Intelligent early warning method and system for abnormal wave based on three-branch bilinear gating fusion model and medium

Abstract

This invention discloses an intelligent early warning method, system, and medium for abnormal waves based on a three-branch bilinear gated fusion model, belonging to the field of marine engineering safety early warning technology. The method includes the following steps: constructing an original dataset containing normal wave samples and abnormal wave samples; constructing abnormal wave early warning samples based on a preset input time window and early warning period; extracting the first, second, and third feature vectors from a specific wavefront sequence using a three-branch feature extraction network; performing adaptive fusion through a bilinear gated fusion module to obtain the final fused features; inputting the final fused features into a classification and discrimination module, and outputting the probability of abnormal wave occurrence through multi-dimensional comprehensive analysis. This invention addresses the problems of scarce abnormal wave samples and highly imbalanced categories by introducing hybrid training, asymmetric loss, and false alarm suppression strategies to improve the stability and generalization ability of the developed model under real sea conditions.

Description

Technical Field

[0001] This invention relates to an intelligent early warning method and system for abnormal waves based on a three-branch bilinear gating fusion model, belonging to the field of marine engineering safety early warning technology, and can be applied to extreme wave risk early warning in scenarios such as ship navigation, marine platform operation, offshore hoisting, personnel transfer and offshore wind power operation and maintenance. Background Technology

[0002] Anomalous waves are a type of extreme ocean wave event with a low probability of occurrence but extremely high destructive power. They typically manifest as unusually large waves that suddenly appear within a short period of time, with amplitudes significantly higher than other waves in the surrounding area. Unlike regular large waves, typhoon waves, or regional high sea state events, anomalous waves are characterized by their suddenness, concealment, localization, and short duration, making them more threatening to ship navigation, offshore platform operations, offshore hoisting, helicopter takeoffs and landings, marine energy development, and special offshore operations. The formation process of anomalous waves may be simultaneously influenced by factors such as linear wave train phase superposition, nonlinear energy focusing, multi-scale wave group modulation, directional diffusion, and background sea state disturbances. Therefore, providing early warnings for anomalous waves is more difficult, requiring the ability to determine whether an anomalous wave event will occur at the monitoring location within a certain future time frame based solely on the measured wavefront sequence within the current time window.

[0003] Currently, there are no effective technical methods for rapid intelligent early warning of abnormal waves among the publicly available technical solutions. Most of the technical solutions focus on the amplitude prediction of ocean wave elements, disastrous large waves, or typhoon waves. The existing patents are as follows: For example, Chinese patent application publication number CN118465872A discloses a coastal wave forecasting method, device, equipment, and medium. This method acquires meteorological data and boundary wave characteristic data of the area to be forecasted, inputs a pre-established wave forecasting model, and outputs wave forecasting results such as significant wave height, wave direction angle, wave period, and spectral peak period. This method mainly solves the problem of continuous numerical forecasting of conventional wave elements in coastal areas, focusing on improving the prediction accuracy of wave parameters such as significant wave height. However, its output is still regional wave statistical elements, and it cannot achieve real-time early warning for specific locations. Chinese patent application publication number CN121435191A discloses a method, device, equipment, medium, and product for jointly predicting the probability of catastrophic large waves and wave height. This method trains a significant wave height prediction model and a large wave occurrence probability prediction model by screening key environmental factors such as mean sea level pressure, zonal wind speed, and meridional wind speed. Although this method involves estimating the "probability of large wave occurrence," its research object is a regional process of catastrophic large waves, and its input mainly relies on environmental factors, still failing to achieve early warning for single-point anomalous waves. Chinese patent application publication number CN121684068A discloses a wave height prediction system and method that integrates dynamic features and physical gating. It outputs significant wave height prediction values ​​through feature selection, dynamic weight allocation, and a physically guided gated feature-aware Transformer prediction model. However, this approach still relies on statistical wave height prediction rather than providing real-time early warning of abnormal waves based on existing wave train information.

[0004] From the perspective of abnormal wave engineering early warning, existing technologies have the following three core shortcomings: ① Warning Target Bias: Most existing intelligent wave prediction methods use continuous variables such as significant wave height, salient wave height, wave steepness, wave period, or typhoon wave height as output targets, essentially belonging to "numerical regression prediction." However, anomalous wave warnings are more concerned with the risk of a sudden super-giant wave occurring within a certain future timeframe. That is, the model needs to determine the probability of anomalous waves occurring locally based on existing local wave train information. If global statistical wave height prediction remains the core, the model can only focus on overall sea state amplitude changes and cannot analyze the event attributes of anomalous waves suddenly generated within local wave groups, making early warning difficult.

[0005] ② Insufficient early warning capability: Although existing deep learning models can handle nonlinear wave time series, their prediction step size is usually short, essentially belonging to linear extrapolation interpolation. They require real-time correction of deviations using measured data, making it impossible to achieve early warning functions with sampling step sizes hundreds of times larger, and even more difficult to accurately predict abnormal wave heights. Furthermore, a single CNN, LSTM, BP neural network, or a single Transformer structure can typically only learn wave sequence information from a single feature perspective, failing to simultaneously consider the influence of the physical mechanisms of abnormal wave generation. Therefore, a single neural network model struggles to simultaneously express the temporal morphological features, multi-scale dynamic features, and physical prior features required for abnormal wave early warning. Summary of the Invention

[0006] The purpose of this invention is to propose an intelligent early warning method and system for abnormal waves based on a three-branch bilinear gated fusion model. Based on the time series information of the preceding wave surface obtained by the buoy, the method comprehensively judges from multiple perspectives whether an abnormal wave event will occur within a preset time period in the future, effectively improving the accuracy, stability and engineering applicability of abnormal wave early warning.

[0007] The intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model described in this invention includes the following steps: S1: Data Acquisition and Labeling: Acquire the sea surface displacement time series collected by the buoy and label the samples to construct an original wave dataset containing normal wave samples and abnormal wave samples; S2: Data preprocessing: The original wave dataset containing normal wave samples and abnormal wave samples is preprocessed to obtain a standardized wave surface time series; S3: Sample Construction: Construct abnormal wave early warning samples based on the preset input time window and early warning period, so that the model input is a fixed-length wavefront sequence before the occurrence of abnormal waves, and the output is the probability of the occurrence of abnormal waves in the future preset period. S4: Three-branch feature extraction: The first feature vector, the second feature vector, and the third feature vector are extracted from the wavefront sequence using a three-branch feature extraction network; S5: Bilinear Gated Fusion: The first feature vector, the second feature vector, and the third feature vector are adaptively fused through the bilinear gated fusion module to obtain the final fused feature, forming a three-branch bilinear gated fusion model; S6: Early warning output: Input the final fused features into the classification and discrimination module, output the probability of abnormal wave occurrence, and convert it into a binary classification early warning result according to the preset discrimination threshold.

[0008] Preferably, the step S1 of acquiring the sea surface displacement time series collected by the buoy and labeling the samples specifically involves: acquiring the sea surface displacement time series data collected by the buoy in the target sea area, constructing an original wave dataset containing normal wave samples and abnormal wave samples, and performing binary classification labeling on the samples according to a preset abnormal wave criterion; the preset abnormal wave criterion is: when the maximum wave height in the wave record is greater than a preset threshold multiple of the significant wave height, it is marked as an abnormal wave event sample, otherwise it is marked as a normal wave sample.

[0009] Preferably, the three-branch feature extraction network in step S4 specifically includes: First branch: After local feature enhancement of the standardized original wavefront sequence, it is divided into multiple overlapping time segments. The local waveform morphology and long-range time dependence are extracted by the time segment Transformer encoder to obtain cross-time waveform morphology features. The second branch: Empirical mode decomposition is performed on the standardized wavefront sequence to obtain a multi-channel mode matrix composed of multiple intrinsic mode functions and residual terms. The energy distribution, mutual modulation and dynamic coupling relationship between wave components at different scales are learned through a lightweight iTransformer encoder to obtain multi-scale dynamic coupling characteristics. The third branch extracts low-dimensional physical feature vectors containing temporal statistical features and spectral features from the original wavefront sequence, and maps them to physical prior features through a lightweight multilayer perceptron; the low-dimensional physical feature vectors include at least several indices such as standard deviation, skewness, kurtosis, range, zero-crossing rate, energy, spectral correlation features, spectral bandwidth, and wave steepness.

[0010] Preferably, in the first branch, the temporal segment length is 64-256 sampling points, the adjacent segment step size is 16-128 sampling points, the number of temporal segment Transformer encoder layers is 2-6, and the number of multi-head attention heads is 4-16; in the second branch, the number of intrinsic mode functions obtained from empirical mode decomposition is 3-8, the number of sieving iterations is 5-20, the threshold for the standard deviation of intrinsic mode function judgment is 0.10-0.30, the number of lightweight iTransformer encoder layers is 2-5, and the dimension of the feedforward network hidden layer is 256-1024.

[0011] Preferably, the bilinear gated fusion process in step S5 specifically includes the following sub-steps: S51: Calculate the dynamic weights between the first and second feature vectors using a gated network to obtain the weighted fusion backbone features; S52: Construct a product consistency term, an absolute difference term, and a low-rank bilinear interaction term between the first and second eigenvectors to enhance the model’s ability to express cross-branch complementary information in the precursor features of anomalous waves. S53: Concatenate the main fusion feature, product consistency term, absolute difference term, and low-rank bilinear interaction term with the third feature vector to obtain the final fusion feature.

[0012] Preferably, the low-rank bilinear interaction term is obtained by linearly projecting the first eigenvector and the second eigenvector onto a trainable low-rank mapping matrix and then multiplying them element by element, with the bilinear interaction rank ranging from 16 to 128.

[0013] Preferably, the warning output in step S6 specifically involves: inputting the final fused features into a classification and discrimination network to obtain the predicted probability of the occurrence of the abnormal wave; converting the probability of the occurrence of the abnormal wave into a binary classification warning result according to a preset discrimination threshold; outputting a "abnormal wave occurs" warning result when the probability of the occurrence of the abnormal wave is greater than or equal to the discrimination threshold; and outputting a "abnormal wave does not occur" warning result when the probability of the occurrence of the abnormal wave is less than the discrimination threshold.

[0014] Preferably, after step S5, a model training step is further included: the three-branch bilinear gated fusion model is optimized using a training strategy suitable for identifying low-probability extreme events; the training strategy includes a training optimization strategy that combines introducing an asymmetric loss function, regularization, and test enhancement; in case of insufficient anomalous wave samples from a single target buoy, anomalous wave samples from neighboring buoys or similar sea areas are introduced together with normal samples from the target buoy to construct a training set.

[0015] The intelligent early warning system for abnormal waves based on a three-branch bilinear gated fusion model described in this invention includes: The data acquisition module is used to acquire the sea surface displacement time series collected by the buoy and label the samples to build an original dataset containing normal wave samples and abnormal wave samples. The data preprocessing module is used to preprocess the original dataset to eliminate data noise and scale differences; The sample construction module is used to construct early warning samples based on a preset input time window and early warning time period, so that the model input is a fixed-length wavefront sequence before the occurrence of abnormal waves, and the output is the probability of the occurrence of abnormal waves in the future preset time period. The three-branch feature extraction module is used to extract waveform morphology features, multi-scale dynamic coupling features, and physical prior features from the preprocessed wavefront sequence, respectively. A bilinear gated fusion module is used to adaptively fuse the waveform morphology features, multi-scale dynamic coupling features and physical prior features to obtain the final fused features. The early warning output module is used to input the final fused features into the classification network and output the probability of abnormal wave occurrence within a preset time period. When the probability is greater than or equal to the preset early warning threshold, an abnormal wave early warning result is generated.

[0016] The computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the steps of the intelligent early warning method for abnormal waves based on a three-branch bilinear gating fusion model.

[0017] Compared with existing technologies, the intelligent early warning method and system for abnormal waves based on a three-branch bilinear gated fusion model of the present invention exhibits the following beneficial effects in terms of technical performance and practical application: 1) This invention directly uses measured wave train information to determine the probability of abnormal waves appearing in a specific area within a preset time period in the future. It can better meet the needs of rapid decision-making on "occurrence / non-occurrence" in the safety early warning scenario of marine engineering operations. It solves the problems of short prediction time and poor accuracy of real-time correction extrapolation model, and the inability of global wave height statistical regression model to achieve local real-time early warning.

[0018] 2) This invention designs a three-branch structure to simultaneously extract time-domain waveform morphological features, multi-scale dynamic coupling features, and physical prior features. The first branch A focuses on the local structure of the waveform and long-range temporal coherence; the second branch B focuses on the cross-scale wave coupling after empirical mode decomposition; and the third branch C supplements statistical and spectral domain physical information. This allows for the simultaneous characterization of complex precursor features before the occurrence of extreme anomalous waves from multiple perspectives, thereby improving the model's prediction accuracy through comprehensive analysis.

[0019] 3) This invention uses a bilinear gating fusion mechanism to adaptively fuse three types of features. It can not only dynamically allocate the weights of feature streams of different depths, but also enhance the complementary expressive ability of cross-branch features through consistency terms, difference terms and low-rank bilinear interaction terms. Compared with single CNN, LSTM, BP networks or simple feature splicing methods, it has a stronger ability to identify abnormal wave states.

[0020] 4) This invention addresses the characteristics of low probability, imbalanced samples, and high cost of false alarms and missed alarms in abnormal waves by introducing asymmetric loss, hard negative sample constraints, and test enhancement strategies. This enables the model to balance the recall rate of abnormal waves and the control of false alarms of normal waves while maintaining a fixed threshold for discrimination, making it more suitable for engineering early warning applications.

[0021] 5) Based on a trained model, this invention can achieve effective early warning without relying on complex three-dimensional wave numerical simulation to reconstruct the future wave field. It has low computational resource requirements and high prediction efficiency. The model is easy to deploy in marine monitoring platforms, ship auxiliary decision-making systems, and marine engineering structure safety early warning systems. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2The following is a comparison of the abnormal wave early warning performance of the TS model and the conventional deep learning model in the embodiments of the present invention. In the figure, (a) represents the prediction ability of the TS model of the present invention under different datasets, (b) represents the prediction ability of the LSTM model under different datasets, (c) represents the prediction ability of the CNN model under different datasets, and (d) represents the prediction ability of the BP model under different datasets. Figure 3 The stability verification diagram of the model of the present invention under different early warning time periods in the embodiments of the present invention; (a) represents the prediction ability of the TS model of the present invention at different times in the D1 dataset, (b) represents the prediction ability of the TS model of the present invention at different times in the D2 dataset, (c) represents the prediction ability of the TS model of the present invention at different times in the D3 dataset, and (d) represents the prediction ability of the TS model of the present invention at different times in the D4 dataset. Figure 4 The figure shows the abnormal wave warning results of the model of the present invention under the condition of severe imbalance of event categories in real sea conditions in the embodiments of the present invention; in the figure, (a) represents the warning capability of the invention under real samples, and (b) represents the number of abnormal wave and normal wave samples under real sea conditions. Figure 5 The ablation experiment results of the three-branch feature extraction structure in the embodiment of the present invention are shown in the figure; (a) represents the correct prediction rate of abnormal waves in each branch of the TS model of the present invention, (b) represents the correct prediction rate of normal waves in each branch of the TS model of the present invention, and (c) represents the overall correct prediction rate of each branch of the TS model of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0024] Example 1: like Figure 1 As shown, the intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model described in this invention includes a sample construction process (A) and an early warning workflow (B). Specifically, it includes the following steps: Sample A construction process: S1: Data Acquisition and Labeling: First, continuous sea level rise and fall time series were acquired to construct an original wave dataset containing normal wave samples and anomalous wave samples. The long time series was divided into multiple short time window sequences of equal length by time window partitioning. Then, the window samples were classified and labeled according to a judgment rule, where the judgment rule for anomalous wave samples is H. max / H s >R0, H max H represents the maximum wave height within the window. sFor significant wave height, R0 is the judgment threshold, and the judgment rule for normal wave samples is H. max / H s ≤R0; The two types of samples are aggregated to form the input sample. The wave data comes from the measured buoy observation system, preferably using quality-controlled sea surface elevation time series data. The original wave records are screened and labeled according to the anomalous wave criteria, where anomalous wave event samples are marked as 1, and normal wave samples are marked as 0.

[0025] S2: Data Preprocessing: The original wave dataset, containing both normal and anomalous wave samples, is preprocessed to obtain a standardized wave surface time series. This includes outlier handling, length alignment, and sample standardization. Infinity values, missing values, and outliers are removed or corrected, and wave sequences from different sources or files are trimmed to a uniform length. Standardization is performed on each input sample to mitigate the impact of energy scale differences across sea states on model training, allowing the model to focus on learning waveform structure, local morphology, multi-scale coupling, and physical prior features.

[0026] S3: Sample Construction: Constructing abnormal wave warning samples based on a preset input time window and warning period. For wave records containing abnormal wave events, the time of the abnormal wave occurrence is placed within the warning interval after the input window, allowing the model to determine whether an abnormal wave will occur in the future based on the wave surface sequence before the abnormal wave occurs. For wave records without abnormal wave events, normal wave samples are constructed using the same input time window length and warning period. This completes the steps of the A sample construction workflow. The model's input is a fixed-length wave surface sequence before the abnormal wave occurs, and the output is the probability of an abnormal wave occurring within a preset future time period.

[0027] B. Early Warning Workflow: S4: Three-branch feature extraction: The three-branch feature extraction model extracts features from three different dimensions simultaneously. The first branch extracts cross-time morphological features, the second branch extracts multi-scale dynamic coupling features, and the third branch extracts physical prior statistical features.

[0028] A first branch is constructed to extract cross-time waveform morphology consistency features. This branch processes the normalized original wavefront time series, first enhancing local waveform features through a one-dimensional convolutional residual structure, then dividing the wave sequence into several overlapping time segments, and extracting the local waveform morphology and long-range time dependence through a Transformer encoder based on these time segments to obtain the first feature vector. This branch is used to identify changes in wave train morphology, wave group structure, and temporal coherence before the occurrence of anomalous waves.

[0029] A second branch is constructed to extract multi-scale dynamic coupling features. This branch performs empirical mode decomposition on the normalized wavefront sequence to obtain multiple intrinsic mode functions and residual terms, which are then combined to form a multi-channel mode matrix. Subsequently, the different mode channels are normalized and input into a lightweight iTransformer encoder to learn the energy distribution, mutual modulation, and dynamic coupling relationships between wave components at different scales, resulting in a second feature vector. This branch is used to characterize the dynamic precursor features of multi-scale non-stationary wave processes preceding anomalous wave occurrences.

[0030] A third branch is constructed for extracting physical prior features. This branch directly extracts statistical and spectral physical features from the original wavefront time series, forming a low-dimensional physical feature vector. These physical features include at least the following indices: standard deviation, skewness, kurtosis, range, zero-crossing rate, energy, spectral correlation, spectral bandwidth, and wave steepness. Subsequently, the physical feature vector is input into a lightweight multilayer perceptron model, mapping it to a physical prior embedding vector that matches the deep features. This branch supplements the wave statistical properties and physical generation mechanism information that are difficult for deep networks to express explicitly.

[0031] S5: Bilinear Gated Fusion: A bilinear gated fusion module is constructed to adaptively fuse the features output from the first, second, and third branches. First, the dynamic weights between the first and second feature vectors are calculated through a gated network to obtain the fused backbone features. Second, consistency terms, difference terms, and low-rank bilinear interaction terms are constructed between the first and second feature vectors to enhance the model's ability to express cross-branch complementary information in the precursor features of anomalous waves. Finally, the above deep interaction features are concatenated with the physical prior embedding vector obtained from the third branch to form the final fused features.

[0032] Let the time-span waveform morphology feature vector of the first branch output be: ; The multi-scale dynamic coupling feature vector output by the second branch is: ; The physical prior embedding vector output by the third branch is: ; in, d For depth feature dimension, d f The physical prior embedding dimension.

[0033] Adaptive fusion via a bilinear gated fusion module specifically includes: First, the first feature vector z 1 and second eigenvectors z2. Concatenate the components and input them into the gating network to obtain the dynamic gating weights: ; Among them, [z1; z [2] represents the vector concatenation operation, φ(·) is a non-linear activation function, preferably the GELU activation function, and σ(·) is the Sigmoid activation function. W g1 , W g2 , b g1 , b g2 These are trainable parameters. The gating weights... g ∈R d This is used to adaptively adjust the contributions of the first and second branches under different sea state conditions.

[0034] Based on the gating weights, the backbone fusion features are obtained: ; Here, ⊙ represents element-wise multiplication. This backbone fusion feature can adaptively select more effective feature sources based on the local morphology, multi-scale energy distribution, and background sea state changes of the input wave sequence.

[0035] To further enhance the complementary expressive power between the two deep branches, this invention constructs the following three types of interaction terms.

[0036] First, construct a product consistency term: ; This term characterizes the consistency of the responses of the first and second branches on the same feature dimensions. When cross-time waveform morphology features and multi-scale dynamic coupling features are enhanced simultaneously on certain dimensions, the product consistency term can amplify this type of common response, thereby improving the model's sensitivity to anomalous wave precursor features.

[0037] Second, construct the absolute difference term: ; This term is used to characterize the feature differences between the first and second branches. Since waveform morphology changes and multi-scale dynamic modulations before the occurrence of anomalous waves do not always occur synchronously, the absolute difference term can preserve the complementary information between different feature flows and avoid the cancellation of effective features due to simple weighted fusion.

[0038] Third, construct low-rank bilinear interaction terms: ; in, U ∈R r*d For trainable low-rank mapping matrices,r The low-rank bilinear interaction term represents the bilinear interaction rank. Through this low-rank bilinear mapping, this invention explicitly models the second-order interaction relationship between the first and second branches while controlling parameter size and computational cost. Compared to directly constructing the complete bilinear matrix, the low-rank bilinear interaction term reduces the risk of model overfitting and is more suitable for engineering early warning scenarios where anomalous wave samples are scarce and class imbalances exist.

[0039] Subsequently, the core fusion features, product consistency term, absolute difference term, low-rank bilinear interaction term, and third-branch physical prior embedding vector are concatenated to obtain the final fusion features: ; in, This is the physical prior feature vector output by the third branch.

[0040] The final fused features are input into the classification and discrimination network to obtain the predicted probability of the abnormal wave occurrence. The classification and discrimination network preferably employs a multilayer perceptron structure, and its calculation process can be expressed as follows: ; ; ; ; Where LN(·) represents layer normalization, Dropou(·) represents random deactivation regularization, W1 and b1 are the trainable parameters of the first fully connected layer, W2 and b2 are the trainable parameters of the second fully connected layer, h1 and h2 are the hidden layer feature vectors, s is the logit value output by the classification module, and p is the probability of an abnormal wave occurring within a preset time period. When p ≥ 0.5, an abnormal wave warning is output; when p < 0.5, no abnormal wave occurs is output.

[0041] Depth feature dimensions of the first and second branches d The number of multi-head attention heads can be 128-512, preferably 320; the number of multi-head attention heads can be 4-16, preferably 8; the number of Transformer encoding layers can be 2-6, preferably 4 layers for the first branch and 3 layers for the second branch; the dimension of the hidden layer of the feedforward network can be 256-1024, preferably 512; the Dropout ratio can be 0.05-0.30, preferably 0.16; the DropPath ratio can be 0-0.20, preferably 0.10.

[0042] In the first branch, the time segment length can be 64-256 sampling points, preferably 112 sampling points; the adjacent segment step size can be 16-128 sampling points, preferably 64 sampling points. The segment length and step size are used to control the extraction density of local waveform morphology information and long-range time dependence information.

[0043] In the second branch, the number of intrinsic mode functions obtained by empirical mode decomposition can be 3-8, preferably 6; the number of screening iterations can be 5-20, preferably 10; and the standard deviation threshold in the intrinsic mode function determination can be 0.10-0.30, preferably 0.20.

[0044] In the third branch, the physical prior feature dimension can be 6-15 dimensions, preferably 9 dimensions, including standard deviation, skewness, kurtosis, range, zero-crossing rate, energy, spectral correlation features, spectral bandwidth, and steepness; the physical prior embedding dimension... d f The value can be 32-128, with 64 being the preferred value.

[0045] In the bilinear gated fusion module, the low-rank bilinear interactive rank r The number of values ​​can be 16-128, with 64 being the preferred value. The gated network preferably adopts a two-layer fully connected structure and outputs the dimension-by-dimensional gate weights through the Sigmoid function.

[0046] S6: Model Training: Following the bilinear gated fusion module, a training optimization strategy combining asymmetric loss function, regularization, and test augmentation is employed to optimize the three-branch bilinear gated fusion model, adapting it to early warning scenarios characterized by low probability of anomalies, scarce samples, and class imbalance. A model training strategy suitable for identifying low-probability extreme events is used to optimize the three-branch bilinear gated fusion model.

[0047] The asymmetric loss function is used to set different loss weights or modulation coefficients for anomalous wave samples and normal wave samples, so that the model pays more attention to the recognition ability of low-probability anomalous wave samples during training and reduces the impact of class imbalance on the classification boundary. Preferably, the asymmetric loss can be in the form of asymmetric focal loss, setting different weight parameters for anomalous wave samples and normal wave samples respectively, and applying different degrees of loss modulation to easily classified samples and difficult-to-classify samples.

[0048] The regularization strategy is used to suppress model overfitting and reduce the risk of false positives for normal wave samples. The regularization strategy includes one or more of the following: hard negative sample constraint, Dropout random deactivation, DropPath random path deactivation, weight decay, layer normalization, and exponential moving average. Specifically, the hard negative sample constraint applies additional penalties to normal wave samples that are assigned a higher probability of anomalous waves by the model, thereby improving the model's ability to distinguish easily confused normal wave samples; the exponential moving average smooths the model parameters during training, improving the model's stability on the validation and test sets.

[0049] The test enhancement strategy is used to improve the stability of the output results during the early warning inference stage. Specifically, the input wavefront sequence can be shifted by one or more small time shifts and then input into the trained three-branch bilinear gated fusion model to obtain multiple classification outputs. These outputs are then weighted and averaged to obtain the final classification output. Subsequently, the final classification output is converted into the probability of anomalous wave occurrence. By using the test enhancement strategy, the influence of the single time window interception position on the early warning results can be reduced, thereby improving the robustness of the anomalous wave early warning output.

[0050] When there are insufficient anomalous wave samples from a single target buoy, anomalous wave samples from neighboring buoys or similar sea areas can be introduced to jointly construct a training set with normal samples from the target buoy, thereby enhancing the model's ability to identify anomalous wave events in the target sea area.

[0051] S7: Early Warning Output: The final fused features are input into the classification and discrimination module, which outputs the probability of abnormal wave occurrence. Based on a preset discrimination threshold, the probability of abnormal wave occurrence is converted into a binary classification early warning result. When the probability of abnormal wave occurrence is greater than or equal to the discrimination threshold, an "abnormal wave has occurred" early warning result is output; when the probability of abnormal wave occurrence is less than the discrimination threshold, an "abnormal wave has not occurred" early warning result is output. Preferably, the discrimination threshold is set to 0.5 to meet the safety-first discrimination requirements in engineering applications.

[0052] In some embodiments, the anomalous wave samples are based on the maximum wave height H. max With significant wave height H s The ratio is used for judgment. When a wave record satisfies the condition that the maximum wave height is greater than R0 times the significant wave height, its corresponding warning sample is marked as an abnormal wave event sample; otherwise, it is marked as a normal wave sample. The sample construction process is used to form a binary classification training objective, enabling the model to learn the risk of abnormal wave events, rather than a continuous regression relationship of specific wave height values.

[0053] In some embodiments, the first branch employs a Transformer encoding structure based on time segments. The normalized wavefront sequence is first enhanced locally by a one-dimensional convolutional residual module, and then divided into multiple overlapping time segments according to a set segment length and stride. Each time segment is linearly mapped and positionally encoded before being input into a multi-layer Transformer encoder, and finally, cross-time waveform morphology consistency features are obtained through global average pooling.

[0054] In some embodiments, the second branch employs a structure combining Empirical Mode Decomposition (EMD) with a lightweight iTransformer. EMD decomposes the original wavefront sequence into several intrinsic mode functions and residual terms to form a multi-channel input reflecting the fluctuation process at different time scales; the lightweight iTransformer encoder further learns the interaction relationships between the various mode channels to capture the cross-scale energy transfer and nonlinear modulation characteristics that exist before the occurrence of anomalous waves.

[0055] In some embodiments, the physical prior features extracted by the third branch include time-domain statistics, frequency-domain energy, linear dispersion, and nonlinear effects. These physical prior features are mapped by a multilayer perceptron model and then participate in the final fusion to enhance the adaptability and interpretability of the prediction model to real ocean wave physics.

[0056] In some embodiments, the bilinear gated fusion module does not simply concatenate the outputs of the three branches. Instead, it adaptively adjusts the contributions of temporal morphological features and multi-scale dynamic coupling features through gating weights. Simultaneously, it introduces low-rank bilinear interaction terms, product consistency terms, and absolute difference terms to enhance the complementary expressive power between different feature flows. This fusion method can dynamically adjust feature weights under different sea states, different buoys, and different background noise conditions.

[0057] In some embodiments, to alleviate the training bias caused by the proportion of abnormal wave samples being much lower than that of normal wave samples, an asymmetric loss function is used to assign different weights to positive and negative samples, and hard negative sample constraints are used to reduce the probability that normal wave samples are misclassified as abnormal waves.

[0058] In some embodiments, the classification and discrimination module outputs the probability of an abnormal wave occurrence. This probability represents the likelihood of an abnormal wave event occurring within a future warning interval, given a preceding wavefront time series and a preset time period. This result can be further converted into a "occurrence" or "non-occurrence" warning signal, used for risk alerts in offshore engineering operations, auxiliary decision-making for ship navigation safety, and operation and maintenance management of marine structures.

[0059] Example 2: like Figure 2 As shown, this embodiment is used to verify the advantages of the model (TS) of the present invention compared with conventional deep learning models.

[0060] Using the same buoy wavefront time series data, the early warning effects of CNN, LSTM, BP neural networks, and the proposed three-branch bilinear gated fusion model (TS) were tested. Each model took a fixed-length wavefront sequence before the occurrence of an abnormal wave as input and output whether the abnormal wave occurred within a preset future time period. D1, D2, D3, and D4 were four different datasets, each including the correct prediction rate for abnormal waves, the correct prediction rate for normal waves, and the overall correct prediction rate. Figure 2 (a) shows the predictive power of the TS model of this invention on four different datasets. Figure 2 (b) shows the predictive power of the LSTM model across four different datasets. Figure 2 (c) shows the predictive power of the CNN model across four different datasets. Figure 2 In the middle (d), the prediction ability of the BP model is shown in four different datasets.

[0061] Among these, CNN, LSTM, and BP neural networks can only extract information from a single dimension. The model of this invention, however, simultaneously extracts cross-time waveform morphology features, multi-scale dynamic coupling features, and physical prior features, and fuses them through a bilinear gating mechanism. Predictive analysis of four sets of data samples (D1~D4) from different sea areas revealed that the model of this invention has a stronger anomalous wave early warning capability compared to CNN, LSTM, and BP networks. It can suppress false alarms of normal waves while maintaining the accuracy of anomalous wave identification, indicating that the three-branch fusion structure is more suitable for anomalous wave binary classification tasks and can meet the engineering requirements for early warning in various sea areas.

[0062] Example 3: like Figure 3 As shown, this embodiment is used to verify the stability of the present invention under different early warning durations.

[0063] During sample construction, the occurrence time of the anomalous wave is set at different time points after the input window, allowing the model to be trained and tested under different lead times. The model input remains the wavefront time series before the occurrence of the anomalous wave, and the output remains whether the anomalous wave occurred.

[0064] Under different lead times, the model's prediction and identification performance for normal and anomalous waves was statistically analyzed. D1, D2, D3, and D4 represent four different constructed datasets, with predictions for anomalous waves, normal waves, and the overall correct prediction rate included in each dataset. Figure 3 (a) shows the predictive power of the TS model of this invention at different times in the D1 dataset. Figure 3 (b) shows the predictive power of the TS model of this invention at different times in the D2 dataset. Figure 3 (c) shows the predictive power of the TS model of this invention at different times in the D3 dataset. Figure 3 (d) shows the prediction capability of the TS model of this invention at different times in the D4 dataset. Experimental results show that as the warning lead time increases, the overall performance fluctuation of the model of this invention is small, and there is no significant performance degradation even when predicting 5 minutes in advance, indicating that the model of this invention can adapt to different prediction time scales. Its stable performance meets the actual engineering needs of ocean wave monitoring and abnormal wave warning.

[0065] Example 4: like Figure 4 As shown, this embodiment is used to verify the early warning capability of the present invention under a real sea state sample ratio. Figure 4 In the figure (a), the early warning capability of the invention is shown in real samples. In the figure (b), the number of abnormal wave and normal wave samples is shown in real sea conditions.

[0066] Since anomalous waves are low-probability events, normal wave samples far outnumber anomalous wave samples in real ocean observations. To simulate real-world application scenarios, an imbalanced dataset of anomalous wave samples and normal wave samples is constructed.

[0067] On this dataset, the model of this invention outputs the probability of anomalous wave occurrence for each input wavefront sequence and provides a "occurred" or "not occurred" judgment result based on a fixed threshold. Experimental results show that even under conditions of severe imbalance in event categories, the model can still maintain accurate prediction ability for anomalous waves and has a certain ability to suppress false alarms of normal waves. Analysis of the global imbalanced dataset shows that the model accurately predicted 164,639 normal wave events, with a total of 217,633 normal wave samples under actual sea conditions, accounting for 75.65% of all normal wave samples; at the same time, it accurately predicted 2,016 anomalous wave events, with a total of 2,818 anomalous wave samples under actual sea conditions. The anomalous wave recall rate is 71.54%, and the overall prediction accuracy reaches 75.6%, which is far higher than the prediction accuracy currently published abroad.

[0068] Example 5: like Figure 5 As shown, this embodiment is used to verify the function of each branch in the three-branch structure of the present invention.

[0069] Control models with some branches removed were constructed, including A using only the first branch (cross-time waveform morphology branch), B using only the second branch (multi-scale dynamic coupling branch), C using only the third branch (physical prior statistical feature branch), and models formed by combinations of different branches (AB, AC, BC, ABC). All control models used the same data partitioning and training methods, and the output was the classification result of anomalous wave "occurrence / non-occurrence".

[0070] Figure 5 In (a), the correct prediction rate of anomalous waves in each branch of the TS model of this invention is represented. Figure 5 (b) represents the correct prediction rate of the normal wave in each branch of the TS model of this invention. Figure 5 In Figure (c), the overall correct prediction rate of each branch of the TS model of this invention is represented. Experimental results show that the multi-scale dynamic coupling branch is the core module for anomalous wave identification, which can effectively extract key multi-scale dynamic features from wave sequences; the cross-time waveform morphology branch mainly enhances the ability to identify normal waves and helps reduce false alarms; the physical prior statistical feature branch can enhance the expression of anomalous wave events and the stability of the model. The complete three-branch structure can achieve a good balance between anomalous wave detection and false alarm control.

[0071] Example 6: This embodiment is a practically deployable intelligent early warning system for abnormal waves built based on the aforementioned early warning method. It supports deployment in three scenarios: nearshore buoy monitoring terminals, offshore platform safety management and control systems, and ship navigation auxiliary decision-making systems. The model parameters and output format can be flexibly configured according to the hardware resources and early warning requirements of different application scenarios.

[0072] The actual operation flow of the system is as follows: The data acquisition module connects to the buoy's real-time communication interface and acquires raw sea surface displacement time series data according to a preset sampling frequency. No additional preprocessing equipment is required, and it supports simultaneous data access from multiple buoys. The data preprocessing module performs data verification and standardization operations on the received raw data, automatically removes outliers, fills in missing values, and completes sequence normalization to eliminate energy scale differences under different sea states. The sample construction module generates a sliding time window input based on a preset input time window, and organizes the latest acquired continuous wavefront data into an input format that the model can accept, without waiting for the complete cycle of data acquisition to be completed; The three-branch feature extraction module performs parallel computation of three types of features, and outputs waveform morphology features, multi-scale dynamic coupling features and physical prior features respectively. The feature computation process can be executed in parallel or serial mode according to the hardware resource configuration. The bilinear gated fusion module completes feature fusion by dynamically adjusting the contribution ratio of different features through adaptive gate weights, and outputs the final fused feature vector. The early warning output module outputs early warning results based on the fusion characteristics. When the probability of an abnormal wave occurring is greater than or equal to a preset threshold, an audible and visual early warning is triggered and pushed to the user terminal. It supports multiple notification methods such as SMS, platform pop-ups, and audible and visual alarms.

[0073] The system performs well in actual operation, with a single-sample inference latency of less than 10ms, which can meet the time response requirements of real-time early warning. The early warning time can be flexibly configured within the range of 1 to 10 minutes according to the needs, adapting to the risk handling time requirements of different operation scenarios. It supports parallel monitoring of multiple buoys, and a single edge computing device can simultaneously access the monitoring data of no less than 20 buoys to realize regional synchronous monitoring of abnormal wave risks.

[0074] Example 7: This embodiment uses a physical storage medium that stores the computer program of the abnormal wave intelligent early warning method. The storage medium can be a physical storage medium such as an SD card, USB flash drive, hard disk, read-only memory (ROM), or programmable read-only memory (PROM). The storage medium form can be flexibly selected according to the hardware interface requirements of the application scenario.

[0075] The computer program stored in the storage medium can be directly loaded into embedded edge computing devices, marine monitoring terminals, and industrial control computers for execution without the need for additional deployment of a runtime environment or reliance on third-party software libraries. When executed, it can realize all the steps of the abnormal wave early warning method of the present invention, including the entire process of data acquisition, preprocessing, feature extraction, feature fusion, and early warning output.

[0076] The advantages of this storage medium are: it does not rely on cloud computing power, all calculations are completed on the local terminal, it supports offline operation, it is suitable for marine environments with no or weak network, and it can effectively avoid the risk of early warning failure caused by network delays, interruptions or data transmission leaks. It is especially suitable for independent deployment applications in offshore operation scenarios.

[0077] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model, characterized in that, Includes the following steps: S1: Data Acquisition and Labeling: Acquire the sea surface displacement time series collected by the buoy and label the samples to construct an original wave dataset containing normal wave samples and abnormal wave samples; S2: Data preprocessing: The original wave dataset containing normal wave samples and abnormal wave samples is preprocessed to obtain a standardized wave surface time series; S3: Sample Construction: Construct abnormal wave early warning samples based on the preset input time window and early warning period, so that the model input is a fixed-length wavefront sequence before the occurrence of abnormal waves, and the output is the probability of the occurrence of abnormal waves in the future preset period. S4: Three-branch feature extraction: The first feature vector, the second feature vector, and the third feature vector are extracted from the wavefront sequence using a three-branch feature extraction network; The three-branch feature extraction network in step S4 includes: First branch: After local feature enhancement of the standardized original wavefront sequence, it is divided into multiple overlapping time segments. The local waveform morphology and long-range time dependence are extracted by the time segment Transformer encoder to obtain the cross-time waveform morphology feature vector. The second branch: Empirical mode decomposition is performed on the standardized wavefront sequence to obtain a multi-channel mode matrix composed of multiple intrinsic mode functions and residual terms. The energy distribution, mutual modulation and dynamic coupling relationship between wave components at different scales are learned through a lightweight iTransformer encoder to obtain a multi-scale dynamic coupling feature vector. The third branch extracts low-dimensional physical feature vectors containing temporal statistical features and spectral features from the original wavefront sequence, and maps them to physical prior features through a lightweight multilayer perceptron; the low-dimensional physical feature vectors include at least several indices such as standard deviation, skewness, kurtosis, range, zero-crossing rate, energy, spectral correlation features, spectral bandwidth, and wave steepness. S5: Bilinear Gated Fusion: The first feature vector, the second feature vector, and the third feature vector are adaptively fused through the bilinear gated fusion module to obtain the final fused feature, forming a three-branch bilinear gated fusion model; The bilinear gated fusion process in step S5 specifically includes the following sub-steps: S51: Calculate the dynamic weights between the first and second feature vectors using a gated network to obtain the weighted fusion backbone features; S52: Construct a product consistency term, an absolute difference term, and a low-rank bilinear interaction term between the first and second eigenvectors to enhance the model’s ability to express cross-branch complementary information in the precursor features of anomalous waves. S53: Concatenate the backbone fusion feature, product consistency term, absolute difference term, and low-rank bilinear interaction term with the third feature vector to obtain the final fusion feature; S6: Early warning output: Input the final fused features into the classification and discrimination module, output the probability of abnormal wave occurrence, and convert it into a binary classification early warning result according to the preset discrimination threshold.

2. The intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model according to claim 1, characterized in that, The step S1 of obtaining and labeling the sea surface displacement time series collected by the buoy specifically involves: obtaining the sea surface displacement time series data collected by the buoy in the target sea area, constructing an original wave dataset containing normal wave samples and abnormal wave samples, and performing binary classification labeling on the samples according to a preset abnormal wave criterion; the preset abnormal wave criterion is: when the maximum wave height in the wave record is greater than a preset threshold multiple of the significant wave height, it is marked as an abnormal wave event sample, otherwise it is marked as a normal wave sample.

3. The intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model according to claim 1, characterized in that: In the first branch, the temporal segment length is 64-256 sampling points, the adjacent segment step size is 16-128 sampling points, the temporal segment Transformer encoder has 2-6 layers, and the multi-head attention head has 4-16. In the second branch, the number of intrinsic mode functions obtained from empirical mode decomposition is 3-8, the number of sieving iterations is 5-20, the intrinsic mode function judgment standard deviation threshold is 0.10-0.30, the number of lightweight iTransformer encoder layers is 2-5, and the hidden layer dimension of the feedforward network is 256-1024.

4. The intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model according to claim 1, characterized in that: The low-rank bilinear interaction term is obtained by linearly projecting the first eigenvector and the second eigenvector onto a trainable low-rank mapping matrix and then multiplying them element by element. The bilinear interaction rank ranges from 16 to 128.

5. The intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model according to claim 1, characterized in that: The warning output in step S6 is specifically as follows: the final fused features are input into the classification and discrimination network to obtain the predicted probability of the abnormal wave occurrence. According to the preset discrimination threshold, the predicted probability of the abnormal wave occurrence is converted into a binary classification warning result. When the probability of the abnormal wave occurrence is greater than or equal to the discrimination threshold, the warning result of "abnormal wave occurrence" is output; when the probability of the abnormal wave occurrence is less than the discrimination threshold, the warning result of "abnormal wave does not occur" is output.

6. The intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model according to claim 1, characterized in that: Step S5 is followed by a model training step: the three-branch bilinear gated fusion model is optimized using a training strategy suitable for low-probability extreme event recognition; the training strategy includes a training optimization strategy that combines an asymmetric loss function, regularization, and test enhancement. To address the issue of insufficient anomalous wave samples from a single target buoy, anomalous wave samples from neighboring buoys or similar sea areas are introduced to jointly construct a training set with normal samples from the target buoy.

7. An intelligent early warning system for abnormal waves based on a three-branch bilinear gated fusion model, characterized in that, include: The data acquisition module is used to acquire the sea surface displacement time series collected by the buoy and label the samples to construct an original dataset containing normal wave samples and anomalous wave samples. The data preprocessing module is used to preprocess the original dataset to eliminate data noise and scale differences; The sample construction module is used to construct early warning samples based on a preset input time window and early warning period, so that the model input is a fixed-length wavefront sequence before the occurrence of abnormal waves, and the output is the probability of the occurrence of abnormal waves in the future preset period. The three-branch feature extraction module is used to extract waveform morphology features, multi-scale dynamic coupling features, and physical prior features from the preprocessed wavefront sequence, respectively. The first branch: After local feature enhancement of the standardized original wavefront sequence, it is divided into multiple overlapping time segments. The local waveform morphology and long-range time dependence are extracted by the temporal segment Transformer encoder to obtain the cross-time waveform morphology feature vector. The second branch: Empirical mode decomposition is performed on the standardized wavefront sequence to obtain a multi-channel mode matrix composed of multiple intrinsic mode functions and residual terms. The energy distribution, mutual modulation and dynamic coupling relationship between wave components at different scales are learned through a lightweight iTransformer encoder to obtain a multi-scale dynamic coupling feature vector. The third branch extracts low-dimensional physical feature vectors containing temporal statistical features and spectral features from the original wavefront sequence, and maps them to physical prior features through a lightweight multilayer perceptron; the low-dimensional physical feature vectors include at least several indices such as standard deviation, skewness, kurtosis, range, zero-crossing rate, energy, spectral correlation features, spectral bandwidth, and wave steepness. A bilinear gated fusion module is used to adaptively fuse the waveform morphology features, multi-scale dynamic coupling features, and physical prior features to obtain the final fused features. A gated network is used to calculate the dynamic weights between the first and second feature vectors to obtain the weighted fused backbone fused features. A product consistency term, an absolute difference term, and a low-rank bilinear interaction term are constructed between the first and second feature vectors to enhance the model's ability to express cross-branch complementary information in the precursor features of anomalous waves. The backbone fused features, product consistency term, absolute difference term, and low-rank bilinear interaction term are concatenated with the third feature vector to obtain the final fused features. The early warning output module is used to input the final fused features into the classification network and output the probability of abnormal wave occurrence within a preset time period. When the probability is greater than or equal to the preset early warning threshold, an abnormal wave early warning result is generated.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the intelligent early warning method for abnormal waves based on a three-branch bilinear gated fusion model as described in any one of claims 1 to 6.

Images