Multi-factor power transmission channel surface deformation prediction method based on InSAR technology and deep learning
By combining InSAR technology and deep learning, using multi-factor data and improved Transformer model, the problems of time series feature extraction and nonlinear relationship capture in surface deformation prediction of transmission channels are solved, achieving more accurate and reliable prediction results, and improving Capability to assess settlement trends and infrastructure impacts.
Patent Information
- Application Number
- CN202410997976.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-24
- Publication Date
- 2024-11-15
AI Technical Summary
The existing surface deformation prediction methods for power transmission channels have poor ability to extract time series deformation characteristics, and it is difficult to capture the complex nonlinear relationship between time step and time series deformation, resulting in insufficient prediction accuracy.
Using multi-factor prediction method based on InSAR technology and deep learning, combining daily average surface temperature, precipitation, vegetation coverage, DEM, slope and lithology data, comprehensive modeling is used to improve the deformation characteristics of time series The accuracy of extraction capability and nonlinear relationship.
It significantly improves the accuracy and reliability of the prediction results, enhances the generalization ability of the model, can predict long-term series deformation more accurately, provides more comprehensive influencing factor considerations and long-term prediction capabilities, and timely evaluates the impact of settlement on infrastructure. .
Smart Images

Figure CN118968292A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of radar remote sensing and image processing, and in particular to a multi-factor transmission channel surface deformation prediction method based on InSAR technology and deep learning. Background Art
[0002] Predicting surface deformation of power transmission channels is of great significance, as it can prevent and control power transmission line disasters, as well as focus on monitoring and maintaining transmission towers that may be affected, to ensure the safe operation of power transmission lines. Deep learning simulates the neural network structure of the human brain to learn and understand data, and uses appropriate neural network structures and large-scale annotated data to automatically detect and locate targets in complex scenarios.
[0003] The existing methods for predicting surface deformation of power transmission channels include the following prediction methods:
[0004] (1) The SVM, LSTM, Elman, and Elman-LSTM models are used to predict the total cumulative displacement. The cumulative total displacement is divided into trend items and fluctuation items according to the principle of time series analysis. Then, the four displacement prediction models are used to predict the displacement values of each item respectively. Finally, the predicted displacement value of the trend item and the predicted displacement value of the fluctuation item are added together to obtain the predicted cumulative total displacement value.
[0005] Deformation is a dynamic process. SVM is a static prediction model that only learns and uses information at the current moment and cannot use historical information. Although LSTM and Elman are dynamic models and perform better than SVM, they also have some disadvantages. The high computational complexity of LSTM makes its internal operating mechanism less intuitive and difficult to explain the decision-making process of the network. It also requires more data for training to avoid overfitting. The Elman model has a strong learning ability, is prone to overfitting training data, and has a slow convergence speed. For large-scale data sets, it will increase the training time of the model. In the experiment, although the Elman-LSTM model has the best overall prediction effect, it is poor at predicting the time series of fluctuation items and trend items.
[0006] (2) Multilayer perceptron neural network (MLPNN), logistic regression (LR), and Bayesian network (BN) are used to predict deformation using eight evaluation factors: slope, profile curvature, plane curvature, stratum lithology, distance from the river, terrain moisture index, distance from the road, and land use type. The MLPNN model has many parameters and is prone to overfitting on the training set. Its performance depends to a large extent on the selection of initial weights, and different initial values may lead to different results. The LR model does not perform well when data features are missing or the feature space is large. The BN network needs to collect a large amount of data in practical applications. Small data will lead to poor results, and in some cases the computational complexity is high.
[0007] (3) The predicted deformation information provided by the deep learning prediction model helps improve the monitoring capability of geological disasters on transmission lines. Since SAR images are not affected by weather and sunlight conditions, have a wide coverage area, and can monitor ground deformation with millimeter-level accuracy, InSAR technology based on SAR data has also been increasingly used in transmission line deformation monitoring in recent years. This has also created a good foundation for long-term series prediction of transmission lines.
[0008] However, the existing traditional prediction models based on InSAR technology rarely combine the environmental factors of the transmission channel, and the training data only uses a single InSAR data. As a result, the ability to extract the InSAR time series deformation features of the transmission channel is poor. In addition, the existing prediction models are not easy to capture the complex nonlinear relationship between the time step and the time series deformation. Therefore, long time series deformation cannot be accurately predicted. The above problems need to be solved urgently. Summary of the invention
[0009] The purpose of the present invention is to solve the problem that the traditional surface deformation prediction method of power transmission channel has poor ability to extract time series deformation features and is not easy to capture the complex nonlinear relationship between time step and time series deformation, resulting in poor prediction accuracy. The present invention provides a multi-factor power transmission channel surface deformation prediction method based on InSAR technology and deep learning.
[0010] A multi-factor transmission channel surface deformation prediction method based on InSAR technology and deep learning includes the following steps:
[0011] Collect the transmission channels of each area to be monitored within a preset period of time, and collect a SAR image containing the transmission towers on each sampling date, as well as the daily average surface temperature data, daily average precipitation data, vegetation coverage data, DEM data, slope data and lithology data; after preprocessing all the collected SAR images, use the small baseline set method in the time series InSAR technology to process each preprocessed SAR image on each sampling date to obtain the surface deformation data on each sampling date;
[0012] The daily average surface temperature data, daily average precipitation data, vegetation coverage data, DEM data, slope data and lithology data corresponding to the 1st to nth sampling dates under the transmission channel of the same monitored area are used as model input data, and the surface deformation data under the transmission channel of the same monitored area under the n+1th to n+mth sampling dates are used as labels to train the deformation prediction model to obtain a trained deformation prediction model; m and n are integers; the 1st to n+mth sampling dates are continuous in time;
[0013] The average daily surface temperature data, average daily precipitation data, vegetation coverage data, DEM data, slope data and lithology data under the current n consecutive sampling dates are input into the trained deformation prediction model for prediction, and the surface deformation data under the next m dates are obtained to complete the surface deformation prediction.
[0014] Preferably, for each power transmission channel in the area to be monitored, a SAR image containing a power transmission tower collected on each sampling date is a SAR image containing a power transmission tower at any time within the sampling date;
[0015] The vegetation coverage data, DEM data, slope data and lithology data on each sampling date are the vegetation coverage data, DEM data, slope data and lithology data at any moment on the sampling date.
[0016] Preferably, the implementation method of preprocessing all SAR images collected under the power transmission channel of the same monitored area includes:
[0017] All SAR images collected under the power transmission channel of the same monitored area are cropped until the monitoring range is the same in all SAR images, and the cropped SAR images are denoised;
[0018] All denoised SAR images under the transmission channel of the same monitored area are registered to complete the preprocessing of the collected SAR images.
[0019] Preferably, the polarization mode of the collected SAR image is VV polarization.
[0020] Preferably, the implementation method of registering all denoised SAR images under the power transmission channel of the same monitored area includes:
[0021] Select any denoised SAR image under the transmission channel of the same area to be monitored as the main image, and select the remaining denoised SAR images under the transmission channel of the same area to be monitored as the slave images;
[0022] The feature points in the slave image are registered to the master image space, specifically: a geometric transformation method is used to make the spatial positions of the slave image consistent with those of the master image.
[0023] Preferably, the geometric transformation method includes translation, rotation and scaling.
[0024] Preferably, the deformation prediction model is implemented using a Transformer model improved based on KAN, specifically:
[0025] The feedforward neural network layers in the encoder and decoder layers of the original Transformer model are replaced with KAN layers to form a KAN-based improved Transformer model.
[0026] Among them, the KAN layer is implemented using a learnable spline basis function.
[0027] Advantages of the present invention:
[0028] The multi-factor transmission channel surface deformation prediction method based on InSAR technology and deep learning proposed in the present invention makes full use of multi-source data such as surface deformation data, daily average precipitation data, daily average surface temperature data, vegetation coverage data, DEM, slope and lithology for comprehensive modeling in the model design stage, and comprehensively considers various factors affecting the surface deformation of the transmission channel. In the model training stage, the model is trained using time-related data collected at preset time intervals within a preset period of time, which improves the accuracy of predicting the complex nonlinear relationship between the time step and the time series deformation. At the same time, the daily average surface temperature data on the sampling date is introduced during training, and the impact of environmental factors on the surface deformation data is comprehensively considered to improve the ability to extract deformation features of the time series.
[0029] In particular, the improved Transformer model based on KAN is used. Compared with traditional methods, this model can adapt to complex data backgrounds more flexibly, accurately capture and model the delicate nonlinear relationship between multiple factors and deformation time series data, as well as the time series deformation data itself, and enhance the generalization ability of the model, thereby effectively improving the accuracy and reliability of the prediction results. In addition, by reducing the number of model parameters, the training and prediction efficiency is significantly improved.
[0030] Compared with traditional transmission channel surface deformation prediction methods, the present invention can obtain deformation prediction results of a longer time series, thereby more comprehensively studying the regional deformation trend related to subsidence, timely evaluating the impact of subsidence on infrastructure, and taking protective measures in advance for transmission towers with high subsidence risks.
[0031] The method proposed in the present invention significantly improves the accuracy, generalization ability and efficiency of the prediction of surface deformation of transmission channels. Compared with traditional transmission channel surface deformation prediction methods, it has more comprehensive consideration of influencing factors and long-term prediction capabilities, enabling researchers to more comprehensively analyze and evaluate deformation trends related to regional subsidence. By introducing multi-source data modeling and the improved Transformer model based on advanced KAN, this method can not only timely evaluate the potential impact of subsidence on infrastructure, but also formulate effective protective measures in advance for transmission towers with higher subsidence risks. This comprehensive method not only improves the ability to extract time series deformation features, but also provides a more accurate and operational prediction tool for transmission channel management and maintenance, and has significant application potential and social value. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 This is a schematic diagram of the principle of training the deformation prediction model;
[0033] Figure 2 This is a structural diagram of the Transformer model improved based on KAN. DETAILED DESCRIPTION
[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0035] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0036] Specific implementation method 1: The following is combined Figure 1 This embodiment describes a method for predicting surface deformation of a multi-factor power transmission channel based on InSAR technology and deep learning, which comprises the following steps:
[0037] Step 1: Collect the transmission channels of each area to be monitored within a preset period of time, and a SAR image containing the transmission towers on each sampling date, as well as the daily average surface temperature data, daily average precipitation data, vegetation coverage data, DEM data, slope data and lithology data; after preprocessing all the collected SAR images, use the small baseline set method in the time series InSAR technology to process each preprocessed SAR image on each sampling date to obtain the surface deformation data on each sampling date;
[0038] Specifically, surface deformation includes surface uplift and surface subsidence;
[0039] In the model design stage, we make full use of multi-source data such as surface deformation data, daily total precipitation, daily average surface temperature, vegetation cover index, DEM, slope and lithology for comprehensive modeling, and comprehensively consider various factors affecting the surface deformation of the transmission channel. In addition, in the model training stage, we use time-related data collected at preset time intervals within a preset period to train the model, which improves the accuracy of the complex nonlinear relationship between the prediction time step and the time series deformation; at the same time, we introduce the daily average surface temperature data on the sampling date during training, comprehensively consider the impact of environmental factors on surface deformation data, and improve the ability to extract deformation features of time series.
[0040] Step 2: Model training:
[0041] The average daily surface temperature data, average daily precipitation data, vegetation coverage data, DEM data, slope data and lithology data corresponding to the 1st to nth sampling dates under the transmission channel of the same monitored area are used as model input data, and the surface deformation data under the transmission channel of the same monitored area under the n+1th to n+mth sampling dates are used as labels to train the deformation prediction model to obtain the trained deformation prediction model; m and n are integers; the 1st to n+mth sampling dates are continuous in time; step three, the average daily surface temperature data, average daily precipitation data, vegetation coverage data, DEM data, slope data and lithology data under the current n consecutive sampling dates are input into the trained deformation prediction model for prediction, and the surface deformation data under the next m dates are obtained to complete the surface deformation prediction.
[0042] Specifically, for each transmission channel in the area to be monitored, a SAR image containing a transmission tower collected on each sampling date is a SAR image containing a transmission tower at any time within the sampling date;
[0043] The vegetation coverage data, DEM data, slope data and lithology data on each sampling date are the vegetation coverage data, DEM data, slope data and lithology data at any moment on the sampling date.
[0044] Furthermore, to ensure data quality and consistency, all SAR images collected under the transmission channel of the same monitored area need to be preprocessed. The specific implementation methods include:
[0045] All SAR images collected under the power transmission channel of the same area to be monitored are cropped until the monitoring range in all SAR images is the same, and the cropped SAR images are denoised; as a preferred method, the polarization mode of the collected SAR images is VV polarization;
[0046] All denoised SAR images under the transmission channel of the same monitored area are registered to complete the preprocessing of the collected SAR images.
[0047] Furthermore, the registration of all denoised SAR images under the transmission channel of the same monitored area is achieved by:
[0048] Select any denoised SAR image under the transmission channel of the same area to be monitored as the main image, and select the remaining denoised SAR images under the transmission channel of the same area to be monitored as the slave images;
[0049] The feature points in the slave image are registered to the master image space, specifically, a geometric transformation method is used to make the spatial position of the slave image consistent with that of the master image. In order to ensure that the slave image and the master image are aligned in the same coordinate system, the geometric transformation method includes translation, rotation and scaling.
[0050] Furthermore, see 2, the deformation prediction model is implemented using the Transformer model improved based on KAN, specifically:
[0051] The feedforward neural network layers in the encoder and decoder layers of the original Transformer model are replaced with KAN layers to form a KAN-based improved Transformer model.
[0052] Among them, the KAN layer is implemented using a learnable spline basis function.
[0053] In this preferred embodiment, the Transformer model is improved based on Kolmogorov-Arnold Networks (KAN), and the learnable spline basis functions of KAN are used to activate and enhance the Transformer time series prediction modeling. The design of KAN is based on the Kolmogorov-Arnold representation theorem, which provides a method to represent any multivariate continuous function as a combination of univariate functions and addition. Inspired by the Kolmogorov-Arnold representation theorem, KAN replaces traditional linear weights with spline-parameterized univariate functions, enabling it to dynamically learn activation patterns. Specifically, in the encoder and decoder layers of the Transformer, the multilayer perceptron (MLP) in the traditional feedforward neural network is replaced by the learnable activation functions in the KAN architecture. In specific implementation, the MLP of each layer can be replaced with the layer structure of KAN, which contains a set of learnable univariate spline functions.
[0054] The advantages of replacing MLP with KAN's layer structure are as follows:
[0055] 1. The MLP layer (MLP) in the traditional Transformer uses a linear weight matrix for linear transformation, and then introduces nonlinearity through a nonlinear activation function (such as ReLU). It usually has a large number of parameters and is prone to overfitting, especially when processing large-scale data. KAN uses learnable splines instead of linear weight matrices. These splines are constructed through basis functions (such as B-splines). By introducing KAN's learnable splines, nonlinear transformation can be achieved with fewer parameters, reducing the number of parameters while ensuring or improving performance, thereby reducing the risk of overfitting and improving the model's expressiveness and generalization capabilities.
[0056] 2. Although the traditional Transformer architecture has advantages in capturing long-distance dependencies, it may not perform well when dealing with highly nonlinear tasks. KAN can more flexibly adapt to the complexity of data through convolutional layers based on learnable splines, capture and model more delicate nonlinear relationships, and thus enhance the performance of Transformer when dealing with complex tasks.
[0057] 3. Compared with MLP, the spline basis function of KAN can be adjusted according to task requirements, making the model more flexible and efficient.
[0058] The parameter efficiency of KAN means that more efficient training and inference can be achieved with the same computing resources.
[0059] In general, the KAN structure layer significantly enhances the flexibility and expressiveness of the neural network by introducing a learnable activation function. The present invention applies KAN to the feedforward neural network part of the Transformer architecture, which can reduce the number of parameters, improve model performance, and enhance generalization ability. This improvement combines the advantages of the two network architectures and helps to build a more efficient and powerful time series prediction model.
[0060] The specific implementation process of the Transformer model improved based on KAN is as follows:
[0061] As shown in Figure 2, the improved Transformer model based on KAN includes an encoder and a decoder:
[0062] The time series deformation prediction model proposed based on the improved Transformer model consists of three layers of encoders and three layers of decoders. There is an input layer, a position encoding layer and three identical encoder layers in the encoder structure. First, the input time series deformation, daily total precipitation, daily average surface temperature, vegetation cover index, DEM, slope and lithology data are mapped to 7 feature vectors through a fully connected network; the position encoding of sine and cosine functions is used to encode the sequence information, and the input vector is added element by element to the position encoding vector. Subsequently, the query vector Q, key vector K and value vector V are obtained through linear mapping. The generated vector is as follows Figure 2 As shown in the encoder structure, the multi-head attention mechanism in the three encoder layers can project the query vector, key vector, and value vector into the subspace one by one, and calculate the correlation between each moment in parallel. After that, the output of the multi-head attention layer will be passed to the feedforward neural network, and the addition and normalization operations will be implemented through the learnable activation function in the KAN structure. After the multi-head attention mechanism and the feedforward neural network, the encoder layer maps the features of the input pixels into a hidden representation of a fixed length. Finally, the final output of the encoder is a D-dimensional vector, where D is a specific value (usually 512), which is determined by the design of the model and the task requirements. This vector contains the encoder's overall understanding and abstract representation of the input, which will be used as the input of the decoder to generate predictions or generation of target time data. The decoder consists of an input layer, three identical decoder layers, and an output layer. As shown Figure 2 As shown in Figure 1, in the third sublayer of the decoder layer structure, the encoder-decoder attention layer, the decoder output uses a self-attention mechanism, which allows the decoder to focus on different parts of the input sequence when generating the sequence to help generate the correct output. The output layer maps the output of the last decoder layer to the target time series data. In order to make the predicted value of the time series deformation based only on the previous data point, forward masking and a position offset are used.
[0063] Figure 2 In the example, if we want to train the model to predict the deformation of m dates in the future based on the factor data of n dates, the encoder input T1 to T n Represents the time series multivariate data of the past n dates, with the same dimension as the number of variables, which is 7; the decoder input T n+1 To T n+m The surface deformation monitoring results (deformation data) on the sampling date are used as labels.
[0064] The time series deformation data for the next m dates is the real value with a dimension of 1; the decoder outputs T n+1 To T n+m It represents the predicted value of the time series deformation data for the next m dates predicted by the model, with a dimension of 1. Based on this training method, the model can learn how to generate future data from historical data more quickly, reduce the accumulation of errors and improve the stability of training. In the actual prediction stage of the model, the multivariate data of the time series of n sampling dates in history is used as input to predict the deformation data for the next m dates.
[0065] In this preferred embodiment, a Transformer model based on KAN improvement is used. Compared with the traditional method, this model can adapt to complex data background more flexibly, accurately capture and model the delicate nonlinear relationship between multiple factors and deformation time series data (surface deformation monitoring results) and time series deformation data itself, and enhance the generalization ability of the model, thereby effectively improving the accuracy and reliability of the prediction results. In addition, by reducing the number of model parameters, the training and prediction efficiency is significantly improved.
[0066] In order to verify the prediction performance of the trained deformation prediction model, in specific applications, the prediction performance of the trained deformation prediction model is achieved by evaluating the performance indicators of the trained deformation prediction model, where the performance indicators are root mean square error, mean absolute error, mean absolute percentage error and residual indicators;
[0067] The surface deformation prediction results of the trained deformation prediction model are compared with the surface deformation monitoring results obtained based on synthetic aperture radar during the same period, and the error index between the surface deformation prediction results and the surface deformation monitoring results is calculated;
[0068] The root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and residual indicators are used to evaluate the prediction performance of the trained deformation prediction model.
[0069] The root mean square error (RMSE) ranges from [0, +∞). When the surface deformation prediction result (predicted value) is completely consistent with the surface deformation monitoring result (true value), it is equal to 0, that is, a perfect model. The larger the error, the larger the value.
[0070]
[0071] in, is the i-th predicted value, y i is the i-th true value;
[0072] The mean absolute error (MAE) ranges from [0, +∞). When the surface deformation prediction result (predicted value) is completely consistent with the surface deformation monitoring result (true value), it is equal to 0, that is, a perfect model. The larger the error, the larger the value.
[0073]
[0074] The mean absolute percentage error range of MAPE is [0, +∞). A mean absolute percentage error of 0% indicates a perfect model, while a mean absolute percentage error greater than 100% indicates a poor model.
[0075]
[0076] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.
Claims
1. A multi-factor transmission channel surface deformation prediction method based on InSAR technology and deep learning, characterized in that: The method comprises the following steps: Collect the transmission channels of each area to be monitored within a preset period of time, and collect a SAR image containing the transmission towers on each sampling date, as well as the daily average surface temperature data, daily average precipitation data, vegetation coverage data, DEM data, slope data and lithology data; after preprocessing all the collected SAR images, use the small baseline set method in the time series InSAR technology to process each preprocessed SAR image on each sampling date to obtain the surface deformation data on each sampling date; The daily average surface temperature data, daily average precipitation data, vegetation coverage data, DEM data, slope data and lithology data corresponding to the 1st to nth sampling dates under the transmission channel of the same monitored area are used as model input data, and the surface deformation data under the transmission channel of the same monitored area under the n+1th to n+mth sampling dates are used as labels to train the deformation prediction model to obtain a trained deformation prediction model; m and n are integers; the 1st to n+mth sampling dates are continuous in time; The average daily surface temperature data, average daily precipitation data, vegetation coverage data, DEM data, slope data and lithology data under the current n consecutive sampling dates are input into the trained deformation prediction model for prediction, and the surface deformation data under the next m dates are obtained to complete the surface deformation prediction.
2. The multi-factor transmission channel surface deformation prediction method based on InSAR technology and deep learning according to claim 1 is characterized in that: For each transmission channel in the area to be monitored, a SAR image containing a transmission tower collected on each sampling date is a SAR image containing a transmission tower at any time within the sampling date; The vegetation coverage data, DEM data, slope data and lithology data on each sampling date are the vegetation coverage data, DEM data, slope data and lithology data at any moment on the sampling date.
3. The multi-factor transmission channel surface deformation prediction method based on InSAR technology and deep learning according to claim 1 is characterized in that: The implementation methods of preprocessing all SAR images collected under the power transmission channel of the same monitored area include: All SAR images collected under the power transmission channel of the same monitored area are cropped until the monitoring range is the same in all SAR images, and the cropped SAR images are denoised; All denoised SAR images under the transmission channel of the same monitored area are registered to complete the preprocessing of the collected SAR images.
4. The multi-factor transmission channel surface deformation prediction method based on InSAR technology and deep learning according to claim 1 is characterized in that: The polarization mode of the collected SAR images is VV polarization.
5. The method for predicting surface deformation of a multi-factor power transmission channel based on InSAR technology and deep learning according to claim 1 or 4, characterized in that: The implementation methods of registering all denoised SAR images under the transmission channel of the same monitored area include: Select any denoised SAR image under the transmission channel of the same area to be monitored as the main image, and select the remaining denoised SAR images under the transmission channel of the same area to be monitored as the slave images; The feature points in the slave image are registered to the master image space, specifically: a geometric transformation method is used to make the spatial positions of the slave image consistent with those of the master image.
6. The method for predicting surface deformation of a multi-factor power transmission channel based on InSAR technology and deep learning according to claim 5 is characterized in that: The geometric transformation methods include translation, rotation and scaling.
7. The method for predicting surface deformation of a multi-factor power transmission channel based on InSAR technology and deep learning according to claim 1 is characterized in that: The deformation prediction model is implemented using the Transformer model improved based on KAN, specifically: The feedforward neural network layers in the encoder and decoder layers of the original Transformer model are replaced with KAN layers to form a KAN-based improved Transformer model. Among them, the KAN layer is implemented using a learnable spline basis function.
Citation Information
Patent Citations
Earth surface deformation monitoring and predicting method
CN115423180A
Multi-environmental-factor ground deformation prediction system and method, electronic equipment and medium
CN118131152A
Cited By
Hyperspectral blind unmixing method
CN119810660A
OCT image choroidal neovascularization segmentation method and system
CN120997226A
Intelligent early warning method and system for scouring of offshore wind power foundation under guidance of physical knowledge
CN121583078A
Foundation pit deformation intelligent prediction and active control method
CN121786710A
