A method and system for detecting a tsunami wave based on a TCN-informer algorithm

Abstract

The application relates to the technical field of marine monitoring and disaster early warning, and discloses a tsunami wave detection method and system based on a TCN-informer algorithm, which comprises the following steps: S1, acquiring original pressure time series data of the seabed, and performing standardization, resampling and segmentation processing on the original pressure time series data to obtain a model input sample sequence; and S2, constructing a TCN-Informer hybrid model in series connection of a TCN feature extractor and an Informer sequence predictor, training the TCN-Informer hybrid model by using historical pressure data without a tsunami event, and enabling the model to learn the change rule of the seabed pressure under normal conditions and the like; the method realizes accurate detection of a tsunami wave signal from original or simply pretreated seabed pressure data by fusing the advantages of the TCN and Informer models.

Description

Technical Field

[0001] This invention relates to the field of marine monitoring and disaster early warning technology, specifically to a tsunami wave detection method and system based on the TCN-informer algorithm. Background Technology

[0002] Currently, global operational tsunami warnings primarily rely on the DART (Deep Sea Tsunami Assessment and Reporting) system. The core detection algorithm of this system is based on polynomial trend fitting of historical seabed pressure data (such as 3-hour data), and determines the presence of a tsunami signal by calculating the residual between the fitted value and the actual observed value. However, this traditional algorithm has several drawbacks: First, the algorithm is essentially a linear extrapolation, making it difficult to characterize the complex dynamic characteristics of seabed pressure changes, and its ability to capture non-stationary, sudden abnormal signals like tsunamis is limited. Second, it heavily relies on preset high-precision tidal model parameters, requiring the subtraction of tidal signals from the tidal model. If the model does not match the actual tidal characteristics of the sea area or the parameters are improperly set, the false alarm rate of the detection results will increase significantly. Third, it requires complex preliminary data preprocessing steps, including multiple rounds of filtering and denoising operations in addition to tidal subtraction, resulting in a cumbersome calculation process and long data processing time, reducing the real-time performance of tsunami warnings. Fourth, it has weak filtering capabilities for high-frequency ocean noise (such as internal waves and swells), easily misinterpreting such noise as tsunami signals, and has poor anti-interference capabilities.

[0003] In recent years, deep learning technology has provided new solutions for the analysis and prediction of time series data. Temporal Convolutional Networks (TCNs) employ dilated causal convolutions and residual connections, which can efficiently and stably extract local and multi-scale dependency features of time series, and have the advantages of parallel computing and high inference efficiency. The Informer model is a Transformer variant optimized for long-series time series prediction. Through the ProbSparse self-attention mechanism and distillation operation, it significantly reduces the computational complexity and memory overhead of long-series modeling while maintaining prediction accuracy, and can effectively capture the global long-term dependencies of time series data.

[0004] However, tsunami wave signals possess unique dual-scale characteristics: their triggering phase manifests as localized, rapid seabed pressure anomalies, representing short-term local features; while the complete waveform propagation of the tsunami wave is superimposed on global, long-period background signals such as tides and low-frequency meteorological noise, representing long-term global features. A single TCN model can only effectively extract local features and struggles to accurately model long-period background signals; a single Informer model lacks sufficient sensitivity to capture short-term, localized, sudden anomaly signals. Therefore, neither of these single models can optimally balance the dual-scale characteristics of tsunami waves, making them unsuitable for direct application in tsunami wave detection.

[0005] Based on the aforementioned technical issues, there is an urgent need to develop a tsunami wave detection method that does not rely excessively on tidal models, has strong anti-interference capabilities, and can simultaneously capture the dual-scale characteristics of tsunami waves, in order to reduce the false alarm rate and false negative rate of detection and improve the timeliness, reliability and automation level of tsunami warnings. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a tsunami wave detection method and system based on the TCN-informer algorithm. This method and system integrate the advantages of TCN and Informer models to accurately detect tsunami wave signals from raw or simply preprocessed seabed pressure data, simplifying the data processing flow, reducing dependence on tidal models, and improving the detection's anti-interference capability and real-time performance.

[0007] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, this invention provides a tsunami wave detection method based on the TCN-informer algorithm, comprising: S1, acquire the raw pressure time series data of the seabed, and perform standardization, resampling and segmentation on the raw pressure time series data to obtain the model input sample sequence; S2, Construct a TCN-Informer hybrid model consisting of a TCN feature extractor and an Informer sequence predictor connected in series. Train the TCN-Informer hybrid model using historical pressure data without tsunami events, so that the model learns the variation pattern of seabed pressure under normal conditions. S3, input the preprocessed real-time pressure data into the trained TCN-Informer hybrid model to obtain the pressure prediction value, calculate the residual between the pressure prediction value and the actual observed pressure value, determine the tsunami signal through residual analysis and generate a warning message. The mathematical expression for the TCN-Informer hybrid model is: , in, Given the historical pressure sequence as input, For the predicted future pressure sequence, This is the TCN feature extractor function. For Informer sequence predictor functions; The TCN feature extractor is composed of multiple layers of dilated causal convolutions stacked together. It performs dilated causal convolution and residual connection processing on the input historical pressure sequence to extract high-level local temporal features. The formula for the dilated causal convolution operation of the l-th layer is: , in, The kernel size is [size]. For the first The expansion factor of the layer For the first The kernel weights of the layer For the first Layer in time The output.

[0008] Preferably, the formula for the residual connection is: , in, To adjust the dimensions of the input through 1×1 convolution, represents the output of the convolutional layer, and Activation(·) is the ReLU activation function.

[0009] Preferably, the Informer sequence predictor includes an Informer encoder, an Informer decoder, and a fully connected layer; The feature sequence output by the TCN feature extractor is input to the Informer encoder, and after encoding, it is transmitted to the Informer decoder. The decoder output is linearly projected through a fully connected layer to obtain the single-point pressure prediction output.

[0010] Preferably, the Informer encoder encodes the feature sequence through the ProbSparse self-attention mechanism to capture long-term global dependencies.

[0011] Preferably, the Informer encoder encodes the feature sequence using the ProbSparse self-attention mechanism to capture long-term global dependencies, including: S221, Calculate the query vector The sparsity measure is given by the formula: , in, query vector Sparsity metric, For the i-th query vector, Let be the dimension of the key vector. Let K be the length of the key sequence. For the j-th key vector, for and dot product, Operate for the maximum value; S222: Select the query vector with the highest sparsity to construct a simplified query matrix, and calculate the attention mechanism using the following formula: V, in, For attention mechanism functions, For a simplified query matrix, This is the original complete bond matrix. For the original complete value matrix, Let be the dimension of the key vector. For activation functions; S223, the encoder of the Informer model compresses the sequence length through a distillation operation. The formula for the distillation operation from layer j to layer j+1 is: , in, For the temporal features of the (j+1)th layer, For the temporal features of the j-th layer, This indicates the output of the attention block. This is a one-dimensional convolution operation. For activation function, For max pooling, the pooling window size is 2.

[0012] Preferably, in S2, the loss function for training the TCN-Informer hybrid model is the mean squared error, and the formula is: MSE=(P pred (t)-P obs (t)) 2 , Where MSE is the mean square error, P pred (t) represents the model-predicted pressure value, P obs (t) represents the actual observed pressure value.

[0013] Preferably, the determination of tsunami signals through residual analysis described in S3 employs a two-step threshold judgment logic, including: S331, If ​​the absolute value of the residual exceeds the set single-point threshold, a primary alarm is triggered; S332, after triggering the primary alarm, calculate the energy of the residual within the sliding window. If the energy continues to exceed the background level during the calm period, it is determined to be a tsunami signal.

[0014] Preferably, the single-point threshold is the pressure residual value corresponding to a 3 cm water depth change; The formula for calculating the energy of the residual within the sliding window is: , Where E(t) is the residual energy within the sliding window at time t, W is the length of the sliding window, and R(ti) is the residual at time ti; The criterion for determining that the energy continuously exceeds the background level during the quiescent period is: The state lasts for 5 sliding windows; Where E(t) is the residual energy within the sliding window at time t, E background The average energy of the background noise. This is the threshold coefficient.

[0015] Preferably, the standardization mentioned in S1 is mean-free standardization, which specifically involves: calculating the mean μ and standard deviation σ of the three-day historical raw pressure data, and processing the raw pressure time series data P_raw(t) according to the formula P(t)=(P_raw(t)-μ) / σ, where P(t) is the standardized pressure data and P_raw(t) is the original pressure time series data; The resampling refers to resampling the standardized pressure data to a fixed frequency of 15 seconds per point. The segmentation refers to dividing the resampled continuous time-series data into a fixed-length sample sequence.

[0016] On the other hand, the present invention also provides a tsunami wave detection system based on the TCN-informer algorithm, used to implement the tsunami wave detection method based on the TCN-informer algorithm described in the first aspect, the system comprising: The data acquisition module is used to acquire raw pressure time series data from the seabed pressure recorder; The preprocessing module, connected to the data acquisition module, is used to perform mean-reduction standardization, resampling, and segmentation on the original pressure time series data to obtain the model input sample sequence. The model training module, connected to the preprocessing module, is used to construct the TCN-Informer hybrid model and train the model using historical pressure data without tsunami events. The detection and early warning module is connected to the preprocessing module and the model training module respectively. It is used to input the preprocessed real-time pressure data into the trained TCN-Informer hybrid model to obtain the pressure prediction value, calculate the residual and perform residual analysis, and generate and send the early warning message after determining the tsunami signal.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. Strong dual-scale feature capture capability: This invention is the first to integrate TCN and Informer models in the field of tsunami wave detection. The TCN feature extractor can keenly capture local and rapid pressure anomalies in the tsunami wave triggering phase, improving the immediacy of detection; the Informer sequence predictor can accurately model and predict the changing patterns of long-period background pressure such as tides, capturing globally dependent features. The combination of the two achieves the optimal balance of dual-scale features of tsunami waves, greatly improving the accuracy of signal detection.

[0018] 2. No need to rely on high-precision tidal models: By performing mean-removal standardization on the raw pressure data, the complex tidal model subtraction steps in traditional methods are replaced, eliminating errors caused by mismatched or improperly set tidal model parameters, simplifying the data preprocessing process, and making the system highly adaptable to different sea areas with different tidal characteristics, greatly improving deployment convenience.

[0019] 3. Excellent anti-interference capability, low false alarm rate and false negative rate: The long sequence modeling capability of the Informer model can accurately separate tsunami signals from background signals such as tides and low-frequency meteorological noise. The TCN model can effectively distinguish tsunami anomalies from high-frequency ocean noise (internal waves, swells). Combined with the two-step threshold judgment logic to analyze the residual sequence, it can effectively avoid misjudgment caused by noise interference. It can still maintain a low false alarm rate and a low false negative rate in strong noise sea areas.

[0020] 4. High real-time performance and high degree of automation: The parallel computing characteristics of TCN are combined with the efficient long sequence processing capabilities of Informer. The model can obtain the future pressure prediction value in one forward propagation, which has high inference efficiency. The entire detection process, from data acquisition, preprocessing, model prediction to residual analysis and early warning generation, can be fully automated without human intervention, which greatly shortens the early warning response time and improves the real-time performance of tsunami warnings.

[0021] 5. Superior feature learning capability: This invention is an end-to-end intelligent detection method. The model can automatically learn the deep temporal features of seabed pressure changes through data-driven methods without manual feature extraction. It can effectively capture the weak and nonlinear precursor features at the initiation stage of tsunami waves, potentially improving the early detection capability of special types of tsunamis such as landslide tsunamis and tsunamis triggered by small-magnitude earthquakes.

[0022] 6. High model efficiency and stability: The cascaded architecture of TCN and Informer ensures prediction accuracy while combining the parallel computing advantages of TCN and the lightweight design advantages of Informer. Compared with a single complex deep learning model, the hybrid model of this invention has better inference speed and lower resource overhead when deployed. Moreover, residual connection and distillation operations effectively solve the gradient vanishing and computational complexity problems of deep networks, and the stability of model training and inference is greatly improved. Attached Figure Description

[0023] Figure 1 This is a flowchart of a tsunami wave detection method based on the TCN-informer algorithm according to an embodiment of the present invention.

[0024] Figure 2This is a flowchart illustrating how S2 encodes the feature sequence using the ProbSparse self-attention mechanism in a tsunami wave detection method based on the TCN-informer algorithm, according to an embodiment of the present invention.

[0025] Figure 3 This is a flowchart of the S3 two-step threshold judgment logic in a tsunami wave detection method based on the TCN-informer algorithm according to an embodiment of the present invention.

[0026] Figure 4 This is a structural block diagram of a tsunami wave detection system based on the TCN-informer algorithm according to another embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[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 invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0029] Please see Figure 1 A preferred embodiment of the present invention provides a tsunami wave detection method based on the TCN-informer algorithm, comprising: S1. Obtain the raw pressure time series data of the seabed, and perform standardization, resampling and segmentation on the raw pressure time series data to obtain the model input sample sequence.

[0030] Raw pressure time series data P_raw(t) is acquired in real-time or near real-time from a bottom pressure recorder (BPR) deployed on the seabed. The raw data undergoes mean normalization, resampling, and segmentation processing sequentially to obtain a fixed-length sample sequence that meets the model input requirements. This replaces the complex tidal subtraction and multi-round filtering steps in traditional methods, simplifying the preprocessing workflow. Specifically, mean normalization involves calculating the mean μ and standard deviation σ of three days of historical raw pressure data, and processing the raw data according to the formula P(t)=(P_raw(t)-μ) / σ to eliminate data bias and normalize dimensions. Resampling involves resampling the normalized pressure data to a fixed frequency of 15 seconds per point to ensure the consistency of the data time series. Segmentation involves dividing the resampled continuous time series data into fixed-length sample sequences as model input.

[0031] S2, construct a TCN-Informer hybrid model consisting of a TCN feature extractor and an Informer sequence predictor connected in series, and train the TCN-Informer hybrid model using historical pressure data without tsunami events, so that the model learns the variation pattern of seabed pressure under normal conditions.

[0032] A TCN-Informer hybrid model, consisting of a TCN feature extractor and an Informer sequence predictor, was constructed. This hybrid model was trained using historical "clean" pressure data without tsunami events, enabling the model to fully learn the variation pattern of seabed pressure from local to global under normal conditions, thus laying the foundation for subsequent anomaly signal detection.

[0033] The mathematical expression for the TCN-Informer hybrid model is: , in, For the input historical pressure sequence ( ), For the predicted future pressure sequence ( ), This is the TCN feature extractor function. For Informer sequence predictor functions, The number of time points in the input sequence. This represents the number of time points in the output sequence. It represents the set of real numbers.

[0034] The TCN feature extractor is composed of multiple layers of dilated causal convolutions. It first performs dilated causal convolution processing on the input historical pressure sequence, and then performs residual connection processing on the convolution result to extract high-level local temporal features.

[0035] The formula for the dilated causal convolution operation in layer l is: , in, The kernel size is [size]. For the first The expansion factor of the layer For the first The kernel weights of the layer For the first Layer in time The output.

[0036] The formula for residual connection is: , in, To adjust the dimensions of the input through 1×1 convolution, represents the output of the convolutional layer, and Activation(·) is the ReLU activation function.

[0037] After passing through multiple TCN residual blocks, an output rich in local features is obtained. ,in This represents the dimension of the model's hidden layers.

[0038] The Informer sequence predictor consists of an Informer encoder, an Informer decoder, and a fully connected layer. The feature sequence output from the TCN is input to the Informer encoder, which uses a ProbSparse self-attention mechanism to capture long-term temporal global dependencies and compresses the sequence length through distillation to reduce computational overhead. The Informer decoder employs generative inference; its input consists of a start marker (taken from the end of the TCN output feature sequence) concatenated with all-zero target placeholders. It acquires global context through masked multi-head attention and encoder-decoder attention, and outputs non-autoregressive predicted features. Finally, the decoder output is linearly projected through the fully connected layer to obtain the single-point pressure prediction value.

[0039] Specifically, such as Figure 2 As shown, the Informer encoder encodes feature sequences using the ProbSparse self-attention mechanism, capturing long-term global dependencies, including: S221, Calculate the query vector The sparsity measure is given by the formula: , in, query vector Sparsity metric, For the i-th query vector, Let be the dimension of the key vector. Let K be the length of the key sequence. For the j-th key vector, for and dot product, Operate for the maximum value; S222: Select the query vector with the highest sparsity to construct a simplified query matrix, and calculate the attention mechanism using the following formula: V, in, For attention mechanism functions, For a simplified query matrix, This is the original complete bond matrix. For the original complete value matrix, Let be the dimension of the key vector. For activation functions; S223, the encoder of the Informer model compresses the sequence length through a distillation operation. The formula for the distillation operation from layer j to layer j+1 is: , in, For the temporal features of the (j+1)th layer, For the temporal features of the j-th layer, This indicates the output of the attention block. This is a one-dimensional convolution operation. For activation function, For max pooling, the pooling window size is 2.

[0040] With this design, the encoder can gradually extract the most representative features from the input sequence after stacking multiple layers, while significantly reducing the computational and memory overhead of long sequence processing.

[0041] The decoder then uses generative inference to output the entire prediction sequence in one go (non-autoregressive). .

[0042] The decoding process is as follows: Input construction: starting from the start marker (Taken from (The last part) is concatenated with a zero target placeholder.

[0043] Masked attention and encoder-decoder attention: The first layer uses masked multi-head attention to prevent information leakage; the second layer interacts with the encoder output to obtain the global context.

[0044] Output: The final predicted future pressure value is obtained through linear projection of the fully connected layer. .

[0045] The hybrid model is trained using historical "clean" stress data that does not include tsunami events. The training objective is to minimize the model's predictions. Compared with actual future observations The differences between them. The loss function for training the TCN-Informer hybrid model uses the mean squared error, and the formula is: , Where MSE is the mean square error, P pred (t) represents the model-predicted pressure value, P obs (t) represents the actual observed pressure value.

[0046] Through training, the model learns the patterns of pressure changes from local to global under normal conditions.

[0047] S3. Input the preprocessed real-time pressure data into the trained TCN-Informer hybrid model to obtain the pressure prediction value, calculate the residual between the pressure prediction value and the actual observed pressure value, determine the tsunami signal through residual analysis and generate a warning message.

[0048] The preprocessed real-time seabed pressure data is input into the trained TCN-Informer hybrid model to obtain the future single-point pressure prediction value P. pred (t). Real-time calculation of predicted pressure value and actual observed pressure value P obs The residual R(t) of the signal is analyzed using a two-step threshold judgment logic to accurately determine the tsunami signal. If the signal is determined to be a tsunami signal, the system automatically generates an early warning message and sends it to the early warning center to achieve early tsunami warning.

[0049] Residual calculation: The residual sequence can effectively reflect the deviation between the actual seabed pressure and the normal background pressure, and the abnormal deviation indicates the existence of potential tsunami signals.

[0050] Specifically, such as Figure 3 As shown, tsunami signals are determined through residual analysis, employing a two-step threshold judgment logic, including: S331, If ​​the absolute value of the residual exceeds the set single-point threshold, a primary alarm is triggered; S332, after triggering the primary alarm, calculate the energy of the residual within the sliding window. If the energy continues to exceed the background level during the calm period, it is determined to be a tsunami signal.

[0051] Furthermore, the single-point threshold can be set to the pressure residual value corresponding to a 3 cm water depth change.

[0052] The formula for calculating the energy of the residual within the sliding window is: , Where E(t) is the residual energy within the sliding window at time t, W is the length of the sliding window, and R(ti) is the residual at time ti.

[0053] The criteria for determining whether energy levels consistently exceed the background level during the quiescent period are: The state lasts for 5 sliding windows; Where E(t) is the residual energy within the sliding window at time t, E background The average energy of the background noise. This is the threshold coefficient.

[0054] Warning generation: After confirming the tsunami signal, a warning message is generated that includes the detection time, the deployment location of the seabed pressure recorder, and the estimated wave height of the tsunami wave, and is sent to the marine disaster early warning center in real time.

[0055] This invention also provides a tsunami wave detection system 100 based on the TCN-informer algorithm, used to implement the tsunami wave detection method based on the TCN-informer algorithm as described in the foregoing embodiments, and possessing corresponding functions and beneficial effects. Figure 4 As shown, the tsunami wave detection system 100 based on the TCN-informer algorithm includes: The data acquisition module 111 is used to acquire raw pressure time series data from the seabed pressure recorder. The raw pressure time series data P_raw(t) can be acquired in real time or near real time using a pressure recorder (BPR) deployed on the seabed.

[0056] The preprocessing module 112, connected to the data acquisition module 111, is used to perform mean-reduction standardization, resampling, and segmentation processing on the original pressure time series data to obtain the model input sample sequence. The specific implementation method is as described in S1, and will not be repeated here.

[0057] The model training module 113, connected to the preprocessing module 112, is used to construct the TCN-Informer hybrid model and train the model using historical pressure data without tsunami events. The specific implementation method is described in S2, and will not be repeated here. The detection and early warning module 114 is connected to the preprocessing module 112 and the model training module 113 respectively. It is used to input the preprocessed real-time pressure data into the trained TCN-Informer hybrid model to obtain the pressure prediction value, calculate the residual and perform residual analysis, and generate and send an early warning message after determining the tsunami signal. The specific implementation method is as described in S3, and will not be repeated here.

[0058] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A tsunami wave detection method based on the TCN-informer algorithm, characterized in that, include: S1, acquire the raw pressure time series data of the seabed, and perform standardization, resampling and segmentation on the raw pressure time series data to obtain the model input sample sequence; S2, Construct a TCN-Informer hybrid model consisting of a TCN feature extractor and an Informer sequence predictor connected in series. Train the TCN-Informer hybrid model using historical pressure data without tsunami events, so that the model learns the variation pattern of seabed pressure under normal conditions. S3, input the preprocessed real-time pressure data into the trained TCN-Informer hybrid model to obtain the pressure prediction value, calculate the residual between the pressure prediction value and the actual observed pressure value, determine the tsunami signal through residual analysis and generate a warning message. The mathematical expression for the TCN-Informer hybrid model is: ， in, Given the historical pressure sequence as input, For the predicted future pressure sequence, This is the TCN feature extractor function. For Informer sequence predictor functions; The TCN feature extractor is composed of multiple layers of dilated causal convolutions stacked together. It performs dilated causal convolution and residual connection processing on the input historical pressure sequence to extract high-level local temporal features. The formula for the dilated causal convolution operation of the l-th layer is: ， in, The kernel size is [size]. For the first The expansion factor of the layer For the first The kernel weights of the layer For the first Layer in time The output.

2. The tsunami wave detection method based on the TCN-informer algorithm according to claim 1, characterized in that, The formula for the residual connection is: ， in, To adjust the dimensions of the input through 1×1 convolution, represents the output of the convolutional layer, and Activation(·) is the ReLU activation function.

3. The tsunami wave detection method based on the TCN-informer algorithm according to claim 2, characterized in that, The Informer sequence predictor includes an Informer encoder, an Informer decoder, and a fully connected layer; The feature sequence output by the TCN feature extractor is input to the Informer encoder, and after encoding, it is transmitted to the Informer decoder. The decoder output is linearly projected through a fully connected layer to obtain the single-point pressure prediction output.

4. The tsunami wave detection method based on the TCN-informer algorithm according to claim 3, characterized in that, The Informer encoder encodes feature sequences using the ProbSparse self-attention mechanism, capturing long-term global dependencies.

5. The tsunami wave detection method based on the TCN-informer algorithm according to claim 4, characterized in that, The Informer encoder encodes the feature sequence using the ProbSparse self-attention mechanism, capturing long-term global dependencies, including: S221, Calculate the query vector The sparsity measure is given by the formula: ， in, For query vector Sparsity metric, For the i-th query vector, Let be the dimension of the key vector. Let K be the length of the key sequence. For the j-th key vector, for and dot product, Operate for the maximum value; S222: Select the query vector with the highest sparsity to construct a simplified query matrix, and calculate the attention mechanism using the following formula: V， in, For attention mechanism functions, For a simplified query matrix, This is the original complete bond matrix. For the original complete value matrix, Let be the dimension of the key vector. For activation functions; S223, the encoder of the Informer model compresses the sequence length through a distillation operation. The formula for the distillation operation from layer j to layer j+1 is: ， in, For the temporal features of the (j+1)th layer, For the temporal features of the j-th layer, Indicates the output of the attention block. This is a one-dimensional convolution operation. For activation function, For max pooling, the pooling window size is 2.

6. The tsunami wave detection method based on the TCN-informer algorithm according to claim 1, characterized in that, In S2, the loss function used for training the TCN-Informer hybrid model is the mean squared error, and the formula is as follows: MSE=(P pred (t)-P obs (t)) 2 , Where MSE is the mean square error, P pred (t) represents the model-predicted pressure value, P obs (t) represents the actual observed pressure value.

7. The tsunami wave detection method based on the TCN-informer algorithm according to claim 1, characterized in that, The method described in S3 for determining tsunami signals through residual analysis employs a two-step threshold judgment logic, including: S331, If ​​the absolute value of the residual exceeds the set single-point threshold, a primary alarm is triggered; S332, after triggering the primary alarm, calculate the energy of the residual within the sliding window. If the energy continues to exceed the background level during the calm period, it is determined to be a tsunami signal.

8. The tsunami wave detection method based on the TCN-informer algorithm according to claim 7, characterized in that, The single-point threshold is the pressure residual value corresponding to a 3 cm water depth change; The formula for calculating the energy of the residual within the sliding window is: ， Where E(t) is the residual energy within the sliding window at time t, W is the length of the sliding window, and R(ti) is the residual at time ti; The criterion for determining that the energy level continuously exceeds the background level during the quiescent period is: The state lasts for 5 sliding windows; Where E(t) is the residual energy within the sliding window at time t, E background The average energy of the background noise. This is the threshold coefficient.

9. The tsunami wave detection method based on the TCN-informer algorithm according to claim 1, characterized in that, The standardization mentioned in S1 is mean-free standardization, which specifically involves: calculating the mean μ and standard deviation σ of the three-day historical raw pressure data, and processing the raw pressure time series data P_raw(t) according to the formula P(t)=(P_raw(t)-μ) / σ, where P(t) is the standardized pressure data and P_raw(t) is the original pressure time series data; The resampling refers to resampling the standardized pressure data to a fixed frequency of 15 seconds per point. The segmentation refers to dividing the resampled continuous time-series data into a fixed-length sample sequence.

10. A tsunami wave detection system based on the TCN-informer algorithm, characterized in that, The system for implementing the tsunami wave detection method based on the TCN-informer algorithm according to any one of claims 1-9 includes: The data acquisition module is used to acquire raw pressure time series data from the seabed pressure recorder; The preprocessing module, connected to the data acquisition module, is used to perform mean-reduction standardization, resampling, and segmentation on the original pressure time series data to obtain the model input sample sequence. The model training module, connected to the preprocessing module, is used to construct the TCN-Informer hybrid model and train the model using historical pressure data without tsunami events. The detection and early warning module is connected to the preprocessing module and the model training module respectively. It is used to input the preprocessed real-time pressure data into the trained TCN-Informer hybrid model to obtain the pressure prediction value, calculate the residual and perform residual analysis, and generate and send the early warning message after determining the tsunami signal.

Images