A method for determining the threshold of landslide early warning factors in high dam and large reservoir areas based on deep learning

Abstract

This invention discloses a deep learning-based method for determining thresholds for early warning landslide factors in high dam and large reservoir areas. The method comprises the following steps: obtaining a structurally unified multi-source landslide monitoring dataset; forming a factor-band separated data tensor; inputting the factor-band separated data tensor into a FEDformer frequency domain decoupling module to form a multi-scale prediction component; inputting the multi-scale prediction component and the factor-band separated data tensor into a residual gated fusion layer to output a residual gated fusion feature tensor; extracting the weight distribution of monitoring factors based on the residual gated fusion feature tensor in a dynamic threshold generation module to form a dynamic landslide early warning threshold; and classifying risk warning levels according to preset risk warning levels. This method enables the dynamic upper and lower thresholds to automatically contract or expand in response to changes in landslide risk, effectively avoiding static threshold hysteresis and misjudgment under extreme operating conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of high dams and large reservoirs, and in particular to a method for determining threshold values ​​of early warning landslide factors in high dam and large reservoir areas based on deep learning. Background Art

[0002] In recent years, large-scale landslides have frequently occurred in the reservoir areas of hydropower projects in the western high mountain canyon regions due to the special topography, complex hydrology and unstable geological conditions. Affected by the seasonal fluctuations in water levels in high dams and large reservoirs, extreme rainfall and rock and soil structures, the landslide deformation process presents a triple coupling feature of slow change in reservoir water level, sudden change in rainfall and lag in seepage pressure.

[0003] In the existing technology, landslide warning mainly relies on single factor thresholds, empirical formulas or simplified multi-factor linear superposition models, which have obvious shortcomings: on the one hand, the complex coupling relationship between multi-source landslide monitoring signals is difficult to effectively decouple using traditional models, which makes the threshold judgment easy to distort, and warning omissions or false alarms often occur due to unreasonable threshold settings. On the other hand, the static threshold setting cannot adapt to the sudden change of reservoir water levels and the complex changing environment of short-term heavy rainfall under extreme working conditions, and it is difficult to reflect the real risk of accelerated landslide deformation in a timely manner. In addition, the existing landslide monitoring system is mainly based on single factor threshold triggering, lacks the ability to segmentally identify trend changes, periodic disturbances and sudden excitations in monitoring data, and cannot dynamically allocate warning weights on a spatiotemporal scale. This leads to poor model interpretability, insufficient timeliness and adaptability, and is difficult to meet the needs of large-scale landslide geological disaster safety monitoring in the reservoir area of ​​western high mountain canyon hydropower projects. Summary of the Invention

[0004] One purpose of the present invention is to propose a method for determining the threshold values ​​of landslide factors for early warning in high dam and large reservoir areas based on deep learning. The present invention enables the dynamic upper threshold and the dynamic lower threshold to automatically shrink or expand according to changes in landslide risk, effectively avoiding the problems of static threshold hysteresis and misjudgment under extreme working conditions.

[0005] A method for determining threshold values ​​of landslide early warning factors in a high dam and large reservoir area based on deep learning according to an embodiment of the present invention includes:

[0006] Collect the multi-source landslide monitoring raw data in the high dam and large reservoir landslide monitoring area, and perform preprocessing operations to obtain a multi-source landslide monitoring dataset with a unified structure;

[0007] The multi-source landslide monitoring data set is input into the factor spectrum analysis layer, and each monitoring factor is subjected to fast Fourier transform. The data are then divided into trend frequency band data, low-frequency periodic frequency band data, and high-frequency mutation frequency band data according to the preset frequency band pass rule to form a factor-frequency band separation data tensor.

[0008] The factor-frequency band separation data tensor is input into the FEDformer frequency domain decoupling module, which decouples the long sequence and outputs the trend component prediction sequence, the period component prediction sequence and the disturbance component prediction sequence, which together constitute the multi-scale prediction component.

[0009] The multi-scale prediction component and the factor-band separation data tensor are input into the residual gated fusion layer to generate the factor-band weight matrix. The input tensor and the gated weighted tensor are fused based on the residual connection structure to output the residual gated fusion feature tensor.

[0010] In the dynamic threshold generation module, the weight distribution of monitoring factors is extracted based on the residual gated fusion feature tensor, and similarity matching is performed with the stored historical near-instability landslide monitoring samples to form a dynamic landslide warning threshold;

[0011] The real-time updated multi-source landslide monitoring data are compared with the dynamic landslide warning threshold, and the risk warning level is divided according to the preset risk warning level.

[0012] Optionally, the multi-source landslide monitoring raw data includes reservoir water level raw data, rainfall raw data, groundwater pressure raw data, surface displacement raw data and microseismic raw data, and the preprocessing operations include time-space alignment, missing segment interpolation, outlier removal and noise suppression operations.

[0013] Optionally, the dividing of trend frequency band data, low-frequency period frequency band data and high-frequency mutation frequency band data includes:

[0014] Each monitoring factor in the multi-source landslide monitoring dataset is taken as input, and each monitoring factor has a corresponding time series data sequence;

[0015] Perform spectrum analysis on the time series data of each monitoring factor to obtain the amplitude response sequence of the monitoring factor at different frequencies;

[0016] The frequency domain amplitude response sequence of each monitoring factor is divided into frequency intervals according to the preset frequency threshold. The frequency intervals include trend frequency bands, low-frequency periodic frequency bands and high-frequency mutation frequency bands.

[0017] In each frequency interval, the frequency domain amplitude response sequence of each monitoring factor is subjected to bandpass filtering to extract the amplitude response component of the monitoring factor in the current frequency interval;

[0018] The amplitude response component of each monitoring factor in each frequency interval obtained by bandpass filtering is restored from the frequency domain to the time domain to obtain the time domain component corresponding to each monitoring factor in different frequency intervals;

[0019] The time domain components of all monitoring factors in all frequency intervals are uniformly combined to obtain the factor-frequency band separation data tensor.

[0020] Optionally, the FEDformer frequency domain decoupling module includes:

[0021] Based on the characteristics of different sensitivity of landslide factors in different frequency bands, the discrete Fourier transform operation is performed on the factor-frequency band separation data tensor along the time dimension to obtain the frequency domain representation tensor.

[0022] Calculate the frequency domain interactive self-attention weight based on the frequency domain representation tensor;

[0023] The frequency domain representation tensor is dynamically weighted and fused across frequency bands using the frequency domain interactive self-attention weights to obtain the frequency domain enhanced feature representation tensor.

[0024] The frequency domain enhanced feature representation tensor is restored from the frequency domain to the time domain through the inverse Fourier transform to obtain the time domain enhanced feature data tensor;

[0025] Based on the time domain enhanced feature data tensor, multi-scale feature aggregation is performed on the trend frequency band, low-frequency periodic frequency band and high-frequency mutation frequency band respectively, and the trend component prediction sequence, periodic component prediction sequence and disturbance component prediction sequence are output respectively, and collectively used as the multi-scale prediction component.

[0026] Optionally, the calculation of the residual gated fusion feature tensor includes:

[0027] Constructing a multi-scale prediction tensor based on the temporal consistency between the multi-scale prediction components and the factor-band separation data tensor;

[0028] Calculate the factor-level gating coefficient for each monitored factor;

[0029] Calculate the band-level gating coefficient for each frequency interval;

[0030] Construct a factor-band weight matrix based on the factor-level gating coefficient and the band-level gating coefficient;

[0031] The factor-band separation data tensor is weighted channel by channel according to the factor-band weight matrix to obtain a gated weighted tensor;

[0032] The gated weighted tensor is element-wise added to the multi-scale prediction tensor to form a residual gated fusion feature tensor.

[0033] Optionally, the generation of the dynamic landslide warning threshold includes:

[0034] The time-varying importance distribution of landslide monitoring factors in each frequency interval is extracted based on the residual gated fusion feature tensor, and the weight distribution tensor of monitoring factors is defined.

[0035] All monitoring factors and elements of all frequency intervals of the monitoring factor weight distribution tensor at the most recent time point are arranged in order to form the current landslide-induced state vector; all monitoring factors and elements of all frequency intervals of each monitoring factor weight distribution tensor stored in the historical near-instability landslide monitoring sample at the corresponding time point are arranged in order to form the corresponding historical near-instability landslide monitoring sample state vector;

[0036] Calculate the cosine similarity between the current landslide-induced state vector and the state vector of each historical near-instability landslide monitoring sample;

[0037] Select the maximum value among all cosine similarities and compare it with the preset landslide state trigger threshold:

[0038] If the maximum cosine similarity is greater than the preset landslide state trigger threshold, the deviation adjustment mechanism is triggered to shrink the interval between the dynamic upper threshold and the dynamic lower threshold;

[0039] If the maximum cosine similarity is not greater than the preset landslide state triggering threshold, the deviation adjustment mechanism will not be triggered, and the conventional rolling statistical strategy will continue to be used to calculate the dynamic threshold;

[0040] In the rolling statistics strategy, a sliding time window of fixed length is set, and the average value of the residual gated fusion feature tensor of each monitoring factor in each frequency interval within the sliding time window is calculated to obtain the predicted mean tensor, and the predicted standard deviation tensor is calculated at the same time;

[0041] When the maximum cosine similarity is greater than the preset landslide state trigger threshold, the first set of upper and lower threshold adjustment coefficients are used to determine the 1-D dynamic upper threshold and the 1-D dynamic lower threshold, so that the interval between the 1-D dynamic upper threshold and the 1-D dynamic lower threshold is reduced;

[0042] When the maximum cosine similarity is not greater than the preset landslide state triggering threshold, the second set of upper and lower threshold adjustment coefficients are used to determine the 2-D dynamic upper threshold and the 2-D dynamic lower threshold, so that the interval between the 2-D dynamic upper threshold and the 2-D dynamic lower threshold is expanded;

[0043] The dynamic upper thresholds and dynamic lower thresholds of all monitoring factors in all frequency intervals are uniformly combined to form a dynamic landslide warning threshold.

[0044] Optionally, the risk warning level determination includes:

[0045] Obtain multi-source landslide monitoring data at the current time point to form the current monitoring data tensor;

[0046] For each monitoring factor and each frequency interval, determine which set of dynamic landslide warning threshold intervals should be used:

[0047] When the maximum cosine similarity is greater than the preset landslide state trigger threshold, the 1-D dynamic upper threshold and the 1-D dynamic lower threshold are used as the threshold intervals of this monitoring factor and this frequency interval at the current time point;

[0048] When the maximum cosine similarity is not greater than the preset landslide state trigger threshold, the 2-D dynamic upper threshold and the 2-D dynamic lower threshold are used as the threshold intervals of this monitoring factor and this frequency interval at the current time point;

[0049] Compare the current monitoring value with the determined dynamic upper and lower thresholds:

[0050] When the current monitoring value is greater than the corresponding dynamic upper threshold, it is recorded as an upper limit;

[0051] When the current monitoring value is less than the corresponding dynamic lower threshold, it is recorded as a lower limit;

[0052] When the current monitoring value is between the dynamic upper threshold and the dynamic lower threshold, it is recorded as a normal state;

[0053] Count the number of times that all monitoring factors exceed the upper limit and the lower limit in all frequency intervals to obtain the total number of violations; count the maximum number of continuous exceeding factors of all monitoring factors in all frequency intervals to obtain the maximum number of continuous exceeding factors; count the exceeding range between the monitoring values ​​of all monitoring factors in all frequency intervals and the corresponding dynamic upper threshold or dynamic lower threshold to obtain the maximum exceeding range of a single monitoring factor, and make a risk warning level judgment, and output the blue warning level, yellow warning level, orange warning level and red warning level.

[0054] Optionally, the risk warning level is determined based on:

[0055] When the current monitoring values ​​of all monitoring factors in all frequency intervals are within the dynamic threshold range, a blue warning level is output;

[0056] When the total number of over-limit factors is less than the first boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, and the maximum number of no consecutive over-limit factors exceeds the boundary threshold, a yellow warning level is output;

[0057] When the total number of over-limit factors is greater than or equal to the first boundary coefficient of the risk warning level multiplied by the total number of monitoring channels and less than the second boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, or the maximum number of consecutive over-limit factors exceeds the boundary threshold, an orange warning level is output;

[0058] When the total number of excess limits is greater than or equal to the second boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, or the maximum excess limit of a single monitoring factor is greater than the extremely high excess limit threshold of a single monitoring factor, a red warning level is output.

[0059] The first dividing coefficient of the risk warning level refers to the dividing standard coefficient used to distinguish between the yellow and orange risk warning levels. The second dividing coefficient of the risk warning level refers to the dividing standard coefficient used to distinguish between the orange and red risk warning levels. The total number of monitoring channels refers to the total number formed by the combination of all monitoring factors and all frequency intervals.

[0060] The beneficial effects of the present invention are:

[0061] The present invention realizes component-level modeling of landslide inducing mechanism through frequency domain-time domain dual-path decoupling and dynamic feature extraction of multi-source landslide monitoring data. At the same time, by constructing factor-frequency band separation data tensor, the multi-source landslide monitoring data is refined in frequency domain decomposition according to trend frequency band, periodic frequency band and high-frequency mutation frequency band. The improved FEDformer frequency domain decoupling module is introduced into the frequency domain interactive self-attention mechanism to realize explicit modeling of the interaction relationship between multi-frequency band characteristics of different landslide disaster factors. It can accurately separate and identify the independent disaster-causing contributions in the triple coupling mechanism of slow change of reservoir water level-sudden change of rainfall-hysteresis of seepage pressure, reduce the distortion risk of setting landslide warning threshold, and improve the sensitivity of the model to the accelerated deformation signal of landslide in complex environment.

[0062] The present invention enables full-process adaptation of threshold generation through a residual gated fusion structure, improves the personalization and interpretability of threshold decisions, adopts a dual-gated structure at the factor level and the frequency band level, and combines residual connections to deeply fuse multi-scale prediction components with factor-frequency band separation data tensors to generate a residual gated fusion feature tensor, and obtains the weight distribution of monitoring factors based on weight normalization. By introducing the similarity matching and deviation adjustment mechanism of historical near-instability samples, the dynamic upper threshold and the dynamic lower threshold can automatically shrink or expand according to changes in landslide risk, effectively avoiding the problems of static threshold hysteresis and misjudgment under extreme working conditions.

[0063] The present invention greatly improves the accuracy and response efficiency of reservoir area landslide safety monitoring through the linkage of dynamic landslide warning threshold tensor and risk level intelligent grading. By performing sliding window statistics on the residual gated fusion feature tensor and configuring multiple sets of threshold adjustment coefficients, the 1-D dynamic upper threshold, 2-D dynamic upper threshold, 1-D dynamic lower threshold and 2-D dynamic lower threshold are generated in real time. In combination with the current monitoring data, multi-factor and multi-band channel-by-channel dynamic over-limit judgment is performed to form a four-level intelligent risk grading system of blue, yellow, orange and red. This ensures that under high-risk conditions such as extreme water level fluctuations or heavy rainfall, the early warning system can respond quickly and output a clear warning level, thereby realizing all-weather, intelligent safety monitoring of landslide disasters in high dams and large reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0065] Figure 1 This is a flow chart of the method for determining the threshold value of early warning landslide factors in high dam and large reservoir areas based on deep learning proposed by the present invention. DETAILED DESCRIPTION

[0066] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0067] refer to Figure 1 , a method for determining the threshold value of landslide factors for early warning in high dam and large reservoir areas based on deep learning, including:

[0068] Collect the multi-source landslide monitoring raw data in the high dam and large reservoir landslide monitoring area, and perform preprocessing operations to obtain a multi-source landslide monitoring dataset with a unified structure;

[0069] The multi-source landslide monitoring data set is input into the factor spectrum analysis layer, and each monitoring factor is subjected to fast Fourier transform. The data are then divided into trend frequency band data, low-frequency periodic frequency band data, and high-frequency mutation frequency band data according to the preset frequency band pass rule to form a factor-frequency band separation data tensor.

[0070] The factor-frequency band separation data tensor is input into the FEDformer frequency domain decoupling module, which decouples the long sequence and outputs the trend component prediction sequence, the period component prediction sequence and the disturbance component prediction sequence, which together constitute the multi-scale prediction component.

[0071] The multi-scale prediction component and the factor-band separation data tensor are input into the residual gated fusion layer to generate the factor-band weight matrix. The input tensor and the gated weighted tensor are fused based on the residual connection structure to output the residual gated fusion feature tensor.

[0072] In the dynamic threshold generation module, the weight distribution of monitoring factors is extracted based on the residual gated fusion feature tensor, and similarity matching is performed with the stored historical near-instability landslide monitoring samples to form a dynamic landslide warning threshold;

[0073] The real-time updated multi-source landslide monitoring data are compared with the dynamic landslide warning threshold, and the risk warning level is divided according to the preset risk warning level.

[0074] In this embodiment, the multi-source landslide monitoring raw data includes reservoir water level raw data, rainfall raw data, groundwater pressure raw data, surface displacement raw data and microseismic raw data, and the preprocessing operations include time and space alignment, missing segment interpolation, outlier removal and noise suppression operations.

[0075] In this embodiment, the trend frequency band data, low-frequency period frequency band data and high-frequency mutation frequency band data are divided, including:

[0076] Each monitoring factor in the multi-source landslide monitoring dataset is taken as input, and each monitoring factor has a corresponding time series data sequence;

[0077] The monitoring factors include reservoir water level monitoring factor, rainfall monitoring factor, groundwater pressure monitoring factor, surface displacement monitoring factor and microseismic monitoring factor. The time series data sequence contains continuous monitoring time points and the monitoring values ​​at each time point.

[0078] Perform spectrum analysis on the time series data of each monitoring factor to obtain the amplitude response sequence of the monitoring factor at different frequencies;

[0079] The spectrum analysis process is to convert the time series data sequence of the monitoring factor from the time domain to the frequency domain. The frequency domain represents the amplitude response of the monitoring factor at different frequencies. The output of the spectrum analysis process is the amplitude response sequence of the monitoring factor at different frequencies.

[0080] The frequency domain amplitude response sequence of each monitoring factor is divided into frequency intervals according to the preset frequency threshold. The frequency intervals include trend frequency bands, low-frequency periodic frequency bands and high-frequency mutation frequency bands.

[0081] The trend frequency band represents the frequency interval within the low-frequency range, the low-frequency cycle frequency band represents the frequency interval within the medium-low frequency range, and the high-frequency mutation frequency band represents the frequency interval within the high-frequency range. The specific boundaries of the frequency intervals are defined by the preset frequency threshold.

[0082] In each frequency interval, the frequency domain amplitude response sequence of each monitoring factor is subjected to bandpass filtering to extract the amplitude response component of the monitoring factor in the current frequency interval;

[0083] The amplitude response component reflects the characteristic changes of the monitoring factor in a specific frequency range.

[0084] The amplitude response component of each monitoring factor in each frequency interval obtained by bandpass filtering is restored from the frequency domain to the time domain to obtain the time domain component corresponding to each monitoring factor in different frequency intervals;

[0085] The time domain component is used to reflect the dynamic characteristics of the monitoring factor changing with time in different frequency ranges.

[0086] The time domain components of all monitoring factors in all frequency intervals are uniformly combined to obtain the factor-frequency band separation data tensor;

[0087] The factor-frequency band separation data tensor is a multidimensional data set indexed by monitoring factors, frequency intervals and time points. Each element in the factor-frequency band separation data tensor represents the decoupled data value of the corresponding monitoring factor in the corresponding frequency interval and the corresponding time point.

[0088] In this embodiment, the FEDformer frequency domain decoupling module includes:

[0089] Based on the characteristics of different sensitivity of landslide factors in different frequency bands, the discrete Fourier transform operation is performed on the factor-frequency band separation data tensor along the time dimension to obtain the frequency domain representation tensor.

[0090] The frequency domain representation tensor is used to reflect the complex amplitude response of each monitoring factor in each frequency interval and at each discrete frequency point.

[0091] Calculate the frequency domain interactive self-attention weight based on the frequency domain representation tensor;

[0092] The frequency domain interactive self-attention weight is used to reflect the correlation strength between each monitoring factor in a certain frequency interval and all other frequency intervals. Indicates the Monitoring factor frequency band and Correlation strength of frequency bands:

[0093] ;

[0094] in, Indicates the The monitoring factor is The first frequency interval The complex amplitude response at each frequency point, Indicates the The monitoring factor is The first frequency interval The complex conjugate amplitude response at each frequency point, is the conjugate operator symbol, is the target frequency interval index to be compared, Represents the complex real part extraction operation, which is used to obtain the real correlation strength of different frequency intervals in the frequency domain, reflecting the actual coupling contribution of the monitoring factor in different frequency band characteristics. is the number of continuous monitoring time points, which is consistent with the time series length of the landslide monitoring data of the high dam and large reservoir. is the number of frequency intervals, S=3, which are trend frequency band, low-frequency cycle frequency band and high-frequency mutation frequency band respectively. All frequency interval indexes .

[0095] The frequency domain representation tensor is dynamically weighted and fused across frequency bands using the frequency domain interactive self-attention weights to obtain the frequency domain enhanced feature representation tensor.

[0096] The frequency domain enhanced feature representation tensor is used to express the enhanced frequency domain response characteristics of each monitoring factor after integrating the information of all other frequency intervals in a certain frequency interval.

[0097] The frequency domain enhanced feature representation tensor is restored from the frequency domain to the time domain through the inverse Fourier transform to obtain the time domain enhanced feature data tensor;

[0098] The time domain enhanced feature data tensor is used to reflect the dynamic feature data of each monitoring factor after integrating the frequency band interaction information in each frequency interval and time point.

[0099] Based on the time-domain enhanced feature data tensor, multi-scale feature aggregation is performed on the trend frequency band, low-frequency periodic frequency band, and high-frequency mutation frequency band, and the trend component prediction sequence, periodic component prediction sequence, and disturbance component prediction sequence are output respectively, and collectively used as the multi-scale prediction component;

[0100] The trend component prediction sequence is used to characterize the trend of slow change characteristics of landslide factors in high dams and large reservoirs. The periodic component prediction sequence is used to characterize the periodic regular change characteristics of landslide factors in high dams and large reservoirs. The disturbance component prediction sequence is used to characterize the sudden disturbance characteristics of landslide factors in high dams and large reservoirs. The multi-scale prediction component is used for accurate prediction of the dynamic determination of the threshold of landslide factors in high dams and large reservoirs.

[0101] This implementation method realizes component-level modeling of landslide inducing mechanism through frequency domain-time domain dual-channel decoupling and dynamic feature extraction of multi-source landslide monitoring data. At the same time, by constructing factor-frequency band separation data tensor, the multi-source landslide monitoring data is refined in frequency domain decomposition according to trend frequency band, periodic frequency band and high-frequency mutation frequency band. The improved FEDformer frequency domain decoupling module is used to introduce the frequency domain interactive self-attention mechanism to realize explicit modeling of the interaction relationship between multi-frequency band characteristics of different landslide disaster factors. It can accurately separate and identify the independent disaster-causing contributions in the triple coupling mechanism of slow change of reservoir water level-sudden change of rainfall-seepage pressure lag, reduce the distortion risk of landslide warning threshold setting, and improve the sensitivity of the model to the accelerated deformation signal of landslide in complex environment.

[0102] In this embodiment, the calculation of the residual gated fusion feature tensor includes:

[0103] Constructing a multi-scale prediction tensor based on the temporal consistency between the multi-scale prediction components and the factor-band separation data tensor;

[0104] The construction process of the multi-scale prediction tensor is as follows: determine that the index of the multi-scale prediction component in the three dimensions of monitoring factor, frequency interval and time point is completely corresponding to the factor-frequency band separation data tensor; and assign the numerical value of the multi-scale prediction component corresponding to the combination as the element value of the multi-scale prediction tensor under the monitoring factor, frequency interval and time point to form a multi-scale prediction tensor. Each element of the multi-scale prediction tensor represents the predicted value of a certain monitoring factor at a certain frequency interval and a certain time point.

[0105] Calculate the factor-level gating coefficient for each monitored factor;

[0106] The factor-level gating coefficient is obtained by the combined effect of the trainable weight matrix and the global statistical vector of the monitoring factor. The global statistical vector of the monitoring factor is the mean of the monitoring factor at all time points and all frequency intervals. The factor-level gating coefficient is mapped to the (0,1) interval through the activation function, which is used to measure the global importance of the monitoring factor in the fusion process.

[0107] Calculate the band-level gating coefficient for each frequency interval;

[0108] The band-level gating coefficient is obtained by the combined effect of the trainable weight matrix and the band statistical vector. The band statistical vector is the mean of the frequency interval over all monitoring factors and all time points. The band-level gating coefficient is mapped to the (0,1) interval through the activation function to measure the global importance of the frequency interval in the fusion process.

[0109] Construct a factor-band weight matrix based on the factor-level gating coefficient and the band-level gating coefficient;

[0110] Each element in the factor-band weight matrix is ​​obtained by multiplying the corresponding factor-level gating coefficient and the frequency band-level gating coefficient. The factor-band weight matrix is ​​used to adjust the channel weights of different monitoring factors and different frequency intervals in the fusion process.

[0111] The factor-band separation data tensor is weighted channel by channel according to the factor-band weight matrix to obtain a gated weighted tensor;

[0112] Each element of the gated weighted tensor is the factor-frequency band separation data corresponding to the monitoring factor, frequency interval and time point multiplied by the factor-frequency band weight corresponding to the monitoring factor and frequency interval.

[0113] Add the gated weighted tensor and the multi-scale prediction tensor element-wise to form a residual gated fusion feature tensor;

[0114] Each element of the residual gated fusion feature tensor is the sum of the elements of the multi-scale prediction tensor and the elements of the gated weighted tensor under the same index. The residual gated fusion feature tensor is consistent with the original monitoring data in terms of monitoring factor, frequency interval and time point dimensions.

[0115] This implementation method enables full-process adaptation of threshold generation through a residual gated fusion structure, improves the personalization and interpretability of threshold decisions, adopts a dual-gated structure at the factor level and frequency band level, and combines residual connections to deeply fuse multi-scale prediction components with factor-frequency band separation data tensors to generate a residual gated fusion feature tensor, and obtains the weight distribution of monitoring factors based on weight normalization. By introducing the similarity matching and deviation adjustment mechanism of historical near-instability samples, the dynamic upper threshold and dynamic lower threshold can automatically shrink or expand according to changes in landslide risk, effectively avoiding the problems of static threshold hysteresis and misjudgment under extreme working conditions.

[0116] In this embodiment, the generation of the dynamic landslide warning threshold includes:

[0117] The time-varying importance distribution of landslide monitoring factors in each frequency interval is extracted based on the residual gated fusion feature tensor, and the weight distribution tensor of monitoring factors is defined.

[0118] The absolute value of each monitoring factor in the residual gated fusion feature tensor at each frequency interval and each time point is extracted, and the absolute values ​​of all monitoring factors and all frequency intervals at the same time point are summed up. The absolute value of each monitoring factor at each frequency interval and each time point is divided by the sum of the absolute values ​​of all monitoring factors and all frequency intervals at the time point to obtain the monitoring factor weight distribution tensor. Each element of the monitoring factor weight distribution tensor is used to measure the importance of the monitoring factor to the overall landslide monitoring status at this frequency interval and this time point.

[0119] All monitoring factors and elements of all frequency intervals of the monitoring factor weight distribution tensor at the most recent time point are arranged in order to form the current landslide-induced state vector; all monitoring factors and elements of all frequency intervals of each monitoring factor weight distribution tensor stored in the historical near-instability landslide monitoring sample at the corresponding time point are arranged in order to form the corresponding historical near-instability landslide monitoring sample state vector;

[0120] Calculate the cosine similarity between the current landslide-induced state vector and the state vector of each historical near-instability landslide monitoring sample;

[0121] The numerator of the cosine similarity is the sum of the products of the corresponding elements of the two state vectors, and the denominator is the product of the square root of the sum of the squares of the current landslide-induced state vector and the square root of the sum of the squares of the state vectors of the historical near-instability landslide monitoring samples. The cosine similarity is used to measure the overall similarity between the current landslide-induced state and the state of each historical near-instability landslide monitoring sample in multiple factor frequency bands.

[0122] Select the maximum value among all cosine similarities and compare it with the preset landslide state trigger threshold:

[0123] If the maximum cosine similarity is greater than the preset landslide state trigger threshold, the deviation adjustment mechanism is triggered to shrink the interval between the dynamic upper threshold and the dynamic lower threshold;

[0124] If the maximum cosine similarity is not greater than the preset landslide state triggering threshold, the deviation adjustment mechanism will not be triggered, and the conventional rolling statistical strategy will continue to be used to calculate the dynamic threshold;

[0125] In the rolling statistics strategy, a sliding time window of fixed length is set, and the average value of the residual gated fusion feature tensor of each monitoring factor in each frequency interval within the sliding time window is calculated to obtain the predicted mean tensor, and the predicted standard deviation tensor is calculated at the same time;

[0126] The predicted standard deviation tensor is calculated as the square root of the mean of the square of the difference between the residual gated fusion feature tensor and its average value for each monitoring factor in each frequency interval within the sliding time window. The predicted mean tensor and the predicted standard deviation tensor represent the average monitoring state and fluctuation intensity of the monitoring factor in the frequency interval within the sliding window, respectively.

[0127] When the maximum cosine similarity is greater than the preset landslide state trigger threshold, the first set of upper and lower threshold adjustment coefficients are used. The predicted mean tensor plus the product of the first set of upper and lower threshold adjustment coefficients and the predicted standard deviation tensor is used as the 1-D dynamic upper threshold. The predicted mean tensor minus the product of the first set of upper and lower threshold adjustment coefficients and the predicted standard deviation tensor is used as the 1-D dynamic lower threshold, so that the interval between the 1-D dynamic upper threshold and the 1-D dynamic lower threshold is contracted.

[0128] When the maximum cosine similarity is not greater than the preset landslide state trigger threshold, the second set of upper and lower threshold adjustment coefficients are used, and the product of the predicted mean tensor plus the second set of upper and lower threshold adjustment coefficients and the predicted standard deviation tensor is used as the 2-D dynamic upper threshold, and the product of the predicted mean tensor minus the second set of upper and lower threshold adjustment coefficients and the predicted standard deviation tensor is used as the 2-D dynamic lower threshold, so that the interval between the 2-D dynamic upper threshold and the 2-D dynamic lower threshold is expanded;

[0129] The dynamic upper thresholds and dynamic lower thresholds of all monitoring factors in all frequency intervals are uniformly combined to form a dynamic landslide warning threshold.

[0130] In another suboptimal embodiment, the generation of a dynamic landslide warning threshold includes:

[0131] The time-varying importance distribution of landslide monitoring factors in each frequency interval is extracted based on the residual gated fusion feature tensor, and the weight distribution tensor of monitoring factors is defined.

[0132] The absolute value of each monitoring factor in the residual gated fusion feature tensor at each frequency interval and each time point is extracted, and the absolute values ​​of all monitoring factors and all frequency intervals at the same time point are summed up. The absolute value of each monitoring factor at each frequency interval and each time point is divided by the sum of the absolute values ​​of all monitoring factors and all frequency intervals at the time point to obtain the monitoring factor weight distribution tensor. Each element of the monitoring factor weight distribution tensor is used to measure the importance of the monitoring factor to the overall landslide monitoring status at this frequency interval and this time point.

[0133] The ROC curve method was used to analyze the influence of each monitoring factor weight distribution tensor on the landslide discrimination results in all samples, and the AUC value was calculated. The input variables with significant landslide early warning discrimination were screened out with AUC ≥ 0.8 as the judgment standard.

[0134] For monitoring factors whose AUC meets the conditions, the Youden index is used for univariate threshold discrimination, and the monitoring factor value corresponding to the maximum value of the Youden index is used as the one-dimensional threshold.

[0135] For combinations of monitoring factors with significant interactive effects, a 2-D input space consisting of two optimal monitoring factors is constructed, and classification and discrimination are performed using classification algorithms including Bayesian networks, decision trees, K-nearest neighbors, binary logistic regression, neural networks, random forests or support vector machines. The optimal classification model is screened with the maximum prediction accuracy, the segmentation boundary corresponding to the two-dimensional threshold is obtained, and the confidence or probability of each prediction result is output.

[0136] The two monitoring factors that have the greatest impact on the accuracy of the results in the optimal binary classifier are used as the horizontal and vertical axes, and all samples are assigned corresponding horizontal and vertical coordinates. The confidence or probability value is used as the color grading standard of the point, and a two-dimensional contour map is drawn. According to the contour line distribution and auxiliary lines, the corresponding critical combination relationship or functional relationship between variables is identified to determine the 2-D threshold.

[0137] If it is necessary to introduce the interaction of three variables, the weight distribution tensors of the three main monitoring factors are normalized based on the prediction results of the optimal binary classifier, and are added and reassigned according to the requirements of the ternary phase diagram, mapped to the equilateral triangle coordinates, and a ternary contour map is drawn. It is graded according to confidence or probability to assist in identifying the 3-D threshold under the three-variable combination.

[0138] The cosine similarity between the current landslide-induced state vector and the state vector of the historical near-instability landslide monitoring sample is calculated in parallel, and the maximum value of all cosine similarities is selected and compared with the preset landslide state trigger threshold:

[0139] If the maximum cosine similarity is greater than the preset landslide state trigger threshold, the deviation adjustment mechanism is triggered to shrink the interval between the dynamic upper threshold and the dynamic lower threshold;

[0140] If the maximum cosine similarity is not greater than the preset landslide state triggering threshold, the deviation adjustment mechanism will not be triggered, and the conventional rolling statistical strategy will continue to be used to calculate the dynamic threshold.

[0141] In the rolling statistical strategy, a sliding time window of fixed length is set, and the average value of the residual gated fusion feature tensor of each monitoring factor in each frequency interval within the sliding time window is calculated respectively to obtain the predicted mean tensor and the predicted standard deviation tensor; and combined with the one-dimensional, two-dimensional or three-dimensional variable discrimination results, different upper and lower threshold adjustment coefficients are used to dynamically shrink or expand the threshold interval of each monitoring factor in each frequency interval.

[0142] The dynamic upper thresholds and dynamic lower thresholds of all monitoring factors in all frequency intervals are uniformly combined to form a dynamic landslide warning threshold.

[0143] In this embodiment, the risk warning level determination includes:

[0144] Obtain multi-source landslide monitoring data at the current time point to form the current monitoring data tensor;

[0145] Each element of the current monitoring data tensor represents the monitoring value of a certain monitoring factor at the current time point in a certain frequency interval. The structure of the current monitoring data tensor is consistent with the structure of the dynamic landslide warning threshold.

[0146] For each monitoring factor and each frequency interval, determine which set of dynamic landslide warning threshold intervals should be used:

[0147] When the maximum cosine similarity is greater than the preset landslide state trigger threshold, the 1-D dynamic upper threshold and the 1-D dynamic lower threshold are used as the threshold intervals of this monitoring factor and this frequency interval at the current time point;

[0148] When the maximum cosine similarity is not greater than the preset landslide state trigger threshold, the 2-D dynamic upper threshold and the 2-D dynamic lower threshold are used as the threshold intervals of this monitoring factor and this frequency interval at the current time point;

[0149] Compare the current monitoring value with the determined dynamic upper and lower thresholds:

[0150] When the current monitoring value is greater than the corresponding dynamic upper threshold, it is recorded as an upper limit;

[0151] When the current monitoring value is less than the corresponding dynamic lower threshold, it is recorded as a lower limit;

[0152] When the current monitoring value is between the dynamic upper threshold and the dynamic lower threshold, it is recorded as a normal state;

[0153] Count the number of times that all monitoring factors exceed the upper limit and the lower limit in all frequency intervals to obtain the total number of violations; count the maximum number of continuous exceeding factors of all monitoring factors in all frequency intervals to obtain the maximum number of continuous exceeding factors; count the exceeding range between the monitoring values ​​of all monitoring factors in all frequency intervals and the corresponding dynamic upper threshold or dynamic lower threshold to obtain the maximum exceeding range of a single monitoring factor, and make a risk warning level judgment, and output the blue warning level, yellow warning level, orange warning level and red warning level.

[0154] In this implementation, the risk warning level is determined based on:

[0155] When the current monitoring values ​​of all monitoring factors in all frequency intervals are within the dynamic threshold range, a blue warning level is output;

[0156] When the total number of over-limit factors is less than the first boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, and the maximum number of no consecutive over-limit factors exceeds the boundary threshold, a yellow warning level is output;

[0157] When the total number of over-limit factors is greater than or equal to the first boundary coefficient of the risk warning level multiplied by the total number of monitoring channels and less than the second boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, or the maximum number of consecutive over-limit factors exceeds the boundary threshold, an orange warning level is output;

[0158] When the total number of excess limits is greater than or equal to the second boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, or the maximum excess limit of a single monitoring factor is greater than the extremely high excess limit threshold of a single monitoring factor, a red warning level is output.

[0159] The first dividing coefficient of the risk warning level refers to the dividing standard coefficient used to distinguish between the yellow and orange risk warning levels. The second dividing coefficient of the risk warning level refers to the dividing standard coefficient used to distinguish between the orange and red risk warning levels. The total number of monitoring channels refers to the total number formed by the combination of all monitoring factors and all frequency intervals. The total number of monitoring channels is equal to the product of the number of monitoring factors and the number of frequency intervals.

[0160] This implementation method significantly improves the accuracy and response efficiency of reservoir area landslide safety monitoring through the linkage between dynamic landslide warning threshold tensors and intelligent risk level grading. By performing sliding window statistics on the residual gated fusion feature tensor and configuring multiple sets of threshold adjustment coefficients, it generates 1-D dynamic upper thresholds, 2-D dynamic upper thresholds, 1-D dynamic lower thresholds, and 2-D dynamic lower thresholds in real time. It also combines current monitoring data to perform multi-factor and multi-band channel-by-channel dynamic over-limit judgment, forming a four-level intelligent risk grading system of blue, yellow, orange, and red. This ensures that under high-risk conditions such as extreme water level fluctuations or heavy rainfall, the early warning system can respond quickly and output clear warning levels, thus realizing all-weather, intelligent safety monitoring of landslide disasters in high dams and large reservoirs.

[0161] Example 1: In a canyon hydropower project reservoir in western China, an on-duty worker at the project safety monitoring center noticed abnormal fluctuations in data from the landslide monitoring point P6, numbered "LS-003," during a routine inspection of the system. The reservoir area employed the system of the present invention.

[0162] In Example 1, the data reported by monitoring point P6 are: reservoir water level 1141.1 meters, rainfall 0 mm, pore pressure 441 kPa, displacement 2.1 mm, and microseismic amplitude 1.05. The system automatically performs a fast Fourier transform on this time series data using the factor spectrum analysis layer, separating components in the trend frequency band, low-frequency periodic frequency band, and high-frequency mutation frequency band. The FEDformer frequency domain decoupling module further aggregates and integrates the component features with historical deformation and instability data from the LS-003 landslide during the period "2024-09-xx to 2024-09-xy." The residual gating layer increases the contribution of reservoir water level from 24% to 32% and the contribution of microseismic from 9% to 21%, based on the sudden increase in the gating weight of P6 in the high-frequency mutation frequency band within the last six hours.

[0163] Influenced by upstream regulation, the reservoir water level rose to 1142.7 meters in just four hours. Meanwhile, the three-hour cumulative rainfall jumped to 16 mm. The displacement at point P6 rose to 3.6 mm, the microseismic amplitude reached 1.12, and the pore pressure reached 456 kPa. The system automatically performed rolling statistics on the residual gated fusion feature tensors for six factors within a 60-hour window, calculating the 1-D dynamic upper threshold for the high-frequency mutation component at P6 to be 3.8 mm and the 1-D dynamic lower threshold to be -3.7 mm. At this time, the traditional static threshold model still maintained 8.0 mm. The current monitored value at point P6 approached the 1-D dynamic upper threshold. The system simultaneously compared the current gated fusion vector at P6 with historical landslide acceleration samples, obtaining a high similarity of 0.92 (higher than the trigger threshold of 0.85). The system immediately narrowed the dynamic upper threshold range and automatically issued a yellow alert to the monitoring center platform.

[0164] A sudden downpour caused dramatic fluctuations in P6 monitoring data: the reservoir water level reached 1143.8 meters, rainfall reached 27 mm per hour, pore pressure rose to 464 kPa, displacement jumped to 6.6 mm, and microseismic amplitude increased to 1.26. The FEDformer module recalculated dynamic thresholds for multiple factor components, tightening the 1-D dynamic upper threshold for the high-frequency mutation component at P6 to 4.3 mm, and the 2-D dynamic upper threshold to 5.1 mm. The current value has exceeded the 1-D dynamic upper threshold for two consecutive hours. The system issued an orange alert, indicating that the high-frequency component at P6 has continuously exceeded the limit. The real-time display showed that the reservoir water level and rainfall contribution had risen to 36% and 41%, respectively. The system automatically issued an emergency inspection work order, and engineers quickly responded and conducted an on-site survey of the area surrounding P6.

[0165] As heavy rainfall continued, the displacement data at point P6 jumped to 9.1 mm. The system again compared the data with near-instability samples, and the similarity reached 0.97. The dynamic threshold range was further tightened, with the upper threshold at 4.1 mm. According to the classification criteria, point P6 had exceeded the upper threshold for four consecutive hours, and the cumulative number of exceeded points accounted for 23% of the total number of channels, exceeding the orange-red threshold. The system automatically reported a red risk and notified the dam operation and disaster prevention leadership team. The system automatically generated a detailed risk analysis report, including the distribution of factor gate weights, the frequency interval exceeded distribution, and a comparative analysis of this event with multiple historical deformation events.

[0166] During the same period, traditional static thresholds (reservoir water level, displacement experience thresholds) were used to judge the data of the same period. Point P6 was not marked as abnormal by the system. A yellow warning was only prompted when the displacement exceeded 8 mm (which was already the actual acceleration stage). The red warning was pushed one day later.

[0167] According to the system recommendations of the present invention, the on-site disposal team carried out emergency reinforcement on the newly discovered seepage area and two new cracks above point P6 and successfully avoided the occurrence of a large-scale instability incident.

[0168] Comparing historical monitoring data from the entire reservoir area, the system of the present invention generated 64 cumulative warnings, including 62 effective warnings, 2 false alarms, and 1 missed alarm. The traditional static threshold method generated 44 cumulative warnings, including 32 effective warnings, 7 false alarms, and 5 missed alarms. The accuracy of landslide deformation warnings for the high dam and large reservoir area reached 97%, compared to 84% for the traditional method. The average lead time was 11 hours compared to 4 hours, respectively.

[0169] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for determining the threshold value of landslide factors for early warning in high dam and large reservoir areas based on deep learning, characterized by: include: Collect the multi-source landslide monitoring raw data in the high dam and large reservoir landslide monitoring area, and perform preprocessing operations to obtain a multi-source landslide monitoring dataset with a unified structure; The multi-source landslide monitoring data set is input into the factor spectrum analysis layer, and each monitoring factor is subjected to fast Fourier transform. The data are then divided into trend frequency band data, low-frequency periodic frequency band data, and high-frequency mutation frequency band data according to the preset frequency band pass rule to form a factor-frequency band separation data tensor. The factor-frequency band separation data tensor is input into the FEDformer frequency domain decoupling module, which decouples the long sequence and outputs the trend component prediction sequence, the period component prediction sequence and the disturbance component prediction sequence, which together constitute the multi-scale prediction component. The multi-scale prediction component and the factor-band separation data tensor are input into the residual gated fusion layer to generate the factor-band weight matrix. The input tensor and the gated weighted tensor are fused based on the residual connection structure to output the residual gated fusion feature tensor. In the dynamic threshold generation module, the weight distribution of monitoring factors is extracted based on the residual gated fusion feature tensor, and similarity matching is performed with the stored historical near-instability landslide monitoring samples to form a dynamic landslide warning threshold; Compare and judge the real-time updated multi-source landslide monitoring data with the dynamic landslide warning threshold, and divide the risk warning level according to the preset risk warning level; The generation of the dynamic landslide warning threshold specifically includes: The time-varying importance distribution of landslide monitoring factors in each frequency interval is extracted based on the residual gated fusion feature tensor, and the weight distribution tensor of monitoring factors is defined. All monitoring factors and elements of all frequency intervals of the monitoring factor weight distribution tensor at the most recent time point are arranged in order to form the current landslide-induced state vector; all monitoring factors and elements of all frequency intervals of each monitoring factor weight distribution tensor stored in the historical near-instability landslide monitoring sample at the corresponding time point are arranged in order to form the corresponding historical near-instability landslide monitoring sample state vector; Calculate the cosine similarity between the current landslide-induced state vector and the state vector of each historical near-instability landslide monitoring sample; Select the maximum value among all cosine similarities and compare it with the preset landslide state trigger threshold: If the maximum cosine similarity is greater than the preset landslide state trigger threshold, the deviation adjustment mechanism is triggered to shrink the interval between the dynamic upper threshold and the dynamic lower threshold; If the maximum cosine similarity is not greater than the preset landslide state triggering threshold, the deviation adjustment mechanism will not be triggered, and the conventional rolling statistical strategy will continue to be used to calculate the dynamic threshold; In the rolling statistics strategy, a sliding time window of fixed length is set, and the average value of the residual gated fusion feature tensor of each monitoring factor in each frequency interval within the sliding time window is calculated to obtain the predicted mean tensor, and the predicted standard deviation tensor is calculated at the same time; When the maximum cosine similarity is greater than the preset landslide state trigger threshold, the first set of upper and lower threshold adjustment coefficients are used to determine the 1-D dynamic upper threshold and the 1-D dynamic lower threshold, so that the interval between the 1-D dynamic upper threshold and the 1-D dynamic lower threshold is reduced; When the maximum cosine similarity is not greater than the preset landslide state triggering threshold, the second set of upper and lower threshold adjustment coefficients are used to determine the 2-D dynamic upper threshold and the 2-D dynamic lower threshold, so that the interval between the 2-D dynamic upper threshold and the 2-D dynamic lower threshold is expanded; The dynamic upper thresholds and dynamic lower thresholds of all monitoring factors in all frequency intervals are uniformly combined to form a dynamic landslide warning threshold.

2. The method for determining the threshold value of landslide factors in high dam and large reservoir area based on deep learning according to claim 1 is characterized in that: The multi-source landslide monitoring raw data includes reservoir water level raw data, rainfall raw data, groundwater pressure raw data, surface displacement raw data and microseismic raw data, and the preprocessing operations include time-space alignment, missing segment interpolation, outlier removal and noise suppression operations.

3. The method for determining the threshold value of landslide factors in high dam and large reservoir area based on deep learning according to claim 1 is characterized in that: The division of trend frequency band data, low-frequency cycle frequency band data and high-frequency mutation frequency band data includes: Each monitoring factor in the multi-source landslide monitoring dataset is taken as input, and each monitoring factor has a corresponding time series data sequence; Perform spectrum analysis on the time series data of each monitoring factor to obtain the amplitude response sequence of the monitoring factor at different frequencies; The frequency domain amplitude response sequence of each monitoring factor is divided into frequency intervals according to the preset frequency threshold. The frequency intervals include trend frequency bands, low-frequency periodic frequency bands and high-frequency mutation frequency bands. In each frequency interval, the frequency domain amplitude response sequence of each monitoring factor is subjected to bandpass filtering to extract the amplitude response component of the monitoring factor in the current frequency interval; The amplitude response component of each monitoring factor in each frequency interval obtained by bandpass filtering is restored from the frequency domain to the time domain to obtain the time domain component corresponding to each monitoring factor in different frequency intervals; The time domain components of all monitoring factors in all frequency intervals are uniformly combined to obtain the factor-frequency band separation data tensor.

4. The method for determining the threshold value of landslide factors in high dam and large reservoir area based on deep learning according to claim 1 is characterized in that: The FEDformer frequency domain decoupling module includes: Based on the characteristics of different sensitivity of landslide factors in different frequency bands, the discrete Fourier transform operation is performed on the factor-frequency band separation data tensor along the time dimension to obtain the frequency domain representation tensor. Calculate the frequency domain interactive self-attention weight based on the frequency domain representation tensor; The frequency domain representation tensor is dynamically weighted and fused across frequency bands using the frequency domain interactive self-attention weights to obtain the frequency domain enhanced feature representation tensor. The frequency domain enhanced feature representation tensor is restored from the frequency domain to the time domain through the inverse Fourier transform to obtain the time domain enhanced feature data tensor; Based on the time domain enhanced feature data tensor, multi-scale feature aggregation is performed on the trend frequency band, low-frequency periodic frequency band and high-frequency mutation frequency band respectively, and the trend component prediction sequence, periodic component prediction sequence and disturbance component prediction sequence are output respectively, and collectively used as the multi-scale prediction component.

5. The method for determining the threshold value of landslide factors for early warning of high dam and large reservoir area based on deep learning according to claim 1 is characterized in that: The calculation of the residual gated fusion feature tensor includes: Constructing a multi-scale prediction tensor based on the temporal consistency between the multi-scale prediction components and the factor-band separation data tensor; Calculate the factor-level gating coefficient for each monitored factor; Calculate the band-level gating coefficient for each frequency interval; Construct a factor-band weight matrix based on the factor-level gating coefficient and the band-level gating coefficient; The factor-band separation data tensor is weighted channel by channel according to the factor-band weight matrix to obtain a gated weighted tensor; The gated weighted tensor is element-wise added to the multi-scale prediction tensor to form a residual gated fusion feature tensor.

6. The method for determining the threshold value of landslide factors in high dam and large reservoir area based on deep learning according to claim 1 is characterized in that: The risk warning level judgment includes: Obtain multi-source landslide monitoring data at the current time point to form the current monitoring data tensor; For each monitoring factor and each frequency interval, determine which set of dynamic landslide warning threshold intervals should be used: When the maximum cosine similarity is greater than the preset landslide state trigger threshold, the 1-D dynamic upper threshold and the 1-D dynamic lower threshold are used as the threshold intervals of this monitoring factor and this frequency interval at the current time point; When the maximum cosine similarity is not greater than the preset landslide state trigger threshold, the 2-D dynamic upper threshold and the 2-D dynamic lower threshold are used as the threshold intervals of this monitoring factor and this frequency interval at the current time point; Compare the current monitoring value with the determined dynamic upper and lower thresholds: When the current monitoring value is greater than the corresponding dynamic upper threshold, it is recorded as an upper limit; When the current monitoring value is less than the corresponding dynamic lower threshold, it is recorded as a lower limit; When the current monitoring value is between the dynamic upper threshold and the dynamic lower threshold, it is recorded as a normal state; Count the number of times that all monitoring factors exceed the upper limit and the lower limit in all frequency intervals to obtain the total number of violations; count the maximum number of continuous exceeding factors of all monitoring factors in all frequency intervals to obtain the maximum number of continuous exceeding factors; count the exceeding range between the monitoring values ​​of all monitoring factors in all frequency intervals and the corresponding dynamic upper threshold or dynamic lower threshold to obtain the maximum exceeding range of a single monitoring factor, and make a risk warning level judgment, and output the blue warning level, yellow warning level, orange warning level and red warning level.

7. The method for determining the threshold value of landslide factors for early warning of high dam and large reservoir area based on deep learning according to claim 6 is characterized in that: The risk warning level is determined based on: When the current monitoring values ​​of all monitoring factors in all frequency intervals are within the dynamic threshold range, a blue warning level is output; When the total number of over-limit factors is less than the first boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, and the maximum number of no consecutive over-limit factors exceeds the boundary threshold, a yellow warning level is output; When the total number of over-limit factors is greater than or equal to the first boundary coefficient of the risk warning level multiplied by the total number of monitoring channels and less than the second boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, or the maximum number of consecutive over-limit factors exceeds the boundary threshold, an orange warning level is output; When the total number of over-limits is greater than or equal to the second boundary coefficient of the risk warning level multiplied by the total number of monitoring channels, or the maximum over-limit amplitude of a single monitoring factor is greater than the extremely high over-limit threshold of a single monitoring factor, a red warning level is output; The first dividing coefficient of the risk warning level refers to the dividing standard coefficient used to distinguish between the yellow and orange risk warning levels. The second dividing coefficient of the risk warning level refers to the dividing standard coefficient used to distinguish between the orange and red risk warning levels. The total number of monitoring channels refers to the total number formed by the combination of all monitoring factors and all frequency intervals.

Images