A method and device for automatically identifying a tsunami based on a BiLSTM method

Abstract

The application relates to a BiLSTM-based tsunami automatic identification method and device, which comprises the following steps: astronomical tide harmonic analysis and band-pass filtering pretreatment are performed on water level station observation data to obtain a difference sequence; the difference sequence is input into BiLSTM in a sliding window mode, and normalization weights of hidden states at all times in the window are calculated through a learnable attention mechanism. On the basis of single-station identification, a multi-station space-time synchronous reasoning layer is further constructed, a tsunami propagation time delay matrix is pre-calculated according to shallow water wave speed and inter-station geographic distance, when the multi-station exceeds a threshold value, the multi-station is upgraded to a regional-level tsunami warning, otherwise, the multi-station is downgraded to a single-station abnormal record. The above method overcomes the defects of incomplete information and poor adaptability of a traditional one-way LSTM with a fixed threshold, improves the accuracy and physical interpretability of tsunami identification, and has the practical advantages of easy engineering deployment.

Description

Technical Field

[0001] This invention relates to the field of computer technology of artificial intelligence, and in particular to an automatic tsunami identification method and device based on BiLSTM. Background Technology

[0002] Tsunamis are highly destructive marine disasters caused by submarine earthquakes, submarine volcanic eruptions, submarine landslides, or other submarine geological activities. Historically, numerous major tsunami events (such as the 2004 Indian Ocean tsunami and the 2011 Tohoku earthquake and tsunami in Japan) have caused enormous casualties and property damage, posing a serious threat to coastal countries and regions. Therefore, efficient and accurate tsunami monitoring methods are of great significance for tsunami prevention and mitigation.

[0003] Existing tsunami monitoring technology relies on traditional statistical analysis methods based on tide gauge data. This involves deploying tide gauge stations near the coast to actively collect water level data after a tsunami event, comparing it with astronomical tide forecasts, and identifying anomalies by setting thresholds. However, traditional threshold methods depend on manual experience and are ill-suited to adapting to complex marine environmental changes. Furthermore, they are easily affected by non-tsunami signals such as storm surges and swells, leading to missed or false alarms. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a BiLSTM-based automatic tsunami identification method and device, which can effectively reduce the false alarm rate and improve the reliability of early warning.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] In a first aspect, embodiments of the present invention provide an automatic tsunami identification method based on BiLSTM, comprising:

[0009] S100: Obtain real-time observation data of two or more water level stations to be predicted within a preset time period, perform astronomical tide stripping and filtering on the real-time observation data, and obtain input data for each water level station used to input the BiLSTM tsunami time series prediction model.

[0010] The BiLSTM tsunami time series prediction model is a model trained on historical water level observation data of a designated water level station after astronomical tide stripping and filtering.

[0011] S200. Input the input data of each water level station into the BiLSTM tsunami time series prediction model, and obtain the prediction data of each water level station output by the BiLSTM tsunami time series prediction model.

[0012] S300: Based on the input data and prediction data of each water level station, obtain the prediction error sequence, smooth the prediction data sequence of all water level stations, calculate the attention enhancement score, and perform tsunami identification based on the attention enhancement score.

[0013] Optionally, S300 includes:

[0014] S301. Based on each water level station, subtract the input data and the predicted data of that water level station to obtain the prediction error sequence of that water level station.

[0015] S302. Using the error information obtained from the model training phase, the prediction error sequence is smoothed over time, and the time-smoothed sequence is smoothed using the exponentially weighted moving average method to obtain the first data sequence for identifying tsunamis.

[0016] S303. Calculate the attention enhancement score based on the first data sequence, obtain the ratio of the attention enhancement score to the tsunami threshold T, match the ratio with the ratio table of historical tsunami events, and trigger the first type of alarm information of the water level station.

[0017] Optionally, S300 further includes:

[0018] S304. When more than one water level station detects a Type I alarm, verify whether the difference in the abnormal trigger times of the station's observation data matches the theoretical tsunami propagation time. If they match, an automatic tsunami alarm is triggered. That is, when an abnormal water level is detected at a water level station, verify whether the difference in the abnormal trigger times of the stations matches the tsunami propagation time. Using the physical propagation laws of tsunamis as constraints, perform a spatiotemporal consistency check on the water level abnormality triggering sequence.

[0019] When the combined results of abnormal water level and the verification results of tsunami propagation time constraints meet the preset judgment conditions, a tsunami alarm will be automatically triggered.

[0020] Optionally, before step S303, the method further includes:

[0021] Obtain the tsunami threshold T. ;

[0022] , These are the global mean and standard deviation obtained during the training phase of the BiLSTM model. Here, the mean and standard deviation can be understood as the mean and standard deviation of the residual data after subtracting the astronomical tide from the water level data and applying bandpass filtering.

[0023] or,

[0024] S302 uses an exponentially weighted moving average method to smooth the time-smoothed sequence, obtaining the first data sequence for tsunami identification, including: ;

[0025] Where α is the smoothing coefficient. It is the EWMA value at time t. It is the prediction error at time t. It is the EWMA value from the previous moment;

[0026] Based on the above formula, we obtain the following: The first data sequence;

[0027] or,

[0028] S303 calculates the attention enhancement score based on the first data sequence. ,include:

[0029] ; ; ;

[0030] in, It is the prediction error at time t. It is the local energy at time t. It is the EWMA smoothing background value; Let be the attention weight at time t; It ultimately enhances abnormal scores. It is the attention enhancement coefficient. It is the original EWMA outlier score.

[0031] Optionally, prior to S200, the method further includes:

[0032] S00A. Based on the historical water level observation data of the designated water level station in the designated area, perform astronomical tide stripping and filtering to obtain the first training data for training the BiLSTM model.

[0033] S00B: Continuously sample the first training data to obtain sample pairs for time-supervised learning; based on the sample pairs, train the BiLSTM model using an end-to-end training method to obtain the BiLSTM tsunami time-series prediction model.

[0034] Optionally, the SOOB includes:

[0035] The trained BiLSTM tsunami time series prediction model is called to perform batch predictions on the validation set data in the sample pairs. The error value between the model prediction residual and the actual residual is calculated for each sample. The mean and standard deviation of the prediction error of the entire validation set are calculated.

[0036] Based on the Gaussian distribution hypothesis, the error value distribution pattern of the fitting validation set is used, and the anomaly detection threshold is automatically calibrated using percentile statistics to obtain the error distribution function;

[0037] The error distribution function can adapt to the differences in hydrological environment in different sea areas and at different water level stations.

[0038] The true residuals mentioned above are the residuals after subtracting the astronomical tide from the water level data and applying bandpass filtering, while the predicted residuals are the results output by BiLSTM.

[0039] Optionally, the SOOA includes:

[0040] Historical water level observation data of a specified water level station in a specified area with a time resolution of 1 minute are acquired; the historical water level observation data are fitted with tidal parameters using the harmonic analysis method to obtain historical water level observation data stripped of astronomical tides.

[0041] The historical water level observation data stripped of astronomical tides are filtered to obtain the first training data after interference signal filtering.

[0042] The first training data is normalized to obtain normalized water level time series data.

[0043] Optionally, SO0B includes:

[0044] For normalized water level time series data, a fixed-length sliding window sampling technique is used to construct sample pairs for time series supervised learning; and the sample pairs are divided into training set and validation set.

[0045] During the training of the BiLSTM model using an end-to-end training method

[0046] Mean squared error (MSE) is selected as the model loss function to quantify the deviation between the model's predicted values ​​and the true values; the Adam adaptive optimizer is used to iteratively update the model weights and bias parameters and adaptively adjust the learning rate.

[0047] The input sequence is fed into a BiLSTM. The forward LSTM processes the sequence sequentially from t=1 to t=N, while the reverse LSTM processes it in reverse order from t=N to t=1. The two states at each time step are fused together. The forward LSTM encodes the historical evolution up to the current time, while the reverse LSTM encodes the current position and subsequent trend. The two states are concatenated to give the model a global perception of the overall shape of the tsunami waveform.

[0048] In addition, embodiments of the present invention also provide an automatic tsunami identification device based on BiLSTM, which includes:

[0049] The data preprocessing module is used to acquire real-time observation data of two or more water level stations to be predicted within a preset time period, and to perform astronomical tide stripping and filtering on the real-time observation data to obtain input data for each water level station used to input the BiLSTM tsunami time series prediction model.

[0050] The BiLSTM tsunami time series prediction model is a model trained on historical water level observation data of a designated water level station after astronomical tide stripping and filtering.

[0051] The model output module is used to input the input data of each water level station into the BiLSTM tsunami time series prediction model and obtain the prediction data of each water level station output by the BiLSTM tsunami time series prediction model.

[0052] The tsunami identification module is used to obtain a prediction error sequence based on the input data and prediction data of each water level station, smooth the prediction data sequences of all water level stations, calculate the attention enhancement score, and perform tsunami identification based on the attention enhancement score.

[0053] Thirdly, embodiments of the present invention also provide an electronic device, comprising:

[0054] Memory, used to store programs or instructions;

[0055] A processor, which, when executing the program or instructions, implements the steps of the BiLSTM-based automatic tsunami identification method according to any one of the first aspects above.

[0056] (III) Beneficial Effects

[0057] The method of this invention can automatically learn the normal variation pattern of nearshore tide levels, overcoming the shortcomings of poor adaptability of traditional threshold methods;

[0058] In the method of this invention embodiment, the preprocessing method combining precise removal of astronomical tides with differential filtering effectively removes background noise and improves the signal-to-noise ratio of abnormal signals.

[0059] Furthermore, in this embodiment of the invention, a BiLSTM deep neural network is used to model the time series of tidal residuals, fully capturing the dynamic features of the time series and realizing the calculation of anomaly scores based on prediction errors.

[0060] In addition, an anomaly scoring smoothing mechanism is introduced in this embodiment of the invention to reduce instantaneous noise interference and improve the stability of detection results; at the same time, a spatiotemporal decision mechanism is established by combining the propagation time constraint of tsunami waves, which significantly reduces the false alarm rate and improves the reliability of the early warning system.

[0061] The method of this invention enables real-time online detection and automated tsunami warning output for practical business applications, providing technical support for coastal disaster prevention and mitigation decision-making. Attached Figure Description

[0062] Figure 1 A flowchart illustrating an automatic tsunami identification method based on BiLSTM provided in an embodiment of the present invention;

[0063] Figure 2A and Figure 2B These are schematic diagrams of an automatic tsunami identification device based on BiLSTM provided in another embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram showing the 2025 Kamchatka tsunami event after some of the steps described in this invention.

[0065] Figure 4 A schematic diagram illustrating the alarm generated using the method of this invention in response to the 2025 Kamchatka tsunami event;

[0066] Figure 5 Heatmap of attention weights for Midway, St. Louis, and Hilo, Hawaii. Detailed Implementation

[0067] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] In recent years, with the rapid development of deep learning technology, recurrent neural networks such as Bidirectional Long Short-Term Memory (BiLSTM) have demonstrated powerful capabilities in time series modeling and anomaly detection. BiLSTM networks can effectively capture long-range dependencies in time series data, and by learning from historical normal data, they can establish accurate models of normal behavior, thereby detecting abnormal deviations.

[0069] The existing technology has the following main drawbacks:

[0070] (1) Traditional threshold methods have poor adaptability. Existing tidal anomaly detection methods based on fixed thresholds cannot automatically adjust the discrimination parameters according to the hydrological and meteorological conditions of different sea areas and seasons, and have insufficient ability to identify abnormal signals in complex marine environments.

[0071] (2) Insufficient removal of astronomical tides and large residual noise interference. Existing methods rely on high-precision harmonic analysis for accurate subtraction of astronomical tides, but lack effective means to filter out complex background noise (such as storm surges and meteorological-driven water level changes), resulting in low signal-to-noise ratio of residual signals and increasing the risk of false detection.

[0072] (3) Lack of time series correlation modeling capability. Traditional statistical methods treat water level time series as independent samples, ignoring the time dependence and dynamic evolution law of water level changes, and cannot effectively model and predict the time series evolution of abnormal signals.

[0073] (4) It is difficult to monitor tsunamis caused by volcanic eruptions and submarine landslides. Existing conventional active tsunami wave monitoring methods cannot identify the sudden process of volcanic eruptions and submarine landslides in real time, and therefore it is difficult to automatically capture tsunami wave signals generated by such causes.

[0074] (5) Failure to fully utilize the physical constraints of tsunami propagation. Most existing detection methods do not take into account the physical laws of tsunami wave propagation speed and direction in the ocean, and cannot effectively utilize the consistency of spatiotemporal propagation to improve the reliability of early warning.

[0075] Example 1

[0076] This embodiment provides an automatic tsunami identification method based on BiLSTM, such as... Figure 1 As shown, the execution subject of the method in this embodiment is any electronic device, and the method in this embodiment may include the following steps:

[0077] S100: Obtain real-time observation data of two or more water level stations to be predicted within a preset time period, and perform astronomical tide stripping and bandpass filtering on the real-time observation data to obtain input data for each water level station used to input the BiLSTM tsunami time series prediction model.

[0078] The BiLSTM tsunami time series prediction model is a model trained on historical water level observation data of a designated water level station after astronomical tide stripping and filtering.

[0079] The input data here can be actual residual data, corresponding to the descriptions of other embodiments below.

[0080] S200. Input the input data of each water level station into the BiLSTM tsunami time series prediction model, and obtain the prediction data of each water level station output by the BiLSTM tsunami time series prediction model.

[0081] The predicted data here are the predicted residuals in the other embodiments described below.

[0082] S300: Based on the input data and prediction data of each water level station, obtain the prediction error sequence, smooth the prediction data sequence of all water level stations, calculate the attention enhancement score, and perform tsunami identification based on the attention enhancement score.

[0083] The above embodiments all represent the model usage stage. Before using the model, it is necessary to train it, specifically as follows: S00A and S00B:

[0084] S00A. Based on the historical water level observation data of a designated water level station within the designated area, perform astronomical tide stripping and filtering to obtain the first training data for training the BiLSTM model.

[0085] For example, historical water level observation data of a specified water level station in a specified area with a time resolution of 1 minute are acquired; the historical water level observation data are fitted with tidal parameters using the harmonic analysis method to obtain historical water level observation data stripped of astronomical tides;

[0086] The historical water level observation data stripped of astronomical tides are filtered to obtain the first training data after interference signal filtering.

[0087] The first training data is normalized to obtain normalized water level time series data.

[0088] S00B: Continuously sample the first training data to obtain sample pairs for time-supervised learning; based on the sample pairs, train the BiLSTM model using an end-to-end training method to obtain the BiLSTM tsunami time-series prediction model.

[0089] In this embodiment, the trained BiLSTM tsunami time series prediction model is also called to perform batch prediction on the validation set data in the sample pair, calculate the error value between the model prediction residual and the actual residual for each sample, and calculate the mean and standard deviation of the prediction error of the entire validation set.

[0090] Based on the Gaussian distribution hypothesis, the error value distribution pattern of the fitting validation set is used, and the anomaly detection threshold is automatically calibrated using percentile statistics to obtain the error distribution function;

[0091] The error distribution function can adapt to the differences in hydrological environment in different sea areas and at different water level stations.

[0092] In this embodiment, for normalized water level time series data, a fixed-length sliding window sampling technique is used to construct sample pairs for time series supervised learning; and the sample pairs are divided into training set and validation set.

[0093] During the training of the BiLSTM model using an end-to-end training method

[0094] Mean squared error (MSE) is selected as the model loss function to quantify the deviation between the model's predicted values ​​and the true values; the Adam adaptive optimizer is used to iteratively update the model weights and bias parameters and adaptively adjust the learning rate.

[0095] The input sequence is fed into a BiLSTM. The forward LSTM processes the sequence sequentially from t=1 to t=N, while the reverse LSTM processes it in reverse order from t=N to t=1. The two states at each time step are fused together. The forward LSTM encodes the historical evolution up to the current time, while the reverse LSTM encodes the current position and subsequent trend. The two states are concatenated to give the model a global perception of the overall shape of the tsunami waveform.

[0096] The method in this embodiment effectively removes background noise and improves the signal-to-noise ratio of anomalous signals by combining precise removal of astronomical tides with difference filtering in the preprocessing process. It also uses a BiLSTM model to model the time series of tide level residuals, fully capturing the dynamic features of the time series and realizing the calculation of anomaly scores based on prediction errors.

[0097] In particular, the above embodiments also introduce an anomaly scoring smoothing mechanism to reduce instantaneous noise interference and improve the stability of detection results; at the same time, combined with the propagation time constraint of tsunami waves, a spatiotemporal decision mechanism is established to significantly reduce the false alarm rate and improve the reliability of the early warning system.

[0098] In this embodiment, step S300 may include the following sub-steps:

[0099] S301. Based on each water level station, subtract the input data and the predicted data of that water level station to obtain the prediction error sequence of that water level station.

[0100] S302. Using the error information obtained from the model training phase, the prediction error sequence is smoothed over time, and the time-smoothed sequence is smoothed using the exponentially weighted moving average method to obtain the first data sequence for identifying tsunamis.

[0101] S303. Calculate the attention enhancement score based on the first data sequence, obtain the ratio of the attention enhancement score to the tsunami threshold T, match the ratio with the ratio table of historical tsunami events, and trigger the first type of alarm information of the water level station.

[0102] The tsunami threshold in this sub-step is obtained based on the global mean and standard deviation obtained during the training phase of the BiLSTM model.

[0103] For example, if the first data sequence in sub-step S302 includes The data sequence, then ;

[0104] Where α is the smoothing coefficient. It is the EWMA value at time t. It is the prediction error at time t. It is the EWMA value from the previous moment;

[0105] Based on the above formula, we obtain the following: The first data sequence;

[0106] Accordingly, in sub-step S303, the attention enhancement score is calculated based on the first data sequence. ,include:

[0107] ; ; ;

[0108] in, It is the prediction error at time t. It is the local energy at time t. It is the EWMA smoothing background value; Let be the attention weight at time t; It ultimately enhances abnormal scores. It is the attention enhancement coefficient. It is the original EWMA outlier score.

[0109] In this embodiment, the alarm information of a single water level station is not used as the standard. In order to more accurately identify tsunamis and prevent interference from regional meteorological factors or isolated noise factors, a spatiotemporal consistency verification method is adopted to judge multiple water level stations. That is, the judgment is made by combining the abnormal water level results and the constraint verification results of tsunami propagation time, as shown in the following sub-step S304, which improves the accuracy of tsunami identification.

[0110] S304. When more than one water level station is detected to have a Class I alarm, check whether the difference in the abnormal triggering time of the station's observation data matches the theoretical tsunami propagation time. If they match, the tsunami alarm will be automatically triggered.

[0111] This embodiment introduces an anomaly scoring smoothing mechanism to reduce instantaneous noise interference and improve the stability of detection results; at the same time, combined with the propagation time constraint of tsunami waves, a spatiotemporal decision mechanism is established to significantly reduce the false alarm rate and improve the reliability of the early warning system.

[0112] The method of this invention enables real-time online detection and automated tsunami warning output for practical business applications, providing technical support for coastal disaster prevention and mitigation decision-making.

[0113] Example 2

[0114] This embodiment provides an automatic tsunami identification device based on BiLSTM, which belongs to the category of intelligent identification application technologies for marine disaster early warning and hydrological monitoring.

[0115] In this embodiment, as Figure 2A and Figure 2B As shown, the overall process is divided into four core functional modules: data preprocessing module, offline model training module, real-time water level data detection module, and intelligent tsunami anomaly alarm module. The overall process achieves accurate, low-false-alarm automatic identification and early warning of tsunami events by combining hydrological data preprocessing, time-series deep learning modeling, data-driven adaptive threshold determination, and spatiotemporal verification of tsunami physical propagation laws.

[0116] Understandably, this embodiment can use historical data of more than one year for training, and data of more than one day can be used for processing, detection, and prediction during real-time water level data detection. The prediction output can also be data results of more than one day.

[0117] The specific implementation process for each module is as follows:

[0118] I. Data Preprocessing Module:

[0119] The data preprocessing module is used to perform quality control, astronomical tide stripping, and bandpass filtering of the raw water level observation data, remove invalid data and interference signals, and extract pure residual time series signals (i.e., real residual data) to provide standardized input data for subsequent model training and real-time detection.

[0120] 1. Acquisition of historical water level data;

[0121] Acquire historical water level observation data from water level stations (with a time span of at least one year, preferably 19 years or more), and perform quality control and imputation of missing values ​​on the raw data to construct a complete and reliable training dataset. For example, in this embodiment, water level stations in the waters from the west coast of the United States to Midway Island and Hawaii can be selected as the research objects. Typically, more than three tide stations (i.e., water level stations) can be selected. In this embodiment, one year of historical data can be used for training.

[0122] The data time resolution was set to 1 minute to ensure effective recording of the arrival time and waveform characteristics of the tsunami wavefront. Simultaneously, basic parameters such as the geographical coordinates (latitude and longitude) of the water level stations and sensor models were collected to ensure the traceability and reliability of the observation data, providing fundamental support for subsequent data analysis and model building.

[0123] Typically, astronomical tides have a 19-year cycle, and theoretically, training a designated water level station can utilize 19 years of historical data. However, due to the incompleteness of historical data, selection can be made based on immediate needs.

[0124] 2. Astronomical tide calculation and difference extraction;

[0125] Ocean water level changes are mainly composed of periodic variations in astronomical tides and non-periodic disturbances such as storm surges, tsunamis, and equipment noise. To isolate periodic astronomical tide interference and accurately extract tsunami anomaly signals, this invention employs harmonic analysis to fit and harmonicly analyze tidal parameters on complete historical water level time-series data. Harmonic constants for each tidal constituent are extracted, and then the predicted astronomical tide values ​​for each time period are calculated. The difference between the measured water level and the predicted astronomical tide values ​​is calculated to extract the water level difference sequence.

[0126] Harmonic analysis was used to fit tidal parameters to historical water level data before subsequent decomposition. The main tidal constituents decomposed in this embodiment include: semi-diurnal tides M2 (major lunar semi-diurnal tide), S2 (major solar semi-diurnal tide), N2 (lunar elliptic semi-diurnal tide), diurnal tides K1 (solar and lunar declination diurnal tide), O1 (major lunar diurnal tide), P1 (major solar diurnal tide), and shallow water tides, totaling no fewer than 37 main tidal constituent harmonic constants.

[0127] The harmonic analysis technique employed in this invention is based on a fundamental astronomical tidal superposition model. This model analyzes the complex sea-level fluctuations caused by the gravitational pull of celestial bodies as a linear superposition of a finite number of tidal constituents with fixed frequencies, amplitudes, and initial phases. The mathematical expression of this fundamental model is as follows:

[0128] (1);

[0129] in, Let t be the predicted tide level relative to mean sea level at time t. The mean sea level deviation is M, where M is the total number of selected tidal constituents. , , Let x be the harmonic amplitude (cm), angular velocity (°), and lag angle (°) of the i-th tidal constituent, respectively. The initial phase (°) of the astronomical scale is based on Greenwich Mean Time. Using this formula (1), the harmonic constants of each tide constituent are determined from historical observation data through harmonic analysis. Based on the harmonic constants, the astronomical tide level at any future time can be calculated by pure astronomical calculation.

[0130] Let the observed water level at time t be... The predicted value of astronomical tides is The water level difference is then defined as: ;

[0131] The resulting difference sequence It includes non-astronomical tidal components such as storm surge, tsunami, and instrument noise.

[0132] 3. Difference filtering processing;

[0133] The water level difference sequence contains high-frequency observation noise and invalid interference signals such as low-frequency sea level changes and seasonal water level fluctuations, which are clearly distinguishable from the tsunami signal frequency band. This embodiment uses a bandpass filtering algorithm to perform fine filtering processing on the difference sequence, taking into account the inherent periodic characteristics of the tsunami signal.

[0134] After completing the astronomical tide removal, the differential water level signal was analyzed. A bandpass filter is applied to extract the tsunami wave component. The filtering process is represented in the time domain as a convolution operation:

[0135] (2);

[0136] in, The impulse response function of the bandpass filter corresponds to the typical period range of tsunami waves (2 minutes to 120 minutes). The integral dummy variable represents the time delay. The estimated tsunami wave value at time t is given by the filter as a weighted average of signals from past and future times. Specifically, in this embodiment, the tsunami wave data is obtained by filtering the historical water level data of one year after subtracting the astronomical tide. The period of the tsunami wave signal can be from 2 minutes to 2 hours.

[0137] The filter passband frequency band strictly matches the effective tsunami signal period, accurately covering the typical fluctuation period of nearshore and offshore tsunamis; at the same time, it completely filters out high-frequency instrument noise with a period of less than 2 minutes and low-frequency slow sea level change signals with a period of more than 2 hours, achieving accurate removal of interference signals, and finally obtaining a clean, high signal-to-noise ratio tsunami-sensitive residual time series, which serves as the standard input data for subsequent bidirectional BiLSTM model modeling.

[0138] In other words, by subtracting the astronomical tide and filtering from one year's historical data in this embodiment, the residual time series sequence for training the model can be obtained. The filtering mentioned above can be a bandpass filter, which is used for frequency selection to suppress noise interference in non-tsunami frequency bands.

[0139] The data preprocessing module in this embodiment can be used during the model training phase or the model usage phase. Both of them perform astronomical tidal harmonic analysis and bandpass filtering on the water level station observation data to extract the final difference time series sequence.

[0140] II. Model Training Module:

[0141] The model training module, based on the clean residual time-series data preprocessed by the data preprocessing module, completes dataset construction, time-series sample partitioning, BiLSTM deep learning model training, error statistics and adaptive anomaly threshold calibration, generating the optimal model and matching threshold parameters that can be used for real-time tsunami detection.

[0142] 4. Data standardization and sliding window;

[0143] To eliminate differences in data dimensions and improve model training convergence speed and prediction accuracy, the complete residual time series dataset after filtering by the data preprocessing module is first subjected to global normalization and standardization. This maps the data uniformly to the BiLSTM model adaptation interval, achieving data standardization preprocessing. In this embodiment, normalization eliminates the impact of differences in dimensions and amplitude imbalances on model convergence. The standardized sequences are then processed using the sliding window method described below.

[0144] To address the continuous nature of water level time series data, a fixed-length sliding window sampling technique is employed to construct a supervised learning sample set. For example, a fixed sliding window length of L=60 can be set, corresponding to 60 consecutive minutes of water level residual time series data. The sliding step size is set to 1 minute, and time series samples are extracted moment by moment. The historical residual sequence within the window for 60 minutes is used as the input feature, and the actual residual value at the next moment of the window is used as the prediction label to construct a one-to-one supervised learning sample pair for time series data.

[0145] The dataset was split according to the time sequence principle. The first 80% of the overall data was selected as the BiLSTM model training set for iterative optimization of BiLSTM model parameters. The last 20% of the data was selected as an independent validation set to monitor the model's generalization ability, suppress overfitting, and ensure the stability of the model's detection accuracy.

[0146] 5. BiLSTM model training;

[0147] In this embodiment, a BiLSTM model is used for training and application.

[0148] Based on the long-distance dependency characteristics of time-series data, a multi-layer stacked BiLSTM long short-term memory neural network model is constructed to adapt to the temporal correlation characteristics of water level time-series data and accurately capture the abnormal water level fluctuation patterns caused by tsunamis.

[0149] The model is trained end-to-end, using historical standardized residual time-series samples captured by a sliding window as input to achieve accurate prediction of the water level residual value at the next time step. The mean squared error (MSE) is used as the model loss function during training to quantify the deviation between the model's predicted values ​​and the true values. The Adam adaptive optimizer is used to iteratively update the model weights and bias parameters, adaptively adjusting the learning rate to accelerate model convergence.

[0150] Throughout the training process, the model's prediction accuracy and generalization performance were monitored in real time using an independent validation set. Overfitting was suppressed through regularization and early stopping mechanisms, and the optimal BiLSTM tsunami time series prediction model was obtained through iterative optimization.

[0151] BiLSTM Basic Gating and Cell State Core Formula:

[0152] A1. Forgotten Gate (decides to discard historical tsunami timeline information) ;

[0153] A2. Input gate (determines the current tsunami level information to be stored in the cell state)

[0154] ;

[0155] A3. Candidate memory cells (temporary tsunami time-series characteristics at the current moment)

[0156] ;

[0157] A4. Cell state update (long-term memory, preserving long-term tsunami-dependent characteristics)

[0158] ;

[0159] A5. Output gate (controls the output of cell state to hidden state) ;

[0160] A6. Hidden layer output (BiLSTM output features at the current time step) ;

[0161] in, Input at time t: Time-series normalized sample of tsunami water level residuals;

[0162] : The hidden state of the LSTM layer at time t-1;

[0163] , , Forget gate, input gate, and output gate output values, ranging from [0,1];

[0164] , : Cell state at the previous moment and the current moment (long-term memory);

[0165] Candidate memory states; , , Portal weight matrix; , , : Portal bias vectors; Sigmoid activation function; Hyperbolic tangent activation function; Hadamard product (element-wise multiplication).

[0166] The input sequence is fed into a BiLSTM. The forward LSTM processes the sequence sequentially from t=1 to t=N; the backward LSTM processes the sequence in reverse order from t=N to t=1. The two states at each time step are fused: the forward state encodes the historical evolution up to the current moment, and the backward state encodes the current position and subsequent trend; the combination of the two states gives the model a global perception of the overall shape of the tsunami waveform.

[0167] 6. Error statistics and threshold determination;

[0168] The optimal BiLSTM model, after training, is used to perform batch predictions on the validation set data. The error between the model's predicted residuals and the actual residuals (i.e., prediction error) is calculated for each sample. Distributional characteristic parameters such as the mean and standard deviation of the prediction errors for the entire validation set are then statistically analyzed. For example... Figure 3 As shown.

[0169] Based on the Gaussian distribution assumption, the prediction error distribution is fitted and the percentile statistics method is used to automatically calibrate the anomaly detection threshold. This method abandons the traditional single judgment mode of fixed threshold and realizes data-driven adaptive threshold setting based on the characteristics of historical data of stations. It adapts to the differences in hydrological environment of different sea areas and different water level stations, improves the adaptability of tsunami anomaly identification, and reduces the risk of missed and false alarms caused by fixed thresholds.

[0170] During the model training phase, the model training module receives the difference sequence output by the data preprocessing module in a sliding window manner. It then uses a bidirectional long short-term memory network to simultaneously model and determine the temporal evolution pattern from both the forward and reverse directions, thus completing the model training and enabling real-time detection and analysis of water level data.

[0171] III. Real-time Detection Module:

[0172] The real-time detection module deploys the trained optimal BiLSTM model and threshold parameters to perform synchronous preprocessing, model inference, error calculation, and anomaly scoring smoothing on the real-time observation data (such as 24 hours of data, two days of data, or one week of data) from the water level station, thereby achieving real-time perception and quantitative assessment of water level anomalies.

[0173] 7. Real-time data input;

[0174] The system receives real-time flow cytometry data from the target water level station and replicates the standardized preprocessing workflow from the model training phase. This includes real-time data quality verification, astronomical tide stripping calculation, bandpass filtering for noise reduction, and data normalization. This ensures that the real-time data preprocessing logic and parameter configuration are completely consistent with the offline training phase, avoiding model inference bias caused by differences in preprocessing rules. Based on the preprocessed real-time residual sequence, a real-time input sliding window with the same dimensionality as the training samples is constructed to provide compliant input data for model inference.

[0175] 8. BiLSTM inference yields predicted values;

[0176] The completed real-time standardized input window is imported into the deployed optimal BiLSTM model for forward inference calculation. Based on the real-time historical time series residual characteristics, the model outputs the predicted value of the water level residual at the current moment, thus completing the fitting prediction of the normal water level time series change pattern.

[0177] For example, the input is a year's worth of data, and the output can be a day's worth of data or a week's worth of data.

[0178] 9. Error smoothing;

[0179] The prediction error between the BiLSTM model's predicted values ​​and the actual observations is calculated (in this embodiment, the prediction error sequence is obtained by subtracting the predicted values ​​obtained from the real-time input data and the prediction inference). Combined with the errors obtained during the training phase, the error sequence at consecutive time points is smoothed using an exponentially weighted moving average (EWMA) method to suppress glitches (i.e., short-term random noise interfering with anomaly detection), resulting in a sequence including... The first data sequence (i.e., the smoothed residual statistics) is transmitted to the anomaly alarm module for processing.

[0180] ;

[0181] Where α is the smoothing coefficient. It is the EWMA value at time t. It is the error value at time t. It is the EWMA value from the previous moment.

[0182] In this embodiment, the real-time detection module is the water level data real-time detection module, which is used to identify the tsunami waveform of the water level data observed at a single station.

[0183] IV. Anomaly Alarm Module:

[0184] The abnormal alarm module combines quantitative judgment of water level anomalies with a dual verification mechanism of spatiotemporal constraints on the physical propagation of tsunamis to accurately identify tsunami events, eliminate interference from ordinary storm surges, equipment malfunctions, and other factors, and finally output reliable tsunami warning signals.

[0185] 10. Anomaly detection;

[0186] Based on the above, including The first data sequence is used to calculate the attention enhancement score.

[0187] The attention enhancement score is calculated using the smoothed EWMA value. :

[0188] ; ; ;

[0189] in, It is the prediction error at time t. It is the local energy at time t. It is the EWMA smoothing background value. Let be the attention weight at time t. It ultimately enhances abnormal scores. It is the attention enhancement coefficient. It is the original EWMA outlier score.

[0190] In this embodiment, the tsunami threshold T can be determined in advance using observational data from historical tsunami events (the calculation process for T is described below), and then the enhanced score can be calculated. The ratio of the ratio to a threshold T is used, and then a tsunami alarm standard is determined based on a ratio case from historical tsunami events (a ratio that can be pre-determined based on observational data from historical tsunami events). An alarm is triggered when the ratio exceeds the alarm standard (this is a single-station anomaly alarm, i.e., the first type of alarm information). To avoid false triggering due to instantaneous spikes, the water level must continuously exceed the alarm standard value for at least 5 minutes before an anomaly can be confirmed. Figure 4 As shown.

[0191] ; ;

[0192] , These are the global mean and standard deviation obtained from training using the BiLSTM method, respectively, where T is the tsunami threshold. The ratio is the ratio of the anomaly score to the threshold T.

[0193] 11. Time constraints of tsunami propagation;

[0194] When an anomaly is detected at a water level station, the difference in the anomaly trigger times at each station is checked to ensure it matches the tsunami propagation time (in this embodiment, inference is made using a multi-station approach; an anomaly at one station does not necessarily mean a tsunami will trigger a real alarm; this embodiment can use data from three or more stations for processing). Using the physical propagation laws of tsunamis as constraints, the spatiotemporal consistency of the water level anomaly trigger sequence is verified, effectively eliminating misjudgments caused by isolated noise events or regional meteorological interference. Figure 5 As shown.

[0195] Assuming there are multiple water level observation stations in the area, the distance between each station is calculated using the following formula:

[0196] ;

[0197] Where R = 6371 km is the average radius of the Earth, ( () represents the longitude and latitude of each station.

[0198] The propagation speed of a tsunami in shallow water is determined by the shallow water wave speed formula:

[0199] ;

[0200] Where g is the acceleration due to gravity. The average water depth between stations is used to calculate the theoretical propagation delay ΔT between stations. ij : i and j are the serial numbers of the water level stations, respectively.

[0201] When an alarm is triggered at a certain station, the data from the corresponding response windows of surrounding stations are monitored, and the theoretical propagation delay condition is determined. For example, the tsunami observation data of each water level station that triggered the alarm are examined to determine whether the propagation delay ΔT obtained based on the water level station observation data matches the theoretical time difference ΔT. ij If they match, it means it's a tsunami.

[0202] 12. Tsunami alarm output;

[0203] By combining the results of the water level anomaly assessment with the tsunami propagation time constraint verification, a tsunami alarm is automatically triggered when both conditions are met. The alarm information is sent to the tsunami warning authority to remind relevant personnel to pay attention and take timely action, and to follow up on the tsunami dynamics, thus achieving rapid and accurate early warning of tsunami disasters.

[0204] The anomaly alarm module in this embodiment pre-calculates the tsunami propagation delay matrix based on shallow water wave velocity and inter-station geographical distance. It performs spatiotemporal alignment and synchronous confirmation counting on the single-station identification signals input by each station within the corresponding delay constraint window. Only when the difference in the anomaly triggering time of each station meets the propagation delay requirement specified by the delay matrix is ​​the station counted as a valid multi-station confirmation. That is, spatiotemporal synchronous reasoning is performed on the multi-station detection results, and normalized weights are calculated through an attention mechanism to enhance tsunami characteristics. When the multi-station confirmation ratio exceeds a set threshold, the regional-level tsunami warning is upgraded and output; otherwise, the single-station anomaly record is downgraded and output.

[0205] Based on the above scheme, this embodiment accurately removes astronomical tides and supplements them with bandpass filtering to highlight tsunami frequency band signals, fundamentally solving the background noise interference problem and laying the foundation for subsequent accurate detection. Furthermore, by utilizing an LSTM network to learn the temporal evolution of normal water level residuals, the real-time prediction error is transformed into a standardized anomaly score, achieving a quantitative assessment of the degree of water level anomaly deviation and overcoming the poor adaptability of traditional fixed threshold methods.

[0206] Applying a moving average or exponential weighted smoothing to the real-time anomaly scoring sequence effectively suppresses false alarms caused by transient noise, significantly reducing the false alarm rate while ensuring detection sensitivity.

[0207] The advantages of this embodiment include: (1) significantly improved detection accuracy: by accurately modeling the normal water level pattern using LSTM, combined with astronomical tide removal and residual filtering, the system’s detection sensitivity and recognition accuracy for real tsunami signals are significantly improved compared to traditional methods, enabling early detection even when the tsunami signal is weak.

[0208] (2) The false alarm rate is greatly reduced: The abnormal scoring smoothing mechanism and the joint decision of propagation time constraint can effectively filter false alarms caused by non-tsunami factors such as local disturbances and instrument failures, and reduce the false alarm rate of the system.

[0209] (3) Strong adaptability, no need for manual parameter adjustment: The adaptive threshold determination method based on statistical distribution can automatically adapt to the hydrological environment of different sea areas and different seasons, without the need for experts to manually set the threshold, and the system maintenance cost is low.

[0210] (4) Strong physical interpretability: The propagation time constraint mechanism introduces the physical prior knowledge of tsunami propagation, which makes the system decision-making process have physical meaning, making it easier for professionals to understand and review the early warning results, and improving the credibility of the system in actual business scenarios.

[0211] In addition, this embodiment also provides an electronic device, which includes: a memory for storing programs or instructions; and a processor, which, when executing the program or instructions, implements the steps of the BiLSTM-based automatic tsunami identification method according to any of the above embodiments.

[0212] The apparatus in Embodiment 2 described above includes a training process. The following provides an apparatus currently in use, such as a BiLSTM-based automatic tsunami identification device, which includes:

[0213] The data preprocessing module is used to acquire real-time observation data of two or more water level stations to be predicted within a preset time period, and to perform astronomical tide stripping and filtering on the real-time observation data to obtain input data for each water level station used to input the BiLSTM tsunami time series prediction model.

[0214] The BiLSTM tsunami time series prediction model is a model trained on historical water level observation data of a designated water level station after astronomical tide stripping and filtering.

[0215] The model output module is used to input the input data of each water level station into the BiLSTM tsunami time series prediction model and obtain the prediction data of each water level station output by the BiLSTM tsunami time series prediction model.

[0216] The tsunami identification module is used to obtain a prediction error sequence based on the input data and prediction data of each water level station, smooth the prediction data sequences of all water level stations, calculate the attention enhancement score, and perform tsunami identification based on the attention enhancement score.

[0217] During use, the aforementioned device can effectively and quickly detect tsunami signals, issue timely alarms, and facilitate prompt handling by relevant personnel to reduce losses.

[0218] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0219] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0220] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0221] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0222] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0223] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. An automatic tsunami identification method based on BiLSTM, characterized in that, include: S100: Obtain real-time observation data of two or more water level stations to be predicted within a preset time period, perform astronomical tide stripping and filtering on the real-time observation data, and obtain input data for each water level station used to input the BiLSTM tsunami time series prediction model. The BiLSTM tsunami time series prediction model is a model trained on historical water level observation data of a designated water level station after astronomical tide stripping and filtering. S200. Input the input data of each water level station into the BiLSTM tsunami time series prediction model, and obtain the prediction data of each water level station output by the BiLSTM tsunami time series prediction model. S300: Based on the input data and prediction data of each water level station, obtain the prediction error sequence, smooth the prediction data sequence of all water level stations, calculate the attention enhancement score, and perform tsunami identification based on the attention enhancement score.

2. The method according to claim 1, characterized in that, The S300 includes: S301. Based on each water level station, subtract the input data and the predicted data of that water level station to obtain the prediction error sequence of that water level station. S302. Using the error information obtained from the model training phase, the prediction error sequence is smoothed over time, and the time-smoothed sequence is smoothed using the exponentially weighted moving average method to obtain the first data sequence for identifying tsunamis. S303. Calculate the attention enhancement score based on the first data sequence, obtain the ratio of the attention enhancement score to the tsunami threshold T, match the ratio with the ratio table of historical tsunami events, and trigger the first type of alarm information of the water level station.

3. The method according to claim 2, characterized in that, The S300 also includes: S304. When more than one water level station is detected to have a Class I alarm, check whether the difference in the abnormal triggering time of the station's observation data matches the theoretical tsunami propagation time. If they match, the tsunami alarm will be automatically triggered.

4. The method according to claim 2, characterized in that, Before step S303, the method further includes: Obtain the tsunami threshold T. ; , These are the global mean and standard deviation obtained during the training phase of the BiLSTM model; or, S302 uses an exponentially weighted moving average method to smooth the time-smoothed sequence, obtaining the first data sequence for tsunami identification, including: ; Where α is the smoothing coefficient. It is the EWMA value at time t. It is the prediction error at time t. It is the EWMA value from the previous moment; Based on the above formula, we obtain the following: The first data sequence; or, S303 calculates the attention enhancement score based on the first data sequence. ,include: ； ； ； in, It is the prediction error at time t. It is the local energy at time t. It is the EWMA smoothing background value; Let be the attention weight at time t; It ultimately enhances abnormal scores. It is the attention enhancement coefficient. It is the original EWMA outlier score.

5. The method according to claim 1, characterized in that, Prior to S200, the method further includes: S00A. Based on the historical water level observation data of the designated water level station in the designated area, perform astronomical tide stripping and filtering to obtain the first training data for training the BiLSTM model. S00B: Continuously sample the first training data to obtain sample pairs for time-supervised learning; based on the sample pairs, train the BiLSTM model using an end-to-end training method to obtain the BiLSTM tsunami time-series prediction model.

6. The method according to claim 5, characterized in that, The SO0B includes: The trained BiLSTM tsunami time series prediction model is called to perform batch predictions on the validation set data in the sample pairs. The error value between the model prediction residual and the actual residual is calculated for each sample. The mean and standard deviation of the prediction error of the entire validation set are calculated. Based on the Gaussian distribution hypothesis, the error value distribution pattern of the fitting validation set is used, and the anomaly detection threshold is automatically calibrated using percentile statistics to obtain the error distribution function; The error distribution function can adapt to the differences in hydrological environment in different sea areas and at different water level stations.

7. The method according to claim 5, characterized in that, The SOOA includes: Historical water level observation data of a specified water level station in a specified area with a time resolution of 1 minute are acquired; the historical water level observation data are fitted with tidal parameters using the harmonic analysis method to obtain historical water level observation data stripped of astronomical tides. The historical water level observation data stripped of astronomical tides are filtered to obtain the first training data after interference signal filtering. The first training data is normalized to obtain normalized water level time series data.

8. The method according to claim 5, characterized in that, S00B includes: For normalized water level time series data, a fixed-length sliding window sampling technique is used to construct sample pairs for time series supervised learning; and the sample pairs are divided into training set and validation set. During the training of the BiLSTM model using an end-to-end training method Mean squared error (MSE) is selected as the model loss function to quantify the deviation between the model's predicted values ​​and the true values; the Adam adaptive optimizer is used to iteratively update the model weights and bias parameters and adaptively adjust the learning rate. The input sequence is fed into a BiLSTM. The forward LSTM processes the sequence sequentially from t=1 to t=N, while the backward LSTM processes it in reverse order from t=N to t=1. The two states at each time step are fused together. The forward LSTM encodes the historical evolution up to the current time, while the backward LSTM encodes the current position and subsequent trend. The two states are concatenated to give the model a global perception of the overall shape of the tsunami waveform.

9. An automatic tsunami identification device based on BiLSTM, characterized in that, include: The data preprocessing module is used to acquire real-time observation data of two or more water level stations to be predicted within a preset time period, and to perform astronomical tide stripping and filtering on the real-time observation data to obtain input data for each water level station used to input the BiLSTM tsunami time series prediction model. The BiLSTM tsunami time series prediction model is a model trained on historical water level observation data of a designated water level station after astronomical tide stripping and filtering. The model output module is used to input the input data of each water level station into the BiLSTM tsunami time series prediction model and obtain the prediction data of each water level station output by the BiLSTM tsunami time series prediction model. The tsunami identification module is used to obtain a prediction error sequence based on the input data and prediction data of each water level station, smooth the prediction data sequences of all water level stations, calculate the attention enhancement score, and perform tsunami identification based on the attention enhancement score.

10. An electronic device, characterized in that, include: Memory, used to store programs or instructions; A processor, which, when executing the program or instructions, implements the steps of the BiLSTM-based automatic tsunami identification method according to any one of claims 1-8.

Images