Marine ecology-oriented time-space diagram neural network anomaly detection method and system

Abstract

The invention relates to the technical field of marine environment monitoring, in particular to a time-space diagram neural network anomaly detection method and system for marine ecology. The method comprises the following steps: acquiring ocean monitoring data; performing data preprocessing on the acquired marine monitoring data; constructing a space-time diagram structure based on the preprocessed marine monitoring data; performing spatio-temporal feature extraction on the constructed spatio-temporal diagram structure by using an anomaly detection model; carrying out feature fusion on the extracted spatial-temporal features; and carrying out local anomaly and global anomaly scoring on the marine ecosystem based on a multi-scale anomaly scoring mechanism. According to the method, a unified data processing framework is designed, multi-source heterogeneous data such as satellite remote sensing data, water quality monitoring data and biological survey data can be processed at the same time, information contained in various kinds of data is fully utilized, and the comprehensiveness and accuracy of anomaly detection are improved. Through the construction of a space-time diagram structure, the incidence relation between different data sources is effectively captured, and more comprehensive feature expression is provided for anomaly detection.

Description

Technical Field

[0001] The present invention relates to the field of marine environment monitoring technology, and in particular to a method and system for detecting anomalies of a spatiotemporal graph neural network for marine ecology. Background Art

[0002] With the influence of biological activities and climate change, marine ecosystems are facing many threats, such as frequent abnormal events such as seawater acidification, eutrophication, red tide outbreaks, and marine garbage pollution. At present, marine ecosystem anomaly detection mainly adopts traditional detection methods based on statistical thresholds, machine learning methods, and deep learning methods. Among them, the traditional detection method based on statistical thresholds determines whether the system is abnormal by setting thresholds of key parameters, but this method is difficult to handle the complex correlation between parameters, and the fixed threshold is difficult to adapt to the dynamic change characteristics of the marine environment. Machine learning methods have advantages in processing multidimensional features, but existing methods often separate time and space features for analysis and cannot fully utilize the spatiotemporal correlation characteristics of data. Although deep learning methods perform well in processing time series data or spatial data, it is difficult to effectively model time dependence and spatial correlation at the same time. In addition, existing detection methods still have many shortcomings in data fusion, model construction, detection mechanism, and system implementation, such as difficulty in effectively integrating multi-source heterogeneous data, lack of explicit modeling of system network structure, and difficulty in achieving early warning of anomalies. Therefore, it is urgent to propose a new marine ecosystem anomaly detection method to solve the above problems. Graph neural networks have attracted widespread attention in recent years due to their powerful structured data processing capabilities. Combining graph neural networks with time series analysis can effectively process the complex spatiotemporal dependencies in marine ecosystems and provide new solutions for marine ecosystem anomaly detection.

[0003] However, there are still the following technical problems in the traditional marine ecosystem anomaly detection methods: (1) How to achieve effective integration of multi-source heterogeneous data in marine ecosystems. The current detection system lacks a unified data processing framework when processing multi-source heterogeneous data such as satellite remote sensing data, water quality monitoring data, and biological survey data. It is difficult to fully utilize the information contained in various types of data, which affects the comprehensiveness and accuracy of anomaly detection.

[0004] (2) How to achieve effective integration of multi-source heterogeneous data in marine ecosystems. The current detection system lacks a unified data processing framework when processing multi-source heterogeneous data such as satellite remote sensing data, water quality monitoring data, and biological survey data. It is difficult to fully utilize the information contained in various types of data, which affects the comprehensiveness and accuracy of anomaly detection.

[0005] (3) How to build a reasonable anomaly detection mechanism to achieve early identification and early warning of abnormal states of marine ecosystems. Existing detection methods mostly focus on discovering anomalies that have already occurred, lack the ability to predict potential anomalies, and cannot provide timely early warning information for the management and protection of marine ecosystems. Summary of the invention

[0006] In order to solve the above-mentioned problems, the present invention provides a spatiotemporal graph neural network anomaly detection method and system for marine ecology.

[0007] In the first aspect, the present invention provides a spatiotemporal graph neural network anomaly detection method for marine ecology, which adopts the following technical solutions: A spatiotemporal graph neural network anomaly detection method for marine ecology, comprising: Obtain ocean monitoring data; Carry out data preprocessing on the acquired ocean monitoring data; Constructing spatiotemporal graph structures based on preprocessed ocean monitoring data; Use anomaly detection models to extract spatiotemporal features of constructed spatiotemporal graph structures; Perform feature fusion on the extracted spatiotemporal features; Scoring of local and global anomalies in marine ecosystems based on a multi-scale anomaly scoring mechanism; Output anomaly detection results.

[0008] Furthermore, the acquired ocean monitoring data is preprocessed, including outlier processing and missing value filling for the ocean monitoring data, wherein the outlier processing includes an improved method based on a time series sliding window. The criterion is to dynamically calculate local statistical features to adapt to the time-varying characteristics of the data. Within the time window w, the identification criterion of outliers is expressed as: in, is the processed data set, x is the observed value, is the mean of the data series, is the standard deviation of the data series, n is the amount of data, and Use sliding window to calculate: in, is the local mean at time t, is the local standard deviation at time t, x i is the observation value at the i-th moment in the window, and t is the current time step.

[0009] Furthermore, the spatiotemporal graph structure is constructed based on the preprocessed ocean monitoring data, including constructing node features based on triple feature vectors, integrating environmental, biological and temporal information; constructing edge relationships of the graph based on node features, wherein a dual edge relationship construction mechanism is used to simultaneously consider spatial position relationships and state similarity, wherein the spatial position relationships include spatial adjacency relationships and dynamic association relationships, and for spatial adjacency relationships, a distance-based Gaussian kernel function is used to define the association strength: in, is the space matrix, d ij is the distance between node i and node j, is the distance scale parameter, controlling the decay rate of spatial correlation, is the distance threshold, and n is the total number of nodes.

[0010] Furthermore, the use of the anomaly detection model to extract spatiotemporal features from the constructed spatiotemporal graph structure includes using the graph attention network of the anomaly detection model to learn spatial features, wherein the influence of important nodes is highlighted by adaptively learning the attention coefficients between nodes, which is expressed as: in, represents the output feature vector of node i, with dimension d, (·) is the activation function, is the attention coefficient, which represents the influence weight of node j on node i. is a learnable feature transformation matrix, f is the dimension of the input feature, is the input feature vector of node j, and the attention coefficient The calculation method is: in, , It is the transformed representation of the node feature, and softmax is a normalization function that ensures that the sum of all attention coefficients is 1.

[0011] Furthermore, the method of using an anomaly detection model to extract spatiotemporal features from the spatiotemporal graph structure also includes extracting temporal features by combining causal convolution and long short-term memory networks, wherein the causal convolution network extracts local temporal patterns: in, is the time series feature output at time t, is the kth convolution kernel parameter, is the input feature at time tk, and then the LSTM network is used to step long-term dependencies: in, is the forward LSTM output feature, Output features for the backward LSTM, is the forward LSTM function, is the backward LSTM function, is the concatenated time series feature vector.

[0012] Furthermore, the feature fusion of the extracted spatiotemporal features includes utilizing an adaptive feature fusion mechanism to achieve an optimal combination of features by learning the importance weights of different features, which is expressed as: Among them, z is the final feature vector after fusion, , are the weight matrices of spatial features and temporal features respectively, , are the spatial and temporal feature vectors, and b is the bias vector.

[0013] Furthermore, the local and global anomaly scoring of marine ecosystems based on the multi-scale anomaly scoring mechanism includes first calculating the local anomaly score of each monitoring point, and reflecting the degree of deviation of the current state of the monitoring point from its historical normal state through the local anomaly score; and proposing a global anomaly scoring mechanism based on local anomaly detection, wherein the anomaly at the system level is identified by comparing the changes in feature distribution, which is expressed as: in, is the probability distribution of the current feature, is the probability distribution of the reference normal state, is the KL divergence loss. The distribution estimation adopts the kernel density estimation method: Where n is the number of samples, h is the smoothness of the control function, K(·) is the Gaussian kernel function, and z i is the ith sample point.

[0014] Second, a spatiotemporal graph neural network anomaly detection system for marine ecology, including: The data acquisition module is configured to acquire ocean monitoring data; The preprocessing module is configured to perform data preprocessing on the acquired ocean monitoring data; A graph construction module, configured to construct a spatiotemporal graph structure based on preprocessed ocean monitoring data; The feature extraction module is configured to extract spatiotemporal features of the constructed spatiotemporal graph structure using the anomaly detection model; The feature fusion module is configured to perform feature fusion on the extracted spatiotemporal features; The scoring module is configured to score local and global anomalies of the marine ecosystem based on a multi-scale anomaly scoring mechanism and output anomaly detection results.

[0015] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device for the spatiotemporal graph neural network anomaly detection method for marine ecology.

[0016] In a fourth aspect, the present invention provides a terminal device including a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being suitable for being loaded by the processor and executing the aforesaid spatiotemporal graph neural network anomaly detection method for marine ecology.

[0017] In summary, the present invention has the following beneficial technical effects: (1) Effective fusion of multi-source heterogeneous data is achieved. The present invention designs a unified data processing framework that can simultaneously process multi-source heterogeneous data such as satellite remote sensing data, water quality monitoring data, and biological survey data, making full use of the information contained in various types of data to improve the comprehensiveness and accuracy of anomaly detection. By constructing a spatiotemporal graph structure, the correlation between different data sources is effectively captured, providing a more comprehensive feature expression for anomaly detection.

[0018] (2) Improved the ability of spatiotemporal feature extraction. The present invention realizes the effective extraction of spatiotemporal features of marine ecosystems by designing a temporal feature extraction module that combines causal convolution and bidirectional LSTM network, as well as a spatial feature learning module based on graph attention mechanism. The adaptive feature fusion mechanism can automatically adjust the importance weights of spatial features and temporal features according to different scenarios, thereby enhancing the adaptability of the model.

[0019] (3) Realize the early identification and early warning of marine ecosystem anomalies. The present invention designs a multi-scale anomaly scoring mechanism and an adaptive early warning mechanism, which can not only detect anomalies that have occurred in a timely manner, but also predict potential anomaly risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of a spatiotemporal graph neural network anomaly detection method for marine ecology according to Example 1 of the present invention. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0022] Example 1 Reference Figure 1 , a spatiotemporal graph neural network anomaly detection method for marine ecology in this embodiment includes: Obtain ocean monitoring data; Carry out data preprocessing on the acquired ocean monitoring data; Constructing spatiotemporal graph structures based on preprocessed ocean monitoring data; Use anomaly detection models to extract spatiotemporal features of constructed spatiotemporal graph structures; Perform feature fusion on the extracted spatiotemporal features; Scoring of local and global anomalies in marine ecosystems based on a multi-scale anomaly scoring mechanism; Output anomaly detection results.

[0023] Specifically: S1 Data acquisition and preprocessing; First, the present invention designs a system data preprocessing process based on the characteristics of multi-source heterogeneous data of marine ecosystems. Since marine monitoring data is easily affected by environmental, equipment and human factors, the original data often contains outliers, which will interfere with the subsequent anomaly detection effect. Therefore, it is necessary to identify and process outliers.

[0024] (1) Outlier processing, Taking into account the dynamic characteristics of marine ecological data, this paper proposes an improved method based on time series sliding window based on traditional statistical methods. This method can better adapt to the time-varying characteristics of data by dynamically calculating local statistical features. Specifically, within the time window w, the identification criterion for outliers is: in, is the processed data set, x is the observed value, is the mean of the data series, is the standard deviation of the data series, n is the amount of data, and Use sliding window to calculate: in, is the local mean at time t, is the local standard deviation at time t, x i is the observation value at the i-th moment in the window, and t is the current time step.

[0025] (2) Missing value filling, Marine monitoring data often have missing values, and simple interpolation methods cannot guarantee the accuracy of the filling results. To this end, the present invention considers the filling strategy with dual constraints of time and space, makes full use of the temporal continuity and spatial correlation of the data, and obtains more reliable filling values ​​by combining the temporal prediction and spatial interpolation results: Among them, x t and x s are the results of temporal interpolation and spatial interpolation, respectively. and is the weight coefficient of the two terms, and .

[0026] S2 spatiotemporal graph structure construction; After completing the data preprocessing, the next step is to construct a graph structure that can effectively express the complex relationships of the marine ecosystem. Considering that the marine ecosystem has obvious spatiotemporal coupling characteristics, the present invention designs a dynamic spatiotemporal graph structure.

[0027] (1) Node feature construction: To comprehensively characterize the state characteristics of the monitoring points, the present invention designs a triple feature vector that integrates environmental, biological and temporal information: in, is the complete feature vector, d is the total feature dimension, is the environmental parameter characteristic, It is a biological indicator characteristic. It is the time characteristic, d=d1+d2+d3.

[0028] Based on the above feature representation, we further construct the edge relationship of the graph, using the dual edge relationship construction mechanism, while considering the spatial position relationship and state similarity. First, the spatial adjacency relationship, using the distance-based Gaussian kernel function to define the association strength: in, is the space matrix, d ij is the distance between node i and node j, is the distance scale parameter, controlling the decay rate of spatial correlation, is the distance threshold, and n is the total number of nodes.

[0029] The second is the dynamic association relationship, which is defined based on the similarity of node features:

[0030] in, is the dynamic incidence matrix, xi and x j is the feature vector of nodes i and j, cos(·) is the cosine similarity function, sim(i,j) is the comprehensive similarity between nodes i and j, is the similarity threshold.

[0031] (2) Temporal dependency modeling In order to capture the temporal evolution characteristics of the system state, the present invention designs a temporal correlation matrix. This matrix not only considers the state similarity, but also introduces a time decay factor to reflect the impact of time distance on the correlation strength: in, is the state vector of node i at time t, is the state vector of node j at time tk, is the time scale parameter, which controls the influence of state differences. is the time interval, is the time decay coefficient, controlling the duration of the time, and k is the time delay step.

[0032] S3 anomaly detection model; Based on the constructed spatiotemporal graph structure, the present invention designs a hierarchical anomaly detection model. The model extracts spatiotemporal features through graph neural networks and implements anomaly detection in combination with a multi-scale anomaly scoring mechanism.

[0033] (1) Spatial feature learning, After constructing the spatiotemporal graph structure, an effective feature extraction mechanism needs to be designed. Considering the differences in the importance of different monitoring points in the marine ecosystem, this paper uses a graph attention network for spatial feature learning. The network adaptively learns the attention coefficients between nodes and highlights the influence of important nodes: in, represents the output feature vector of node i, with dimension d, (·) is the activation function, is the attention coefficient, which represents the influence weight of node j on node i. is a learnable feature transformation matrix, f is the dimension of the input feature, is the input feature vector of node j. Attention coefficient The calculation method is: in, , It is the transformed representation of the node feature, and softmax is a normalization function that ensures that the sum of all attention coefficients is 1.

[0034] Through the above steps, the model can automatically identify and focus on key nodes and relationships in anomaly detection.

[0035] (2) Temporal feature learning, Abnormalities in marine ecosystems often manifest as changes in temporal patterns. In order to capture such abnormal temporal features, the present invention designs a temporal feature extraction module that combines causal convolution and long short-term memory networks.

[0036] First, local temporal patterns are extracted through a causal convolutional network: in, is the time series feature output at time t, is the kth convolution kernel parameter, is the input feature at time tk, and then the LSTM network is used to step long-term dependencies: in, is the forward LSTM output feature, Output features for the backward LSTM, is the forward LSTM function, is the backward LSTM function, is the concatenated time series feature vector.

[0037] (3) Feature fusion and anomaly detection, In order to comprehensively utilize spatial and temporal features for anomaly detection, this paper designs an adaptive feature fusion mechanism. This mechanism achieves the optimal combination of features by learning the importance weights of different features: Among them, z is the final feature vector after fusion, , are the weight matrices of spatial features and temporal features respectively, , are the spatial and temporal feature vectors, and b is the bias vector.

[0038] S4 Multi-scale anomaly scoring mechanism; Considering that marine ecosystem anomalies may occur at both local and global scales, this paper proposes a multi-scale anomaly scoring mechanism.

[0039] (1) Local abnormality score, First, the local anomaly score of each monitoring point is calculated. This score reflects the degree of deviation between the current state of the monitoring point and its historical normal state: in, is the importance weight of the i-th dimension feature, is the i-th dimension feature value of the current observation, and is the historical normal mean of the i-th dimension feature. In order to improve the stability of detection, the present invention adopts a local statistical feature update mechanism based on a sliding window: in, is the update rate parameter, which is used to control the degree of retention of historical information, t is the current time step, w is the size of the sliding window, is the historical observation value of the i-th dimension feature in the window.

[0040] (2) Global anomaly score, Based on local anomaly detection, the present invention further designs a global anomaly scoring mechanism. This mechanism identifies system-level anomalies by comparing changes in feature distribution: in, is the probability distribution of the current feature, is the probability distribution of the reference normal state, is the KL divergence loss. The distribution estimation adopts the kernel density estimation method: Where n is the number of samples, h is the smoothness of the control function, K(·) is the Gaussian kernel function, and z i is the ith sample point.

[0041] The abnormality detection of the marine ecosystem is completed through the above steps S1-S4.

[0042] Example 2 This embodiment provides a spatiotemporal graph neural network anomaly detection system for marine ecology, including: The data acquisition module is configured as follows: A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device for a method for detecting anomalies of a spatiotemporal graph neural network for marine ecology.

[0043] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to implement a spatiotemporal graph neural network anomaly detection method for marine ecology.

[0044] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A spatiotemporal graph neural network anomaly detection method for marine ecology, characterized in that: include: Obtain ocean monitoring data; Carry out data preprocessing on the acquired ocean monitoring data; Constructing spatiotemporal graph structures based on preprocessed ocean monitoring data; Use anomaly detection models to extract spatiotemporal features of constructed spatiotemporal graph structures; Perform feature fusion on the extracted spatiotemporal features; Scoring of local and global anomalies in marine ecosystems based on a multi-scale anomaly scoring mechanism; Output anomaly detection results.

2. According to claim 1, a method for detecting anomalies of a spatiotemporal graph neural network for marine ecology is characterized in that: include: The data preprocessing of the acquired ocean monitoring data includes outlier processing and missing value filling of the ocean monitoring data, wherein the outlier processing includes an improved method based on a time series sliding window. The criterion is to dynamically calculate local statistical features to adapt to the time-varying characteristics of the data. Within the time window w, the identification criterion of outliers is expressed as: in, is the processed data set, x is the observed value, is the mean of the data series, is the standard deviation of the data series, n is the amount of data, and Use sliding window to calculate: in, is the local mean at time t, is the local standard deviation at time t, x i is the observation value at the i-th moment in the window, and t is the current time step.

3. The method for detecting anomalies of a spatiotemporal graph neural network for marine ecology according to claim 2 is characterized in that: include: The construction of the spatiotemporal graph structure based on the preprocessed ocean monitoring data includes constructing node features based on triple feature vectors and integrating environmental, biological and temporal information; The edge relationship of the graph is constructed based on the node features. The dual edge relationship construction mechanism is used to consider the spatial position relationship and state similarity at the same time. The spatial position relationship includes spatial adjacency relationship and dynamic association relationship. For the spatial adjacency relationship, the distance-based Gaussian kernel function is used to define the association strength: in, is the space matrix, d ij is the distance between node i and node j, is the distance scale parameter, controlling the decay rate of spatial correlation, is the distance threshold, and n is the total number of nodes.

4. The method for detecting anomalies of a spatiotemporal graph neural network for marine ecology according to claim 3 is characterized in that: include: The anomaly detection model is used to extract spatiotemporal features from the spatiotemporal graph structure, including using the graph attention network of the anomaly detection model to learn spatial features, wherein the influence of important nodes is highlighted by adaptively learning the attention coefficients between nodes, which is expressed as: in, represents the output feature vector of node i, with dimension d, (·) is the activation function, is the attention coefficient, which represents the influence weight of node j on node i. is a learnable feature transformation matrix, f is the dimension of the input feature, is the input feature vector of node j, and the attention coefficient The calculation method is: in, , It is the transformed representation of the node feature, and softmax is a normalization function that ensures that the sum of all attention coefficients is 1.

5. The method for detecting anomalies of a spatiotemporal graph neural network for marine ecology according to claim 4 is characterized in that: include: Further, the use of the anomaly detection model to extract spatiotemporal features from the constructed spatiotemporal graph structure also includes extracting temporal features by combining causal convolution and long short-term memory networks, wherein the causal convolution network extracts local temporal patterns: in, is the time series feature output at time t, is the kth convolution kernel parameter, is the input feature at time tk, and then the LSTM network is used to step long-term dependencies: in, is the forward LSTM output feature, Output features for the backward LSTM, is the forward LSTM function, is the backward LSTM function, is the concatenated time series feature vector.

6. The method for detecting anomalies of a spatiotemporal graph neural network for marine ecology according to claim 5, characterized in that: include: The feature fusion of the extracted spatiotemporal features includes utilizing an adaptive feature fusion mechanism to achieve an optimal combination of features by learning the importance weights of different features, which is expressed as: Among them, z is the final feature vector after fusion, , are the weight matrices of spatial features and temporal features respectively, , are the spatial and temporal feature vectors, and b is the bias vector.

7. The method for detecting anomalies of a spatiotemporal graph neural network for marine ecology according to claim 6 is characterized in that: include: The local anomaly and global anomaly scoring of the marine ecosystem based on the multi-scale anomaly scoring mechanism includes first calculating the local anomaly score of each monitoring point, and reflecting the degree of deviation of the current state of the monitoring point from its historical normal state through the local anomaly score; and proposing a global anomaly scoring mechanism based on local anomaly detection, wherein the anomaly at the system level is identified by comparing the changes in feature distribution, which is expressed as: in, is the probability distribution of the current feature, is the probability distribution of the reference normal state, It is the KL divergence loss, and the distribution estimation adopts the kernel density estimation method: Where n is the number of samples, h is the smoothness of the control function, K(·) is the Gaussian kernel function, and z i is the ith sample point.

8. A spatiotemporal graph neural network anomaly detection system for marine ecology, characterized in that: include: The data acquisition module is configured to acquire ocean monitoring data; The preprocessing module is configured to perform data preprocessing on the acquired ocean monitoring data; A graph construction module, configured to construct a spatiotemporal graph structure based on preprocessed ocean monitoring data; The feature extraction module is configured to extract spatiotemporal features of the constructed spatiotemporal graph structure using the anomaly detection model; The feature fusion module is configured to perform feature fusion on the extracted spatiotemporal features; The scoring module is configured to score local and global anomalies of the marine ecosystem based on a multi-scale anomaly scoring mechanism and output anomaly detection results.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

10. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method as claimed in claim 1 .

Images