Marine environment abnormal data detection method and system based on multi-sensor fusion
Through the combination of multi-sensor fusion and graph neural network, the refinement problem of marine environmental data detection in severe weather is solved, the accuracy and reliability of data detection are improved, and solid support is provided for forecasting and disaster reduction.
Patent Information
- Application Number
- CN202411133085.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-12-06
AI Technical Summary
The existing technology cannot obtain refined observation data in severe weather conditions such as typhoons and storm surges, and cannot be used for emergency observation, providing important data and auxiliary decision-making support for forecasting and disaster reduction.
Using multi-sensor fusion method, data is obtained through marine biological, chemical and physical signal acquisition modules, data fusion and anomaly detection are used using weighted average algorithm and graph neural network (GNN), and static graphs are constructed for node feature analysis and prediction, and abnormal points are identified. .
It improves the accuracy and system reliability of marine environmental abnormal data detection, is suitable for emergency observation in severe weather, provides important data and decision-making support for forecasting and disaster reduction, and ensures the safety of people's lives and property.
Smart Images

Figure CN119089347A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ocean data processing, and in particular to a method and system for detecting abnormal data in an ocean environment based on multi-sensor fusion. Background Art
[0002] In the work of disaster early warning, how to comprehensively and quickly obtain the real-time observation of sea level, wind speed, wind direction and air pressure and other hydrological and meteorological observation data during the disaster is undoubtedly a strong technical support for improving the accuracy of early warning. Conventional data acquisition systems can collect hydrological and meteorological observation data in real time.
[0003] Chinese patent application number CN202210641348.3 discloses an anomaly detection method for marine online monitoring data, comprising the following steps: S1, performing point anomaly labeling and sequence anomaly labeling on marine historical monitoring data to obtain a test data set; S2, based on the test data set, selecting the optimal parameters of threshold detection, mean square error detection, mutation point detection, weighted moving average fitting detection, Holt-Winter fitting detection, isolated forest detection and weighted anomaly score through testing; S3, performing data quality detection on the marine online monitoring data and preprocessing; S4, using the threshold detection, mean square error detection, mutation point detection, weighted moving average fitting detection and Holt-Winter fitting detection with the selected optimal parameters to perform point anomaly detection and labeling on the preprocessed marine online monitoring data; using the isolated forest detection with the selected optimal parameters to perform sequence anomaly detection and labeling on the preprocessed marine online monitoring data; this method can ensure the anomaly detection effect of marine online monitoring data.
[0004] Chinese patent application number CN202310023123.6 discloses anomaly detection methods, systems, equipment and media for ocean space observation data, which belongs to the field of ocean space data and data anomaly detection technology. Obtain different depth data observed by each observation point in the ocean space and form them into a data set; after excluding low-quality depth data in the data set, normalize the data set to determine whether the different depth data observed by a certain observation point is abnormal. If abnormal, scale the different depth data observed by the observation point, and then scale the data set, reduce the dimension of the data set, and form the data set into an image form; zero-padding the image to construct a CoordConv layer; input the image into a convolutional neural network model, and based on the constructed CoordConv layer, construct a Bicubic+CoordConv+Pool convolutional neural network model; train the constructed convolutional neural network model; input the data to be detected into the trained convolutional neural network model to detect outliers.
[0005] However, the above schemes are unable to obtain detailed observation data in severe weather conditions such as typhoons and storm surges, and cannot be used for emergency observation to provide important data and auxiliary decision-making support for forecasting and disaster reduction. Summary of the invention
[0006] In order to solve the above technical problems, the present invention provides a method and system for detecting abnormal data in an ocean environment based on multi-sensor fusion.
[0007] To achieve the above object, the present invention is implemented according to the following technical solutions:
[0008] One of the purposes of the present invention is to provide a method for detecting abnormal data in an ocean environment based on multi-sensor fusion, comprising the following steps:
[0009] S1. Obtain the observation elements, i.e., high-frequency signals emitted by marine organisms, through several marine biological signal acquisition modules; obtain the observation elements, i.e., water temperature, salinity, and turbidity of seawater, through several marine chemical signal acquisition modules; obtain the observation elements, i.e., flow velocity and direction of seawater, through several marine physical signal acquisition modules;
[0010] S2. Fusion of the same type of observation elements acquired by several marine biological signal acquisition modules / marine chemical signal acquisition modules / marine physical signal acquisition modules;
[0011] S3. Use the statistical features of the Node-Level nodes of the GNN graph neural network to detect abnormal nodes by merging multiple observation elements of the same type.
[0012] Furthermore, the step S2 specifically includes:
[0013] S21, synchronize timestamp: obtain local time through network time protocol NTP, synchronize to marine biological signal acquisition module, marine chemical signal acquisition module, marine physical signal acquisition module, and set as the system time of the current marine biological signal acquisition module, marine chemical signal acquisition module, and marine physical signal acquisition module, take the current system time as the acquisition time of each observation element, and synchronize the timestamp of all data for subsequent analysis;
[0014] S22. Data cleaning: Analyze the missing values including Nan values. If Nan values exist continuously in a large range, use the dropna method to delete the missing values including Nan.
[0015] S23, feature extraction analysis: extract the statistical features of each observation element, including mean, maximum, minimum and standard deviation, and use MinMaxScaler for normalization;
[0016] S24, data fusion: Use weighted average to process the data fusion of multiple observation elements of the same type, combine the information of n observation elements of the same type, and process the differences in the measurement of observation elements of the same type. The variances of each observation element of the same type are Total variance Calculate the weight value B of each observation element of the same type i , the fusion value is:
[0017]
[0018] Where: x i B is the measured value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module; i is the weight value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module, which is calculated by the total variance and measurement value of the same type of observation elements. The calculation formula is: k is the number of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module.
[0019] Furthermore, the step S3 specifically includes:
[0020] S31. Construct a static graph, where nodes represent subsequences, each edge contains the relationship between subsequences, and add time relationships to the graph; where Iput data: multidimensional channel of input time series, each observation element represents the corresponding value of each dimension, N observation elements correspond to N-dimensional fusion data, and the multidimensional vector is expressed as X t , Graph Node: Set the graph node to the current observation element X i , adjacent node X j ,The edges between nodes represent the continuity and information between the current ,observation element and other observation elements;
[0021] S32, GNN training: Use the neighbor node information related to the current node to represent the current node corresponding vector Use ɑ for weighting and transform it through the ReLU nonlinear activation layer. The result is used as the input of the dense layer to get the current time t sequence for prediction and get the predicted value And the mean square error (MSE) is used to evaluate the error between the current predicted value and the true observed value;
[0022] S33, GNN prediction: For new time series data, the anomaly score at each time point is calculated through the GNN model. For different factors, iPerform standardization, compare the normalized results of the anomaly scores of different factors, and take the result with the largest difference as the final anomaly score;
[0023] S34, anomaly identification: a threshold of anomaly score is set according to the training results, and points above this threshold are considered as anomalies; points with anomaly scores higher than the threshold are identified as anomalies;
[0024] S35. Abnormal value processing: mark abnormal values, record the acquisition time and value of abnormal values, and use Mark array to store and record abnormal values.
[0025] The second object of the present invention is to provide a marine environment abnormal data detection system based on multi-sensor fusion, comprising:
[0026] A power supply module, used to supply power to the system;
[0027] The ocean data acquisition module includes a marine biological signal acquisition module, a marine chemical signal acquisition module, and a marine physical signal acquisition module; a passive acoustic sensor is provided inside the marine biological signal acquisition module for monitoring the high-frequency signals emitted by the observed elements, namely, marine organisms; the marine chemical signal acquisition module is a temperature-salinity-depth turbidity meter, which is used to monitor the observed elements, namely, the conductivity, temperature, depth and turbidity of the water body in the ocean, and calculate the salinity S of the seawater according to the known temperature T and conductivity K through the formula S=1.3888*K-0.02478*K*T-6171.9; a mechanical wind sensor and an acoustic Doppler current profiler are provided inside the marine physical signal acquisition module for measuring the velocity and direction of the observed elements, namely, the seawater;
[0028] A remote communication module, used to transmit the real-time observation data collected by the ocean data collection module to the data processing module;
[0029] The data processing module integrates the same type of observation elements obtained by several marine biological signal acquisition modules / marine chemical signal acquisition modules / marine physical signal acquisition modules; specific implementation:
[0030] S21, synchronize timestamp: obtain local time through network time protocol NTP, synchronize to marine biological signal acquisition module, marine chemical signal acquisition module, marine physical signal acquisition module, and set as the system time of the current marine biological signal acquisition module, marine chemical signal acquisition module, and marine physical signal acquisition module, take the current system time as the acquisition time of each observation element, and synchronize the timestamp of all data for subsequent analysis;
[0031] S22. Data cleaning: Analyze the missing values including Nan values. If Nan values exist continuously in a large range, use the dropna method to delete the missing values including Nan.
[0032] S23, feature extraction analysis: extract the statistical features of each observation element, including mean, maximum, minimum and standard deviation, and use MinMaxScaler for normalization;
[0033] S24, data fusion: Use weighted average to process the data fusion of multiple observation elements of the same type, combine the information of n observation elements of the same type, and process the differences in the measurement of observation elements of the same type. The variances of each observation element of the same type are Total variance Calculate the weight value B of each observation element of the same type i , the fusion value is:
[0034]
[0035] Where: x i B is the measured value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module; i is the weight value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module, which is calculated by the total variance and measurement value of the same type of observation elements. The calculation formula is: k is the number of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module;
[0036] The data transmission storage module is used to generate a data file from the data processed by the data processing module, including a data packet header, real-time observation data after quality control, data packet length, a data check code generated according to the total length of the data and data characters, and a data packet end symbol, and then transmit the data to the analysis and decision support module in real time;
[0037] The analysis and decision support module uses the statistical characteristics of the GNN graph neural network Node-Level nodes to detect abnormal nodes by integrating multiple observation elements of the same type; specific implementation:
[0038] S31. Construct a static graph, where nodes represent subsequences, each edge contains the relationship between subsequences, and add time relationships to the graph; where Iput data: multidimensional channel of input time series, each observation element represents the corresponding value of each dimension, N observation elements correspond to N-dimensional fusion data, and the multidimensional vector is expressed as X t , Graph Node: Set the graph node to the current observation element X i , adjacent node X j ,The edges between nodes represent the continuity and information between the current ,observation element and other observation elements;
[0039] S32, GNN training: Use the neighbor node information related to the current node to represent the current node corresponding vector Use ɑ for weighting and transform it through the ReLU nonlinear activation layer. The result is used as the input of the dense layer to get the current time t sequence for prediction and get the predicted value And the mean square error (MSE) is used to evaluate the error between the current predicted value and the true observed value;
[0040] S33, GNN prediction: For new time series data, the anomaly score at each time point is calculated through the GNN model: For different factors, i Perform standardization, compare the normalized results of the anomaly scores of different factors, and take the result with the largest difference as the final anomaly score;
[0041] S34, anomaly identification: a threshold of anomaly score is set according to the training results, and points above this threshold are considered as anomalies; points with anomaly scores higher than the threshold are identified as anomalies;
[0042] S35. Abnormal value processing: mark abnormal values, record the acquisition time and value of abnormal values, and use Mark array to store and record abnormal values.
[0043] Furthermore, the power supply module is mains electricity or a solar battery.
[0044] Furthermore, the data transmission and storage module transmits data in real time via a satellite or wireless network.
[0045] Compared with the existing technology, the present invention integrates multiple sensors for refined observation, combines the information obtained from different sensors of marine biology, marine chemistry, and marine physics with information from multiple data sources using a weighted average multi-sensor data fusion algorithm, and then outputs the fused data as a time series data stream in a unified format; analyzes the collected data, integrates the GNN graph neural network to detect abnormal data on the time series, removes errors and abnormal data, thereby improving the accuracy of abnormal data detection and the reliability of the system; the present invention is designed for typhoons, storm surges and other severe disaster weather, and is suitable for emergency observation in severe weather, providing important data and decision-making support for forecasting and disaster reduction, and providing a solid and powerful guarantee for protecting people's property and lives. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a structural block diagram of the marine environment abnormal data detection system based on multi-sensor fusion of the present invention.
[0047] Figure 2 This is the workflow diagram for the analysis and decision support module. DETAILED DESCRIPTION
[0049] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the invention.
[0050] Example 1
[0051] like Figure 1 As shown, this embodiment exemplarily shows a marine environment abnormal data detection system based on multi-sensor fusion, including:
[0052] Power supply module 1, used to power the system, the power supply module can choose mains electricity or solar battery;
[0053] The ocean data acquisition module 2 includes a marine biological signal acquisition module, a marine chemical signal acquisition module, and a marine physical signal acquisition module; the marine biological signal acquisition module is internally provided with a passive acoustic sensor for monitoring the high-frequency signals emitted by the observed elements, namely, marine organisms; the marine chemical signal acquisition module is a temperature-salinity-depth turbidity meter for monitoring the observed elements, namely, the conductivity, temperature, depth and turbidity of the water body in the ocean, and calculating the salinity S of the seawater according to the known temperature T and conductivity K through the formula S=1.3888*K-0.02478*K*T-6171.9; the marine physical signal acquisition module is internally provided with a mechanical wind sensor and an acoustic Doppler current profiler for measuring the velocity and direction of the observed elements, namely, the seawater;
[0054] Remote communication module 3, used to transmit the real-time observation data collected by the ocean data collection module 2 to the data processing module;
[0055] Data processing module 4 integrates the same type of observation elements obtained by several marine biological signal acquisition modules / marine chemical signal acquisition modules / marine physical signal acquisition modules; specifically, it performs:
[0056] S21, synchronize timestamp: obtain local time through network time protocol NTP, synchronize to marine biological signal acquisition module, marine chemical signal acquisition module, marine physical signal acquisition module, and set as the system time of the current marine biological signal acquisition module, marine chemical signal acquisition module, and marine physical signal acquisition module, take the current system time as the acquisition time of each observation element, and synchronize the timestamp of all data for subsequent analysis;
[0057] S22. Data cleaning: Analyze the missing values including Nan values. If Nan values exist continuously in a large range, use the dropna method to delete the missing values including Nan.
[0058] S23, feature extraction analysis: extract the statistical features of each observation element, including mean, maximum, minimum and standard deviation, and use MinMaxScaler for normalization;
[0059] S24, data fusion: Use weighted average to process the data fusion of multiple observation elements of the same type, combine the information of n observation elements of the same type, and process the differences in the measurement of observation elements of the same type. The variances of each observation element of the same type are Total variance Calculate the weight value B of each observation element of the same type i , the fusion value is:
[0060]
[0061] Where: x i B is the measured value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module; i is the weight value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module, which is calculated by the total variance and measurement value of the same type of observation elements. The calculation formula is: k is the number of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module;
[0062] The data transmission storage module 5 is used to generate a data file from the data processed by the data processing module, including a data packet header, real-time observation data after quality control, data packet length, a data check code generated according to the total length of the data and data characters, and a data packet end symbol, and then transmit the data in real time to the analysis and decision support module via a satellite or wireless network;
[0063] Analysis and decision support module 6 uses the statistical features of the GNN graph neural network Node-Level nodes to detect abnormal nodes by integrating multiple observation elements of the same type; Figure 2 As shown, the specific implementation is:
[0064] S31. Construct a static graph, where nodes represent subsequences, each edge contains the relationship between subsequences, and add time relationships to the graph; where Iput data: multidimensional channel of input time series, each observation element represents the corresponding value of each dimension, N observation elements correspond to N-dimensional fusion data, and the multidimensional vector is expressed as X t , Graph Node: Set the graph node to the current observation element X i , adjacent node X j ,The edges between nodes represent the continuity and information between the current ,observation element and other observation elements;
[0065] S32, GNN training: Use the neighbor node information related to the current node to represent the current node corresponding vector Use ɑ for weighting and transform it through the ReLU nonlinear activation layer. The result is used as the input of the dense layer to get the current time t sequence for prediction and get the predicted value And the mean square error (MSE) is used to evaluate the error between the current predicted value and the true observed value;
[0066] S33, GNN prediction: For new time series data, the anomaly score at each time point is calculated through the GNN model: For different factors, i Perform standardization, compare the normalized results of the anomaly scores of different factors, and take the result with the largest difference as the final anomaly score;
[0067] S34, anomaly identification: a threshold of anomaly score is set according to the training results, and points above this threshold are considered as anomalies; points with anomaly scores higher than the threshold are identified as anomalies;
[0068] S35. Abnormal value processing: mark abnormal values, record the acquisition time and value of abnormal values, and use Mark array to store and record abnormal values.
[0069] Example 2
[0070] The system of Example 1 can be used to perform a method for detecting abnormal data in an ocean environment based on multi-sensor fusion, comprising the following steps:
[0071] S1. Acquire high-frequency signals emitted by marine organisms through several marine biological signal acquisition modules; acquire the water temperature, salinity, and turbidity of seawater through several marine chemical signal acquisition modules; and acquire the flow velocity and direction in seawater through several marine physical signal acquisition modules;
[0072] S2. Fusion of the same type of observation elements acquired by several marine biological signal acquisition modules / marine chemical signal acquisition modules / marine physical signal acquisition modules;
[0073] S21, synchronize timestamp: obtain local time through network time protocol NTP, synchronize to marine biological signal acquisition module, marine chemical signal acquisition module, marine physical signal acquisition module, and set as the system time of the current marine biological signal acquisition module, marine chemical signal acquisition module, and marine physical signal acquisition module, take the current system time as the acquisition time of each observation element, and synchronize the timestamp of all data for subsequent analysis;
[0074] S22. Data cleaning: Analyze the missing values including Nan values. If Nan values exist continuously in a large range, use the dropna method to delete the missing values including Nan.
[0075] S23, feature extraction analysis: extract the statistical features of each observation element, including mean, maximum, minimum and standard deviation, and use MinMaxScaler for normalization;
[0076] S24, data fusion: Use weighted average to process the data fusion of multiple observation elements of the same type, combine the information of n observation elements of the same type, and process the differences in the measurement of observation elements of the same type. The variances of each observation element of the same type are Total variance Calculate the weight value B of each observation element of the same type i , the fusion value is:
[0077]
[0078] Where: x i B is the measured value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module; i is the weight value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module, which is calculated by the total variance and measurement value of the same type of observation elements. The calculation formula is: k is the number of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module;
[0079] S3, using the statistical features of the GNN graph neural network Node-Level nodes to detect abnormal nodes by combining multiple fused observation elements of the same type;
[0080] S31. Construct a static graph, where nodes represent subsequences, each edge contains the relationship between subsequences, and add time relationships to the graph; where Iput data: multidimensional channel of input time series, each observation element represents the corresponding value of each dimension, N observation elements correspond to N-dimensional fusion data, and the multidimensional vector is expressed as X t , Graph Node: Set the graph node to the current observation element X i , adjacent node X j ,The edges between nodes represent the continuity and information between the current ,observation element and other observation elements;
[0081] S32, GNN training: Use the neighbor node information related to the current node to represent the current node corresponding vector Use ɑ for weighting and transform it through the ReLU nonlinear activation layer. The result is used as the input of the dense layer to get the current time t sequence for prediction and get the predicted value And the mean square error (MSE) is used to evaluate the error between the current predicted value and the true observed value;
[0082] S33, GNN prediction: For new time series data, the anomaly score at each time point is calculated through the GNN model: For different factors, i Perform standardization, compare the normalized results of the anomaly scores of different factors, and take the result with the largest difference as the final anomaly score;
[0083] S34, anomaly identification: a threshold of anomaly score is set according to the training results, and points above this threshold are considered as anomalies; points with anomaly scores higher than the threshold are identified as anomalies;
[0084] S35. Abnormal value processing: mark abnormal values, record the acquisition time and value of abnormal values, and use Mark array to store and record abnormal values.
[0085] In summary, the present invention integrates multiple sensors for refined observation, combines the information obtained from different sensors of marine biology, marine chemistry, and marine physics with information from multiple data sources using a weighted average multi-sensor data fusion algorithm, and then outputs the fused data as a time series data stream in a unified format; analyzes the collected data, integrates the GNN graph neural network to perform abnormal data detection on the time series, removes errors and abnormal data, thereby improving the accuracy of abnormal data detection and the reliability of the system; the present invention is designed for typhoons, storm surges and other severe disaster weather, and is suitable for emergency observations in severe weather, providing important data and decision-making support for forecasting and disaster reduction, and providing a solid and powerful guarantee for protecting people's property and lives.
[0086] The technical solution of the present invention is not limited to the above-mentioned specific embodiments. All technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.
Claims
1. A method for detecting abnormal data in marine environment based on multi-sensor fusion, characterized in that: The following steps are involved: S1. Obtain the observation elements, i.e., high-frequency signals emitted by marine organisms, through several marine biological signal acquisition modules; obtain the observation elements, i.e., water temperature, salinity, and turbidity of seawater, through several marine chemical signal acquisition modules; obtain the observation elements, i.e., flow velocity and direction of seawater, through several marine physical signal acquisition modules; S2. Fusion of the same type of observation elements acquired by several marine biological signal acquisition modules / marine chemical signal acquisition modules / marine physical signal acquisition modules; S3. Use the statistical features of the Node-Level nodes of the GNN graph neural network to detect abnormal nodes by merging multiple observation elements of the same type.
2. The method for detecting abnormal data in the marine environment based on multi-sensor fusion according to claim 1 is characterized in that: The step S2 specifically includes: S21, synchronize timestamp: obtain local time through network time protocol NTP, synchronize to marine biological signal acquisition module, marine chemical signal acquisition module, marine physical signal acquisition module, and set as the system time of the current marine biological signal acquisition module, marine chemical signal acquisition module, and marine physical signal acquisition module, take the current system time as the acquisition time of each observation element, and synchronize the timestamp of all data for subsequent analysis; S22. Data cleaning: Analyze the missing values including Nan values. If Nan values exist continuously in a large range, use the dropna method to delete the missing values including Nan. S23, feature extraction and analysis: extract the statistical features of each observation element, analyze the mean, maximum, minimum and standard deviation, and use MinMaxScaler for normalization; S24, data fusion: Use weighted average to process the data fusion of multiple observation elements of the same type, combine the information of n observation elements of the same type, and process the differences in the measurement of observation elements of the same type. The variances of each observation element of the same type are Total variance Calculate the weight value B of each observation element of the same type i , the fusion value is: Where: x i B is the measured value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module; i is the weight value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module, which is calculated by the total variance and measurement value of the same type of observation elements. The calculation formula is: k is the number of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module.
3. The method for detecting abnormal data in the marine environment based on multi-sensor fusion according to claim 1 is characterized in that: The step S3 specifically includes: S31. Construct a static graph, where nodes represent subsequences, each edge contains the relationship between subsequences, and add time relationships to the graph; where Iput data: multidimensional channel of input time series, each observation element of the same type represents the corresponding value of each dimension, N observation elements of the same type correspond to N-dimensional fusion data, and the multidimensional vector is expressed as X t , Graph Node: Set the graph node to the current observation element X i , adjacent node X j ,The edges between nodes represent the continuity and information between the current ,observation element and other observation elements; S32, GNN training: Use the neighbor node information related to the current node to represent the current node corresponding vector Use ɑ for weighting and transform it through the ReLU nonlinear activation layer. The result is used as the input of the dense layer to get the current time t sequence for prediction and get the predicted value And the mean square error (MSE) is used to evaluate the error between the current predicted value and the true observed value; S33, GNN prediction: For new time series data, the anomaly score at each time point is calculated through the GNN model: For different factors, i Perform standardization, compare the normalized results of the anomaly scores of different factors, and take the result with the largest difference as the final anomaly score; S34, anomaly identification: a threshold of anomaly score is set according to the training results, and points above this threshold are considered as anomalies; points with anomaly scores higher than the threshold are identified as anomalies; S35. Abnormal value processing: mark abnormal values, record the acquisition time and value of abnormal values, and use Mark array to store and record abnormal values.
4. A marine environment abnormal data detection system based on multi-sensor fusion, characterized in that: include: A power supply module, used to supply power to the system; The ocean data acquisition module includes a marine biological signal acquisition module, a marine chemical signal acquisition module, and a marine physical signal acquisition module; a passive acoustic sensor is provided inside the marine biological signal acquisition module for monitoring the high-frequency signals emitted by the observed elements, namely, marine organisms; the marine chemical signal acquisition module is a temperature-salinity-depth turbidity meter, which is used to monitor the observed elements, namely, the conductivity, temperature, depth and turbidity of the water body in the ocean, and calculate the salinity S of the seawater according to the known temperature T and conductivity K through the formula S=1.3888*K-0.02478*K*T-6171.9; a mechanical wind sensor and an acoustic Doppler current profiler are provided inside the marine physical signal acquisition module for measuring the velocity and direction of the observed elements, namely, the seawater; A remote communication module, used to transmit the real-time observation data collected by the ocean data collection module to the data processing module; The data processing module is used to fuse the same type of observation elements obtained by several marine biological signal acquisition modules / marine chemical signal acquisition modules / marine physical signal acquisition modules; the specific implementation is: S21, synchronize timestamp: obtain local time through network time protocol NTP, synchronize to marine biological signal acquisition module, marine chemical signal acquisition module, marine physical signal acquisition module, and set as the system time of the current marine biological signal acquisition module, marine chemical signal acquisition module, and marine physical signal acquisition module, take the current system time as the acquisition time of each observation element, and synchronize the timestamp of all data for subsequent analysis; S22. Data cleaning: Analyze the missing values including Nan values. If Nan values exist continuously in a large range, use the dropna method to delete the missing values including Nan. S23, feature extraction analysis: extract the statistical features of each observation element, including mean, maximum, minimum and standard deviation, and use MinMaxScaler for normalization; S24, data fusion: Use weighted average to process the data fusion of multiple observation elements of the same type, combine the information of n observation elements of the same type, and process the differences in the measurement of observation elements of the same type. The variances of each observation element of the same type are Total variance Calculate the weight value B of each observation element of the same type i , the fusion value is: Where: x i B is the measured value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module; i is the weight value of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module, which is calculated by the total variance and measurement value of the same type of observation elements. The calculation formula is: k is the number of each observation element in the marine biological signal acquisition module / marine chemical signal acquisition module / marine physical signal acquisition module; The data transmission storage module is used to generate a data file from the data processed by the data processing module, including a data packet header, real-time observation data after quality control, data packet length, a data check code generated according to the total length of the data and data characters, and a data packet end symbol, and then transmit the data to the analysis and decision support module in real time; The analysis and decision support module uses the statistical characteristics of the GNN graph neural network Node-Level nodes to detect abnormal nodes by integrating multiple observation elements of the same type; specific implementation: S31, construct a static graph, where nodes represent subsequences, each edge contains the relationship between subsequences, and add the time relationship to the graph; Where Iput data: multidimensional channel of input time series, each observation element represents the corresponding value of each dimension, N observation elements correspond to N-dimensional fusion data, and the multidimensional vector is expressed as X t , Graph Node: Set the graph node to the current observation element X i , adjacent node X j ,The edges between nodes represent the continuity and information between the current ,observation element and other observation elements; S32, GNN training: Use the neighbor node information related to the current node to represent the current node corresponding vector Use ɑ for weighting and transform it through the ReLU nonlinear activation layer. The result is used as the input of the dense layer to get the current time t sequence for prediction and get the predicted value And the mean square error (MSE) is used to evaluate the error between the current predicted value and the true observed value; S33, GNN prediction: For new time series data, the anomaly score at each time point is calculated through the GNN model: For different factors, i Perform standardization, compare the normalized results of the anomaly scores of different factors, and take the result with the largest difference as the final anomaly score; S34, anomaly identification: a threshold of anomaly score is set according to the training results, and points above this threshold are considered as anomalies; points with anomaly scores higher than the threshold are identified as anomalies; S35. Abnormal value processing: mark abnormal values, record the acquisition time and value of abnormal values, and use Mark array to store and record abnormal values.
5. The marine environment abnormal data detection system based on multi-sensor fusion according to claim 4 is characterized in that: The power supply module is commercial electricity or a solar battery.
6. The marine environment abnormal data detection system based on multi-sensor fusion according to claim 4 is characterized in that: The data transmission and storage module transmits data in real time via a satellite or wireless network.
