A celestial body intelligent detection system based on synchronous networking

Abstract

The present invention relates to the field of intelligent celestial body detection, specifically to an intelligent celestial body detection system based on synchronous networking, comprising an array of synchronous networked observation stations, a data processing module, a networked data intelligent analysis module and a celestial body classification module; the present invention adopts an array of synchronous networked observation stations to conduct comprehensive detection and analysis of celestial body types with lattice data, thereby reducing detection deviation and increasing classification accuracy; the present invention designs a multi-graph fusion dilated convolution model based on wavelet transform, considers short-term fluctuations and time series characteristics of data, designs a multi-graph fusion dilated convolution model to capture short-term fluctuation characteristics, processes approximate coefficients and captures time series characteristics, conducts comprehensive analysis and classification, and obtains accurate results of celestial body detection classification; the present invention designs a dynamic time regularization processing method based on interpolation, calculates the similarity between time series lattice observation data, and fills in missing data with the average value of the most similar values, thereby increasing the integrity of time series lattice observation data.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent celestial body detection, and specifically is a celestial body intelligent detection system based on synchronous networking. Background Art

[0002] Astronomical detection refers to the classification and summarization of celestial bodies detected during astronomical detection missions. In the fields of astronomy and astrophysics, astronomical detection aims to obtain information about celestial objects through various observation methods in order to study the nature and evolution of the universe. However, the existing astronomical detection target classification system uses a single astronomical detection telescope to detect the position of celestial targets, which has instability and detection deviation, and the detection subdivision is insufficient, resulting in unclear and inaccurate detection information. Summary of the invention

[0003] In order to overcome the defects of the prior art, the present invention adopts a synchronous networked observation station array, transmits and receives detection signals at multiple points, and uses the large-aperture time-series dot matrix observation data to conduct comprehensive detection and analysis of celestial body types, thereby reducing detection deviations and increasing classification accuracy. In view of the problem of insufficient subdivision of the prior art detection, the present invention creatively designs a multi-image fusion dilated convolution model based on wavelet transform, considers the two main characteristics of time-series dot matrix observation data, namely short-term fluctuations and time series characteristics, and uses wavelet transform to convert the time series dot matrix observation data into a multi-image fusion dilated convolution model based on wavelet transform. The lattice observation data is decomposed into approximate coefficients and detail coefficients, and a multi-image fusion dilated convolution model is designed to process the detail coefficients and capture the short-term fluctuation characteristics. The convolutional long short-term memory network is used to process the approximate coefficients and capture the timing characteristics. The time series lattice observation data is analyzed and classified through the comprehensive analysis of the short-term fluctuation characteristics and the timing characteristics to obtain the results of celestial body detection classification, thereby increasing the accuracy of detection classification. The present invention creatively designs a dynamic time warping processing method based on interpolation, uses a dynamic time warping algorithm to calculate the similarity between time series lattice observation data, and fills in the missing data through the average of the K most similar values, thereby increasing the integrity of the time series lattice observation data.

[0004] The present invention provides a celestial body intelligent detection system based on synchronous networking, comprising a synchronous networking observation station array, a data processing module, a networking data intelligent analysis module and a celestial body classification module;

[0005] The synchronous networked observation station array selects radio telescopes at multiple locations to form a synchronous network to observe known celestial bodies, form large-aperture time-series dot matrix observation data, and send the time-series dot matrix observation data to the data processing module;

[0006] The data processing module pre-processes the time series point matrix observation data and sends it to the networking data analysis module;

[0007] The network data intelligent analysis module uses a multi-image fusion dilated convolution model based on wavelet transform to extract spatiotemporal features from the time series lattice observation data, and obtains the matching results of spatiotemporal features and known celestial body types;

[0008] The celestial body classification module adopts a synchronous networked observation station array and a data processing module to observe unknown celestial bodies, and processes the returned large-aperture time-series point array observation data, adopts a networked data intelligent analysis module to extract spatiotemporal features, and obtains the category of the position celestial body based on the matching results of the spatiotemporal features and the celestial body type.

[0009] Furthermore, in the networking data intelligent analysis module, a data processing of a multi-graph fusion dilated convolution model based on wavelet transform specifically includes the following steps:

[0010] Step S1: Undirected graph construction, defining the lattice observation data as an undirected graph G=(V,E,A), where V represents the set of observation station nodes in the time series lattice observation data, E represents the set of undirected edges between nodes in the undirected graph G, and A is the adjacency matrix of the undirected graph;

[0011] Step S2: Node definition and feature extraction, define each observation station node in the time series lattice observation data as a node, and define the time series lattice observation data observed by each node as the node attribute feature: ; ;

[0012] in, Represents node n in the historical time series The time series observation data collected in the Represents node n at time step Observation data at time Indicates that all nodes are in the historical time series The time series lattice observation data within Indicates the length of the time series;

[0013] Step S3: Dynamic time warping based on interpolation is performed on the node attribute features between all nodes based on difference to obtain an interpolated time series, which specifically includes the following steps:

[0014] Step S31: Calculate the similarity between nodes using the dynamic time warping algorithm: ;

[0015] in, and Represents nodes With Node Node attribute characteristics, Representative Node With Node The difference between the node attribute characteristics of represents the minimum cumulative cost of the path between nodes, Representative Node With Node The similarity between the node attribute features;

[0016] Step S32: Setting a threshold and performing interpolation processing: ; ;

[0017] In the formula, Represents the node attribute feature with the highest similarity under the preset threshold. is the preset threshold, which means the selection The node attribute features with the highest similarity , represent The weighted average of the attribute features of the nodes with the highest similarity, Represents the time series point matrix observation data after interpolation processing, that is, the interpolation time series;

[0018] Step S4: wavelet transform, input the interpolation time series into a low-pass filter and a high-pass filter to obtain a low-pass filter component and a high-pass filter component: ; ;

[0019] In the formula, Represents the node attribute characteristics in the time series lattice observation data after interpolation processing, represents low-pass filter processing, represents the low-pass filter component, represents high-pass filter processing, represents the high-pass filter component;

[0020] Step S5: multi-level transformation, performing wavelet transformation on the low-pass filter component again to obtain a secondary low-pass filter component and a secondary high-pass filter component, and repeating this operation m times, that is, performing wavelet transformation on the low-pass filter component obtained each time again to obtain a final low-pass filter component and a final high-pass filter component;

[0021] Step S6: adjacency matrix design and feature extraction, feature extraction and fusion based on the adjacency matrix are performed on the final high-pass filter component and the final low-pass filter component to obtain the spatiotemporal features of the time series lattice observation data, specifically including the following steps:

[0022] Step S61: construct a spatial correlation similarity matrix based on the final high-pass filter component: ;

[0023] In the formula, Represents the square of the spatial distance between nodes in the final high-pass filtered component, represents the standard deviation of the spatial distance, To adjust the preset parameters of the matrix distribution and sparsity, is an exponential function, represents the spatial correlation similarity matrix;

[0024] Step S62: construct a dynamic similarity matrix: ;

[0025] In the formula, and Represents the node in the final high-pass filter component With Node Node attribute characteristics, and Represent the Pearson correlation coefficient and the Spearman correlation coefficient, respectively. and To control the weight factor of the two correlation coefficients, is the counting operator, Representation Node The node attribute characteristics in the historical time series The average value within Represents the node in the final high-pass filter component With Node The square of the absolute value of the difference between the node attribute characteristics;

[0026] Step S63: extracting multi-graph spatial features, using a graph convolutional neural network model to extract features from the spatial correlation similarity matrix and the dynamic similarity matrix to obtain spatial features and dynamic similarity features;

[0027] Step S64: feature fusion and dilated convolution optimization, the spatial features and dynamic similarity features are fused and optimized based on dilated convolution to obtain the temporal spatial features: ; ; ; ;

[0028] In the formula, and Represent spatial features and dynamic similarity features respectively, is the intermediate fusion feature, represents the dilated convolution operation, represents a nonlinear activation function, represents the dilated convolution feature, represents the activation function, , and represents the learnable parameters, represents the optimization adjustment factor, Represents temporal spatial characteristics;

[0029] Step S65: spatiotemporal feature extraction, using a convolutional long short-term memory network to extract features from the low-pass filter component to obtain short-term fluctuation features, and then fusing the short-term fluctuation features with the temporal space features to obtain spatiotemporal features.

[0030] The beneficial results achieved by the present invention using the above scheme are as follows:

[0031] (1) In view of the problems of instability and detection deviation in the existing technology when detecting celestial body targets through a single celestial body detection telescope, the present invention adopts a synchronous networked observation station array, transmits and receives detection signals at multiple points, and uses large-aperture time-series dot matrix observation data to conduct comprehensive detection and analysis of celestial body types, thereby reducing detection deviation and increasing classification accuracy;

[0032] (2) In order to solve the problem of insufficient subdivision of detection in the prior art, the present invention creatively designs a multi-image fusion dilated convolution model based on wavelet transform, taking into account the two main characteristics of time series lattice observation data, namely short-term fluctuation and time series characteristics, and uses wavelet transform to decompose the time series lattice observation data into approximate coefficients and detail coefficients. A multi-image fusion dilated convolution model is designed to process the detail coefficients and capture the short-term fluctuation characteristics, and a convolutional long short-term memory network is used to process the approximate coefficients and capture the time series characteristics. The time series lattice observation data is analyzed and classified by combining the short-term fluctuation characteristics and the time series characteristics, and the results of celestial body detection classification are obtained, thereby increasing the accuracy of detection classification.

[0033] (3) The present invention creatively designs a dynamic time warping processing method based on interpolation, uses a dynamic time warping algorithm to calculate the similarity between time series lattice observation data, and fills in missing data by taking the average of the K most similar values, thereby increasing the integrity of the time series lattice observation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A module diagram of a celestial body intelligent detection system based on synchronous networking provided by the present invention;

[0035] Figure 2 Schematic diagram of the data processing flow of a multi-image fusion dilated convolution model based on wavelet transform.

[0036] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0038] Example 1, see Figure 1 The present invention provides a celestial body intelligent detection system based on synchronous networking, including a synchronous networking observation station array, a data processing module, a networking data intelligent analysis module and a celestial body classification module;

[0039] The synchronous networked observation station array selects radio telescopes at multiple locations to form a synchronous network to observe known celestial bodies, form large-aperture time-series dot matrix observation data, and send the time-series dot matrix observation data to the data processing module;

[0040] The data processing module pre-processes the time series point matrix observation data and sends it to the networking data analysis module;

[0041] The network data intelligent analysis module uses a multi-image fusion dilated convolution model based on wavelet transform to extract spatiotemporal features from the time series lattice observation data, and obtains the matching results of spatiotemporal features and known celestial body types;

[0042] The celestial body classification module adopts a synchronous networked observation station array and a data processing module to observe unknown celestial bodies, and processes the returned large-aperture time-series point array observation data, adopts a networked data intelligent analysis module to extract spatiotemporal features, and obtains the category of the position celestial body based on the matching results of the spatiotemporal features and the celestial body type.

[0043] Example 2, see Figure 2 This embodiment is based on the above embodiment. In the networking data intelligent analysis module, a data processing method of a multi-graph fusion dilated convolution model based on wavelet transform specifically includes the following steps:

[0044] Step S1: Undirected graph construction, defining the lattice observation data as an undirected graph G=(V,E,A), where V represents the set of observation station nodes in the time series lattice observation data, E represents the set of undirected edges between nodes in the undirected graph G, and A is the adjacency matrix of the undirected graph;

[0045] Step S2: Node definition and feature extraction, define each observation station node in the time series lattice observation data as a node, and define the time series lattice observation data observed by each node as the node attribute feature: ; ;

[0046] in, Represents node n in the historical time series The time series observation data collected in the Represents node n at time step Observation data at time Indicates that all nodes are in the historical time series The time series lattice observation data within Indicates the length of the time series;

[0047] Step S3: Performing dynamic time warping based on interpolation to perform dynamic time warping based on differences on the node attribute features between all nodes to obtain an interpolated time series;

[0048] Step S4: wavelet transform, input the interpolation time series into a low-pass filter and a high-pass filter to obtain a low-pass filter component and a high-pass filter component: ; ;

[0049] In the formula, Represents the node attribute characteristics in the time series lattice observation data after interpolation processing, represents low-pass filter processing, represents the low-pass filter component, represents high-pass filter processing, represents the high-pass filter component;

[0050] Step S5: multi-level transformation, performing wavelet transformation on the low-pass filter component again to obtain a secondary low-pass filter component and a secondary high-pass filter component, and repeating this operation 4 times, that is, performing wavelet transformation on the low-pass filter component obtained each time again to obtain a final low-pass filter component and a final high-pass filter component;

[0051] Step S6: adjacency matrix design and feature extraction, feature extraction and fusion based on the adjacency matrix are performed on the final high-pass filter component and the final low-pass filter component to obtain the spatiotemporal characteristics of the time series lattice observation data.

[0052] Embodiment 3: This embodiment is based on the above embodiment, and step S3 specifically includes the following steps:

[0053] Step S31: Calculate the similarity between nodes using the dynamic time warping algorithm: ;

[0054] in, and Represents nodes With Node Node attribute characteristics, Representative Node With Node The difference between the node attribute characteristics of represents the minimum cumulative cost of the path between nodes, Representative Node With Node The similarity between the node attribute features;

[0055] Step S32: Setting a threshold and performing interpolation processing: ; ;

[0056] In the formula, Represents the node attribute feature with the highest similarity under the preset threshold. is the preset threshold, which means the selection The node attribute features with the highest similarity , represent The weighted average of the attribute features of the nodes with the highest similarity, Represents the time series point matrix observation data after interpolation processing, that is, the interpolated time series.

[0057] Embodiment 4: This embodiment is based on the above embodiment, and step S6 specifically includes the following steps:

[0058] Step S61: construct a spatial correlation similarity matrix based on the final high-pass filter component: ;

[0059] In the formula, Represents the square of the spatial distance between nodes in the final high-pass filtered component, represents the standard deviation of the spatial distance, To adjust the preset parameters of the matrix distribution and sparsity, is an exponential function, represents the spatial correlation similarity matrix;

[0060] Step S62: construct a dynamic similarity matrix: ;

[0061] In the formula, and Represents the node in the final high-pass filter component With Node Node attribute characteristics, and Represent the Pearson correlation coefficient and the Spearman correlation coefficient, respectively. and To control the weight factor of the two correlation coefficients, is the counting operator, Representation Node The node attribute characteristics in the historical time series The average value within Represents the node in the final high-pass filter component With Node The square of the absolute value of the difference between the node attribute characteristics;

[0062] Step S63: extracting multi-graph spatial features, using a graph convolutional neural network model to extract features from the spatial correlation similarity matrix and the dynamic similarity matrix to obtain spatial features and dynamic similarity features;

[0063] Step S64: feature fusion and dilated convolution optimization, the spatial features and dynamic similarity features are fused and optimized based on dilated convolution to obtain the temporal spatial features: ; ; ; ;

[0064] In the formula, and Represent spatial features and dynamic similarity features respectively, is the intermediate fusion feature, represents the dilated convolution operation, represents a nonlinear activation function, represents the dilated convolution feature, represents the activation function, , and represents the learnable parameters, represents the optimization adjustment factor, Represents temporal spatial characteristics;

[0065] Step S65: spatiotemporal feature extraction, using a convolutional long short-term memory network to extract features from the final low-pass filter component to obtain short-term fluctuation features, and then fusing the short-term fluctuation features with the temporal space features to obtain spatiotemporal features.

[0066] Embodiment 5: This embodiment is based on the above embodiment. In step S65, a convolutional long short-term memory network is used to extract features of the final low-pass filter component. Specifically, the following steps are included:

[0067] Data preprocessing:

[0068] The low-pass filter components were normalized, smoothed, and missing value filling preprocessing steps were performed to ensure data quality and stability;

[0069] Data reshaping:

[0070] Reshape the low-pass filter component into a three-dimensional data format suitable for the ConvLSTM model input, that is, (batch_size, time_steps, input_dim), where batch_size represents the batch size, time_steps represents the time step, and input_dim represents the feature dimension of each time step;

[0071] Build the ConvLSTM model:

[0072] Construct a ConvLSTM model, including convolutional layers, LSTM layers, and fully connected layers;

[0073] Model training:

[0074] Train the ConvLSTM model and optimize the model parameters to minimize the loss function;

[0075] Feature extraction:

[0076] During the training process of the ConvLSTM model, learning is performed to extract spatiotemporal features in the low-pass filtered component.

[0077] Embodiment 6: Based on the above embodiment, this embodiment adopts other graph convolution feature networks to replace a multi-graph fusion dilated convolution model based on wavelet transform for feature extraction, and conducts experimental analysis:

[0078] Pretreatment before the experiment:

[0079] The construction of the spectrum graph includes the following steps:

[0080] Step Q1: Prepare time series data:

[0081] Obtain the time series data that needs to be processed, that is, the low-pass filter component;

[0082] Step Q2: Preprocessing time series data:

[0083] Preprocess the low-pass filter components, including removing noise, filling missing values, and normalizing to ensure data quality and stability;

[0084] Step Q3: Select the spectrum analysis method:

[0085] Determine the spectrum analysis method to be used, and adopt Fourier transform as the spectrum analysis method;

[0086] Step Q4: Apply spectrum analysis method:

[0087] The low-pass filtered component is input into the selected spectrum analysis method to calculate the spectrum diagram of the time series data;

[0088] Step Q5: Spectrum graph construction:

[0089] The spectrum data obtained by spectrum analysis is visualized as a spectrum graph, with frequency as the horizontal axis and amplitude as the vertical axis to form a spectrum graph;

[0090] Step Q6: Spectrum graph visualization:

[0091] The constructed spectrum is visualized and displayed graphically to visually observe the characteristics of the low-pass filtering component in the frequency domain;

[0092] Experimental results analysis:

[0093] Graph convolutional feature network selection: GCN, GAT, GIN, T-GCN and STGCN models are selected as graph convolutional feature networks to replace a multi-graph fusion dilated convolution model based on wavelet transform for feature extraction, and the results are analyzed:

[0094] Model: GCN; RMSE: 8.37; MAD: 8.34; ACC: 0.79;

[0095] Model: GAT; RMSE: 9.54; MAD: 7.12; ACC: 0.85;

[0096] Model: GIN; RMSE: 7.53; MAD: 5.23; ACC: 0.78;

[0097] Model: T-GCN; RMSE: 6.37; MAD: 3.28; ACC: 0.89;

[0098] Model: STGCN; RMSE: 6.01; MAD: 3.02; ACC: 0.92;

[0099] Model: the model of the present invention; RMSE: 4.56; MAD: 2.73; ACC: 0.94;

[0100] By analyzing the celestial body classification detection indicators of different models under the same time series, it can be seen that compared with the GCN, GAT, GIN, T-GCN and STGCN models, a multi-graph fusion dilated convolution model based on wavelet transform has made significant progress. The analysis shows that the GCN model is insufficient in extracting temporal features. GAT and GIN introduce attention mechanisms and node mechanisms for optimization based on the GCN model, but they still fail to change the defects of temporal feature extraction. T-GCN introduces GRU to capture spatial and temporal features, but its accuracy is low. STGCN captures relevant features by combining one-dimensional convolution and gating mechanisms, but it relies on a fixed polynomial kernel, which limits its adaptability to complex celestial environment conditions.

[0101] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0102] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0103] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A celestial body intelligent detection system based on synchronous networking, characterized in that: It includes a synchronous networked observatory array, a data processing module, a networked data intelligent analysis module, and a celestial body classification module; The synchronous networked observation station array selects radio telescopes at multiple locations to form a synchronous network to observe known celestial bodies, form large-aperture time-series dot matrix observation data, and send the time-series dot matrix observation data to the data processing module; The data processing module pre-processes the time series point matrix observation data and sends it to the networking data analysis module; The network data intelligent analysis module uses a multi-image fusion dilated convolution model based on wavelet transform to extract spatiotemporal features from the time series lattice observation data, and obtains the matching results of spatiotemporal features and known celestial body types; The celestial body classification module adopts a synchronous networked observation station array and a data processing module to observe unknown celestial bodies, and processes the returned large-aperture time-series point array observation data, adopts a networked data intelligent analysis module to extract spatiotemporal features, and obtains the category of the position celestial body based on the matching results of the spatiotemporal features and the celestial body type.

2. The celestial body intelligent detection system based on synchronous networking according to claim 1, characterized in that: In the networking data intelligent analysis module, a data processing of a multi-graph fusion dilated convolution model based on wavelet transform specifically includes the following steps: Step S1: Undirected graph construction, defining the lattice observation data as an undirected graph G=(V,E,A), where V represents the set of observation station nodes in the time series lattice observation data, E represents the set of undirected edges between nodes in the undirected graph G, and A is the adjacency matrix of the undirected graph; Step S2: Node definition and feature extraction, define each observation station node in the time series lattice observation data as a node, and define the time series lattice observation data observed by each node as the node attribute feature: ； ； in, Represents node n in the historical time series The time series observation data collected in the Represents node n at time step Observation data at time Indicates that all nodes are in the historical time series The time series lattice observation data within Indicates the length of the time series; Step S3: Performing dynamic time warping based on interpolation to perform dynamic time warping based on differences on the node attribute features between all nodes to obtain an interpolated time series; Step S4: wavelet transform, input the interpolation time series into a low-pass filter and a high-pass filter to obtain a low-pass filter component and a high-pass filter component; Step S5: multi-level transformation, performing wavelet transformation on the low-pass filter component again to obtain a secondary low-pass filter component and a secondary high-pass filter component, and repeating this operation 4 times, that is, performing wavelet transformation on the low-pass filter component obtained each time again to obtain a final low-pass filter component and a final high-pass filter component; Step S6: adjacency matrix design and feature extraction, feature extraction and fusion based on the adjacency matrix are performed on the final high-pass filter component and the final low-pass filter component to obtain the spatiotemporal characteristics of the time series lattice observation data.

3. The celestial body intelligent detection system based on synchronous networking according to claim 2 is characterized in that: Step S3 specifically includes the following steps: Step S31: Calculate the similarity between node attributes through a dynamic time warping algorithm; Step S32: Setting the threshold and performing interpolation processing: ； ； In the formula, Represents the node attribute feature with the highest similarity under the preset threshold. is the preset threshold, which means the selection The node attribute features with the highest similarity , represent The weighted average of the attribute features of the nodes with the highest similarity, Represents the time series point matrix observation data after interpolation processing, that is, the interpolated time series.

4. The celestial body intelligent detection system based on synchronous networking according to claim 3 is characterized in that: Step S6 specifically includes the following steps: Step S61: constructing a spatial correlation similarity matrix according to the final high-pass filtering component; Step S62: construct a dynamic similarity matrix: ； In the formula, and Represents the node in the final high-pass filter component With Node Node attribute characteristics, and Represent the Pearson correlation coefficient and the Spearman correlation coefficient, respectively. and To control the weight factor of the two correlation coefficients, is the counting operator, Representation Node The node attribute characteristics in the historical time series The average value within Represents the node in the final high-pass filter component With Node The square of the absolute value of the difference between the node attribute characteristics; Step S63: extracting multi-graph spatial features, using a graph convolutional neural network model to extract features from the spatial correlation similarity matrix and the dynamic similarity matrix to obtain spatial features and dynamic similarity features; Step S64: feature fusion and dilated convolution optimization, fusing the spatial features and the dynamic similarity features and optimizing based on the dilated convolution to obtain the temporal spatial features; Step S65: spatiotemporal feature extraction, using a convolutional long short-term memory network to extract features from the final low-pass filter component to obtain short-term fluctuation features, and then fusing the short-term fluctuation features with the temporal space features to obtain spatiotemporal features.

Images