Ocean extreme wave measurement method based on multi-source data driving

Abstract

The invention discloses an ocean extreme wave measurement method based on multi-source data driving, and the method comprises the steps: obtaining multi-source heterogeneous monitoring data of a target sea area, which at least comprises buoy motion attitude data collected in real time and multi-source original wave observation data; calculating a working condition response index reflecting the stability degree of the buoy, and establishing a mapping relation between the working condition response index and the data credibility to obtain a real-time reliability coefficient for the multi-source original wave observation data; constructing a dynamic observation variance by using the real-time reliability coefficient, and performing Bayesian fusion processing on the multi-source original wave observation data to obtain a wave element fusion observation value; and inputting the wave element fusion observation value into a pre-trained extreme wave prediction model, and outputting an extreme wave element prediction result of the target sea area. According to the method, the problems of data distortion and model physical consistency deficiency caused by strenuous movement of the buoy under the extreme sea condition are effectively solved, and the prediction precision of extreme wave elements is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of ocean monitoring, and particularly relates to a method for measuring extreme ocean waves based on multi-source data driving. BACKGROUND

[0002] As key equipment for obtaining wave elements, the real-time and accuracy of the observation data of the ocean monitoring buoy are crucial for disaster prevention and mitigation. However, in extreme sea conditions, the buoy will produce violent pitching, rolling and heaving movements, and even submersion due to water increase, which puts high requirements on the long-term stability and data continuity of the observation system.

[0003] Currently, for the monitoring and prediction of ocean wave elements, the original observation data is preprocessed by statistical methods, such as using the threshold method based on Z-Score to remove outliers, and using Kalman filtering or standard Bayesian estimation to fuse multi-source data. In terms of prediction models, pure data-driven algorithms such as radial basis function (RBF) neural network or support vector machine (SVM) are often used to minimize the mean square error to fit the wave evolution law in historical data.

[0004] However, the existing technology has a significant physical mechanism missing and data trust mismatch problem when dealing with extreme sea conditions. Mainly reflected in: the existing fusion method usually assumes that the sensor has a static or statistically fixed measurement variance, ignoring the strong coupling relationship between sensor reliability and buoy instantaneous motion posture. When the buoy is violently shaken by extreme big waves, the sensor is actually in a physical saturation or high noise state, and if it is still fused according to the conventional weight, it will cause distorted data to pollute the fusion result. In addition, pure data-driven models lack physical constraints, and when predicting extreme values (tail samples) with sparse samples, they are prone to output prediction values that violate the positive correlation physical law between anchor tension and wave height under similar sea conditions, resulting in poor generalization ability and low reliability of the model in the most critical extreme working conditions. SUMMARY

[0005] The application aims to provide a method for measuring extreme ocean waves based on multi-source data driving to solve the above problems existing in the prior art.

[0006] The technical scheme is a method for measuring extreme ocean waves based on multi-source data driving, comprising:

[0007] Obtaining multi-source heterogeneous monitoring data of a target sea area, at least including real-time collected buoy motion posture data and multi-source original wave observation data;

[0008] A working condition response index reflecting the stability of the buoy is calculated based on the buoy motion posture data, and a mapping relationship between the working condition response index and the data reliability is established to obtain a real-time reliability coefficient for the multi-source original wave observation data;

[0009] A dynamic observation variance is constructed using the real-time reliability coefficient, and based on this, a Bayesian fusion processing is performed on the multi-source original wave observation data to obtain a wave element fusion observation value;

[0010] The wave element fusion observation value is input into a pre-trained extreme wave prediction model to output an extreme wave element prediction result of the target sea area.

[0011] Beneficial effects, the present application effectively solves the problem of data distortion caused by the violent movement of the buoy and the lack of model physical consistency in extreme sea conditions, and improves the prediction accuracy of the extreme wave element. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 A step flowchart of a marine extreme wave measurement method based on multi-source data driving provided by the embodiment of the present application.

[0013] Figure 2 A step flowchart of calculating the working condition response index provided by the embodiment of the present application.

[0014] Figure 3 A step flowchart of obtaining the real-time reliability coefficient provided by the embodiment of the present application.

[0015] Figure 4 A step flowchart of obtaining the wave element fusion observation value provided by the embodiment of the present application. DETAILED DESCRIPTION

[0016] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0017] It should be noted that the terms first, second, etc. in the description of the present application are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms include and have and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to the clearly listed steps or units, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0018] As shown in Figure 1 A multi-source data-driven ocean extreme wave measurement method includes the following steps:

[0019] Obtain multi-source heterogeneous monitoring data of the target sea area, and the multi-source heterogeneous monitoring data at least includes real-time collected buoy motion attitude data and multi-source original wave observation data.

[0020] In other words, obtain multi-source heterogeneous monitoring data containing real-time collected buoy motion attitude data and multi-source original wave observation data in the target sea area, perform preprocessing to obtain preprocessed multi-source heterogeneous monitoring data.

[0021] Specifically, a marine monitoring network including a plurality of sensor terminals is constructed. The buoy motion attitude data is usually derived from an inertial measurement unit (IMU) or an attitude heading reference system (AHRS) installed inside the marine buoy, and specifically includes high-frequency time series data such as roll angle (Roll), pitch angle (Pitch), and vertical heave displacement (Heave), with a sampling frequency preferably set to 1 Hz to 10 Hz to capture instantaneous violent shaking. The multi-source original wave observation data covers data sources of different observation principles, such as measured wave height data from an accelerometer wave meter carried by the buoy, and remotely sensed wave height data from a satellite altimeter or an unmanned aerial synthetic aperture radar (SAR). Optionally, the multi-source heterogeneous monitoring data can also include real-time collected anchor rope tension data and buoy observation data, as well as remotely sensed image data. After obtaining these data, perform basic data cleaning operations, such as standardizing each channel data using the Z-Score method.

[0022] It should be noted that considering that the real giant wave signal under extreme sea conditions may deviate greatly from the historical mean, the standardization process preferably adopts a dynamic threshold strategy, i.e. in the extreme sea condition early warning state, automatically relax the abnormal exclusion threshold from the conventional 3 times standard deviation to 6 times standard deviation or even higher, to maximize the retention of real extreme physical signals and avoid misjudgment of physical extreme values as sensor noise.

[0023] A working condition response index reflecting the stability of the buoy is calculated based on the buoy motion posture data, and a mapping relationship between the working condition response index and the data reliability is established to obtain a real-time reliability coefficient for the multi-source original wave observation data.

[0024] In this embodiment, the working condition response index is a comprehensive index for quantifying the severity of the physical environment in which the buoy is currently located, and can identify whether the buoy is in a state of violent motion beyond the normal working range, because such a state often leads to sensor saturation or measurement distortion. Instead of simply using the instantaneous posture value at a certain moment, the system statistically analyzes the posture changes over a period of time to obtain a more representative stability evaluation. For reliability mapping, a differentiated processing strategy is preferably adopted. For buoy observation data, the working condition response index is used as a reliability evaluation index, and a first real-time reliability coefficient is calculated through a logic function, meaning that the more violent the buoy sways, the lower the reliability weight of the wave data collected by the buoy. For remote sensing image data, the time lag of the data acquisition time relative to the current time is calculated, and the time lag is used as a reliability evaluation index to calculate a second real-time reliability coefficient through a logic function. The non-real-time nature of remote sensing data is considered, i.e., the older the data, the lower its reference value for the current time. It can dynamically assign trust weights to data of different sources according to their physical states.

[0025] A dynamic observation variance is constructed using the real-time reliability coefficient, and a Bayesian fusion process is performed on the multi-source original wave observation data based on the dynamic observation variance to obtain a wave element fused observation value.

[0026] Specifically, the Bayesian inference framework is used to achieve optimal estimation of multi-source data. The dynamic observation variance is a bridge connecting the physical state and the statistical model, and the real-time reliability coefficient can be used as a scaling factor for the observation variance. When the reliability coefficient of a certain data source tends to zero, the corresponding dynamic observation variance will tend to infinity. In the Bayesian fusion formula, the data item with a larger variance has a smaller contribution weight to the posterior mean. Conversely, a data source with high reliability will be assigned a smaller variance, thus dominating the fusion result. An adaptive weighting mechanism based on physical perception is achieved: when the buoy is violently swaying but the remote sensing data is relatively new, the system will automatically trust the remote sensing data more than the buoy data; when the buoy is stable but the remote sensing data is severely lagging, the system will automatically trust the buoy data more than the remote sensing data. For the case of complete failure of some sensors leading to data missing, the system can directly set the reliability coefficient to zero, and the Bayesian fusion algorithm can naturally handle such missing items without complex manual filling.

[0027] The wave element fused observation value is input into a pre-trained extreme wave prediction model to output a prediction result of extreme wave elements in the target sea area.

[0028] In the present embodiment, the extreme wave prediction model preferably employs a machine learning model with non-linear mapping capability, such as a radial basis function (RBF) neural network. The extreme wave prediction model is pre-trained in an offline stage using a large amount of historical extreme sea state data, and its input includes not only the fused wave observation values, but also other environmental features that have been screened. The output prediction results specifically include key wave elements such as significant wave height (Hs), spectral peak period (Tp), and mean wave direction in a future period of time (e.g. 10 to 30 minutes). By using data fused from multiple sources with high reliability as input, the extreme wave prediction model can reduce the prediction bias caused by the failure or noise interference of a single sensor, and especially in extreme adverse environments such as typhoon passage, it can provide more robust wave prediction than traditional single data source methods.

[0029] As shown in Figure 2 In one possible implementation, the multi-source heterogeneous monitoring data further includes real-time collected anchor rope tension data; a working condition response index reflecting the stability of the buoy is calculated based on the buoy motion posture data, including:

[0030] A sliding time window statistical analysis is performed on the buoy motion posture data, the standard deviation of the posture in the window is calculated, and the posture statistical characteristic quantity is obtained;

[0031] The ratio of the anchor rope tension data to the preset limit tension is calculated, and the tension load ratio is obtained;

[0032] The working condition response index is generated based on the weighted combination of the posture statistical characteristic quantity and the tension load ratio.

[0033] Specifically, in order to accurately capture the dynamic response of the buoy, a sliding time window is set, and the window length is preferably set to 10 minutes and the sliding step is 1 minute. In each time window, the standard deviation of the roll (Roll), pitch (Pitch) and heave (Heave) sequence is calculated respectively. The standard deviation can effectively reflect the dispersion degree of the data around the mean value, i.e. the severity of the buoy shaking. At the same time, the anchor rope tension is introduced as another key dimension. The anchor rope tension reflects the mechanical load of the sea wave on the buoy system, and when the tension approaches the design limit, it often means that the sea state is extremely severe. The tension load ratio is the ratio of the current measured tension mean value to the anchor rope design breaking force. The combination of the posture statistical characteristic quantity and the tension load ratio can construct a multi-dimensional working condition evaluation system, avoiding possible misjudgment of a single posture index.

[0034] In one preferred implementation, the working condition response index is calculated by the following formula:

[0035] EWI(t) = a std roll (t) + b std pitch(t) + γ · std heave (t) + δ · (T(t) / T max );

[0036] wherein EWI(t) represents the working condition response index at time t, std roll (t), std pitch (t), std heave (t) respectively represent the standard deviation of pitch, the standard deviation of roll and the standard deviation of heave in the attitude statistical feature quantity, T(t) represents the anchor rope tension data at time t, T max represents the preset limit tension, and α, β, γ, δ are pre-configured non-negative weight coefficients.

[0037] In the embodiment, a specific linear weighting calculation formula of the working condition response index is given. The selection of each weight coefficient α, β, γ, δ depends on the specific type of the buoy and the characteristics of the deployment sea area. As a preferred embodiment, α = 0.3, β = 0.3, γ = 0.2, δ = 0.2 can be set, indicating that the contribution of pitch and roll to working condition determination is slightly greater than that of heave and tension.

[0038] For ease of understanding, a specific numerical calculation case is provided below. Assuming that at a certain time t, the attitude data obtained by sliding window statistics is as follows: the standard deviation of pitch std roll (t) = 15 degrees, the normalized value is 0.5; the standard deviation of roll std pitch (t) = 10 degrees, the normalized value is 0.3; the standard deviation of heave std heave (t) = 2.5 meters, the normalized value is 0.8; at the same time, the measured anchor rope tension T(t) = 160 kN, and the design limit tension T max of the anchor rope of the buoy is 200 kN. The tension load ratio is calculated as: Ratio tension = 160 / 200 = 0.8. The above numerical values are substituted into the formula for calculation: EWI(t) = 0.3*0.5 + 0.3*0.3 + 0.2*0.8 + 0.2*0.8 = 0.15 + 0.09 + 0.16 + 0.16 = 0.56. In this numerical calculation case, the calculated working condition response index is 0.56. The system can compare this value with the preset threshold, for example, the preset threshold is 0.5. Since 0.56 is greater than 0.5, the system will determine that the buoy is in a severe motion working condition at the current time, and then trigger the subsequent reliability reduction mechanism. This enables the system to accurately perceive the severity of the environment, providing a solid physical basis for subsequent data fusion.

[0039] As Figure 3As shown, according to one aspect of the present application, a mapping relationship between the working condition response index and the data reliability is established, to obtain a real-time reliability coefficient for the multi-source original wave observation data, including:

[0040] A logic function is constructed, which defines a negative correlation between the working condition response index and the real-time reliability coefficient;

[0041] The working condition response index is mapped to a normalized interval using the logic function, and the real-time reliability coefficient corresponding to each multi-source original wave observation data is calculated; wherein the greater the value of the working condition response index, the closer the corresponding real-time reliability coefficient is to zero.

[0042] In the present embodiment, a conversion mechanism is defined, i.e. how to convert the physical state quantity into a statistical weight, i.e. how to convert the working condition response index into the reliability coefficient. Preferably, the logic function selects a variant of the S-type function (Sigmoid function) to achieve a non-linear mapping effect. For example, for buoy observation data, its reliability coefficient c buoy (t) can be represented as:

[0043] c buoy (t) = 1 / (1+exp(k*(EWI(t)-E0)));

[0044] wherein k is a parameter controlling the steepness of the curve, and E0 is a working condition threshold. When the working condition response index EWI(t) is much smaller than E0, the denominator approaches 1, and the reliability coefficient approaches 1, indicating that the data is highly reliable; when the working condition response index EWI(t) exceeds E0, the denominator increases rapidly, and the reliability coefficient quickly decreases to approach 0, indicating that the data is unreliable. In line with the physical law that the measurement error of the sensor will increase exponentially when the sea state severity exceeds a certain limit.

[0045] In an exemplary embodiment, the multi-source heterogeneous monitoring data includes buoy observation data and remote sensing image data, and the real-time reliability coefficient corresponding to each multi-source original wave observation data is calculated, including:

[0046] For buoy observation data, the working condition response index is taken as the first reliability evaluation index, and the first real-time reliability coefficient is calculated through the logic function;

[0047] For remote sensing image data, the time lag of the data acquisition time relative to the current time is calculated, and the time lag is taken as the second reliability evaluation index, and the second real-time reliability coefficient is calculated through the logic function.

[0048] Specifically, for remote sensing image data, the main source of unreliability is not the motion attitude, but the time lag of the data. Therefore, the second real-time reliability coefficient c remoteThe logic function of (t) can be expressed in exponentially decaying form: c remote (t) = exp(-λ*Δt); where Δt is the time lag in minutes; and λ is the attenuation coefficient. To illustrate this more clearly, a numerical demonstration will be provided using the above numerical calculation example. Assume the current time is T. now The calculated EWI(t) is 0.56. Parameters are set as k=10 and E0=0.5. Substituting these values ​​into the buoy reliability formula: c buoy =1 / (1+exp(10*(0.56-0.5))) =1 / (1+exp(0.6))c buoy ≈1 / (1+1.82)≈0.35. The result shows that due to the relatively violent movement of the buoy, its data reliability is only 0.35. Assume the system has just received a wave field map retrieved from a satellite, acquired at time T. now -2 minutes, i.e., Δt=2. Set the attenuation coefficient λ=0.1. Substitute into the remote sensing reliability formula: c remote =exp(-0.1*2)=exp(-0.2)≈0.82. The result shows that although the data is delayed by 2 minutes, its reliability is still as high as 0.82. remote (0.82) is much larger than c buoy The value of (0.35) indicates that in subsequent data fusion steps, the system will assign greater weight to remote sensing data, while primarily ignoring the buoy's own observations. This effectively avoids the risk of overall prediction failure due to the deterioration of a single data source, demonstrating the flexibility in handling heterogeneous data.

[0049] like Figure 4 As shown, in one embodiment of this application, a dynamic observation variance is constructed using a real-time reliability coefficient, and Bayesian fusion processing is performed on multi-source raw wave observation data based on the dynamic observation variance to obtain fused wave element observations, including:

[0050] Obtain the pre-configured baseline variance corresponding to each multi-source raw wave observation data;

[0051] The dynamic observation variance is obtained by calculating the ratio of the baseline variance to the real-time reliability coefficient.

[0052] Based on the dynamic observation variance and the preset prior distribution parameters, the Bayesian posterior mean is calculated and used as the fused observation value of wave elements.

[0053] In particular, in traditional Bayesian fusion, the measurement variance of each sensor is usually assumed to be a fixed value. However, this assumption is no longer valid in extreme sea conditions. This embodiment introduces a pre-configured baseline variance, which is determined by the long-term statistical error of each sensor in factory calibration or calm sea conditions. For example, the baseline variance of an accelerometer wave gauge can be set to 0.04 square meters, while the baseline variance of remote sensing inversion can be set to 0.25 square meters. The baseline variance is dynamically inflated using the real-time reliability coefficient. The calculation formula is expressed as the baseline variance divided by the real-time reliability coefficient. When the real-time reliability coefficient approaches 1, the dynamic observation variance approaches the baseline variance, indicating that the data is at a normal accuracy level; when the real-time reliability coefficient drops to 0.1 due to the severe rocking of the buoy, the dynamic observation variance will be expanded by 10 times. The dramatic inflation of the variance means that the data source becomes extremely flat in the probability distribution, and the amount of effective information it contains is drastically reduced, and it is automatically marginalized in subsequent fusion calculations. This embodiment discloses the internal mechanism of converting data reliability into statistical variance.

[0054] In a preferred embodiment, the Bayesian posterior mean is calculated by the following formula:

[0055] H fused (t) = (μ0 / σ0 2 +Σ k [ H k (t) / σ k 2 (t) ]) / (1 / σ0 2 +Σ k [1 / σ k 2 (t)]);

[0056] where H fused (t) represents the wave element fusion observation value at time t, μ0 and σ0 2 represent the mean and variance of the preset prior distribution respectively, H k (t) represents the original wave observation data of the kth data source, σ k 2 (t) represents the dynamic observation variance corresponding to the kth data source, and Σ represents the summation of all effective data sources.

[0057] In this embodiment, the analytical solution formula of Bayesian maximum a posteriori estimation is given, which is expressed as precision weighted average, that is, the weight of each item of data is proportional to the inverse of its variance. In order to more clearly show the superiority of the algorithm in rejecting abnormal data, a detailed numerical example is provided. It is assumed that the system is predicting significant wave height H, and the preset prior distribution parameters are: historical mean μ0= 4 meters, prior variance σ0 2= 1. Current time t receives two observations: data source one (buoy wave gauge): observation H1 = 10 meters. Due to the buoy's posture is violent, calculate its real-time reliability coefficient c1 = 0.1. Known its reference variance σ base1 2 = 0.25. Data source two (remote sensing data): observation H2 = 6 meters. Since the data is new and not affected by posture, calculate its real-time reliability coefficient c2 = 0.9. Known its reference variance σ base2 2 = 0.25. Calculate the dynamic observation variance: buoy dynamic variance σ1 2 = 0.25 / 0.1 = 2.5. Remote sensing dynamic variance σ2 2 = 0.25 / 0.9 ≈ 0.278. Calculate the numerator of the Bayesian fusion formula (weighted sum): prior term: 4 / 1 = 4. Buoy term: 10 / 2.5 = 4. Remote sensing term: 6 / 0.278 ≈ 21.58. The sum of the numerator = 4 + 4 + 21.58 = 29.58. Calculate the denominator (weight sum): prior weight: 1 / 1 = 1. Buoy weight: 1 / 2.5 = 0.4. Remote sensing weight: 1 / 0.278 ≈ 3.60. The sum of the denominator = 1 + 0.4 + 3.60 = 5.0. Get the fusion result: H fused = 29.58 / 5.0 ≈ 5.92 meters. It can be seen that although the buoy reported a 10-meter wave value, which may be a false measurement due to the sensor itself rolling with the wave, but due to its extremely low reliability, the final fusion result 5.92 meters is very close to the observation value 6 meters of remote sensing. If the traditional arithmetic average is used, the result will be (10+6) / 2 = 8 meters, which will bring a huge prediction error. This embodiment successfully uses physical state information to suppress noise interference. In addition, for the case where a certain data source is completely lost, for example, H k is null, in actual engineering implementation, the corresponding reliability coefficient can be directly set to 0, so that this term naturally disappears in the summation formula, ensuring the robustness of the algorithm.

[0058] In one possible embodiment, before inputting the wave element fused observation value into the pre-trained extreme wave prediction model, further comprising:

[0059] calculating a sample statistical weight based on the working condition response index, the sample statistical weight including a weighted term for the extreme working condition;

[0060] calculating a weighted correlation coefficient between the pre-set candidate input feature and the wave element fused observation value, the weighted correlation coefficient being weighted processed on the covariance by using the sample statistical weight;

[0061] The candidate input features with a weighted correlation coefficient greater than a preset threshold are determined as an effective feature set inputting into the extreme wave prediction model.

[0062] Specifically, when constructing the input vector of the extreme wave prediction model, dozens of candidate features are usually extracted from the original monitoring data, such as acceleration, angular velocity, anchor rope tension, wind speed, etc. on each axis. The traditional feature screening method often uses the Pearson correlation coefficient. However, in marine data, 99% of the samples belong to ordinary sea conditions with calm wind and waves, and only 1% belong to extreme sea conditions such as typhoon. The features screened by the traditional method are often features that perform well in calm sea conditions, while ignoring key features that show strong correlation only in extreme sea conditions. For example, horizontal acceleration has good correlation with wave height in small waves, but is often filled with irregular collision noise in huge waves; on the contrary, anchor rope tension does not change significantly in small waves, but can accurately reflect wave energy in huge waves. In order to solve this problem, the sample statistical weight w(t) is introduced in this embodiment. Exemplarily, the calculation formula of the sample statistical weight can be set as:

[0063] w(t) = 1 + λ·EWI(t);

[0064] where λ is an adjustment coefficient, for example, set to 10. It means that the weight of the extreme sample with a working condition response index of 1 in the correlation analysis is 11 times that of the ordinary sample. In calculating the weighted correlation coefficient, the weight is used to weight the covariance matrix. As an example, suppose the system is evaluating the vertical heave velocity as a candidate feature. In ordinary correlation analysis, its coefficient may be only 0.6. But after introducing the weighting term for extreme working conditions, since the feature maintains a high degree of synchronization with the significant wave height during typhoon, its weighted correlation coefficient may be improved to 0.85. If the preset threshold is set to 0.8, the feature will be successfully selected. On the contrary, for some sensor channels that fail in extreme working conditions, their weighted correlation coefficients will be significantly lower than the ordinary correlation coefficients, and thus will be excluded. By constructing a simplified feature set dedicated to large wave prediction, both the model calculation amount and the anti-interference ability are improved.

[0065] In an exemplary embodiment, suppose the system is evaluating the correlation between the vertical heave velocity and the significant wave height, and there are 100 historical samples. The statistical weight of each sample is calculated. Set the adjustment coefficient λ = 10. For the i-th sample, if its working condition response index EWI i = 0.8, the weight w i = 1 + 10 × 0.8 = 9; if EWI i = 0.1, the weight w i = 1 + 10 × 0.1 = 2. Calculate the weighted covariance. Set the heave velocity sequence as V, and the significant wave height sequence as H, and the weighted mean values are respectively: Vw =∑(w i ×V i ) / ∑w i , H w =∑(w i ×H i ) / ∑w i . The weighted covariance Cov w (V, H) = ∑(w i ×(V i -V w )×(H i -H w )) / ∑w i . Assuming that through the above calculation, the ordinary Pearson correlation coefficient is 0.6, and the weighted correlation coefficient is 0.85. Since 0.85 is greater than the preset threshold 0.8, the feature is included in the effective feature set. It shows that although the feature has general correlation in the ordinary sample, it has strong synchronism with the wave height in the extreme sea state, and should be retained.

[0066] The embodiment solves the problem that the traditional correlation analysis is invalid when the samples are unbalanced, and ensures that the selected features are still effective in the extreme sea state.

[0067] According to one aspect of the present application, the extreme wave prediction model is a radial basis function neural network, and the extreme wave prediction model is obtained by pre-training through a particle swarm optimization algorithm; the particle swarm optimization algorithm uses a fitness function to evaluate the quality of particles, and the fitness function includes a data error term and a physical constraint penalty term, and the physical constraint penalty term is used to constrain the model prediction result to comply with a preset physical monotonicity rule.

[0068] In the embodiment, a machine learning training framework that fuses physical mechanisms is constructed. The RBF neural network is used as a basic model, and its structure includes an input layer, a hidden layer and an output layer. In order to determine the number, position and bandwidth parameters of the hidden layer center, a particle swarm optimization algorithm (PSO) is used for global optimization. In PSO, each particle represents a set of potential model parameter configurations. The particle swarm optimization algorithm needs to construct a fitness function to evaluate the quality of particles, and the fitness function not only requires the model prediction value to be as close to the true value as possible (data error term), but also forces the input-output relationship of the model to comply with the physical common sense of ocean engineering (physical constraint penalty term). It effectively prevents the phenomenon of overfitting or violation of physical laws in the extreme value interval of sparse samples in pure data-driven models.

[0069] In a preferred implementation, the calculation process of the data error term includes:

[0070] The loss weight of the pre-stored training sample is determined based on the working condition response index, and the loss weight increases with the increase of the working condition response index;

[0071] The error squares between the predicted values of the calculation model and the real observed values are weighted and summed by using loss weights to obtain a data error term.

[0072] Specifically, the calculation of the data error term follows the tail risk concern principle. The ordinary mean square error (MSE) treats all samples equally, resulting in the model often sacrificing the prediction accuracy of extreme values to obtain the optimal overall mean. By introducing the loss weight, the optimization algorithm is forced to focus on the dangerous samples with higher EWI. For example, the prediction error of the typhoon sample with EWI=0.9 will be amplified by 10 times or even more in the fitness function. The trained model may have a slightly larger error when predicting a 1-meter wave, but will be very accurate when predicting a 10-meter wave, meeting the actual needs of disaster prevention and reduction.

[0073] In another preferred implementation, the calculation process of the physical constraint penalty term includes:

[0074] Selecting a sample pair with similar working conditions and different anchor rope tensions from the pre-stored training data set;

[0075] For each sample pair, determining whether the predicted wave height of the extreme wave prediction model for the anchor rope tension sample higher than the preset threshold is less than the predicted wave height of the anchor rope tension sample lower than the preset threshold;

[0076] If yes, it is determined that the sample pair violates the physical monotonicity, and the square of the predicted wave height difference is calculated as a penalty value added to the physical constraint penalty term.

[0077] Exemplarily, selecting a sample pair with similar working conditions and different anchor rope tensions from the training data set includes: calculating the absolute value of the difference between the working condition response indices of any two training samples; if the absolute value is less than a preset working condition tolerance threshold, and the anchor rope tension difference of the two training samples is greater than a tension difference threshold, the two training samples are selected as a sample pair.

[0078] Specifically, although it is generally believed that the greater the anchor rope tension, the greater the wave height, but it may fail in the mixed surge (long period) and wind wave (short period) or different wave directions. Therefore, in addition to requiring similar working condition response indices (for example, the difference is less than 0.05), the embodiment also preferably adds the similarity judgment of wave period and wave direction when constructing the sample pair. Exemplarily, the system traverses the training set to find a sample pair (A, B) that satisfies the following conditions: similar working conditions: |EWI A - EWI B | < 0.05; similar periods: |Tp A - Tp B | < 1 second; similar wave directions: |Dir A- Dir B | < 30 degrees; tension difference is significant: T A > T B + 20 kN; wherein Dir A and Dir B are the wave propagation directions of sample A and sample B; T A and T B are the anchor tensions of sample A and sample B. Only when all the above conditions are met, sample A and sample B are considered to be under the control of similar physical mechanisms, and at this time, the physical law that the greater the tension, the greater the wave height has a strong constraint. If under this condition, the model predicts that H pred_A < H pred_B , that is, the greater the tension, the smaller the predicted wave height, which means that the model has learned the wrong mapping relationship; wherein H pred_A and H pred_B represent the predicted wave heights of sample A and sample B by the extreme wave prediction model. At this time, the system calculates (H pred_B -H pred_A ) 2 as a penalty term R phys , and adds it to the total fitness function after multiplying it by a larger penalty coefficient λ phys .

[0079] For example, the tension of sample A is 150 kN, and that of sample B is 100 kN, and the periods and wave directions of the two are consistent. The model currently predicts that the wave height of sample A is 6 meters, and that of sample B is 7 meters. Obviously, it violates the physical common sense. The penalty value (7 - 6) 2 =1. Assuming λ phys =100, the fitness function increases by 100. In the next iteration of PSO, the particle will have to move in the direction of correcting this error until the model can correctly output the wave height of sample A greater than sample B. The prior knowledge of ocean dynamics is successfully embedded in the weights of the neural network.

[0080] In an embodiment of the present application, it further comprises improving the adaptability of the extreme wave prediction model to extreme sea conditions through an online feedback mechanism, comprising: inputting the wave element fusion observation value into the pre-trained extreme wave prediction model to output the extreme wave element prediction result of the target sea area.

[0081] Specifically, in the online running phase, the system receives data streams from various sensors in real time. Once the working condition response index (EWI) of the buoy is monitored to exceed a preset warning threshold, for example, 0.6, the system enters an extreme prediction mode. At this time, the wave element fused observation value will be loaded into the pre-trained RBF neural network model. In order to improve real-time performance, the inference process of this model is usually deployed on a high-performance server on the shore or an embedded edge computing module inside the buoy. The output prediction result not only contains the numerical value of the significant wave height and the spectral peak period, but preferably also contains the confidence interval of the prediction result. The confidence interval is calculated based on the real-time reliability coefficient of the input data. When the reliability of the input data is low, the model will output a wider confidence interval to alert the decision maker to the risk.

[0082] Further, an extreme sample library-based model iterative optimization mechanism is also provided. Since extreme sea conditions (such as typhoons) belong to small probability events, the coverage of the initial training data is often insufficient. Therefore, the system maintains a dynamically updated extreme sample library during operation. Whenever the system detects a high-risk sea condition with EWI greater than a certain threshold (for example, 0.8), the system automatically packages and stores all original monitoring data, fused data, and real wave elements verified by high-precision means during this period to form new standard extreme samples; the high-precision means can be post-download non-delay satellite data or passing ship reports. When the number of new samples in the extreme sample library reaches a preset update batch, or before the arrival of the typhoon season every year, the system will automatically trigger an offline retraining process. The offline retraining process calls the PSO-RBF training method with physical constraints to fine-tune the model parameters using the expanded sample library. The model can continuously learn new sea condition characteristics, such as adapting to new wave spectrum characteristics caused by climate change, thereby realizing the continuous evolution of system performance. This allows the system to maintain high accuracy in complex and changing marine environments for a long time.

[0083] Optionally, during system operation, the following boundary condition and abnormal situation handling strategies should also be considered: when the working condition response index EWI exceeds 0.95, it indicates that the buoy may be at the edge of overturning or submersion, at which time the system enters an extreme conservative mode, forcibly sets the reliability coefficient of all data collected by the buoy to below 0.05, mainly relies on remote sensing data or historical priori for prediction, and sends an alarm information to the shore-based monitoring center. When the anchor rope tension T(t) is detected to exceed 0.9 x T maxWhen the abnormal state is detected, the system automatically triggers a mooring safety warning and records the abnormal state in a log for post-analysis and equipment maintenance decisions. If valid remote sensing data updates cannot be received for 3 consecutive sampling periods, the system automatically switches to a single-source degraded mode, uses only buoy data, appropriately relaxes the reliability evaluation threshold, and labels the data source limited flag in the output prediction results, prompting downstream users to pay attention to the increased uncertainty of the prediction results.

[0084] In some computing resource-limited implementations, a segmented discrete function can be used instead of a continuous logical function (Sigmoid) for the reliability mapping mechanism. Specifically, the working condition response index EWI can be divided into several discrete intervals. For example, when EWI is less than 0.3, the real-time reliability coefficient is set to 1.0 (high reliability); when EWI is between 0.3 and 0.7, the real-time reliability coefficient is set to 0.5 (medium reliability); and when EWI is greater than 0.7, the real-time reliability coefficient is set to 0.1 (low reliability). Although the smoothness is sacrificed, the floating point operation load of the embedded chip on the buoy is reduced, which is suitable for low-power ocean monitoring nodes.

[0085] In other implementations, a linear average method based on reliability weighting can be used as an alternative for Bayesian fusion processing. Specifically, instead of explicitly constructing a dynamic observation variance, the real-time reliability coefficient is directly normalized as a weight factor. The specific calculation formula can be represented as:

[0086] H fused (t) =Σ(w k (t) * H k (t));

[0087] where the weight w k (t) = c k (t) / Σc i (t), c k (t) is the real-time reliability coefficient of the kth data source, and Σc i (t) is the sum of the reliability coefficients of all data sources. This embodiment can be a special case of Bayesian fusion when the prior distribution is unknown or the prior variance tends to infinity, suitable for newly developed sea areas where the historical wave statistics of the sea area cannot be accurately obtained, i.e., the prior mean μ0 cannot be determined, and can quickly realize the weighted fusion of multi-source data.

[0088] In addition, during the normal monitoring period when the sea conditions are relatively stable, i.e. EWI is long-term below 0.2, the global Pearson correlation coefficient can be used instead of the weighted correlation coefficient. In this mode, the system does not distinguish between extreme samples and ordinary samples, but uses all historical data to calculate the linear correlation between the features and the target variable. As a simplified working mode, i.e. energy saving mode, only when EWI mutation is detected, the extreme working condition weighted screening mode is switched to. Both ensure the efficiency of normal operation and ensure the prediction accuracy in extreme sea conditions.

[0089] Optionally, in addition to wave period and wave direction, flow rate can also be introduced as an auxiliary judgment index for screening conditions of physical constraint sample pairs. In strong current sea areas, the influence of flow rate on anchor line tension cannot be ignored. Therefore, when constructing sample pairs, a limit condition that the flow rate difference is less than 0.2 meters / second can be added. This ensures that the change of anchor line tension is mainly caused by waves, enhancing the effectiveness of physical monotonicity constraints. Such adaptive adjustments for specific environmental factors should be considered within the scope of protection of the present application.

[0090] In one possible embodiment, it further comprises: evaluating the reliability of the extreme wave prediction model, verifying the prediction performance of the model on the tail extreme samples, specifically: in order to verify the performance of the model after physical constraint training, the collected historical measured ocean buoy wave data needs to be scientifically divided. The system divides the cleaned data set into training set and test set according to the time sequence order, and the preferred division ratio is 7:3. It needs to be specially pointed out that, in order to prevent data leakage, the time sequence principle is strictly followed during division, i.e. using the past 70% data to predict the future 30% data, rather than randomly shuffling. The training set is used to perform PSO-RBF parameter optimization process, while the test set is not involved in training, and is only used for final evaluation of model performance. In the selection of evaluation indexes, in addition to the conventional root mean square error (RMSE) and determination coefficient, the extreme sample average absolute percentage error (E-MAPE) can also be introduced as a key index. The conventional RMSE is easily diluted by a large number of small wave samples, resulting in a high overall accuracy of the model, but a large deviation in large waves. E-MAPE only calculates the error for samples with a working condition response index greater than 0.5 in the test set. For example, for the prediction of significant wave height, RMSE is calculated to measure overall stability, and determination coefficient is calculated to measure trend fitting degree. At the same time, E-MAPE is focused on, if the index is less than 15%, the model is considered to have the value of actual engineering application in extreme sea conditions. Objectively prove that the physical constraint mechanism improves the generalization ability of the model in the extreme value interval, rather than just fitting the average state.

[0091] The embodiment successfully solves the three problems of untrustworthy, discontinuous and unphysical ocean monitoring data in extreme sea conditions by constructing a reliability evaluation system based on physical state perception, combining dynamic Bayesian fusion and physical constraint machine learning. Whether it is the core mathematical model or the above-mentioned various engineering variants, its essence is to deeply embed the physical laws of ocean dynamics into the data-driven prediction process, realizing the organic integration of physical mechanism and artificial intelligence.

[0092] In an embodiment of the present application, the multi-source data-driven ocean extreme wave measurement method can also be: collecting multi-source heterogeneous data of the target sea area, including ocean buoy monitoring data, remote sensing image data, hydrological observation data and buoy motion attitude data. The Z-Score method is used to identify and process outliers of the data set, and all data are normalized to eliminate the influence of dimension and improve data consistency. Based on the Bayesian data fusion algorithm, multi-source information is integrated, the motion attitude response parameters of the ocean buoy (including pitch, roll, heave and anchor rope tension) are selected as the model input features, and the wave element (significant wave height, mean period, spectral peak period and wave speed) is output as the target variable. The Pearson correlation coefficient is used to analyze the correlation between variables, and redundant and low correlation parameters are removed to optimize the input feature set. A machine learning model with radial basis function (RBF) as the core is established, and appropriate loss function, optimizer and hyperparameters are configured. The root mean square error (RMSE), determination coefficient (R 2 ) and mean absolute percentage error (MAPE) are used to compare and evaluate the prediction performance of various machine learning models, and the benchmark structure of the RBF model is preliminarily determined. The particle swarm optimization algorithm (PSO) is introduced to automatically optimize the hyperparameters of the RBF model, such as network layers, neuron number, activation function and loss function. In the training process, regularization techniques are used to suppress overfitting and improve model generalization ability. The mapping relationship between buoy motion attitude and anchor rope tension is learned using the training set, and a multi-source data-driven PSO-RBF extreme wave element prediction model is constructed. The measured ocean buoy wave data is divided into training set and test set in the ratio of 7:3, and RMSE and R 2 As an evaluation index, the prediction accuracy of the model for significant wave height, mean period, spectral peak period and wave speed is quantitatively analyzed to verify the reliability and accuracy of the PSO-RBF model. The trained PSO-RBF model is applied to actual extreme wave element prediction, and feedback optimization is performed by continuously accessing measured data, dynamically adjusting hyperparameters, and realizing continuous improvement and iterative evolution of model performance.

[0093] According to an aspect of the present application, a computer device is provided for performing the method of measuring ocean extreme wave based on multi-source data driving according to any one of the above embodiments. The computer device can be embodied as a server of a shore-based data center, an embedded industrial computer inside an ocean buoy, or an intelligent gateway with edge computing capability. From the hardware architecture, the device includes a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0094] Specifically, the memory can include a high-speed random access memory (RAM) and can also include a non-volatile memory (Non-volatile Memory), such as at least one magnetic disk storage device, a flash memory device, or other solid-state memory device. The memory stores an operating system (such as Linux, Windows, or an embedded RTOS) and program modules for executing the core algorithm of the present application. These program modules include but are not limited to: a working condition response index calculation module, a reliability mapping module, a Bayesian fusion module, a feature screening module, and a PSO-RBF prediction reasoning module.

[0095] The processor can be a central processing unit (CPU), or a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic device. When the processor executes the computer program, the following steps are implemented: acquiring multi-source heterogeneous monitoring data of a target sea area; calculating a working condition response index based on buoy motion posture data and determining a real-time reliability coefficient; constructing a dynamic observation variance using the real-time reliability coefficient and performing Bayesian fusion; and inputting the fusion result into a pre-trained extreme wave prediction model to obtain a prediction result.

[0096] In addition, the computer device also includes a communication interface for data interaction with external devices. For example, receiving attitude and tension data sent by a remote sea buoy through a satellite communication module (such as a Beidou short message terminal), receiving remote sensing data of a weather satellite through a fiber network, and sending the final wave prediction result to the maritime bureau or disaster prevention and mitigation command center through an Ethernet interface. Make the mathematical model into actual disaster prevention and mitigation productivity.

[0097] Overall, the multi-source data-driven ocean extreme wave measurement method comprises: acquiring heterogeneous monitoring data containing the motion posture of the buoy; calculating a working condition response index reflecting the smoothness of the buoy, and mapping the real-time reliability coefficient of each data source accordingly; using the real-time reliability coefficient to construct a dynamic observation variance, and performing adaptive fusion on the multi-source wave data through a Bayesian algorithm; and inputting the fusion result into a pre-trained extreme wave prediction model to obtain a prediction value. The prediction model is trained using a particle swarm algorithm, and its fitness function includes a physical constraint penalty term for forcing the model to comply with the monotonicity rule of anchor line tension and wave height.

[0098] The application adopts a physical state-aware dynamic fusion strategy. By calculating the working condition response index containing the standard deviations of pitch, roll, heave and anchor line tension, the degree of violent motion of the buoy is quantified and nonlinearly mapped into a real-time reliability coefficient. Using this coefficient to construct a dynamic observation variance, in the Bayesian fusion process, when extreme sea conditions cause the buoy to move violently, the variance of the affected data source is automatically inflated and its weight is reduced. The problem of sensor saturation or physical failure in the traditional method is solved, and the contaminated observation data is effectively eliminated. A physically constrained model training strategy is adopted. A physical constraint penalty term is introduced into the fitness function of the model training, which forces the model output to comply with the physical monotonicity rule that the anchor line tension is positively correlated with the wave height under similar sea conditions. By selecting specific sample pairs and punishing the prediction results that violate physical common sense, the model is forced to follow the physical law in the extreme value interval (tail samples) where the sample is sparse. The problem of poor generalization ability and easy output of non-physical results of existing models is solved.

[0099] The above describes the preferred embodiments of the application, but the application is not limited to the specific details of the above embodiments. Within the technical concept of the application, various equivalent transformations of the technical solutions of the application can be made, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. A method for measuring extreme ocean waves based on multi-source data, characterized in that, include: Acquire multi-source heterogeneous monitoring data of the target sea area, including at least real-time collected buoy motion attitude data and multi-source raw wave observation data; The working condition response index, which reflects the stability of the buoy, is calculated based on the buoy motion attitude data. A mapping relationship between the working condition response index and data reliability is established to obtain the real-time reliability coefficient for multi-source raw wave observation data. A dynamic observation variance is constructed using the real-time reliability coefficient. Based on this, Bayesian fusion processing is performed on the multi-source raw wave observation data to obtain the fused observation values ​​of wave elements. The wave element fusion observations are input into a pre-trained extreme wave prediction model, and the extreme wave element prediction results for the target sea area are output.

2. The method according to claim 1, characterized in that, Multi-source heterogeneous monitoring data also includes real-time collected anchor rope tension data; The calculation of the operating condition response index includes: Perform sliding time window statistical analysis on the buoy motion attitude data, calculate the attitude standard deviation within the window, and obtain the attitude statistical characteristics. Calculate the ratio of the anchor rope tension data to the preset limit tension to obtain the tension load ratio; The working condition response index is generated based on a weighted combination of attitude statistical characteristics and tension load ratio.

3. The method according to claim 2, characterized in that, The operating condition response index is calculated using the following formula: EWI(t)=α·std roll (t)+β·std pitch (t)+γ·std heave (t)+δ·(T(t) / T max ); Where EWI(t) represents the operating condition response exponent at time t, std roll (t), std pitch (t), std heave (t) represent the standard deviations of pitch, roll, and heave in the attitude statistics, respectively, and T(t) represents the anchor rope tension data at time t. max This represents the preset limit tension, where α, β, γ, and δ are pre-configured non-negative weighting coefficients.

4. The method according to claim 1, characterized in that, The real-time reliability coefficient is obtained, including: Construct a logic function, wherein the operating condition response index and the real-time reliability coefficient are negatively correlated; By using a logistic function to map the working condition response index to a normalized interval, the real-time reliability coefficient corresponding to each multi-source raw wave observation data is calculated. The larger the value of the operating condition response index, the closer the corresponding real-time reliability coefficient is to zero.

5. The method according to claim 1, characterized in that, The obtained wave element fusion observations include: Obtain the pre-configured baseline variance corresponding to each multi-source raw wave observation data; The dynamic observation variance is obtained by calculating the ratio of the baseline variance to the real-time reliability coefficient. Based on the dynamic observation variance and the preset prior distribution parameters, the Bayesian posterior mean is calculated and used as the fused observation value of wave elements.

6. The method according to claim 1, characterized in that, Before inputting the fused wave element observations into the pre-trained extreme wave prediction model, the following steps are also included: The statistical weights of the samples are calculated based on the operating condition response index, including weighting terms for extreme operating conditions; Calculate the weighted correlation coefficient between the preset candidate input features and the fused observations of wave elements. The weighted correlation coefficient uses sample statistical weights to weight the covariance. Candidate input features with a weighted correlation coefficient greater than a preset threshold are identified as the effective feature set for input extreme wave prediction models.

7. The method according to claim 1, characterized in that, The extreme wave prediction model is a radial basis function neural network, which is pre-trained using a particle swarm optimization algorithm. The particle swarm optimization algorithm uses a fitness function to evaluate the quality of particles. The fitness function includes a data error term and a physical constraint penalty term. The physical constraint penalty term is used to constrain the model prediction results to conform to a preset physical monotonicity law.

8. The method according to claim 7, characterized in that, The calculation process for the physical constraint penalty term includes: Select sample pairs with similar working conditions and different anchor rope tensions from the pre-stored training dataset; For each sample pair, determine whether the predicted wave height of the extreme wave prediction model for anchor rope tension samples above the preset threshold is less than the predicted wave height for anchor rope tension samples below the preset threshold. If so, the sample is determined to violate physical monotonicity, and the square of the predicted wave height difference is calculated as a penalty value and added to the physical constraint penalty term.

9. The method according to claim 7, characterized in that, The calculation process for the data error term includes: The loss weights of the pre-stored training samples are determined based on the operating condition response index, and the loss weights increase as the operating condition response index increases. The squared error between the model's predicted values ​​and the actual observed values ​​is calculated, and the squared errors are weighted and summed using loss weights to obtain the data error term.

10. The method according to claim 4, characterized in that, Multi-source heterogeneous monitoring data includes buoy observation data and remote sensing image data. The real-time reliability coefficient corresponding to each multi-source raw wave observation data is calculated, including: Based on buoy observation data, the operating condition response index is used as the first reliability evaluation index, and the first real-time reliability coefficient is calculated through a logic function. For remote sensing image data, the time lag between the data acquisition time and the current time is calculated and used as the second reliability evaluation index. The second real-time reliability coefficient is obtained by calculating through a logical function.

Images