross-scene generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic target recognition (UATR) technology, specifically relating to an underwater acoustic target recognition method, system, and storage medium based on the transfer of general acoustic knowledge from a large speech model. Background Technology

[0002] Underwater acoustic target recognition (UATR) is an important task in marine research and operations, supporting applications such as biodiversity monitoring, search and rescue, seabed mapping, and maritime safety (Xie Y, Ren J, Xu J. Unraveling complex data diversity in underwater acoustic target recognition through convolution-based mixture of experts[J]. Expert Systemswith Applications, 2024, 249: 123431.). However, building a highly reliable underwater acoustic target recognition system still faces two major challenges: First, data scarcity. The acquisition and annotation of large-scale underwater acoustic data is costly (Tian S, Bai D, Zhou J, et al. Few-shot learning for joint model in underwateracoustic target recognition[J]. Scientific Reports, 2023, 13(1): 17502.), and the data distribution varies significantly under different acoustic environments, making it difficult to directly reuse data to form a unified dataset; Second, domain complexity. Underwater acoustic signals are affected by non-stationary noise, multipath effects, and spatiotemporal variations, resulting in unstable signal characteristics and posing a significant challenge to the generalization ability of the model.

[0003] Existing research mainly addresses the above problems through two types of strategies: one is the data center approach, such as data augmentation techniques like spectral masking, pitch transformation, and generative synthesis, which attempts to expand the diversity of the limited training set; the other is the model center approach, including noise-robust feature extraction, domain adaptation, and adversarial training, to improve the model's generalization ability under different acoustic conditions (Domingos LCF, Santos PE, Skelton PSM, et al. A survey of underwater acoustic data classification methods using deep learning forshoreline surveillance[J]. Sensors, 2022, 22(6): 2181.). Although these methods improve performance to some extent, they are often limited by the limited scale of pre-trained data in the underwater acoustics field, making it difficult to capture general acoustic patterns that can be transferred across domains.

[0004] In recent years, the emergence of SLM has provided new ideas for solving these problems. SLM, pre-trained on large-scale human speech corpora, can learn rich acoustic priors and highly generalized representations, and has demonstrated strong transfer capabilities in various tasks such as automatic speech recognition, speaker recognition, emotion recognition, and environmental sound classification (Mohamed A, Lee H, Borgholt L, et al. Self-supervised speech representation learning: A review[J]. IEEE Journal of Selected Topics in Signal Processing, 2022, 16(6): 1179-1210.). Although human speech and underwater acoustic signals have significant differences in physical properties, they still share spectral patterns and temporal structures at a low level, which makes it possible to transfer the acoustic knowledge of SLM for underwater acoustic target recognition.

[0005] However, current research applying SLM to underwater acoustics remains limited. Existing work often only transfers a portion of the encoder or freezes most model parameters, lacking systematic adaptation and end-to-end optimization for underwater acoustic tasks, thus failing to fully unleash the potential of pre-trained models. Therefore, there is an urgent need for an end-to-end method that can fully utilize the general acoustic knowledge of SLM and combine it with a small amount of underwater acoustic data for efficient fine-tuning, in order to significantly improve the accuracy, robustness, and cross-domain generalization ability of underwater acoustic target recognition. Based on this need, this invention proposes to utilize the feature extraction and acoustic coding capabilities of large speech models, through data standardization, structural modification, and end-to-end few-sample fine-tuning, to transform a general speech model into a high-performance underwater acoustic target recognition model. Summary of the Invention

[0006] The purpose of this invention is to provide an underwater acoustic target recognition method, system, and storage medium based on the transfer of general acoustic knowledge from a large speech model, which solves the technical problems of low model recognition accuracy, poor robustness, and weak generalization ability in existing underwater acoustic target recognition technologies due to the scarcity of training data and the complex and variable marine environment.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] Underwater target recognition methods based on general acoustic knowledge transfer from large speech models include:

[0009] Underwater acoustic audio data from different acoustic environments are collected, preprocessed to obtain underwater acoustic target recognition dataset D, and dataset D is divided into training set, validation set and test set;

[0010] Obtain the network structure of the general speech model SLM and split it according to the component structure. Keep the acoustic encoder and its input format adapted feature extractor pre-trained on a large-scale human speech corpus, and remove the decoder part.

[0011] The acoustic decoder is replaced with a lightweight classifier, and the output range R of the classifier is made consistent with the number of categories in the underwater acoustic target recognition dataset D to form a new large-scale speech classification model.

[0012] The large speech classification model is fine-tuned end-to-end using the training set of the underwater acoustic target recognition dataset D, and the optimal model is selected through the validation set to obtain the large speech model UATR-SLM for underwater acoustic target recognition.

[0013] Input the test set of dataset D into the trained underwater acoustic target recognition speech model UATR-SLM to perform underwater acoustic target detection and output the predicted category results.

[0014] Furthermore, the process of collecting underwater acoustic audio data from different acoustic environments, classifying, resampling, and segmenting the data, and matching the segmented audio segments with corresponding category labels specifically includes:

[0015] For two open-source underwater audio files from completely different acoustic environments, classify them according to ship type or recognized classification rules;

[0016] For audio files that have been categorized, perform uniform resampling to ensure consistent sampling rates.

[0017]

[0018] in, Represents the original discrete audio signal at the original sampling rate. The k-th sample value; This represents the resampled discrete audio signal at the target sampling rate. The m-th sample value; Representative standard The function is the impulse response of an ideal low-pass filter;

[0019] The resampled audio file is length-segmented to obtain several audio segments, and corresponding category labels are matched to form dataset D;

[0020] Divide D into training, validation and test sets in an 8:1:1 ratio.

[0021] Furthermore, the SLM model is decomposed according to its component structure, specifically including:

[0022] The model structure is disassembled, separating the acoustic encoder from the decoder;

[0023] The hierarchical stacking relationship of the acoustic encoder remains unchanged, and it is adapted to the feature extractor of the SLM model to ensure that the input feature format is consistent with that of the pre-training stage.

[0024] Furthermore, the lightweight classifier is a linear classifier, specifically including:

[0025] Determine the dimension N of the acoustic encoder output vector and the number of target categories M based on the number of categories in dataset D;

[0026] Initialize the parameters of a linear layer with shape (N, M), assign weights randomly, and use it as a classifier;

[0027] After connecting the linear classifier to the acoustic encoder, the high-dimensional feature vector output by the encoder is received and mapped to the prediction result of the target category.

[0028] Furthermore, the large speech classification model input from the training set is fine-tuned end-to-end, and the loss is calculated using the cross-entropy loss function to assist backpropagation in updating the model parameters, specifically as follows:

[0029] Set hyperparameters for model training, including learning rate, batch size, and number of training epochs;

[0030] Based on the batch size, the audio samples in the training set are processed in batches, and then these data are input into the large speech classification model for training.

[0031] Forward propagation is performed on the model to predict audio categories, the loss value is calculated using the cross-entropy loss function, and the parameters of the large speech classification model are updated using the backpropagation algorithm. The backpropagation process is as follows:

[0032]

[0033] in, The logit value represents the j-th class, used to calculate the predicted probability of that class; These represent model parameters, which are continuously updated during model training to minimize the loss. This represents the learning rate, which controls the step size for each parameter update. Let represent the predicted probability of class i, which is obtained by normalizing logit using the softmax function, and c represents the total number of classes in the classification task. Cross-entropy loss is used to measure the difference between the predicted probability distribution and the true label distribution. This indicates that the loss function is applied to the parameters. The gradient indicates the direction of parameter updates.

[0034] After each round of parameter updates, the adjusted model is used to calculate the current validation loss using the validation set. After all training is completed, the model with the smallest loss on the validation set is retained as the final underwater acoustic target recognition speech model UATR-SLM.

[0035] Furthermore, the process of performing underwater acoustic target detection and outputting predicted category results specifically includes:

[0036] Extract the underwater acoustic audio files from the test set corresponding to dataset D, feed the audio files into the UATR-SLM model and start the model prediction;

[0037] The model processes underwater acoustic audio files and outputs confidence probabilities for different categories in the final linear layer. The category with the highest probability is used as the prediction result for underwater acoustic target detection.

[0038] A computer device / apparatus / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of an underwater acoustic target recognition method based on general acoustic knowledge transfer from a large speech model.

[0039] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of an underwater target recognition method based on general acoustic knowledge transfer from a large speech model.

[0040] A computer program product includes a computer program / instructions that, when executed by a processor, implement steps for an underwater target recognition method based on general acoustic knowledge transfer from a large speech model.

[0041] An electronic device, comprising:

[0042] Memory, used to store computer programs;

[0043] A processor is used to execute the computer program to implement an underwater target recognition method based on general acoustic knowledge transfer from a large speech model.

[0044] The beneficial effects of this invention are as follows:

[0045] 1. This invention innovatively applies a large-scale speech model trained on a massive human speech corpus to the field of underwater acoustic target recognition, effectively solving the two long-standing problems of "data scarcity" and "cross-domain difficulties" in the underwater acoustic field. This method eliminates the need to collect and label massive amounts of expensive underwater acoustic data for pre-training; instead, it relies more on the strong acoustic prior knowledge of existing speech models plus a small amount of data for fine-tuning, significantly reducing the development threshold and cost of high-performance underwater acoustic recognition systems.

[0046] 2. The UATR-SLM framework proposed in this invention achieves top-tier recognition performance through a simple and effective architecture that reuses the speech feature extraction pipeline, adapts a large speech model as an acoustic encoder, and combines it with a lightweight classifier. To better demonstrate the universality of this invention, two different datasets, D1 and D2, were constructed using the aforementioned method for constructing underwater acoustic target recognition datasets. The recognition accuracy of this invention on both test sets exceeds 99%, significantly outperforming the traditional ResNet baseline model and other advanced domain-specific models, setting a new performance benchmark for underwater acoustic target recognition technology.

[0047] 3. This invention demonstrates superior robustness and cross-domain generalization ability, solving the problem of unstable performance of traditional models in variable marine environments. In the variable signal length test, a high accuracy of 95.87% can be achieved using only a 1-second audio clip, far exceeding the model trained from scratch. More importantly, in the zero-shot cross-domain test (i.e., trained on the D1 dataset and tested on the same class samples in the completely unseen D2 dataset), the accuracy reaches as high as 96.67%, proving that this invention can learn universal acoustic essential features independent of specific environments.

[0048] 4. The method provided by this invention possesses broad versatility and scalability, and is not limited to specific large-scale speech models. It establishes a new paradigm for applying fundamental speech models to non-speech acoustic fields, applicable to a variety of existing and future large-scale speech models. Research has demonstrated the enormous potential of this cross-domain knowledge transfer, providing important reference and guidance for solving the data scarcity problem in other acoustic signal processing applications (such as marine biological acoustic monitoring and mechanical fault diagnosis). Attached Figure Description

[0049] Figure 1 An overview diagram of the underwater acoustic target recognition method based on general acoustic knowledge transfer from a large speech model provided by this invention;

[0050] Figure 2 Diagram of underwater acoustic dataset construction. Detailed Implementation

[0051] The present invention will now be further described with reference to the accompanying drawings.

[0052] The purpose of this invention is to provide an underwater acoustic target recognition method, system, and storage medium based on the transfer of general acoustic knowledge from a large speech model. This invention aims to solve the technical problems of low model recognition accuracy, poor robustness, and weak generalization ability in existing underwater acoustic target recognition technologies due to the scarcity of training data and the complex and variable marine environment.

[0053] like Figure 1 Figure 2 As shown, the method of the present invention includes the following steps:

[0054] Step 1: Construct underwater acoustic target recognition dataset D, obtain underwater acoustic audio data in different acoustic environments, and preprocess them respectively to obtain the required dataset;

[0055] The collected underwater acoustic audio data is classified, resampled, and segmented. The segmented audio segments are then matched with corresponding category labels. Specifically, this includes:

[0056] Step 1.1: Extract underwater acoustic audio files and classify the two types of samples according to ship type or recognized classification rules;

[0057] Step 1.2: Perform uniform resampling on the dataset after class division to ensure consistent sampling rates. The resampling process can be represented as:

[0058]

[0059] in, Represents the original discrete audio signal at the original sampling rate. The k-th sample value; This represents the resampled discrete audio signal at the target sampling rate. The m-th sample value; Representative standard The function is the impulse response of an ideal low-pass filter;

[0060] Step 1.3: The resampled audio file is segmented into several audio segments of 5 seconds each, and corresponding category labels are matched to form dataset D. Finally, dataset D is divided into training, validation, and test sets in an 8:1:1 ratio.

[0061] Step 2: Retain the acoustic encoder of the general speech large model (SLM) and its input format adaptation feature extractor, and remove the decoder part;

[0062] Step 2.1: Disassemble the model structure and separate the acoustic encoder from the decoder;

[0063] Step 2.2: Keep the hierarchical stacking relationship of the acoustic encoder unchanged and adapt it to the feature extractor of the original model to ensure that the input feature format is consistent with the pre-training stage.

[0064] Step 3: Replace the original decoder with a lightweight classifier, and make the output range R of the classifier consistent with the number of categories in the underwater acoustic target recognition dataset D, thereby forming a new large-scale speech classification model;

[0065] Step 3.1: Determine the dimension N of the acoustic encoder output vector and determine the number of target categories M based on the number of categories in dataset D;

[0066] Step 3.2: Initialize the parameters of a linear layer with shape (N, M), assign weights randomly, and use it as a classifier;

[0067] Step 3.3: After connecting the linear classifier to the acoustic encoder, receive the high-dimensional feature vector output by the encoder and map it to the prediction result of the target category.

[0068] Step 4: Use the underwater acoustic target recognition dataset D to fine-tune the large speech classification model end-to-end to obtain the high-performance underwater acoustic target recognition speech model UATR-SLM;

[0069] The training set is input into a large speech classification model for training, and the loss is calculated using the cross-entropy loss function to assist in backpropagation to update the model parameters. Then, the optimal model is obtained through the validation set. Specifically:

[0070] Step 4.1: Set hyperparameters for model training, including learning rate, batch size, training epochs, etc.

[0071] Step 4.2: Based on the batch size, process the audio samples in the training set into batches, and then input these data batches into the large speech classification model for training;

[0072] Step 4.3: Perform forward propagation on the model to predict the audio category, calculate the loss value using the cross-entropy loss function, and update the large speech classification model SLM using the backpropagation algorithm. classification Parameters. The entire backpropagation process can be represented by the following function:

[0073]

[0074] in, The logit value represents the i-th class, used to calculate the predicted probability of that class; This represents the model parameters, which are continuously updated during model training to minimize the loss; This represents the learning rate, which controls the step size for each parameter update. The predicted probability of class i is obtained by normalizing logit using the softmax function; Cross-entropy loss is used to measure the difference between the predicted probability distribution and the true label distribution. This indicates that the loss function is applied to the parameters. The gradient indicates the direction of parameter updates;

[0075] Step 4.4: After each round of parameter updates, the adjusted model is used to calculate the current validation loss using the validation set. After the entire training process is completed, the model with the smallest loss on the validation set is retained as the final underwater acoustic target recognition speech model UATR-SLM.

[0076] Step 5: Input the underwater acoustic audio files in the test set of dataset D from step 4 into the trained underwater acoustic target recognition speech model UATR-SLM for underwater acoustic target detection.

[0077] The underwater audio clips from the test set are fed into the UATR-SLM to obtain the predicted class results of the final linear layer output, which are then compared with the class of the true source:

[0078] Step 5.1: Extract the underwater acoustic audio files from the test set in dataset D, input the audio files into UATR-SLM and start the model prediction;

[0079] Step 5.2: The model processes the underwater audio file and outputs the confidence probabilities for different categories in the final linear layer, taking the category with the highest probability as the model's prediction result;

[0080] Step 5.3: Compare the predicted category results with the actual category of the current underwater acoustic audio file to determine whether the current prediction result is accurate;

[0081] Step 5.4: Compare all predicted results with the actual results and calculate the model's recognition accuracy.

[0082] Step 6: Compare our trained UATR-SLM model with the ResNet baseline model and some excellent underwater acoustic target recognition models from recent years;

[0083] To better demonstrate the effectiveness of this invention, two datasets, D1 and D2, with completely different acoustic environments were constructed using the method described in step 1 above. The accuracy of each model on its corresponding test set after training on the training sets of datasets D1 and D2 was statistically analyzed, as well as its accuracy under varying audio lengths or zero-sample cross-domain conditions (i.e., training on dataset D1 and testing on samples of the same class in the completely unseen dataset D2).

[0084] Step 6.1: On the underwater acoustic target recognition datasets D1 and D2, which are completely different acoustic environments, the recognition accuracy of the present invention on its corresponding test set after training on the training set exceeds 99%, which is significantly better than the traditional ResNet series baseline models (92%~96%) and other advanced domain-specific models (94%~98%), setting a new performance benchmark for underwater acoustic target recognition technology.

[0085] Step 6.2: The model trained using our method in the variable audio length scenario was trained on the D1 dataset (5s long audio) and tested on audio of (1~20s). It achieved a high accuracy of 95.87% in 1-second audio segments, far exceeding the ResNet series baseline (27%~33%) we selected, demonstrating a strong ability to adapt to variable audio length.

[0086] Step 6.3: In zero-shot cross-domain scenarios, our model achieved an accuracy of 96.67% on all samples of the same class in the D2 dataset, far exceeding the ResNet series baseline (53%~70%), demonstrating extremely strong cross-domain generalization ability.

[0087] This patent provides an efficient, robust, and cost-effective underwater acoustic target recognition method. By cleverly combining mature technologies in the field of speech recognition with underwater acoustic applications, it significantly reduces reliance on specialized domain data while elevating recognition performance to industry-leading levels, demonstrating strong environmental adaptability and generalization.

[0088] In particular, in some preferred embodiments of the present invention, a computer device is also provided, including a memory and a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the underwater acoustic target recognition method based on general acoustic knowledge transfer of a large speech model as described in any of the above embodiments.

[0089] In some other preferred embodiments of the present invention, a computer-readable storage medium is also provided, on which a computer program / instruction is stored, wherein when the computer program is executed by a processor, the steps of the underwater acoustic target recognition method based on the transfer of general acoustic knowledge of a large speech model described in any of the above embodiments are implemented.

[0090] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the underwater acoustic target recognition method embodiment based on the transfer of general acoustic knowledge of speech large model described above, which will not be repeated here.

[0091] Computer-readable storage media encompass a variety of types, including persistent and non-persistent, portable and fixed. These media store information using different technologies, and the content can be machine instructions, data structures, program modules, or other types of data. Some typical examples of computer storage media include: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), various types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory and other storage technologies, optical storage media such as CD-ROM and digital video disc (DVD), magnetic storage devices such as magnetic tape and disks, and other non-transferable media used to store information accessible to computing devices. It is important to note that the computer-readable media described herein do not include temporary storage media, such as modulated data signals and carrier waves.

[0092] Those skilled in the art will further recognize that the operation of the module can be achieved using existing technical protocols or programs, without relying on new computer programs themselves. The units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0093] The steps of the methods or algorithms described in conjunction with the embodiments disclosed in this invention can be implemented in hardware, software modules executed by a processor, or a combination of both. The software modules can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for underwater acoustic target recognition based on general acoustic knowledge transfer from a large speech model, characterized in that: include: Underwater acoustic audio data from different acoustic environments are collected, preprocessed to obtain underwater acoustic target recognition dataset D, and dataset D is divided into training set, validation set and test set; Obtain the network structure of the general speech model SLM and split it according to the component structure. Keep the acoustic encoder and its input format adapted feature extractor pre-trained on a large-scale human speech corpus, and remove the decoder part. The acoustic decoder is replaced with a lightweight classifier, and the output range R of the classifier is made consistent with the number of categories in the underwater acoustic target recognition dataset D, forming a new large-scale speech classification model. The large speech classification model is fine-tuned end-to-end using the training set of the underwater acoustic target recognition dataset D, and the optimal model is selected through the validation set to obtain the large speech model UATR-SLM for underwater acoustic target recognition. Input the test set of the underwater acoustic target recognition dataset D into the trained underwater acoustic target recognition speech model UATR-SLM to perform underwater acoustic target detection and output the predicted category results.

2. The underwater acoustic target recognition method based on general acoustic knowledge transfer from a large speech model according to claim 1, characterized in that, The process involves collecting underwater acoustic audio data from different acoustic environments, classifying, resampling, and segmenting the data, and then matching the segmented audio segments with corresponding category labels. Specifically, this includes: Underwater acoustic audio files for different acoustic environments are categorized according to ship type or recognized classification rules. For audio files that have been categorized, perform uniform resampling to ensure consistent sampling rates. in, Represents the original discrete audio signal at the original sampling rate. The k-th sample value; This represents the resampled discrete audio signal at the target sampling rate. The m-th sample value; Representative standard The function is the impulse response of an ideal low-pass filter; The resampled audio file is length-segmented to obtain several audio segments, and corresponding category labels are matched to them to form an underwater acoustic target recognition dataset D; The underwater acoustic target recognition dataset D is divided into training, validation, and test sets in an 8:1:1 ratio.

3. The underwater acoustic target recognition method based on general acoustic knowledge transfer from a large speech model according to claim 2, characterized in that, The SLM model is broken down according to its component structure, specifically including: The SLM model structure is disassembled, separating the acoustic encoder and decoder; The hierarchical stacking relationship of the acoustic encoder remains unchanged, and it is adapted to the feature extractor of the SLM model to ensure that the input feature format is consistent with that of the pre-training stage.

4. The underwater acoustic target recognition method based on general acoustic knowledge transfer from a large speech model according to claim 3, characterized in that, The lightweight classifier is a linear classifier, specifically including: Determine the dimension N of the acoustic encoder output vector and the number of target categories M based on the number of categories in dataset D; Initialize the parameters of a linear layer with shape (N, M), assign weights randomly, and use it as a classifier; After connecting the linear classifier to the acoustic encoder, the high-dimensional feature vector output by the encoder is received and mapped to the prediction result of the target category.

5. The underwater acoustic target recognition method based on general acoustic knowledge transfer from a large speech model according to claim 4, characterized in that, The large speech classification model in D, which is trained on the input set, undergoes end-to-end fine-tuning. The loss is calculated using the cross-entropy loss function to assist backpropagation in updating the model parameters. Specifically: Set hyperparameters for model training, including learning rate, batch size, and number of training epochs; Based on the batch size, the audio samples in the training set are processed in batches, and then these data are input into the large speech classification model for training. Forward propagation is performed on the model to predict audio categories, the loss value is calculated using the cross-entropy loss function, and the parameters of the large speech classification model are updated using the backpropagation algorithm. The backpropagation process is as follows: in, The logit value represents the j-th class, used to calculate the predicted probability of that class; These represent model parameters, which are continuously updated during model training to minimize the loss. This represents the learning rate, which controls the step size for each parameter update. Let represent the predicted probability of class i, which is obtained by normalizing logit using the softmax function, and c represents the total number of classes in the classification task. Cross-entropy loss is used to measure the difference between the predicted probability distribution and the true label distribution. This indicates that the loss function is applied to the parameters. The gradient indicates the direction of parameter updates; After each round of parameter updates, the adjusted model is used to calculate the current validation loss using the validation set. After all training is completed, the model with the smallest loss on the validation set is retained as the final underwater acoustic target recognition speech model UATR-SLM.

6. The underwater acoustic target recognition method based on general acoustic knowledge transfer from a large speech model according to claim 5, characterized in that, The process of detecting underwater acoustic targets and outputting predicted category results specifically includes: Extract the underwater acoustic audio files from the test set in dataset D, feed the audio files into the UATR-SLM model, and start the model prediction. The model processes underwater acoustic audio files and outputs confidence probabilities for different categories in the final linear layer. The category with the highest probability is used as the prediction result for underwater acoustic target detection.

7. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method of claim 6.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 6.

9. A computer program product, comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 6.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in claim 6.

Images