Ship navigation risk early warning method based on unbalanced marine weather data enhancement

Abstract

The application discloses a ship navigation risk early warning method based on unbalanced offshore meteorological data enhancement, which firstly determines the ship navigation risk grade, and divides the unbalanced data set of target sea area meteorological disaster-causing elements corresponding thereto. Secondly, an adversarial learning model composed of a deep neural network generator and an evidence reasoning discriminator is designed to balance the unbalanced data set under different risk grades, and a Gaussian distribution model is established to describe the feature distribution of meteorological disaster-causing elements under different risk grades. Then, the risk grade reliability distribution of meteorological disaster-causing elements is calculated through the Gaussian distribution model. Finally, the weighted average method is adopted to fuse the risk grade reliability distribution, and the risk mode with the highest reliability after fusion is selected as the navigation risk grade of the current target sea area. The application can generate a small number of meteorological data conforming to the real distribution, effectively solve the data imbalance problem, and improve the extreme weather risk warning capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship navigation safety and weather warning, and particularly relates to a ship navigation risk warning method based on unbalanced offshore weather data enhancement, which is especially suitable for dynamic risk assessment and avoidance decision of navigation path under extreme weather conditions such as typhoon. BACKGROUND

[0002] Current ship navigation risk warning mainly relies on meteorological forecast data and fixed threshold determination method, by analyzing meteorological elements such as wind speed, wave height, visibility, and combining with navigation safety standards to divide the navigation risk level. However, this method has significant limitations: first, the frequency of extreme weather (such as typhoon, strong convection) is much lower than that of normal weather, resulting in a small amount of meteorological data samples of high risk level, making it difficult for traditional statistical analysis and machine learning models to effectively model. In addition, the existing method mostly uses static risk threshold (such as wind speed ≥ 20 m / s is high risk), which fails to fully consider the synergistic effect of wind, wave, flow and other multi-weather elements, as well as the adaptability of different ship types to extreme weather, resulting in lack of pertinence in risk assessment, which may cause misjudgment or failure of avoidance decision.

[0003] In order to improve the ability of navigation risk warning, in recent years, researchers have tried to introduce machine learning methods (such as support vector machine SVM, random forest RF) for risk prediction. However, due to the serious shortage of extreme weather data compared with normal weather data, the model is prone to majority class (normal weather) in the training process, thus weakening the recognition accuracy of extreme weather. Some studies use oversampling methods (such as SMOTE) to alleviate the problem of data imbalance, but offshore weather data has complex nonlinear and spatiotemporal dynamic characteristics, and the traditional generated samples may lead to distortion of data distribution, reducing the generalization ability of the model. SUMMARY

[0004] To address the aforementioned problems, this invention proposes a ship navigation risk warning method based on imbalanced marine meteorological data enhancement. The method first determines the ship navigation risk level and, for each risk level, divides the target sea area into imbalanced datasets of meteorological disaster-causing elements. It then designs an adversarial learning model composed of a deep neural network (DNN) generator and an evidence reasoning (ER) discriminator to enhance the balance of the meteorological disaster-causing element datasets under different risk levels. A risk level pattern of meteorological disaster-causing elements is established based on a Gaussian model to describe the feature distribution of meteorological disaster-causing elements under different risk levels. Real-time acquisition of meteorological disaster-causing element data for the target sea area is performed, and the risk level reliability distribution of the meteorological disaster-causing elements is calculated based on the aforementioned risk level pattern. The risk level reliability distribution of the meteorological disaster-causing elements in the target sea area is fused to obtain the comprehensive risk level reliability distribution of the target sea area. Finally, based on the comprehensive risk level reliability distribution, a ship navigation risk warning result for the corresponding target sea area is generated.

[0005] The proposed method for ship navigation risk early warning based on unbalanced marine meteorological data enhancement includes the following steps:

[0006] Step 1: Determine the risk level of ship navigation, and for each risk level, divide the corresponding imbalanced dataset of meteorological disaster-causing factors in the target sea area.

[0007] Step 2: Design an adversarial learning model consisting of a deep neural network (DNN) generator and an evidence reasoning (ER) discriminator to enhance the balance of the imbalanced datasets under different risk levels and obtain a fused dataset.

[0008] Step 3: Establish a Gaussian distribution model for the fused dataset based on the Gaussian model to describe the characteristic distribution of meteorological disaster-causing factors under different risk levels.

[0009] Step 4: Acquire real-time meteorological disaster-causing factor data for the target sea area, and calculate the risk level confidence distribution of meteorological disaster-causing factors using a Gaussian distribution model.

[0010] Step 5: Integrate the risk level reliability distribution of meteorological disaster-causing factors in the target sea area to obtain the comprehensive risk level reliability distribution of the target sea area.

[0011] Step 6: Based on the comprehensive risk level confidence distribution, generate ship navigation risk warning results for the corresponding target sea area.

[0012] The application provides a ship navigation risk early warning method based on unbalanced marine weather data enhancement.

[0013] The application has the following beneficial effects:

[0014] I. An adversarial learning model composed of a deep neural network (DNN) generator and an evidence reasoning (ER) discriminator can generate minority class (high risk level) weather data that conforms to the real distribution, effectively solve the data imbalance problem, and improve the extreme weather risk early warning capability.

[0015] II. By fusing the risk level confidence distribution of the meteorological disaster elements, the comprehensive risk level of the target sea area is calculated, the misjudgment caused by a single meteorological index is avoided, and the early warning result is more comprehensive and reliable. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a ship navigation risk early warning process block diagram based on unbalanced marine weather data enhancement;

[0017] Figure 2 is a DNN generator and ER discriminator interactive adversarial process diagram. DETAILED DESCRIPTION

[0018] The application provides a ship navigation risk early warning method based on unbalanced marine weather data enhancement, and its flow chart is as shown in Figure 1 The method comprises the following steps:

[0019] (1) Determine the ship navigation risk level, and divide the unbalanced data set of the meteorological disaster elements of the target sea area corresponding to each risk level.

[0020] (2) Design an adversarial learning model composed of a deep neural network (DNN) generator and an evidence reasoning (ER) discriminator, balance the meteorological disaster element data set under the different risk levels, and obtain a fusion data set.

[0021] (3) Based on the Gaussian model, a Gaussian distribution model of the fusion dataset is established to describe the characteristic distribution of meteorological disaster-causing factors under different risk levels.

[0022] (4) Acquire meteorological disaster-causing factor data of the target sea area in real time, and calculate the risk level confidence distribution of meteorological disaster-causing factors through Gaussian distribution model.

[0023] (5) By integrating the risk level reliability distribution of meteorological disaster-causing factors in the target sea area, the comprehensive risk level reliability distribution of the target sea area is obtained.

[0024] (6) Based on the comprehensive risk level confidence distribution, generate ship navigation risk warning results for the corresponding target sea area.

[0025] The specific description of step (1) is as follows:

[0026] (1-1) Based on the minimum reasonable feasibility principle in risk analysis theory and technology, the risk level of ship navigation is divided into N risk modes D={d1,...,d n ,...,d N}, and each risk model has J distinct attributes J = {c1,...,c j ,...,c J Then, the dataset of marine meteorological disaster-causing factors can be represented as U={x1,...,x}. i ,...,x I}, where a single sample vector can be represented as x i =[c1(i),...,c j (i),...,c J (i),d n ],d n ∈D, c j (i) is the j-th attribute c j The value of . For risk model d n The sample vector set below, The risk model is d n The number of sample vectors.

[0027] (1-2) Using the last risk pattern d N sample vector set Using the k-means algorithm as a benchmark, two sets of sample vectors are obtained. and Based on the reference values ​​of each attribute, we obtain K cluster centers corresponding to the j-th attribute. Sort these K cluster centers in ascending order to obtain the set of reference values ​​for the j-th attribute.

[0028] (1-3) Using reference value set A j Calculate the matching degree of the input sample. and likelihood

[0029]

[0030] in c represents all sample values ​​of a certain attribute j (i) Matching reference value Furthermore, this sample was identified as risk pattern d. n The sum of matching degrees, c j (i) identified as The likelihood function,

[0031] The final reference evidence matrix (evidence reasoning discriminator) REM j It can be represented as:

[0032]

[0033] in This means the sample attribute value is c. j Time corresponding category d n The reliability.

[0034] To make it easier to understand, step (1) is illustrated here, and the process is as follows:

[0035] The navigation risk level of the target sea area can be divided into three risk modes: D = {d1, d2, d3}, corresponding to extremely high risk (posing a serious threat to ship navigation, operations should be completely avoided in this state), relatively high risk (posing a relatively high threat to ship navigation, navigation and operations in this state should be avoided as much as possible for extended periods), and some risk (ships can navigate and operate in this state). Each risk mode has two distinct attributes corresponding to the wind speed c1 and wave height c2 of the target sea area. The sample vector set formed by the three risk modes can be represented as follows:

[0036]

[0037] The quantitative relationships among the various sample vector sets are as follows:

[0038] This shows a significant imbalance in the distribution of risk patterns d1 and d3, requiring preprocessing to address the data imbalance issue. The k-means algorithm is used to obtain the sample vector sets respectively. Reference values ​​for wind speed c1 and wave height c2 (A1, A2):

[0039] A1 = [-1.6822, -0.8266, -0.1143, 0.5997, 1.4616, 4.0866]

[0040] A2 = [-1.6129, -0.7126, 0.0330, 0.7996, 1.7591, 5.1194]

[0041] Calculate the matching degree of the input sample and further construct the reference evidence matrix (evidence reasoning discriminator) REM j As follows:

[0042]

[0043] Similarly, the data imbalance problem needs to be preprocessed between risk patterns d2 and d3, and the corresponding reference value set and evidence matrix can also be obtained, which is similar in form to the reference value set (A1, A2) and reference evidence matrix (REM1, REM2) constructed by balancing risk patterns d2 and d3.

[0044] The specific description of the step (2) is as follows:

[0045] (2-1) Based on the evidence reasoning discriminator REM j of step (1-3), the single sample vector x i can be further obtained according to formula (4) n,j The activated evidence:

[0046]

[0047] Wherein is the matching degree of the input sample, p i represents the confidence of x n being diagnosed as risk pattern d j .

[0048] After the ER rule fuses the activated evidence e = [e(1),..., e(j),... e(J)] of the J attributes of x i , the fused confidence distribution O(x n ) = {(d n,e(J) , p n ) | n = 1,..., N} can be obtained by formula (5) and formula (6):

[0049]

[0050] Where e(j) = (d n,j , p j ), w jFor ER fusion weights, iteratively update Get g, F, and E are the power sets of set D.

[0051] (2-2) Based on the fusion confidence distribution value p n,e(J) Minimize the discriminator loss and update the ER fusion weights w. j r j And the evidence reasoning discriminator REM j The loss function is:

[0052]

[0053] in This represents the ideal reliability distribution corresponding to the input sample.

[0054] (2-3) The generator employs a deep neural network (DNN) structure, which includes an input layer, hidden layers, and an output layer, for generating simulated data.

[0055] The input layer receives a random noise vector. Used to receive low-dimensional random noise vector z, dimension 1, number of

[0056] The hidden layer processes the received noise vector in a fully connected manner, and consists of a fully connected layer and a nonlinear activation function, used for feature extraction and nonlinear mapping of the input:

[0057]

[0058] Where W1 and W2 are the weight matrices of the hidden layer, b1 and b2 are the bias vectors of the hidden layer, and h1 and h2 are the outputs of the first and second hidden layers, respectively.

[0059] The output layer is used to map high-dimensional features to the output dimension of the target data space, where the output dimension is a set of simulated data with the same amount of random noise.

[0060] This indicates that it can be used to enhance d n The number of simulation datasets under risk model is

[0061] (2-4) Using the updated evidence reasoning discriminant REM j The simulation data can be obtained through equations (4)-(6). sample vector Fusion reliability distribution P n,e(J) To minimize the generator loss, the loss function is:

[0062]

[0063] wherein represents the ideal belief distribution of the risk pattern d n under the simulated data.

[0064] Repeat step (2-3) using the optimized generator to continuously generate augmented data

[0065] (2-5) Merge the augmented data generated in step (2-4) belonging to the risk level pattern d n into the original data set to obtain a balanced fusion data set, as shown in . Figure 2

[0066] For ease of understanding, step (2) is illustrated as follows:

[0067] Based on the evidence reasoning discriminators REM1, REM2 obtained in step (1), the belief distribution of the fusion of individual sample vectors in O(x i ) = {(d n , p n,e(2) ) | n = 1, 2, 3} can be obtained:

[0068]

[0069] Based on the fused belief distribution, the loss of the evidence reasoning discriminators is minimized, and the ER fusion weights w j , r j and REM1, REM2 are updated; in the loss function, B represents the ideal belief distribution of the training sample:

[0070]

[0071] The generator receives a random noise vector with a dimension of 1, and maps it through a hidden layer and an output layer to obtain simulated data with a dimension of 2

[0072]

[0073] Using the updated evidence matrix evidence reasoning discriminators: REM1, REM2, the sample vector of the simulated data , the fusion belief distribution P n,e(2) of the sample vector ​Represented as:

[0074]

[0075] The optimized generator is used to continuously generate enhanced data using equations (8) to (10).

[0076]

[0077] Enhanced data under the generated risk model d1 Compared with the original dataset By merging, a balanced merged dataset is obtained; similarly, augmented data can be obtained. Used to balance the original dataset After balancing, the number of sample vectors is the same across all three risk models, which is...

[0078] The specific description of step (3) is as follows:

[0079] (3-1) Based on the balanced fusion dataset from step 2-5, calculate the mean values ​​of meteorological disaster-causing factors respectively. and standard deviation

[0080]

[0081] Where c j (i) represents the original data. A single sample vector, To enhance data A single sample vector.

[0082] (3-2) Based on the mean and standard deviation of the meteorological disaster-causing factors, construct Gaussian distribution models for the meteorological disaster-causing factors respectively.

[0083] To make it easier to understand, step (3) is illustrated here as follows:

[0084] The sample sets for each risk level model after balancing are as follows: Risk Level I Risk Level II Risk Level III

[0085] Based on the mean and standard deviation of wind speed and wave height in the above balanced risk level pattern dataset, Gaussian distribution models for wind speed and wave height are constructed as follows:

[0086]

[0087] The Gaussian distribution model of the wind speed and wave height of the risk mode d1 is taken as the probability distribution model of the wind speed and wave height under the corresponding risk level mode:

[0088]

[0089] The distribution models of the remaining risk modes are similar.

[0090] The specific description of the step (4) is as follows:

[0091] (4-1) Real-time acquisition of the meteorological disaster-causing element data c(t) of the target sea area.

[0092] (4-2) The meteorological disaster-causing element data are substituted into the Gaussian distribution models under each risk level mode in the step (3-2) respectively to obtain the corresponding reliabilities under different risk level modes, which are respectively denoted as

[0093] (4-3) The reliabilities of the meteorological disaster-causing element under different risk levels are normalized to obtain the grade reliability distribution of the meteorological disaster-causing element of the target sea area

[0094] For the convenience of understanding, the process is illustrated as follows:

[0095] For the wind speed and wave height data c(t) = [15.5466, 3.9158] acquired online, the wind speed and wave height are respectively substituted into the probability distribution models of the step (3-2) to obtain

[0096]

[0097] The reliabilities of the wind speed and wave height under different risk levels are normalized to obtain the normalized grade reliability distribution

[0098]

[0099] The specific description of the step (5) is as follows:

[0100] According to the grade reliability distribution of the target sea area disaster-causing element, the weighted average method is adopted to fuse the J attributes, and the comprehensive grade reliability distribution of the target sea area is calculated, and the fusion formula is as follows:

[0101]

[0102] The fused grade reliability distribution is denoted as ε j The fusion weight of each disaster-causing element is denoted as

[0103] For ease of understanding, step (5) is illustrated here, the process is as follows:

[0104] The normalized level confidence distribution obtained in step (4) is fused Weighted average fusion, calculate the comprehensive level confidence distribution of the target sea area (here take ε1=0.4, ε2=0.6):

[0105]

[0106] The specific description of step (6) is as follows:

[0107] According to the fusion level confidence distribution obtained in step (5), the risk mode with the highest confidence is selected as the navigation risk level of the current target sea area according to the principle of maximum confidence, and the calculation formula is as follows:

[0108]

[0109] Among them, Indicates the navigation risk level of the current target sea area.

[0110] For ease of understanding, step (6) is illustrated here, for the online acquisition of target sea area wind speed and wave height data, According to the principle of maximum confidence, the comprehensive navigation risk level of the current sea area is d2, that is, the risk is large, and the ship navigation is threatened, and it should be avoided as much as possible to navigate and operate in this state for a long time.

[0111] The following will introduce the embodiment of the method of the application in detail in combination with the drawings:

[0112] As Figure 1 The flow chart of the application is shown, and the core part of the application is: first, determine the ship navigation risk level, and divide the unbalanced data set of the target sea area meteorological disaster-causing elements corresponding to each risk level; design an adversarial learning model composed of a deep neural network (DNN) generator and an evidence reasoning (ER) discriminator, balance the meteorological disaster-causing element data sets under different risk levels, and obtain a fusion data set; based on the Gaussian model, establish a risk level mode of the fusion data set for describing the feature distribution of the meteorological disaster-causing elements under different risk levels; real-time acquisition of the meteorological disaster-causing element data of the target sea area, and based on the above risk level mode, calculate the risk level confidence distribution of the meteorological disaster-causing elements; fuse the risk level confidence distribution of the target sea area meteorological disaster-causing elements to obtain the comprehensive risk level confidence distribution of the target sea area; according to the comprehensive risk level confidence distribution, generate the ship navigation risk warning result of the corresponding target sea area.

[0113] Taking the Zhujiajian to Dongji Island route in Zhoushan as an example, the wind speed and wave height data collected during the typhoon (Kanadi) and low-risk periods were used to obtain the wind speed and wave height data of the target sea area online, According to the principle of maximum credibility, the current sea area comprehensive navigation risk level is d2, that is, the risk is relatively large, and the ship navigation is threatened, and the ship should be avoided as much as possible to navigate and operate in this state for a long time.

[0114] The wind speed and wave height data set in the example is divided into training set and test set according to the ratio of 8:2, wherein 80% of the training data is enhanced by the ER-DNN model, and a Gaussian distribution model is constructed based on the enhanced data. Then, the remaining 20% of the original data is used as an independent test set, and the key steps (4) to (6) in the application are repeated to verify the performance of the model. The experimental results show that, as shown in Table 1, the early warning performance of the data enhanced by the ER-DNN in the high-risk mode (d1, d2) is significantly improved, and the F1 value (F1-score) is obviously improved. This method effectively enhances the recognition ability of the model to the high-risk navigation state, makes the early warning result more accurate and reliable, and provides a better technical guarantee for the ship navigation safety.

[0115] Table 1 Test sample data classification results

[0116]

Claims

1. A method for ship navigation risk early warning based on unbalanced marine meteorological data enhancement, characterized in that, Includes the following steps: Step 1: Determine the risk level of ship navigation, and for each risk level, divide the corresponding imbalanced dataset of meteorological disaster-causing factors in the target sea area. The specific implementation is as follows: Step 1-1: The ship's navigation risk level is classified as follows: Risk Model And each risk model has J different attributes. The dataset of marine meteorological disaster-causing factors is as follows: A single sample vector is , It is the first Attributes The value; Risk model The sample vector set below, For risk model The number of sample vectors; Steps 1-2, using the last risk model sample vector set Using the k-means algorithm as a benchmark, two sets of sample vectors are obtained. Based on the reference values ​​of each attribute, we obtain K cluster centers corresponding to the j-th attribute. We then sort these K cluster centers in ascending order to obtain the set of reference values ​​for the j-th attribute. ; Steps 1-3: Utilizing the reference value set Calculate the matching degree of the input sample. and likelihood : ； ； Represents all sample values ​​of a certain attribute Matching reference value Furthermore, this sample was identified as a risk pattern. The sum of matching degrees, express identified as The likelihood function, Evidence Reasoning Discriminator REM j Represented as: ； in This represents the sample attribute value. Time corresponding category Reliability; Step 2: Design an adversarial learning model consisting of a deep neural network generator and an evidence reasoning discriminator to enhance the balance of imbalanced datasets under different risk levels and obtain a fused dataset. The specific implementation is as follows: Step 2-1: Evidence-based reasoning discriminator To obtain a single sample vector Evidence of activation: ； in express Diagnosed as a risk pattern Reliability; After ER rule fusion of Activation evidence of each attribute Then, the fused confidence distribution was obtained. : ； ； in , , , For ER fusion weights, iteratively update Get , where g, F, and E are the power sets of set D; Step 2-2: Based on the fused confidence distribution value Minimize the discriminator loss and update the ER fusion weights. , And the evidence reasoning discriminator REM j ; Steps 2-3: The generator uses a deep neural network structure to generate simulated data. The input layer of a deep neural network receives a random noise vector. Dimension 1, quantity 1 ; The hidden layers of the deep neural network process the received noise vector in a fully connected manner. They consist of fully connected layers and nonlinear activation functions, used for feature extraction and nonlinear mapping of the input to obtain the output. The output layer of the deep neural network outputs a set of simulated data with the same dimension as the amount of random noise. ； This indicates that it can be used for enhancement. The number of simulation datasets under risk model is , ; Steps 2-4: Use the updated evidence reasoning discriminator REM j To obtain simulation data sample vector Fusion reliability distribution Minimize the generator loss; Repeat steps 2-3 using the optimized generator to continuously generate enhanced data. , ; Steps 2-5: The generated [items / properties] belong to... Enhanced data in risk level mode Compared with the original dataset Merge the datasets to obtain a balanced, merged dataset. Step 3: Establish a Gaussian distribution model for the fused dataset based on the Gaussian model to describe the characteristic distribution of meteorological disaster-causing factors under different risk levels; Step 4: Acquire real-time meteorological disaster-causing factor data for the target sea area, and calculate the risk level confidence distribution of meteorological disaster-causing factors using a Gaussian distribution model; Step 5: Use the weighted average method to fuse the risk level reliability distribution to obtain the comprehensive risk level reliability distribution of the target sea area. Select the risk pattern with the highest reliability as the navigation risk level of the current target sea area and generate a ship navigation risk warning result for the target sea area.

2. The ship navigation risk early warning method based on unbalanced marine meteorological data enhancement according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3-1: Based on the balanced fused dataset, calculate the mean values ​​of meteorological disaster-causing factors respectively. and standard deviation ; Step 3-2: Based on the mean and standard deviation of meteorological disaster-causing factors, construct Gaussian distribution models for each meteorological disaster-causing factor. .

3. The ship navigation risk early warning method based on unbalanced marine meteorological data enhancement according to claim 2, characterized in that, The specific process of step 4 is as follows: Step 4-1: Acquire real-time meteorological disaster-causing factor data for the target sea area. ; Step 4-2: Substitute the meteorological disaster-causing factor data into the Gaussian distribution model under each risk level mode in Step 3-2 to obtain the corresponding confidence levels under different risk level modes, denoted as follows: ; Step 4-3: The reliability of the meteorological disaster-causing factors under different risk levels is normalized to obtain the level reliability distribution of the meteorological disaster-causing factors in the target sea area. .

Images