Typhoon path inversion method and system based on ocean background noise spectrum characteristics

Abstract

The application relates to the technical field of marine disaster early warning, and discloses a typhoon path inversion method and system based on marine background noise spectrum characteristics, which comprises the following steps: S1, constructing an LSTM-Attention feature fusion model; S2, collecting noise data and meteorological data of at least three nodes, once every 5 minutes; the noise data comprises 8-20 Hz infrasound wave signals, and the meteorological data comprises air pressure data, wind speed data and temperature data; S3, performing screening processing on the noise data to obtain effective infrasound wave signals; and S4, performing spectrum analysis and feature extraction on the effective infrasound wave signals to obtain infrasound wave features and the like; the method and system are coupled with meteorological parameters through marine background noise infrasound wave features (8-20 Hz) to model, a physical coupling mechanism of a 'typhoon infrasound wave-air pressure gradient-sea surface wind field' is established, and typhoon path deviation can be early warned for 6-12 hours.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine disaster warning, in particular to a typhoon path inversion method and system based on marine background noise spectrum characteristics. BACKGROUND

[0002] Marine engineering scenarios such as offshore wind farms and ocean observation stations have an urgent need for real-time perception of typhoon dynamics. Existing typhoon monitoring technologies have significant defects: 1. The update cycle of satellite cloud images and meteorological buoy data generally reaches or exceeds 30 minutes. This time lag makes it difficult for the system to capture sudden changes in typhoon paths, including sudden changes in direction or sudden acceleration, which are key dynamic characteristics, seriously affecting the real-time nature of early warning; there is a significant blind spot in near-shore area monitoring. When strong rainfall occurs near the eye wall of a typhoon, traditional radar signals are easily disturbed, causing path prediction errors to often exceed 50 kilometers. This precision gap is particularly pronounced when a typhoon approaches land, directly threatening the accuracy of disaster prevention decisions in coastal areas; the data dimensions of existing monitoring methods are relatively single, mainly relying on changes in single parameters such as air pressure or wind speed, lacking a comprehensive representation mechanism for the physical processes of typhoon and sea surface interaction, and unable to fully reflect the energy transfer and structural evolution of the typhoon system, resulting in inadequate response of prediction models to complex environments. SUMMARY

[0003] To overcome the defects of the prior art, the present application provides a typhoon path inversion method and system based on marine background noise spectrum characteristics, which can realize intelligent early warning of real-time inversion of typhoon paths and is suitable for scenarios such as offshore wind farms and ocean observation stations that require early perception of typhoon dynamics, solving the problems of low time resolution and large near-shore blind area of traditional monitoring methods.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] The typhoon path inversion method based on the ocean background noise spectrum characteristics comprises the following steps: S1, constructing an LSTM-Attention feature fusion model; S2, collecting noise data and meteorological data of at least three nodes, once every 5 minutes; the noise data comprises 8-20Hz infrasound signals, and the meteorological data comprises air pressure data, wind speed data and temperature data; S3, performing screening processing on the noise data to obtain effective infrasound signals; S4, performing spectrum analysis and feature extraction on the effective infrasound signals to obtain infrasound features, wherein the infrasound features comprise a center frequency offset and a frequency band energy ratio; S5, performing screening and verification on the meteorological data to obtain accurate air pressure data, wind speed data and temperature data; S6, calculating an air pressure gradient and a wind shear index according to the accurate air pressure data and the wind speed data as meteorological parameters; S7, performing time stamp unification on the infrasound features and the meteorological parameters; S8, inputting the infrasound features and the meteorological parameters after time stamp unification as feature values into the LSTM-Attention feature fusion model, and the LSTM-Attention feature fusion model outputs the coordinates of the typhoon center and the moving speed of the typhoon; S9, calculating the average value of the typhoon coordinates and the moving speed of the typhoon obtained by each node as the final coordinates of the typhoon center and the final moving speed of the typhoon; in S8, the process that the LSTM-Attention feature fusion model outputs the coordinates of the typhoon center and the moving speed of the typhoon comprises the following steps: S81, performing time series data format conversion on the infrasound features and the meteorological parameters, and performing data normalization verification and adjustment to obtain a data tensor; S82, inputting the data tensor into a first-layer LSTM to extract features, and outputting a first-layer LSTM feature vector; S83, inputting the data tensor into a second-layer LSTM to extract features, and outputting a deep-layer feature vector; S84, performing attention weight calculation, normalizing the attention weight, performing feature weighting, splicing the attention output result and the deep-layer feature vector to obtain a fused feature vector; S85, initializing a full connection network, inputting the fused feature vector into the full connection network, performing full connection layer feature mapping, and outputting the coordinates of the typhoon center and the moving speed of the typhoon.

[0006] In the application, preferably, S3 comprises the following steps: S31, removing invalid data, and correcting the infrasound signal amplitude according to the ambient temperature at the collection time; S32, processing the infrasound signal by using a wavelet threshold denoising algorithm; S33, verifying the signal-to-noise ratio of the infrasound signal, and correcting the infrasound signal paragraph higher than the signal-to-noise ratio threshold to obtain effective infrasound signals.

[0007] In the application, preferably, S4 comprises: S41, performing wavelet packet decomposition on the effective infrasound wave signal to generate a plurality of sub-bands, and calculating the energy of each sub-band; S42, counting the energy proportion of all sub-bands, and sorting the sub-bands according to the energy proportion from large to small; S43, accumulating the energy proportion, extracting the sub-band with cumulative proportion greater than 80% or extracting the sub-band of 11-14Hz, and merging the sub-bands into a main frequency band; S44, calculating the center frequency f_current of the main frequency band, comparing the center frequency f_current with the benchmark center frequency f_benchmark in the non-typhoon period, and obtaining the center frequency offset Δf = f_current - f_benchmark; and S45, calculating the total energy E of the main frequency band 主 In the total energy E of the full frequency band （8-20）Hz , the proportion of the main frequency band is obtained, and the frequency band energy ratio R = E 主 / E （8-20）Hz .

[0008] In the application, preferably, S5 comprises: S51, removing outliers in the air pressure data, wind speed data and temperature data; S52, filling in missing values in the air pressure data, wind speed data and temperature data; and S53, performing consistency check on the air pressure data, wind speed data and temperature data, and the data passing the check is accurate data.

[0009] In the application, preferably, in S6, the air pressure gradient is calculated in the following manner: calculating the straight line distance D_ij between adjacent nodes, calculating the air pressure difference ΔP_ij of adjacent nodes, obtaining the air pressure gradient _ij = ΔP_ij / D_ij of each adjacent node, taking the maximum value as the air pressure gradient of the collection period, and retaining 2 decimal places; and the wind shear index α is calculated in the following manner: for each node, obtaining the wind speed v 10 at the height of 10m, the wind speed v 100 at the height of 100m, calculating α according to the logarithmic wind profile model, α = ln(v 100 / v 10 ) / ln (100 / 10) = ln (v 100 / v 10 ) / 2.303, if v 10 = 0, then α is 0.25, and if v 100 ≤ v 10 , then the wind speed data is re-collected.

[0010] In the application, preferably, the S7 comprises: S71, extracting the time stamp of infrasound wave characteristics and weather parameters collection, and converting into "year-month-day-hour-minute-second" format; S72, if the time stamp difference of infrasound wave characteristics and weather parameters in the same collection cycle is greater than 1 minute, it is determined as "time misalignment", and the cycle data is removed.

[0011] In the application, preferably, it further comprises: S10, introducing set Kalman filter to fuse satellite cloud image data, and correcting the error of final coordinates of typhoon center and final typhoon moving speed through dynamic weight distribution.

[0012] In the application, preferably, it further comprises: S11, if the path offset angle is greater than 25° or the moving speed increment is greater than 10 km / h, a first-level early warning is triggered, and the wind turbine of offshore wind farm is prompted to stop and the ship is prompted to evacuate urgently; S12, if 15°< the path offset angle is less than or equal to 25° or 5 km / h< the moving speed increment is less than or equal to 10 km / h, a second-level early warning is triggered, and the wind turbine is prompted to start power reduction operation and the ship is notified to prepare to evacuate.

[0013] The system for inverting typhoon path based on the spectral characteristics of ocean ambient noise comprises: a collection module for collecting noise data and meteorological data of at least three nodes, once every 5 minutes; the noise data comprises 8-20Hz infrasound signals, and the meteorological data comprises air pressure data, wind speed data and temperature data; an acoustic signal screening module for screening and processing the noise data to obtain effective infrasound signals; an acoustic signal analysis module for performing spectral analysis and feature extraction on the effective infrasound signals to obtain infrasound features, the infrasound features comprising a center frequency offset and a frequency band energy ratio; a meteorological screening module for screening and verifying the meteorological data to obtain accurate air pressure data, wind speed data and temperature data; a meteorological calculation module for calculating the air pressure gradient and the wind shear index as meteorological parameters according to the accurate air pressure data and wind speed data; a timestamp unification module for unifying the timestamps of the infrasound features and the meteorological parameters; a neural network module for inputting the timestamp-unified infrasound features and meteorological parameters as feature values into an LSTM-Attention feature fusion model, the LSTM-Attention feature fusion model outputting the coordinates of the typhoon center and the moving speed of the typhoon; an output module for averaging the typhoon coordinates and the moving speed of the typhoon obtained by each node to obtain the final coordinates of the typhoon center and the final moving speed of the typhoon; a correction module for introducing ensemble Kalman filtering to fuse satellite cloud image data, and correcting the errors of the final coordinates of the typhoon center and the final moving speed of the typhoon through dynamic weight distribution; and an early warning module for triggering a first-level early warning and prompting the shutdown of wind turbines of offshore wind farms and the emergency evacuation of ships if the path offset angle is greater than 25° or the moving speed increment is greater than 10km / h, and triggering a second-level early warning and prompting the start of wind turbine power reduction operation and the notification of ship evacuation preparation if 15°<path offset angle≤25° or 5km / h<moving speed increment≤10km / h; in the neural network module, the process of the LSTM-Attention feature fusion model outputting the coordinates of the typhoon center and the moving speed of the typhoon according to the infrasound features and the meteorological parameters comprises: performing time series data format conversion on the infrasound features and the meteorological parameters, and performing data normalization verification and adjustment to obtain a data tensor; inputting the data tensor into a first-layer LSTM for feature extraction to output a first-layer LSTM feature vector; inputting the data tensor into a second-layer LSTM for feature extraction to output a deep-layer feature vector; performing attention weight calculation, normalizing the attention weight, and performing feature weighting, splicing the Attention output result and the deep-layer feature vector to obtain a fused feature vector; initializing a full connection network, inputting the fused feature vector into the full connection network, performing full connection layer feature mapping, and outputting the coordinates of the typhoon center and the moving speed of the typhoon.

[0014] In the application, preferably, the sound signal screening module comprises: an invalid elimination unit for eliminating invalid data and correcting the infrasound signal amplitude according to the ambient temperature at the collection time; a denoising unit for processing the infrasound signal by using a wavelet threshold denoising algorithm; and a verification unit for verifying the signal-to-noise ratio of the infrasound signal, correcting the infrasound signal segment higher than the signal-to-noise ratio threshold, and obtaining the effective infrasound signal; the sound signal analysis module comprises: a decomposition unit for wavelet packet decomposition of the effective infrasound signal to generate a plurality of sub-bands and calculate the energy of each sub-band; a statistical unit for statistically counting the energy proportion of all sub-bands and sorting them from large to small according to the energy proportion; a merging unit for accumulating the energy proportion and extracting the sub-band with a cumulative proportion greater than 80% or the sub-band of 11-14 Hz and merging them into a main frequency band; a center frequency offset calculation unit for calculating the center frequency f_current of the main frequency band, comparing it with the reference center frequency f_benchmark in the absence of a typhoon, and obtaining the center frequency offset Δf = f_current - f_benchmark; and a frequency band energy ratio calculation unit for calculating the total energy E 主 of the main frequency band, calculating the proportion of the total energy E （8-20）Hz in the full-band, and obtaining the frequency band energy ratio R = E 主 / E （8-20）Hz ; the weather screening module comprises: an abnormality elimination unit for eliminating abnormal values in the pressure data, wind speed data and temperature data; a missing value filling unit for filling missing values in the pressure data, wind speed data and temperature data; and a verification unit for consistency verification of the pressure data, wind speed data and temperature data, wherein the data passing the verification is accurate data; in the weather calculation module, the pressure gradient is calculated as follows: the straight-line distance D_ij between adjacent nodes is calculated, the pressure difference ΔP_ij of adjacent nodes is calculated, the pressure gradient _ij = ΔP_ij / D_ij of each adjacent node is obtained, the maximum value is taken as the pressure gradient of the collection period, and 2 decimal places are reserved; the wind shear index α is calculated as follows: for each node, the 10m height wind speed v 10 , the 100m height wind speed v 100 are obtained, the logarithmic wind profile model is used to calculate α, α = ln(v 100 / v 10 ) / ln (100 / 10) = ln (v 100 / v 10 ) / 2.303, if v 10 =0, then α takes 0.25, and if v 100 ≤v 10If the wind speed data is not collected, the time stamp unification module re-collects the wind speed data; the time stamp unification module comprises: a unification unit configured to extract the collection time stamps of infrasound wave features and meteorological parameters, and unify the collection time stamps into a "year-month-day-hour-minute-second" format; and a misalignment elimination unit configured to determine that the time is misaligned if the time stamp difference between the infrasound wave features and the meteorological parameters in the same collection period is greater than 1 minute, and eliminate the data in the period.

[0015] Compared with the prior art, the method and system have the following beneficial effects:

[0016] The method and system of the present application establish a physical coupling mechanism of "typhoon infrasound wave-air pressure gradient-sea surface wind field" by coupling modeling of infrasound wave features (8-20 Hz) of ocean ambient noise and meteorological parameters, realize 6-12 hours early warning of typhoon path deviation, supplement near-shore monitoring blind area data by using 8-20 Hz ocean ambient noise infrasound waves (low attenuation and long-distance propagation characteristics), build multi-dimensional data input by coupling infrasound wave spectrum features and meteorological parameters, improve the comprehensive representation ability of the model to typhoon dynamics, and realize real-time inversion by using a hybrid architecture of LSTM-Attention+ensemble Kalman filter, which takes into account both time series dependence capture and prediction bias correction, to provide "instant strong wind event" emergency response decision support (such as wind turbine shutdown and ship evacuation) for a wind farm. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the typhoon path inversion method based on the spectrum features of ocean ambient noise according to an embodiment of the present application.

[0018] Figure 2 The flowchart of S3 in the typhoon path inversion method based on the spectrum features of ocean ambient noise according to an embodiment of the present application.

[0019] Figure 3 The flowchart of S4 in the typhoon path inversion method based on the spectrum features of ocean ambient noise according to an embodiment of the present application.

[0020] Figure 4 The flowchart of S5 in the typhoon path inversion method based on the spectrum features of ocean ambient noise according to an embodiment of the present application.

[0021] Figure 5 The flowchart of S7 in the typhoon path inversion method based on the spectrum features of ocean ambient noise according to an embodiment of the present application.

[0022] Figure 6 The flowchart of S8 in the typhoon path inversion method based on the spectrum features of ocean ambient noise according to an embodiment of the present application.

[0023] Figure 7 The flowchart of the typhoon path inversion method based on the spectrum features of ocean ambient noise according to another embodiment of the present application.

[0024] Figure 8 Figure 1 is a structural schematic diagram of a typhoon path inversion system based on the spectral characteristics of ocean ambient noise according to another embodiment of the present application.

[0025] In the drawings: 1, acquisition module; 2, acoustic signal screening module; 3, acoustic signal analysis module; 4, weather screening module; 5, weather calculation module; 6, timestamp unification module; 7, neural network module; 8, output module; 9, correction module; 10, early warning module. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0027] It should be noted that when a component is referred to as being "fixed" to another component, it can be directly on the other component or there can be intervening components. When a component is referred to as being "connected" to another component, it can be directly connected to the other component or there can be intervening components. When a component is referred to as being "disposed" on another component, it can be directly on the other component or there can be intervening components. The terms "vertical", "horizontal", "left", "right", and similar terms as used herein are for purposes of description only.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0029] Referring to Figure 1 A typhoon path inversion method based on the spectral characteristics of ocean ambient noise is provided in a preferred embodiment of the present application, comprising:

[0030] A typhoon path inversion method based on the spectral characteristics of ocean ambient noise, comprising:

[0031] S1, constructing an LSTM-Attention feature fusion model.

[0032] The LSTM-Attention feature fusion model takes four-dimensional time series features as input, takes the coordinates (longitude, latitude) of the typhoon center and the typhoon moving speed as output, and adopts a double-layer LSTM+Attention architecture.

[0033] Input layer structure: The normalized four-dimensional features (Δf, R, , α) are reconstructed into a time series input matrix with dimensions [batch_size, 12, 4] (batch_size=32, optimized based on GPU memory) according to the time steps timesteps=12 (corresponding to 1 hour of historical data, 5 minutes / time * 12 times = 60 minutes). Among them, Δf is the center frequency offset, R is the frequency band energy ratio, is the pressure gradient, and α is the wind shear index.

[0034] First-layer LSTM (64 units) structure: Its function is to extract the short-term time series dependence of individual features (such as the continuous offset trend of Δf within 1 hour); the activation function uses tanh (hidden layer) + sigmoid (gate), dropout=0.2 (to prevent overfitting), and returns the hidden state of each time step (dimension [32, 12, 64]).

[0035] Second-layer LSTM (32 units) structure: Its function is to capture the deep dependence between multiple features (such as the cooperative change of Δf and ); the activation function and dropout are the same as the first layer, the input is the output of the first-layer LSTM, and the hidden state of the final time step is returned (dimension [32, 32]).

[0036] Attention mechanism enhancement: Its function is to focus on the moment of Δf mutation (typhoon turning precursor, Δf mutation is defined as the absolute value of the difference between adjacent time steps ≥2Hz); its calculation process is as follows:

[0037] (1) Define the Query vector (32 dimensions, trainable), Key vector (first-layer LSTM output, [32, 12, 64]), and Value vector (consistent with Key);

[0038] (2) Calculate the attention score: score(Q, K_i)=W_v・tanh(W_q・Q + W_k・K_i) (W_q / W_k / W_v are trainable weight matrices);

[0039] (3) Mutation time step weight enhancement: multiply the score of the Δf mutation time step by a factor of 1.5;

[0040] (4) Normalization and weighting: normalize the weight with softmax, and weighted sum to get the Attention output (dimension [32, 32]).

[0041] Output layer: Fully connected network (16 units ReLU activation → 3 units linear output), output the longitude (λ, °), latitude (φ, °), and moving speed (v, km / h) of the typhoon center in the future 1-6 hours.

[0042] Training data: Typhoon monitoring data of a certain wind farm for one year (365 days) (containing more than 10,000 time series samples).

[0043] Training parameters: MSE loss function (suitable for continuous value prediction), Adam optimizer (initial learning rate 0.001, decay 10% every 5 epochs), early stopping strategy (stop if the validation set MSE does not decrease for 5 consecutive epochs, save the optimal model).

[0044] Model deployment: Export in ONNX format and deploy to an edge computing node (such as a wind farm control room server), single sample inference time ≤ 0.1 seconds.

[0045] S2, collect noise data and meteorological data from at least 3 nodes, every 5 minutes; noise data includes 8-20Hz infrasound signals, meteorological data includes air pressure data, wind speed data and temperature data.

[0046] In this embodiment, the application scenario is a certain offshore wind farm, and 4 nodes are used for data collection. 4 noise measuring instruments and 4 meteorological stations are placed in the corner positions of the wind farm area to form a noise measurement and meteorological parameter collection matrix. The noise measuring instrument (sampling rate 100Hz, center frequency 10Hz) is used to capture 8-20Hz infrasound; the meteorological station is used to monitor air pressure, temperature, humidity and wind speed data.

[0047] Start the noise measuring instrument of the 4 nodes, ensure that the equipment is in an unobstructed and electromagnetic interference-free state, and collect 8-20Hz infrasound signals in real time. Save the time domain data every 5 minutes automatically, and record the environmental temperature at the time of collection (for subsequent signal temperature compensation).

[0048] Synchronize the start of the 4-node multi-parameter meteorological station, the 10m / 100m height wind speed sensor, the air pressure sensor, and the temperature and humidity sensor work simultaneously. After collection is completed, automatically mark the data integrity (such as "complete", "missing wind speed", "missing air pressure") to facilitate subsequent screening.

[0049] S3, screen and process the noise data to obtain effective infrasound signals.

[0050] Specifically, as shown in Figure 2 S3 includes:

[0051] S31, invalid data is removed, and the infrasound signal amplitude is corrected according to the ambient temperature at the collection time.

[0052] Invalid data is removed, that is, data of device failure (such as sensor offline and signal amplitude of 0) and environmental interference exceeding the standard (such as 1-5Hz strong vibration signal generated by sudden fan failure) at the collection time is deleted.

[0053] The infrasound signal amplitude is corrected according to the ambient temperature at the collection time, and the signal amplitude correction coefficient is 0.98 when the temperature changes by 1°C. The formula is: A_compensated=A_original*0.98^(T-25), T is the ambient temperature, and A_compensated is the corrected value of the infrasound signal amplitude, and A_original is the original value of the infrasound signal amplitude.

[0054] S32, the wavelet threshold denoising algorithm is used to process the infrasound signal.

[0055] The wavelet threshold denoising algorithm is used, and the specific steps are as follows:

[0056] (1) db4 wavelet base is selected, 5-minute time domain signal (30000 data points) is decomposed by 4 layers of wavelet, and approximate coefficients and detail coefficients are obtained.

[0057] (2) The detail coefficients are processed by "soft threshold" (threshold = 0.5*max(detail coefficient absolute value)) to suppress high-frequency noise (such as 50Hz power frequency interference).

[0058] (3) The signal is reconstructed: the approximate coefficients and the detail coefficients after processing are inversely transformed by wavelet to obtain the 8-20Hz pure infrasound signal after denoising.

[0059] S33, the signal-to-noise ratio of the infrasound signal is verified, and the infrasound signal segment higher than the signal-to-noise ratio threshold is corrected to obtain the effective infrasound signal.

[0060] Specifically, the 5-minute signal after denoising is segmented by 1 minute to obtain 5 sub-signals (6000 data points per segment), the signal-to-noise ratio (SNR) of each sub-signal is calculated, and the sub-signal with SNR<10dB is determined as invalid, and the mean value of the effective sub-signals before and after it is replaced to ensure that each signal meets the subsequent spectrum analysis requirements.

[0061] S4, the effective infrasound signal is subjected to spectrum analysis and feature extraction to obtain infrasound features, including center frequency offset and frequency band energy ratio.

[0062] Specifically, as shown in Figure 3 S4 includes:

[0063] S41, perform wavelet packet decomposition on the effective infrasound signal to generate several sub-bands, and calculate the energy of each sub-band.

[0064] In wavelet packet decomposition, each effective sub-signal segment undergoes 6 levels of wavelet packet decomposition, generating 2^6=64 sub-bands covering a frequency range of 8-20Hz. The energy of each sub-band is calculated (formula: E=Σ|x_i|). 2 (where x_i is the signal sampling point within the sub-frequency band).

[0065] S42, calculate the energy percentage of all sub-bands and sort them from largest to smallest energy percentage.

[0066] S43, cumulative energy percentage, extract sub-bands with a cumulative percentage >80% or extract sub-bands of 11-14Hz, and merge them into the main frequency band.

[0067] In practical applications, the 11-14Hz frequency band can be used as the default main frequency band. If the cumulative proportion is not met, the frequency band range can be dynamically adjusted.

[0068] S44. Calculate the center frequency f_current of the main frequency band and compare it with the reference center frequency f_benchmark when there is no typhoon to obtain the center frequency offset Δf = f_current - f_benchmark.

[0069] S45, Calculate the total energy E of the main frequency band. 主 Total energy E across the entire frequency band （8-20）Hz The proportion of the frequency band energy ratio R=E is obtained by calculating the proportion of the frequency band energy ratio. 主 / E （8-20）Hz .

[0070] S5 filters and verifies meteorological data to obtain accurate air pressure, wind speed, and temperature data.

[0071] Specifically, such as Figure 4 As shown, S5 includes:

[0072] S51 removes outliers from air pressure, wind speed, and temperature data.

[0073] Using the three-standard-deviation principle, data with air pressure greater than 1050 hPa or less than 950 hPa, wind speed greater than 50 m / s, temperature less than -10℃ or greater than 40℃, and humidity less than 0% or greater than 100% were deleted.

[0074] S52 fills in missing values ​​in air pressure, wind speed, and temperature data.

[0075] The rules for filling in the gaps are:

[0076] (1) If a parameter of a single node is missing (e.g., air pressure is missing), the "distance weighted mean" of the parameters of the remaining three nodes (weight = 1 / linear distance between nodes, sum normalized to 1) is calculated;

[0077] (2) If multiple nodes are missing the same parameter (less than or equal to two nodes), a "time series interpolation method" (based on the effective data of the previous and subsequent five collection periods) is used;

[0078] (3) If more than three nodes are missing the same parameter, mark this collection period as "invalid meteorological data", suspend subsequent meteorological parameter calculation, and wait for the next collection period.

[0079] S53, consistency check is performed on the air pressure data, wind speed data and temperature data, and the data passing the check is accurate data.

[0080] Specifically, compare the 10m and 100m height wind speeds of the same node. Under normal circumstances, the 100m wind speed > 10m wind speed (difference > 0.5m / s). If it does not meet the condition, the wind speed data of the node is re-collected. Compare the air pressures of adjacent nodes. The difference is normally within the range of 0.1-2hPa. If it exceeds the range, check whether the weather station has problems such as air leakage and obstruction, and re-collect after repair.

[0081] S6, according to the accurate air pressure data and wind speed data, the air pressure gradient and the wind shear index are calculated as meteorological parameters.

[0082] Specifically, the air pressure gradient is calculated as follows:

[0083] Calculate the linear distance D_ij between adjacent nodes, calculate the air pressure difference ΔP_ij of adjacent nodes, and get the air pressure gradient _ij = ΔP_ij / D_ij, take the maximum value as the air pressure gradient of this collection period, and keep 2 decimal places.

[0084] The calculation method of the wind shear index a is as follows:

[0085] For each node, get the 10m height wind speed v 10 , 100m height wind speed v 100 , calculate a according to the logarithmic wind profile model, a = ln (v 100 / v 10 ) / ln (100 / 10) = ln (v 100 / v 10 ) / 2.303, if v 10 = 0, then a = 0.25, if v 100 ≤ v 10 , then re-collect the wind speed data.

[0086] S7, timestamping the infrasound wave features and the meteorological parameters.

[0087] Specifically, as shown in Figure 5 S7 includes:

[0088] S71, extracting the collection timestamps of the infrasound wave features and the meteorological parameters, and converting them into the format of "year-month-day-hour: minute: second".

[0089] S72, if the timestamp difference between the infrasound wave features and the meteorological parameters in the same collection period is greater than 1 minute, determining that the time is inaccurate, and removing the data of the period.

[0090] S8, inputting the infrasound wave features and the meteorological parameters after timestamping as feature values into the LSTM-Attention feature fusion model, and outputting the coordinates of the typhoon center and the typhoon moving speed by the LSTM-Attention feature fusion model.

[0091] Specifically, as shown in Figure 6 In S8, the process of outputting the coordinates of the typhoon center and the typhoon moving speed by the LSTM-Attention feature fusion model according to the infrasound wave features and the meteorological parameters includes:

[0092] S81, performing time series data format conversion on the infrasound wave features and the meteorological parameters, and performing data normalization verification and adjustment to obtain a data tensor.

[0093] Receiving the preprocessed four-dimensional feature data (Δf, R, , α), each feature containing continuous time series data. According to the time step timesteps=12 (corresponding to 1 hour of historical data, 1 data point every 5 minutes), the feature data is cut into fixed-length time series segments. The input data tensor is constructed, with a shape of [number of samples, timesteps=12, number of features=4], where each sample represents a feature sequence of 12 consecutive time steps.

[0094] Check the normalization of the input data to ensure that all feature values are within the range of [-3, 3] (consistent with the Z-Score standardization result). Truncate the abnormal values outside the range (values less than -3 are set to -3, and values greater than 3 are set to 3). Convert the input data into a tensor format that can be processed by the model (such as Tensor in PyTorch or Tensor in TensorFlow), and set the data type to float32.

[0095] S82, inputting the data tensor into the first layer LSTM for feature extraction, and outputting the first layer LSTM feature vector.

[0096] Initialize the first-layer LSTM network (64 neurons) with the following parameters:

[0097] Input dimension: 4 (corresponding to 4 features);

[0098] Hidden layer dimension: 64;

[0099] Activation functions: tanh for hidden states and sigmoid for gate units, Dropout rate: 0.2 (to prevent overfitting);

[0100] Return sequences: True (return outputs at each time step for subsequent Attention mechanism).

[0101] Input the input tensor [samples, 12, 4] into the first-layer LSTM for forward propagation: calculate the activation values of the forget gate, input gate, and output gate to dynamically control information retention and update; update the cell state and hidden state to capture short-term temporal dependencies in the feature sequence.

[0102] Output the result of the first-layer LSTM with a shape of [samples, 12, 64], i.e., a 64-dimensional feature vector at each time step, which is the first-layer LSTM feature vector.

[0103] S83, input the data tensor into the second-layer LSTM for feature extraction, and output the deep feature vector.

[0104] Initialize the second-layer LSTM network (32 neurons) with the following parameters:

[0105] Input dimension: 64 (receive the output of the first-layer LSTM);

[0106] Hidden layer dimension: 32;

[0107] Activation functions: consistent with the first layer (tanh + sigmoid);

[0108] Dropout rate: 0.2;

[0109] Return sequences: False (only return the output at the last time step).

[0110] Input the output of the first-layer LSTM [samples, 12, 64] into the second-layer LSTM for forward propagation: further learn the long-term dependencies in the feature sequence and integrate the relevance between multiple features; compress the feature dimension from 64 to 32.

[0111] Output the result of the second-layer LSTM with a shape of [samples, 32], i.e., a 32-dimensional deep feature vector for each sample.

[0112] S84, Attention weight calculation is performed, attention weight is normalized, feature weighting is performed, the Attention output result is spliced with the deep feature vector, and a fused feature vector is obtained.

[0113] Define the key vector of the Attention mechanism:

[0114] Query vector: a 32-dimensional trainable vector randomly initialized, representing the feature type that the model focuses on;

[0115] Key vector: the output of the first layer LSTM [sample number, 12, 64], representing the feature information of each time step;

[0116] Value vector: same as Key vector, i.e. the output of the first layer LSTM [sample number, 12, 64].

[0117] Calculate the similarity (attention score) between Query and each time step Key: map Query and Key to the same dimension through linear transformation; use additive attention to calculate the score.

[0118] Enhance the attention score at the Δf mutation moment: detect the mutation moment of Δf in the input sequence (the absolute value of the difference between adjacent time steps Δf is greater than or equal to 2Hz); multiply the attention score at the mutation moment by a weight of 1.5 times to enhance the influence of the key signal.

[0119] Use the softmax function to normalize the attention scores of the 12 time steps to obtain the weight value of each time step, which is in the shape of [sample number, 12], and the sum of the weights is 1; based on the attention weight, the Value vector is weighted and summed to obtain the output of the Attention mechanism, which is in the shape of [sample number, 64]; the Attention output is spliced with the output of the second layer LSTM to obtain a fused feature vector, which is in the shape of [sample number, 64+32=96].

[0120] S85, initialize the fully connected network, input the fused feature vector into the fully connected network, perform feature mapping of the fully connected layer, and output the coordinates of the typhoon center and the typhoon moving speed.

[0121] Specifically, initialize the fully connected network:

[0122] First layer (hidden layer): 16 neurons, using ReLU activation function;

[0123] Second layer (output layer): 3 neurons, no activation function (output continuous value).

[0124] Input the fused feature vector [sample number, 96] into the fully connected network:

[0125] Hidden layer calculation: 96-dimensional features are mapped to 16-dimensional features through linear transformation and ReLU activation;

[0126] Output layer calculation: 16-dimensional features are mapped to 3-dimensional features corresponding to 3 predicted targets;

[0127] Output the prediction results in the shape of [sample number, 3], which respectively represent:

[0128] the longitude (°) of the typhoon center in the future 1-6 hours;

[0129] the latitude (°) of the typhoon center in the future 1-6 hours;

[0130] the moving speed (km / h) of the typhoon in the future 1-6 hours.

[0131] S9, the typhoon coordinates and the typhoon moving speed obtained by each node are averaged to obtain the final coordinates of the typhoon center and the final typhoon moving speed.

[0132] In a preferred embodiment of the present application, as shown in Figure 7 the typhoon path inversion method based on the spectral characteristics of ocean ambient noise further comprises:

[0133] S10, satellite cloud image data is introduced to fuse the set Kalman filter, and the errors of the final coordinates of the typhoon center and the final typhoon moving speed are corrected through dynamic weight distribution.

[0134] When the typhoon eye wall has strong rainfall, the model prediction is prone to deviation, and satellite cloud image data (updated every 30 minutes) needs to be fused to correct:

[0135] (1) taking the prediction results (i.e. the final coordinates of the typhoon center and the final typhoon moving speed) of the LSTM-Attention as the initial value, and the typhoon position extracted from the satellite cloud image as the observation value;

[0136] (2) calculating the prediction error covariance and the observation error covariance, and dynamically distributing the weight (the weight of the satellite data is 0.3 when there is no rainfall, and the weight is 0.5 when there is rainfall);

[0137] (3) outputting the corrected typhoon path to ensure that the prediction error is less than or equal to 20 km.

[0138] In a preferred embodiment of the present application, as shown in Figure 7 the typhoon path inversion method based on the spectral characteristics of ocean ambient noise further comprises:

[0139] S11, if the path offset angle is greater than 25° or the moving speed increment is greater than 10 km / h, a first-level warning is triggered, prompting the wind turbine of the offshore wind farm to shut down and the ship to evacuate urgently.

[0140] S12, if 15° < path offset angle ≤ 25° or 5 km / h < speed increment ≤ 10 km / h, triggering a secondary early warning, prompting to start the fan to reduce power operation, and informing the ship to prepare to evacuate.

[0141] In another embodiment of the present application, a typhoon path inversion system based on the spectral characteristics of ocean ambient noise is also provided, as shown in the figure, comprising: Figure 8

[0142] The acquisition module 1 is used for acquiring noise data and meteorological data of at least three nodes, and the acquisition is performed once every 5 minutes; the noise data includes 8-20Hz infrasound signals, and the meteorological data includes air pressure data, wind speed data and temperature data.

[0143] The sound signal screening module 2 is used for screening and processing the noise data to obtain effective infrasound signals.

[0144] The sound signal screening module 2 comprises:

[0145] The invalid elimination unit is used for eliminating invalid data, and the amplitude of the infrasound signal is corrected according to the ambient temperature at the acquisition time;

[0146] The denoising unit is used for processing the infrasound signal by using a wavelet threshold denoising algorithm;

[0147] The verification unit is used for verifying the signal-to-noise ratio of the infrasound signal, correcting the infrasound signal segment higher than the signal-to-noise ratio threshold, and obtaining the effective infrasound signal.

[0148] The sound signal analysis module 3 is used for performing spectral analysis and feature extraction on the effective infrasound signal to obtain infrasound features, and the infrasound features include a center frequency offset and a frequency band energy ratio.

[0149] The sound signal analysis module 3 comprises:

[0150] The decomposition unit is used for performing wavelet packet decomposition on the effective infrasound signal to generate a plurality of sub-bands, and calculating the energy of each sub-band;

[0151] The statistical unit is used for statistically calculating the energy proportion of all sub-bands, and sorting the sub-bands according to the energy proportion from large to small;

[0152] The merging unit is used for accumulating the energy proportion, extracting the sub-band with a cumulative proportion greater than 80% or extracting the sub-band of 11-14Hz, and merging the sub-bands into a main frequency band;

[0153] The center frequency offset calculation unit is used for calculating the center frequency f_current of the main frequency band, comparing the center frequency f_current with the reference center frequency f_benchmark in the absence of a typhoon, and obtaining the center frequency offset Δf = f_current - f_benchmark.​

[0154] a band energy ratio calculation unit for calculating a total energy E of a main frequency band 主 a ratio in the total energy E of the full frequency band （8-20）Hz , obtaining a band energy ratio R = E 主 / E （8-20）Hz .

[0155] a weather screening module 4 for screening and verifying weather data to obtain accurate barometric pressure data, wind speed data and temperature data.

[0156] The weather screening module 4 comprises:

[0157] an abnormal value elimination unit for eliminating abnormal values in the barometric pressure data, wind speed data and temperature data;

[0158] a missing value filling unit for filling missing values in the barometric pressure data, wind speed data and temperature data;

[0159] a verification unit for performing consistency verification on the barometric pressure data, wind speed data and temperature data, and the data passing the verification is accurate data.

[0160] a weather calculation module 5 for calculating a barometric pressure gradient and a wind shear index as weather parameters according to the accurate barometric pressure data and wind speed data.

[0161] In the weather calculation module 5, the barometric pressure gradient is calculated in the following manner:

[0162] a straight-line distance D_ij between adjacent nodes is calculated, a barometric pressure difference ΔP_ij of adjacent nodes is calculated, and a barometric pressure gradient _ij = ΔP_ij / D_ij of each adjacent node is obtained, the maximum value is taken as the barometric pressure gradient of the collection period, and 2 decimal places are reserved;

[0163] The calculation method of the wind shear index a is as follows:

[0164] For each node, a 10m height wind speed v 10 , a 100m height wind speed v 100 is obtained, a is calculated according to a logarithmic wind profile model, a = ln (v 100 / v 10 ) / ln (100 / 10) = ln (v 100 / v 10 ) / 2.303, if v 10 = 0, a is taken as 0.25, and if v 100 ≤ v 10 , wind speed data is re-collected.

[0165] The timestamp unification module 6 is used to unify the timestamps of infrasound characteristics and meteorological parameters.

[0166] Timestamp unification module 6 includes:

[0167] A unified unit is used to extract the acquisition timestamps of infrasound features and meteorological parameters, and convert them into a unified format of "year-month-day-hour:minute:second".

[0168] The inaccuracy elimination unit is used to determine that if the difference between the timestamp of the infrasound characteristics and the meteorological parameters is greater than 1 minute within the same acquisition period, it is considered to be time inaccuracy and the data of that period is eliminated.

[0169] Neural network module 7 is used to input the infrasound features and meteorological parameters after unifying the timestamps as feature values ​​into the LSTM-Attention feature fusion model. The LSTM-Attention feature fusion model outputs the coordinates of the typhoon center and the typhoon's movement speed.

[0170] Output module 8 is used to calculate the average value of the typhoon coordinates and typhoon movement speed obtained from each node, and use it as the final coordinates and final typhoon movement speed of the typhoon center.

[0171] Correction module 9 is used to introduce ensemble Kalman filter fusion satellite cloud image data and correct the errors in the final coordinates of the typhoon center and the final typhoon movement speed through dynamic weight allocation.

[0172] The early warning module 10 is used to trigger a Level 1 early warning if the path deviation angle is greater than 25° or the speed increment is greater than 10km / h, prompting the wind turbines of the offshore wind farm to shut down and the ships to evacuate urgently; if 15° < path deviation angle ≤ 25° or 5km / h < speed increment ≤ 10km / h, it triggers a Level 2 early warning, prompting the wind turbines to start operating at reduced power and notifying the ships to prepare for evacuation.

[0173] In neural network module 7, the process by which the LSTM-Attention feature fusion model outputs the coordinates of the typhoon center and the typhoon's movement speed based on infrasound features and meteorological parameters includes:

[0174] The infrasound characteristics and meteorological parameters are converted into time-series data formats, and the data is normalized, verified, and adjusted to obtain data tensors;

[0175] Input the data tensor into the first LSTM layer for feature extraction, and output the first LSTM layer feature vector.

[0176] The data tensor is input into the second LSTM layer for feature extraction, and the deep feature vector is output.

[0177] The attention weight calculation is performed, the attention weight is normalized, feature weighting is performed, the attention output result is spliced with the deep feature vector, and a fused feature vector is obtained;

[0178] The initialized full connection network is used to input the fused feature vector into the full connection network, full connection layer feature mapping is performed, and the coordinates of the typhoon center and the typhoon moving speed are output.

[0179] The above description is a detailed description of the preferred embodiments of the application, but the embodiments are not intended to limit the scope of the patent application. Any equivalent changes or modifications made under the technical spirit of the application should be included in the scope of the patent.

Claims

1. A method for retrieving typhoon track based on spectral characteristics of ocean ambient noise, characterized in that, The application relates to a method for tracking a typhoon center, comprising the following steps: S1, constructing an LSTM-Attention feature fusion model; S2, collecting noise data and meteorological data of at least three nodes, once every 5 minutes; the noise data comprises 8-20Hz infrasound signals, and the meteorological data comprises air pressure data, wind speed data and temperature data; S3, performing screening processing on the noise data to obtain effective infrasound signals; S4, performing frequency spectrum analysis and feature extraction on the effective infrasound signals to obtain infrasound features, wherein the infrasound features comprise a center frequency offset and a frequency band energy ratio; S5, performing screening and verification on the meteorological data to obtain accurate air pressure data, wind speed data and temperature data; S6, calculating an air pressure gradient and a wind shear index according to the accurate air pressure data and the wind speed data as meteorological parameters; S7, performing time stamp unification on the infrasound features and the meteorological parameters; S8, inputting the infrasound features and the meteorological parameters after time stamp unification as feature values into the LSTM-Attention feature fusion model, and outputting coordinates of a typhoon center and a typhoon moving speed by the LSTM-Attention feature fusion model; S9, calculating average values of the typhoon coordinates and the typhoon moving speed obtained by each node as final coordinates of the typhoon center and a final typhoon moving speed; In the S8, the process of outputting the coordinates of the typhoon center and the typhoon moving speed by the LSTM-Attention feature fusion model according to the infrasound features and the meteorological parameters comprises the following steps: S81, performing time sequence data format conversion on the infrasound features and the meteorological parameters, and performing data normalization verification and adjustment to obtain a data tensor; S82, inputting the data tensor into a first-layer LSTM to perform feature extraction, and outputting a first-layer LSTM feature vector; S83, inputting the data tensor into a second-layer LSTM to perform feature extraction, and outputting a deep-layer feature vector; S84, performing attention weight calculation, normalizing the attention weight, performing feature weighting, splicing an attention output result and the deep-layer feature vector, and obtaining a fused feature vector; S85, initializing a full connection network, inputting the fused feature vector into the full connection network, performing full connection layer feature mapping, and outputting the coordinates of the typhoon center and the typhoon moving speed; The S4 comprises the following steps: S41, performing wavelet packet decomposition on the effective infrasound signals to generate a plurality of sub-bands, and calculating energy of each sub-band; S42, counting energy proportions of all the sub-bands, and sorting the sub-bands according to the energy proportions from large to small; S43, accumulating the energy proportions, extracting a sub-band with a cumulative proportion greater than 80% or extracting a 11-14Hz sub-band, and merging the sub-bands into a main frequency band; S44, calculating a center frequency f_current of the main frequency band, comparing the center frequency f_current with a benchmark center frequency f_benchmark without a typhoon, and obtaining a center frequency offset Delta f = f_current - f_benchmark; S45, calculate the total energy E of the main frequency band 主 In the proportion of the total energy E of the full frequency band （8-20）Hz , get the frequency band energy ratio R = E 主 / E （8-20）Hz .

2. The typhoon track inversion method according to claim 1, characterized in that, The S3 comprises the following steps: S31, removing invalid data, and correcting an infrasound signal amplitude according to an ambient temperature at a collection time; S32, processing the infrasound signal by using a wavelet threshold denoising algorithm; S33, verifying the signal-to-noise ratio of the infrasound wave signal, correcting the infrasound wave signal segment higher than the signal-to-noise ratio threshold, and obtaining an effective infrasound wave signal.

3. The typhoon track inversion method according to claim 1, wherein, The S5 comprises: S51, removing outliers in the air pressure data, wind speed data and temperature data; S52, filling in missing values in the air pressure data, wind speed data and temperature data; S53, performing consistency verification on the air pressure data, wind speed data and temperature data, and the data passing the verification is accurate data.

4. The typhoon track inversion method of claim 1, wherein, In the S6, The calculation method of the air pressure gradient ∇P is: Calculate the straight-line distance D_ij between adjacent nodes, calculate the air pressure difference ΔP_ij of adjacent nodes, obtain the air pressure gradient ∇P_ij of each adjacent node = ΔP_ij / D_ij, and take the maximum value as the air pressure gradient ∇P of the collection period; The calculation method of the wind shear index α is: For each node, get the 10m height wind speed v 10 , 100m height wind speed v 100 , calculate a according to the logarithmic wind profile model, a = ln(v 100 / v 10 ) / ln (100 / 10) = ln(v 100 / v 10 ) / 2.303, if v 10 =0, then a takes 0.25, if v 100 ≤v 10 , then re-collect wind speed data.

5. The typhoon track inversion method according to claim 1, wherein, The S7 comprises: S71, extracting the collection time stamps of the infrasound wave features and the meteorological parameters, and uniformly converting them into the "year-month-day-hour: minute: second" format; S72, if the time stamp difference of the infrasound wave features and the meteorological parameters in the same collection period is greater than 1 minute, it is determined that the time is inaccurate, and the data of the period is removed.

6. The typhoon track inversion method according to claim 1, wherein, Further comprising: S10, introducing set Kalman filtering to fuse satellite cloud image data, and correcting the errors of the final coordinates of the typhoon center and the final typhoon moving speed through dynamic weight distribution.

7. The typhoon track inversion method according to claim 6, wherein, Further comprising: S11, if the path offset angle is greater than 25° or the speed increment is greater than 10 km / h, a first-level warning is triggered, prompting the wind turbine of the offshore wind farm to shut down and the ship to evacuate urgently; S12, if 15°< path offset angle ≤25° or 5 km / h< speed increment ≤10 km / h, a second-level warning is triggered, prompting the wind turbine to start power reduction operation and notifying the ship to prepare to evacuate.

8. A typhoon track inversion system based on the spectral characteristics of ocean ambient noise, characterized in that, Comprise: A collection module for collecting noise data and meteorological data of at least 3 nodes, once every 5 minutes; the noise data includes 8-20Hz infrasound wave signal, and the meteorological data includes air pressure data, wind speed data and temperature data; An acoustic signal screening module for screening and processing the noise data to obtain effective infrasound wave signal; An acoustic signal analysis module for performing frequency spectrum analysis and feature extraction on the effective infrasound wave signal to obtain infrasound wave features, the infrasound wave features including center frequency offset and frequency band energy ratio; A meteorological screening module for screening and verifying the meteorological data to obtain accurate air pressure data, wind speed data and temperature data; A meteorological calculation module for calculating the air pressure gradient and the wind shear index as meteorological parameters according to the accurate air pressure data and wind speed data; A time stamp unification module for unifying the time stamps of the infrasound wave features and the meteorological parameters; A neural network module for inputting the infrasound wave features and the meteorological parameters after time stamp unification as characteristic values into an LSTM-Attention feature fusion model, and outputting the coordinates of the typhoon center and the typhoon moving speed by the LSTM-Attention feature fusion model; An output module for averaging the typhoon coordinates and the typhoon moving speed obtained by each node to obtain the final coordinates of the typhoon center and the final typhoon moving speed. The correction module is used for introducing the ensemble Kalman filter to fuse satellite cloud image data, and correcting errors of final coordinates of the typhoon center and final moving speed of the typhoon through dynamic weight distribution; The early warning module is used for triggering a first-level early warning and prompting a shutdown of a wind turbine of a sea wind farm and an emergency evacuation of a ship if the path deviation angle is greater than 25° or the moving speed increment is greater than 10 km / h, and triggering a second-level early warning and prompting a start of a power reduction operation of the wind turbine and a notification of a preparation for evacuation of the ship if 15° is less than the path deviation angle and is less than or equal to 25° or 5 km / h is less than the moving speed increment and is less than or equal to 10 km / h; In the neural network module, a process in which the LSTM-Attention feature fusion model outputs the coordinates of the typhoon center and the moving speed of the typhoon according to the infrasound wave features and the meteorological parameters comprises: The infrasound wave features and the meteorological parameters are converted into a time series data format, and data normalization verification and adjustment are performed to obtain a data tensor; The data tensor is input into a first-layer LSTM for feature extraction, and a first-layer LSTM feature vector is output; The data tensor is input into a second-layer LSTM for feature extraction, and a deep-layer feature vector is output; Attention weight calculation is performed, the attention weight is normalized, feature weighting is performed, the Attention output result is spliced with the deep-layer feature vector, and a fused feature vector is obtained; A full-connection network is initialized, the fused feature vector is input into the full-connection network, full-connection layer feature mapping is performed, and the coordinates of the typhoon center and the moving speed of the typhoon are output; The sound signal screening module comprises: An invalid elimination unit is configured to eliminate invalid data and correct the amplitude of the infrasound wave signal according to the ambient temperature at the collection time; A denoising unit is configured to process the infrasound wave signal by using a wavelet threshold denoising algorithm; A verification unit is configured to verify the signal-to-noise ratio of the infrasound wave signal, correct the infrasound wave signal segment with a signal-to-noise ratio higher than a threshold, and obtain valid infrasound wave signals; The sound signal analysis module comprises: A decomposition unit is configured to perform wavelet packet decomposition on the valid infrasound wave signal, generate a plurality of sub-bands, and calculate the energy of each sub-band; A statistical unit is configured to statistically analyze the energy proportion of all the sub-bands and sort the sub-bands according to the energy proportion from large to small; A merging unit is configured to accumulate the energy proportion, extract a sub-band with a cumulative proportion greater than 80% or extract a sub-band of 11-14 Hz, and merge the sub-bands into a main frequency band; A center frequency offset calculation unit is configured to calculate the center frequency f_current of the main frequency band, compare the center frequency f_current with a benchmark center frequency f_benchmark in the absence of a typhoon, and obtain a center frequency offset Δf = f_current - f_benchmark; a band energy ratio calculating unit for calculating a total energy E of the main band 主 In the total energy E of the full band （8-20）Hz The proportion of the main band, the band energy ratio R = E 主 / E （8-20）Hz ; The meteorological screening module comprises: An abnormality elimination unit is configured to eliminate abnormal values in the pressure data, the wind speed data and the temperature data; A missing value filling unit is configured to fill in missing values in the pressure data, the wind speed data and the temperature data; A verification unit is configured to perform consistency verification on the pressure data, the wind speed data and the temperature data, and the data passing the verification is accurate data; In the meteorological calculation module, the pressure gradient ∇P is calculated in the following manner: A straight-line distance D_ij between adjacent nodes is calculated, a barometric pressure difference ΔP_ij of adjacent nodes is calculated, a barometric pressure gradient ∇P_ij of each adjacent node is obtained, a maximum value is taken as a barometric pressure gradient ∇P of the collection period, and 2 decimal places are reserved; The calculation method of the wind shear index a is: For each node, get the 10m height wind speed v 10 , 100m height wind speed v 100 , calculate a according to the logarithmic wind profile model, a = ln(v 100 / v 10 ) / ln (100 / 10) = ln(v 100 / v 10 ) / 2.303, if v 10 =0, then a takes 0.25, if v 100 ≤v 10 , then re-collect wind speed data; The timestamp unification module comprises: A unification unit is configured to extract the collection timestamps of the infrasound wave features and the meteorological parameters, and convert them into the "year-month-day-hour: minute: second" format; An out-of-alignment elimination unit is configured to determine that the time is out of alignment if the timestamp difference between the infrasound wave features and the meteorological parameters in the same collection period is greater than 1 minute, and eliminate the data in the period.

Images