Method and device for inverting sea surface wind speed through marine environment noise and storage medium

Abstract

The invention discloses a sea surface wind speed inversion method and device based on marine environment noise, and a storage medium. The method comprises the following steps: S1, obtaining marine noise data; s2, acquiring wind speed data according to time and latitude and longitude information in the ocean noise data; s3, audio data in the ocean noise data are processed, and center frequency point noise spectrum level data are obtained; s4, summarizing the noise spectrum level data of the center frequency point and the corresponding wind speed data into a neural network data set through time and latitude and longitude information; step S5, constructing an initial Elman network; s6, performing iterative optimization on the network weight and bias of the Elman network by adopting a cuckoo search algorithm to generate optimized network parameters; step S7, initializing an Elman network by using the optimized network parameters, and generating a CS-Elman model; training a CS-Elman model by using the neural network data set; and step S8, carrying out quantitative evaluation on the CS-Elman model. The method can achieve the high-precision prediction of the sea surface wind speed.

Description

Technical Field

[0001] This invention relates to a wind speed inversion method based on Elman neural networks, and more particularly to a method, device, and storage medium for inverting sea surface wind speed based on marine environmental noise. Background Technology

[0002] Marine environmental noise, as an important carrier reflecting the dynamic processes at the air-sea interface, exhibits a significant statistical correlation between its energy distribution in the 10Hz to 10kHz frequency band and changes in sea surface wind speed. This has become a crucial research direction for non-contact inversion of marine wind fields in recent years. Currently widely used sea surface wind speed monitoring technologies, such as buoy observation, GNSS-R remote sensing, and radar inversion, while achieving some success in local or wide-area observations, still face bottlenecks in high-resolution, high-precision real-time monitoring, including limited data coverage, high equipment costs, and strong environmental dependence.

[0003] Artificial neural networks, especially dynamic network structures with short-term memory capabilities such as Elman networks, have been applied to many geoscientific variable estimation tasks due to their superior performance in modeling complex nonlinear dynamic systems. However, due to problems such as gradient vanishing and sensitivity to initial weights, Elman networks are prone to getting trapped in local optima in wind speed inversion and have insufficient generalization ability. Summary of the Invention

[0004] This invention proposes a machine learning-based method for retrieving sea surface wind speed from marine environmental noise. This method utilizes marine environmental noise data collected by an underwater glider. After extracting multi-band spectral features through 1 / 3 octave band analysis, it incorporates water depth and buoyancy status as input variables. The method is then trained using an Elman network optimized with the Cuckoo Search algorithm to achieve high-precision prediction of sea surface wind speed. This invention introduces the Cuckoo Search (CS) algorithm, which is based on bird nest parasitism and the Levy flight model. This algorithm possesses excellent global optimization capabilities and can be used to optimize the neural network weights and bias parameters, thereby enhancing model convergence and stability.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for inverting sea surface wind speed from marine environmental noise includes the following steps:

[0007] Step S1: Acquire ocean noise data;

[0008] Step S2: Obtain wind speed data based on the time and latitude / longitude information in the ocean noise data;

[0009] Step S3: Process the audio data in the ocean noise data to obtain the center frequency noise spectrum data;

[0010] Step S4: The center frequency noise spectrum data is used to summarize the corresponding wind speed data into a neural network dataset using the time and latitude / longitude information;

[0011] Step S5: Construct the initial Elman network;

[0012] Step S6: The network weights and biases of the Elman network are iteratively optimized using the Cuckoo Search algorithm to generate optimized network parameters;

[0013] Step S7: Initialize the Elman network using the optimized network parameters to generate a CS-Elman model; train the CS-Elman model using the neural network dataset;

[0014] Step S8: Quantitatively evaluate the CS-Elman model.

[0015] Preferably, in step S3, the audio data is subjected to one-third octave band analysis to obtain multiple center frequency noise spectrum data of 10 to 10000Hz at depths of 50 to 800m in each profile, and the center frequency noise spectrum data is written into an Excel spreadsheet using MATLAB.

[0016] Preferably, in step S5, the input layer is 33-dimensional, the output layer is 1-dimensional, and the number of hidden layer nodes is set to 8 based on error analysis experience.

[0017] Preferably, in step S8, the CS-Elman model is used to perform inference predictions on the training set, the test set, and the independent sample group respectively; the mean absolute error, mean square error, root mean square error, and mean absolute percentage error are calculated.

[0018] In a second aspect, an electronic device includes a processor and a memory; the processor is connected to the memory; the memory is used to store executable program code; the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, for performing the steps of the method provided as in the first aspect or any possible implementation thereof.

[0019] Thirdly, a computer-readable storage medium having a computer program stored thereon, the computer-readable storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform the method provided as in the first aspect or any possible implementation thereof.

[0020] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0021] (1) By introducing the CS algorithm to globally optimize the weights and biases of the Elman network, the traditional Elman network is effectively avoided from getting trapped in local optima and unstable training, thus improving the convergence speed and stability of the model.

[0022] (2) Using marine environmental noise data measured by underwater gliders, spectral features of 31 center frequency points from 10Hz to 10kHz were extracted by 1 / 3 octave band processing, and combined with multi-dimensional input variables such as depth and attitude, so that the model can effectively learn complex nonlinear relationships.

[0023] (3) It shows excellent inversion performance on the training set, test set and independent sample group. The prediction error is mainly concentrated in the range of [-0.5, 0.5) m / s, with a sample ratio of more than 80%. The mean absolute error (MAE) is 0.37798, the mean square error (MSE) is 0.24646, the root mean square error (RMSE) is 0.49644, and the mean absolute percentage error (MAPE) is as low as 3.99%.

[0024] (4) Compared with traditional methods such as BP, SSA-BP and unoptimized Elman network, the method of the present invention performs better in terms of accuracy of wind speed inversion, concentration of error distribution and generalization ability, and is suitable for wind speed monitoring scenarios with high precision and real-time requirements.

[0025] In summary, this invention combines the cuckoo search algorithm with the Elman neural network model to propose a sea surface wind speed inversion method based on the CS-Elman neural network. This method improves the accuracy of wind speed inversion and can handle large amounts of noise spectral data. Furthermore, its MAE, MSE, and RMSE parameters outperform those of the Elman neural network, meeting the model error requirements in sea surface wind speed inversion applications. This invention not only improves the accuracy and robustness of the wind speed inversion model but also expands the application value of marine noise data in non-contact wind speed inversion. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the Elman neural network structure in Embodiment 1 of the present invention;

[0027] Figure 2 The flowchart of the CS-Elman algorithm in Embodiment 1 of the present invention is shown below;

[0028] Figure 3 This is a comparison chart of different node count parameters in the hidden layer of Embodiment 1 of the present invention;

[0029] Figure 4 This is a comparison chart of the prediction results of each model in Embodiment 1 of the present invention;

[0030] Figure 5A comparison of the prediction errors of each model in Embodiment 1 of the present invention;

[0031] Figure 6 This is a comparison chart of model parameters in Embodiment 1 of the present invention. Detailed Implementation

[0032] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0033] Example 1:

[0034] A method for inverting sea surface wind speed from marine environmental noise includes the following steps:

[0035] Step S1: Acquire ocean noise data;

[0036] This example uses .dat format files of marine environmental noise data collected by an underwater glider in the South China Sea. The depth, time, latitude and longitude, and profile number contained in the data are extracted using MATLAB and stored in an Excel spreadsheet. The audio information is also extracted and saved as a WAV file, thus obtaining the raw data based on marine environmental noise.

[0037] Step S2: This invention is mainly applied to sea surface wind speed inversion. Therefore, standard wind speed values ​​(true values) for the corresponding time and latitude / longitude region are needed to train the network model and to compare with the inversion results to calculate errors. The standard wind speed values ​​(true values) for the corresponding time and latitude / longitude region are downloaded from the European Centre for Medium-Range Weather Forecasts (ECMWF) platform. This yields the required standard wind speed values ​​(true values).

[0038] Step S3: This invention performs one-third octave band analysis on WAV format audio files. In noise measurement, the one-third octave band spectrum analysis method can more comprehensively reflect the spectral characteristics of the noise source, providing a more complete understanding of the sound source generation mechanism and proposing optimal noise reduction strategies. The human ear's ability to distinguish audio frequencies can be described by frequency bands, and the one-third octave band is considered a frequency band division method that better suits human hearing characteristics. For a set of sound signals, if one-third octave band analysis is required, one must first be familiar with the method for selecting the center frequency of the one-third octave band, and then use Fourier transform to convert the time-domain audio signal to the frequency domain to calculate the sound pressure level of each center frequency. The steps for processing a specific audio signal using the one-third octave band are roughly as follows: First, use MATLAB to read the audio file information, perform a Fourier transform, select 31 center frequency points within the range of 10–10000 Hz according to the one-third octave band division method, and perform spectrum analysis on data collected from different profiles. A profile refers to a continuous set of data collected along the depth direction during the process of an underwater glider descending from the sea surface to a target depth (e.g., 1000 meters) and then surfacing back to the surface. During the ascent and descent phases of each profile, samples are uniformly taken from a depth range of 50m to 800m, extracting 31 sets of noise spectral level data for the corresponding depth. Ultimately, one-third octave band spectral data for all profiles within the depth range of 50–800m is obtained.

[0039] Step S4: Data aggregation. Noise spectrum data at multiple center frequency points (10–10000Hz) at depths of 50–800m across each profile are matched with corresponding wind speed data using time and latitude / longitude information and aggregated as the dataset for subsequent neural network processing.

[0040] Step S5, as follows Figure 1 As shown, an Elman neural network is constructed. The main structure of an Elman neural network is a feedforward connection, consisting of three parts: an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is set to 33, corresponding to the spectral features + depth + pose dimension; the number of nodes in the output layer is 1, corresponding to the predicted wind speed. The number of nodes in the hidden layer is based on an empirical formula:

[0041]

[0042] Where n is the number of hidden layer nodes, i is the number of input nodes, o is the number of output nodes, and e is a constant ∈ [1:10]. Since the number of input nodes is 33 and the number of output nodes is 1, the range of hidden layer nodes is 6–15. To establish the optimal network model, this study adopts an adaptive strategy for selecting the optimal number of nodes. The optimal number of hidden layer nodes is selected by calculating and comparing network performance parameters under different node numbers, such as… Figure 3The analysis results are shown, resulting in a selection of 8 hidden layer nodes. An Elman neural network was created, and the dataset was normalized. The transfer functions for the hidden and output layers were set to tansig and purelin, respectively. The network used the tansig function as the activation function for the hidden layers and the purelin linear function for the output layer. The training algorithm was the Levenberg-Marquardt (LM) backpropagation algorithm. Initial weights and biases were generated using the Cuckoo algorithm, without random initialization, to improve convergence efficiency and model stability.

[0043] Step S6: Optimize using the Cuckoo Search (CS) algorithm. Initialize the population size to 10, the maximum number of generations to 50, and the fitness function to be the mean squared error (MSE) of the network output. During optimization, each individual represents a set of Elman network weights and biases. The Cuckoo Search generates new individual solutions through Levy flights and updates individuals using a local random walk and greedy selection strategy. If the fitness of the current individual's new solution is better than the previous generation's solution, the new solution is retained; otherwise, the original state is maintained. Some individuals are randomly replaced under the control of a discovery probability pa = 0.25 to enhance population diversity and avoid local convergence. Finally, a set of optimal initial network parameter configurations that minimizes the MSE is obtained. Figure 2 The diagram shows the optimization flowchart of the CS-Elman model, illustrating the optimization path of the cuckoo search algorithm during network initialization, including steps such as nest initialization, fitness evaluation, and Levy flight update.

[0044] Step S7: Initialize the Elman network using the network parameters optimized by the CS algorithm, and train it using the normalized sample set. The dataset contains 1308 samples, of which 1066 are used as the training set, 150 as the test set, and 92 independent samples are used for performance verification. The number of training epochs is set to 1000, the learning rate is 0.01, the minimum performance gradient is 10^-6, and the target error is set to 10^-5. An early stopping strategy is used during training to prevent overfitting, and the best model is retained for subsequent inversion after the training error converges.

[0045] Step S8: Perform model inference predictions on the training set, test set, and independent sample groups respectively, and calculate four indicators: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). Compare this method with the traditional BP network, the SSA-BP optimized network, and the unoptimized original Elman network. Figure 6As shown, the CS-Elman model outperforms the comparative models in terms of inversion accuracy, stability, and error distribution concentration, especially with a sample proportion exceeding 80% within the prediction error range [-0.5, 0.5) m / s. Furthermore, its MAPE is as low as 3.99%, demonstrating stronger generalization ability.

[0046] Figure 4 The image shows the wind speed inversion results for selected time periods using CS-Elman, BP, SSA-BP, and Elman network models trained on a noise spectrum dataset. The red curve, WindVelocity, represents the actual time-wind speed values, while the others are the inversion results from each neural network model. Figure 3 As shown, the overall trend of the inversion results of each model is very similar to the actual wind speed value, and they all have good inversion effects. In comparison, although the CS-Elman model has a slightly larger inversion error at some time points, the overall wind speed inversion result curve is less different from the actual wind speed value compared with other models. Figure 5 The inversion errors of each model were segmented into four segments based on their numerical values: [-2, -0.5), [-0.5, 0), [0, 0.5), and [0.5, 2). It is evident that the CS-Elman model has the largest proportion of inversion errors in the segments [-0.5, 0) and [0, 0.5), indicating that the CS-Elman model's inversion errors are concentrated within a smaller range, resulting in better inversion performance.

[0047] The specific design principles and workflow of this invention have been described in detail above. Those skilled in the art can make numerous modifications and variations based on the concept of this invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of this invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for inverting sea surface wind speed from marine environmental noise, characterized in that, Includes the following steps: Step S1: Acquire ocean noise data; Step S2: Obtain wind speed data based on the time and latitude / longitude information in the ocean noise data; Step S3: Process the audio data in the ocean noise data to obtain the center frequency noise spectrum data; Step S4: The center frequency noise spectrum data is used to summarize the corresponding wind speed data into a neural network dataset using the time and latitude / longitude information; Step S5: Construct the initial Elman network; Step S6: The network weights and biases of the Elman network are iteratively optimized using the Cuckoo Search algorithm to generate optimized network parameters; Step S7: Initialize the Elman network using the optimized network parameters to generate a CS-Elman model; train the CS-Elman model using the neural network dataset; Step S8: Quantitatively evaluate the CS-Elman model.

2. The method for inverting sea surface wind speed from marine environmental noise according to claim 1, characterized in that, In step S3, the audio data is subjected to one-third octave band analysis to obtain multiple center frequency noise spectrum data of 10 to 10000Hz at depths of 50 to 800m in each profile, and the center frequency noise spectrum data is written into an Excel spreadsheet using MATLAB.

3. The method for inverting sea surface wind speed from marine environmental noise according to claim 1, characterized in that, In step S5, the input layer is 33-dimensional, the output layer is 1-dimensional, and the number of hidden layer nodes is set to 8 based on error analysis experience.

4. The method for inverting sea surface wind speed from marine environmental noise according to claim 1, characterized in that, In step S8, the CS-Elman model is used to perform inference predictions on the training set, the test set, and the independent sample group respectively; the mean absolute error, mean square error, root mean square error, and mean absolute percentage error are calculated.

5. An electronic device, comprising a processor and a memory; the processor being connected to the memory; the memory being used to store executable program code; the processor running a program corresponding to the executable program code by reading the executable program code stored in the memory, for performing the steps of the method as described in claim 1.

6. A computer-readable storage medium having a computer program stored thereon, the computer-readable storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform the method of claim 1.

Images