Sea level variation rate reconstruction method based on geometric algebraic feature learning
By employing a method based on geometric algebraic feature learning, a unified representation of ocean three-dimensional dynamic data is achieved. Using multi-vector encoding and convolutional gated recurrent units, the problem of neglecting the relationship between vector fields and scalar fields in traditional methods is solved, thus realizing high-precision reconstruction of sea level variability.
Patent Information
- Application Number
- CN202512018450.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-03-06
AI Technical Summary
Existing methods for simulating sea level variability are unable to accurately reproduce local dynamics. Traditional methods ignore the complex relationship between vector fields and scalar fields, leading to a decrease in the accuracy of the model under complex nonlinear dynamics.
A method based on geometric algebraic feature learning is adopted. By uniformly representing marine three-dimensional dynamic data in geometric algebraic space, multi-vector encoding, geometric algebraic convolution gated recurrent units and dual-path convolution feature enhancement modules are used to extract spatiotemporal features and reconstruct the context, and finally reconstruct sea level variability.
It improves the accuracy and robustness of sea-level variability reconstruction, enabling better capture of local and global dynamic characteristics and providing scientific support for coastal risk assessment and climate adaptation strategies.
Smart Images
Figure CN121615508A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of marine science and machine learning, and in particular to a method for reconstructing sea level variability based on geometric algebraic feature learning. Background Technology
[0002] Regional sea-level variability is driven by complex three-dimensional ocean dynamic processes. Its data typically originates from observations by various platforms such as satellites, buoys, and ships, forming a multidimensional spatiotemporal dataset containing scalar fields (e.g., temperature, salinity) and vector fields (e.g., eastward and northward current velocities). These variables are subject to geometrical constraints. Effectively extracting key features driving sea-level variability from this type of multidimensional data with its complex internal structure and constructing high-precision simulation models is a current challenge in the fields of marine science and geographic information science.
[0003] Traditional methods for simulating sea-level variability primarily rely on dynamic scale models (such as general circulation models, GCMs) or machine learning techniques (such as convolutional neural networks (CNNs) and long short-term memory networks (LSTMs). GCMs, limited by computational complexity, coarse resolution (approximately 100 km grid), and the reduction in vertical layers in shallow water, struggle to accurately reproduce local dynamics. Machine learning methods capture spatiotemporal features through convolution and recursion mechanisms, treating input variables (such as temperature, salinity, and seawater velocity) as independent channels and simply concatenating them along the channel axes. However, this independent processing neglects the complex relationships within the vector field and between the vector field and the scalar field, limiting the model's ability to extract meaningful representations, particularly leading to decreased accuracy in simulations under complex nonlinear dynamics. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide a sea level variability reconstruction method based on geometric algebraic feature learning that considers the inherent geometric relationships between variables and improves the accuracy and robustness of regional sea level variability reconstruction.
[0005] Technical solution: The sea level variability reconstruction method based on geometric algebraic feature learning described in this invention includes...
[0006] (1) In the geometric algebra space, the preprocessed marine three-dimensional dynamic data are uniformly represented based on generalized primitives to form a multi-vector code;
[0007] (2) Multi-vector encoding is input into geometric algebraic convolution gated recurrent units to extract spatiotemporal features and obtain spatiotemporal feature sequences;
[0008] (3) Introduce a time stamp matrix, reconstruct the spatiotemporal feature sequence in context, input it into the dual-path convolution feature enhancement module, extract local and global temporal features, and obtain the enhanced feature sequence;
[0009] (4) The enhanced feature sequence is input into a multi-layer gated recurrent unit and decoded step by step. The decoded sequence is input into a linear mapping layer, projected onto the target dimension, and outputs the sea level variability simulation result sequence to complete the reconstruction of sea level change.
[0010] Furthermore, in step (1), the preprocessing of the marine three-dimensional dynamic data is as follows:
[0011] Ocean three-dimensional dynamic data sequence The shape is defined as [T, C*D*H*W], where T represents the total selected time length, C represents the number of ocean variables, D represents the number of depth layers, and H and W represent the spatial grid dimensions, respectively.
[0012] The outlier data is calculated by subtracting the climate average data from the monthly average data to obtain the outlier sequence for each variable; the proportion of valid data in each depth layer is calculated, and if the proportion of valid data is less than 80%, the depth layer is removed; sporadic missing values in the depth layer are filled with the average value of that layer; normalization is performed; and the preprocessed ocean three-dimensional dynamic data sequence is obtained.
[0013] Furthermore, ocean variables include potential temperature, salinity, eastward seawater flow velocity, and northward seawater flow velocity.
[0014] Furthermore, in step (1), a unified representation is performed to form a multi-vector code, as follows:
[0015] Preserve the number of variable channels and spatial dimensions of the preprocessed ocean three-dimensional dynamic data sequence, split the time series data into batches, and determine the total time length. Divide into multiple subsequences to obtain a five-dimensional tensor;
[0016] Transform the five-dimensional tensor into a multi-vector encoding in geometric algebraic space. :
[0017]
[0018] in, It is a multi-vector encoding at a time step t. Salinity For temperature, The eastward seawater flow velocity. This refers to the northward seawater current speed; As scalar primitives, and For vector primitives, It is a two-vector basis.
[0019] Furthermore, in step (2), the geometric algebraic convolution gated cyclic unit operator satisfies:
[0020]
[0021] in, This represents the output of a geometric algebraic convolution gated loop unit. This represents the geometric algebraic convolution gated cyclic unit operator. The geometric algebraic convolution gated recurrent unit operator represents a multi-vector input at one time step t, and includes geometric algebraic convolution operations. And gating mechanisms.
[0022] Furthermore, in step (2), the gating mechanism is as follows:
[0023]
[0024] in, Indicates an update to the door. This indicates that the door is being reset. Indicates the candidate hidden state. Indicates the final hidden state. This indicates that the input is a multi-vector at one time step t. This represents the hidden state at time step t-1. It is the Sigmoid activation function. The hyperbolic tangent activation function is used. For geometric algebraic operations with different convolution kernels, This represents the Hadamard product.
[0025] Furthermore, step (3) is as follows:
[0026] The spatiotemporal feature sequence and the time stamp matrix are fused together in the feature dimension by concatenation operation to form an embedded temporal context representation. The embedding dimension is an extension of the original feature dimension.
[0027] The reconstructed sequence is input into the dual-path convolutional feature enhancement module to extract local and global temporal features. Path 1 is used to capture global dynamics, and path 2 is used to capture local dynamics. The output of path 1 is processed by residuals and then added to the output of path 2 to obtain the enhanced feature sequence.
[0028] Furthermore, the function expression for path 1 is as follows:
[0029]
[0030] in, Indicates the input tensor. Indicates the output tensor. The Sigmoid activation function is used, MaxPool(●) represents the max pooling operation, and Conv1D(●) represents the one-dimensional convolution operation.
[0031] Furthermore, the function expression for path 2 is as follows:
[0032]
[0033] in, Indicates the input tensor. This represents the output tensor, and AvgPool(●) represents the average pooling operation.
[0034] Furthermore, path 1 suffers feature loss after multi-layer convolutional pooling, so a residual mapping operator is introduced. :
[0035]
[0036] in, Indicates the mapped output. This indicates that path 1 outputs a tensor;
[0037] The final path fusion output is
[0038]
[0039] in, This represents the output after fusion. This indicates that path 2 outputs a tensor. This represents the output tensor mapped by path 1, where C is the channel after convolution.
[0040] Beneficial Effects: Compared with existing technologies, the significant advantages of this invention are: 1. This invention uses geometric algebra to uniformly represent multidimensional ocean variables, preserving the interaction between temperature, salinity, and velocity components, achieving high-precision reconstruction of regional sea-level variability time series, while improving the model's adaptability to local and global dynamic features, providing more reliable scientific support for coastal risk assessment and climate adaptation strategies; 2. This invention introduces geometric algebraic convolutional gated recurrent units, temporal context embedding, and dual-path feature enhancement modules, enabling it to more fully extract the spatial structure and temporal dependence features of sea-level variability, thereby improving reconstruction accuracy. Attached Figure Description
[0041] Figure 1 This is a flowchart of the present invention;
[0042] Figure 2 This is a schematic diagram of sea level variation at a certain coastal station. Detailed Implementation
[0043] The sea level variability reconstruction method based on geometric algebraic feature learning described in this invention includes the following steps:
[0044] (1) Ocean data preprocessing and geometric algebraic multivectorization representation
[0045] (11) Collect monthly mean sea level data for designated tidal stations, with a time range from January 1993 to December 2023; remove abnormal observation data: if the sea level height change between two adjacent months exceeds 500 mm, the record is judged as abnormal and deleted; remove tidal stations with missing data for more than 5 years to ensure data continuity; exclude tidal stations located in narrow bays, semi-enclosed waters or estuaries to avoid topographic effects; divide the remaining data into missing values and non-missing values, and perform detrending processing on the non-missing value data, retaining only seasonal and periodic components; use a 13-month moving average low-pass filter to extract interannual variation signals from the preprocessed data and normalize them.
[0046] Collect three-dimensional ocean dynamic data (such as potential temperature, salinity, and eastward and northward seawater current velocities), calculate outlier data: subtract climate average data from monthly average data to obtain the outlier sequence for each variable; calculate the proportion of valid data for each depth layer, and remove the depth layer if the proportion of valid data is less than 80%; fill in the missing values in the depth layer with the average value of that layer; and perform normalization.
[0047] The input ocean physics tensor in geometric algebra The geometric algebra encoding in space is defined as follows:
[0048]
[0049] in, It is a multi-vector encoding at a time step t. The physical quantity components to be encoded. It is a generalization primitive.
[0050] Original sequence The shape is defined as [T, C*D*H*W], where T represents the total selected time length, C represents the number of ocean variables, D represents the number of depth layers, and H and W represent the spatial grid dimensions, respectively. To accommodate batch processing and model training requirements, the original sequence... Convert to batch input using the following steps. :
[0051] (12) Batch processing: Divide the time series data into batches of size B, where B represents the number of samples in each batch, thus introducing the batch dimension.
[0052] (13) Sequence length division: Divide the total time length Divide into multiple subsequences, the length of which is defined as This ensures that each subsequence independently represents a continuous segment of time step.
[0053] (14) Channel and spatial dimension preservation: Preserve the variable channel number and depth layer combination of the original sequence. The spatial dimensions H and W are used as feature dimensions for subsequent processing.
[0054] (15) Tensor Reconstruction: Through the above batching and sequence partitioning, the data is reconstructed into a five-dimensional tensor. .data As the initial input, it is first processed through geometric algebra encoding to adapt to subsequent processing. Processing. The encoding process is as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Mapping to geometric algebra The encoding formula for multiple vectors in space is:
[0055]
[0056] in, It is a multi-vector encoding at a time step t. For salinity, For temperature, This represents the east / northwest direction seawater flow velocity. As scalar primitives, and For vector primitives, It is a two-vector basis.
[0057] (2) Spatiotemporal feature extraction based on geometric algebraic convolution gated recurrent units
[0058] (21) The expression for the geometric algebraic convolution gated loop unit operator is:
[0059]
[0060] in, This represents the output of a geometric algebraic convolution gated loop unit. This represents the geometric algebraic convolution gated cyclic unit operator. This indicates that the input is a multi-vector at one time step t, and the operator includes a geometric algebraic convolution operation. And gating mechanisms.
[0061] (211) Definition of Geometric Algebraic Convolution Operator
[0062]
[0063] in, This represents the output of the geometric algebraic convolution operator. This represents the geometric algebraic convolution operation. This indicates a multi-vector input.
[0064] (212) Gating mechanism
[0065]
[0066] in, Indicates an update to the door. This indicates that the door is being reset. Indicates the candidate hidden state. Indicates the final hidden state. This indicates that the input is a multi-vector at one time step t. This represents the hidden state at time step t-1. It is the Sigmoid activation function. The hyperbolic tangent activation function is used. For geometric algebraic operations with different convolution kernels, This represents the Hadamard product.
[0067] Initial hidden state The encoded data in step 1 Input geometric algebraic convolution gated recurrent unit Perform timing processing to generate output. and update the hidden state. This allows for the gradual evolution of sequence data. Initial input and first-step output:
[0068] Initial time step hour, Input :
[0069]
[0070] Geometric Algebraic Convolution Operations Applied to and The gating mechanism updates the hidden state:
[0071]
[0072] Output As This refers to the input of the sequence at the next time step. The evolution of subsequent time steps:
[0073] For subsequent time steps Input of the current time step Based on the output of the previous time step Update, recursive calculation:
[0074]
[0075] Gating mechanism updates hidden state:
[0076]
[0077] Output As That is, the input data of the next time step is used to derive... Stacking R represents the number of channels updated after convolution.
[0078] (3) Embedding of temporal context information and dual-path feature enhancement
[0079] (31) Timing Context Embedding Module
[0080] Based on the output tensor extracted by the Geometric Algebraic Convolution Gated Recurrent Unit (GAConvGRU), a time-stamped matrix is introduced to reconstruct the context of the sequence. Let the output of GAConvGRU be:
[0081]
[0082] Where B is the batch size, L is the time step, and F is the feature dimension. Let the time stamp matrix be:
[0083]
[0084] Where B is the batch size and L is the time step. This represents the external time series label dimension for each time step.
[0085] The temporal context embedding operator is:
[0086]
[0087] in This indicates a splicing operation. This represents the dimension after embedding.
[0088] (32) Dual-path convolutional feature enhancement module
[0089] To simultaneously extract local and global temporal features, this step designs a dual-path convolutional enhancement module. The function expression for path 1 is as follows:
[0090]
[0091] in, Indicates the input tensor. Indicates the output tensor. The Sigmoid activation function is used, MaxPool(●) represents the max pooling operation, and Conv1D(●) represents the one-dimensional convolution operation.
[0092] The function expression for path 2 is:
[0093]
[0094] in, Indicates the input tensor. Indicates the output tensor. The Sigmoid activation function is used, AvgPool(●) represents the average pooling operation, and Conv1D(●) represents the one-dimensional convolution operation.
[0095] To avoid feature loss after multiple convolutional pooling steps in path 1, a residual mapping operator is introduced. :
[0096]
[0097] in, Indicates the mapped output. This indicates that path 1 outputs a tensor.
[0098] The final path fusion output is:
[0099]
[0100] in, This represents the output after fusion. This indicates that path 2 outputs a tensor. This represents the output tensor mapped by path 1, where C is the channel after convolution.
[0101] At the model input end, a time stamp matrix is simultaneously received. This matrix describes the external time series information (year, month, day) for each time step. In step 2, this matrix is not used; (Step 2...) Processing Restructuring ,in The time stamp is entered into the time context embedding module. With time stamp matrix It is obtained by concatenation along the feature dimension. , This represents the dimension after embedding.
[0102] Will Input path 1, get Then Input path 2, get ,Will After residual mapping After that, I received Output of path 2 Adding them together, we get , as subsequent input.
[0103] (4) Timing Decoding and Analog Output
[0104] (41) Timing Decoding
[0105] Gated cyclic unit (GRU) operator This operator is used for time-step decoding of the input tensor and is expressed as follows:
[0106]
[0107] in, Indicates the decoding sequence. Represents a time series.
[0108] (42) Linear layer
[0109] Linear mapping operator This operator is used to perform projection onto the target dimension, and its expression is:
[0110]
[0111] Indicates linear output. This represents the decoded sequence.
[0112] Using the gated cyclic unit (GRU) operator In step 3 Perform time-step decoding:
[0113]
[0114] in, The dimension representing the hidden state of a gated recurrent unit (GRU).
[0115] Decoding sequence Input to linear mapping operator :
[0116]
[0117] in, This represents the sequence of simulation results for sea level variability.
[0118] (5) Experimental verification
[0119] (51) Experimental environment
[0120] The experiment was conducted on a system running Windows 11 Professional, equipped with a 13th generation Intel® Core™ i9-13900 CPU with a clock speed of 2.00 GHz. The system was configured with an NVIDIA RTX A4000 GPU and 64GB of RAM. The programming environment used was PyTorch version 1.12.1 and CUDA version 12.2.
[0121] (52) Data source
[0122] The experimental data came from the following two main data sources:
[0123] Permanent Mean Sea Level Service (PSMSL)
[0124] The data is provided by PSMSL, which was founded in 1933 and is supported by the International Union of Geodesy and Geophysics (IUGG) and the Intergovernmental Oceanographic Commission (IOC) of UNESCO. It is an important component of the Global Sea Level Observation System (GLOSS).
[0125] The data are derived from monthly mean sea level records from global tidal observation stations, based on water levels measured by tide gauges, and referenced to a revised local reference datum. Short-term fluctuations caused by tides are excluded.
[0126] Data from the SANTA MONICA (MUNICIPAL PIER) tidal station, located on a coastline, were selected, covering the period from January 1993 to December 2023. The station is situated in a relatively open marine area, avoiding the influence of narrow bays or river estuaries.
[0127] Copernicus Marine Environmental Monitoring Service (CMEMS)
[0128] The data comes from CMEMS’ global ocean eddy resolution reanalysis product GLO RYS12V1, with a horizontal resolution of 1 / 12° (approximately 8 km) and 50 standard depth layers in the vertical direction, covering the period from 1993 to the present.
[0129] The core model is based on the NEMO platform, and the surface-driven data comes from the ERA-Interim and ERA5 reanalysis data of the European Centre for Medium-Range Weather Forecasts (ECMWF).
[0130] The specific data products used include:
[0131] Monthly average data: cmems_mod_glo_phy_my_0.083deg_P1M-m (https: / / data.marine.copernicus.eu / product / GLOBAL_MULTIYEAR_PHY_001_030 / download?dataset=cmems_mod_glo_phy_my_0.083deg_P1M-m_202311)
[0132] Transition period monthly average data: cmems_mod_glo_phy_myint_0.083deg_P1M-m (https: / / data.marine.copernicus.eu / product / GLOBAL_MULTIYEAR_PHY_001_030 / download?dataset=cmems_mod_glo_phy_myint_0.083deg_P1M-m_202311)
[0133] Climate average data: cmems_mod_glo_phy_my_0.083deg-climatology_P1M-m(https: / / data.marine.copernicus.eu / product / GLOBAL_MULTIYEAR_PHY_001_030 / download?dataset=cmems_mod_glo_phy_my_0.083deg-climatology_P1M-m_202311)
[0134] The analysis period spanned from January 1993 to December 2023. Selected physical quantities included seawater salinity (so, unit: g / kg), seawater potential temperature (thetao, unit: °C), eastward seawater current velocity (uo, unit: m / s), and northward seawater current velocity (vo, unit: m / s). The study area was limited to the ocean region approximately 10 to 100 kilometers away from the tidal station, with a depth range of 0 to 700 meters.
[0135] (53) Model parameter settings
[0136] In this embodiment, the network modules are configured as follows:
[0137] Input time step L=12;
[0138] The hidden dimension of the geometric algebraic convolution gated recurrent unit (GAConvGRU) is set to 64, the kernel size is (3,3), and the number of layers is 1.
[0139] Feature Dimensions of Context Embedding Layer The dropout rate is set to 0.1.
[0140] In the dual-path convolutional enhancement structure, the kernel size of path one is 7, and the kernel size of path two is 11, both using 32 channels;
[0141] The GRU decoding layer has 128 hidden units, and the output layer is a one-dimensional linear projection.
[0142] (54) Training strategies
[0143] The training process uses mean squared error (MSE) as the loss function, and the optimizer is Adam with an initial learning rate of [missing value]. The training batch size is 16, and the maximum number of iterations is 100. The dataset is divided into training, validation, and testing sets in a 6:3:1 ratio. This sequential partitioning method is used to avoid time leakage and ensure the continuity of the time series data.
[0144] (55) Evaluation indicators
[0145] This invention selects mean absolute error (MAE) and root mean square error (RMSE) as modeling evaluation indicators.
[0146] in, This is the actual sea level variability value. These are the sea level variability values simulated by the model. It refers to the number of samples.
[0147]
[0148] These are real sea level variability values. These are the sea level variability values simulated by the model. It refers to the number of samples.
[0149] To verify the effectiveness of the TSGAConvGRU model proposed in this invention, it was compared with the following four typical time series prediction models:
[0150] GRU – Classical gated recurrent unit model; LSTM – Long Short-Term Memory network model; SOFTS – A state-of-the-art fusion model for multivariate time series; Transformer – A time series modeling method based on attention mechanism.
[0151] The experimental results on a certain sea level reconstruction task are shown in Table 1. The units for both RMSE and MAE are mm. The TSGAConvGRU model proposed in this invention is significantly better than traditional recursive models (GRU, LSTM), attention mechanism models (Transformer), and multivariate time series models (SOFTS). In particular, in a certain experimental task, the RMSE of TSGAConvGRU was reduced to 11.34 and the MAE was reduced to 9.32, which are 21.3% and 17.4% higher than the second-best model GRU, respectively. This shows that the method of this invention can more fully capture the multivariate geometric algebraic features and contextual information in sea level time series data, and has stronger prediction and reconstruction capabilities. Figure 1This is a reconstruction of sea level variability at a coastal station from 1993 to 2023 using the TSGAConvGRU model. The model's predicted sea level anomalies (in millimeters) are highly consistent with observed values over continuous time steps, indicating the overall reliability of the model. The model robustly reconstructs the tidal station data through multi-vector correlation based on geometric algebra.
[0152] Table 1
[0153] Model RMSE MAE SOFTS 17.44 13.72 GRU 14.41 11.28 LSTM 16.40 13.35 Transformer 19.51 15.22 TSGAConvGRU (Method of this invention) 11.34 9.32
Claims
1. A sea surface height variability reconstruction method based on geometric algebra feature learning, characterized in that, Comprising (1) In the geometric algebra space, the preprocessed marine stereodynamic data is uniformly represented based on the generalized primitive, forming a multi-vector code; (2) The multi-vector code input geometric algebra convolution gated recurrent unit, extracts spatio-temporal features, and obtains a spatio-temporal feature sequence; (3) Introducing a time marker matrix, after context reconstruction of the spatio-temporal feature sequence, inputting a double-path convolution feature enhancement module, extracting local and global time sequence features, and obtaining an enhanced feature sequence; (4) The enhanced feature sequence is input into a multi-layer gated recurrent unit, and the decoding sequence is input into a linear mapping layer to project to the target dimension, and output the sea level change simulation result sequence, completing the reconstruction of the sea level change.
2. The method of claim 1, wherein, In step (1), the marine stereodynamic data preprocessing is as follows: Ocean three-dimensional dynamic data sequence The shape is defined as [T, C*D*H*W], where T represents the total selected time length, C represents the number of ocean variables, D represents the number of depth layers, and H and W represent the spatial grid dimensions, respectively. Calculate the abnormal data, subtract the monthly average data from the climate average data to obtain the abnormal value sequence of each variable; Calculate the proportion of valid data at each depth layer, if the proportion of valid data is less than 80%, the depth layer is removed; For sporadic missing values in the depth layer, the average value of the layer is used for filling; Normalization; Obtain the preprocessed marine stereodynamic data sequence.
3. The method of claim 2, wherein, The marine variables include potential temperature, salinity, eastward sea water velocity and northward sea water velocity.
4. The method of claim 3, wherein, In step (1), the uniform representation is as follows: The variable channel number and spatial dimension of the pretreated marine stereodynamic data sequence are reserved, the time sequence data is segmented by batches, the total time length is divided into multiple subsequences to obtain a five-dimensional tensor; Encoding a five-dimensional tensor as a multivector in a geometric algebra space : ； wherein, is a multi-vector encoding of a time step t, is salinity, is temperature, is an eastward sea water velocity, is a northward sea water velocity; is a scalar basis element, and is a vector basis element, is a bi-vector basis.
5. The method of claim 1, wherein, In step (2), the geometric algebra convolution gated recurrent unit operator satisfies: ； wherein, denotes the output of a geometric algebra convolution gated recurrent unit, denotes the operator of a geometric algebra convolution gated recurrent unit, denotes the input of a multi-vector at a time step t, the operator of a geometric algebra convolution gated recurrent unit comprises a geometric algebra convolution operation and a gating mechanism.
6. The method of claim 5, wherein, In step (2), the gating mechanism is as follows: ； wherein, denotes an update gate, denotes a reset gate, denotes a candidate hidden state, denotes a final hidden state, denotes a multi-vector input at a time step t, denotes a hidden state at time step t-1, is a Sigmoid activation function, is a hyperbolic tangent activation function, is a geometric algebra operation for different convolution kernels, denotes a Hadamard product.
7. The method of claim 1, wherein, Step (3) is as follows: The spatio-temporal feature sequence and the time marker matrix are fused in the feature dimension through the concatenation operation to form the embedded time sequence context representation, and the embedding dimension is the expansion of the original feature dimension; The reconstructed sequence is input into a double-path convolution feature enhancement module to extract local and global time sequence features, wherein path 1 is used to capture global dynamics, and path 2 is used to capture local dynamics, the output of path 1 is processed through residual, and then added to the output of path 2, and the fusion obtains an enhanced feature sequence.
8. The method of claim 7, wherein, The function expression of path 1 is ； wherein, denotes an input tensor, denotes an output tensor, is a Sigmoid activation function, MaxPool(·) denotes a max-pooling operation, and Conv1D(·) denotes a one-dimensional convolution operation.
9. The method of claim 8, wherein, The function expression of path 2 is ； wherein, denotes an input tensor, denotes an output tensor, AvgPool(·) denotes an average pooling operation.
10. The method of claim 9, wherein, Path 1 introduces a residual mapping operator after multi-layer convolutional pooling : ； wherein, represents a mapping output, represents a path 1 output tensor; The final path fusion output is ； wherein, represents the output after fusion, represents the path 2 output tensor, represents the path 1 mapped output tensor, and C is the channel after convolution.