Marine environment noise forecasting method based on LightGBM algorithm

Abstract

The invention provides an ocean environment noise forecasting method based on a LightGBM algorithm. The ocean environment noise forecasting method mainly solves the problem that in the prior art, a model and a method for forecasting noise through machine learning are low in forecasting accuracy. According to the method, on the basis of actually-measured acoustic experiment data and meteorological data in a public database, noise spectrum levels, calculated by the actually-measured acoustic experiment data, of different frequency points serve as label variables, the meteorological data serve as characteristic variables, hyper-parameter optimization is conducted on a LightGBM model through a grid search and cross validation method, and the optimal LightGBM model and hyper-parameters are obtained; the method has the advantages of high training speed, high accuracy and the like during data processing, and prediction precision can be effectively improved while prediction time is shortened.

Description

Technical Field

[0001] The present invention belongs to the field of marine environmental noise prediction, and mainly relates to a marine environmental noise prediction method based on the LightGBM algorithm. Background Art

[0002] Marine environmental noise is a kind of interference background field in the underwater acoustic channel. When using the sonar equation to predict the sonar range, it is required to accurately evaluate the marine environmental noise level. In sonar signal processing, from the perspective of anti-interference, it is also required to fully grasp the spatiotemporal statistical characteristics of the environmental noise field, find and use the spatiotemporal statistical difference characteristics between the signal field and the noise field, so that the sonar system can obtain a higher signal-to-noise ratio and anti-interference ability. In addition, marine environmental noise also contains a lot of information about seawater sound speed, sea surface meteorological parameters, seabed bottom acoustic parameters, etc. The sea surface, water body and seabed parameters can be inverted from the noise data, which is of great significance to the acquisition and guarantee of battlefield environmental information.

[0003] Machine learning methods draw heavily on knowledge about the human brain, statistics, and applied mathematics. Thanks to more powerful computers, larger data sets, and technologies that can train deeper networks, their popularity and applicability have greatly developed. In the context of big data and deep learning, machine learning methods have been successfully applied to various research fields, including speech recognition, language processing, and image processing, and can effectively solve problems in pattern recognition, classification, clustering, function approximation, etc. At present, there have been significant research results in the application of machine learning in ocean acoustics. In the detection of toothed whale positioning echo noise and the identification of biological noise types, the convolutional neural network (CNN) and deep neural network methods can realize automatic detection and identification with an accuracy rate of more than 90%; when environmental parameters such as sea surface temperature, sea surface height, longitude and latitude, seabed topography, month, sea surface current velocity, and empirical orthogonal reconstruction parameters are used as model input parameters, the self-organizing map network (SOM) model can be used to reconstruct the deep ocean velocity field and sound velocity profile; in the study of the acoustic characteristics of seabed sediments, on the basis of the existing BP neural network prediction, the genetic algorithm is used to optimize its initial weights and thresholds to construct a sound velocity prediction model based on water content and porosity for sound velocity prediction; wind speed and wave height, which dominate the intensity of noise sources, are used as input parameters of the machine learning training model, and the artificial neural network (ANN) method can be used to predict the shallow sea low-frequency noise spectrum level curve in the wind speed range of 2.5 to 8.5 m / s.

[0004] In summary, the prediction accuracy of noise prediction models and methods using machine learning at home and abroad is relatively low, so a prediction method is needed to accurately predict marine environmental noise. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art and solve the problem of low prediction accuracy of noise prediction models and methods, the present invention provides a method for predicting marine environmental noise based on the LightGBM algorithm.

[0006] A method for predicting ocean environmental noise based on the LightGBM algorithm includes the following steps:

[0007] S1, construct data set A and perform preprocessing;

[0008] S2, construct training set D and test set T;

[0009] S3, initialize and train the LightGBM model;

[0010] S4. Optimize the hyperparameters of the LightGBM model;

[0011] The grid search and cross-validation methods are used to optimize the hyperparameters of the LightGBM model trained with the training set to obtain the optimal LightGBM model for each label variable, thereby determining the optimal values ​​of the number of leaf nodes num_leaves, the maximum depth of the tree max_depth, the learning rate learning_rate, and the number of decision trees n_estimators to obtain the optimal hyperparameters.

[0012] Furthermore, the dataset A includes acoustic experimental data based on field measurements, the ERA5 Hourly Data On Single Levels From 1979 To Present dataset of the European Centre for Mesoscale Weather Forecasts, and the global NCEP reanalysis dataset of the National Center for Atmospheric Research.

[0013] Furthermore, the preprocessing is:

[0014] The power spectrum of the time domain signal stored in each sample is estimated, and the 1 / 3 octave power spectrum is calculated. The calculation steps are as follows:

[0015] Assume the sampling frequency is f s , the number of sampling points is N, intercept the time domain signal x(n) in file t, n=0,…N-1 for calculation;

[0016] First, perform an M-point fast Fourier transform on x(n) with a frequency resolution of f s / M, spectrum for:

[0017]

[0018] In the formula, f m =m×f s / M is the frequency value corresponding to each point; The unit is W;

[0019] The unit of power spectrum is W / Hz. The power spectrum p(f m ) is:

[0020]

[0021] The power spectrum p(f c ) is calculated as:

[0022]

[0023] where f a is the lower frequency, f c is the center frequency, f b is the upper limit frequency; the power spectrum p(f c ) is in W / Hz;

[0024] The power spectrum level NL(f c ) is calculated as:

[0025] NL(f c )=10log(I(f c ) / I0)-M p -m r (4)

[0026] Among them, M p is the sound pressure sensitivity of the hydrophone, m r is the amplification factor of the receiving system, I0 is the reference sound pressure,

[0027] I0=1μPa,I(f c ) is the sound pressure at the center frequency; the power spectrum level NL(f c ) in dB re 1μPa 2 / Hz.

[0028] Furthermore, the steps of constructing the training set D and the test set T are:

[0029] The preprocessed dataset A is divided into a training set D and a test set T in a ratio of 8:2. The training set D is used for training the LightGBM model, and the test set T is used for evaluating the performance of the LightGBM model.

[0030] The power spectrum levels of different frequency points calculated from the measured acoustic experimental data are used as label variables, and the meteorological data are used as feature variables; the meteorological data include 10m wind speed, significant wave height, rainfall rate and atmospheric pressure.

[0031] Furthermore, the steps of initializing and training the LightGBM model are:

[0032] Set the initialization model and use the training data set D for training;

[0033] The hyperparameters of the LightGBM model include the number of decision trees, learning rate, number of leaf nodes, and maximum depth of the tree; the initial values ​​of the number of decision trees, learning rate, number of leaf nodes, and maximum depth of the tree are:

[0034] The initial value of the number of decision trees n_estimators is 100;

[0035] The initial value of the learning rate learning_rate is 0.1;

[0036] The initial value of the number of leaf nodes num_leaves is 31;

[0037] The initial value of the maximum depth of the tree, max_depth, is -1.

[0038] The beneficial effects of the present invention are:

[0039] The present invention provides a method for predicting marine environmental noise based on the LightGBM algorithm. LightGBM is a fast, distributed, high-performance gradient boosting framework based on the decision tree algorithm, which can be used for sorting, classification, regression and more machine learning tasks. The LightGBM algorithm was used for underwater noise prediction for the first time. The LightGBM algorithm has the advantages of not being easy to overfit, fast training speed and high accuracy when processing data, and is used to train the marine environmental noise prediction model, which can effectively improve the prediction accuracy while reducing the prediction time. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the process of the present invention;

[0041] Figure 2 It is a schematic diagram of the basic structure of the decision tree in the present invention;

[0042] Figure 3 It is a schematic diagram of the GBDT framework in the present invention;

[0043] Figure 4 It is a schematic diagram of the layer-by-layer growth leaf-wise strategy in the present invention;

[0044] Figure 5 It is a schematic diagram of the layer-wise growth level-wise strategy in the present invention;

[0045] Figure 6 It is a schematic diagram of the histogram algorithm in the present invention;

[0046] Figure 7 It is a schematic diagram of data parallelism in the present invention;

[0047] Figure 8 It is a schematic diagram of the experiment in the present invention;

[0048] Fig. 9 These are 24 frequency point spectrum level diagrams in the present invention; (a) is a frequency point spectrum level diagram of 50Hz, 200Hz, 800Hz and 3150Hz; (b) is a frequency point spectrum level diagram of 63Hz, 250Hz, 1000Hz and 4000Hz; (c) is a frequency point spectrum level diagram of 80Hz, 315Hz, 1250Hz and 5000Hz; (d) is a frequency point spectrum level diagram of 100Hz, 400Hz, 1600Hz and 6300Hz; (e) is a frequency point spectrum level diagram of 125Hz, 500Hz, 2000Hz and 8000Hz;

[0049] Fig.10 These are four types of meteorological data graphs in the present invention; (a) is a 10m wind speed meteorological data graph; (b) is a wave height meteorological data graph; (c) is an atmospheric pressure meteorological data graph; (d) is a rainfall rate meteorological data graph;

[0050] Fig.11 It is a schematic diagram of 5-fold cross validation in the present invention;

[0051] Fig.12 It is a graph showing the change in the loss function of the LightGBM model training in the present invention;

[0052] Fig.13 These are scatter plots comparing the true values ​​and predicted values ​​at different frequencies in the present invention; (a) is a scatter plot comparing the true values ​​and predicted values ​​at 80 Hz; (b) is a scatter plot comparing the true values ​​and predicted values ​​at 800 Hz; (c) is a scatter plot comparing the true values ​​and predicted values ​​at 2000 Hz; and (d) is a scatter plot comparing the true values ​​and predicted values ​​at 6300 Hz. DETAILED DESCRIPTION

[0053] The LightGBM (Light Gradient Boosting Machine) algorithm is a gradient boosting decision tree algorithm that is not prone to overfitting, has fast training speed and high accuracy when processing data, and has not yet been used in the field of underwater noise prediction. Therefore, using the LightGBM algorithm to train the marine environmental noise prediction model can effectively improve the prediction accuracy while reducing the prediction time.

[0054] This paper mainly introduces the following important parameters: the number of leaf nodes num_leaves is a unique parameter in the LightGBM model. The more leaf node trees there are, the higher the data partitioning and complexity that the model can represent, but at the same time it increases the risk of overfitting; the maximum depth max_depth determines the maximum depth of each tree. Trees with greater depth can better capture the complex structure of the data. Setting a reasonable depth can help the model find a balance between expressiveness and generalization ability; the learning rate learning_rate determines the size of the step size each time the model is updated. A smaller learning rate can usually improve the generalization ability of the model, but it takes longer to train; the number of trees n_estimators determines the number of trees in the model. The more trees there are, the higher the complexity of the data that the model can capture, and at the same time it increases the risk of overfitting.

[0055] A method for predicting ocean environmental noise based on the LightGBM algorithm includes the following steps:

[0056] S1. Build dataset A;

[0057] Dataset A includes acoustic experimental data based on field measurements, the ERA5 Hourly Data On Single Levels From 1979 To Present dataset of the European Centre for Medium-Range Weather Forecasts (ECMWF), and the global NCEP National Centers for Environmental Prediction reanalysis dataset of the National Center for Atmospheric Research (NCAR);

[0058] Preprocess the data set A, estimate the power spectrum of the time domain signal stored in each sample, and calculate the 1 / 3 octave power spectrum. The calculation steps are as follows:

[0059] Assume the sampling frequency is f s , the number of sampling points is N, intercept the time domain signal x(n) in file t, n=0,…N-1 for calculation;

[0060] First, perform an M-point fast Fourier transform on x(n) with a frequency resolution of f s / M, spectrum for:

[0061]

[0062] In the formula, f m =m×fs / M is the frequency value corresponding to each point; The unit is W;

[0063] The unit of power spectrum is W / Hz. The power spectrum p(f m ) is:

[0064]

[0065] The power spectrum p(f c ) is calculated as:

[0066]

[0067] where f a is the lower frequency, f c is the center frequency, f b is the upper limit frequency; the power spectrum p(f c ) is in W / Hz;

[0068] The power spectrum level NL(f c ) is calculated as:

[0069] NL(f c )=10log(I(f c ) / I0)-M p -m r (4)

[0070] Among them, M p is the sound pressure sensitivity of the hydrophone, m r is the amplification factor of the receiving system, I0 is the reference sound pressure,

[0071] I0=1μPa,I(f c ) is the sound pressure at the center frequency; the power spectrum level NL(f c ) in dB re 1μPa 2 / Hz;

[0072] S2, construct training set D and test set T;

[0073] The preprocessed dataset A is divided into a training set D and a test set T in a ratio of 8:2. The training set D is used for training the LightGBM model, and the test set T is used for evaluating the performance of the LightGBM model.

[0074] The power spectrum levels of different frequency points calculated from the measured acoustic experimental data are used as label variables, and the meteorological data are used as feature variables; the meteorological data include 10m wind speed, significant wave height, rainfall rate and atmospheric pressure;

[0075] S3. Initialize the LightGBM model;

[0076] Set the initialization model and use the training data set D for training;

[0077] The hyperparameters of the LightGBM model include the number of decision trees, learning rate, number of leaf nodes, and maximum depth of the tree;

[0078] The initial value of the number of decision trees n_estimators is 100;

[0079] The initial value of the learning rate learning_rate is 0.1;

[0080] The initial value of the number of leaf nodes num_leaves is 31;

[0081] The initial value of the maximum depth of the tree, max_depth, is -1;

[0082] S4. Optimize the hyperparameters of the LightGBM model;

[0083] Use grid search and cross-validation methods to optimize the hyperparameters of the LightGBM model trained with the training set, and obtain the optimal LightGBM model for each label variable, so as to determine the optimal values ​​of the number of leaf nodes num_leaves, the maximum depth of the tree max_depth, the learning rate learning_rate, and the number of decision trees n_estimators parameters, and obtain the optimal hyperparameter combination;

[0084] S5. Evaluate the performance of the LightGBM model;

[0085] Evaluate the performance of the LightGBM model trained under the optimal hyperparameter combination for each label variable;

[0086] Calculate the root mean square error (RMSE) between the predicted value and the actual value of each label variable, and use logarithmic transformation to calculate RMSE to reduce the impact of outliers;

[0087] The root mean square error RMSE calculation formula is:

[0088]

[0089] where observed t represents the measured value of the noise spectrum level, predicted t Represents the predicted value of the noise spectrum level, T represents the number of samples; the smaller the RMSE, the better the performance of the trained LightGBM model.

[0090] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0091] The object of the present invention is to provide a method for predicting ocean environmental noise based on the LightGBM algorithm, comprising the following steps:

[0092] S1. Build a dataset A based on the measured acoustic experimental data and the public databases such as the ERA5 Hourly Data On Single Levels From 1979 To Present dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), and the global NCEP National Centers for Environmental Prediction reanalysis dataset jointly developed by the National Center for Atmospheric Research (NCAR), to form samples of the training set D and the test set T. The measured acoustic experimental data needs to be preprocessed, and the power spectrum of the time domain signal stored in each sample is estimated, and the 1 / 3 octave power spectrum is calculated. The calculation method is as follows:

[0093] Assume the sampling frequency is f s , the number of sampling points is N, and the time domain signal x(n) in file t is intercepted, n=0,…N-1 for calculation. First, perform M-point fast Fourier transform on x(n), with a frequency resolution of f s / M, the spectrum (W) is:

[0094]

[0095] In the formula, f m =m×f s / M is the frequency value corresponding to each point. Then calculate the power spectrum (W / Hz)

[0096]

[0097] Next, calculate the lower frequency f a , center frequency f c and upper frequency f b , power spectrum (W / Hz) within the defined frequency band:

[0098]

[0099] where f a is the lower frequency, f c is the center frequency, f b is the upper limit frequency; the power spectrum p(f c) is in W / Hz;

[0100] Finally, consider the acoustic pressure sensitivity M of the hydrophone p , the magnification of the receiving system is m r , reference sound pressure I0 = 1μPa power spectrum level at the center frequency (dB re 1μPa 2 / Hz) is:

[0101] NL(f c )=10log(I(f c ) / I0)-M p -m r (4)

[0102] S2. Divide the dataset A into training set D and test set T in a ratio of 8:2, where training set D is used for training the LightGBM model and test set T is used for evaluating the model performance. The noise spectrum levels at different frequencies calculated from the measured acoustic experimental data are used as label variables, and meteorological data (such as 10m wind speed, effective wave height, rainfall rate, and atmospheric pressure) are used as feature variables.

[0103] S3. For the training data set D, set the initialization model. Use K-fold KFold for cross-validation to divide the data into K folds for training and validation. The training data will be divided into K parts, namely D1, D2, ... Dk. Each part is used as a validation set in turn, and the rest is used as training data. In each KFold fold, all labels will be trained and predicted, and a LightGBM model will be trained for each label.

[0104] S4. Use grid search GridSearchCV to optimize the hyperparameters of the LightGBM model, find the best parameter combination grid_search, determine the optimal values ​​of parameters such as the number of leaf nodes num_leaves, the maximum depth of the tree max_depth, the learning rate learning_rate, the number of trees n_estimators, etc., and evaluate and verify the model through the test set T to obtain the optimal model multi_target_model.

[0105] There are two basic methods for selecting hyperparameters: manual selection and automatic selection. Manual selection of hyperparameters requires understanding the role of hyperparameters and how machine learning models achieve good generalization. Automatic hyperparameter selection algorithms greatly reduce the time to learn, but require higher computational costs.

[0106] The main goal of manually searching for hyperparameters is to adjust the effective capacity of the model to match the complexity of the task. The ideal learning algorithm should only need to input a data set and output the learned function without manually adjusting hyperparameters.

[0107] At present, the main methods for automatically selecting hyperparameters include grid search, random search, Bayesian optimization, etc. Grid search is when the hyperparameters are numerical (ordered), and the minimum and maximum elements of each list can be conservatively selected based on the experience of previous similar experiments to ensure that the optimal solution is very likely to be within the selected range. Random search can converge to good values ​​of hyperparameters faster through simple programming, without wasted experiments. Bayesian optimization is an approximate approximation method that uses various surrogate functions to fit the relationship between hyperparameters and model evaluation, and then selects promising hyperparameter combinations for iteration, and finally obtains the best hyperparameter combination.

[0108] Each method has its applicable scenarios and advantages and disadvantages. Grid search is suitable for tuning fewer hyperparameters. The present invention mainly adjusts and optimizes the above four parameters, so the grid search method is used.

[0109] S5. Evaluate the performance of the LightGBM model trained under the optimal hyperparameter combination. Calculate the root mean square error (RMSE) between the predicted value and the actual value of each label, and use logarithmic transformation to calculate RMSE to reduce the impact of outliers.

[0110] Based on the RMSE calculated for each fold, the average RMSE of all folds can be calculated and used as the final evaluation indicator of model performance. The formula for calculating the root mean square error RMSE is:

[0111]

[0112] where observed t represents the measured value of the noise spectrum level, predicted t represents the predicted value of the noise spectrum level, and T represents the number of samples.

[0113] Since the LightGBM algorithm model used in this invention is an implementation of a gradient decision tree (GBDT), its essential principle is to use the base classifier (decision tree) training integration to obtain the optimal model. The basic structure of the decision tree is as follows Figure 2 As shown. The structural framework of the gradient decision tree (GBDT) is shown in Figure 3 As shown in the figure, LightGBM mainly includes a decision tree algorithm based on Histogram, a decision tree growth strategy based on Leaf-Wise, one-sided gradient sampling (GOSS), mutually exclusive feature bundling (EFB), and direct support for categorical features (Categorical Feature).

[0114] (1) Based on the Leaf-Wise decision tree growth strategy, each time in the current leaf node, the leaf node with the largest split gain is found for splitting, rather than splitting all nodes, which can improve accuracy. Figure 4 As shown in the leaf-wise strategy diagram, even if the left subtree is much deeper than the right subtree, as long as the gradient division of the left subtree is still superior to that of the right subtree, the division will continue in the left subtree.

[0115] (2) The Histogram algorithm replaces the pre-sorted algorithm of XGBoost. The Histogram algorithm first bins the eigenvalues, discretizes the continuous floating-point eigenvalues ​​into k integers to form bins, and constructs a histogram with a width of k. When traversing the data, the statistics are accumulated in the histogram based on the discretized values ​​as indexes. After traversing the data once, the histogram accumulates the required statistics. According to the discrete values ​​of the histogram, the optimal split point is traversed. The specific steps are as follows: Figure 6 As shown in the histogram algorithm diagram. The LightGBM algorithm uses a histogram-based split point selection algorithm, and the splitting criterion is minimum variance, that is, maximizing variance gain:

[0116]

[0117] Among them, O represents a training set of a fixed node, j represents the segmentation feature, d represents the corresponding split point, and n O =∑I[x i ∈O],

[0118] (3) The single-sided gradient sampling GOSS algorithm reduces the complexity of calculating the objective function gain by sampling samples. The whole process is as follows:

[0119] 1) First calculate a predicted value preds based on the model;

[0120] 2) Calculate the loss between the predicted value preds and the true value, which may be the root mean square error or other;

[0121] 3) Sort the calculated losses (each sample will have a loss), and the sorted array is sorted;

[0122] 4) Select the top N samples in sorted, which are called large gradient data;

[0123] 5) Randomly select some of the remaining samples with relatively accurate predictions (small gradient data), where a is the proportion of large gradient data selected and b is the proportion of small gradient data selected;

[0124] 6) Multiply the small gradient sample by a weight coefficient (1-a) / b;

[0125] 7) Use the selected large gradient data, small gradient data and weights to train a new weak learner;

[0126] 8) Add this weak learner to the models and repeat the process again.

[0127] (4) EFB mutually exclusive feature bundling mainly includes two steps: feature clustering and feature merging.

[0128] Feature clustering treats each feature as a node, adds edges between non-mutually exclusive nodes, and converts the feature bundle problem into a graph coloring problem. The specific process is as follows:

[0129] 1) Take features as nodes and conflicts between features (the number of samples where two features are simultaneously non-zero) as edge weights to construct a graph;

[0130] 2) Sort the features in descending order according to node degree;

[0131] 3) Traverse the features in order and add them to the existing bundle (if the conflict count of the bundle after adding is less than the threshold) or create a new bundle for it.

[0132] Feature merging: Multiple mutually exclusive features are merged into one feature by adding an offset. For example, if the value range of feature A is (0,10) and the value range of feature B is (0,20), then an offset of 10 can be added to feature B. After merging, the value range of B is (10,30), and the value range of the new feature is (0,30).

[0133] (5) Parallel learning;

[0134] The LightGBM algorithm supports feature parallelism and data parallelism. It uses distributed reduction to assign the task of histogram merging to different machines, reducing communication and computing overhead, and using histograms to accelerate training to further reduce computing time. The specific process is as follows Figure 7 The data parallel diagram is shown in the figure.

[0135] The technical solution of the present invention is described in detail with reference to the accompanying drawings in the embodiments of the present invention.

[0136] The present invention is a method for predicting ocean environmental noise based on the LightGBM algorithm, comprising the following steps:

[0137] S1. Build dataset D, based on the acoustic experimental data of the eastern Indian Ocean and the ERA5 dataset provided by the European Center for Mesoscale Weather Prediction, a public database, and the global NCEP reanalysis dataset jointly developed by the National Center for Atmospheric Research (NCAR), which contains four meteorological data of 10m wind speed, significant wave height, rainfall rate and atmospheric pressure, to form samples of the training set and test set. The acoustic experimental data of the present invention mainly selected data recorded by a hydrophone at a station in the eastern Indian Ocean from January 15 to January 22, 2021, with a water depth of about 460m. The experimental schematic diagram is as follows Figure 8 As shown. A total of 1984 sets of files from January 15 to January 18, 2021 were selected as the training set, and 625 sets of files from January 21 to January 22 were selected as the test set. The normal file length is 131.072 seconds, and each file is a sample.

[0138] After the sample selection is completed, it is preprocessed. First, the obvious identifiable noise is eliminated, and the more obvious line spectrum interference and impulse noise are removed. Secondly, the power spectrum of the time domain signal stored in each sample is estimated, and the 1 / 3 octave power spectrum is calculated in the range of 50Hz to 10000Hz. Fig. 9 The 24 frequency spectrum points are shown.

[0139] In order to train the Lightgbm model, in addition to the 24 frequency spectrum levels calculated from the measured acoustic experimental data as label variables, four meteorological data including 10m wind speed, significant wave height, rainfall rate and atmospheric pressure are also required as feature variables.

[0140] The wave height in the meteorological data is mainly derived from the ERA5 HourlyData On Single Levels From 1979To Present dataset provided by the European Centre for Mesoscale Weather Forecasts. The 10m wind speed, rainfall rate, and atmospheric pressure in the meteorological data are derived from the global NCEP reanalysis dataset jointly developed by the National Center for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR). Fig.10 shown.

[0141] S2. Divide the dataset A into training set D and test set T in a ratio of 8:2, where training set D is used for training the LightGBM model and test set T is used for evaluating the model performance. The 24 frequency spectrum levels calculated from the measured acoustic experimental data are used as label variables, and four meteorological data, 10m wind speed, significant wave height, rainfall rate and atmospheric pressure, are used as feature variables.

[0142] S3. For each training data set D, set the initialization model. Use 5-fold KFold for cross-validation to divide the data into 5 folds for training and validation. The training data will be divided into 5 equal parts, such as Fig.11, as shown in the 5-fold cross validation diagram, each part is used as a validation set in turn, and the rest is used as training data. Ensure that the data is disrupted during each training. At the same time, multiple training and validation can reduce the model's dependence on a single training set and reduce the risk of overfitting. In each KFold fold, all labels will be trained and predicted, and a LightGBM regression model lgbm will be trained for each label.

[0143] S4. Use cross-validation GridSearchCV to optimize the hyperparameters of the lgbm model, find the best parameter combination grid_search, determine the optimal values ​​of parameters such as the number of leaf nodes num_leaves, the maximum depth of the tree max_depth, the learning rate learning_rate, the number of trees n_estimators, etc. in the model, and evaluate and verify the model through the test set T to obtain the optimal model multi_target_model.

[0144] The optimization range of each parameter is shown in the following table:

[0145] Hyperparameters Hyperparameter meaning Optimization range num_leaves Number of leaf nodes [25，31，35] max_depth Maximum Depth [-1，90，100] learning_rate Learning Rate [0.1，0.2] n_estimators Number of trees [90，100，110]

[0146] S5. Evaluate the performance of the model trained under the optimal hyperparameter combination. Calculate the root mean square error (RMSE) between the predicted value and the actual value of each label, and use logarithmic transformation to calculate RMSE to reduce the impact of outliers. Based on the RMSE calculated for each fold, the average RMSE of all folds can be calculated and used as the final evaluation indicator of model performance. The calculation results are as follows: Fig.12 The model training loss function graph is shown.

[0147] In order to more clearly show the training results of the model, a scatter plot comparing the actual value and the predicted value of different frequencies (80Hz, 800Hz, 2000Hz, 6300Hz) is given, as shown in the figure: Fig.13 Scatter plot comparing the true and predicted values ​​for different frequencies is shown.

[0148] The above contents are further detailed descriptions of the present invention in combination with specific optimized implementation methods, and the specific implementation of the present invention cannot be considered to be limited to these descriptions. Under the premise of obviously departing from the idea of ​​the present invention, any simple promotion or substitution made shall be deemed to fall within the protection scope of the present invention.

Claims

1. A method for predicting ocean environmental noise based on the LightGBM algorithm, characterized in that: The steps include: S1, construct data set A and perform preprocessing; S2, construct training set D and test set T; S3, initialize and train the LightGBM model; S4. Optimize the hyperparameters of the LightGBM model; The grid search and cross-validation methods are used to optimize the hyperparameters of the LightGBM model trained with the training set to obtain the optimal LightGBM model for each label variable, thereby determining the optimal values ​​of the number of leaf nodes num_leaves, the maximum depth of the tree max_depth, the learning rate learning_rate, and the number of decision trees n_estimators to obtain the optimal hyperparameters.

2. According to claim 1, a method for predicting ocean environmental noise based on the LightGBM algorithm is characterized in that: The dataset A includes acoustic experimental data based on field measurements, the ERA5HourlyDataOnSingleLevelsFrom1979ToPresent dataset of the European Centre for Mesoscale Weather Forecasts, and the global NCEP reanalysis dataset of the National Center for Atmospheric Research.

3. According to claim 1, a method for predicting ocean environmental noise based on the LightGBM algorithm is characterized in that: The steps of the pretreatment are: The power spectrum of the time domain signal stored in each sample is estimated, and the 1 / 3 octave power spectrum is calculated. The calculation steps are as follows: Assume the sampling frequency is f s , the number of sampling points is N, intercept the time domain signal x(n) in file t, n=0,…N-1 for calculation; First, perform an M-point fast Fourier transform on x(n) with a frequency resolution of f s / M, spectrum for: In the formula, f m =m×f s / M is the frequency value corresponding to each point; The unit is W; The unit of power spectrum is W / Hz. The power spectrum p(f m ) is: The power spectrum p(f c ) is calculated as: where f a is the lower frequency, f c is the center frequency, f b is the upper limit frequency; the power spectrum p(f c ) is in W / Hz; The power spectrum level NL(f c ) is calculated as: NL(f c )=10log(I(f c ) / I0)-M p -m r (4) Among them, M p is the sound pressure sensitivity of the hydrophone, m r is the amplification factor of the receiving system, I0 is the reference sound pressure, I0=1μPa, I(f c ) is the sound pressure at the center frequency; the power spectrum level NL(f c ) in dB re 1μPa 2 / Hz.

4. The method for predicting ocean environmental noise based on the LightGBM algorithm according to claim 1 is characterized in that: The steps of constructing the training set D and the test set T are: The preprocessed dataset A is divided into a training set D and a test set T in a ratio of 8:

2. The training set D is used for training the LightGBM model, and the test set T is used to evaluate the performance of the LightGBM model. The power spectrum levels of different frequency points calculated from the measured acoustic experimental data are used as label variables, and the meteorological data are used as feature variables; the meteorological data include 10m wind speed, significant wave height, rainfall rate and atmospheric pressure.

5. The method for predicting ocean environmental noise based on the LightGBM algorithm according to claim 1 is characterized in that: The steps of initializing and training the LightGBM model are: Set the initialization model and use the training data set D for training; The hyperparameters of the LightGBM model include the number of decision trees, learning rate, number of leaf nodes, and maximum depth of the tree; the initial values ​​of the number of decision trees, learning rate, number of leaf nodes, and maximum depth of the tree are: The initial value of the number of decision trees n_estimators is 100; The initial value of the learning rate learning_rate is 0.1; The initial value of the number of leaf nodes num_leaves is 31; The initial value of the maximum depth of the tree, max_depth, is -1.

Images