A mountain landslide monitoring and risk early warning method for power transmission key equipment

Abstract

The application discloses a power transmission key equipment-oriented mountain landslide monitoring and risk early warning method, relates to the technical field of intelligent geological disaster monitoring and multi-source time sequence data fusion analysis, and comprises the following steps: constructing a dynamic alignment weight matrix based on the multi-source heterogeneous data of the mountain area of the power transmission key equipment; weighting and reconstructing a low-frequency characteristic sequence by using the dynamic alignment weight matrix, and mapping the reconstructed characteristics to the time axis of high-frequency characteristics through a back projection operation to generate a multi-scale fusion characteristic matrix with semantic alignment; inputting the multi-scale fusion characteristic matrix into an Informer encoder to output a hidden state sequence; adopting a progressive decoding strategy to perform multi-step displacement prediction on the hidden state sequence to output a deformation evolution trend graph; and determining a comprehensive landslide risk index based on the deformation evolution trend graph and the spatiotemporal distribution characteristics of the multi-source static characteristics to complete landslide risk early warning. The application can improve the precision of power transmission key equipment mountain landslide monitoring and risk early warning in a complex environment.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring of geological disasters and multi-source time-series data fusion analysis technology, and in particular to a method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment. Background Technology

[0002] With the continuous improvement of power grid construction, a large number of key power transmission equipment are deployed in mountainous and hilly areas with complex terrain, forming a wide-area power corridor system. These areas have fragile geological conditions and are prone to landslides and collapses due to factors such as topography, soil and rock characteristics, heavy rainfall, and engineering disturbances. These landslides and collapses can easily cause foundation settlement, tower tilting, and structural damage to power equipment, leading to line tripping, large-scale power outages, and cascading power grid failures, seriously threatening the safe and stable operation of the power grid. Therefore, conducting long-term, high-precision automated monitoring and early warning of slopes along mountainous power corridors is a key technological requirement for power grid disaster prevention and mitigation.

[0003] Current power line landslide monitoring mainly relies on conventional methods such as manual line inspection, GNSS point monitoring, sensor deployment, and drone aerial photography. These methods can only achieve localized and discrete monitoring, and have problems such as limited monitoring range, high cost, and difficulty in continuous all-weather monitoring. They are not suitable for monitoring scenarios with long-distance, linear distribution of transmission lines and scattered potential hazard points. InSAR technology, with its advantages of all-weather, wide-area, and high-resolution monitoring, has been gradually applied to slope deformation monitoring. It can achieve non-contact, long-term observation of power corridors. However, due to interference from vegetation obstruction, atmospheric delay, and observation angle, single InSAR data has high noise and poor coherence, making it difficult to fully reflect the evolution of landslides.

[0004] Existing research often integrates multi-source data on topography, geology, meteorology, and deformation, combined with deep learning models for landslide prediction and assessment. However, multi-source data exhibits cross-scale differences in characteristics such as high-frequency deformation, cumulative meteorological data, and static geological data. Existing preprocessing methods, such as uniform interpolation and simple splicing, easily lead to distortion of low-frequency driving factors and loss of early subtle deformation signals. Furthermore, existing models do not distinguish between high- and low-frequency feature channels, making them susceptible to being dominated by high-frequency deformation data, hindering the characterization of the entire landslide evolution cycle. Moreover, risk assessments often employ static weighting methods, failing to dynamically adapt to the evolving state of the disaster, resulting in insufficient accuracy and interpretability.

[0005] In summary, existing technologies suffer from shortcomings such as poor cross-scale fusion of multi-source heterogeneous data, insufficient collaborative modeling capabilities for high and low frequency features, and lack of dynamic risk assessment mechanisms. These deficiencies make it difficult to meet the high-precision early warning requirements for power transmission equipment in mountainous areas. There is an urgent need to develop a landslide dynamic monitoring and early warning technology solution that integrates InSAR and deep learning. Summary of the Invention

[0006] The purpose of this application is to provide a method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment, which can improve the accuracy of landslide monitoring and risk warning in mountainous areas where key power transmission equipment is located.

[0007] To achieve the above objectives, this application provides the following solution: This application provides a method for monitoring and risk early warning of landslides in mountainous areas for key power transmission equipment, including: An input matrix was constructed based on multi-source heterogeneous data from key power transmission equipment in mountainous areas. The input matrix is ​​divided into high-frequency feature sequences and low-frequency feature sequences according to the differences in time scale. Based on the temporal distribution characteristics of high-frequency feature sequences and the envelope characteristics of low-frequency feature sequences, a dynamic alignment weight matrix is ​​constructed. The low-frequency feature sequence is reconstructed by weighting the dynamic alignment weight matrix, and the reconstructed features are mapped to the time axis of high-frequency features through back projection operation to generate a semantically aligned multi-scale fusion feature matrix. Train the Informer encoder and Informer decoder; The multi-scale fused feature matrix is ​​input into the Informer encoder, which outputs a hidden state sequence. The Informer decoder is invoked, and a progressive decoding strategy is used to perform multi-step displacement prediction on the hidden state sequence, outputting a deformation evolution trend map. Based on the deformation evolution trend map and the spatiotemporal distribution characteristics of multi-source static features, a comprehensive landslide risk index is determined to complete the landslide risk early warning.

[0008] Optionally, the multi-source heterogeneous data includes: high-frequency dynamic data, static data, and low-frequency dynamic data; The high-frequency dynamic data is: the original time-series data of the InSAR deformation monitoring points; The static data is: elevation data; Optionally, an input matrix is ​​constructed based on multi-source heterogeneous data from mountainous areas containing key power transmission equipment, specifically including: The high-frequency dynamic data is processed using a first preprocessing method to obtain a land surface deformation observation sequence; the first preprocessing method includes, in sequence: neighborhood time series mean interpolation method, z-score anomaly detection algorithm, local quadratic polynomial regression smoothing method, and wavelet denoising algorithm. Static data and low-frequency dynamic data are standardized and normalized; the standardization and normalization process includes, in turn: Z-score standardization method, min-max normalization method, robust standardization method and mean-variance consistency test method. Based on the acquisition frequency and update cycle of the surface deformation observation sequence, a fixed time window sampling mechanism is constructed to perform sliding window slicing on the surface deformation observation sequence, generating high-density time embedding vectors for high-frequency channels; By using piecewise constant interpolation and consistency verification methods, low-frequency dynamic data are expanded into a high-granularity dataset aligned with the land surface deformation observation sequence; The trend envelope feature vector of the low-frequency channel is extracted from a high-granularity dataset using a moving average filter. The high-density time embedding vector of the high-frequency channel and the trend envelope feature vector of the low-frequency channel are concatenated along the time axis using a time axis index matching method to generate an input matrix. The time axis index matching method includes, in sequence: channel concatenation algorithm, missing value compensation interpolation algorithm, feature standardization processing and matrix dimension verification algorithm.

[0009] Optionally, the input matrix can be divided into high-frequency feature sequences and low-frequency feature sequences according to differences in time scale, specifically including: Based on the time-frequency characteristics, the input matrix is ​​separated in the frequency domain to obtain high-frequency dynamic feature sequences and low-frequency static feature sequences. Based on the sampling frequency and time resolution of high-frequency dynamic data, a sliding window local sequence extraction is performed on the high-frequency dynamic feature sequence to obtain high-frequency local fluctuation mode features. Piecewise constant interpolation is performed on the low-frequency static feature sequence to obtain the extended low-frequency feature sequence; The moving average filtering algorithm is used to extract low-frequency trend envelope features from the expanded low-frequency feature sequence. Channel segmentation is performed on the high-frequency local fluctuation pattern features and the low-frequency trend envelope features to generate high-frequency feature sequences and low-frequency feature sequences; both the high-frequency feature sequences and the low-frequency feature sequences are structured data.

[0010] Based on the temporal distribution characteristics of high-frequency feature sequences and the envelope characteristics of low-frequency feature sequences, a dynamic alignment weight matrix is ​​constructed, specifically including: Based on the temporal distribution characteristics of high-frequency feature sequences, a set of temporal convolution kernels with shared weights is used to perform convolution processing on high-frequency local fluctuation pattern features to extract the temporal dynamic representation of high-frequency features. Perform a sliding window slicing operation on the low-frequency trend envelope of the low-frequency feature sequence to generate a low-frequency local context fragment aligned with the time window of the high-frequency features; By using a multi-head attention mechanism, the temporal dynamic representation of high-frequency features is matched with low-frequency local context fragments in terms of semantics and temporal position to obtain a relevance score that reflects cross-scale correlation. Based on the relevance score, a dynamic alignment weight matrix is ​​constructed using a second preprocessing method; the second preprocessing method includes, in sequence: a normalization algorithm, a weight smoothing algorithm, an importance recalibration algorithm, a matrix normalization method, and an additive attention weight tuning algorithm.

[0011] The low-frequency feature sequence is reconstructed using a dynamic alignment weight matrix, and the reconstructed features are mapped to the time axis of high-frequency features through back projection, generating a semantically aligned multi-scale fusion feature matrix, specifically including: Based on the dynamic alignment weight matrix, a weighted mapping operation is performed on the feature vector of each time step in the low-frequency feature sequence to obtain the preliminary aligned representation of the low-frequency features on the high-frequency time axis. Based on the initial alignment representation, a differentiable back projection transformation function is used to map the weighted low-frequency features from the local window back to the original high-frequency time axis, thereby constructing a reconstructed low-frequency feature sequence aligned with the high-frequency features in the time dimension. Based on the local fluctuation pattern of high-frequency feature sequences, the attention weight distribution corresponding to each time step is calculated. Based on the attention weight distribution, the temporal alignment accuracy of reconstructed low-frequency features is dynamically adjusted; The reconstructed low-frequency features and high-frequency features after dynamic adjustment are spliced ​​together in the channel dimension to obtain the initial multi-scale fusion feature matrix; Normalization is performed on the initial multi-scale fusion feature matrix to obtain a semantically aligned and time-synchronized multi-scale fusion feature matrix.

[0012] Training the Informer encoder and Informer decoder specifically includes: Based on the gradient distribution characteristics of the multi-scale fused feature matrix, the original gradient magnitudes of the high-frequency and low-frequency feature channels during backpropagation are obtained using an automatic differentiation mechanism. Calculate the gradient energy density of each channel based on the original gradient magnitudes of the high-frequency and low-frequency characteristic channels. A dynamic normalization coefficient generation function is constructed based on the statistical results of gradient energy density distribution. The gradient magnitudes of the high-frequency and low-frequency feature channels are normalized using a dynamic normalization coefficient generation function to generate normalized gradient vectors. The parameter update operation is performed based on the normalized gradient vector, and the updated model parameters are fed back to the Informer encoder and Informer decoder.

[0013] The principle of the Informer encoder is as follows: Perform a linear mapping operation on the multi-scale fused feature matrix to generate a query vector dominated by high-frequency features; The key vector and value vector are generated by performing key vector and value vector generation operations on the weighted and reconstructed multi-scale fusion features. Multi-head self-attention mechanism computation is performed based on query vector, key vector and value vector: using the correlation scores between multi-scale features, the model is guided to identify co-evolving feature interaction patterns at different time scales, and output attention-weighted latent state feature matrix. Perform positional encoding enhancement on the hidden state feature matrix; The hidden state feature matrix after the position encoding enhancement operation is processed by a feedforward neural network to extract high-order nonlinear feature representations as the output hidden state sequence.

[0014] The Informer decoder is invoked, and a progressive decoding strategy is employed to perform multi-step displacement prediction on the hidden state sequence, outputting a deformation evolution trend map, specifically including: The Informer decoder is invoked, and based on the hidden state sequence, an attention mechanism is used to calculate the temporal correlation weight between the current hidden state and the historical multi-scale fusion features, generating a context-aware decoding initial state vector. The initial state vector is processed by a multi-layer gated recurrent unit to generate the latent variable output for the current prediction time step; each gated recurrent unit in the multi-layer gated recurrent unit introduces a time decay factor to adjust the retention ratio of historical information; Based on the latent variable output of the current prediction time step, the single-step displacement prediction value of the corresponding monitoring point is generated through mapping by the fully connected layer, and the single-step displacement prediction is fed back to the gated recurrent unit decoder as the input of the next time step to generate a progressive displacement sequence. The progressive displacement sequence is smoothed by a sliding window based on the prediction error covariance matrix to obtain the displacement prediction sequence. Based on the displacement prediction sequence, and combined with the spatial coordinates and geological attribute labels of each monitoring point, a deformation evolution trend map with spatiotemporal semantics is generated.

[0015] Based on the spatiotemporal distribution characteristics of deformation evolution trend maps and multi-source static features, a comprehensive landslide risk index is determined to complete landslide risk early warning, specifically including: A first-order difference operation was performed on the deformation evolution trend map to obtain the displacement acceleration time series characteristics of each monitoring point; Based on deformation evolution trend maps, geological lithology maps, and elevation models, a spatial distribution layer of geological vulnerability is generated; the geological vulnerability is obtained by weighted calculation based on lithology type, slope, and historical deformation rate. A meteorological triggering index is constructed using time-series data of meteorological factors. The meteorological triggering index is generated by performing a normalized weighted method based on daily rainfall, soil moisture content change rate and temperature anomaly values. A multi-dimensional landslide risk factor matrix was constructed based on displacement acceleration, geological vulnerability, and meteorological triggering index; Z-score standardization was performed on the multi-dimensional landslide risk factor matrix; A dynamic weight vector is determined using a dynamic weight allocation mechanism model. The dynamic weight allocation mechanism model is obtained by training a multilayer perceptron model based on a historical landslide event sample set. The dynamic weight allocation mechanism model is used to identify the relative contribution of different risk factors at different times and spatial locations, and to generate a dynamic weight vector that changes with time and region. The standardized multi-dimensional landslide risk factor matrix is ​​weighted and fused using dynamic weight vectors to generate a spatiotemporally continuous comprehensive landslide risk index. The landslide risk level corresponding to the interval of the comprehensive landslide risk index is determined as the current landslide risk level. Drive the early warning and emergency response mechanism corresponding to the current landslide risk level.

[0016] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method for monitoring and risk early warning of landslides in mountainous areas for key power transmission equipment. This method constructs high-frequency and low-frequency feature sequences based on multi-source heterogeneous data from mountainous areas containing key power transmission equipment. Based on the temporal distribution characteristics of the high-frequency feature sequences and the envelope characteristics of the low-frequency feature sequences, a dynamic alignment weight matrix is ​​constructed. The low-frequency feature sequences are weighted and reconstructed using the dynamic alignment weight matrix, and the reconstructed features are mapped to the time axis of the high-frequency features through back-projection, generating a semantically aligned multi-scale fusion feature matrix. An Informer encoder and Informer decoder are trained. The multi-scale fusion feature matrix is ​​input into the Informer encoder, which outputs a hidden state sequence. The Informer decoder is invoked, and a progressive decoding strategy is used to perform multi-step displacement prediction on the hidden state sequence, outputting a deformation evolution trend map. Based on the deformation evolution trend map and the spatiotemporal distribution characteristics of the multi-source static features, a comprehensive landslide risk index is determined, completing the landslide risk early warning. This significantly improves the accuracy of landslide monitoring in mountainous areas with power transmission corridor equipment under complex environments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a flowchart of a method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment, as described in one embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] In one exemplary embodiment, such as Figure 1 As shown, a method for monitoring and risk early warning of landslides in mountainous areas for key power transmission equipment is provided, including: Step 101: Construct an input matrix based on multi-source heterogeneous data from mountainous areas with key power transmission equipment. The multi-source heterogeneous data includes: high-frequency dynamic data, static data, and low-frequency dynamic data. High-frequency dynamic data consists of raw time-series data from InSAR deformation monitoring points. Static data consists of elevation data. Low-frequency dynamic data includes lithological data and meteorological data.

[0022] Step 101 specifically includes: The high-frequency dynamic data is processed using the first preprocessing method to obtain the surface deformation observation sequence. The first preprocessing method includes, in sequence: neighborhood time series mean interpolation method, z-score anomaly detection algorithm, local quadratic polynomial regression smoothing method, and wavelet denoising algorithm.

[0023] Static data and low-frequency dynamic data are standardized and normalized. The standardization and normalization processes include: Z-score standardization, min-max normalization, robust standardization, and mean-variance consistency test.

[0024] Based on the acquisition frequency and update cycle of the surface deformation observation sequence, a fixed time window sampling mechanism is constructed to perform sliding window slicing on the surface deformation observation sequence, generating high-density time embedding vectors for high-frequency channels.

[0025] By using piecewise constant interpolation and consistency verification methods, low-frequency dynamic data are expanded into a high-granularity dataset aligned with the land surface deformation observation sequence.

[0026] The trend envelope feature vector of the low-frequency channel is extracted from the high-granularity dataset using a moving average filter.

[0027] The input matrix is ​​generated by concatenating the high-density temporal embedding vector of the high-frequency channel with the trend envelope feature vector of the low-frequency channel along the time axis using a time-axis index matching method. The time-axis index matching method includes, in sequence: channel concatenation algorithm, missing value compensation interpolation algorithm, feature normalization processing, and matrix dimension verification algorithm.

[0028] Step 101 constructs the input matrix of the time-coding branch based on InSAR deformation time-series data and multi-source heterogeneous datasets of elevation, lithology, and meteorological factors. The high-frequency channel uses fixed-time-window sampling to generate a high-density time embedding, while the low-frequency channel is expanded to a uniform time granularity through piecewise constant interpolation and its trend envelope is extracted using a moving average filter. Specifically, this includes: S1.1: Perform missing value imputation and outlier removal on the raw time series data of InSAR deformation monitoring points to obtain a complete and stable surface deformation observation sequence.

[0029] For the raw time-series data of InSAR deformation monitoring points, an original data matrix is ​​established based on the observation sampling records and the output format of the monitoring equipment to ensure that the data covers all time steps and spatial monitoring point indices, so as to form a structured input that can be used for subsequent missing value analysis.

[0030] The neighborhood time series mean interpolation method (parameters: window length = 3, interpolation weight decreases with time distance) is used to perform interpolation operations on the missing values ​​detected in the original time series, so as to restore the time continuity and reduce the risk of time series breakage caused by data loss.

[0031] Furthermore, the z-score anomaly detection algorithm (parameters: threshold=3, mean and standard deviation are calculated based on the most recent 30 time steps) is used to identify anomaly observations that significantly deviate from the statistical distribution in the sequence and obtain an outlier index set.

[0032] Furthermore, a local quadratic polynomial regression smoothing method (parameters: fitting order = 2, smoothing window length = 5) is used to reconstruct and replace the window containing outliers to ensure that the deformation curve remains physically interpretable on a short-term scale and to generate a stable monitoring sequence.

[0033] Furthermore, a wavelet denoising algorithm (basis function: Daubechies-4, decomposition level = 3, threshold rule = soft threshold) is introduced to suppress noise in the filled and replaced time series data, extract stable surface deformation signal components, and output a stationary sequence matrix for subsequent modeling.

[0034] Through continuous processing using algorithms such as neighborhood interpolation, outlier detection and replacement, and wavelet denoising, the original InSAR time series data from the previous step is transformed into a complete and stable surface deformation observation sequence, achieving a dual guarantee of temporal continuity and data stability, and providing a reliable foundation for constructing high-frequency channel input for the time coding branch.

[0035] S1.2: Based on multi-source static and low-frequency dynamic data of elevation, lithology, and meteorological factors, standardization and normalization processing is performed to eliminate the interference of different physical dimensions on subsequent model training.

[0036] Based on multi-source static and low-frequency dynamic data of elevation, lithology, and meteorological factors, the Z-score standardization method (parameters: mean μ and standard deviation σ of each feature dimension) is used to achieve dimensionless transformation of data with different physical dimensions.

[0037] Furthermore, by using the min-max normalization method (parameters: the minimum value min and the maximum value max of a specific feature), the numerical values ​​of each data point are compressed to a uniform range of 0-1, and the feature matrix after interval mapping is obtained.

[0038] Furthermore, a robust standardization method (parameters: median of the feature and interquartile range (IQR)) is adopted to scale meteorological sequences sensitive to outliers and generate standardized results with enhanced noise resistance.

[0039] Furthermore, the mean-variance consistency test method (parameters: setting significance level α and test statistic F) is used to verify the variance balance of the data before and after standardization, and output the dataset that passes the test.

[0040] By combining Z-score, min-max normalization, and robust standardization, the multi-source static and low-frequency dynamic data from the previous step are transformed into unified input features that are balanced in numerical scale and distribution and can be directly entered into the time coding branch. This eliminates interference from different physical dimensions and improves the subsequent feature fusion effect.

[0041] For example, in an embodiment focusing on a transmission line monitoring area in the mountainous region of southwest China, the input multi-source static data includes elevation model values ​​ranging from 500 meters to 1500 meters, lithology code values ​​ranging from 1 to 5, and low-frequency meteorological dynamic data including daily average rainfall of 0 to 120 mm. When Z-score normalization is used, the mean elevation μ is 1025.4 meters, and the standard deviation σ is 150.2 meters. The formula is then applied. The dimensionless elevation values ​​were found to be distributed in the interval (-3.5, -3.2). Min-max normalization was applied to the lithological codes, with min=1 and max=5, using the formula... The normalized lithological eigenvalues ​​were obtained, ranging from 0 to 1. Meteorological precipitation was robustly standardized, with a median of 42.5 mm and an IQR of 30.2 mm, and the formula was applied. The rainfall scaling results were obtained to resist outliers. After the mean-variance consistency test, the difference in the coefficient of variation of the three types of data was controlled within 0.05, which significantly improved the fusionability of data from different sources in the same input matrix. The obtained standardized features effectively eliminated the differences in dimensions and enhanced the representation ability of low-frequency dynamic features in subsequent models.

[0042] S1.3: Based on the acquisition frequency and update cycle of InSAR data, a fixed time window sampling mechanism is constructed to perform sliding window slicing on the deformation time series data to generate high-density time embedding vectors.

[0043] S1.4: For low-frequency characteristic data such as lithological variation trends and daily rainfall, perform piecewise constant interpolation to expand its temporal granularity to a unified time axis aligned with high-frequency InSAR sequences.

[0044] For low-frequency dynamic characteristic sequences such as lithological variation trends and daily rainfall after normalization, a piecewise constant interpolation method (parameters: time base is high-frequency InSAR sampling period, interpolation strategy is zero-order preservation) is used to extend the time series of low-frequency data to achieve the unification of time granularity with high-frequency sequences.

[0045] Furthermore, based on the original set of low-frequency timestamps and the target set of high-frequency timestamps, a time mapping index matrix is ​​constructed (parameter: the mapping rule is the nearest time point matching), which realizes the static preservation interpolation of low-frequency feature values ​​at each high-frequency sampling point, and obtains an extended low-frequency feature set aligned with the high-frequency time axis.

[0046] Furthermore, a time window boundary constraint algorithm (parameter: window length equal to an integer multiple of the high-frequency sampling interval) is used to segment and label the expanded low-frequency features in different time periods to ensure that the interpolated data does not introduce spurious smoothing effects near the abrupt change points of geological events.

[0047] Furthermore, by using a value consistency verification method (parameter: the allowable error threshold is set to 5% of the original low-frequency observation mean square error), the deviation of abrupt change points caused by time axis mapping error in the interpolation results is detected and corrected.

[0048] By combining piecewise constant interpolation with consistency verification, the original low-frequency feature sequence is transformed into a high-granularity dataset that is fully aligned with the high-frequency InSAR time series, achieving the technical effect of cross-scale time series feature consistency construction.

[0049] For example, in a landslide monitoring task for a power transmission line in a mountainous area of ​​southwest China, the original sampling period for lithological change trend data was once per quarter, the sampling period for daily rainfall data was once per day, and the sampling period for high-frequency InSAR deformation data was once every 6 days. In this scenario, the interpolation strategy was set to zero-order preservation, and the target time axis was a high-frequency timestamp sequence with a step size of 6 days from the start date to the end date. By constructing a nearest-neighbor matching index matrix, the most recent observation value of the quarterly lithological data was copied to multiple corresponding high-frequency time points on the target time axis; for daily rainfall data, the system extended the most recent daily rainfall record to all high-frequency time points covered by that day. The interpolated lithological sequence was segmented at the boundary of a time window with a length of 12 days, and the interpolation of abrupt change points with deviations exceeding the original quarterly mean square error of 0.05 mm was corrected to the mean of adjacent segments using a consistency check method. Using the above processing, the extended time granularity of lithological and meteorological low-frequency features is perfectly matched with the InSAR data, providing high-fidelity and high-consistency multi-source time series data input for subsequent moving average filter extraction of trend envelope. This significantly improves the effect of multi-scale feature fusion and spatiotemporal characterization accuracy in model training and landslide risk prediction.

[0050] S1.5: Based on the interpolated low-frequency feature sequence, a moving average filter is applied to extract its trend envelope, so as to highlight the cumulative effect and changing trend of low-frequency factors in the time dimension.

[0051] S1.6: Align and concatenate the high-density temporal embedding vector of the high-frequency channel with the trend envelope feature vector of the low-frequency channel along the time axis to generate a multi-source heterogeneous feature input matrix, which serves as the structured input for the temporal coding branch.

[0052] The input conditions include the high-frequency channel high-density time embedding vector obtained by step S1.3 and the low-frequency channel trend envelope feature vector obtained by step S1.5. Both have been sampled, interpolated and filtered respectively and are in a coordinate system with a unified time axis.

[0053] The time axis index matching method (parameter: timestamp precision is on the hour level) is used to realize the hourly sample position correspondence between high frequency channel vectors and low frequency channel vectors in the time dimension.

[0054] Furthermore, through a channel splicing algorithm (parameters: splicing direction is the feature dimension, output dimension is high-frequency dimension + low-frequency dimension), feature fusion of high-frequency temporal embedding and low-frequency trend envelope at a unified sample location is achieved, and a preliminary multi-source heterogeneous input matrix is ​​obtained.

[0055] Furthermore, by using a missing value compensation interpolation algorithm (parameter: cubic spline interpolation), missing features generated during the synchronization process of the two channels are filled, ensuring the integrity of the spliced ​​matrix.

[0056] Furthermore, feature standardization (parameter: Z-score method) is used to achieve numerical scale uniformity of multi-channel features, as shown in the following formula: .

[0057] Where x' is the standardized feature value, x is the original feature value, μ is the mean of the full sample of the channel, and σ is the standard deviation of the full sample of the channel.

[0058] The matrix dimension verification algorithm (parameters: the output dimension target is T×(hf+lf), T is the unified time step, hf is the high-frequency feature dimension, and lf is the low-frequency feature dimension) transforms the result of the previous step into a structured matrix that meets the input requirements of the time coding branch, thereby achieving semantic and temporal alignment of multi-source time series before encoding.

[0059] For example, in the monitoring scenario of key power transmission equipment in mountainous power corridors, the high-frequency channel input is a 64-dimensional time embedding vector obtained by sliding window sampling of InSAR displacement data with a sampling frequency of 0.5 days. The low-frequency channel input is an 8-dimensional trend envelope vector obtained by piecewise constant interpolation and moving average of daily rainfall. During the time axis index matching process, the unified time axis length T is set to 720 steps. The high-frequency and low-frequency vectors are constructed according to hourly timestamps to establish the positional correspondence, and the initial output matrix dimension is 720×72. Cubic spline interpolation is used to compensate for approximately 3% of missing positions that occur during the synchronization process to ensure that the matrix is ​​free of gaps. Subsequently, Z-score standardization is performed. The sample mean of the high-frequency channel is 1.25 mm and the standard deviation is 0.15 mm, while the sample mean of the low-frequency channel is 12.6 mm and the standard deviation is 3.4 mm. After normalization, the feature scales of the two channels are consistent. The final output structured matrix has a dimension of 720×72, which meets the input requirements of the time-encoded branch, and shows a significant improvement in the temporal synchronization and semantic fusion of high-frequency and low-frequency information in the subsequent cross-scale dynamic alignment module.

[0060] Step 102: Divide the input matrix into high-frequency feature sequences and low-frequency feature sequences according to the time scale difference.

[0061] Step 102 specifically includes: The input matrix is ​​separated in the frequency domain based on its time-frequency characteristics to obtain high-frequency dynamic feature sequences and low-frequency static feature sequences.

[0062] Based on the sampling frequency and time resolution of high-frequency dynamic data, a sliding window local sequence extraction is performed on the high-frequency dynamic feature sequence to obtain high-frequency local fluctuation pattern features.

[0063] A piecewise constant interpolation operation is performed on the low-frequency static feature sequence to obtain the extended low-frequency feature sequence.

[0064] The moving average filtering algorithm is used to extract low-frequency trend envelope features from the expanded low-frequency feature sequence.

[0065] Channel segmentation is performed on the high-frequency local fluctuation pattern features and low-frequency trend envelope features to generate high-frequency feature sequences and low-frequency feature sequences. Both high-frequency and low-frequency feature sequences are structured data.

[0066] Step 102: For the input matrix of the time-encoding branch, divide it into high-frequency feature sequences and low-frequency feature sequences according to the time scale differences. The high-frequency feature sequences contain local fluctuation patterns of InSAR deformation data, while the low-frequency feature sequences characterize lithological change trends and the cumulative effects of meteorological factors, specifically including: S2.1: Perform frequency domain separation processing on the multidimensional feature vector output by the time-encoded branch based on the time-frequency characteristics to distinguish between high-frequency dynamic features and low-frequency static features.

[0067] S2.2: Based on the sampling frequency and time resolution of InSAR deformation data, a sliding window local sequence extraction is performed on the high-frequency channel features to generate a local fluctuation mode feature matrix with high time density.

[0068] S2.3: Perform piecewise constant interpolation on the time series of lithological factors and meteorological variables to expand their time granularity to be consistent with the high-frequency channel while preserving their low-frequency trend characteristics.

[0069] S2.4: Based on the extended low-frequency feature sequence, the trend envelope is extracted using a moving average filtering algorithm to remove short-term disturbances and enhance the characterization ability of its long-term evolution features.

[0070] Based on the low-frequency feature sequence processed by piecewise constant interpolation, a moving average filtering algorithm is adopted (parameters: window length W, weighting factor set {α}). i This achieves smooth suppression of short-term perturbations in the sequence. Furthermore, a weighted moving average formula is used to linearly combine the values ​​at each time step of the low-frequency sequence, as follows: .

[0071] in, Let be the value of the low-frequency feature at time ti. For the corresponding weighting factors, The window length is used, and weight normalization is employed to ensure the stability of the amplitude of the filtered output.

[0072] Furthermore, by using a recursive update method to reduce computational overhead, the current moving average is represented as a weighted sum of the average of the previous time step and the newly entered data, thus achieving low-latency trend extraction.

[0073] Furthermore, by using a center window approach to center the time window at the current sample, the statistical contributions of data before and after the prediction period are balanced, thereby improving the expected consistency of the trend envelope.

[0074] Furthermore, by using an adaptive window length adjustment function, the W value is dynamically expanded or shortened according to the rate of change of low-frequency factors, ensuring that trend extraction remains timely and accurate in geologically slow-changing environments.

[0075] By using a moving average filtering algorithm, the interpolated low-frequency feature sequence is transformed into smooth trend envelope data that retains only long-term evolution components, thereby achieving the expected technical effect of enhancing the ability to represent long-term changes in low-frequency features.

[0076] For example, in a key power transmission equipment area in the southwestern mountainous region, a moving average filter is applied to the daily rainfall sequence after piecewise constant interpolation, with a window length W = 15 days and a weighting factor of uniform distribution α. i =1 / W. Calculate the trend value for the 100th day using the above formula, where... Let be the rainfall on day (100-i) (in mm). Substituting each value into the formula and solving, the trend envelope value is 12.6 mm. Under different meteorological seasons, by adjusting the window length to 30 days or 7 days to address the changes between the dry and rainy seasons, the trend envelope curve shows a broad and gently rising shape during the rainy season, while maintaining low-amplitude stable fluctuations during the dry season. Validation results show that this trend envelope can significantly improve the explanatory power of low-frequency factors in long-term landslide risk prediction. The correlation coefficient between the filtered trend envelope and the long-term displacement changes of the landslide body is significantly higher than that of the unfiltered data, thus improving the overall temporal synergy of multi-source features.

[0077] S2.5: Channel segmentation is performed on the high-frequency local fluctuation pattern features and low-frequency trend envelope features to generate structured high-frequency feature sequences and low-frequency feature sequences, which serve as inputs to the cross-scale dynamic alignment module.

[0078] Step 103: Construct a dynamic alignment weight matrix based on the temporal distribution characteristics of high-frequency feature sequences and the envelope characteristics of low-frequency feature sequences.

[0079] Step 103 specifically includes: Based on the temporal distribution characteristics of high-frequency feature sequences, a set of temporal convolution kernels with shared weights is used to perform convolution processing on high-frequency local fluctuation pattern features to extract the temporal dynamic representation of high-frequency features.

[0080] Perform a sliding window slicing operation on the low-frequency trend envelope of the low-frequency feature sequence to generate a low-frequency local context fragment aligned with the high-frequency feature time window.

[0081] By utilizing a multi-head attention mechanism, the temporal dynamic representation of high-frequency features is matched with low-frequency local context fragments in terms of semantics and temporal location to obtain a relevance score that reflects cross-scale correlation.

[0082] Based on the relevance scores, a dynamic alignment weight matrix is ​​constructed using a second preprocessing method. This second preprocessing method includes, in sequence: a normalization algorithm, a weight smoothing algorithm, an importance recalibration algorithm, a matrix normalization method, and an additive attention weight tuning algorithm.

[0083] Step 103, based on the temporal distribution characteristics of high-frequency feature sequences and the envelope characteristics of low-frequency feature sequences, constructs a differentiable cross-scale alignment module. This module contains a group of temporal convolutional kernels with shared weights. It calculates the correlation score between the high-frequency sequence at each time point and the corresponding window of the low-frequency sequence through an attention mechanism, generating a dynamic alignment weight matrix. Specifically, this includes: S3.1: Based on the temporal distribution characteristics of high-frequency feature sequences, a set of temporal convolution kernels with shared weights is used to perform convolution processing on high-frequency local fluctuation patterns in order to extract the temporal dynamic representation of high-frequency features.

[0084] S3.2: Perform a sliding window slicing operation on the envelope features of the low-frequency feature sequence to generate low-frequency local context fragments aligned with the time window of the high-frequency features, which are then used for subsequent cross-scale correlation modeling.

[0085] S3.3: Based on the high-frequency feature convolution output and low-frequency local context fragments, a multi-head attention mechanism is constructed to calculate the correlation score between the high-frequency sequence at each time point and the corresponding window of the low-frequency sequence, so as to quantify the semantic matching degree between cross-scale features.

[0086] Based on the high-frequency feature convolution output and the low-frequency local context fragment input conditions, a multi-head attention mechanism (parameters: number of attention heads h, query dimension dq, key dimension dk, value dimension dv) is adopted to realize semantic matching modeling of cross-scale features.

[0087] The high-dimensional time-step embedding vector generated by high-frequency feature convolution output is projected using a linear mapping function to obtain a query vector group with multi-head attention mechanism, so as to preserve the temporal semantic representation of high-frequency local fluctuation patterns.

[0088] Furthermore, a linear mapping function is used through low-frequency local context fragments. and Generate key vector groups and value vector groups respectively. The key vectors are used to calculate the relevance with the query vector, and the value vectors are used for feature weighted fusion.

[0089] Employing the scaled dot product attention algorithm (parameter: scaling factor) The correlation matrix is ​​calculated using the following formula: .

[0090] Where S is the time step relevance score matrix, Q is the query vector group, K is the key vector group, and KT is the transpose matrix of the key vectors.

[0091] Furthermore, by applying the Softmax normalization function to each row of the correlation matrix S, an exponential normalization is obtained to obtain the probability distribution matrix A, and a proportional representation of the correlation at each time point is achieved: .

[0092] By performing matrix multiplication on the value vector group V and the attention distribution matrix A, a cross-scale feature weighted representation matrix O is generated, which simultaneously preserves high-frequency local dynamics and low-frequency trend semantics at the time granularity.

[0093] The weighted representation matrix O of each attention head output is concatenated column by column and restored to the original embedding dimension through linear transformation to form a comprehensive multi-head cross-scale correlation modeling output.

[0094] Through the multi-head attention mechanism described above, the convolutional features from the previous step are precisely matched with the low-frequency context in terms of semantics and temporal location, resulting in a relevance score matrix that reflects cross-scale correlation, thus achieving the mapping and alignment of features at different time scales.

[0095] For example, in a landslide risk prediction scenario for a power transmission line in a mountainous area of ​​Southwest China, the vector dimension dq of the high-frequency feature convolution output is set to 64, the dimension dv of the low-frequency context segment is set to 64, and the number of attention heads h is 4. The query vector Q is obtained by linearly mapping a high-frequency embedding sequence of InSAR deformation from five consecutive days. The key vector K and value vector V are generated by combining lithological envelope characteristics and meteorological trend characteristics within the same time window. , The scaling factor is generated by the formula. The calculated correlation matrix S, after Softmax normalization, yields an attention distribution matrix A that exhibits a significant peak at the time step with drastic deformation. The weighted output O obtained by multiplying the attention distribution matrix A by the value vector matrix V significantly improves the accuracy of capturing landslide evolution trends in subsequent time-based prediction models. Furthermore, in practical validation, it demonstrates a more robust multi-step displacement prediction effect, a narrower error range, and enhanced model sensitivity to extreme weather events.

[0096] S3.4: Based on the relevance score output by the multi-head attention mechanism, a dynamic alignment weight matrix is ​​generated. This weight matrix reflects the semantic importance distribution of low-frequency features on the high-frequency time axis, so as to achieve dynamic weighted mapping of cross-scale features.

[0097] Based on the cross-scale correlation score matrix output by the multi-head attention mechanism, a normalization processing algorithm (parameter: temperature coefficient τ of the Softmax function) is used to realize the probabilistic distribution mapping of the correlation score, so as to ensure that the sum of weights at each time position satisfies the probability constraint.

[0098] Furthermore, through a weighted smoothing algorithm (parameter: window width w) s The Gaussian smoothing kernel type is used to perform time-series smoothing on the normalized correlation scores and obtain a weight distribution result with continuous characteristics.

[0099] Furthermore, an importance recalibration algorithm is employed (parameter: high-frequency feature energy threshold E). th This method re-weights the smoothed weight distribution and generates a semantic importance coefficient matrix of low-frequency features on the high-frequency time axis to enhance the contribution of low-frequency features in highly relevant periods. A matrix normalization method (parameter: L2 norm constraint) is used to normalize each row of the importance coefficient matrix, ensuring that the weight magnitudes at different time steps are comparable on a uniform scale, thus forming a preliminary dynamically aligned weight matrix.

[0100] Furthermore, based on the additive attention weight tuning algorithm (parameter: tuning ratio α), the initial dynamic alignment weight matrix and the original attention score matrix are fused to generate the final dynamic alignment weight matrix. The temporal correlation structure of the original attention mechanism is preserved through weight tuning, while a cross-scale semantic matching enhancement effect is introduced.

[0101] Through the above-mentioned multi-level weight generation and optimization process, the correlation score result of the previous step is transformed into a differentiable dynamic alignment weight matrix, realizing the dynamic weighted mapping of low-frequency features on the high-frequency time axis, and improving the model's ability to synchronize time with heterogeneous features across scales.

[0102] For example, in the application of landslide risk prediction for key power transmission equipment in a mountainous area of ​​southwest China, a cross-scale correlation score matrix (size 120×720) was constructed for high-frequency InSAR deformation sequences (sampling interval 6 days, sequence length 120) and low-frequency daily rainfall sequences (sampling interval 1 day, sequence length 720). This matrix was output by a multi-head attention mechanism with 8 attention heads and a temperature coefficient τ of 0.8. The score matrix was Softmax normalized and smoothed using a Gaussian smoothing kernel (standard deviation σ=2.0) on the time axis to obtain a continuous weight distribution. The deformation energy value corresponding to each high-frequency time step was calculated, and importance recalibration was performed with Eth=2.5mm as the threshold. The weight coefficient of time steps above the threshold was increased by 30%, and the weight coefficient of time steps below the threshold was decreased by 10%. The recalibrated weight matrix was then L2 norm normalized to make the norm of each row of weight vectors 1. Finally, the optimized weight matrix and the original attention score matrix were fused together at a ratio of α=0.6 to generate a 120×720 dynamic alignment weight matrix. This matrix was used for subsequent low-frequency feature weighted fusion processing. Experimental results show that the model significantly improves the prediction accuracy of landslide evolution trends during high rainfall seasons and periods of high landslide incidence, with the RMSE value reduced to 0.42mm, demonstrating good cross-scale feature perception capabilities.

[0103] S3.5: The dynamic alignment weight matrix and the low-frequency feature sequence are weighted and fused to generate a semantically aligned low-frequency feature enhancement representation, which serves as the input basis for the subsequent construction of the multi-scale fusion feature matrix.

[0104] Step 104: Use the dynamic alignment weight matrix to reconstruct the low-frequency feature sequence with weights, and use the back projection operation to map the reconstructed features to the time axis of the high-frequency features to generate a semantically aligned multi-scale fusion feature matrix.

[0105] Step 104 specifically includes: Based on the dynamic alignment weight matrix, a weighted mapping operation is performed on the feature vector of each time step in the low-frequency feature sequence to obtain the preliminary aligned representation of the low-frequency features on the high-frequency time axis.

[0106] Based on the initial alignment representation, a differentiable back-projection transformation function is used to map the weighted low-frequency features from the local window back to the original high-frequency time axis, thus constructing a reconstructed low-frequency feature sequence aligned with the high-frequency features in the time dimension.

[0107] Based on the local fluctuation patterns of high-frequency feature sequences, the attention weight distribution corresponding to each time step is calculated.

[0108] Based on the attention weight distribution, the temporal alignment accuracy of reconstructed low-frequency features is dynamically adjusted.

[0109] The reconstructed low-frequency features and high-frequency features after dynamic adjustment are spliced ​​together in the channel dimension to obtain the initial multi-scale fusion feature matrix.

[0110] Normalization is performed on the initial multi-scale fusion feature matrix to obtain a semantically aligned and time-synchronized multi-scale fusion feature matrix.

[0111] Step 104 uses a dynamic alignment weight matrix to reconstruct the low-frequency feature sequence using weighted reassembly, and then maps the reconstructed features to the time axis of high-frequency features through back projection, generating a semantically aligned multi-scale fusion feature matrix. The weight distribution of low-frequency features is dynamically adjusted by the local fluctuation patterns of high-frequency features. Specifically, this includes: S4.1: Based on the dynamic alignment weight matrix output by the cross-scale alignment module, a weighted mapping operation is performed on the feature vector of each time step in the low-frequency feature sequence to obtain the preliminary alignment representation of the low-frequency features on the high-frequency time axis.

[0112] S4.2: A differentiable back-projection transformation function is used to map the weighted low-frequency features from the local window back to the original high-frequency time axis, so as to construct a reconstructed low-frequency feature sequence that is aligned with the high-frequency features in the time dimension.

[0113] S4.3: Based on the local fluctuation pattern of high-frequency feature sequences, calculate the attention weight distribution corresponding to each time step, and use this distribution to adjust the time alignment accuracy of reconstructing low-frequency features, so as to achieve dynamic control of the low-frequency feature weight distribution.

[0114] S4.4: The reconstructed low-frequency features after dynamic adjustment are spliced ​​together with the high-frequency features in the channel dimension to form a multi-scale fusion feature matrix in a unified dimension, so as to preserve the co-evolutionary relationship between high-frequency and low-frequency information.

[0115] S4.5: Perform normalization processing on the multi-scale fused feature matrix to eliminate scale differences between different feature channels, and output a semantically aligned and time-synchronized multi-scale fused feature matrix for use by the subsequent Informer encoder.

[0116] Train the Informer encoder and Informer decoder.

[0117] Based on the gradient distribution characteristics of the multi-scale fused feature matrix, the original gradient magnitudes of the high-frequency and low-frequency feature channels during backpropagation are obtained using an automatic differentiation mechanism.

[0118] The gradient energy density of each channel is calculated based on the original gradient magnitudes of the high-frequency and low-frequency characteristic channels.

[0119] A dynamic normalized coefficient generation function is constructed based on the statistical results of gradient energy density distribution.

[0120] The gradient magnitudes of the high-frequency and low-frequency feature channels are normalized using a dynamic normalization coefficient generation function to generate normalized gradient vectors.

[0121] The parameter update operation is performed based on the normalized gradient vector, and the updated model parameters are fed back to the Informer encoder and Informer decoder.

[0122] A landslide risk time-series prediction model constructed using the Informer encoder and decoder is trained. During the model training phase, a gradient equalization strategy is implemented based on the gradient distribution characteristics of multi-scale fusion features. This involves dynamically normalizing the gradient magnitudes of high-frequency and low-frequency feature channels to prevent high-frequency signals from dominating the parameter update process. Specifically, this includes: S4-1: Based on the gradient distribution characteristics of the multi-scale fusion feature matrix, the gradient magnitudes of the high-frequency and low-frequency feature channels during backpropagation are obtained using an automatic differentiation mechanism to obtain the original gradient distribution of each channel.

[0123] S4-2: Calculate the gradient energy density of each channel based on the original gradient magnitudes of the high-frequency and low-frequency characteristic channels to quantify the difference in contribution intensity between high-frequency and low-frequency signals to parameter updates.

[0124] S4-3: Based on the statistical results of the gradient energy density distribution, a dynamic normalization coefficient generation function is constructed. This function dynamically adjusts the normalization factor according to the relative proportion of channel gradient energy to suppress the dominant role of high-frequency signals in the parameter update process.

[0125] Based on the statistical results of the gradient energy density distribution of the high-frequency and low-frequency feature channels output from the previous sub-step, an adaptive coefficient optimization method (parameters: channel energy density vector, smoothing time window length) is used to construct the initial structure of the normalized coefficient generation function.

[0126] Furthermore, by using a proportional-maintaining constraint optimization algorithm (parameters: energy density ratio, upper and lower bound coefficient thresholds), the dynamic adjustment function form of the normalization factor is optimized, and a set of basic weight adjustment factors is obtained.

[0127] Furthermore, an exponential mapping smoothing mechanism is adopted (parameter: smoothing factor λ ranges from [0.01, 0.2]) to achieve a smooth transition of the normalized coefficients during continuous gradient iteration and generate a smoothed coefficient time series.

[0128] Furthermore, based on the coefficient time series, a nonlinear suppression function design method (parameters: suppression function type, inflection point control parameter μ) is applied to construct a suppression curve for the high-frequency gradient amplitude and generate a low-frequency gradient amplification curve, which is used to control the relative contribution of the two types of channels in parameter updates.

[0129] By using a weighted coefficient fusion processing method, the normalized coefficients are combined with the original gradient magnitude vector to form a channel gradient contribution balance vector, thereby effectively suppressing the dominant role of high-frequency signals.

[0130] For example, in a training batch containing 256 high-frequency channels and 64 low-frequency channels, the average energy density of the high-frequency gradients is 0.85, and the average energy density of the low-frequency gradients is 0.35. The energy density ratio is calculated using a proportionality-preserving constraint optimization algorithm. The resulting ratio is 2.4286. Upper and lower bound threshold coefficients are set to 0.8 and 1.25, respectively, generating initial adjustment factors of 0.92 for the high-frequency channel and 1.18 for the low-frequency channel. An exponential smoothing function is then applied. Under the condition of λ=0.05, the adjustment factor is smoothed to avoid drastic fluctuations in coefficients during training. A nonlinear inhibition function is used. Under the condition of μ=0.7, the adjustment coefficient of the high-frequency channel was further reduced to 0.88, while the adjustment coefficient of the low-frequency channel was increased to 1.22. Through final fusion processing, the high-frequency gradient contribution balance value range [0.86, 0.90] and the low-frequency gradient contribution balance value range [1.20, 1.25] were obtained. The validation results show that under the same training rounds, the participation of the low-frequency feature weights of the model is significantly improved, while the dominance of the high-frequency channel gradient is significantly reduced, and the model convergence curve is more stable.

[0131] S4-4: The gradient magnitudes of the high-frequency and low-frequency feature channels are normalized using a dynamic normalization coefficient generation function to generate normalized gradient vectors, thereby achieving a balance of contribution of each channel in the parameter update process.

[0132] S4-5: Perform parameter update operations based on the normalized gradient vector, and feed the updated model parameters back to the Informer structure to improve the convergence stability and prediction performance of the model during multi-scale feature fusion training.

[0133] Step 105: Input the multi-scale fused feature matrix into the hidden state sequence output by the Informer encoder.

[0134] The Informer encoder works on the following principle: Perform a linear mapping operation on the multi-scale fused feature matrix to generate a query vector dominated by high-frequency features.

[0135] The key vector and value vector are generated by performing key vector and value vector generation operations on the weighted and reconstructed multi-scale fusion features.

[0136] Multi-head self-attention mechanism computation is performed based on query vector, key vector and value vector: using the correlation scores between multi-scale features, the model is guided to identify co-evolutionary feature interaction patterns at different time scales, and output attention-weighted latent state feature matrix.

[0137] Perform positional encoding enhancement on the hidden state feature matrix.

[0138] The hidden state feature matrix after the position encoding enhancement operation is processed by a feedforward neural network to extract high-order nonlinear feature representations as the output hidden state sequence.

[0139] Step 105 inputs the multi-scale fused feature matrix into the Informer encoder. The Informer encoder is improved so that the query vector of the self-attention mechanism is generated primarily by high-frequency features, and the key vector and value vector are constructed from the weighted fused multi-scale features, guiding the model to focus on extracting key spatiotemporal correlation features during multi-scale co-evolution periods. Specifically, this includes: S5.1: Perform a linear mapping operation on the multi-scale fused feature matrix to generate a query vector dominated by high-frequency features. Based on the local fluctuation patterns of the high-frequency feature sequence, project them onto a unified semantic space using a learnable weight matrix to obtain a high-frequency dominated query vector Query, which guides the model to focus on key time points in deformation evolution.

[0140] Based on the multi-scale fusion feature matrix, the high-frequency feature sequence is selected as the dominant data source for generating the query vector. The high-frequency feature sequence contains a high-temporal-density feature embedding vector obtained by encoding the local fluctuation pattern of InSAR deformation data.

[0141] A linear mapping method is used to project high-frequency feature sequences onto a unified semantic space step by step, thereby realizing a linear transformation from high-dimensional feature vectors to the query space.

[0142] Furthermore, through the weight matrix W q The learnable parameters are automatically optimized during training, so that the projected query vector retains the salient patterns of high-frequency deformation signals in the semantic space while suppressing noise components.

[0143] Furthermore, channel normalization is performed on the projection results through a normalization layer to ensure that the numerical scale of the query vector is consistent at different times, thereby eliminating the interference of input distribution differences on attention calculation.

[0144] Furthermore, a positional coding enhancement method is employed to enhance the temporal positional information vector PE generated by the temporal coding branch. qThe query vector after linear projection is added element-wise to form a query matrix that is sensitive to temporal position, which is used to emphasize key time points of deformation evolution in subsequent self-attention mechanisms.

[0145] By using the above-mentioned linear mapping and positional encoding fusion processing method, high-frequency feature sequences are transformed into query vector data with a unified semantic space and containing temporal positional sensitive information, thereby achieving the expected technical effect of guiding the Informer encoder to focus on key evolutionary moments in multi-scale features.

[0146] S5.2: Perform key and value vector generation operations on the weighted reconstructed multi-scale fused features. Based on the concatenated representation of low-frequency and high-frequency features under the control of the dynamically aligned weight matrix, key and value vectors are generated through independent linear transformation layers to capture the contextual dependencies between multi-scale features.

[0147] Based on the semantically aligned and time-synchronized multi-scale fusion feature matrix obtained from S4.5, an independent linear transformation layer (parameters: weight matrix size matches the fusion feature dimension, bias vector initialized to zero) is used to perform feature subspace mapping on the fusion matrix, realizing the transformation from the original high-dimensional feature space to the key vector generation space.

[0148] Furthermore, another set of independent linear transformation layers (parameters: weight matrix matched with the value generation space, bias vector initialized with a small random perturbation) is used to perform a mapping on the fusion matrix to obtain the projected features for value vector generation, thus realizing a linear mapping from the original fusion features to the value vector space.

[0149] Furthermore, for the key vector generation process, a feature regulation mechanism based on a dynamic alignment weight matrix is ​​introduced (parameter: the weight matrix comes from the output of S3.4). Before the linear mapping of the key vector, the feature intensity of the low-frequency feature channel is scaled channel by channel according to the alignment weight, so as to realize the semantic weighting of low-frequency features in the key vector construction process.

[0150] Furthermore, for the value vector generation process, the same dynamic weight control is adopted, and the low-frequency feature components in the fusion matrix are multiplied by the corresponding weight coefficients before the mapping calculation, so as to retain the temporal context contribution of low-frequency features in the value vector and enhance the dependency representation ability of multi-scale features.

[0151] Through the above linear transformation and dynamic weight adjustment, the fusion matrix of the previous step is transformed into a key vector Key and a value vector Value that meet the input requirements of the Informer self-attention mechanism, thereby realizing high-dimensional semantic mapping of multi-scale feature context dependencies.

[0152] For example, in a landslide prediction task for a power transmission line in a mountainous area of ​​southwest China, the multi-scale fusion feature matrix has a dimension of 128, with high-frequency features occupying 64 dimensions and low-frequency features occupying 64 dimensions. The linear transformation layer weight matrix generated from key vectors has a shape of 128×64, and the linear transformation layer weight matrix generated from value vectors also has a shape of 128×64. The dynamically aligned weight matrix comes from the output of S3.4 and has a dimension of 64×64, with scaling applied row-by-row on the low-frequency feature channels. For example, if the weight coefficient of a certain low-frequency feature at the 10th time step is 0.85, all low-frequency feature values ​​corresponding to that time step are multiplied by 0.85 during key vector generation. The same weight scaling is used in the value vector generation process. After transformation, the generated key matrix has a dimension of 64×64, and the value matrix has a dimension of 64×64. In the subsequent self-attention mechanism calculation, this significantly improves the model's sensitivity to the interaction of features at different time scales. Validation results show that the RMSE index for multi-step displacement prediction is significantly reduced compared to the uncontrolled state.

[0153] S5.3: Based on the generated query vector, key vector, and value vector, a multi-head self-attention mechanism is executed. By utilizing the correlation scores between multi-scale features, the model is guided to identify co-evolutionary feature interaction patterns at different time scales, outputting an attention-weighted latent state feature matrix, thereby enhancing the model's spatiotemporal perception of the landslide evolution process.

[0154] Based on a query vector (Query) generated primarily from high-frequency features, and key and value vectors (Key and Value) generated from weighted fusion of multi-scale features, a multi-head self-attention computation method is employed (parameters: number of attention heads h, scaling factor...). ), of which (d k (The key vector dimension is used to model multi-scale feature associations.) Further, a scaling dot product attention algorithm (parameters: Query, Key matrix, Value matrix) is used to calculate the matching weights between each high-frequency feature time step and the multi-scale fused feature time step, obtaining the relevance score matrix within the time scale. Further, a Softmax normalization function (parameter: relevance score matrix) is used to achieve a probabilistic distribution of attention weights at each time step, highlighting the feature interaction intensity corresponding to highly relevant time positions. Further, a weighted product operation is performed on the normalized attention weight matrix and the Value matrix to generate an enhanced latent state representation matrix in the multi-scale feature context space, preserving the cross-timescale co-evolutionary pattern. Further, the multi-head attention outputs are concatenated in the head dimension and synthesized into a unified-dimensional latent state feature matrix through a linear mapping layer, ensuring the fusion and information integration of different head attention patterns.

[0155] .

[0156] Where Query is the Q matrix, Key is the K matrix, and Value is the V matrix. The feature dimension of the Key is defined. Through multi-head parallel computation, the above Attention operation is executed independently in different subspaces and then mapped to a unified semantic space after concatenation, thereby achieving the fusion of multi-scale information.

[0157] By using a multi-head self-attention algorithm, the Query, Key, and Value vectors from the previous step are transformed into a latent state feature matrix containing cross-timescale co-evolutionary patterns, thereby achieving the expected technical effect of improving the Informer encoder's spatiotemporal perception of the landslide evolution process.

[0158] For example, when processing landslide monitoring data of a power transmission line in a mountainous area of ​​southwestern China, the number of attention heads h is set to 8, and the key vector dimension is... The value is 64, and the scaling factor is... Using the normalized multi-scale fusion feature matrix as input, the generated Query matrix has dimensions [1024, 64], Key matrix has dimensions [1024, 64], and Value matrix has dimensions [1024, 64]. A scaled dot product attention algorithm is used to calculate the relevance score matrix with dimensions [1024, 1024]. After normalization using the Softmax function, the weights of highly relevant time steps are significantly increased, which is reflected in the weighted Value matrix enhancement representation, resulting in an output latent state feature matrix with dimensions [1024, 64]. In multi-head parallel mode, the latent state matrices from eight different subspaces are concatenated and merged into a high-dimensional representation of [1024, 512], which is then compressed to [1024, 256] through a linear mapping layer for subsequent location encoding enhancement and feedforward network processing. Under this configuration, the encoder can significantly improve the cross-timescale feature alignment and key pattern recognition capabilities in landslide precursor displacement sequences, and the output feature sequence maintains high stability and discriminability in subsequent prediction tasks.

[0159] S5.4: Perform positional encoding enhancement on the latent state feature matrix output by the multi-head self-attention mechanism. Based on the temporal positional information output by the temporal encoding branch, a learnable positional embedding vector is introduced to compensate for the insufficient ability of the attention mechanism to model temporal order and improve the model's accuracy in temporal modeling of deformation evolution trends.

[0160] For the hidden state feature matrix output by the multi-head self-attention mechanism, a positional encoding enhancement method (parameters: learnable positional embedding dimension and initialization strategy) is adopted to explicitly inject temporal sequence information, thereby making up for the shortcomings of the attention mechanism in absolute temporal sequence perception.

[0161] Furthermore, the temporal position information output by the time-coded branch is parsed by the position coding generation function (parameters: sine / cosine fundamental frequency range, embedding vector length) to obtain the original position representation vector of each time step, and retain its index identifier on the time axis.

[0162] Furthermore, a learnable position embedding mapping algorithm (parameters: training rounds, learning rate) is adopted to project the original position representation vector onto the same feature dimension space as the hidden state feature matrix, generating a position embedding matrix to achieve element-wise matching between the position vector and the hidden state features.

[0163] Furthermore, by employing a vector addition fusion strategy, the position embedding matrix and the hidden state feature matrix are added element by element to obtain an enhanced hidden state feature sequence containing position encoding, thereby introducing semantic constraints of temporal and positional information into the feature representation.

[0164] Furthermore, to prevent the positional encoding from introducing value range biases at different scales, a normalization method is used to perform layer normalization on the fused feature sequence to maintain the consistency of multi-scale feature distribution. The output latent state feature matrix enhanced by positional encoding is then used as the input to the encoder's subsequent feedforward network.

[0165] By using a positional encoding enhancement algorithm, the multi-head self-attention output of the previous step is transformed into high-dimensional feature data with explicit temporal index semantics, thereby improving the model's ability to perceive temporal order in landslide deformation trend modeling.

[0166] For example, in a landslide monitoring task of a power transmission line in a mountainous area of ​​southwest China, the Informer encoder's hidden state feature matrix dimension is set to 512, the time step length is 128, and the location embedding matrix dimension is also 512. It is initialized using a mixture of sine and cosine functions, with a fundamental frequency range of 0.0001-0.1. The embedding vector training epochs are 50, and the learning rate is 0.001. In this scenario, the time step index provided by the time encoding branch is parsed into 128 original location vectors by the location encoding generation function, each vector having a length of 512. Through the location embedding mapping algorithm, these vectors are projected onto the hidden state feature dimension space and added element-wise with the corresponding hidden features to form a location-enhanced feature matrix. After normalization, the values ​​of this matrix are distributed between -1.0 and 1.0, effectively ensuring feature stability under different location step lengths. In the model performance evaluation, the introduction of location coding enhancement significantly improved the consistency between the predicted future displacement sequence and the dominant frequency of the actual monitoring data in terms of temporal trend. It also enhanced the ability to capture small-amplitude, gradual deformation processes, effectively identifying early signs of potential landslides and providing more accurate risk warnings for power grid operation safety.

[0167] S5.5: The enhanced hidden state feature matrix is ​​processed by a feedforward neural network (FFN) to extract high-order nonlinear feature representations. A two-layer fully connected network and activation function are used to perform nonlinear transformations on the features, outputting a high-dimensional semantic feature sequence from the encoder, providing a stable and discriminative feature representation for subsequent progressive decoding.

[0168] For the latent state feature matrix enhanced by position encoding, a two-layer fully connected feedforward neural network (FFN) (parameters: the output dimension of the first layer is 4 times the original input dimension, and the activation function is ReLU; the output dimension of the second layer is reduced back to the original input dimension, and the activation function is linear) is used to construct a high-order nonlinear feature representation.

[0169] Furthermore, by using matrix multiplication and bias vector addition, the input feature vector H is mapped to a high-dimensional space to generate an intermediate feature representation. The nonlinear enhanced feature F is then output through the ReLU activation function.

[0170] Step 106: Call the Informer decoder and use a progressive decoding strategy to perform multi-step displacement prediction on the hidden state sequence, and output the deformation evolution trend map.

[0171] Step 106 specifically includes: The Informer decoder is invoked, and based on the hidden state sequence, an attention mechanism is used to calculate the temporal correlation weight between the current hidden state and the historical multi-scale fused features, generating a context-aware decoding initial state vector.

[0172] The initial state vector is processed by a multi-level gated recurrent unit to generate the latent variable output for the current prediction time step. Each gated recurrent unit in the multi-level gated recurrent unit introduces a time decay factor to adjust the proportion of historical information retained.

[0173] Based on the latent variable output of the current prediction time step, the single-step displacement prediction value of the corresponding monitoring point is generated through mapping by the fully connected layer, and the single-step displacement prediction is fed back to the gated recurrent unit decoder as the input of the next time step to generate a progressive displacement sequence.

[0174] The progressive displacement sequence is smoothed by a sliding window based on the prediction error covariance matrix to obtain the displacement prediction sequence.

[0175] Based on the displacement prediction sequence, and combined with the spatial coordinates and geological attribute labels of each monitoring point, a deformation evolution trend map with spatiotemporal semantics is generated.

[0176] Step 106 employs a progressive decoding strategy to perform multi-step displacement prediction on the hidden state sequence output by the Informer encoder. It combines historical multi-scale fusion states with current latent variables to generate predicted values ​​for future displacement sequences, outputting the deformation evolution trend of each monitoring point on the landslide. Specifically, this includes: S6.1: Based on the hidden state sequence output by the Informer encoder, an attention mechanism is used to calculate the temporal correlation weight between the current hidden state and the historical multi-scale fusion features, so as to generate a context-aware decoding initial state vector.

[0177] S6.2: Perform multi-level gated recurrent unit (GRU) processing on the decoded initial state vector to generate the latent variable output for the first prediction time step, where each GRU unit introduces a time decay factor to adjust the retention ratio of historical information to enhance the model's stability for long sequence predictions.

[0178] For the initial decoding state vector generated by the attention mechanism, a multi-layer gated recurrent unit (GRU) structure is adopted (parameters: number of hidden layer units h=128, number of layers L=3, activation function is tanh, update gate and reset gate use sigmoid function) to realize the recursive update and feature extraction functions of the temporal state.

[0179] Furthermore, by introducing a time decay factor α (with a range of 0.85 to 0.98, dynamically adjusted according to the length of the historical sequence and the noise level), an exponential decay is applied to the historical state term in the GRU state update formula, thereby adjusting the retention ratio of remote historical information and obtaining a time-sensitive recursive state vector adapted to long sequence prediction.

[0180] For example, when predicting a landslide in a power transmission section in a mountainous area in Southwest China, the initial decoded state vector dimension is set to 256. A three-layer stacked GRU structure is adopted, with 128 hidden units per layer. The time decay factor α is determined by the input sequence length of 200 steps and the noise variance of 0.25, and is set to 0.92. During the GRU update gate calculation, the historical state term is multiplied by a weight of 0.92 to suppress redundant information from distant historical data. In the output of the first-layer GRU, the update gate value z is calculated to be approximately 0.64 using the above formula. The candidate state and the historical state are fused according to their corresponding proportions to obtain the hidden variable vector h of the first prediction step. t Its response to surface deformation sensitive characteristics shows a synchronous presentation of short-period disturbance patterns and long-term subsidence trends. In subsequent testing phases, the RMSE of the prediction sequence driven by this latent variable remained within 0.48 mm, demonstrating significantly improved prediction stability and the ability to capture precursors of landslides.

[0181] S6.3: Based on the latent variable output of the current prediction time step, the single-step displacement prediction value of the corresponding monitoring point is generated through mapping by the fully connected layer, and the prediction value is fed back to the GRU decoder as the input of the next time step to realize the generation of progressive displacement sequence.

[0182] S6.4: Perform sliding window smoothing on the displacement sequence generated by multi-step prediction, and optimize the transition continuity between time steps based on the prediction error covariance matrix to reduce the impact of high-frequency noise on the deformation trend prediction results.

[0183] S6.5: Based on the smoothed future displacement sequence, combined with the spatial coordinates and geological attribute labels of each monitoring point, a deformation evolution trend map with spatiotemporal semantics is generated as the core input for subsequent landslide risk index (LRI) modeling.

[0184] Based on the future displacement prediction sequence that has been smoothed by a sliding window, a spatial index mapping method (parameter: unique spatial coordinates of monitoring points) is used to construct a one-to-one correspondence between the prediction sequence and the location of the monitoring points in geographic space.

[0185] Furthermore, through a geological attribute label parsing algorithm (parameters: lithology category code, slope value, historical deformation rate level), the stability-related static factors of each monitoring point are processed in a structured and vectorized manner, and a geological attribute feature matrix that can be used for map construction is obtained.

[0186] Furthermore, a spatiotemporal feature fusion method (parameters: timestamp, three-dimensional spatial coordinates, fusion weight matrix) is adopted to jointly encode the temporal dimension of the smooth displacement sequence and the spatial dimension of the monitoring point's geographical location, forming a set of spatial features with time identifiers.

[0187] Furthermore, spatial continuity processing is performed on the fused feature set using a two-dimensional or three-dimensional raster interpolation algorithm (parameters: spatial resolution Δx, Δy, Δz, interpolation kernel function type), generating complete spatiotemporal grid data at a given spatial resolution.

[0188] Furthermore, using a trend graph generation algorithm (parameters: color mapping scheme, displacement amplitude stratification threshold), the spatiotemporal grid data is transformed into a visualized semantic graph data structure, where color stratification corresponds to different displacement change trend levels.

[0189] Through the above-mentioned map generation and processing method, the smoothed future displacement sequence is transformed into a comprehensive spatiotemporal semantic map containing time evolution trend, spatial displacement distribution and geological attribute labels, realizing the generation of core input data that can directly drive the modeling of the comprehensive landslide risk index (LRI).

[0190] For example, in a power transmission line area in a mountainous region of southwest China, a spatial resolution of Δx=10m and Δy=10m was selected. The smoothed displacement prediction sequence for the next 30 days of 60 monitoring points was spatially indexed and mapped to their WGS74 coordinates. Geological attribute labels included lithology codes [granite=01, gneiss=02], slope (°), and historical deformation rate levels [low=1, medium=2, high=3]. The displacement sequence timestamps (unit: days) and the three-dimensional coordinates of the monitoring points were fused using a weight matrix W. f (Dimensions: 4×4, rows correspond to time, longitude, latitude, and altitude, columns correspond to the fused feature dimensions) Linear combination encoding is performed to form a complete spatiotemporal feature set. Bilinear interpolation (10m×10m raster resolution) is used to generate a regional displacement distribution grid, and color mapping is performed through threshold layering [displacement change <0.5mm = green, 0.5mm~1mm = yellow, >1mm = red] to construct a 30-frame dynamic map of future displacement evolution trends. Application results show that this map can intuitively depict the differences in displacement evolution corresponding to different geological attributes, exhibits good spatial continuity, and can provide stable and information-rich input data for subsequent LRI models.

[0191] Step 107: Determine the comprehensive landslide risk index based on the deformation evolution trend map and the spatiotemporal distribution characteristics of multi-source static features, and complete the landslide risk early warning.

[0192] Step 107 specifically includes: A first-order difference operation was performed on the deformation evolution trend map to obtain the displacement acceleration time series characteristics of each monitoring point.

[0193] Based on deformation evolution trend maps, geological lithology maps, and elevation models, a spatial distribution layer of geological vulnerability is generated. Geological vulnerability is calculated by weighting lithology type, slope, and historical deformation rate.

[0194] A meteorological triggering index was constructed using time-series data of meteorological factors. The meteorological triggering index was generated by applying a normalized weighted method based on daily rainfall, soil moisture content change rate, and temperature anomalies.

[0195] A multi-dimensional landslide risk factor matrix was constructed based on displacement acceleration, geological vulnerability, and meteorological triggering index.

[0196] Z-score standardization was performed on the multidimensional landslide risk factor matrix.

[0197] A dynamic weight allocation mechanism model is used to determine the dynamic weight vector. This model is obtained by training a multilayer perceptron model based on a historical landslide event sample set. The dynamic weight allocation mechanism model is used to identify the relative contributions of different risk factors at different times and spatial locations, generating a dynamic weight vector that varies with time and region.

[0198] By using dynamic weight vectors to weight and fuse the standardized multi-dimensional landslide risk factor matrix, a spatiotemporally continuous comprehensive landslide risk index is generated.

[0199] The landslide risk level corresponding to the interval of the comprehensive landslide risk index is determined as the current landslide risk level.

[0200] Drive the early warning and emergency response mechanism corresponding to the current landslide risk level.

[0201] Step 107, based on the spatiotemporal distribution characteristics of deformation prediction sequences and multi-source static features, constructs a comprehensive landslide risk index (LRI) calculation model. This model integrates multi-dimensional assessment indicators such as displacement acceleration, geological vulnerability, and meteorological triggering factors through a dynamic weight allocation mechanism. Specifically, it includes: S7.1: Perform a first-order difference operation on the future displacement prediction sequence output by the Informer model to obtain the displacement acceleration time series characteristics of each monitoring point, which is used to characterize the dynamic evolution rate of the landslide body.

[0202] S7.2: Based on InSAR deformation data, geological lithology maps and elevation models, a spatial distribution layer of geological vulnerability is generated. Geological vulnerability is calculated by weighting lithology type, slope and historical deformation rate to reflect regional landslide sensitivity.

[0203] S7.3: Construct a meteorological trigger index using time-series data of meteorological factors. This index is generated using a normalized weighted method based on daily rainfall, soil moisture content change rate and temperature anomaly values ​​to quantify the intensity of external environment inducing landslides.

[0204] S7.4: Based on three assessment indicators—displacement acceleration, geological vulnerability, and meteorological triggering index—a multi-dimensional landslide risk factor matrix is ​​constructed, and Z-score standardization is performed on each factor to eliminate the influence of dimensional differences and form a unified risk assessment input space.

[0205] S7.5: Design a dynamic weight allocation mechanism, train a multilayer perceptron (MLP) model based on a historical landslide event sample set to identify the relative contribution of different risk factors at different times and spatial locations, and generate a dynamic weight vector that changes with time and region.

[0206] S7.6: The standardized multi-dimensional factor matrix and dynamic weight vector are weighted and fused to generate a spatiotemporally continuous comprehensive landslide risk index (LRI), which is used to quantify the landslide risk level of each monitoring point at different time periods.

[0207] S7.7: Based on the numerical distribution of the comprehensive landslide risk index (LRI), risk level thresholds are defined, and a piecewise linear mapping function is used to map the LRI values ​​to the risk probability range of 0 to 1, outputting the risk warning level to drive subsequent early warning and emergency response mechanisms.

[0208] The method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment provided in this application constructs a multi-source heterogeneous dataset containing InSAR deformation time series, elevation, lithology and meteorological factors. A dual-channel time coding matrix is ​​generated by sampling with a fixed time window and piecewise interpolation. A cross-scale dynamic alignment module is designed. The correlation between high-frequency fluctuations and low-frequency trends is calculated by using an attention mechanism. A weight matrix is ​​generated to reconstruct low-frequency features and map them to the high-frequency time axis. An Informer architecture is adopted, and a gradient equalization strategy is implemented during training. Combined with progressive decoding, the displacement prediction sequence is output. Finally, displacement acceleration, geological vulnerability and meteorological factors are integrated to construct a comprehensive landslide risk index model.

[0209] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0210] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for monitoring and risk early warning of landslides in mountainous areas for key power transmission equipment, characterized in that, include: An input matrix was constructed based on multi-source heterogeneous data from key power transmission equipment in mountainous areas. The input matrix is ​​divided into high-frequency feature sequences and low-frequency feature sequences according to the differences in time scale. Based on the temporal distribution characteristics of high-frequency feature sequences and the envelope characteristics of low-frequency feature sequences, a dynamic alignment weight matrix is ​​constructed. The low-frequency feature sequence is reconstructed by weighting the dynamic alignment weight matrix, and the reconstructed features are mapped to the time axis of high-frequency features through back projection operation to generate a semantically aligned multi-scale fusion feature matrix. The multi-scale fused feature matrix is ​​input into the Informer encoder, which outputs a hidden state sequence. The Informer decoder is invoked, and a progressive decoding strategy is used to perform multi-step displacement prediction on the hidden state sequence, outputting a deformation evolution trend map. Based on the deformation evolution trend map and the spatiotemporal distribution characteristics of multi-source static features, a comprehensive landslide risk index is determined to complete the landslide risk early warning.

2. The method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment according to claim 1, characterized in that, The multi-source heterogeneous data includes: high-frequency dynamic data, static data, and low-frequency dynamic data; The high-frequency dynamic data is: the original time-series data of the InSAR deformation monitoring points; The static data is: elevation data; The low-frequency dynamic data includes lithological data and meteorological data.

3. The method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment according to claim 2, characterized in that, An input matrix is ​​constructed based on multi-source heterogeneous data from mountainous areas containing key power transmission equipment, specifically including: The high-frequency dynamic data is processed using a first preprocessing method to obtain a land surface deformation observation sequence; the first preprocessing method includes, in sequence: neighborhood time series mean interpolation method, z-score anomaly detection algorithm, local quadratic polynomial regression smoothing method, and wavelet denoising algorithm. Static data and low-frequency dynamic data are standardized and normalized; the standardization and normalization process includes, in turn: Z-score standardization method, min-max normalization method, robust standardization method and mean-variance consistency test method. Based on the acquisition frequency and update cycle of the surface deformation observation sequence, a fixed time window sampling mechanism is constructed to perform sliding window slicing on the surface deformation observation sequence, generating high-density time embedding vectors for high-frequency channels; By using piecewise constant interpolation and consistency verification methods, low-frequency dynamic data are expanded into a high-granularity dataset aligned with the land surface deformation observation sequence; The trend envelope feature vector of the low-frequency channel is extracted from a high-granularity dataset using a moving average filter. The high-density time embedding vector of the high-frequency channel and the trend envelope feature vector of the low-frequency channel are concatenated along the time axis using a time axis index matching method to generate an input matrix. The time axis index matching method includes, in sequence: channel concatenation algorithm, missing value compensation interpolation algorithm, feature standardization processing and matrix dimension verification algorithm.

4. The method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment according to claim 2, characterized in that, The input matrix is ​​divided into high-frequency feature sequences and low-frequency feature sequences according to the time scale difference, specifically including: Based on the time-frequency characteristics, the input matrix is ​​separated in the frequency domain to obtain high-frequency dynamic feature sequences and low-frequency static feature sequences. Based on the sampling frequency and time resolution of high-frequency dynamic data, a sliding window local sequence extraction is performed on the high-frequency dynamic feature sequence to obtain high-frequency local fluctuation mode features. Piecewise constant interpolation is performed on the low-frequency static feature sequence to obtain the extended low-frequency feature sequence; The moving average filtering algorithm is used to extract low-frequency trend envelope features from the expanded low-frequency feature sequence. Channel segmentation is performed on the high-frequency local fluctuation pattern features and the low-frequency trend envelope features to generate high-frequency feature sequences and low-frequency feature sequences; both the high-frequency feature sequences and the low-frequency feature sequences are structured data.

5. The method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment according to claim 2, characterized in that, Based on the temporal distribution characteristics of high-frequency feature sequences and the envelope characteristics of low-frequency feature sequences, a dynamic alignment weight matrix is ​​constructed, specifically including: Based on the temporal distribution characteristics of high-frequency feature sequences, a set of temporal convolution kernels with shared weights is used to perform convolution processing on high-frequency local fluctuation pattern features to extract the temporal dynamic representation of high-frequency features. Perform a sliding window slicing operation on the low-frequency trend envelope of the low-frequency feature sequence to generate a low-frequency local context fragment aligned with the time window of the high-frequency features; By using a multi-head attention mechanism, the temporal dynamic representation of high-frequency features is matched with low-frequency local context fragments in terms of semantics and temporal position to obtain a relevance score that reflects cross-scale correlation. Based on the relevance score, a dynamic alignment weight matrix is ​​constructed using a second preprocessing method; the second preprocessing method includes, in sequence: a normalization algorithm, a weight smoothing algorithm, an importance recalibration algorithm, a matrix normalization method, and an additive attention weight tuning algorithm.

6. The method for monitoring and risk early warning of landslides in mountainous areas for key power transmission equipment according to claim 2, characterized in that, The low-frequency feature sequence is reconstructed using a dynamic alignment weight matrix, and the reconstructed features are mapped to the time axis of high-frequency features through back projection, generating a semantically aligned multi-scale fusion feature matrix, specifically including: Based on the dynamic alignment weight matrix, a weighted mapping operation is performed on the feature vector of each time step in the low-frequency feature sequence to obtain the preliminary aligned representation of the low-frequency features on the high-frequency time axis. Based on the initial alignment representation, a differentiable back projection transformation function is used to map the weighted low-frequency features from the local window back to the original high-frequency time axis, thereby constructing a reconstructed low-frequency feature sequence aligned with the high-frequency features in the time dimension. Based on the local fluctuation pattern of high-frequency feature sequences, the attention weight distribution corresponding to each time step is calculated. Based on the attention weight distribution, the temporal alignment accuracy of reconstructed low-frequency features is dynamically adjusted; The reconstructed low-frequency features and high-frequency features after dynamic adjustment are spliced ​​together in the channel dimension to obtain the initial multi-scale fusion feature matrix; Normalization is performed on the initial multi-scale fusion feature matrix to obtain a semantically aligned and time-synchronized multi-scale fusion feature matrix.

7. The method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment according to claim 2, characterized in that, Before inputting the multi-scale fused feature matrix into the Informer encoder to output the hidden state sequence, the following steps are also included: Train the Informer encoder and Informer decoder; Training the Informer encoder and Informer decoder specifically includes: Based on the gradient distribution characteristics of the multi-scale fused feature matrix, the original gradient magnitudes of the high-frequency and low-frequency feature channels during backpropagation are obtained using an automatic differentiation mechanism. Calculate the gradient energy density of each channel based on the original gradient magnitudes of the high-frequency and low-frequency characteristic channels. A dynamic normalization coefficient generation function is constructed based on the statistical results of gradient energy density distribution. The gradient magnitudes of the high-frequency and low-frequency feature channels are normalized using a dynamic normalization coefficient generation function to generate normalized gradient vectors. The parameter update operation is performed based on the normalized gradient vector, and the updated model parameters are fed back to the Informer encoder and Informer decoder.

8. The method for monitoring and risk early warning of landslides in mountainous areas for key power transmission equipment according to claim 2, characterized in that, The principle of the Informer encoder is as follows: Perform a linear mapping operation on the multi-scale fused feature matrix to generate a query vector dominated by high-frequency features; The key vector and value vector are generated by performing key vector and value vector generation operations on the weighted and reconstructed multi-scale fusion features. Multi-head self-attention mechanism computation is performed based on query vector, key vector and value vector: using the correlation scores between multi-scale features, the model is guided to identify co-evolving feature interaction patterns at different time scales, and output attention-weighted latent state feature matrix. Perform positional encoding enhancement on the hidden state feature matrix; The hidden state feature matrix after the position encoding enhancement operation is processed by a feedforward neural network to extract high-order nonlinear feature representations as the output hidden state sequence.

9. The method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment according to claim 2, characterized in that, The Informer decoder is invoked, and a progressive decoding strategy is employed to perform multi-step displacement prediction on the hidden state sequence, outputting a deformation evolution trend map, specifically including: The Informer decoder is invoked, and based on the hidden state sequence, an attention mechanism is used to calculate the temporal correlation weight between the current hidden state and the historical multi-scale fusion features, generating a context-aware decoding initial state vector. The initial state vector is processed by a multi-layer gated recurrent unit to generate the latent variable output for the current prediction time step; each gated recurrent unit in the multi-layer gated recurrent unit introduces a time decay factor to adjust the retention ratio of historical information; Based on the latent variable output of the current prediction time step, the single-step displacement prediction value of the corresponding monitoring point is generated through mapping by the fully connected layer, and the single-step displacement prediction is fed back to the gated recurrent unit decoder as the input of the next time step to generate a progressive displacement sequence. The progressive displacement sequence is smoothed by a sliding window based on the prediction error covariance matrix to obtain the displacement prediction sequence. Based on the displacement prediction sequence, and combined with the spatial coordinates and geological attribute labels of each monitoring point, a deformation evolution trend map with spatiotemporal semantics is generated.

10. The method for monitoring and risk warning of landslides in mountainous areas for key power transmission equipment according to claim 2, characterized in that, Based on the spatiotemporal distribution characteristics of deformation evolution trend maps and multi-source static features, a comprehensive landslide risk index is determined to complete landslide risk early warning, specifically including: A first-order difference operation was performed on the deformation evolution trend map to obtain the displacement acceleration time series characteristics of each monitoring point; Based on deformation evolution trend maps, geological lithology maps, and elevation models, a spatial distribution layer of geological vulnerability is generated; the geological vulnerability is obtained by weighted calculation based on lithology type, slope, and historical deformation rate. A meteorological triggering index is constructed using time-series data of meteorological factors. The meteorological triggering index is generated by performing a normalized weighted method based on daily rainfall, soil moisture content change rate and temperature anomaly values. A multi-dimensional landslide risk factor matrix was constructed based on displacement acceleration, geological vulnerability, and meteorological triggering index; Z-score standardization was performed on the multi-dimensional landslide risk factor matrix; A dynamic weight vector is determined using a dynamic weight allocation mechanism model. The dynamic weight allocation mechanism model is obtained by training a multilayer perceptron model based on a historical landslide event sample set. The dynamic weight allocation mechanism model is used to identify the relative contribution of different risk factors at different times and spatial locations, and to generate a dynamic weight vector that changes with time and region. The standardized multi-dimensional landslide risk factor matrix is ​​weighted and fused using dynamic weight vectors to generate a spatiotemporally continuous comprehensive landslide risk index. The landslide risk level corresponding to the interval of the comprehensive landslide risk index is determined as the current landslide risk level. Drive the early warning and emergency response mechanism corresponding to the current landslide risk level.

Images