A marine environment monitoring and early warning method

Abstract

This invention discloses a marine environmental monitoring and early warning method, belonging to the field of marine environmental monitoring technology. The method includes collecting measured time-series data of marine environmental parameters at a target location within a first time period; determining the cluster center point with the smallest distance to the measured characteristic parameters of the target location, denoted as the target cluster center point; when the distance between the measured characteristic parameters of the target location and its neighboring cluster centers is less than a first threshold, the target location is considered an anomaly; at neighboring locations adjacent to the anomaly location, determining the cluster center point with the smallest distance to the measured characteristic parameters within the first time period, denoted as the neighboring cluster center point; calculating a consensus degree based on the cluster consistency between the anomaly location and its neighboring locations; when the consensus degree is greater than a consensus threshold, setting the early warning type of the anomaly location to the early warning type corresponding to the target cluster center point. This method effectively suppresses noise in the sensor-collected data and increases the accuracy of early warning identification.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring technology, and more specifically, to a method for marine environmental monitoring and early warning. Background Technology

[0002] With the development of the marine economy, the demand for real-time monitoring of water quality and ecological status in areas such as ports, waterways, nearshore aquaculture areas, and ecologically sensitive areas is constantly increasing.

[0003] A search revealed that Chinese patent CN119848650A discloses a method for real-time monitoring and early warning of the marine environment. This method involves collecting historical marine environment data, constructing a historical marine environment data feature matrix, using a clustering analysis algorithm to classify the marine environment data in the historical marine environment data feature matrix to obtain a marine environment type center data matrix, and setting a set of abnormal marine environment data types. Sensors are deployed in the marine environment to collect real-time marine environment data, which is then preprocessed to obtain preprocessed real-time marine environment data. Features are extracted from the preprocessed real-time marine environment data to construct a real-time marine environment data feature set, and the types of real-time marine environment data are calculated. An early warning is then issued based on the types of real-time marine environment data.

[0004] The aforementioned methods rely solely on current regional marine environmental data for early warning, exhibiting poor noise suppression capabilities for sensor-collected data and requiring improved accuracy. Therefore, we propose a new method for marine environmental monitoring and early warning. Summary of the Invention

[0005] 1. Technical problems to be solved The purpose of this invention is to provide a marine environmental monitoring and early warning method to solve the problems mentioned in the background art.

[0006] 2. Technical Solution This invention is achieved through the following technical solution: A method for marine environmental monitoring and early warning includes the following steps: S1. Collect historical time-series data of marine environmental parameters under different early warning types; S2. The first data processing method is used to process historical time series data under different warning types to obtain historical feature parameters; S3. Using the DBSCAN clustering algorithm, cluster analysis is performed on the feature parameters under different warning types to obtain no less than the number of cluster centers for each warning type, and a one-to-one correspondence is established between the cluster centers and the warning types. S4. Collect measured time-series data of marine environmental parameters at the target location within the first time period; S5. The measured time-series data of marine environmental parameters at the target location are processed using the first data processing method to obtain the measured characteristic parameters of the target location. S6. Determine the cluster center point that has the smallest distance to the measured feature parameters of the target point and denot it as the target cluster center point. When the distance between the measured feature parameters of the target point and the neighboring cluster center points is less than the first threshold, the target point is regarded as an abnormal point. S7. At neighboring points adjacent to the outlier, determine the cluster center with the smallest distance to the measured feature parameters within the first time period, and denot it as the neighboring cluster center. Calculate the consensus degree based on the clustering consistency between the outlier and its neighboring points; the formula for calculating the consensus degree is as follows: ; ; Where n is the number of neighboring points. Let i be the consensus level of the i-th neighboring point. Let be the distance between the measured feature parameters of the i-th neighboring point and the target cluster center. and These are the minimum and maximum distances between the nearest point and all cluster centers, respectively. S8. When the consensus level is greater than the consensus threshold, the warning type of the abnormal point is set to the warning type corresponding to the target cluster center point. The consensus threshold is calculated using the following formula: ; In the formula, and These represent the minimum and maximum distances between the measured feature parameters of neighboring points and the target cluster center point, respectively, within the first time period; α is a weighting coefficient with a positive value. The distance between the measured feature parameters of the target point and the target cluster center point.

[0007] The first data processing method includes the following steps: A1. Filter the time series data of various marine environmental parameters; A2. Calculate the rate of change and variance of time series data for each marine environmental parameter; A3. Construct a feature matrix using the rate of change and variance of time-series data of various marine environmental parameters, and input the feature matrix into a neural network model to obtain feature vectors; A4. Calculate the correlation coefficient between the time series data of any two marine environmental parameters and form a correlation matrix; A5. Combine the feature vectors with the correlation matrix to form feature parameters.

[0008] As an optional solution to the technical solution of this application, step S8 further includes generating a warning level, which is calculated using the following formula: ; In the formula, f(n) is the measured characteristic parameter. This represents the similarity between time series data i and time series data j within a time window n. α1 and α2 are weighting coefficients representing the similarity between the time series data of parameter i and the time series data of parameter j within a time window n-1.

[0009] As an optional solution to the technical solution of this application, in step S1, the marine environmental parameters include temperature, salinity, pH value, dissolved oxygen, turbidity, wind speed and air pressure parameters.

[0010] As an optional solution to the technical solution of this application, in step S3, the early warning types include: red tide early warning, pollution leak early warning, and low oxygen early warning; the method for corresponding cluster centers with early warning types includes: B1. Label the marine environmental status corresponding to the historical characteristic parameters of each time period with the warning type, which includes red tide warning, pollution leakage warning and low oxygen warning; B2. By setting the neighborhood radius ε and the minimum number of samples MinPts, at least three clusters are formed; B3. Count the frequency of each warning type in the historical samples within each cluster; B4. The warning type that appears most frequently is determined as the warning type corresponding to this cluster; B5. Establish a one-to-one correspondence between cluster centers and early warning types.

[0011] As an optional solution to the technical solution of this application, in step A3, the dimension of the feature vector is smaller than the dimension of the feature matrix, and the neural network model is trained in the following manner: C1. A decoding layer is cascaded after the neural network model to form a complete autoencoder training model. The output dimension of the decoding layer is the same as the dimension of the feature matrix. C2. Construct a feature matrix based on historical marine environmental parameters and use it as input for training the autoencoder model; C3. Use the variance between the output of the autoencoder training model and the constructed feature matrix as the loss value of the training model. Based on the loss value, use the backpropagation algorithm to iteratively optimize the network parameters of the training model. C4. When the number of training iterations reaches the preset threshold, or the loss value converges to a stable state, stop training, fix the network parameters, remove the decoding layer, and you will get the trained neural network model.

[0012] As an optional solution to the technical solution of this application, in step A4, the correlation coefficient is calculated using the following formula: ; In the formula, W is the length of the time series data, and X... i (K), X j (K) represents the data of parameters i and j at position k. , Let be the mean of the time series data for parameters i and j.

[0013] As an optional solution to the technical solution of this application, in step S6, the first threshold is the sum of the mean and standard deviation of the distances from each historical feature parameter to its cluster center point under the same warning type.

[0014] As an optional solution to the technical solution of this application, in step S7, the distance between the neighboring points and the target point does not exceed 1km, and the number of neighboring points is 3-5.

[0015] 3. Beneficial effects Compared with the prior art, the beneficial effects of the present invention are: 1) This invention introduces a clustering consensus analysis mechanism based on neighboring points. When a target point is identified as an anomaly, the consensus degree is further calculated based on the clustering consistency between the anomaly point and its neighboring points, and this is used to confirm the warning type of the target point. Since marine environmental parameters have a certain correlation in spatial distribution, by performing consensus analysis on multiple neighboring points, abnormal data caused by single-point sensor noise or transient interference can be effectively identified, thereby reducing the probability of misjudgment and improving the accuracy of warning identification.

[0016] 2) This invention reduces the dimensionality of the feature matrix composed of marine environmental parameters by constructing a neural network model, and uses an autoencoder to learn the latent feature structure of the data, thereby extracting more representative feature vectors. Compared with methods that directly use the original features for analysis, this approach can reduce data dimensionality, reduce redundant information, and effectively suppress the impact of noise on feature representation, thus improving the stability and reliability of data analysis.

[0017] 3) By combining the above-mentioned spatial consistency verification mechanism with the neural network feature extraction method, the present invention can achieve more accurate and stable anomaly identification and early warning judgment in complex marine environmental monitoring scenarios, thereby improving the overall performance of the marine environmental monitoring and early warning system. Attached Figure Description

[0018] Figure 1 This is a logic block diagram of a marine environmental monitoring and early warning method; Figure 2 This is a graph showing the change of mean squared error loss (MSELoss) during the autoencoder training process in this invention with the number of training iterations. Detailed Implementation

[0019] The technical solution of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0020] Please see Figure 1 This embodiment provides a marine environmental monitoring and early warning method, which uses a target point and its neighboring points as monitoring objects. The distance between the neighboring points and the target point does not exceed 1 km, and the number of neighboring points is 3. The early warning types include red tide warning, pollution leak warning, and low oxygen warning. The method includes the following steps: S1. Collect historical time-series data of marine environmental parameters under different early warning types: ; T(n) is the water temperature at time window n, S(n) is the salinity at time window n, DO(n) is the dissolved oxygen at time window n, Tu(n) is the turbidity at time window n, W(n) is the wind speed at time window n, and P(n) is the air pressure at time window n. In this scheme, the time window length is 1 hour, and 6 consecutive time windows are selected. The collected historical time series data are as follows: Table 1: Example of historical time-series data for marine environmental parameters:

[0021] S2. The first data processing method is used to process historical time-series data under different warning types to obtain historical feature parameters; the first data processing method includes the following steps: A1. Filter the time series data of various marine environmental parameters; ; Where L is the length of the filtering window.

[0022] A2. Calculate the rate of change and variance of time series data for each marine environmental parameter; ; ; In the formula, W is the length of the time series data.

[0023] A3. Construct a feature matrix X(n) = [ΔT(n),σT²(n),ΔS(n),σS²(n),ΔDO(n),σDO²(n),ΔTu(n),σTu²(n),ΔW(n),σW²(n),ΔP(n),σP²(n)] using the rate of change and variance of time series data of various marine environmental parameters, and input the feature matrix into a neural network model to obtain the feature vector f(n); The neural network model takes a feature matrix X(n) as input and outputs a feature vector f(n), where the dimension of the feature vector f(n) is smaller than the dimension of the feature matrix X(n). The training steps of the neural network model include: C1. A decoding layer is cascaded after the neural network model to form a complete autoencoder training model. The output dimension of the decoding layer is the same as the dimension of the feature matrix X(n). C2. Construct a feature matrix X(n) based on historical marine environmental parameters and use it as the input for training the autoencoder model; C3. The mean squared error between the reconstructed feature matrix output by the autoencoder training model and the input feature matrix is ​​used as the loss value L of the training model. Based on the loss value L, the network parameters of the training model are iteratively optimized using the backpropagation algorithm. The loss value L adopts the mean squared error loss function, the expression of which is: ; Where N is the number of training samples, and X(n) is the historical feature matrix within the nth time window. This is the output of the autoencoder training model.

[0024] This embodiment records the mean squared error loss (MSELoss) during the autoencoder training process and plots its curve as a function of the number of training iterations, as shown below. Figure 2 As shown. Figure 2 In the diagram, the horizontal axis represents the number of training iterations, and the vertical axis represents the mean squared error loss value. As training progresses, the loss value gradually decreases and eventually stabilizes, indicating that the autoencoder training model has converged and can effectively learn the latent feature structure of the marine environmental parameter feature matrix, thus providing a stable feature vector for subsequent feature extraction.

[0025] C4. When the number of training iterations reaches the preset threshold, or the loss value converges to a stable state, stop training, fix the network parameters, remove the decoding layer, and you will get the trained neural network model.

[0026] Preferably, the neural network model includes three fully connected layers and an output layer. The dimension of the input layer is the same as the dimension of the feature matrix X(n), which is 12 dimensions. The three fully connected layers have 10, 8, and 6 neurons respectively, used to extract latent features from the feature matrix layer by layer and achieve feature compression. The output layer contains 5 neurons, used to output the dimensionality-reduced feature vector f(n). The activation function of each fully connected layer is the ReLU function, and the decoding layer adopts a network structure symmetrical to the encoding layer.

[0027] A4. Calculate the correlation coefficient between the time series data of any two marine environmental parameters to form a correlation matrix: ; This correlation matrix reflects the interrelationships of various marine environmental parameters within a time window; the correlation coefficients are calculated using the following formula: ; In the formula, W is the length of the time series data, and X... i (K), X j (K) represents the data of parameters i and j at position k. , Let be the mean of the time series data for parameters i and j.

[0028] A5. Combine the eigenvector f(n) with the correlation matrix R to form the eigenparameter z(n); Specifically, the upper triangular elements of the correlation matrix R are expanded and concatenated with the feature vector f(n) to form a 20-dimensional feature parameter z(n).

[0029] S3. Using the DBSCAN clustering algorithm, cluster analysis is performed on the feature parameters under different warning types to obtain no less than the number of cluster centers for each warning type, and a one-to-one correspondence is established between the cluster centers and the warning types.

[0030] The methods for mapping cluster centers to early warning types include: B1. Label the marine environmental status corresponding to the historical characteristic parameters of each time period with the warning type, which includes red tide warning, pollution leakage warning and low oxygen warning; B2. By setting the neighborhood radius ε and the minimum number of samples MinPts, at least three clusters are formed; wherein, the neighborhood radius ε is determined according to the distance distribution between historical feature parameters, preferably 1.2 to 1.8 times the average distance of each historical feature parameter to its nearest neighbor feature parameter, and in this embodiment, ε=0.5; the minimum number of samples MinPts is determined according to the dimension of the feature parameters, preferably 2 to 4 times the number of feature dimensions, and in this embodiment, the feature parameter dimension is 20, so MinPts is 40-60, preferably 60.

[0031] By setting the neighborhood radius ε and the minimum number of samples MinPts, the DBSCAN algorithm is used to perform density clustering on historical feature parameters, resulting in several naturally formed clusters and a small number of noise points. The number of clusters is naturally determined by the distribution density of historical feature parameters in the feature space, without presetting the number of clusters.

[0032] B3. Count the frequency of each warning type in the historical samples within each cluster; B4. The warning type that appears most frequently is determined as the warning type corresponding to this cluster; B5. Establish a one-to-one correspondence between cluster centers and early warning types.

[0033] In this embodiment, examples of cluster center points corresponding to different warning types are as follows: The representative feature vector of the cluster corresponding to the red tide warning is:

[0034] ; The representative feature vector of the cluster corresponding to the pollution leak early warning is:

[0035] ; The representative feature vector of the cluster corresponding to the low oxygen warning is:

[0036] ; The Euclidean distance between different cluster centers is greater than a preset intra-class distance threshold, thereby ensuring the distinguishability of different warning types in the feature space; in this embodiment, the intra-class distance threshold is preferably 0.3.

[0037] S4. Collect measured time-series data of marine environmental parameters at the target location within the first time period. Some of the measured time-series data are as follows: Table 2: Example of measured time-series data for target locations:

[0038] S5. The measured time-series data of marine environmental parameters at the target location are processed using the first data processing method to obtain the measured characteristic parameters of the target location: z(n)=[0.62,0.48,0.71,0.55,0.83,0.79,0.65,0.72,0.60,0.81,0.67,0.74,0.63,0.69,0.71,0.58,0.77,0.66,0.73,0.68]; S6. The calculated distances of the measured characteristic parameters of the target location to the three cluster centers (red tide warning, pollution leak warning, and low oxygen warning) are d1=0.83, d2=0.41, and d3=0.96, respectively. The distance between the cluster center with the smallest distance to the measured characteristic parameters is less than the first threshold of 0.5, and should be regarded as an abnormal location. The first threshold is determined based on the distribution of distances between historical characteristic parameters and their corresponding cluster centers, and is the sum of the mean and standard deviation of the distances from each historical characteristic parameter to its cluster center under the same warning type.

[0039] S7. For the three neighboring points adjacent to the outlier, calculate the distance between each neighboring point and the three cluster centers, using the following formula: ; Where Dij represents the distance between the measured feature parameter of the i-th neighboring point and the j-th cluster center point, m is the dimension of the feature parameter, zi(k) is the k-th dimension component of the measured feature parameter of the i-th neighboring point, and Cj(k) is the k-th dimension component of the j-th cluster center point.

[0040] In this embodiment, the calculation results are shown in the following example: The distances between the first nearest neighbor and the three cluster centroids are as follows: D11=0.58, D12=0.36, D13=0.82; The distances between the second nearest neighbor and the three cluster centroids are as follows: D21=0.62, D22=0.40, D23=0.79; The distances between the third nearest neighbor and the three cluster centroids are as follows: D31=0.55, D32=0.38, D33=0.85.

[0041] Therefore, the minimum distances of the three neighboring points are D12, D22 and D32, respectively, and the corresponding cluster center points are all the second cluster center point C2. That is, the neighboring cluster center points of the three neighboring points are all C2.

[0042] The cluster center with the smallest distance to its neighboring points in the first time period is identified and denoted as the neighboring cluster center. The consensus degree is calculated based on the clustering consistency between outlier points and their neighboring points, using the following formula: ; in, Let be the distance between the measured feature parameters of the i-th neighboring point and the target cluster center. and These are the minimum and maximum distances between the nearest neighbor and all cluster centers, respectively; in this embodiment, the consensus degrees corresponding to the three nearest neighbors are as follows: , , The consensus level is taken as the average of the consensus levels of three neighboring points, that is: ; The consensus scores were: c1=0.68; c2=0.72; c3=0.75, with the average of 0.72.

[0043] S8. When the consensus level is greater than the consensus threshold, the warning type of the abnormal point is set to the warning type corresponding to the target cluster center point. The consensus threshold is calculated using the following formula: ; In the formula, and These are the minimum and maximum distances between the measured feature parameters of neighboring points and the target cluster center point, respectively, within the first time period; α is a positive weighting coefficient, preferably α=1.0; The distance between the measured feature parameters of the target point and the target cluster center point.

[0044] The consensus threshold is calculated to be 0.65. Since 0.72 > 0.65, the warning type for the target point can be confirmed as a pollution leak warning.

[0045] The warning level is calculated using the following formula: ; In the formula, f(n) is the measured characteristic parameter, ρ ij (n) represents the similarity between time series data i and time series data j within the time window n, ρ ij (n-1) represents the similarity between the time series data of parameter i and the time series data of parameter j within the time window n-1, and α1 and α2 are weight coefficients, α1=0.7 and α2=0.3.

[0046] Based on the above parameters, the warning level index of the target location in the first time period is calculated. In this embodiment, the similarity term Σρ ij (n)=0.81, Σρ ij (n-1)=0.55, substituting into the formula, we get: S(n)=α1×0.81+α2×0.55=0.73.

[0047] Table 3: Correspondence between Warning Levels and Warning Level Indicator S(n)

[0048] According to Table 3, the warning level index of the target location in the first time period corresponds to Level III warning.

Claims

1. A method for marine environmental monitoring and early warning, characterized in that: Includes the following steps: S1. Collect historical time-series data of marine environmental parameters under different early warning types; S2. The first data processing method is used to process historical time series data under different warning types to obtain historical feature parameters; S3. Using the DBSCAN clustering algorithm, cluster analysis is performed on the feature parameters under different warning types to obtain no less than the number of cluster centers for each warning type, and a one-to-one correspondence is established between the cluster centers and the warning types. S4. Collect measured time-series data of marine environmental parameters at the target location within the first time period; S5. The measured time-series data of marine environmental parameters at the target location are processed using the first data processing method to obtain the measured characteristic parameters of the target location. S6. Determine the cluster center point that has the smallest distance to the measured feature parameters of the target point and denot it as the target cluster center point. When the distance between the measured feature parameters of the target point and the neighboring cluster center points is less than the first threshold, the target point is regarded as an abnormal point. S7. At neighboring points adjacent to the outlier, determine the cluster center with the smallest distance to the measured feature parameters within the first time period, and denot it as the neighboring cluster center. Calculate the consensus degree based on the clustering consistency between the outlier and its neighboring points; the formula for calculating the consensus degree is as follows: ; ; Where n is the number of neighboring points. Let be the consensus level of the i-th neighboring point. Let be the distance between the measured feature parameters of the i-th neighboring point and the target cluster center. and These are the minimum and maximum distances between the nearest point and all cluster centers, respectively. S8. When the consensus level is greater than the consensus threshold, the warning type of the abnormal point is set to the warning type corresponding to the target cluster center point. The consensus threshold is calculated using the following formula: ； In the formula, and These represent the minimum and maximum distances between the measured feature parameters of neighboring points and the target cluster center point, respectively, within the first time period; α is a weighting coefficient with a positive value. The distance between the measured feature parameters of the target point and the target cluster center point.

2. The marine environmental monitoring and early warning method according to claim 1, characterized in that: The first data processing method includes the following steps: A1. Filter the time series data of various marine environmental parameters; A2. Calculate the rate of change and variance of time series data for each marine environmental parameter; A3. Construct a feature matrix using the rate of change and variance of time-series data of various marine environmental parameters, and input the feature matrix into a neural network model to obtain feature vectors; A4. Calculate the correlation coefficient between the time series data of any two marine environmental parameters and form a correlation matrix; A5. Combine the feature vectors with the correlation matrix to form feature parameters.

3. The marine environmental monitoring and early warning method according to claim 1, characterized in that: Step S8 further includes generating a warning level, which is calculated using the following formula: ; In the formula, f(n) is the measured characteristic parameter. This represents the similarity between time series data i and time series data j within a time window n. α1 and α2 are weighting coefficients representing the similarity between the time series data of parameter i and the time series data of parameter j within a time window n-1.

4. The marine environmental monitoring and early warning method according to claim 1, characterized in that: In step S1, the marine environmental parameters include temperature, salinity, pH value, dissolved oxygen, turbidity, wind speed, and air pressure.

5. A marine environmental monitoring and early warning method according to claim 1, characterized in that: In step S3, the warning types include: red tide warning, pollution leak warning, and low oxygen warning; the method for mapping cluster centers to warning types includes: B1. Label the marine environmental status corresponding to the historical characteristic parameters of each time period with the warning type, which includes red tide warning, pollution leakage warning and low oxygen warning; B2. By setting the neighborhood radius ε and the minimum number of samples MinPts, at least three clusters are formed; B3. Count the frequency of each warning type in the historical samples within each cluster; B4. The warning type that appears most frequently is determined as the warning type corresponding to this cluster; B5. Establish a one-to-one correspondence between cluster centers and early warning types.

6. A marine environmental monitoring and early warning method according to claim 2, characterized in that: In step A3, the dimension of the feature vector is smaller than the dimension of the feature matrix, and the neural network model is trained in the following manner: C1. A decoding layer is cascaded after the neural network model to form a complete autoencoder training model. The output dimension of the decoding layer is the same as the dimension of the feature matrix. C2. Construct a feature matrix based on historical marine environmental parameters and use it as input for training the autoencoder model; C3. Use the variance between the output of the autoencoder training model and the constructed feature matrix as the loss value of the training model. Based on the loss value, use the backpropagation algorithm to iteratively optimize the network parameters of the training model. C4. When the number of training iterations reaches the preset threshold, or the loss value converges to a stable state, stop training, fix the network parameters, remove the decoding layer, and you will get the trained neural network model.

7. A marine environmental monitoring and early warning method according to claim 2, characterized in that: In step A4, the correlation coefficient is calculated using the following formula: ； In the formula, W is the length of the time series data, and X... i (K), X j (K) represents the data of parameters i and j at position k. , Let be the mean of the time series data for parameters i and j.

8. A marine environmental monitoring and early warning method according to claim 1, characterized in that: In step S6, the first threshold is the sum of the mean and standard deviation of the distances from each historical feature parameter to its cluster center point under the same warning type.

9. A marine environmental monitoring and early warning method according to claim 1, characterized in that: In step S7, the distance between the neighboring points and the target point does not exceed 1km, and the number of neighboring points is 3-5.

Images