A method for predicting ocean ambient noise based on a deep neural network

Abstract

In order to solve the technical problems that the existing marine environmental noise prediction method based on deep neural network cannot meet the requirements of large-scale data fast modeling, cannot accurately capture the long-range evolution law of noise sequence and cannot meet the real-time prediction demand, the present application provides a marine environmental noise prediction method based on CTTM network.The present application adopts the noise prediction model constructed and trained to predict the marine environmental noise, the noise prediction model is constructed according to the specific data flow sequence of 'local short-term change feature extraction-global long-term dependence modeling-cross-scale time sequence structure enhancement-key time step weighted aggregation-noise spectrum level output', and the prediction task can be completed under the condition of only relying on current and historical time window meteorological data, so that the requirements of large-scale data fast modeling, long-range evolution law capturing and real-time prediction demand can be met.

Description

Technical Field

[0001] This invention relates to a method for predicting marine environmental noise, and more particularly to a method for predicting marine environmental noise based on deep neural networks. Background Technology

[0002] Marine environmental noise prediction is a crucial part of marine scientific research, playing a vital role in marine noise monitoring, acoustic environment assessment, maritime navigation safety, and early warning and emergency response to marine disasters. In recent years, with the rapid development of deep learning technology, the application of deep neural networks in marine environmental noise prediction has increased significantly. Among existing marine environmental noise prediction methods, LSTM, GRU, and their bidirectional variants (BiLSTM, BiGRU) are widely used due to their ability to capture time-series dependencies. However, these models still have the following significant shortcomings in practical applications:

[0003] First, LSTM and GRU use a serial recursive computational structure, which makes it difficult to achieve efficient parallel training. This results in slow training speed when processing high sampling rate and long time-series ocean noise data, and cannot meet the needs of rapid modeling of large-scale data.

[0004] Second, although LSTM and GRU alleviate the gradient vanishing problem through gating mechanisms, their ability to model dependencies over extremely long time spans and multi-scale periodic changes (such as tides, intermittent ship passage, and sudden changes in wind and waves) remains limited, making it difficult to accurately capture long-range evolution patterns in noise sequences.

[0005] Third, LSTM and GRU are weak at capturing abrupt changes in high-frequency noise and perform poorly on strong non-stationary signals common in marine environments (such as rainstorm noise and thruster transient noise). Although bidirectional variants of LSTM and GRU can enhance feature capture capabilities by introducing future information, they significantly increase computational load and output latency, making it difficult to meet the application requirements of real-time prediction. Summary of the Invention

[0006] To address the technical challenges of existing deep neural network-based marine environmental noise prediction methods, such as their inability to rapidly model large-scale data, accurately capture the long-range evolution of noise sequences, and meet real-time prediction requirements, this invention proposes a marine environmental noise prediction method based on a CTTM network. The noise prediction model in this method processes data according to the following flow sequence: "Meteorological feature sequence X—Short-term variation feature C—Global dependency feature T—Cross-scale temporal feature D—Global aggregation feature Z—Noise spectral level prediction value Y". It should be noted that the modules in this invention are processed sequentially according to the aforementioned data flow sequence. Unlike recurrent neural networks such as LSTM and GRU, the temporal feature processing module in this invention does not need to recursively update the hidden state step-by-step in chronological order. Instead, it processes features within the time window through one-dimensional convolution, multi-head self-attention, dilated convolution, and attention-weighted aggregation, thereby reducing the recursive computation dependency in the time dimension.

[0007] Specifically, the CNN local feature extraction module extracts short-term variation features within the local temporal neighborhood through one-dimensional convolution and max pooling, reducing the temporal dimension of subsequent processing and helping to improve sampling rate and modeling efficiency under long-term data. The Transformer encoding module establishes global dependencies between different time steps within the time window based on the short-term variation features C. The TCN sequence convolution module expands the temporal receptive field and extracts cross-scale temporal features through dilated convolution, enabling the model to simultaneously represent local short-term disturbances, long-term dependencies, and cross-scale periodic changes, thereby improving the ability to capture the long-term evolution and non-stationary changes of marine environmental noise. The multi-head attention pooling module assigns contribution weights and performs weighted aggregation of hidden features at different time steps, so that time steps that contribute more to the prediction results have a higher proportion in the global aggregated features. The fully connected layer output module finally maps the global aggregated features to the noise spectral level prediction value corresponding to the target frequency. Therefore, this invention can complete the prediction task by relying only on current and historical meteorological data, avoiding the dependence of bidirectional recursive structures such as BiLSTM and BiGRU on future time step information and the resulting output delay, thus taking into account the needs of rapid modeling of large-scale data, long-range evolution law capture and real-time prediction.

[0008] The technical solution adopted in this invention is as follows:

[0009] A method for predicting marine environmental noise based on deep neural networks, characterized by the following steps:

[0010] Real-time collection of four types of meteorological data in the target area: rainfall, sea surface pressure, sea surface temperature, and wind speed;

[0011] The four types of meteorological data collected were processed by interpolation of missing values, moving average and standardization, and assembled into meteorological feature sequences according to the time window set in the model training stage, and then input into the trained noise prediction model.

[0012] The noise prediction model outputs the predicted value of the marine environmental noise spectrum level in the target area at the target frequency or frequency band.

[0013] The noise prediction model is a CTTM network-based model, comprising a CNN local feature extraction module, a Transformer encoding module, a TCN sequence convolution module, a multi-head attention pooling module, and a fully connected layer output module. The CNN local feature extraction module extracts short-term variation features in the local temporal neighborhood from meteorological feature sequences. The Transformer encoding module models the correlation between different time steps in the short-term variation features using a multi-head self-attention mechanism, obtaining and outputting global dependency features. The TCN sequence convolution module captures and outputs cross-scale temporal features from the global dependency features using multiple serially stacked dilated convolutional blocks with different dilation rates. The multi-head attention pooling module assigns attention weights and performs weighted aggregation on features at different time steps in the cross-scale temporal features using multiple parallel pooling attention heads, and concatenates the weighted aggregation results of all pooling attention heads along the feature dimension to output the global aggregated features. The fully connected layer output module maps the global aggregated features to obtain and output the predicted noise spectral level value corresponding to the target frequency.

[0014] Furthermore, the CNN local feature extraction module includes multiple parallel convolutional modules and a concatenation module; each convolutional module includes a serially connected one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer; the one-dimensional convolutional layer is used to extract local change information of meteorological variables within adjacent time steps from the meteorological feature sequence; the batch normalization layer normalizes the local change information; the ReLU activation layer performs nonlinear mapping on the normalized features; the max pooling layer downsamples the ReLU activated features in the time dimension to obtain short-term change features; the one-dimensional convolutional layers in different convolutional modules have different scales; the concatenation module is used to concatenate the short-term change features output by each convolutional module along the feature dimension to obtain the concatenated short-term change features, which is the output of the entire CNN local feature extraction module.

[0015] Furthermore, the CNN local feature extraction module includes a convolutional module; a single convolutional module includes a serially connected one-dimensional convolutional layer, a batchNormalization normalization layer, a ReLU activation layer, and a max pooling layer; the one-dimensional convolutional layer is used to extract local change information of meteorological variables within adjacent time steps from the meteorological feature sequence; the batchNormalization normalization layer normalizes the local change information; the ReLU activation layer performs nonlinear mapping on the normalized features; the max pooling layer downsamples the ReLU activated features in the time dimension to obtain short-term change features, which are the output of the entire CNN local feature extraction module.

[0016] Furthermore, the Transformer encoding module includes a multi-head self-attention sub-layer, a first residual connection layer, a first normalization layer, a feedforward network sub-layer, a second residual connection layer, and a second normalization layer;

[0017] Multi-head self-attention sublayers are used to extract global attention features between different time steps in short-term change features C. ;

[0018] The first residual connection layer is used to connect the input features of the multi-head self-attention sub-layer with the global attention features output by the multi-head self-attention sub-layer. The fusion is performed to obtain the fusion feature C+. ;

[0019] The first normalization layer affects the fused features C+. Normalization is performed to obtain feature U;

[0020] The feedforward sublayer is used to perform a nonlinear mapping on the feature U output by the first normalization layer to obtain the feature. ;

[0021] The second residual connection layer will nonlinearly mapped features The features are fused with the input of the feedforward network sublayer to obtain the features. ;

[0022] The second normalization layer is used to process the features. Normalization is performed to obtain global dependency features.

[0023] Furthermore, the multi-head self-attention quantum layer is composed of It consists of a parallel attention head, a second splicing module, and a linear transformation layer;

[0024] Each attention head performs the following operations independently:

[0025] The feature representations of the short-term change features at each time step in time window L1 are respectively mapped to query vector Query, key vector Key, and value vector Value;

[0026] At time step t, calculate the correlation between its query vector Query and the key vector Key corresponding to all L1 time steps within time window L1, and obtain the attention weight of time step t relative to all L1 time steps.

[0027] The attention features extracted by the attention head at time step t are obtained by weighting and summing the value vectors Value corresponding to all L1 time steps based on the attention weights. h is the number of the attention head;

[0028] The time window L1 is determined by the time window length L of the meteorological feature sequence X and the pooling window size of the max pooling layer in the CNN local feature extraction module, L1=L / p, where p is the pooling window size of the max pooling layer;

[0029] The second stitching module, at each time step of time window L1, will... Attention features obtained from each attention head splicing;

[0030] The linear transformation layer performs a linear transformation on the concatenated features to obtain the global attention features at each time step of time window L1. , =1,2,…,L1.

[0031] Furthermore, the TCN module extracts cross-scale temporal features from global dependency features through N serially stacked dilated convolutional blocks with progressively increasing dilation rates; a single dilated convolutional block includes a serially connected dilated one-dimensional convolutional layer, a ReLU activation function, a Dropout layer, and a residual connection layer.

[0032] Furthermore, the structure of the multi-head attention pooling module includes multiple pooling attention heads set in parallel and a feature concatenation unit;

[0033] Each pooling attention head performs the following operations independently:

[0034] Attention scores are calculated for the hidden feature vectors at each time step in the cross-scale time series features. The attention scores are then converted into contribution weights of the pooled attention head at different time steps. Based on the calculated contribution weights, the hidden feature vectors at each time step in the cross-scale time series features are weighted and summed, and the weighted aggregation result is output.

[0035] The feature concatenation unit concatenates the weighted aggregation results corresponding to all pooling attention heads to obtain the global aggregated features.

[0036] Furthermore, the training method for the noise prediction model based on the CTTM network is as follows:

[0037] Step 1: Collect data;

[0038] Four types of meteorological data were acquired: rainfall, sea surface pressure, sea surface temperature, and wind speed. At the same time, marine environmental noise data corresponding to the time of the above four types of meteorological data were collected. The marine environmental noise data were provided by giving the noise spectrum level in each pre-divided frequency band.

[0039] Step 2: Preprocessing meteorological data;

[0040] Missing values ​​in the four types of meteorological data were filled using linear interpolation.

[0041] The four types of meteorological data were processed by moving average to filter out short-term fluctuations, resulting in smoothed meteorological data.

[0042] The smoothed meteorological data is standardized to convert it to a uniform scale;

[0043] Step 3: Time-align meteorological data with marine environmental noise data;

[0044] The four types of meteorological data after standardization were time-aligned with the marine environmental noise data according to the same timestamp, based on a unified sampling interval, to ensure that the meteorological data in each sample is consistent with the target noise spectrum level in time.

[0045] Step 4: Construct the dataset;

[0046] Set a fixed-length time window and expand the aligned meteorological data and ocean noise data with the same time window.

[0047] Feature vectors are constructed using four types of meteorological data within each time window;

[0048] The meteorological feature vector and its corresponding target noise spectral level are combined into a set of sample pairs, and all sample pairs are used to construct the dataset.

[0049] Step 5: Divide the dataset into training and testing sets;

[0050] The dataset is divided into a training set and a test set according to a set ratio;

[0051] Step 6: Train the model;

[0052] The noise prediction model based on CTTM was trained using the training set;

[0053] Step 7, Model Performance Evaluation;

[0054] The prediction performance of the CTTM-based noise prediction model is evaluated using a test set. If the prediction performance does not meet the preset requirements, return to step 6 to continue training or adjust the model parameters until the prediction performance meets the requirements.

[0055] Furthermore, the model training employs the Adam optimizer and the Huber loss function, combined with learning rate scheduling, early stopping, and learning rate decay strategies to ensure the robustness and generalization ability of the noisy prediction model. The learning rate adopts a strategy combining warm-up and cosine decay, linearly increasing in the first 5 training rounds to ensure stable convergence of the model, and then gradually decreasing the learning rate according to the cosine function.

[0056] This invention also proposes a marine environmental noise prediction model based on deep neural networks. The special feature of this marine environmental noise prediction model is that it is the noise prediction model based on CTTM network used in the above prediction method.

[0057] The beneficial effects of this invention are:

[0058] 1. This invention constructs and trains a noise prediction model based on a CTTM network for predicting the spectral level of marine environmental noise. The CTTM network comprises, in the order of data processing, a CNN local feature extraction module, a Transformer encoding module, a TCN sequence convolution module, a multi-head attention pooling module, and a fully connected layer output module, connected sequentially. Although CNN, Transformer, TCN, and attention pooling are all independently usable network structures, this invention does not simply string these modules together. Instead, it addresses the characteristics of marine environmental noise prediction, such as short-term meteorological disturbances, long-term dependencies, multi-scale periodic changes, and uneven contributions from different time steps. The model is constructed according to a specific data flow sequence: "local short-term change feature extraction—global long-term dependency modeling—cross-scale temporal structure enhancement—weighted aggregation of key time steps—noise spectral level output."

[0059] Specifically, the CNN local feature extraction module first extracts short-term variation features within the local temporal neighborhood from the meteorological feature sequence, thus characterizing short-term meteorological disturbances such as sudden increases in rainfall, abrupt changes in air pressure, and changes in wind speed. The Transformer encoding module establishes global dependencies between different time steps based on these short-term variation features. The TCN sequence convolution module expands the temporal receptive field through dilated convolution, further extracting cross-scale temporal features corresponding to tidal cycles, wave cycles, and long-term evolution. The multi-head attention pooling module assigns contribution weights and performs weighted aggregation of hidden features at different time steps, strengthening information from key time steps and relatively suppressing information from weaker time steps. Finally, the fully connected layer output module maps the globally aggregated features to the noise spectral level prediction value corresponding to the target frequency or frequency band. Thus, a progressive feature transfer relationship is formed between the modules, with the features output by the previous module serving as input for further processing by the next module. This avoids problems that easily arise when simply splicing together multiple mature network modules, such as inconsistent feature dimensions, difficulty in fusing local and global features, and averaging of key time step information. If the data is not processed in the specific order of the data stream mentioned above, and instead a single module or the outputs of multiple modules are simply spliced ​​together for prediction, it will be difficult to simultaneously take into account short-term mutation capture, long-term dependency modeling, cross-scale periodic expression, and key time step screening. The model's ability to represent changes in noise in complex marine environments will be significantly reduced.

[0060] One of the technical innovations of this invention is that, based on the feature evolution law in the marine environmental noise prediction task, CNN, Transformer, TCN and multi-head attention pooling are combined and progressively processed according to a specific sequential data flow. This allows the meteorological feature sequence to sequentially complete the extraction of local short-term changes, global dependency modeling, cross-scale temporal enhancement and key time step aggregation, rather than simply stacking mature modules. Therefore, this structure can improve the model's ability to model non-stationary marine environmental noise, long-term and short-term coupling changes and differences in contributions at different time steps while maintaining high parallel computing power.

[0061] 2. The noise prediction model based on the CTTM (CNN-Transformer-TCN-MHAP) network proposed in this invention is designed to address the characteristics of underwater acoustic signals, such as limited sample size, strong randomness, and significant non-stationarity. This is because shallow models alone cannot fully express the complex temporal relationship between meteorological variables and noise spectral levels; simply increasing network depth would increase the risk of overfitting under small sample conditions. This invention addresses the limited sample size by employing Dropout and Gaussian noise enhancement to reduce overfitting; it addresses randomness and anomalous fluctuations by using Huber loss to reduce the impact of outliers; and it addresses non-stationarity and multi-scale variations by using CNN, Transformer, and TCN to progressively extract short-term, global, and cross-scale temporal features, thereby improving the accuracy of marine environmental noise spectral level prediction.

[0062] 3. The noise prediction model based on the CTTM (CNN-Transformer-TCN-MHAP) network proposed in this invention does not rely on the assumption of sample independence, nor does it require constraints on the prior distribution of the signal. Specifically, the CNN local feature extraction module extracts short-term variation features within the local temporal neighborhood from the time-aligned meteorological feature sequence; the Transformer encoding module captures global dependencies within the entire time window through a multi-head self-attention mechanism; the TCN sequence convolution module further captures cross-scale temporal structure features through dilated convolution; and the multi-head attention pooling module automatically learns the contribution weights of different time slices, thereby enhancing key moment information in the final prediction. Due to the limited number of underwater acoustic signal samples, this invention also incorporates techniques such as Dropout and Gaussian noise data augmentation to control overfitting and improve the generalization ability of the noise prediction model for different underwater noise scenarios.

[0063] 4. During model training, this invention utilizes the Huber loss function as the energy function, combined with the Adam optimizer and learning rate scheduling strategy, to ensure smooth convergence of the network during training and avoid excessive influence of outliers on gradient updates. After training, the model not only accurately predicts underwater noise spectrum levels but also exhibits strong robustness, maintaining stable performance under small sample conditions, providing effective technical support for underwater acoustic environment analysis, noise monitoring, and target recognition. Attached Figure Description

[0064] Figure 1 This is a diagram of the noise prediction model architecture based on the CTTM network used in this invention.

[0065] Figure 2 This is a flowchart of the marine environmental noise prediction method based on CTTM network of the present invention (the CNN local feature extraction module shows only one convolution module, and the TCN sequence convolution module shows only one dilated convolution block).

[0066] Figure 3 This is a comparison chart of the accuracy of prediction results. Detailed Implementation

[0067] The present invention will be further described in detail below with reference to the accompanying drawings.

[0068] The marine environmental noise prediction method based on deep neural networks proposed in this invention includes the following steps:

[0069] Step 1: Construct a noise prediction model based on the CTTM network;

[0070] Reference Figure 1 The noise prediction model based on the CTTM network includes a CNN local feature extraction module, a Transformer encoding module, a TCN sequence convolution module, a multi-head attention pooling module, and a fully connected layer output module connected in sequence.

[0071] ①CNN Local Feature Extraction Module

[0072] The CNN local feature extraction module receives the meteorological feature sequence X from the input noise prediction model based on the CTTM network, and extracts short-term variation features C within the local time neighborhood from the meteorological feature sequence X. The meteorological feature sequence X consists of four types of meteorological data—preprocessed and time-aligned rainfall, sea surface pressure, sea surface temperature, and wind speed—according to a preset time window L. The meteorological feature sequence X is not the unprocessed original meteorological observation sequence.

[0073] Let the preset time window length be L (L also represents the number of sampling points within the time window, i.e., the number of time steps), then the meteorological feature sequence X can be represented as:

[0074]

[0075] In the formula, Indicates the first Meteorological feature vectors corresponding to each time step Meteorological feature vectors By the The four types of meteorological variables corresponding to each time step can be represented as follows:

[0076]

[0077] In the formula, Indicates the first Rainfall at each time step, Indicates the first Sea surface pressure at each time step Indicates the first Sea surface temperature at each time step Indicates the first Wind speed at each time step This indicates the number of time steps contained within the preset time window.

[0078] Specifically, the CNN local feature extraction module includes several convolutional modules; each convolutional module includes a serially connected one-dimensional convolutional layer, a batchNormalization normalization layer, a ReLU activation layer, and a max pooling layer. The one-dimensional convolutional layer performs local temporal neighborhood convolution operations on the meteorological feature sequence X to extract local change information of meteorological variables within adjacent time steps. This local temporal change information can characterize short-term meteorological changes occurring within the local temporal neighborhood, such as sudden increases in rainfall, abrupt changes in air pressure, and increased wind speed. The batchNormalization normalization layer normalizes the local change information output by the one-dimensional convolutional modules. The ReLU activation layer performs a non-linear mapping on the normalized features. The max pooling layer downsamples the ReLU-activated features in the time dimension, thereby obtaining short-term change features in the local temporal neighborhood. If the CNN local feature extraction module contains multiple convolutional modules, these modules operate independently and in parallel, with different scales for the one-dimensional convolutional layers. In this case, the CNN local feature extraction module also includes a concatenation module, which is used to process the short-term variation features output by all convolutional modules. Concatenating along the feature dimension yields the concatenated features, which represent the short-term variation features C output by the entire CNN local feature extraction module. If the CNN local feature extraction module contains only one convolutional module, the short-term variation features output by that convolutional module... This refers to the short-term change features C output by the local feature extraction module of the entire CNN.

[0079] Scenario 1:

[0080] When there are multiple convolutional modules, the feature extraction process of the i-th convolutional module can be represented as:

[0081]

[0082]

[0083] In the formula, This represents the kernel size of the one-dimensional convolutional layer in the i-th convolutional module. This represents the one-dimensional convolution operation of the one-dimensional convolutional layer in the i-th convolutional module. This indicates BatchNormalization, a normalization process. Represents the ReLU activation function. This indicates max pooling. This represents the local convolutional features extracted by the i-th convolutional module.

[0084] At each time step, the concatenation module combines the local convolutional features output by different convolutional modules. By concatenating the features along the feature dimension, we obtain the short-term change features of the local temporal neighborhood output by the CNN local feature extraction module at each time step. Its expression is:

[0085]

[0086] In the formula, This represents the short-term change features of the local temporal neighborhood output by the CNN local feature extraction module at time step t, where t=1,2,…,L; This indicates a feature concatenation operation. These represent the kernel sizes used in the one-dimensional convolutional layers of the 1st, 2nd, ..., mth convolutional modules, respectively. Indicates the number of convolutional modules.

[0087] Then, the short-term change feature C output by the CNN local feature extraction module within the preset time window L is:

[0088] .

[0089] Therefore, the CNN local feature extraction module can convert local meteorological change information at different time scales in the meteorological feature sequence X into short-term change features C in the local time neighborhood.

[0090] Scenario 2:

[0091] When there is only one convolutional module, the feature extraction process can be represented as:

[0092]

[0093]

[0094] In the formula, X represents the meteorological feature sequence input to the CNN local feature extraction module; This represents the convolution operation of a one-dimensional convolutional layer in the convolution module; This indicates BatchNormalization normalization processing; Represents the ReLU activation function; This indicates max pooling. The short-term change features of the local temporal neighborhood output by the convolution module are represented as the short-term change features C of the local temporal neighborhood output by the CNN local feature extraction module.

[0095] Therefore, the CNN local feature extraction module can convert local meteorological change information at the same time scale in the meteorological feature sequence X into short-term change features C in the local time neighborhood. This short-term change feature C is used to characterize the short-term change pattern of the meteorological feature sequence X in the local time neighborhood and serves as the basic feature for the subsequent Transformer encoding module to perform global dependency modeling.

[0096] The advantages of the two different CNN local feature extraction modules mentioned above are as follows:

[0097] When the CNN local feature extraction module includes only one convolutional module, the model structure is simpler, with fewer parameters and less computation, and the feature extraction speed is faster. This makes it suitable for applications with high real-time requirements, limited computing resources, or relatively concentrated timescales of weather changes. This structure can extract the main short-term change features in the local time neighborhood with lower complexity, which helps improve the efficiency of model deployment.

[0098] When a CNN's local feature extraction module includes multiple convolutional modules, each module can employ different kernel sizes to extract local meteorological change information from different time scales. Smaller kernels are better at capturing sudden changes in a short period, while larger kernels are better at capturing local change trends over a longer period. The local convolutional features output by multiple modules can be concatenated to form a richer representation of short-term changes. Therefore, this structure has stronger expressive power and robustness to complex meteorological disturbances and local changes of varying durations, but it correspondingly increases the number of parameters and computational cost.

[0099] ②Transformer encoding module

[0100] The Transformer encoding module receives the short-term change features C of the local temporal neighborhood output by the CNN local feature extraction module, and models the correlation between different time steps in the short-term change features C to obtain the global dependency features T.

[0101] Specifically, the Transformer encoding module includes a multi-head self-attention sub-layer, a first residual connection layer, a first normalization layer, a feedforward network sub-layer, a second residual connection layer, and a second normalization layer.

[0102] Multi-head self-attention sublayers are used to extract global attention features between different time steps in short-term change features C. .

[0103] Multi-head self-attention quantum layer It consists of a parallel attention head, a second splicing module, and a linear transformation layer.

[0104] Each attention head independently performs the following operations: It maps the feature representations of the short-term change feature C at each time step within time window L1 to query vectors Query, key vectors Key, and value vectors Value, respectively. At time step t, it calculates the correlation between its query vector Query and the key vectors Key corresponding to all L1 time steps within time window L1, obtaining the attention weight of time step t relative to all L1 time steps. Then, it performs a weighted summation of the value vectors Value corresponding to all L1 time steps based on this attention weight, obtaining the attention feature extracted by this attention head at time step t. , where h is the number of the attention head. The time window L1 is determined by the time window length L of the meteorological feature sequence X and the pooling window size of the max pooling layer in the CNN local feature extraction module, L1 = L / p, where p is the pooling window size of the max pooling layer.

[0105] Therefore, the h-th attention head (h=1,2,…, Based on the short-term change features C extracted by the local temporal neighborhood of the CNN local feature extraction module, it can further establish cross-temporal dependencies between different time steps within the time window L1, thereby forming attention features from a local perspective that include global contextual information within the time window L1. .

[0106] The second stitching module, at each time step of time window L1, will... Attention features obtained from each attention head By concatenating the features and then applying a linear transformation layer to the concatenated features, we can obtain the global attention features at each time step of time window L1. , =1,2,…,L1.

[0107] The first residual connection layer is used to connect the input features (i.e., short-term variation features C) of the multi-head self-attention sub-layer with the global attention features output by the multi-head self-attention sub-layer. The fusion is performed to obtain the fusion feature C+ This preserves the original temporal features and mitigates information attenuation during deep feature transformation; the first normalization layer applies the fused features C+ Normalization is performed to obtain feature U, which improves the stability of the model training process; the feedforward sublayer is used to perform nonlinear mapping on the feature U output by the first normalization layer to obtain feature U. The second residual connection layer maps the nonlinearly mapped features. The features are fused with the input of the feedforward network sublayer (i.e., feature U) to obtain the feature. The second normalization layer is used to normalize the features. After normalization, the global dependency feature T is obtained, which can be represented as: , This represents the global dependency feature at time step t. The global dependency feature T serves as the input to the subsequent TCN sequence convolution module.

[0108] In summary, the feature extraction process of the Transformer encoding module can be represented by the following formula:

[0109] The calculation process for the h-th attention head:

[0110]

[0111]

[0112] The calculation process of splicing and linear transformation:

[0113]

[0114] The computational process of residual connections, layer normalization, and feedforward network sublayer processing:

[0115]

[0116]

[0117] In the formula,

[0118] This represents the short-term variation features in the local temporal neighborhood output by the CNN local feature extraction module;

[0119] This represents the h-th attention head in a multi-head self-attention sub-layer;

[0120] This represents the total number of attention heads in a multi-head self-attention sub-layer;

[0121] This represents the query vector mapping matrix corresponding to the h-th attention head;

[0122] This represents the key vector mapping matrix corresponding to the h-th attention head;

[0123] This represents the value vector mapping matrix corresponding to the h-th attention head;

[0124] This represents the query vector obtained by mapping the short-term change feature C to the h-th attention head;

[0125] This represents the key vector obtained by mapping the h-th attention head to the short-term change feature C;

[0126] This represents the value vector obtained by mapping the h-th attention head to the short-term change feature C;

[0127] Represents the key vector matrix Transpose of;

[0128] The dimension of the key vector is used to represent the... The calculation results are scaled to avoid the dot product result being too large;

[0129] This represents the normalization function used to convert relevance scores into attention weights;

[0130] This represents the attention feature output by the h-th attention head at the t-th time step;

[0131] This indicates a concatenation operation, used to concatenate the outputs of multiple attention heads along the feature dimension;

[0132] This represents the linear transformation matrix at the output of the multi-head self-attention sublayer;

[0133] This represents the global attention feature output by the multi-head self-attention sublayer at time step t;

[0134] Presentation layer normalization processing;

[0135] Representing global attention features Intermediate features obtained by connecting the input features C via residuals and layer normalization;

[0136] This represents a feedforward network sublayer.

[0137] ③TCN Sequence Convolution Module

[0138] The TCN sequence convolution module receives the global dependency feature T output by the Transformer encoding module. It then extracts cross-scale temporal features from the global dependency feature T using N (N>1) serially stacked dilated convolutional blocks, yielding cross-scale temporal features D. These cross-scale temporal features D serve as input to the subsequent multi-head attention pooling module.

[0139] A single dilated convolutional block consists of a dilated one-dimensional convolutional layer, a ReLU activation function, a Dropout layer, and a residual connection layer. The dilated one-dimensional convolutional layer expands the temporal receptive field by introducing a dilation interval between the sampling positions of the convolutional kernel, performing convolution operations on the feature representations at different time intervals in the input globally dependent feature T, thereby obtaining convolutional features without significantly increasing the number of parameters and computational cost. The ReLU activation function is used to perform a nonlinear mapping on the convolutional features output by the dilated one-dimensional convolutional layer to obtain intermediate features. Dropout layers are used for intermediate features. Random deactivation is performed to reduce the risk of overfitting; the residual connection layer is used to fuse the input features of the dilated convolution block with the features obtained after random deactivation to maintain the stability of temporal feature propagation, thereby obtaining fused features. , This represents the number of the dilated convolution block. ,2,…,N; The fused features output by the last dilated convolutional block at time step t. Let be the cross-scale temporal features output by the entire TCN sequence convolutional module at time step t. Therefore, the cross-scale temporal features output by the entire TCN sequence convolutional module... .

[0140] The dilation rates of the dilated one-dimensional convolutional layers within the N dilated convolutional blocks differ and increase progressively. Low-dilation-rate convolutional layers primarily extract short-term variations between adjacent or recent time steps, while high-dilation-rate layers extract long-term variations between more distant time steps. Unlike ordinary convolution, which extracts features only within adjacent time steps, dilated convolutional blocks introduce intervals between kernel sampling positions, enabling convolution operations to cover data across longer time spans, thus expanding the temporal receptive field without significantly increasing the number of parameters or computational cost. Building upon this, through the progressive processing of multiple dilated convolutional blocks with progressively increasing dilation rates, the TCN sequence convolution module can further extract temporal variations at different time scales from the globally dependent feature T, resulting in a cross-scale temporal feature D that simultaneously includes short-term disturbance information and long-term evolutionary information. Therefore, the TCN sequence convolution module can characterize the superimposed effects of different time scale factors such as tidal cycles and wave cycles on the noise spectrum level and improve the model's adaptability to non-stationary events such as sudden approach of ship noise and increased wave intensity.

[0141] The feature extraction process of the TCN sequence convolution module can be expressed by the following formula:

[0142]

[0143]

[0144]

[0145] in,

[0146] Indicates the number of the dilated convolution block; This indicates the number of dilated convolutional blocks in the TCN sequence convolution module; Indicates the first The input features of a dilated convolution block, when hour, ,when hour, For the first The convolutional features output by each dilated convolutional block; Indicates the first Among the input features of the dilated convolution block, the one located at the... Feature representation of each time step; Indicates the first The dilation rate of the dilated one-dimensional convolutional layer in a dilated convolutional block; Indicates the first The kernel size of the dilated one-dimensional convolutional layer in a dilated convolutional block; Indicates the first In the dilated convolutional block, the convolutional kernel in the dilated one-dimensional convolutional layer is... Weight parameters corresponding to each position; Indicates the first The bias term of the dilated one-dimensional convolutional layer in each dilated convolutional block; Represents a non-linear activation function; Indicates the first In the dilated convolutional block, the one-dimensional convolutional layer is dilated at the _th ... Intermediate features are obtained by nonlinear activation of the convolutional features extracted at each time step. Indicates random deactivation; Indicates the first The input features of the dilated one-dimensional convolutional layer in each dilated convolutional block are subjected to dimension matching mapping to... Consistent dimensions; when feature dimensions are consistent ; Indicates the first The dilated convolution block in the _th ... Features output at each time step; D represents the cross-scale temporal features D output by the TCN sequence convolution module within the time window L1.

[0147] As can be seen from the above feature extraction process, in the first... At each time step, a single dilated convolutional block in the TCN sequence convolutional module does not only utilize the global dependency features of the current time step. Instead, it is based on the expansion rate. Simultaneously utilize The feature representations from multiple time points are weighted and calculated. As the dilation rate increases, the convolution operation can cover more distant historical time steps, thus obtaining a larger temporal receptive field. This structure can expand the temporal receptive field without significantly increasing the number of parameters, thereby improving the model's ability to represent non-stationary events such as tidal cycles, wind and wave changes, and sudden approach of ship noise.

[0148] ④ Multi-head Attention Pooling Module (MHAP)

[0149] The multi-head attention pooling module receives the cross-scale temporal features D output by the TCN sequence convolution module, and performs attention weight allocation and weighted aggregation on the feature vectors at different time steps in the cross-scale temporal features D to obtain the global aggregated features Z. This global aggregated feature Z serves as the input to the output module of the subsequent fully connected layer, used to generate the predicted value of the marine environmental noise spectrum corresponding to the target frequency or frequency band.

[0150] Specifically, the multi-head attention pooling module includes parallel settings. One pooling attention head and one feature concatenation unit.

[0151] Each pooling attention head calculates the attention score for the hidden feature vector at each time step in the cross-scale temporal feature D, and obtains the contribution weight of each time step under that pooling attention head through Softmax normalization. This processing method assigns a larger contribution weight to time steps that correspond to key moments that contribute significantly to the prediction of the noise spectrum, and their hidden features are strengthened during the weighted aggregation process; conversely, time steps with smaller contribution weights correspond to moments with weaker prediction contributions, and their hidden features are relatively suppressed during the weighted aggregation process.

[0152] For the The pooled attention head, which is for the first The attention score of the hidden feature vectors at each time step can be expressed as:

[0153]

[0154] In the formula, Indicates the first The pooled attention head is the first Feature vectors at each time step The calculated attention score; Indicates the first A pooled attention head with a learnable query vector; Indicates the first The weight matrix of each pooling attention head is a hyperparameter of the model, which is determined through model training and updating. Indicates the first Bias term for each pooled attention head; Representing the cross-scale temporal features D, the first... Feature vectors at each time step; Represents the hyperbolic tangent activation function; superscript This indicates transpose.

[0155] After obtaining the attention score, the first The pooling attention head is calculated using Softmax normalization. Contribution weights corresponding to each time step:

[0156]

[0157] In the formula, Indicates the first The pooled attention head is the first Feature vectors at each time step The assigned contribution weight; Indicates the first The pooled attention head is the first The attention score is calculated from the hidden feature vectors at each time step. Indicates the time step index; 1 indicates the number of time steps contained within the time window; Represented by natural constant An exponential function with base 0.

[0158] Based on the above contribution weights, the first Each pooling attention head performs a weighted summation of the feature vectors at each time step in the cross-scale temporal feature D, yielding the weighted aggregation result corresponding to that pooling attention head:

[0159]

[0160] In the formula, Indicates the first The weighted aggregation result of the output of each pooling attention head, that is, the pooling result of that pooling attention head.

[0161] Finally, the feature concatenation unit concatenates the weighted aggregation results of all pooling attention heads along the feature dimension to obtain the global aggregated feature Z output by the multi-head attention pooling module:

[0162]

[0163] In the formula, Z represents the global aggregated feature output by the multi-head attention pooling module; Indicates feature concatenation operation; Indicates the first The weighted aggregation result of the pooled attention head outputs; This represents the total number of pooled attention heads in the multi-head attention pooling module.

[0164] In the above processing, the multi-head attention pooling module can automatically learn the contribution of different time steps to the noise spectrum prediction result based on the cross-scale temporal features D output by the TCN sequence convolution module, and perform weighted aggregation of time step information from different feature subspaces through multiple pooling attention heads. The resulting global aggregated feature Z integrates the key information in the cross-scale temporal features and can be used as the input to the output module of the subsequent fully connected layer, thereby improving the model's prediction stability and expressive ability for marine environmental noise spectrum levels.

[0165] ⑤ Fully Connected Layer Output Module (Dense)

[0166] The fully connected layer output module receives the global aggregated feature Z output by the multi-head attention pooling module and performs feature mapping processing on the global aggregated feature Z to obtain the predicted value of the marine environmental noise spectrum level corresponding to the target frequency or target frequency band.

[0167] As mentioned above, the global aggregated feature Z is the overall feature representation obtained by the multi-head attention pooling module after weighted aggregation of cross-scale temporal features, containing key temporal information that contributes to noise spectrum level prediction within the time window L1. The fully connected layer output module consists of a fully connected layer and an output layer; the fully connected layer can be one or more layers, used to perform linear transformation and nonlinear mapping on the input global aggregated feature Z to obtain intermediate features H; the output layer maps the intermediate features H to the noise spectrum level prediction value space, thereby obtaining the noise spectrum level prediction result Y corresponding to the target frequency or target frequency band.

[0168] The mapping process of the output module of the fully connected layer can be represented as:

[0169]

[0170] In the formula, This represents the global aggregated feature output by the multi-head attention pooling module; This represents the feature mapping function corresponding to the output module of the fully connected layer; This represents the predicted value of the marine environmental noise spectrum level corresponding to the target frequency or target frequency band.

[0171] In one alternative implementation, when the fully connected layer is a single layer, the calculation process can be expressed as follows:

[0172]

[0173]

[0174] In the formula, This represents the intermediate features of the hidden fully connected layer output; and These represent the weight matrix and bias term of the output fully connected layer, respectively. and These represent the weight matrix and bias term of the output layer, respectively. This represents the predicted noise spectral level output by the fully connected layer output module.

[0175] In another alternative implementation, when there are two fully connected layers, the calculation process can be expressed as follows:

[0176]

[0177]

[0178]

[0179] In the formula, This represents the intermediate features output by the first fully connected layer; This represents the intermediate features output by the second fully connected layer; and These represent the weight matrix and bias term of the first fully connected layer, respectively. and These represent the weight matrix and bias term of the second fully connected layer, respectively. and represents the weight matrix and bias term of the output layer, respectively; Y represents the scalar noise spectrum prediction value corresponding to the target frequency output by the fully connected layer output module.

[0180] Step 2: Construct the training and test sets;

[0181] Step 2.1, Data Acquisition;

[0182] Four types of meteorological data were acquired: rainfall, sea surface pressure, sea surface temperature, and wind speed. Simultaneously, marine environmental noise data corresponding to the time periods of these four meteorological data were collected. The marine environmental noise data was not presented as raw time-domain waveforms or a single total sound level, but rather as noise spectral levels within each pre-defined frequency band. The noise spectral level within each frequency band serves as the target output corresponding to the four types of meteorological data, and its physical meaning is defined by the following formula:

[0183]

[0184] In the formula, The noise spectral level at frequency f; is the sound pressure power spectral density.

[0185] Step 2.2, meteorological data preprocessing;

[0186] Because meteorological data and marine environmental noise data differ in sampling rate, physical dimensions, and timestamps, it is necessary to perform unified preprocessing on the four types of meteorological data to facilitate their subsequent alignment with the time series of marine environmental noise.

[0187] Meteorological data preprocessing includes:

[0188] Step 2.2.1, missing value interpolation;

[0189] Considering the potential for missing values ​​in various meteorological data, linear interpolation is used to fill in these missing values, ensuring the continuity of the meteorological data over time. For any given time... Missing values, utilizing adjacent time steps The value at that location is estimated as follows:

[0190]

[0191] In the formula, For a moment The estimated value of the meteorological data; and They are time points Meteorological data at the location.

[0192] Step 2.2.2, moving average;

[0193] To mitigate the impact of high-frequency disturbances on the noise prediction model, window lengths of [missing information] were applied to various types of meteorological data. The moving average process is used to filter out short-term fluctuations, resulting in smoothed meteorological data that exhibits a more stable trend in the changes of physical quantities.

[0194]

[0195] In the formula, The meteorological data is smoothed after processing with a moving average. For the data index within the sliding window; For a moment The original meteorological observation data.

[0196] Step 2.2.3, standardization process;

[0197] Since the four types of meteorological data—rainfall, sea surface pressure, sea surface temperature, and wind speed—have different dimensions, they need to be standardized to convert them to a unified scale.

[0198] Standardization can be achieved using Z-score standardization or Min-Max normalization.

[0199] Step 2.3: Time-align the meteorological data with the marine environmental noise data;

[0200] During the training and testing dataset construction phase, to form supervised learning samples, the preprocessed four types of meteorological data were time-aligned with the marine environmental noise data using the same timestamp at a uniform sampling interval, ensuring that the meteorological data in each sample was temporally consistent with the target noise spectrum level. In the actual deployment and prediction phase, only the four types of meteorological data were collected, and the input meteorological features were constructed using the same meteorological preprocessing method as in the training phase, without time alignment with the unknown noise data.

[0201] Step 2.4, construct the dataset;

[0202] First, a fixed-length time window is set, and the aligned meteorological data and ocean noise data are expanded using the same time window.

[0203] Then, meteorological feature vectors are constructed using the four types of meteorological data within each time window. :

[0204]

[0205] In the formula, They represent the first and second, respectively. Rainfall, sea surface pressure, sea surface temperature, and wind speed data within a specific time window; This represents a meteorological feature vector composed of the above four types of meteorological variables, used to characterize the environmental state within this time segment.

[0206] Meteorological feature vector Four types of meteorological data are used for a continuous time segment to help a noise prediction model learn the temporal dependencies between meteorological variables and between meteorological variables and noise spectral levels within that time segment.

[0207] Finally, the meteorological feature vector Each pair of samples is composed of its corresponding target noise spectral level Y, and all sample pairs constitute the dataset. The meteorological feature vectors within each sample pair... As input features, the target noise spectral level Y serves as the target output (label) corresponding to the input features; where , The selected target frequency or frequency band.

[0208] Step 2.5: Divide the dataset into training and test sets;

[0209] The dataset constructed in step 2.4 is divided into a training set and a test set according to a set ratio (e.g., 7:3) for training and testing the noise prediction model.

[0210] To enhance the robustness of the noise prediction model, this invention will also train a set of meteorological feature vectors. Gaussian noise was added to four types of meteorological data to obtain Gaussian noise-enhanced data. This enhanced data was then merged into the training set to expand the input features. Finally, the enhanced data was paired with the target outputs (labels) that matched the original data. The enhanced data merged into the training set consists of new, slightly perturbed input feature data samples that can simulate random interference present in the real environment, thus improving the robustness of the noise prediction model.

[0211] Step 3: Train the noise prediction model;

[0212] The noise prediction model based on CTTM constructed in step 1 is trained using the training set. Training is stopped when the training error meets the preset conditions (such as loss convergence or EarlyStopping trigger).

[0213] Preferably, the model training uses the Adam optimizer, with an initial learning rate of 1e-4, a Huber loss function, and a training batch size of 16. The learning rate scheduling, Early Stopping, and ReduceLROnPlateau strategies are combined to ensure the robustness and generalization ability of the noisy prediction model.

[0214] The Huber loss function uses squared error for small errors and linear error for larger errors, and its expression is:

[0215]

[0216] In the formula, The value of Huber's loss function; The noise spectral level prediction value output by the noise prediction model; True noise spectrum level (label); The error threshold is set according to the required prediction accuracy.

[0217] The Huber loss function measures the error between the predicted value and the true noise spectrum (label). When the error is small, it ensures gradient smoothing, and when the error is large, it reduces the impact of outliers on model training.

[0218] The optimization method employs the backpropagation algorithm to update all network parameters in the noise prediction model. Specifically, the convolutional layers and Dense mapping layers in the CNN local feature extraction module, the multi-head self-attention sub-layers and feedforward sub-layers in the Transformer encoding module, the dilated convolutional layers in the TCN sequence convolution module, the attention weight parameters in the multi-head attention pooling module, and the weight parameters in the fully connected layers are all updated based on the gradient of the Huber loss function.

[0219] During the backpropagation process of the convolutional layer, the gradient change of the Huber loss function with respect to the convolutional kernel is first calculated. :

[0220]

[0221] In the formula, This represents the gradient change. The value of Huber's loss function; These are the weight parameters of the convolutional layer; This indicates the calculation of partial derivatives.

[0222] The weights of the convolutional layer are updated along the gradient direction, and the update formula is:

[0223]

[0224] In the formula, The learning rate; For training steps; For the first The weight parameters of the step; For the first The weight parameters of the step.

[0225] The backpropagation of the multi-head attention pooling module follows the chain rule, achieving joint optimization of attention weights and input features by calculating the gradient of the weighted aggregation operation. The parameters of the multi-head self-attention and feedforward networks in the Transformer encoding module are also updated using a similar backpropagation method.

[0226] The learning rate employs a strategy combining warm-up and cosine decay. It is linearly increased during the first five training epochs to ensure stable convergence of the noisy prediction model. Then, the learning rate is gradually decreased according to a cosine function, as shown in the mathematical expression below:

[0227]

[0228] in, Indicates the first The learning rate corresponding to each training cycle; This is the current training round number; This indicates the number of training rounds that the learning rate warm-up phase lasts. This represents the total number of training rounds for the model; The initial learning rate is denoted as . During the warm-up phase, the learning rate increases linearly with the number of training epochs. After the warm-up phase, a cosine annealing strategy is used to gradually reduce the learning rate, thereby improving the model training stability and convergence performance.

[0229] Step 4: Evaluate the performance of the noise prediction model; if the performance does not meet the preset requirements, return to step 3 to continue training or adjust the model parameters; if the performance meets the preset requirements, proceed to step 5.

[0230] After the noise prediction model is trained, the test set is input into the noise prediction model trained in step 3. The model outputs results based on the noise prediction model and the evaluation index is used to evaluate the model's prediction performance.

[0231] Evaluation metrics include mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). By comparing the differences between the predicted values ​​output by the noise prediction model and the actual noise spectrum level, the accuracy and robustness of the noise prediction model in underwater acoustic noise prediction tasks are verified.

[0232] The mathematical expressions for mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) are as follows:

[0233]

[0234]

[0235]

[0236] In the formula, For the first The true noise spectrum (label) of each input feature data sample. To input the first noise prediction model After inputting a sample of feature data, the predicted value output by the noise prediction model is... This represents the mean of the true noise spectrum (label). The total number of input feature data samples.

[0237] Step 5: Prediction of marine environmental noise;

[0238] The noise prediction model, whose performance met the evaluation requirements, was deployed in practice. Real-time collection of four types of meteorological data—rainfall, sea surface pressure, sea surface temperature, and wind speed—was conducted in the target area. The collected meteorological data underwent missing value interpolation, moving average, and standardization using the same method as in step 2.2, and was then assembled into a meteorological feature sequence X according to the time window set during the training phase. This sequence was then input into the noise prediction model. During actual prediction, since the marine environmental noise data at the time of prediction was unknown, time alignment between the meteorological and noise data was not performed. The noise prediction model outputs the predicted spectral level of marine environmental noise in the target area at the target frequency or frequency band.

[0239] Implementation examples and verification of their technical effects:

[0240] The noise prediction model structure in this embodiment is as follows:

[0241] The CNN local feature extraction module includes a convolutional module; the one-dimensional convolutional layers in this module have a kernel size of 3, 32 output channels, a stride of 1, and use same padding; the max pooling layers in this module have a pooling window size of 2 and a stride of 2. After processing by the CNN local feature extraction module, the meteorological feature sequence X yields short-term variation features C in a local temporal neighborhood of dimension (L / 2, 32).

[0242] The Transformer encoding module has 4 attention heads in its multi-head self-attention sublayer with an embedding dimension of 128. The feedforward network sublayer includes a fully connected layer of dimension 256, a ReLU activation function, and a fully connected layer of dimension 128. The short-term variation feature C of dimension (L / 2, 32) is processed by the Transformer encoding module to obtain the global dependency feature T of dimension (L / 2, 128).

[0243] The TCN sequence convolution module consists of four dilated convolutional blocks stacked sequentially. The output channels of the dilated one-dimensional convolutional layers in each of the four blocks are all 64, the kernel size is 3, and the dilation rates are 1, 2, 4, and 8 respectively. The Dropout value of the Dropout layer in each of the four blocks is 0.2. The global dependency feature T of dimension (L / 2, 128) is processed by the TCN sequence convolution module to obtain the cross-scale temporal feature D of dimension (L / 2, 64).

[0244] The multi-head attention pooling module includes four pooling attention heads and one feature concatenation unit; the cross-scale temporal feature D with dimension (L / 2, 64) is processed by the multi-head attention pooling module to obtain the global aggregated feature Z with dimension (1, 128).

[0245] The fully connected layer output module consists of two fully connected layers and one output layer. The first fully connected layer contains 128 neurons; the second fully connected layer contains 64 neurons; and the output layer contains one output neuron. The global aggregated feature Z with dimensions (1,128) is processed by the fully connected layer output module to obtain the noise spectrum prediction value Y corresponding to the target frequency.

[0246] After the noise prediction model in this embodiment is trained and tested to pass the test, it is used to predict noise, and the results are as follows. Figure 3 As shown. Figure 3 The paper also presents prediction results for commonly used underwater acoustic prediction depth network models such as LSTM, BiLSTM, and BiGRU. Through... Figure 3 As can be seen, the prediction accuracy of this embodiment is superior to that of other network models.

Claims

1. A method for predicting marine environmental noise based on deep neural networks, characterized in that, Including the following steps: Real-time collection of four types of meteorological data in the target area: rainfall, sea surface pressure, sea surface temperature, and wind speed; The four types of meteorological data collected were processed by interpolation of missing values, moving average and standardization, and assembled into meteorological feature sequences according to the time window set in the model training stage, and then input into the trained noise prediction model. The noise prediction model outputs the predicted value of the marine environmental noise spectrum level in the target area at the target frequency or frequency band. The noise prediction model is a noise prediction model based on the CTTM network, including a CNN local feature extraction module, a Transformer encoding module, a TCN sequence convolution module, a multi-head attention pooling module, and a fully connected layer output module. The CNN local feature extraction module is used to extract short-term change features in the local temporal neighborhood from the meteorological feature sequence. The Transformer encoding module models the correlation between different time steps in the short-term change features through a multi-head self-attention mechanism, and obtains and outputs global dependency features. The TCN sequence convolution module captures and outputs cross-scale temporal features from the global dependency features by using multiple dilated convolution blocks stacked in sequence with different dilation rates; the multi-head attention pooling module assigns attention weights and performs weighted aggregation on features at different time steps in the cross-scale temporal features by using multiple parallel pooling attention heads, and then concatenates the weighted aggregation results of all pooling attention heads along the feature dimension to output the global aggregated features. The fully connected layer output module is used to map the global aggregated features to obtain and output the noise spectrum prediction value corresponding to the target frequency.

2. The marine environmental noise prediction method based on deep neural networks according to claim 1, characterized in that, The CNN local feature extraction module includes multiple parallel convolutional modules and a concatenation module. Each convolutional module includes a serially connected one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer. The one-dimensional convolutional layer is used to extract local variation information of meteorological variables within adjacent time steps from the meteorological feature sequence. The batch normalization layer normalizes the local variation information. The ReLU activation layer performs nonlinear mapping on the normalized features. The max pooling layer downsamples the ReLU activated features in the time dimension to obtain short-term variation features. The one-dimensional convolutional layers in different convolutional modules have different scales. The concatenation module is used to concatenate the short-term variation features output by each convolutional module along the feature dimension to obtain the concatenated short-term variation features, which is the output of the entire CNN local feature extraction module.

3. The marine environmental noise prediction method based on deep neural networks according to claim 1, characterized in that, The CNN local feature extraction module includes a convolutional module; a single convolutional module includes a serially connected one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer; the one-dimensional convolutional layer is used to extract local change information of meteorological variables within adjacent time steps from the meteorological feature sequence; the batch normalization layer normalizes the local change information; the ReLU activation layer performs nonlinear mapping on the normalized features; the max pooling layer downsamples the ReLU activated features in the time dimension to obtain short-term change features, which is the output of the entire CNN local feature extraction module.

4. The marine environmental noise prediction method based on deep neural networks according to claim 1, characterized in that, The Transformer encoding module includes a multi-head self-attention sub-layer, a first residual connection layer, a first normalization layer, a feedforward network sub-layer, a second residual connection layer, and a second normalization layer; Multi-head self-attention sublayers are used to extract global attention features between different time steps in short-term change features C. ; The first residual connection layer is used to connect the input features of the multi-head self-attention sub-layer with the global attention features output by the multi-head self-attention sub-layer. The fusion is performed to obtain the fusion feature C+. ; The first normalization layer affects the fused features C+. Normalization is performed to obtain feature U; The feedforward sublayer is used to perform a nonlinear mapping on the feature U output by the first normalization layer to obtain the feature. ; The second residual connection layer will nonlinearly mapped features The features are fused with the input of the feedforward network sublayer to obtain the features. ; The second normalization layer is used to process the features. Normalization is performed to obtain global dependency features.

5. The marine environmental noise prediction method based on deep neural networks according to claim 4, characterized in that, Multi-head self-attention quantum layer It consists of a parallel attention head, a second splicing module, and a linear transformation layer; Each attention head performs the following operations independently: The feature representations of the short-term change features at each time step in time window L1 are respectively mapped to query vector Query, key vector Key, and value vector Value; At time step t, calculate the correlation between its query vector Query and the key vector Key corresponding to all L1 time steps within time window L1, and obtain the attention weight of time step t relative to all L1 time steps. The attention features extracted by the attention head at time step t are obtained by weighting and summing the value vectors Value corresponding to all L1 time steps based on the attention weights. h is the number of the attention head; The time window L1 is determined by the time window length L of the meteorological feature sequence X and the pooling window size of the max pooling layer in the CNN local feature extraction module, L1=L / p, where p is the pooling window size of the max pooling layer; The second stitching module, at each time step of time window L1, will... Attention features obtained from each attention head splicing; The linear transformation layer performs a linear transformation on the concatenated features to obtain the global attention features at each time step of time window L1. , =1,2,…,L1.

6. The marine environmental noise prediction method based on deep neural networks according to claim 1, characterized in that, The TCN module extracts cross-scale temporal features from global dependency features through N serially stacked dilated convolutional blocks with progressively increasing dilation rates. A single dilated convolutional block includes a serially connected dilated one-dimensional convolutional layer, a ReLU activation function, a Dropout layer, and a residual connection layer.

7. The marine environmental noise prediction method based on deep neural networks according to any one of claims 1-6, characterized in that, The structure of the multi-head attention pooling module includes multiple pooling attention heads set in parallel and a feature concatenation unit; Each pooling attention head performs the following operations independently: Attention scores are calculated for the hidden feature vectors at each time step in the cross-scale time series features. The attention scores are then converted into contribution weights of the pooled attention head at different time steps. Based on the calculated contribution weights, the hidden feature vectors at each time step in the cross-scale time series features are weighted and summed, and the weighted aggregation result is output. The feature concatenation unit concatenates the weighted aggregation results corresponding to all pooling attention heads to obtain the global aggregated features.

8. The marine environmental noise prediction method based on deep neural networks according to claim 1, characterized in that, The training method for the noise prediction model based on the CTTM network is as follows: Step 1: Collect data; Four types of meteorological data were acquired: rainfall, sea surface pressure, sea surface temperature, and wind speed. At the same time, marine environmental noise data corresponding to the time of the above four types of meteorological data were collected. The marine environmental noise data were provided by giving the noise spectrum level in each pre-divided frequency band. Step 2: Preprocessing meteorological data; Missing values ​​in the four types of meteorological data were filled using linear interpolation. The four types of meteorological data were processed by moving average to filter out short-term fluctuations, resulting in smoothed meteorological data. The smoothed meteorological data is standardized to convert it to a uniform scale; Step 3: Time-align meteorological data with marine environmental noise data; The four types of meteorological data after standardization were time-aligned with the marine environmental noise data according to the same timestamp, based on a unified sampling interval, to ensure that the meteorological data in each sample is consistent with the target noise spectrum level in time. Step 4: Construct the dataset; Set a fixed-length time window and expand the aligned meteorological data and ocean noise data with the same time window. Feature vectors are constructed using four types of meteorological data within each time window; The meteorological feature vector and its corresponding target noise spectral level are combined into a set of sample pairs, and all sample pairs are used to construct the dataset. Step 5: Divide the dataset into training and testing sets; The dataset is divided into a training set and a test set according to a set ratio; Step 6: Train the model; The noise prediction model based on CTTM was trained using the training set; Step 7, Model Performance Evaluation; The prediction performance of the CTTM-based noise prediction model is evaluated using a test set. If the prediction performance does not meet the preset requirements, return to step 6 to continue training or adjust the model parameters until the prediction performance meets the requirements.

9. The marine environmental noise prediction method based on deep neural networks according to claim 8, characterized in that, The model training uses the Adam optimizer and the Huber loss function. It combines learning rate scheduling, early stopping, and learning rate decay strategies to ensure the robustness and generalization ability of the noisy prediction model. The learning rate adopts a strategy combining warm-up and cosine decay. It is linearly increased in the first 5 training rounds to make the model converge stably, and then the learning rate is gradually decreased according to the cosine function.

10. A marine environmental noise prediction model based on deep neural networks, characterized in that, The noise prediction model based on the CTTM network used in any of the prediction methods described in claims 1-9.

Images