A deep-sea reverberation weak signal recognition method based on a deep learning model

Abstract

This invention relates to the field of deep-sea signal processing and recognition technology, and in particular to a method for recognizing weak reverberation signals in the deep sea based on a deep learning model. The method includes: collecting reverberation data containing weak signals using underwater sensors; removing sensor noise and high-frequency interference from the reverberation data using a wavelet threshold denoising algorithm; normalizing the reverberation data to map its signal amplitude to a preset range; constructing a convolutional neural network model containing convolutional layers, pooling layers, and fully connected layers; dividing the normalized reverberation data into training, validation, and test sets; inputting the weak reverberation signal to be identified into the trained convolutional neural network model after denoising and normalization; and outputting the signal and recognition result from the convolutional neural network model. This invention eliminates a large amount of tedious manual feature engineering, reduces human intervention, and improves recognition efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of deep-sea signal processing and recognition technology, specifically to a method for recognizing weak reverberation signals in the deep sea based on a deep learning model. Background Technology

[0002] In the deep-sea environment, received signals are constantly interfered with by ocean reverberation, making the identification of weak signals extremely difficult. Traditional signal identification methods have significant shortcomings in dealing with such complex situations. For example, traditional transform domain analysis methods such as Fourier transform cannot effectively capture the detailed features of weak signals against a complex reverberant background, making it difficult to accurately extract the target signal from a large amount of interference. Classification algorithms based on manually designed features, such as support vector machines, heavily rely on prior knowledge and tedious manual feature engineering. In the deep-sea environment, with its variable environmental parameters and complex signal characteristics, manually designed features cannot comprehensively cover all situations, resulting in poor algorithm adaptability and low identification accuracy and efficiency. With the continuous deepening of marine development and research, the need for accurate identification of weak signals in deep-sea reverberation is becoming increasingly urgent, requiring an innovative method to overcome the current technological bottlenecks. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a method for identifying weak deep-sea reverberation signals based on a deep learning model.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying weak deep-sea reverberation signals based on a deep learning model, comprising the following steps: Step 1: Using underwater sensors, collect reverberation data containing weak signals at specific time intervals and sampling frequencies in different deep-sea areas and sea conditions; Step 2: Use wavelet threshold denoising algorithm to remove sensor noise and high-frequency interference from the reverberation data, and then normalize the reverberation data after removing sensor noise and high-frequency interference to map its signal amplitude to the [0,1] interval; Step 3: Build a convolutional neural network model containing three sets of convolutional layers, pooling layers, and fully connected layers; Step 4: Divide the normalized reverberation data from Step 2 into training, validation, and test sets according to the proportions. Based on the data in the training, validation, and test sets, use the cross-entropy loss function as the optimization objective and the Adam optimizer to dynamically adjust the learning rate to train the convolutional neural network model. Step 5: After denoising and normalizing the weak deep-sea reverberation signal to be identified, input it into the trained convolutional neural network model. The convolutional neural network model outputs the signal and identifies the result, determining whether it is the target weak signal and its signal category.

[0005] In some embodiments, the sampling frequency in step one is set between 50kHz and 500kHz, the sampling duration is 8 to 15 seconds, and multiple samplings are performed under at least three different combinations of environmental parameters such as water depth, water temperature, and salinity.

[0006] In some embodiments, in the wavelet thresholding denoising algorithm in step two, the db3-db5 wavelet basis is selected, the number of decomposition layers is set to 4-6, and a threshold is set based on the minimax criterion to remove wavelet coefficients smaller than the threshold.

[0007] In some embodiments, after normalizing the weak deep-sea reverberation signal to be identified in step five, the method further includes performing data augmentation operations on the weak deep-sea reverberation signal to be identified. The data augmentation operations include, but are not limited to, time axis translation, amplitude scaling, and adding Gaussian white noise, or one or more combinations thereof.

[0008] In some embodiments, the convolutional layer in step three extracts local features of the normalized reverberation data using convolutional kernels of different sizes; the pooling layer downsamples the local features extracted by the convolutional layer and outputs a feature map of the local features; and the fully connected layer converts the feature map output by the pooling layer into a one-dimensional vector.

[0009] In some embodiments, in the three sets of convolutional layers in step three, the first set of convolutional layers uses 32-64 convolutional kernels of size 3×3, the second set of convolutional layers uses 64-128 convolutional kernels of size 5×5, and the third set of convolutional layers uses 128-256 convolutional kernels of size 5×5, and a ReLU activation function is added after each convolutional layer; the pooling layer uses 2×2 max pooling operation.

[0010] In some embodiments, the training set in step four is used to train the convolutional neural network model; the validation set is used to monitor the training process of the convolutional neural network model, evaluate the performance of the convolutional neural network model, and prevent overfitting; and the test set is used to evaluate the generalization ability of the trained convolutional neural network model on unknown data.

[0011] In some embodiments, when dividing the normalized reverberation data in step four, the ratio of training set, validation set, and test set is 65%–75%, 10%–20%, and 10%–20%, respectively.

[0012] In some embodiments, during step four of training the convolutional neural network model, when the model's accuracy on the validation set no longer improves for 3-7 consecutive rounds, the most recently saved optimal model parameters are loaded, and training is stopped.

[0013] Compared with the prior art, the beneficial effects of the present invention are: I. This invention utilizes the automatic feature extraction capability of deep learning models, eliminating a large amount of tedious manual feature engineering, reducing manual intervention, and improving recognition efficiency and accuracy. Second, CNN models have powerful feature extraction capabilities, enabling them to effectively capture the complex features of weak deep-sea reverberation signals and have better adaptability in complex and ever-changing deep-sea environments. Third, by adopting transfer learning technology, the model training time is shortened, the dependence on a large number of samples is reduced, and the generalization ability of the model is improved, enabling the model to maintain good recognition performance in different deep-sea scenarios.

[0014] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. The embodiments of this application will provide a detailed description and understanding of the application. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method 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. Example 1

[0017] Please see Figure 1 This invention provides a technical solution: a method for identifying weak reverberation signals in deep sea based on a deep learning model, comprising the following steps: Step 1: Using underwater sensors, collect reverberation data containing weak signals at specific time intervals and sampling frequencies in different deep-sea areas and sea conditions; Step 2: A wavelet thresholding denoising algorithm is used to remove sensor noise and high-frequency interference from the reverberation data. A db3-db5 wavelet basis can be selected, as these basis bases exhibit good time-frequency localization characteristics when processing non-stationary deep-sea reverberation signals. The signal is decomposed into 4-6 layers. For each sub-band, a threshold is set based on the different characteristics of noise and signal in the wavelet coefficients. The threshold is set using a minimax criterion. Noise wavelet coefficients are typically small and uniformly distributed, while signal wavelet coefficients are relatively large and concentrated in specific locations. By setting wavelet coefficients below the threshold to zero and retaining coefficients above the threshold, an inverse wavelet transform is performed to effectively remove sensor noise and high-frequency interference from the environment, resulting in a relatively clean signal. The reverberation data after removing sensor noise and high-frequency interference is then normalized. For each signal sample... The normalization formula is: ,in and These are the minimum and maximum values ​​of the signal sample in the entire dataset, respectively. This operation maps all signal amplitudes to the [0,1] interval, ensuring that signals with different amplitudes have the same scale in subsequent model processing, thus avoiding adverse effects of amplitude differences on model training and recognition results. Furthermore, after data preprocessing, data augmentation operations are performed, including one or more combinations of time axis shifting, amplitude scaling, and adding Gaussian white noise, to further expand data diversity and improve model robustness. Step 3: Construct a convolutional neural network model containing three sets of convolutional layers, pooling layers, and fully connected layers. The first convolutional layer uses 32-64 kernels of size 3×3 or 5×5. These kernels slide across the input signal to extract local features. The parameters of the kernels are continuously adjusted during training to learn the most effective feature patterns for identifying weak signals. The second convolutional layer uses 64-128 kernels of size 3×3 or 5×5 to further extract more advanced and complex features. The third convolutional layer uses 128-256 kernels of size 3×3 or 5×5 to deeply extract key hidden feature information from the signal. A ReLU activation function is added after each convolutional layer, with the formula: The ReLU function enhances the nonlinear expressiveness of the model, enabling it to learn more complex functional relationships. The pooling layers employ 2×2 or 3×3 max pooling operations, placed between the convolutional layers. Max pooling selects the maximum value within the corresponding neighborhood as the output, thus preserving key features while downsampling the convolutional layer output, reducing data volume, effectively lowering computational complexity, and mitigating the risk of overfitting. Step 4: Divide the normalized reverberation data from Step 2 into training, validation, and test sets proportionally. Based on the data in these sets, use the cross-entropy loss function as the optimization objective. For multi-class classification problems, the formula for the cross-entropy loss function is: , where is the real label (if it is the first (If the value is 1, then it is 0; otherwise, it is 1), indicating that the model predicts the [number]th [class]. The probability of a class is determined by the number of samples. The Adam optimizer dynamically adjusts the learning rate to train the convolutional neural network model. The Adam optimizer combines the advantages of Adagrad and Adadelta, adaptively adjusting the learning rate. The initial learning rate is set to 0.0005-0.0015. As training progresses, the learning rate is decayed to 0.7-0.9 times its original value every 8-12 training epochs to balance the model's convergence speed and accuracy. Through continuous iterative training, the model parameters are adjusted to continuously reduce the model's loss value on the training set. Simultaneously, performance metrics such as accuracy and recall are monitored on the validation set to prevent overfitting. Step 5: After denoising and normalizing the weak deep-sea reverberation signal to be identified, input it into the trained convolutional neural network model. The convolutional neural network model outputs the signal and identifies the result, determining whether it is the target weak signal and its signal category.

[0018] In step one, the sampling frequency is set between 5kHz and 50kHz, the sampling duration is 8-15 seconds, and multiple samplings are performed under at least three different combinations of environmental parameters such as water depth, water temperature, and salinity.

[0019] Specifically, specialized underwater sensors, such as high-sensitivity hydrophones, were used to collect data in various deep-sea areas. These areas covered different environmental conditions, including varying depths, temperatures, and salinities, while also considering different sea states, such as calm seas, moderate waves, and areas under severe storm conditions. The sensor's sampling frequency f was set between 5kHz and 50kHz, and data was collected at fixed time intervals to ensure the complete capture of weak signals and their time-varying characteristics. The duration T of each data collection was set to 8–15 seconds to ensure rich signal information. To enhance data diversity and the model's generalization ability, multiple data collections were conducted under at least three different combinations of environmental parameters, including depth, temperature, and salinity, accumulating a large number of reverberation data samples containing weak signals.

[0020] The specific implementation of the above technical solution is as follows: Data Acquisition: Utilizing specialized underwater sensors, such as high-sensitivity hydrophones, data was collected in various deep-sea areas. These areas encompassed diverse environmental conditions including water depth, temperature, and salinity, while also considering different sea states, including calm waters, moderate waves, and areas experiencing severe storms. The sensor sampling frequency was set between 50kHz and 500kHz, with data collected at fixed time intervals to ensure the complete capture of weak signals and their time-varying characteristics. The duration T of each data acquisition was set to 8–15 seconds to guarantee the inclusion of rich signal information. To enhance data diversity and the model's generalization ability, multiple acquisitions were conducted under at least three different combinations of environmental parameters such as water depth, temperature, and salinity, accumulating a large number of reverberation data samples containing weak signals. Denoising: A wavelet thresholding algorithm is used to purify the original data. A db3-db5 wavelet basis can be selected, as these basis bases exhibit good time-frequency localization characteristics when processing non-stationary deep-sea reverberant signals. The signal is decomposed into 4-6 layers. In each sub-band, a threshold is set based on the different characteristics of noise and signal in the wavelet coefficients. The threshold is set using a minimax criterion. Noise wavelet coefficients are typically small and uniformly distributed, while signal wavelet coefficients are relatively large and concentrated in specific locations. By setting wavelet coefficients smaller than the threshold to zero and retaining coefficients larger than the threshold, an inverse wavelet transform is performed to effectively remove sensor noise and high-frequency interference from the environment, resulting in a relatively clean signal. Normalization operation: The amplitude of the denoised signal is normalized using the min-max normalization method. For each signal sample... The normalization formula is: ,in and These are the minimum and maximum values ​​of the signal sample in the entire dataset, respectively. This operation maps all signal amplitudes to the [0,1] interval, ensuring that signals with different amplitudes have the same scale in subsequent model processing, thus avoiding adverse effects of amplitude differences on model training and recognition results. Furthermore, after data preprocessing, data augmentation operations are performed, including one or more combinations of time axis shifting, amplitude scaling, and adding Gaussian white noise, to further expand the diversity of the data and improve the robustness of the model. Model Architecture: A model based on a Convolutional Neural Network (CNN) is constructed. This model consists of multiple convolutional layers, pooling layers, and fully connected layers. The first convolutional layer uses 32-64 convolutional kernels of size 3×3 or 5×5. These kernels slide across the input signal to extract local features. The parameters of the convolutional kernels are continuously adjusted during training to learn the most effective feature patterns for identifying weak signals. The second convolutional layer uses 64-128 convolutional kernels of size 3×3 or 5×5 to further extract more advanced and complex features. The third convolutional layer uses 128-256 convolutional kernels of size 3×3 or 5×5 to deeply extract key hidden feature information from the signal. A ReLU activation function is added after each convolutional layer, with the formula: The ReLU function enhances the nonlinear expressiveness of the model, enabling it to learn more complex functional relationships. The pooling layers employ 2×2 or 3×3 max pooling operations, placed between the convolutional layers. Max pooling selects the maximum value within the corresponding neighborhood as the output, thus preserving key features while downsampling the convolutional layer output, reducing data volume, effectively lowering computational complexity, and mitigating the risk of overfitting. Model Training: During model training, the cross-entropy loss function is used as the optimization objective. For multi-class classification problems, the formula for the cross-entropy loss function is: ,in Is it a real label (if it is the first) (If the value is 1, then it is 0; otherwise, it is 1), indicating that the model predicts the [number]th [class]. The probability of a class This refers to the sample size. The Adam optimizer is used to update the model parameters. The Adam optimizer combines the advantages of Adagrad and Adadelta, adaptively adjusting the learning rate. The initial learning rate is set to 0.0005-0.0015. As training progresses, the learning rate is decayed to 0.7-0.9 times its original value every 8-12 training epochs to balance the model's convergence speed and accuracy. Through continuous iterative training and adjustment of model parameters, the model's loss value on the training set is continuously reduced. Simultaneously, performance metrics such as accuracy and recall are monitored on the validation set to prevent overfitting. Dataset partitioning: The preprocessed data is divided into training, validation, and test sets in proportions of 65%–75%, 10%–20%, and 10%–20%, respectively. The training set is used for training the model's parameters. The validation set is used to monitor the model's training process, evaluate its performance, and prevent overfitting. When the model's performance (e.g., accuracy) on the validation set no longer improves, the model is considered to have reached a good training state, and training is stopped. The test set is used to evaluate the generalization ability of the finally trained model on unknown data. During model training, model parameters are saved periodically. When the model's accuracy on the validation set no longer improves for 3-7 consecutive rounds, the most recently saved optimal model parameters are loaded, and training is stopped.

[0021] Transfer learning: Pre-trained model parameters from other related fields (such as audio signal processing) are used to initialize some layers of this model. The parameters of the first 2-4 convolutional layers in the pre-trained model are directly transferred to this model. These pre-trained parameters have learned some general signal features, which can help the model converge faster and reduce training time and sample requirements. After parameter transfer, the entire model undergoes at least 10 rounds of fine-tuning training to adapt it to the task of identifying weak deep-sea reverberation signals. Further training on the training set, adjusting the transferred parameters and the parameters of newly added layers, further optimizes model performance and improves its generalization ability on the task of identifying weak deep-sea reverberation signals. Signal Recognition and Result Output: The weak deep-sea reverberation signal to be identified is processed using the same preprocessing steps as the training data, including denoising, normalization, and data augmentation. The processed signal is then input into a trained deep learning model. Based on the feature patterns learned during training, the model performs layer-by-layer calculations and analysis on the input signal. First, features are extracted through convolutional and pooling layers, then classification is performed through fully connected layers, finally outputting the signal recognition result. The result is presented in the form of a probability distribution; that is, the model outputs the probability value of the signal belonging to each category, and the category with the highest probability value is selected as the final recognition result, determining whether the signal is the target weak signal and its corresponding signal category. Example 2

[0022] This embodiment uses a practical application example to illustrate the implementation process of the method of the present invention, including: Implementation Environment and Preparation This example was conducted on a computer equipped with an NVIDIA GPU (such as a GeForce RTX 3090), an Intel Core i9 processor, and 32GB of RAM. The operating system was Ubuntu 20.04 LTS, and the deep learning framework used was TensorFlow 2.5.

[0023] Data Acquisition and Preprocessing Data Acquisition: High-sensitivity hydrophones were used as underwater sensors to collect data in multiple deep-sea areas of the Pacific, Atlantic, and Indian Oceans. In each area, different combinations of water depth (e.g., 1000m, 2000m, 3000m), water temperature (e.g., 1℃, 2℃, 3℃), and salinity (e.g., 34‰, 35‰, 36‰) were selected, covering sea conditions including calm surfaces, moderate waves, and severe storms. The sampling frequency was set to 200kHz, with data collected at 0.1-second intervals for 10-second periods, resulting in 2000 sets of reverberation data containing weak signals.

[0024] Denoising: A wavelet thresholding denoising algorithm is used, selecting the db4 wavelet basis to decompose the signal into 5 layers. A threshold is set based on the minimax criterion, and wavelet coefficients smaller than the threshold are removed, effectively removing sensor noise and high-frequency environmental interference to obtain a clean signal.

[0025] Normalization and Data Augmentation: The min-max normalization method is used to map the amplitude of the denoised signal to the [0,1] interval. Data augmentation operations are then performed, including time axis shifting (random shifting within the range of -0.05 seconds to 0.05 seconds), amplitude scaling (random scaling between 0.8 and 1.2 times), and adding Gaussian white noise (mean 0, standard deviation 0.05) to further expand data diversity and improve model robustness.

[0026] Model building and training Model Architecture: A convolutional neural network model consisting of three sets of convolutional layers, pooling layers, and fully connected layers is constructed. The first set of convolutional layers uses 48 3×3 convolutional kernels, the second set uses 96 5×5 convolutional kernels, and the third set uses 192 5×5 convolutional kernels, with a ReLU activation function added after each convolutional layer. The pooling layers use 2×2 max pooling. In the fully connected layers, the first fully connected layer contains 128 neurons, and the second fully connected layer sets 8 neurons according to the actual signal classification (assuming 8 classes), and connects to a Softmax layer to convert the output into class probabilities.

[0027] Model Training: The preprocessed data was divided into training, validation, and test sets in a 70%, 15%, and 15% ratio, respectively. The cross-entropy loss function was used as the optimization objective, employing the Adam optimizer with an initial learning rate of 0.001. Every 10 training epochs, the learning rate was decayed to 0.8 times its original value. During training, iterative training was continuously performed, adjusting model parameters to continuously reduce the loss value on the training set. Simultaneously, performance metrics such as accuracy and recall were monitored on the validation set to prevent overfitting. When the model's accuracy on the validation set no longer improved after 5 consecutive epochs, the most recently saved optimal model parameters were loaded, and training was stopped. Furthermore, using pre-trained model parameters from the audio signal processing domain, the first three convolutional layers of this model were initialized. After parameter transfer, the entire model underwent 15 epochs of fine-tuning training to further optimize model performance.

[0028] Signal recognition and result evaluation Signal Recognition: New weak deep-sea reverberation signals to be identified are acquired and processed using the same preprocessing steps as the training data, including denoising, normalization, and data augmentation. The processed signals are then input into a trained convolutional neural network model. The model calculates and analyzes the input signals layer by layer, extracting features through convolutional and pooling layers, and then classifying them through fully connected layers. The final output signal recognition result is presented in the form of a probability distribution. The category with the highest probability value is selected as the final recognition result to determine whether the signal is the target weak signal and its corresponding signal category.

[0029] Results Evaluation: The trained model was evaluated using a test set. After multiple tests, the model in this embodiment achieved a 93% accuracy rate in recognizing weak deep-sea reverberant signals, a significant improvement over traditional methods, demonstrating the effectiveness and superiority of this method in practical applications. Furthermore, analysis of signal recognition results under different sea states and environmental parameters revealed that the model maintains high stability and accuracy under various conditions, further validating its generalization ability.

[0030] The above embodiments demonstrate in detail the specific operation process and practical effects of the deep-sea reverberation weak signal identification method based on a deep learning model, providing strong practical evidence for the application of this invention. In practical applications, the model parameters and data acquisition and processing procedures can be appropriately adjusted according to the specific deep-sea environment and signal characteristics to achieve the best identification effect.

[0031] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying weak reverberation signals in deep sea based on a deep learning model, characterized in that: Includes the following steps: Step 1: Using underwater sensors, collect reverberation data containing weak signals at specific time intervals and sampling frequencies in different deep-sea areas and sea conditions; Step 2: Use wavelet threshold denoising algorithm to remove sensor noise and high-frequency interference from the reverberation data, and then normalize the reverberation data after removing sensor noise and high-frequency interference to map its signal amplitude to the [0,1] interval; Step 3: Build a convolutional neural network model containing three sets of convolutional layers, pooling layers, and fully connected layers; Step 4: Divide the normalized reverberation data from Step 2 into training, validation, and test sets according to the proportions. Based on the data in the training, validation, and test sets, use the cross-entropy loss function as the optimization objective and the Adam optimizer to dynamically adjust the learning rate to train the convolutional neural network model. Step 5: After denoising and normalizing the weak deep-sea reverberation signal to be identified, input it into the trained convolutional neural network model. The convolutional neural network model outputs the signal and identifies the result, determining whether it is the target weak signal and its signal category.

2. The method for identifying weak deep-sea reverberation signals based on a deep learning model according to claim 1, characterized in that: In step one, the sampling frequency is set between 50kHz and 500kHz, the sampling duration is 8 to 15 seconds, and multiple samplings are performed under at least three different combinations of environmental parameters such as water depth, water temperature, and salinity.

3. The method for identifying weak deep-sea reverberation signals based on a deep learning model according to claim 1, characterized in that: In the wavelet threshold denoising algorithm in step two, the db3-db5 wavelet basis is selected, the number of decomposition layers is set to 4-6, and the threshold is set based on the minimax criterion to remove wavelet coefficients smaller than the threshold.

4. The method for identifying weak deep-sea reverberation signals based on a deep learning model according to claim 1, characterized in that: In step five, after normalizing the identified deep-sea reverberation weak signal, a data augmentation operation is also performed on the identified deep-sea reverberation weak signal. The data augmentation operation includes, but is not limited to, one or more combinations of time axis translation, amplitude scaling, and adding Gaussian white noise.

5. The method for identifying weak deep-sea reverberation signals based on a deep learning model according to claim 1, characterized in that: In step three, the convolutional layer extracts local features of the normalized reverberation data using convolutional kernels of different sizes; the pooling layer downsamples the local features extracted by the convolutional layer and outputs feature maps of the local features; and the fully connected layer converts the feature maps output by the pooling layer into one-dimensional vectors.

6. The method for identifying weak deep-sea reverberation signals based on a deep learning model according to claim 1, characterized in that: In step three, the three sets of convolutional layers use 32-64 3×3 kernels in the first set, 64-128 5×5 kernels in the second set, and 128-256 5×5 kernels in the third set. A ReLU activation function is added after each convolutional layer. The pooling layer uses 2×2 max pooling.

7. The method for identifying weak deep-sea reverberation signals based on a deep learning model according to claim 1, characterized in that: The training set in step four is used to train the convolutional neural network model; the validation set is used to monitor the training process of the convolutional neural network model, evaluate the performance of the convolutional neural network model, and prevent overfitting; the test set is used to evaluate the generalization ability of the trained convolutional neural network model on unknown data.

8. The method for identifying weak deep-sea reverberation signals based on a deep learning model according to claim 1, characterized in that: In step four, when dividing the normalized reverberation data, the proportions of the training set, validation set, and test set are 65%–75%, 10%–20%, and 10%–20%, respectively.

9. A method for identifying weak deep-sea reverberation signals based on a deep learning model according to claim 1, characterized in that: In step four, during the training of the convolutional neural network model, when the model's accuracy on the validation set no longer improves for 3-7 consecutive rounds, the most recently saved optimal model parameters are loaded, and training is stopped.

Images